diff --git "a/test.csv" "b/test.csv"
deleted file mode 100644--- "a/test.csv"
+++ /dev/null
@@ -1,61 +0,0 @@
-query,positive,negative
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I need to train a GAN that can generate high-quality outputs for various applications. To do this, I need a system that is able to store and manage the massive amount of training data. Since I have sufficient computing resources, my plan is to use distributed training methods to handle the huge amount of data. In order to minimize the communication cost in the distributed system, I think an algorithm using quantized gradient might work. Additionally, I also want the algorithm to be able to avoid convergence issues, which I believe can be handled by the gradient compression method. With such an algorithm, I can have a GAN that is able to be trained on various dataset including wikipedia, github and other sources from the internet, and try to develop a model that is able to have conversations with human naturally.","['Training deep neural networks on large datasets containing high-dimensional\ndata requires a large amount of computation. A solution to this problem is\ndata-parallel distributed training, where a model is replicated into several\ncomputational nodes that have access to different chunks of the data. This\napproach, however, entails high communication rates and latency because of the\ncomputed gradients that need to be shared among nodes at every iteration. The\nproblem becomes more pronounced in the case that there is wireless\ncommunication between the nodes (i.e. due to the limited network bandwidth). To\naddress this problem, various compression methods have been proposed including\nsparsification, quantization, and entropy encoding of the gradients. Existing\nmethods leverage the intra-node information redundancy, that is, they compress\ngradients at each node independently. In contrast, we advocate that the\ngradients across the nodes are correlated and propose methods to leverage this\ninter-node redundancy to improve compression efficiency. Depending on the node\ncommunication protocol (parameter server or ring-allreduce), we propose two\ninstances of the LGC approach that we coin Learned Gradient Compression (LGC).\nOur methods exploit an autoencoder (i.e. trained during the first stages of the\ndistributed training) to capture the common information that exists in the\ngradients of the distributed nodes. We have tested our LGC methods on the image\nclassification and semantic segmentation tasks using different convolutional\nneural networks (ResNet50, ResNet101, PSPNet) and multiple datasets (ImageNet,\nCifar10, CamVid). The ResNet101 model trained for image classification on\nCifar10 achieved an accuracy of 93.57%, which is lower than the baseline\ndistributed training with uncompressed gradients only by 0.18%.', ""A large portion of data mining and analytic services use modern machine\nlearning techniques, such as deep learning. The state-of-the-art results by\ndeep learning come at the price of an intensive use of computing resources. The\nleading frameworks (e.g., TensorFlow) are executed on GPUs or on high-end\nservers in datacenters. On the other end, there is a proliferation of personal\ndevices with possibly free CPU cycles; this can enable services to run in\nusers' homes, embedding machine learning operations. In this paper, we ask the\nfollowing question: Is distributed deep learning computation on WAN connected\ndevices feasible, in spite of the traffic caused by learning tasks? We show\nthat such a setup rises some important challenges, most notably the ingress\ntraffic that the servers hosting the up-to-date model have to sustain.\n  In order to reduce this stress, we propose adaComp, a novel algorithm for\ncompressing worker updates to the model on the server. Applicable to stochastic\ngradient descent based approaches, it combines efficient gradient selection and\nlearning rate modulation. We then experiment and measure the impact of\ncompression, device heterogeneity and reliability on the accuracy of learned\nmodels, with an emulator platform that embeds TensorFlow into Linux containers.\nWe report a reduction of the total amount of data sent by workers to the server\nby two order of magnitude (e.g., 191-fold reduction for a convolutional network\non the MNIST dataset), when compared to a standard asynchronous stochastic\ngradient descent, while preserving model accuracy."", 'Although distributed machine learning has opened up many new and exciting\nresearch frontiers, fragmentation of models and data across different machines,\nnodes, and sites still results in considerable communication overhead, impeding\nreliable training in real-world contexts.\n  The focus on gradients as the primary shared statistic during training has\nspawned a number of intuitive algorithms for distributed deep learning;\nhowever, gradient-centric training of large deep neural networks (DNNs) tends\nto be communication-heavy, often requiring additional adaptations such as\nsparsity constraints, compression, quantization, and more, to curtail\nbandwidth.\n  We introduce an innovative, communication-friendly approach for training\ndistributed DNNs, which capitalizes on the outer-product structure of the\ngradient as revealed by the mechanics of auto-differentiation. The exposed\nstructure of the gradient evokes a new class of distributed learning algorithm,\nwhich is naturally more communication-efficient than full gradient sharing. Our\napproach, called distributed auto-differentiation (dAD), builds off a marriage\nof rank-based compression and the innate structure of the gradient as an\nouter-product. We demonstrate that dAD trains more efficiently than other state\nof the art distributed methods on modern architectures, such as transformers,\nwhen applied to large-scale text and imaging datasets. The future of\ndistributed learning, we determine, need not be dominated by gradient-centric\nalgorithms.', 'Distributed training is an effective way to accelerate the training process\nof large-scale deep learning models. However, the parameter exchange and\nsynchronization of distributed stochastic gradient descent introduce a large\namount of communication overhead. Gradient compression is an effective method\nto reduce communication overhead. In synchronization SGD compression methods,\nmany Top-k sparsification based gradient compression methods have been proposed\nto reduce the communication. However, the centralized method based on the\nparameter servers has the single point of failure problem and limited\nscalability, while the decentralized method with global parameter exchanging\nmay reduce the convergence rate of training. In contrast with Top-$k$ based\nmethods, we proposed a gradient compression method with globe gradient vector\nsketching, which uses the Count-Sketch structure to store the gradients to\nreduce the loss of the accuracy in the training process, named global-sketching\nSGD (gs-SGD). The gs-SGD has better convergence efficiency on deep learning\nmodels and a communication complexity of O($\\log d*\\log P$), where $d$ is the\nnumber of model parameters and P is the number of workers. We conducted\nexperiments on GPU clusters to verify that our method has better convergence\nefficiency than global Top-$k$ and Sketching-based methods. In addition, gs-SGD\nachieves 1.3-3.1x higher throughput compared with gTop-$k$, and 1.1-1.2x higher\nthroughput compared with original Sketched-SGD.', 'Large-scale distributed training of Deep Neural Networks (DNNs) on\nstate-of-the-art platforms is expected to be severely communication\nconstrained. To overcome this limitation, numerous gradient compression\ntechniques have been proposed and have demonstrated high compression ratios.\nHowever, most existing methods do not scale well to large scale distributed\nsystems (due to gradient build-up) and/or fail to evaluate model fidelity (test\naccuracy) on large datasets. To mitigate these issues, we propose a new\ncompression technique, Scalable Sparsified Gradient Compression (ScaleCom),\nthat leverages similarity in the gradient distribution amongst learners to\nprovide significantly improved scalability. Using theoretical analysis, we show\nthat ScaleCom provides favorable convergence guarantees and is compatible with\ngradient all-reduce techniques. Furthermore, we experimentally demonstrate that\nScaleCom has small overheads, directly reduces gradient traffic and provides\nhigh compression rates (65-400X) and excellent scalability (up to 64 learners\nand 8-12X larger batch sizes over standard training) across a wide range of\napplications (image, language, and speech) without significant accuracy loss.', 'Large amount of data is often required to train and deploy useful machine\nlearning models in industry. Smaller enterprises do not have the luxury of\naccessing enough data for machine learning, For privacy sensitive fields such\nas banking, insurance and healthcare, aggregating data to a data warehouse\nposes a challenge of data security and limited computational resources. These\nchallenges are critical when developing machine learning algorithms in\nindustry. Several attempts have been made to address the above challenges by\nusing distributed learning techniques such as federated learning over disparate\ndata stores in order to circumvent the need for centralised data aggregation.\nThis paper proposes an improved algorithm to securely train deep neural\nnetworks over several data sources in a distributed way, in order to eliminate\nthe need to centrally aggregate the data and the need to share the data thus\npreserving privacy. The proposed method allows training of deep neural networks\nusing data from multiple de-linked nodes in a distributed environment and to\nsecure the representation shared during training. Only a representation of the\ntrained models (network architecture and weights) are shared. The algorithm was\nevaluated on existing healthcare patients data and the performance of this\nimplementation was compared to that of a regular deep neural network trained on\na single centralised architecture. This algorithm will pave a way for\ndistributed training of neural networks on privacy sensitive applications where\nraw data may not be shared directly or centrally aggregating this data in a\ndata warehouse is not feasible.', 'With the rapid increase of big data, distributed Machine Learning (ML) has\nbeen widely applied in training large-scale models. Stochastic Gradient Descent\n(SGD) is arguably the workhorse algorithm of ML. Distributed ML models trained\nby SGD involve large amounts of gradient communication, which limits the\nscalability of distributed ML. Thus, it is important to compress the gradients\nfor reducing communication. In this paper, we propose FastSGD, a Fast\ncompressed SGD framework for distributed ML. To achieve a high compression\nratio at a low cost, FastSGD represents the gradients as key-value pairs, and\ncompresses both the gradient keys and values in linear time complexity. For the\ngradient value compression, FastSGD first uses a reciprocal mapper to transform\noriginal values into reciprocal values, and then, it utilizes a logarithm\nquantization to further reduce reciprocal values to small integers. Finally,\nFastSGD filters reduced gradient integers by a given threshold. For the\ngradient key compression, FastSGD provides an adaptive fine-grained delta\nencoding method to store gradient keys with fewer bits. Extensive experiments\non practical ML models and datasets demonstrate that FastSGD achieves the\ncompression ratio up to 4 orders of magnitude, and accelerates the convergence\ntime up to 8x, compared with state-of-the-art methods.', 'Large-scale distributed training requires significant communication bandwidth\nfor gradient exchange that limits the scalability of multi-node training, and\nrequires expensive high-bandwidth network infrastructure. The situation gets\neven worse with distributed training on mobile devices (federated learning),\nwhich suffers from higher latency, lower throughput, and intermittent poor\nconnections. In this paper, we find 99.9% of the gradient exchange in\ndistributed SGD is redundant, and propose Deep Gradient Compression (DGC) to\ngreatly reduce the communication bandwidth. To preserve accuracy during\ncompression, DGC employs four methods: momentum correction, local gradient\nclipping, momentum factor masking, and warm-up training. We have applied Deep\nGradient Compression to image classification, speech recognition, and language\nmodeling with multiple datasets including Cifar10, ImageNet, Penn Treebank, and\nLibrispeech Corpus. On these scenarios, Deep Gradient Compression achieves a\ngradient compression ratio from 270x to 600x without losing accuracy, cutting\nthe gradient size of ResNet-50 from 97MB to 0.35MB, and for DeepSpeech from\n488MB to 0.74MB. Deep gradient compression enables large-scale distributed\ntraining on inexpensive commodity 1Gbps Ethernet and facilitates distributed\ntraining on mobile. Code is available at:\nhttps://github.com/synxlin/deep-gradient-compression.', 'Communication overhead severely hinders the scalability of distributed\nmachine learning systems. Recently, there has been a growing interest in using\ngradient compression to reduce the communication overhead of the distributed\ntraining. However, there is little understanding of applying gradient\ncompression to adaptive gradient methods. Moreover, its performance benefits\nare often limited by the non-negligible compression overhead. In this paper, we\nfirst introduce a novel adaptive gradient method with gradient compression. We\nshow that the proposed method has a convergence rate of\n$\\mathcal{O}(1/\\sqrt{T})$ for non-convex problems. In addition, we develop a\nscalable system called BytePS-Compress for two-way compression, where the\ngradients are compressed in both directions between workers and parameter\nservers. BytePS-Compress pipelines the compression and decompression on CPUs\nand achieves a high degree of parallelism. Empirical evaluations show that we\nimprove the training time of ResNet50, VGG16, and BERT-base by 5.0%, 58.1%,\n23.3%, respectively, without any accuracy loss with 25 Gb/s networking.\nFurthermore, for training the BERT models, we achieve a compression rate of\n333x compared to the mixed-precision training.', ""Communication overhead is a major bottleneck hampering the scalability of\ndistributed machine learning systems. Recently, there has been a surge of\ninterest in using gradient compression to improve the communication efficiency\nof distributed neural network training. Using 1-bit quantization, signSGD with\nmajority vote achieves a 32x reduction on communication cost. However, its\nconvergence is based on unrealistic assumptions and can diverge in practice. In\nthis paper, we propose a general distributed compressed SGD with Nesterov's\nmomentum. We consider two-way compression, which compresses the gradients both\nto and from workers. Convergence analysis on nonconvex problems for general\ngradient compressors is provided. By partitioning the gradient into blocks, a\nblockwise compressor is introduced such that each gradient block is compressed\nand transmitted in 1-bit format with a scaling factor, leading to a nearly 32x\nreduction on communication. Experimental results show that the proposed method\nconverges as fast as full-precision distributed momentum SGD and achieves the\nsame testing accuracy. In particular, on distributed ResNet training with 7\nworkers on the ImageNet, the proposed algorithm achieves the same testing\naccuracy as momentum SGD using full-precision gradients, but with $46\\%$ less\nwall clock time."", 'Gradient quantization is an emerging technique in reducing communication\ncosts in distributed learning. Existing gradient quantization algorithms often\nrely on engineering heuristics or empirical observations, lacking a systematic\napproach to dynamically quantize gradients. This paper addresses this issue by\nproposing a novel dynamically quantized SGD (DQ-SGD) framework, enabling us to\ndynamically adjust the quantization scheme for each gradient descent step by\nexploring the trade-off between communication cost and convergence error. We\nderive an upper bound, tight in some cases, of the convergence error for a\nrestricted family of quantization schemes and loss functions. We design our\nDQ-SGD algorithm via minimizing the communication cost under the convergence\nerror constraints. Finally, through extensive experiments on large-scale\nnatural language processing and computer vision tasks on AG-News, CIFAR-10, and\nCIFAR-100 datasets, we demonstrate that our quantization scheme achieves better\ntradeoffs between the communication cost and learning performance than other\nstate-of-the-art gradient quantization methods.', 'We consider machine learning applications that train a model by leveraging\ndata distributed over a trusted network, where communication constraints can\ncreate a performance bottleneck. A number of recent approaches propose to\novercome this bottleneck through compression of gradient updates. However, as\nmodels become larger, so does the size of the gradient updates. In this paper,\nwe propose an alternate approach to learn from distributed data that quantizes\ndata instead of gradients, and can support learning over applications where the\nsize of gradient updates is prohibitive. Our approach leverages the dependency\nof the computed gradient on data samples, which lie in a much smaller space in\norder to perform the quantization in the smaller dimension data space. At the\ncost of an extra gradient computation, the gradient estimate can be refined by\nconveying the difference between the gradient at the quantized data point and\nthe original gradient using a small number of bits. Lastly, in order to save\ncommunication, our approach adds a layer that decides whether to transmit a\nquantized data sample or not based on its importance for learning. We analyze\nthe convergence of the proposed approach for smooth convex and non-convex\nobjective functions and show that we can achieve order optimal convergence\nrates with communication that mostly depends on the data rather than the model\n(gradient) dimension. We use our proposed algorithm to train ResNet models on\nthe CIFAR-10 and ImageNet datasets, and show that we can achieve an order of\nmagnitude savings over gradient compression methods. These communication\nsavings come at the cost of increasing computation at the learning agent, and\nthus our approach is beneficial in scenarios where communication load is the\nmain problem.', 'Federated learning (FL) is an emerging technique for training machine\nlearning models using geographically dispersed data collected by local\nentities. It includes local computation and synchronization steps. To reduce\nthe communication overhead and improve the overall efficiency of FL, gradient\nsparsification (GS) can be applied, where instead of the full gradient, only a\nsmall subset of important elements of the gradient is communicated. Existing\nwork on GS uses a fixed degree of gradient sparsity for i.i.d.-distributed data\nwithin a datacenter. In this paper, we consider adaptive degree of sparsity and\nnon-i.i.d. local datasets. We first present a fairness-aware GS method which\nensures that different clients provide a similar amount of updates. Then, with\nthe goal of minimizing the overall training time, we propose a novel online\nlearning formulation and algorithm for automatically determining the\nnear-optimal communication and computation trade-off that is controlled by the\ndegree of gradient sparsity. The online learning algorithm uses an estimated\nsign of the derivative of the objective function, which gives a regret bound\nthat is asymptotically equal to the case where exact derivative is available.\nExperiments with real datasets confirm the benefits of our proposed approaches,\nshowing up to $40\\%$ improvement in model accuracy for a finite training time.', ""The performance and efficiency of distributed training of Deep Neural\nNetworks highly depend on the performance of gradient averaging among all\nparticipating nodes, which is bounded by the communication between nodes. There\nare two major strategies to reduce communication overhead: one is to hide\ncommunication by overlapping it with computation, and the other is to reduce\nmessage sizes. The first solution works well for linear neural architectures,\nbut latest networks such as ResNet and Inception offer limited opportunity for\nthis overlapping. Therefore, researchers have paid more attention to minimizing\ncommunication. In this paper, we present a novel gradient compression framework\nderived from insights of real gradient distributions, and which strikes a\nbalance between compression ratio, accuracy, and computational overhead. Our\nframework has two major novel components: sparsification of gradients in the\nfrequency domain, and a range-based floating point representation to quantize\nand further compress gradients frequencies. Both components are dynamic, with\ntunable parameters that achieve different compression ratio based on the\naccuracy requirement and systems' platforms, and achieve very high throughput\non GPUs. We prove that our techniques guarantee the convergence with a\ndiminishing compression ratio. Our experiments show that the proposed\ncompression framework effectively improves the scalability of most popular\nneural networks on a 32 GPU cluster to the baseline of no compression, without\ncompromising the accuracy and convergence speed."", 'Compressed communication, in the form of sparsification or quantization of\nstochastic gradients, is employed to reduce communication costs in distributed\ndata-parallel training of deep neural networks. However, there exists a\ndiscrepancy between theory and practice: while theoretical analysis of most\nexisting compression methods assumes compression is applied to the gradients of\nthe entire model, many practical implementations operate individually on the\ngradients of each layer of the model. In this paper, we prove that layer-wise\ncompression is, in theory, better, because the convergence rate is upper\nbounded by that of entire-model compression for a wide range of biased and\nunbiased compression methods. However, despite the theoretical bound, our\nexperimental study of six well-known methods shows that convergence, in\npractice, may or may not be better, depending on the actual trained model and\ncompression ratio. Our findings suggest that it would be advantageous for deep\nlearning frameworks to include support for both layer-wise and entire-model\ncompression.', 'A standard approach in large scale machine learning is distributed stochastic\ngradient training, which requires the computation of aggregated stochastic\ngradients over multiple nodes on a network. Communication is a major bottleneck\nin such applications, and in recent years, compressed stochastic gradient\nmethods such as QSGD (quantized SGD) and sparse SGD have been proposed to\nreduce communication. It was also shown that error compensation can be combined\nwith compression to achieve better convergence in a scheme that each node\ncompresses its local stochastic gradient and broadcast the result to all other\nnodes over the network in a single pass. However, such a single pass broadcast\napproach is not realistic in many practical implementations. For example, under\nthe popular parameter server model for distributed learning, the worker nodes\nneed to send the compressed local gradients to the parameter server, which\nperforms the aggregation. The parameter server has to compress the aggregated\nstochastic gradient again before sending it back to the worker nodes. In this\nwork, we provide a detailed analysis on this two-pass communication model and\nits asynchronous parallel variant, with error-compensated compression both on\nthe worker nodes and on the parameter server. We show that the\nerror-compensated stochastic gradient algorithm admits three very nice\nproperties: 1) it is compatible with an \\emph{arbitrary} compression technique;\n2) it admits an improved convergence rate than the non error-compensated\nstochastic gradient methods such as QSGD and sparse SGD; 3) it admits linear\nspeedup with respect to the number of workers. The empirical study is also\nconducted to validate our theoretical results.', 'Distributed stochastic gradient descent (SGD) approach has been widely used\nin large-scale deep learning, and the gradient collective method is vital to\nensure the training scalability of the distributed deep learning system.\nCollective communication such as AllReduce has been widely adopted for the\ndistributed SGD process to reduce the communication time. However, AllReduce\nincurs large bandwidth resources while most gradients are sparse in many cases\nsince many gradient values are zeros and should be efficiently compressed for\nbandwidth saving. To reduce the sparse gradient communication overhead, we\npropose Sparse-Sketch Reducer (S2 Reducer), a novel sketch-based sparse\ngradient aggregation method with convergence guarantees. S2 Reducer reduces the\ncommunication cost by only compressing the non-zero gradients with count-sketch\nand bitmap, and enables the efficient AllReduce operators for parallel SGD\ntraining. We perform extensive evaluation against four state-of-the-art methods\nover five training models. Our results show that S2 Reducer converges to the\nsame accuracy, reduces 81\\% sparse communication overhead, and achieves 1.8$\n\\times $ speedup compared to state-of-the-art approaches.']","['Training large neural networks requires distributing learning across multiple\nworkers, where the cost of communicating gradients can be a significant\nbottleneck. signSGD alleviates this problem by transmitting just the sign of\neach minibatch stochastic gradient. We prove that it can get the best of both\nworlds: compressed gradients and SGD-level convergence rate. The relative\n$\\ell_1/\\ell_2$ geometry of gradients, noise and curvature informs whether\nsignSGD or SGD is theoretically better suited to a particular problem. On the\npractical side we find that the momentum counterpart of signSGD is able to\nmatch the accuracy and convergence speed of Adam on deep Imagenet models. We\nextend our theory to the distributed setting, where the parameter server uses\nmajority vote to aggregate gradient signs from each worker enabling 1-bit\ncompression of worker-server communication in both directions. Using a theorem\nby Gauss we prove that majority vote can achieve the same reduction in variance\nas full precision distributed SGD. Thus, there is great promise for sign-based\noptimisation schemes to achieve fast communication and fast convergence. Code\nto reproduce experiments is to be found at https://github.com/jxbz/signSGD .', 'Due to the substantial computational cost, training state-of-the-art deep\nneural networks for large-scale datasets often requires distributed training\nusing multiple computation workers. However, by nature, workers need to\nfrequently communicate gradients, causing severe bottlenecks, especially on\nlower bandwidth connections. A few methods have been proposed to compress\ngradient for efficient communication, but they either suffer a low compression\nratio or significantly harm the resulting model accuracy, particularly when\napplied to convolutional neural networks. To address these issues, we propose a\nmethod to reduce the communication overhead of distributed deep learning. Our\nkey observation is that gradient updates can be delayed until an unambiguous\n(high amplitude, low variance) gradient has been calculated. We also present an\nefficient algorithm to compute the variance with negligible additional cost. We\nexperimentally show that our method can achieve very high compression ratio\nwhile maintaining the result model accuracy. We also analyze the efficiency\nusing computation and communication cost models and provide the evidence that\nthis method enables distributed deep learning for many scenarios with commodity\nenvironments.', 'In this paper, we present a communication-efficient federated learning\nframework inspired by quantized compressed sensing. The presented framework\nconsists of gradient compression for wireless devices and gradient\nreconstruction for a parameter server (PS). Our strategy for gradient\ncompression is to sequentially perform block sparsification, dimensional\nreduction, and quantization. Thanks to gradient sparsification and\nquantization, our strategy can achieve a higher compression ratio than one-bit\ngradient compression. For accurate aggregation of the local gradients from the\ncompressed signals at the PS, we put forth an approximate minimum mean square\nerror (MMSE) approach for gradient reconstruction using the\nexpectation-maximization generalized-approximate-message-passing (EM-GAMP)\nalgorithm. Assuming Bernoulli Gaussian-mixture prior, this algorithm\niteratively updates the posterior mean and variance of local gradients from the\ncompressed signals. We also present a low-complexity approach for the gradient\nreconstruction. In this approach, we use the Bussgang theorem to aggregate\nlocal gradients from the compressed signals, then compute an approximate MMSE\nestimate of the aggregated gradient using the EM-GAMP algorithm. We also\nprovide a convergence rate analysis of the presented framework. Using the MNIST\ndataset, we demonstrate that the presented framework achieves almost identical\nperformance with the case that performs no compression, while significantly\nreducing communication overhead for federated learning.', 'Parallel implementations of stochastic gradient descent (SGD) have received\nsignificant research attention, thanks to excellent scalability properties of\nthis algorithm, and to its efficiency in the context of training deep neural\nnetworks. A fundamental barrier for parallelizing large-scale SGD is the fact\nthat the cost of communicating the gradient updates between nodes can be very\nlarge. Consequently, lossy compression heuristics have been proposed, by which\nnodes only communicate quantized gradients. Although effective in practice,\nthese heuristics do not always provably converge, and it is not clear whether\nthey are optimal.\n  In this paper, we propose Quantized SGD (QSGD), a family of compression\nschemes which allow the compression of gradient updates at each node, while\nguaranteeing convergence under standard assumptions. QSGD allows the user to\ntrade off compression and convergence time: it can communicate a sublinear\nnumber of bits per iteration in the model dimension, and can achieve\nasymptotically optimal communication cost. We complement our theoretical\nresults with empirical data, showing that QSGD can significantly reduce\ncommunication cost, while being competitive with standard uncompressed\ntechniques on a variety of real tasks.\n  In particular, experiments show that gradient quantization applied to\ntraining of deep neural networks for image classification and automated speech\nrecognition can lead to significant reductions in communication cost, and\nend-to-end training time. For instance, on 16 GPUs, we are able to train a\nResNet-152 network on ImageNet 1.8x faster to full accuracy. Of note, we show\nthat there exist generic parameter settings under which all known network\narchitectures preserve or slightly improve their full accuracy when using\nquantization.', 'Training large machine learning models requires a distributed computing\napproach, with communication of the model updates being the bottleneck. For\nthis reason, several methods based on the compression (e.g., sparsification\nand/or quantization) of updates were recently proposed, including QSGD\n(Alistarh et al., 2017), TernGrad (Wen et al., 2017), SignSGD (Bernstein et\nal., 2018), and DQGD (Khirirat et al., 2018). However, none of these methods\nare able to learn the gradients, which renders them incapable of converging to\nthe true optimum in the batch mode, incompatible with non-smooth regularizers,\nand slows down their convergence. In this work we propose a new distributed\nlearning method --- DIANA --- which resolves these issues via compression of\ngradient differences. We perform a theoretical analysis in the strongly convex\nand nonconvex settings and show that our rates are superior to existing rates.\nOur analysis of block-quantization and differences between $\\ell_2$ and\n$\\ell_\\infty$ quantization closes the gaps in theory and practice. Finally, by\napplying our analysis technique to TernGrad, we establish the first convergence\nrate for this method.', 'We study distributed algorithms for expected loss minimization where the\ndatasets are large and have to be stored on different machines. Often we deal\nwith minimizing the average of a set of convex functions where each function is\nthe empirical risk of the corresponding part of the data. In the distributed\nsetting where the individual data instances can be accessed only on the local\nmachines, there would be a series of rounds of local computations followed by\nsome communication among the machines. Since the cost of the communication is\nusually higher than the local machine computations, it is important to reduce\nit as much as possible. However, we should not allow this to make the\ncomputation too expensive to become a burden in practice. Using second-order\nmethods could make the algorithms converge faster and decrease the amount of\ncommunication needed. There are some successful attempts in developing\ndistributed second-order methods. Although these methods have shown fast\nconvergence, their local computation is expensive and could enjoy more\nimprovement for practical uses. In this study we modify an existing approach,\nDANE (Distributed Approximate NEwton), in order to improve the computational\ncost while maintaining the accuracy. We tackle this problem by using iterative\nmethods for solving the local subproblems approximately instead of providing\nexact solutions for each round of communication. We study how using different\niterative methods affect the behavior of the algorithm and try to provide an\nappropriate tradeoff between the amount of local computation and the required\namount of communication. We demonstrate the practicality of our algorithm and\ncompare it to the existing distributed gradient based methods such as SGD.', 'Deep Neural Networks have gained significant attraction due to their wide\napplicability in different domains. DNN sizes and training samples are\nconstantly growing, making training of such workloads more challenging.\nDistributed training is a solution to reduce the training time.\nHigh-performance distributed training platforms should leverage\nmulti-dimensional hierarchical networks, which interconnect accelerators\nthrough different levels of the network, to dramatically reduce expensive NICs\nrequired for the scale-out network. However, it comes at the expense of\ncommunication overhead between distributed accelerators to exchange gradients\nor input/output activation. In order to allow for further scaling of the\nworkloads, communication overhead needs to be minimized. In this paper, we\nmotivate the fact that in training platforms, adding more intermediate network\ndimensions is beneficial for efficiently mitigating the excessive use of\nexpensive NIC resources. Further, we address different challenges of the DNN\ntraining on hierarchical networks. We discuss when designing the interconnect,\nhow to distribute network bandwidth resources across different dimensions in\norder to (i) maximize BW utilization of all dimensions, and (ii) minimizing the\noverall training time for the target workload. We then implement a framework\nthat, for a given workload, determines the best network configuration that\nmaximizes performance, or performance-per-cost.', 'We consider large scale distributed optimization over a set of edge devices\nconnected to a central server, where the limited communication bandwidth\nbetween the server and edge devices imposes a significant bottleneck for the\noptimization procedure. Inspired by recent advances in federated learning, we\npropose a distributed stochastic gradient descent (SGD) type algorithm that\nexploits the sparsity of the gradient, when possible, to reduce communication\nburden. At the heart of the algorithm is to use compressed sensing techniques\nfor the compression of the local stochastic gradients at the device side; and\nat the server side, a sparse approximation of the global stochastic gradient is\nrecovered from the noisy aggregated compressed local gradients. We conduct\ntheoretical analysis on the convergence of our algorithm in the presence of\nnoise perturbation incurred by the communication channels, and also conduct\nnumerical experiments to corroborate its effectiveness.', 'Generative models are becoming increasingly popular in the literature, with\nGenerative Adversarial Networks (GAN) being the most successful variant, yet.\nWith this increasing demand and popularity, it is becoming equally difficult\nand challenging to implement and consume GAN models. A qualitative user survey\nconducted across 47 practitioners show that expert level skill is required to\nuse GAN model for a given task, despite the presence of various open source\nlibraries. In this research, we propose a novel system called AuthorGAN, aiming\nto achieve true democratization of GAN authoring. A highly modularized library\nagnostic representation of GAN model is defined to enable interoperability of\nGAN architecture across different libraries such as Keras, Tensorflow, and\nPyTorch. An intuitive drag-and-drop based visual designer is built using\nnode-red platform to enable custom architecture designing without the need for\nwriting any code. Five different GAN models are implemented as a part of this\nframework and the performance of the different GAN models are shown using the\nbenchmark MNIST dataset.', 'While the availability of large datasets is perceived to be a key requirement\nfor training deep neural networks, it is possible to train such models with\nrelatively little data. However, compensating for the absence of large datasets\ndemands a series of actions to enhance the quality of the existing samples and\nto generate new ones. This paper summarizes our winning submission to the\n""Data-Centric AI"" competition. We discuss some of the challenges that arise\nwhile training with a small dataset, offer a principled approach for systematic\ndata quality enhancement, and propose a GAN-based solution for synthesizing new\ndata points. Our evaluations indicate that the dataset generated by the\nproposed pipeline offers 5% accuracy improvement while being significantly\nsmaller than the baseline.', 'Generative adversarial networks (GANs) have been recently adopted for\nsuper-resolution, an application closely related to what is referred to as\n""downscaling"" in the atmospheric sciences: improving the spatial resolution of\nlow-resolution images. The ability of conditional GANs to generate an ensemble\nof solutions for a given input lends itself naturally to stochastic\ndownscaling, but the stochastic nature of GANs is not usually considered in\nsuper-resolution applications. Here, we introduce a recurrent, stochastic\nsuper-resolution GAN that can generate ensembles of time-evolving\nhigh-resolution atmospheric fields for an input consisting of a low-resolution\nsequence of images of the same field. We test the GAN using two datasets, one\nconsisting of radar-measured precipitation from Switzerland, the other of cloud\noptical thickness derived from the Geostationary Earth Observing Satellite 16\n(GOES-16). We find that the GAN can generate realistic, temporally consistent\nsuper-resolution sequences for both datasets. The statistical properties of the\ngenerated ensemble are analyzed using rank statistics, a method adapted from\nensemble weather forecasting; these analyses indicate that the GAN produces\nclose to the correct amount of variability in its outputs. As the GAN generator\nis fully convolutional, it can be applied after training to input images larger\nthan the images used to train it. It is also able to generate time series much\nlonger than the training sequences, as demonstrated by applying the generator\nto a three-month dataset of the precipitation radar data. The source code to\nour GAN is available at https://github.com/jleinonen/downscaling-rnn-gan.', 'Generative Adversarial Networks (GANs) have received a great deal of\nattention due in part to recent success in generating original, high-quality\nsamples from visual domains. However, most current methods only allow for users\nto guide this image generation process through limited interactions. In this\nwork we develop a novel GAN framework that allows humans to be ""in-the-loop"" of\nthe image generation process. Our technique iteratively accepts relative\nconstraints of the form ""Generate an image more like image A than image B"".\nAfter each constraint is given, the user is presented with new outputs from the\nGAN, informing the next round of feedback. This feedback is used to constrain\nthe output of the GAN with respect to an underlying semantic space that can be\ndesigned to model a variety of different notions of similarity (e.g. classes,\nattributes, object relationships, color, etc.). In our experiments, we show\nthat our GAN framework is able to generate images that are of comparable\nquality to equivalent unsupervised GANs while satisfying a large number of the\nconstraints provided by users, effectively changing a GAN into one that allows\nusers interactive control over image generation without sacrificing image\nquality.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'In this paper, we present a distributed variant of adaptive stochastic\ngradient method for training deep neural networks in the parameter-server\nmodel. To reduce the communication cost among the workers and server, we\nincorporate two types of quantization schemes, i.e., gradient quantization and\nweight quantization, into the proposed distributed Adam. Besides, to reduce the\nbias introduced by quantization operations, we propose an error-feedback\ntechnique to compensate for the quantized gradient. Theoretically, in the\nstochastic nonconvex setting, we show that the distributed adaptive gradient\nmethod with gradient quantization and error-feedback converges to the\nfirst-order stationary point, and that the distributed adaptive gradient method\nwith weight quantization and error-feedback converges to the point related to\nthe quantized level under both the single-worker and multi-worker modes. At\nlast, we apply the proposed distributed adaptive gradient methods to train deep\nneural networks. Experimental results demonstrate the efficacy of our methods.', 'Generative Adversarial Networks (GAN) have many potential medical imaging\napplications, including data augmentation, domain adaptation, and model\nexplanation. Due to the limited memory of Graphical Processing Units (GPUs),\nmost current 3D GAN models are trained on low-resolution medical images, these\nmodels either cannot scale to high-resolution or are prone to patchy artifacts.\nIn this work, we propose a novel end-to-end GAN architecture that can generate\nhigh-resolution 3D images. We achieve this goal by using different\nconfigurations between training and inference. During training, we adopt a\nhierarchical structure that simultaneously generates a low-resolution version\nof the image and a randomly selected sub-volume of the high-resolution image.\nThe hierarchical design has two advantages: First, the memory demand for\ntraining on high-resolution images is amortized among sub-volumes. Furthermore,\nanchoring the high-resolution sub-volumes to a single low-resolution image\nensures anatomical consistency between sub-volumes. During inference, our model\ncan directly generate full high-resolution images. We also incorporate an\nencoder with a similar hierarchical structure into the model to extract\nfeatures from the images. Experiments on 3D thorax CT and brain MRI demonstrate\nthat our approach outperforms state of the art in image generation. We also\ndemonstrate clinical applications of the proposed model in data augmentation\nand clinical-relevant feature extraction.', 'Recent years have witnessed the rapid progress of generative adversarial\nnetworks (GANs). However, the success of the GAN models hinges on a large\namount of training data. This work proposes a regularization approach for\ntraining robust GAN models on limited data. We theoretically show a connection\nbetween the regularized loss and an f-divergence called LeCam-divergence, which\nwe find is more robust under limited training data. Extensive experiments on\nseveral benchmark datasets demonstrate that the proposed regularization scheme\n1) improves the generalization performance and stabilizes the learning dynamics\nof GAN models under limited training data, and 2) complements the recent data\naugmentation methods. These properties facilitate training GAN models to\nachieve state-of-the-art performance when only limited training data of the\nImageNet benchmark is available.', 'Sentiment analysis is a task that may suffer from a lack of data in certain\ncases, as the datasets are often generated and annotated by humans. In cases\nwhere data is inadequate for training discriminative models, generate models\nmay aid training via data augmentation. Generative Adversarial Networks (GANs)\nare one such model that has advanced the state of the art in several tasks,\nincluding as image and text generation. In this paper, I train GAN models on\nlow resource datasets, then use them for the purpose of data augmentation\ntowards improving sentiment classifier generalization. Given the constraints of\nlimited data, I explore various techniques to train the GAN models. I also\npresent an analysis of the quality of generated GAN data as more training data\nfor the GAN is made available. In this analysis, the generated data is\nevaluated as a test set (against a model trained on real data points) as well\nas a training set to train classification models. Finally, I also conduct a\nvisual analysis by projecting the generated and the real data into a\ntwo-dimensional space using the t-Distributed Stochastic Neighbor Embedding\n(t-SNE) method.', 'Modern large scale machine learning applications require stochastic\noptimization algorithms to be implemented on distributed computational\narchitectures. A key bottleneck is the communication overhead for exchanging\ninformation such as stochastic gradients among different workers. In this\npaper, to reduce the communication cost we propose a convex optimization\nformulation to minimize the coding length of stochastic gradients. To solve the\noptimal sparsification efficiently, several simple and fast algorithms are\nproposed for approximate solution, with theoretical guaranteed for sparseness.\nExperiments on $\\ell_2$ regularized logistic regression, support vector\nmachines, and convolutional neural networks validate our sparsification\napproaches.', 'In federated learning, communication cost is often a critical bottleneck to\nscale up distributed optimization algorithms to collaboratively learn a model\nfrom millions of devices with potentially unreliable or limited communication\nand heterogeneous data distributions. Two notable trends to deal with the\ncommunication overhead of federated algorithms are gradient compression and\nlocal computation with periodic communication. Despite many attempts,\ncharacterizing the relationship between these two approaches has proven\nelusive. We address this by proposing a set of algorithms with periodical\ncompressed (quantized or sparsified) communication and analyze their\nconvergence properties in both homogeneous and heterogeneous local data\ndistribution settings. For the homogeneous setting, our analysis improves\nexisting bounds by providing tighter convergence rates for both strongly convex\nand non-convex objective functions. To mitigate data heterogeneity, we\nintroduce a local gradient tracking scheme and obtain sharp convergence rates\nthat match the best-known communication complexities without compression for\nconvex, strongly convex, and nonconvex settings. We complement our theoretical\nresults and demonstrate the effectiveness of our proposed methods by several\nexperiments on real-world datasets.', ""The present paper develops a novel aggregated gradient approach for\ndistributed machine learning that adaptively compresses the gradient\ncommunication. The key idea is to first quantize the computed gradients, and\nthen skip less informative quantized gradient communications by reusing\noutdated gradients. Quantizing and skipping result in `lazy' worker-server\ncommunications, which justifies the term Lazily Aggregated Quantized gradient\nthat is henceforth abbreviated as LAQ. Our LAQ can provably attain the same\nlinear convergence rate as the gradient descent in the strongly convex case,\nwhile effecting major savings in the communication overhead both in transmitted\nbits as well as in communication rounds. Empirically, experiments with real\ndata corroborate a significant communication reduction compared to existing\ngradient- and stochastic gradient-based algorithms."", 'Machine learning systems based on deep neural networks (DNNs) produce\nstate-of-the-art results in many applications. Considering the large amount of\ntraining data and know-how required to generate the network, it is more\npractical to use third-party DNN intellectual property (IP) cores for many\ndesigns. No doubt to say, it is essential for DNN IP vendors to provide test\ncases for functional validation without leaking their parameters to IP users.\nTo satisfy this requirement, we propose to effectively generate test cases that\nactivate parameters as many as possible and propagate their perturbations to\noutputs. Then the functionality of DNN IPs can be validated by only checking\ntheir outputs. However, it is difficult considering large numbers of parameters\nand highly non-linearity of DNNs. In this paper, we tackle this problem by\njudiciously selecting samples from the DNN training set and applying a\ngradient-based method to generate new test cases. Experimental results\ndemonstrate the efficacy of our proposed solution.', 'Sign-based algorithms (e.g. signSGD) have been proposed as a biased gradient\ncompression technique to alleviate the communication bottleneck in training\nlarge neural networks across multiple workers. We show simple convex\ncounter-examples where signSGD does not converge to the optimum. Further, even\nwhen it does converge, signSGD may generalize poorly when compared with SGD.\nThese issues arise because of the biased nature of the sign compression\noperator. We then show that using error-feedback, i.e. incorporating the error\nmade by the compression operator into the next step, overcomes these issues. We\nprove that our algorithm EF-SGD with arbitrary compression operator achieves\nthe same rate of convergence as SGD without any additional assumptions. Thus\nEF-SGD achieves gradient compression for free. Our experiments thoroughly\nsubstantiate the theory and show that error-feedback improves both convergence\nand generalization. Code can be found at\n\\url{https://github.com/epfml/error-feedback-SGD}.', 'Leveraging powerful deep learning techniques, the end-to-end (E2E) learning\nof communication system is able to outperform the classical communication\nsystem. Unfortunately, this communication system cannot be trained by deep\nlearning without known channel. To deal with this problem, a generative\nadversarial network (GAN) based training scheme has been recently proposed to\nimitate the real channel. However, the gradient vanishing and overfitting\nproblems of GAN will result in the serious performance degradation of E2E\nlearning of communication system. To mitigate these two problems, we propose a\nresidual aided GAN (RA-GAN) based training scheme in this paper. Particularly,\ninspired by the idea of residual learning, we propose a residual generator to\nmitigate the gradient vanishing problem by realizing a more robust gradient\nbackpropagation. Moreover, to cope with the overfitting problem, we reconstruct\nthe loss function for training by adding a regularizer, which limits the\nrepresentation ability of RA-GAN. Simulation results show that the trained\nresidual generator has better generation performance than the conventional\ngenerator, and the proposed RA-GAN based training scheme can achieve the\nnear-optimal block error rate (BLER) performance with a negligible\ncomputational complexity increase in both the theoretical channel model and the\nray-tracing based channel dataset.', ""The Generative Models have gained considerable attention in the field of\nunsupervised learning via a new and practical framework called Generative\nAdversarial Networks (GAN) due to its outstanding data generation capability.\nMany models of GAN have proposed, and several practical applications emerged in\nvarious domains of computer vision and machine learning. Despite GAN's\nexcellent success, there are still obstacles to stable training. The problems\nare due to Nash-equilibrium, internal covariate shift, mode collapse, vanishing\ngradient, and lack of proper evaluation metrics. Therefore, stable training is\na crucial issue in different applications for the success of GAN. Herein, we\nsurvey several training solutions proposed by different researchers to\nstabilize GAN training. We survey, (I) the original GAN model and its modified\nclassical versions, (II) detail analysis of various GAN applications in\ndifferent domains, (III) detail study about the various GAN training obstacles\nas well as training solutions. Finally, we discuss several new issues as well\nas research outlines to the topic."", 'Large-scale distributed optimization is of great importance in various\napplications. For data-parallel based distributed learning, the inter-node\ngradient communication often becomes the performance bottleneck. In this paper,\nwe propose the error compensated quantized stochastic gradient descent\nalgorithm to improve the training efficiency. Local gradients are quantized to\nreduce the communication overhead, and accumulated quantization error is\nutilized to speed up the convergence. Furthermore, we present theoretical\nanalysis on the convergence behaviour, and demonstrate its advantage over\ncompetitors. Extensive experiments indicate that our algorithm can compress\ngradients by a factor of up to two magnitudes without performance degradation.', 'Deep neural networks (DNNs) have been extremely successful in solving many\nchallenging AI tasks in natural language processing, speech recognition, and\ncomputer vision nowadays. However, DNNs are typically computation intensive,\nmemory demanding, and power hungry, which significantly limits their usage on\nplatforms with constrained resources. Therefore, a variety of compression\ntechniques (e.g. quantization, pruning, and knowledge distillation) have been\nproposed to reduce the size and power consumption of DNNs. Blockwise knowledge\ndistillation is one of the compression techniques that can effectively reduce\nthe size of a highly complex DNN. However, it is not widely adopted due to its\nlong training time. In this paper, we propose a novel parallel blockwise\ndistillation algorithm to accelerate the distillation process of sophisticated\nDNNs. Our algorithm leverages local information to conduct independent\nblockwise distillation, utilizes depthwise separable layers as the efficient\nreplacement block architecture, and properly addresses limiting factors (e.g.\ndependency, synchronization, and load balancing) that affect parallelism. The\nexperimental results running on an AMD server with four Geforce RTX 2080Ti GPUs\nshow that our algorithm can achieve 3x speedup plus 19% energy savings on VGG\ndistillation, and 3.5x speedup plus 29% energy savings on ResNet distillation,\nboth with negligible accuracy loss. The speedup of ResNet distillation can be\nfurther improved to 3.87 when using four RTX6000 GPUs in a distributed cluster.', 'Recently there has been a surge of research on improving the communication\nefficiency of distributed training. However, little work has been done to\nsystematically understand whether the network is the bottleneck and to what\nextent.\n  In this paper, we take a first-principles approach to measure and analyze the\nnetwork performance of distributed training. As expected, our measurement\nconfirms that communication is the component that blocks distributed training\nfrom linear scale-out. However, contrary to the common belief, we find that the\nnetwork is running at low utilization and that if the network can be fully\nutilized, distributed training can achieve a scaling factor of close to one.\nMoreover, while many recent proposals on gradient compression advocate over\n100x compression ratio, we show that under full network utilization, there is\nno need for gradient compression in 100 Gbps network. On the other hand, a\nlower speed network like 10 Gbps requires only 2x--5x gradients compression\nratio to achieve almost linear scale-out. Compared to application-level\ntechniques like gradient compression, network-level optimizations do not\nrequire changes to applications and do not hurt the performance of trained\nmodels. As such, we advocate that the real challenge of distributed training is\nfor the network community to develop high-performance network transport to\nfully utilize the network capacity and achieve linear scale-out.', 'As deep learning is showing unprecedented success in medical image analysis\ntasks, the lack of sufficient medical data is emerging as a critical problem.\nWhile recent attempts to solve the limited data problem using Generative\nAdversarial Networks (GAN) have been successful in generating realistic images\nwith diversity, most of them are based on image-to-image translation and thus\nrequire extensive datasets from different domains. Here, we propose a novel\nmodel that can successfully generate 3D brain MRI data from random vectors by\nlearning the data distribution. Our 3D GAN model solves both image blurriness\nand mode collapse problems by leveraging alpha-GAN that combines the advantages\nof Variational Auto-Encoder (VAE) and GAN with an additional code discriminator\nnetwork. We also use the Wasserstein GAN with Gradient Penalty (WGAN-GP) loss\nto lower the training instability. To demonstrate the effectiveness of our\nmodel, we generate new images of normal brain MRI and show that our model\noutperforms baseline models in both quantitative and qualitative measurements.\nWe also train the model to synthesize brain disorder MRI data to demonstrate\nthe wide applicability of our model. Our results suggest that the proposed\nmodel can successfully generate various types and modalities of 3D whole brain\nvolumes from a small set of training data.', 'Learning predictive models from interaction with the world allows an agent,\nsuch as a robot, to learn about how the world works, and then use this learned\nmodel to plan coordinated sequences of actions to bring about desired outcomes.\nHowever, learning a model that captures the dynamics of complex skills\nrepresents a major challenge: if the agent needs a good model to perform these\nskills, it might never be able to collect the experience on its own that is\nrequired to learn these delicate and complex behaviors. Instead, we can imagine\naugmenting the training set with observational data of other agents, such as\nhumans. Such data is likely more plentiful, but represents a different\nembodiment. For example, videos of humans might show a robot how to use a tool,\nbut (i) are not annotated with suitable robot actions, and (ii) contain a\nsystematic distributional shift due to the embodiment differences between\nhumans and robots. We address the first challenge by formulating the\ncorresponding graphical model and treating the action as an observed variable\nfor the interaction data and an unobserved variable for the observation data,\nand the second challenge by using a domain-dependent prior. In addition to\ninteraction data, our method is able to leverage videos of passive observations\nin a driving dataset and a dataset of robotic manipulation videos. A robotic\nplanning agent equipped with our method can learn to use tools in a tabletop\nrobotic manipulation setting by observing humans without ever seeing a robotic\nvideo of tool use.', 'Deep Neural Networks (DNNs) have achieved im- pressive accuracy in many\napplication domains including im- age classification. Training of DNNs is an\nextremely compute- intensive process and is solved using variants of the\nstochastic gradient descent (SGD) algorithm. A lot of recent research has\nfocussed on improving the performance of DNN training. In this paper, we\npresent optimization techniques to improve the performance of the data parallel\nsynchronous SGD algorithm using the Torch framework: (i) we maintain data\nin-memory to avoid file I/O overheads, (ii) we present a multi-color based MPI\nAllreduce algorithm to minimize communication overheads, and (iii) we propose\noptimizations to the Torch data parallel table framework that handles\nmulti-threading. We evaluate the performance of our optimizations on a Power 8\nMinsky cluster with 32 nodes and 128 NVidia Pascal P100 GPUs. With our\noptimizations, we are able to train 90 epochs of the ResNet-50 model on the\nImagenet-1k dataset using 256 GPUs in just 48 minutes. This significantly\nimproves on the previously best known performance of training 90 epochs of the\nResNet-50 model on the same dataset using 256 GPUs in 65 minutes. To the best\nof our knowledge, this is the best known training performance demonstrated for\nthe Imagenet- 1k dataset.', 'While pre-trained language models (e.g., BERT) have achieved impressive\nresults on different natural language processing tasks, they have large numbers\nof parameters and suffer from big computational and memory costs, which make\nthem difficult for real-world deployment. Therefore, model compression is\nnecessary to reduce the computation and memory cost of pre-trained models. In\nthis work, we aim to compress BERT and address the following two challenging\npractical issues: (1) The compression algorithm should be able to output\nmultiple compressed models with different sizes and latencies, in order to\nsupport devices with different memory and latency limitations; (2) The\nalgorithm should be downstream task agnostic, so that the compressed models are\ngenerally applicable for different downstream tasks. We leverage techniques in\nneural architecture search (NAS) and propose NAS-BERT, an efficient method for\nBERT compression. NAS-BERT trains a big supernet on a search space containing a\nvariety of architectures and outputs multiple compressed models with adaptive\nsizes and latency. Furthermore, the training of NAS-BERT is conducted on\nstandard self-supervised pre-training tasks (e.g., masked language model) and\ndoes not depend on specific downstream tasks. Thus, the compressed models can\nbe used across various downstream tasks. The technical challenge of NAS-BERT is\nthat training a big supernet on the pre-training task is extremely costly. We\nemploy several techniques including block-wise search, search space pruning,\nand performance approximation to improve search efficiency and accuracy.\nExtensive experiments on GLUE and SQuAD benchmark datasets demonstrate that\nNAS-BERT can find lightweight models with better accuracy than previous\napproaches, and can be directly applied to different downstream tasks with\nadaptive model sizes for different requirements of memory or latency.', ""Generative Adversarial Networks (GANs) have become a dominant class of\ngenerative models. In recent years, GAN variants have yielded especially\nimpressive results in the synthesis of a variety of forms of data. Examples\ninclude compelling natural and artistic images, textures, musical sequences,\nand 3D object files. However, one obvious synthesis candidate is missing. In\nthis work, we answer one of deep learning's most pressing questions: GAN you do\nthe GAN GAN? That is, is it possible to train a GAN to model a distribution of\nGANs? We release the full source code for this project under the MIT license."", ""BERT based ranking models have achieved superior performance on various\ninformation retrieval tasks. However, the large number of parameters and\ncomplex self-attention operation come at a significant latency overhead. To\nremedy this, recent works propose late-interaction architectures, which allow\npre-computation of intermediate document representations, thus reducing the\nruntime latency. Nonetheless, having solved the immediate latency issue, these\nmethods now introduce storage costs and network fetching latency, which limits\ntheir adoption in real-life production systems.\n  In this work, we propose the Succinct Document Representation (SDR) scheme\nthat computes highly compressed intermediate document representations,\nmitigating the storage/network issue. Our approach first reduces the dimension\nof token representations by encoding them using a novel autoencoder\narchitecture that uses the document's textual content in both the encoding and\ndecoding phases. After this token encoding step, we further reduce the size of\nentire document representations using a modern quantization technique.\n  Extensive evaluations on passage re-reranking on the MSMARCO dataset show\nthat compared to existing approaches using compressed document representations,\nour method is highly efficient, achieving 4x-11.6x better compression rates for\nthe same ranking quality."", 'This article is in the context of gradient compression. Gradient compression\nis a popular technique for mitigating the communication bottleneck observed\nwhen training large machine learning models in a distributed manner using\ngradient-based methods such as stochastic gradient descent. In this article,\nassuming a Gaussian distribution for the components in gradient, we find the\nrate distortion trade-off of gradient quantization schemes such as Scaled-sign\nand Top-K, and compare with the Shannon rate distortion limit. A similar\ncomparison with vector quantizers also is presented.', 'Although the distributed machine learning methods can speed up the training\nof large deep neural networks, the communication cost has become the\nnon-negligible bottleneck to constrain the performance. To address this\nchallenge, the gradient compression based communication-efficient distributed\nlearning methods were designed to reduce the communication cost, and more\nrecently the local error feedback was incorporated to compensate for the\ncorresponding performance loss. However, in this paper, we will show that a new\n""gradient mismatch"" problem is raised by the local error feedback in\ncentralized distributed training and can lead to degraded performance compared\nwith full-precision training. To solve this critical problem, we propose two\nnovel techniques, 1) step ahead and 2) error averaging, with rigorous\ntheoretical analysis. Both our theoretical and empirical results show that our\nnew methods can handle the ""gradient mismatch"" problem. The experimental\nresults show that we can even train faster with common gradient compression\nschemes than both the full-precision training and local error feedback\nregarding the training epochs and without performance loss.', 'Conditional Generative Adversarial Networks (cGANs) have enabled controllable\nimage synthesis for many vision and graphics applications. However, recent\ncGANs are 1-2 orders of magnitude more compute-intensive than modern\nrecognition CNNs. For example, GauGAN consumes 281G MACs per image, compared to\n0.44G MACs for MobileNet-v3, making it difficult for interactive deployment. In\nthis work, we propose a general-purpose compression framework for reducing the\ninference time and model size of the generator in cGANs. Directly applying\nexisting compression methods yields poor performance due to the difficulty of\nGAN training and the differences in generator architectures. We address these\nchallenges in two ways. First, to stabilize GAN training, we transfer knowledge\nof multiple intermediate representations of the original model to its\ncompressed model and unify unpaired and paired learning. Second, instead of\nreusing existing CNN designs, our method finds efficient architectures via\nneural architecture search. To accelerate the search process, we decouple the\nmodel training and search via weight sharing. Experiments demonstrate the\neffectiveness of our method across different supervision settings, network\narchitectures, and learning methods. Without losing image quality, we reduce\nthe computation of CycleGAN by 21x, Pix2pix by 12x, MUNIT by 29x, and GauGAN by\n9x, paving the way for interactive image synthesis.', 'The Gradient Boosted Tree (GBT) algorithm is one of the most popular machine\nlearning algorithms used in production, for tasks that include Click-Through\nRate (CTR) prediction and learning-to-rank. To deal with the massive datasets\navailable today, many distributed GBT methods have been proposed. However, they\nall assume a row-distributed dataset, addressing scalability only with respect\nto the number of data points and not the number of features, and increasing\ncommunication cost for high-dimensional data. In order to allow for scalability\nacross both the data point and feature dimensions, and reduce communication\ncost, we propose block-distributed GBTs. We achieve communication efficiency by\nmaking full use of the data sparsity and adapting the Quickscorer algorithm to\nthe block-distributed setting. We evaluate our approach using datasets with\nmillions of features, and demonstrate that we are able to achieve multiple\norders of magnitude reduction in communication cost for sparse data, with no\nloss in accuracy, while providing a more scalable design. As a result, we are\nable to reduce the training time for high-dimensional data, and allow more\ncost-effective scale-out without the need for expensive network communication.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', ""The proliferation of big data has brought an urgent demand for\nprivacy-preserving data publishing. Traditional solutions to this demand have\nlimitations on effectively balancing the tradeoff between privacy and utility\nof the released data. Thus, the database community and machine learning\ncommunity have recently studied a new problem of relational data synthesis\nusing generative adversarial networks (GAN) and proposed various algorithms.\nHowever, these algorithms are not compared under the same framework and thus it\nis hard for practitioners to understand GAN's benefits and limitations. To\nbridge the gaps, we conduct so far the most comprehensive experimental study\nthat investigates applying GAN to relational data synthesis. We introduce a\nunified GAN-based framework and define a space of design solutions for each\ncomponent in the framework, including neural network architectures and training\nstrategies. We conduct extensive experiments to explore the design space and\ncompare with traditional data synthesis approaches. Through extensive\nexperiments, we find that GAN is very promising for relational data synthesis,\nand provide guidance for selecting appropriate design solutions. We also point\nout limitations of GAN and identify future research directions."", 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', ""Communication cost is one major bottleneck for the scalability for\ndistributed learning. One approach to reduce the communication cost is to\ncompress the gradient during communication. However, directly compressing the\ngradient decelerates the convergence speed, and the resulting algorithm may\ndiverge for biased compression. Recent work addressed this problem for\nstochastic gradient descent by adding back the compression error from the\nprevious step. This idea was further extended to one class of variance reduced\nalgorithms, where the variance of the stochastic gradient is reduced by taking\na moving average over all history gradients. However, our analysis shows that\njust adding the previous step's compression error, as done in existing work,\ndoes not fully compensate the compression error. So, we propose\nErrorCompensatedX, which uses the compression error from the previous two\nsteps. We show that ErrorCompensatedX can achieve the same asymptotic\nconvergence rate with the training without compression. Moreover, we provide a\nunified theoretical analysis framework for this class of variance reduced\nalgorithms, with or without error compensation."", 'Humans can naturally learn to execute a new task by seeing it performed by\nother individuals once, and then reproduce it in a variety of configurations.\nEndowing robots with this ability of imitating humans from third person is a\nvery immediate and natural way of teaching new tasks. Only recently, through\nmeta-learning, there have been successful attempts to one-shot imitation\nlearning from humans; however, these approaches require a lot of human\nresources to collect the data in the real world to train the robot. But is\nthere a way to remove the need for real world human demonstrations during\ntraining? We show that with Task-Embedded Control Networks, we can infer\ncontrol polices by embedding human demonstrations that can condition a control\npolicy and achieve one-shot imitation learning. Importantly, we do not use a\nreal human arm to supply demonstrations during training, but instead leverage\ndomain randomisation in an application that has not been seen before:\nsim-to-real transfer on humans. Upon evaluating our approach on pushing and\nplacing tasks in both simulation and in the real world, we show that in\ncomparison to a system that was trained on real-world data we are able to\nachieve similar results by utilising only simulation data.', 'GANS are powerful generative models that are able to model the manifold of\nnatural images. We leverage this property to perform manifold regularization by\napproximating the Laplacian norm using a Monte Carlo approximation that is\neasily computed with the GAN. When incorporated into the feature-matching GAN\nof Improved GAN, we achieve state-of-the-art results for GAN-based\nsemi-supervised learning on the CIFAR-10 dataset, with a method that is\nsignificantly easier to implement than competing methods.', 'Due to limited communication resources at the client and a massive number of\nmodel parameters, large-scale distributed learning tasks suffer from\ncommunication bottleneck. Gradient compression is an effective method to reduce\ncommunication load by transmitting compressed gradients. Motivated by the fact\nthat in the scenario of stochastic gradients descent, gradients between\nadjacent rounds may have a high correlation since they wish to learn the same\nmodel, this paper proposes a practical gradient compression scheme for\nfederated learning, which uses historical gradients to compress gradients and\nis based on Wyner-Ziv coding but without any probabilistic assumption. We also\nimplement our gradient quantization method on the real dataset, and the\nperformance of our method is better than the previous schemes.', 'Due to the increasing amount of data on the internet, finding a\nhighly-informative, low-dimensional representation for text is one of the main\nchallenges for efficient natural language processing tasks including text\nclassification. This representation should capture the semantic information of\nthe text while retaining their relevance level for document classification.\nThis approach maps the documents with similar topics to a similar space in\nvector space representation. To obtain representation for large text, we\npropose the utilization of deep Siamese neural networks. To embed document\nrelevance in topics in the distributed representation, we use a Siamese neural\nnetwork to jointly learn document representations. Our Siamese network consists\nof two sub-network of multi-layer perceptron. We examine our representation for\nthe text categorization task on BBC news dataset. The results show that the\nproposed representations outperform the conventional and state-of-the-art\nrepresentations in the text classification task on this dataset.', 'Sufficient supervised information is crucial for any machine learning models\nto boost performance. However, labeling data is expensive and sometimes\ndifficult to obtain. Active learning is an approach to acquire annotations for\ndata from a human oracle by selecting informative samples with a high\nprobability to enhance performance. In recent emerging studies, a generative\nadversarial network (GAN) has been integrated with active learning to generate\ngood candidates to be presented to the oracle. In this paper, we propose a\nnovel model that is able to obtain labels for data in a cheaper manner without\nthe need to query an oracle. In the model, a novel reward for each sample is\ndevised to measure the degree of uncertainty, which is obtained from a\nclassifier trained with existing labeled data. This reward is used to guide a\nconditional GAN to generate informative samples with a higher probability for a\ncertain label. With extensive evaluations, we have confirmed the effectiveness\nof the model, showing that the generated samples are capable of improving the\nclassification performance in popular image classification tasks.', 'Generative adversarial networks (GANs) are very popular to generate realistic\nimages, but they often suffer from the training instability issues and the\nphenomenon of mode loss. In order to attain greater diversity in GAN\nsynthesized data, it is critical to solving the problem of mode loss. Our work\nexplores probabilistic approaches to GAN modelling that could allow us to\ntackle these issues. We present Prb-GANs, a new variation that uses dropout to\ncreate a distribution over the network parameters with the posterior learnt\nusing variational inference. We describe theoretically and validate\nexperimentally using simple and complex datasets the benefits of such an\napproach. We look into further improvements using the concept of uncertainty\nmeasures. Through a set of further modifications to the loss functions for each\nnetwork of the GAN, we are able to get results that show the improvement of GAN\nperformance. Our methods are extremely simple and require very little\nmodification to existing GAN architecture.', 'Neural networks have proven their capabilities by outperforming many other\napproaches on regression or classification tasks on various kinds of data.\nOther astonishing results have been achieved using neural nets as data\ngenerators, especially in settings of generative adversarial networks (GANs).\nOne special application is the field of image domain translations. Here, the\ngoal is to take an image with a certain style (e.g. a photography) and\ntransform it into another one (e.g. a painting). If such a task is performed\nfor unpaired training examples, the corresponding GAN setting is complex, the\nneural networks are large, and this leads to a high peak memory consumption\nduring, both, training and evaluation phase. This sets a limit to the highest\nprocessable image size. We address this issue by the idea of not processing the\nwhole image at once, but to train and evaluate the domain translation on the\nlevel of overlapping image subsamples. This new approach not only enables us to\ntranslate high-resolution images that otherwise cannot be processed by the\nneural network at once, but also allows us to work with comparably small neural\nnetworks and with limited hardware resources. Additionally, the number of\nimages required for the training process is significantly reduced. We present\nhigh-quality results on images with a total resolution of up to over 50\nmegapixels and emonstrate that our method helps to preserve local image details\nwhile it also keeps global consistency.', 'Communication overhead is the key challenge for distributed training.\nGradient compression is a widely used approach to reduce communication traffic.\nWhen combining with parallel communication mechanism method like pipeline,\ngradient compression technique can greatly alleviate the impact of\ncommunication overhead. However, there exists two problems of gradient\ncompression technique to be solved. Firstly, gradient compression brings in\nextra computation cost, which will delay the next training iteration. Secondly,\ngradient compression usually leads to the decrease of convergence accuracy.', 'Data sets are growing in complexity thanks to the increasing facilities we\nhave nowadays to both generate and store data. This poses many challenges to\nmachine learning that are leading to the proposal of new methods and paradigms,\nin order to be able to deal with what is nowadays referred to as Big Data. In\nthis paper we propose a method for the aggregation of different Bayesian\nnetwork structures that have been learned from separate data sets, as a first\nstep towards mining data sets that need to be partitioned in an horizontal way,\ni.e. with respect to the instances, in order to be processed. Considerations\nthat should be taken into account when dealing with this situation are\ndiscussed. Scalable learning of Bayesian networks is slowly emerging, and our\nmethod constitutes one of the first insights into Gaussian Bayesian network\naggregation from different sources. Tested on synthetic data it obtains good\nresults that surpass those from individual learning. Future research will be\nfocused on expanding the method and testing more diverse data sets.', 'The performance of neural decoders can degrade over time due to\nnonstationarities in the relationship between neuronal activity and behavior.\nIn this case, brain-machine interfaces (BMI) require adaptation of their\ndecoders to maintain high performance across time. One way to achieve this is\nby use of periodical calibration phases, during which the BMI system (or an\nexternal human demonstrator) instructs the user to perform certain movements or\nbehaviors. This approach has two disadvantages: (i) calibration phases\ninterrupt the autonomous operation of the BMI and (ii) between two calibration\nphases the BMI performance might not be stable but continuously decrease. A\nbetter alternative would be that the BMI decoder is able to continuously adapt\nin an unsupervised manner during autonomous BMI operation, i.e. without knowing\nthe movement intentions of the user.\n  In the present article, we present an efficient method for such unsupervised\ntraining of BMI systems for continuous movement control. The proposed method\nutilizes a cost function derived from neuronal recordings, which guides a\nlearning algorithm to evaluate the decoding parameters. We verify the\nperformance of our adaptive method by simulating a BMI user with an optimal\nfeedback control model and its interaction with our adaptive BMI decoder. The\nsimulation results show that the cost function and the algorithm yield fast and\nprecise trajectories towards targets at random orientations on a 2-dimensional\ncomputer screen. For initially unknown and non-stationary tuning parameters,\nour unsupervised method is still able to generate precise trajectories and to\nkeep its performance stable in the long term. The algorithm can optionally work\nalso with neuronal error signals instead or in conjunction with the proposed\nunsupervised adaptation.', 'Many engineering problems require the prediction of\nrealization-to-realization variability or a refined description of modeled\nquantities. In that case, it is necessary to sample elements from unknown\nhigh-dimensional spaces with possibly millions of degrees of freedom. While\nthere exist methods able to sample elements from probability density functions\n(PDF) with known shapes, several approximations need to be made when the\ndistribution is unknown. In this paper the sampling method, as well as the\ninference of the underlying distribution, are both handled with a data-driven\nmethod known as generative adversarial networks (GAN), which trains two\ncompeting neural networks to produce a network that can effectively generate\nsamples from the training set distribution. In practice, it is often necessary\nto draw samples from conditional distributions. When the conditional variables\nare continuous, only one (if any) data point corresponding to a particular\nvalue of a conditioning variable may be available, which is not sufficient to\nestimate the conditional distribution. This work handles this problem using an\na priori estimation of the conditional moments of a PDF. Two approaches,\nstochastic estimation, and an external neural network are compared here for\ncomputing these moments; however, any preferred method can be used. The\nalgorithm is demonstrated in the case of the deconvolution of a filtered\nturbulent flow field. It is shown that all the versions of the proposed\nalgorithm effectively sample the target conditional distribution with minimal\nimpact on the quality of the samples compared to state-of-the-art methods.\nAdditionally, the procedure can be used as a metric for the diversity of\nsamples generated by a conditional GAN (cGAN) conditioned with continuous\nvariables.', 'Extracting information from full documents is an important problem in many\ndomains, but most previous work focus on identifying relationships within a\nsentence or a paragraph. It is challenging to create a large-scale information\nextraction (IE) dataset at the document level since it requires an\nunderstanding of the whole document to annotate entities and their\ndocument-level relationships that usually span beyond sentences or even\nsections. In this paper, we introduce SciREX, a document level IE dataset that\nencompasses multiple IE tasks, including salient entity identification and\ndocument level $N$-ary relation identification from scientific articles. We\nannotate our dataset by integrating automatic and human annotations, leveraging\nexisting scientific knowledge resources. We develop a neural model as a strong\nbaseline that extends previous state-of-the-art IE models to document-level IE.\nAnalyzing the model performance shows a significant gap between human\nperformance and current baselines, inviting the community to use our dataset as\na challenge to develop document-level IE models. Our data and code are publicly\navailable at https://github.com/allenai/SciREX', 'With the increase of information, document classification as one of the\nmethods of text mining, plays vital role in many management and organizing\ninformation. Document classification is the process of assigning a document to\none or more predefined category labels. Document classification includes\ndifferent parts such as text processing, term selection, term weighting and\nfinal classification. The accuracy of document classification is very\nimportant. Thus improvement in each part of classification should lead to\nbetter results and higher precision. Term weighting has a great impact on the\naccuracy of the classification. Most of the existing weighting methods exploit\nthe statistical information of terms in documents and do not consider semantic\nrelations between words. In this paper, an automated document classification\nsystem is presented that uses a novel term weighting method based on semantic\nrelations between terms. To evaluate the proposed method, three standard\nPersian corpuses are used. Experiment results show 2 to 4 percent improvement\nin classification accuracy compared with the best previous designed system for\nPersian documents.', 'We present a learned image compression system based on GANs, operating at\nextremely low bitrates. Our proposed framework combines an encoder,\ndecoder/generator and a multi-scale discriminator, which we train jointly for a\ngenerative learned compression objective. The model synthesizes details it\ncannot afford to store, obtaining visually pleasing results at bitrates where\nprevious methods fail and show strong artifacts. Furthermore, if a semantic\nlabel map of the original image is available, our method can fully synthesize\nunimportant regions in the decoded image such as streets and trees from the\nlabel map, proportionally reducing the storage cost. A user study confirms that\nfor low bitrates, our approach is preferred to state-of-the-art methods, even\nwhen they use more than double the bits.', 'How to leverage cross-document interactions to improve ranking performance is\nan important topic in information retrieval (IR) research. However, this topic\nhas not been well-studied in the learning-to-rank setting and most of the\nexisting work still treats each document independently while scoring. The\nrecent development of deep learning shows strength in modeling complex\nrelationships across sequences and sets. It thus motivates us to study how to\nleverage cross-document interactions for learning-to-rank in the deep learning\nframework. In this paper, we formally define the permutation-equivariance\nrequirement for a scoring function that captures cross-document interactions.\nWe then propose a self-attention based document interaction network and show\nthat it satisfies the permutation-equivariant requirement, and can generate\nscores for document sets of varying sizes. Our proposed methods can\nautomatically learn to capture document interactions without any auxiliary\ninformation, and can scale across large document sets. We conduct experiments\non three ranking datasets: the benchmark Web30k, a Gmail search, and a Google\nDrive Quick Access dataset. Experimental results show that our proposed methods\nare both more effective and efficient than baselines.', ""In information retrieval (IR) and related tasks, term weighting approaches\ntypically consider the frequency of the term in the document and in the\ncollection in order to compute a score reflecting the importance of the term\nfor the document. In tasks characterized by the presence of training data (such\nas text classification) it seems logical that the term weighting function\nshould take into account the distribution (as estimated from training data) of\nthe term across the classes of interest. Although `supervised term weighting'\napproaches that use this intuition have been described before, they have failed\nto show consistent improvements. In this article we analyse the possible\nreasons for this failure, and call consolidated assumptions into question.\nFollowing this criticism we propose a novel supervised term weighting approach\nthat, instead of relying on any predefined formula, learns a term weighting\nfunction optimised on the training set of interest; we dub this approach\n\\emph{Learning to Weight} (LTW). The experiments that we run on several\nwell-known benchmarks, and using different learning methods, show that our\nmethod outperforms previous term weighting approaches in text classification."", 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', 'Generative Adversarial Networks (GAN) have received wide attention in the\nmachine learning field for their potential to learn high-dimensional, complex\nreal data distribution. Specifically, they do not rely on any assumptions about\nthe distribution and can generate real-like samples from latent space in a\nsimple manner. This powerful property leads GAN to be applied to various\napplications such as image synthesis, image attribute editing, image\ntranslation, domain adaptation and other academic fields. In this paper, we aim\nto discuss the details of GAN for those readers who are familiar with, but do\nnot comprehend GAN deeply or who wish to view GAN from various perspectives. In\naddition, we explain how GAN operates and the fundamental meaning of various\nobjective functions that have been suggested recently. We then focus on how the\nGAN can be combined with an autoencoder framework. Finally, we enumerate the\nGAN variants that are applied to various tasks and other fields for those who\nare interested in exploiting GAN for their research.', 'This is a tutorial and survey paper on Generative Adversarial Network (GAN),\nadversarial autoencoders, and their variants. We start with explaining\nadversarial learning and the vanilla GAN. Then, we explain the conditional GAN\nand DCGAN. The mode collapse problem is introduced and various methods,\nincluding minibatch GAN, unrolled GAN, BourGAN, mixture GAN, D2GAN, and\nWasserstein GAN, are introduced for resolving this problem. Then, maximum\nlikelihood estimation in GAN are explained along with f-GAN, adversarial\nvariational Bayes, and Bayesian GAN. Then, we cover feature matching in GAN,\nInfoGAN, GRAN, LSGAN, energy-based GAN, CatGAN, MMD GAN, LapGAN, progressive\nGAN, triple GAN, LAG, GMAN, AdaGAN, CoGAN, inverse GAN, BiGAN, ALI, SAGAN,\nFew-shot GAN, SinGAN, and interpolation and evaluation of GAN. Then, we\nintroduce some applications of GAN such as image-to-image translation\n(including PatchGAN, CycleGAN, DeepFaceDrawing, simulated GAN, interactive\nGAN), text-to-image translation (including StackGAN), and mixing image\ncharacteristics (including FineGAN and MixNMatch). Finally, we explain the\nautoencoders based on adversarial learning including adversarial autoencoder,\nPixelGAN, and implicit autoencoder.', ""This paper studies a distributed multi-agent convex optimization problem. The\nsystem comprises multiple agents in this problem, each with a set of local data\npoints and an associated local cost function. The agents are connected to a\nserver, and there is no inter-agent communication. The agents' goal is to learn\na parameter vector that optimizes the aggregate of their local costs without\nrevealing their local data points. In principle, the agents can solve this\nproblem by collaborating with the server using the traditional distributed\ngradient-descent method. However, when the aggregate cost is ill-conditioned,\nthe gradient-descent method (i) requires a large number of iterations to\nconverge, and (ii) is highly unstable against process noise. We propose an\niterative pre-conditioning technique to mitigate the deleterious effects of the\ncost function's conditioning on the convergence rate of distributed\ngradient-descent. Unlike the conventional pre-conditioning techniques, the\npre-conditioner matrix in our proposed technique updates iteratively to\nfacilitate implementation on the distributed network. In the distributed\nsetting, we provably show that the proposed algorithm converges linearly with\nan improved rate of convergence than the traditional and adaptive\ngradient-descent methods. Additionally, for the special case when the minimizer\nof the aggregate cost is unique, our algorithm converges superlinearly. We\ndemonstrate our algorithm's superior performance compared to prominent\ndistributed algorithms for solving real logistic regression problems and\nemulating neural network training via a noisy quadratic model, thereby\nsignifying the proposed algorithm's efficiency for distributively solving\nnon-convex optimization. Moreover, we empirically show that the proposed\nalgorithm results in faster training without compromising the generalization\nperformance."", ""The Electromyography (EMG) signal is the electrical activity produced by\ncells of skeletal muscles in order to provide a movement. The non-invasive\nprosthetic hand works with several electrodes, placed on the stump of an\namputee, that record this signal. In order to favour the control of prosthesis,\nthe EMG signal is analyzed with algorithms based on machine learning theory to\ndecide the movement that the subject is going to do. In order to obtain a\nsignificant control of the prosthesis and avoid mismatch between desired and\nperformed movements, a long training period is needed when we use the\ntraditional algorithm of machine learning (i.e. Support Vector Machines). An\nactual challenge in this field concerns the reduction of the time necessary for\nan amputee to learn how to use the prosthesis. Recently, several algorithms\nthat exploit a form of prior knowledge have been proposed. In general, we refer\nto prior knowledge as a past experience available in the form of models. In our\ncase an amputee, that attempts to perform some movements with the prosthesis,\ncould use experience from different subjects that are already able to perform\nthose movements. The aim of this work is to verify, with a computational\ninvestigation, if for an amputee this kind of previous experience is useful in\norder to reduce the training time and boost the prosthetic control.\nFurthermore, we want to understand if and how the final results change when the\nprevious knowledge of intact or amputated subjects is used for a new amputee.\nOur experiments indicate that: (1) the use of experience, from other subjects\nalready trained to perform a task, makes us able to reduce the training time of\nabout an order of magnitude; (2) it seems that an amputee that tries to learn\nto use the prosthesis doesn't reach different results when he/she exploits\nprevious experience of amputees or intact."", 'We propose a Label Propagation based algorithm for weakly supervised text\nclassification. We construct a graph where each document is represented by a\nnode and edge weights represent similarities among the documents. Additionally,\nwe discover underlying topics using Latent Dirichlet Allocation (LDA) and\nenrich the document graph by including the topics in the form of additional\nnodes. The edge weights between a topic and a text document represent level of\n""affinity"" between them. Our approach does not require document level\nlabelling, instead it expects manual labels only for topic nodes. This\nsignificantly minimizes the level of supervision needed as only a few topics\nare observed to be enough for achieving sufficiently high accuracy. The Label\nPropagation Algorithm is employed on this enriched graph to propagate labels\namong the nodes. Our approach combines the advantages of Label Propagation\n(through document-document similarities) and Topic Modelling (for minimal but\nsmart supervision). We demonstrate the effectiveness of our approach on various\ndatasets and compare with state-of-the-art weakly supervised text\nclassification approaches.', 'Generative models are known to be difficult to assess. Recent works,\nespecially on generative adversarial networks (GANs), produce good visual\nsamples of varied categories of images. However, the validation of their\nquality is still difficult to define and there is no existing agreement on the\nbest evaluation process. This paper aims at making a step toward an objective\nevaluation process for generative models. It presents a new method to assess a\ntrained generative model by evaluating the test accuracy of a classifier\ntrained with generated data. The test set is composed of real images.\nTherefore, The classifier accuracy is used as a proxy to evaluate if the\ngenerative model fit the true data distribution. By comparing results with\ndifferent generated datasets we are able to classify and compare generative\nmodels. The motivation of this approach is also to evaluate if generative\nmodels can help discriminative neural networks to learn, i.e., measure if\ntraining on generated data is able to make a model successful at testing on\nreal settings. Our experiments compare different generators from the\nVariational Auto-Encoders (VAE) and Generative Adversarial Network (GAN)\nframeworks on MNIST and fashion MNIST datasets. Our results show that none of\nthe generative models is able to replace completely true data to train a\ndiscriminative model. But they also show that the initial GAN and WGAN are the\nbest choices to generate on MNIST database (Modified National Institute of\nStandards and Technology database) and fashion MNIST database.', ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'We consider a request processing system composed of organizations and their\nservers connected by the Internet.\n  The latency a user observes is a sum of communication delays and the time\nneeded to handle the request on a server. The handling time depends on the\nserver congestion, i.e. the total number of requests a server must handle. We\nanalyze the problem of balancing the load in a network of servers in order to\nminimize the total observed latency. We consider both cooperative and selfish\norganizations (each organization aiming to minimize the latency of the\nlocally-produced requests). The problem can be generalized to the task\nscheduling in a distributed cloud; or to content delivery in an\norganizationally-distributed CDNs.\n  In a cooperative network, we show that the problem is polynomially solvable.\nWe also present a distributed algorithm iteratively balancing the load. We show\nhow to estimate the distance between the current solution and the optimum based\non the amount of load exchanged by the algorithm. During the experimental\nevaluation, we show that the distributed algorithm is efficient, therefore it\ncan be used in networks with dynamically changing loads.\n  In a network of selfish organizations, we prove that the price of anarchy\n(the worst-case loss of performance due to selfishness) is low when the network\nis homogeneous and the servers are loaded (the request handling time is high\ncompared to the communication delay). After relaxing these assumptions, we\nassess the loss of performance caused by the selfishness experimentally,\nshowing that it remains low.\n  Our results indicate that a network of servers handling requests can be\nefficiently managed by a distributed algorithm. Additionally, even if the\nnetwork is organizationally distributed, with individual organizations\noptimizing performance of their requests, the network remains efficient.', 'Adam is shown not being able to converge to the optimal solution in certain\ncases. Researchers recently propose several algorithms to avoid the issue of\nnon-convergence of Adam, but their efficiency turns out to be unsatisfactory in\npractice. In this paper, we provide new insight into the non-convergence issue\nof Adam as well as other adaptive learning rate methods. We argue that there\nexists an inappropriate correlation between gradient $g_t$ and the\nsecond-moment term $v_t$ in Adam ($t$ is the timestep), which results in that a\nlarge gradient is likely to have small step size while a small gradient may\nhave a large step size. We demonstrate that such biased step sizes are the\nfundamental cause of non-convergence of Adam, and we further prove that\ndecorrelating $v_t$ and $g_t$ will lead to unbiased step size for each\ngradient, thus solving the non-convergence problem of Adam. Finally, we propose\nAdaShift, a novel adaptive learning rate method that decorrelates $v_t$ and\n$g_t$ by temporal shifting, i.e., using temporally shifted gradient $g_{t-n}$\nto calculate $v_t$. The experiment results demonstrate that AdaShift is able to\naddress the non-convergence issue of Adam, while still maintaining a\ncompetitive performance with Adam in terms of both training speed and\ngeneralization.', ""Nowadays, the Internet represents a vast informational space, growing\nexponentially and the problem of search for relevant data becomes essential as\nnever before. The algorithm proposed in the article allows to perform natural\nlanguage queries on content of the document and get comprehensive meaningful\nanswers. The problem is partially solved for English as SQuAD contains enough\ndata to learn on, but there is no such dataset in Russian, so the methods used\nby scientists now are not applicable to Russian. Brain2 framework allows to\ncope with the problem - it stands out for its ability to be applied on small\ndatasets and does not require impressive computing power. The algorithm is\nillustrated on Sberbank of Russia Strategy's text and assumes the use of a\nneuromodel consisting of 65 mln synapses. The trained model is able to\nconstruct word-by-word answers to questions based on a given text. The existing\nlimitations are its current inability to identify synonyms, pronoun relations\nand allegories. Nevertheless, the results of conducted experiments showed high\ncapacity and generalisation ability of the suggested approach."", 'Complex deep learning models now achieve state of the art performance for\nmany document retrieval tasks. The best models process the query or claim\njointly with the document. However for fast scalable search it is desirable to\nhave document embeddings which are independent of the claim. In this paper we\nshow that knowledge distillation can be used to encourage a model that\ngenerates claim independent document encodings to mimic the behavior of a more\ncomplex model which generates claim dependent encodings. We explore this\napproach in document retrieval for a fact extraction and verification task. We\nshow that by using the soft labels from a complex cross attention teacher\nmodel, the performance of claim independent student LSTM or CNN models is\nimproved across all the ranking metrics. The student models we use are 12x\nfaster in runtime and 20x smaller in number of parameters than the teacher', 'Neural network-based methods for abstractive summarization produce outputs\nthat are more fluent than other techniques, but which can be poor at content\nselection. This work proposes a simple technique for addressing this issue: use\na data-efficient content selector to over-determine phrases in a source\ndocument that should be part of the summary. We use this selector as a\nbottom-up attention step to constrain the model to likely phrases. We show that\nthis approach improves the ability to compress text, while still generating\nfluent summaries. This two-step process is both simpler and higher performing\nthan other end-to-end content selection models, leading to significant\nimprovements on ROUGE for both the CNN-DM and NYT corpus. Furthermore, the\ncontent selector can be trained with as little as 1,000 sentences, making it\neasy to transfer a trained summarizer to a new domain.', 'The concept of a decentralized ledger usually implies that each node of a\nblockchain network stores the entire blockchain. However, in the case of\npopular blockchains, which each weigh several hundreds of GB, the large amount\nof data to be stored can incite new or low-capacity nodes to run lightweight\nclients. Such nodes do not participate to the global storage effort and can\nresult in a centralization of the blockchain by very few nodes, which is\ncontrary to the basic concepts of a blockchain.\n  To avoid this problem, we propose new low storage nodes that store a reduced\namount of data generated from the blockchain by using erasure codes. The\nproperties of this technique ensure that any block of the chain can be easily\nrebuild from a small number of such nodes. This system should encourage low\nstorage nodes to contribute to the storage of the blockchain and to maintain\ndecentralization despite of a globally increasing size of the blockchain. This\nsystem paves the way to new types of blockchains which would only be managed by\nlow capacity nodes.', 'In this paper, we proposed a new technique, {\\em variance controlled\nstochastic gradient} (VCSG), to improve the performance of the stochastic\nvariance reduced gradient (SVRG) algorithm. To avoid over-reducing the variance\nof gradient by SVRG, a hyper-parameter $\\lambda$ is introduced in VCSG that is\nable to control the reduced variance of SVRG. Theory shows that the\noptimization method can converge by using an unbiased gradient estimator, but\nin practice, biased gradient estimation can allow more efficient convergence to\nthe vicinity since an unbiased approach is computationally more expensive.\n$\\lambda$ also has the effect of balancing the trade-off between unbiased and\nbiased estimations. Secondly, to minimize the number of full gradient\ncalculations in SVRG, a variance-bounded batch is introduced to reduce the\nnumber of gradient calculations required in each iteration. For smooth\nnon-convex functions, the proposed algorithm converges to an approximate\nfirst-order stationary point (i.e.\n$\\mathbb{E}\\|\\nabla{f}(x)\\|^{2}\\leq\\epsilon$) within\n$\\mathcal{O}(min\\{1/\\epsilon^{3/2},n^{1/4}/\\epsilon\\})$ number of stochastic\ngradient evaluations, which improves the leading gradient complexity of\nstochastic gradient-based method SCS\n$(\\mathcal{O}(min\\{1/\\epsilon^{5/3},n^{2/3}/\\epsilon\\})$. It is shown\ntheoretically and experimentally that VCSG can be deployed to improve\nconvergence.', 'Novel contexts may often arise in complex querying scenarios such as in\nevidence-based medicine (EBM) involving biomedical literature, that may not\nexplicitly refer to entities or canonical concept forms occurring in any fact-\nor rule-based knowledge source such as an ontology like the UMLS. Moreover,\nhidden associations between candidate concepts meaningful in the current\ncontext, may not exist within a single document, but within the collection, via\nalternate lexical forms. Therefore, inspired by the recent success of\nsequence-to-sequence neural models in delivering the state-of-the-art in a wide\nrange of NLP tasks, we develop a novel sequence-to-set framework with neural\nattention for learning document representations that can effect term transfer\nwithin the corpus, for semantically tagging a large collection of documents. We\ndemonstrate that our proposed method can be effective in both a supervised\nmulti-label classification setup for text categorization, as well as in a\nunique unsupervised setting with no human-annotated document labels that uses\nno external knowledge resources and only corpus-derived term statistics to\ndrive the training. Further, we show that semi-supervised training using our\narchitecture on large amounts of unlabeled data can augment performance on the\ntext categorization task when limited labeled data is available. Our approach\nto generate document encodings employing our sequence-to-set models for\ninference of semantic tags, gives to the best of our knowledge, the\nstate-of-the-art for both, the unsupervised query expansion task for the TREC\nCDS 2016 challenge dataset when evaluated on an Okapi BM25--based document\nretrieval system; and also over the MLTM baseline (Soleimani et al, 2016), for\nboth supervised and semi-supervised multi-label prediction tasks on the\ndel.icio.us and Ohsumed datasets. We will make our code and data publicly\navailable.', 'Graph Neural Networks (GNNs) have received significant attention due to their\nstate-of-the-art performance on various graph representation learning tasks.\nHowever, recent studies reveal that GNNs are vulnerable to adversarial attacks,\ni.e. an attacker is able to fool the GNNs by perturbing the graph structure or\nnode features deliberately. While being able to successfully decrease the\nperformance of GNNs, most existing attacking algorithms require access to\neither the model parameters or the training data, which is not practical in the\nreal world.\n  In this paper, we develop deeper insights into the Mettack algorithm, which\nis a representative grey-box attacking method, and then we propose a\ngradient-based black-box attacking algorithm. Firstly, we show that the Mettack\nalgorithm will perturb the edges unevenly, thus the attack will be highly\ndependent on a specific training set. As a result, a simple yet useful strategy\nto defense against Mettack is to train the GNN with the validation set.\nSecondly, to overcome the drawbacks, we propose the Black-Box Gradient Attack\n(BBGA) algorithm. Extensive experiments demonstrate that out proposed method is\nable to achieve stable attack performance without accessing the training sets\nof the GNNs. Further results shows that our proposed method is also applicable\nwhen attacking against various defense methods.', 'Document clustering is a text mining technique used to provide better\ndocument search and browsing in digital libraries or online corpora. A lot of\nresearch has been done on biomedical document clustering that is based on using\nexisting ontology. But, associations and co-occurrences of the medical concepts\nare not well represented by using ontology. In this research, a vector\nrepresentation of concepts of diseases and similarity measurement between\nconcepts are proposed. They identify the closest concepts of diseases in the\ncontext of a corpus. Each document is represented by using the vector space\nmodel. A weight scheme is proposed to consider both local content and\nassociations between concepts. A Self-Organizing Map is used as document\nclustering algorithm. The vector projection and visualization features of SOM\nenable visualization and analysis of the clusters distributions and\nrelationships on the two dimensional space. The experimental results show that\nthe proposed document clustering framework generates meaningful clusters and\nfacilitate visualization of the clusters based on the concepts of diseases.', 'Traditional Relational Topic Models provide a way to discover the hidden\ntopics from a document network. Many theoretical and practical tasks, such as\ndimensional reduction, document clustering, link prediction, benefit from this\nrevealed knowledge. However, existing relational topic models are based on an\nassumption that the number of hidden topics is known in advance, and this is\nimpractical in many real-world applications. Therefore, in order to relax this\nassumption, we propose a nonparametric relational topic model in this paper.\nInstead of using fixed-dimensional probability distributions in its generative\nmodel, we use stochastic processes. Specifically, a gamma process is assigned\nto each document, which represents the topic interest of this document.\nAlthough this method provides an elegant solution, it brings additional\nchallenges when mathematically modeling the inherent network structure of\ntypical document network, i.e., two spatially closer documents tend to have\nmore similar topics. Furthermore, we require that the topics are shared by all\nthe documents. In order to resolve these challenges, we use a subsampling\nstrategy to assign each document a different gamma process from the global\ngamma process, and the subsampling probabilities of documents are assigned with\na Markov Random Field constraint that inherits the document network structure.\nThrough the designed posterior inference algorithm, we can discover the hidden\ntopics and its number simultaneously. Experimental results on both synthetic\nand real-world network datasets demonstrate the capabilities of learning the\nhidden topics and, more importantly, the number of topics.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Latent topic models have been successfully applied as an unsupervised topic\ndiscovery technique in large document collections. With the proliferation of\nhypertext document collection such as the Internet, there has also been great\ninterest in extending these approaches to hypertext [6, 9]. These approaches\ntypically model links in an analogous fashion to how they model words - the\ndocument-link co-occurrence matrix is modeled in the same way that the\ndocument-word co-occurrence matrix is modeled in standard topic models. In this\npaper we present a probabilistic generative model for hypertext document\ncollections that explicitly models the generation of links. Specifically, links\nfrom a word w to a document d depend directly on how frequent the topic of w is\nin d, in addition to the in-degree of d. We show how to perform EM learning on\nthis model efficiently. By not modeling links as analogous to words, we end up\nusing far fewer free parameters and obtain better link prediction results.', 'In both the fields of computer science and medicine there is very strong\ninterest in developing personalized treatment policies for patients who have\nvariable responses to treatments. In particular, I aim to find an optimal\npersonalized treatment policy which is a non-deterministic function of the\npatient specific covariate data that maximizes the expected survival time or\nclinical outcome. I developed an algorithmic framework to solve multistage\ndecision problem with a varying number of stages that are subject to censoring\nin which the ""rewards"" are expected survival times. In specific, I developed a\nnovel Q-learning algorithm that dynamically adjusts for these parameters.\nFurthermore, I found finite upper bounds on the generalized error of the\ntreatment paths constructed by this algorithm. I have also shown that when the\noptimal Q-function is an element of the approximation space, the anticipated\nsurvival times for the treatment regime constructed by the algorithm will\nconverge to the optimal treatment path. I demonstrated the performance of the\nproposed algorithmic framework via simulation studies and through the analysis\nof chronic depression data and a hypothetical clinical trial. The censored\nQ-learning algorithm I developed is more effective than the state of the art\nclinical decision support systems and is able to operate in environments when\nmany covariate parameters may be unobtainable or censored.', 'Document categorization is a technique where the category of a document is\ndetermined. In this paper three well-known supervised learning techniques which\nare Support Vector Machine(SVM), Na\\""ive Bayes(NB) and Stochastic Gradient\nDescent(SGD) compared for Bengali document categorization. Besides classifier,\nclassification also depends on how feature is selected from dataset. For\nanalyzing those classifier performances on predicting a document against twelve\ncategories several feature selection techniques are also applied in this\narticle namely Chi square distribution, normalized TFIDF (term\nfrequency-inverse document frequency) with word analyzer. So, we attempt to\nexplore the efficiency of those three-classification algorithms by using two\ndifferent feature selection techniques in this article.', 'Modern entity linking systems rely on large collections of documents\nspecifically annotated for the task (e.g., AIDA CoNLL). In contrast, we propose\nan approach which exploits only naturally occurring information: unlabeled\ndocuments and Wikipedia. Our approach consists of two stages. First, we\nconstruct a high recall list of candidate entities for each mention in an\nunlabeled document. Second, we use the candidate lists as weak supervision to\nconstrain our document-level entity linking model. The model treats entities as\nlatent variables and, when estimated on a collection of unlabelled texts,\nlearns to choose entities relying both on local context of each mention and on\ncoherence with other entities in the document. The resulting approach rivals\nfully-supervised state-of-the-art systems on standard test sets. It also\napproaches their performance in the very challenging setting: when tested on a\ntest set sampled from the data used to estimate the supervised systems. By\ncomparing to Wikipedia-only training of our model, we demonstrate that modeling\nunlabeled documents is beneficial.', 'Information on different fields which are collected by users requires\nappropriate management and organization to be structured in a standard way and\nretrieved fast and more easily. Document classification is a conventional\nmethod to separate text based on their subjects among scientific text, web\npages and digital library. Different methods and techniques are proposed for\ndocument classifications that have advantages and deficiencies. In this paper,\nseveral unsupervised and supervised document classification methods are studied\nand compared.', 'We introduce a novel type of text representation that preserves the 2D layout\nof a document. This is achieved by encoding each document page as a\ntwo-dimensional grid of characters. Based on this representation, we present a\ngeneric document understanding pipeline for structured documents. This pipeline\nmakes use of a fully convolutional encoder-decoder network that predicts a\nsegmentation mask and bounding boxes. We demonstrate its capabilities on an\ninformation extraction task from invoices and show that it significantly\noutperforms approaches based on sequential text or document images.', 'Often, when dealing with real-world recognition problems, we do not need, and\noften cannot have, knowledge of the entire set of possible classes that might\nappear during operational testing. In such cases, we need to think of robust\nclassification methods able to deal with the ""unknown"" and properly reject\nsamples belonging to classes never seen during training. Notwithstanding,\nexisting classifiers to date were mostly developed for the closed-set scenario,\ni.e., the classification setup in which it is assumed that all test samples\nbelong to one of the classes with which the classifier was trained. In the\nopen-set scenario, however, a test sample can belong to none of the known\nclasses and the classifier must properly reject it by classifying it as\nunknown. In this work, we extend upon the well-known Support Vector Machines\n(SVM) classifier and introduce the Open-Set Support Vector Machines (OSSVM),\nwhich is suitable for recognition in open-set setups. OSSVM balances the\nempirical risk and the risk of the unknown and ensures that the region of the\nfeature space in which a test sample would be classified as known (one of the\nknown classes) is always bounded, ensuring a finite risk of the unknown. In\nthis work, we also highlight the properties of the SVM classifier related to\nthe open-set scenario, and provide necessary and sufficient conditions for an\nRBF SVM to have bounded open-space risk.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.', ""Neural networks with deep architectures have demonstrated significant\nperformance improvements in computer vision, speech recognition, and natural\nlanguage processing. The challenges in information retrieval (IR), however, are\ndifferent from these other application areas. A common form of IR involves\nranking of documents--or short passages--in response to keyword-based queries.\nEffective IR systems must deal with query-document vocabulary mismatch problem,\nby modeling relationships between different query and document terms and how\nthey indicate relevance. Models should also consider lexical matches when the\nquery contains rare terms--such as a person's name or a product model\nnumber--not seen during training, and to avoid retrieving semantically related\nbut irrelevant results. In many real-life IR tasks, the retrieval involves\nextremely large collections--such as the document index of a commercial Web\nsearch engine--containing billions of documents. Efficient IR methods should\ntake advantage of specialized IR data structures, such as inverted index, to\nefficiently retrieve from large collections. Given an information need, the IR\nsystem also mediates how much exposure an information artifact receives by\ndeciding whether it should be displayed, and where it should be positioned,\namong other results. Exposure-aware IR systems may optimize for additional\nobjectives, besides relevance, such as parity of exposure for retrieved items\nand content publishers. In this thesis, we present novel neural architectures\nand methods motivated by the specific needs and challenges of IR tasks."", 'We present a novel Bayesian topic model for learning discourse-level document\nstructure. Our model leverages insights from discourse theory to constrain\nlatent topic assignments in a way that reflects the underlying organization of\ndocument topics. We propose a global model in which both topic selection and\nordering are biased to be similar across a collection of related documents. We\nshow that this space of orderings can be effectively represented using a\ndistribution over permutations called the Generalized Mallows Model. We apply\nour method to three complementary discourse-level tasks: cross-document\nalignment, document segmentation, and information ordering. Our experiments\nshow that incorporating our permutation-based model in these applications\nyields substantial improvements in performance over previously proposed\nmethods.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I want a model to produce a prediction distribution closely approximating a delta function centered at a single value, but  I also need to make sure the optimal prediction distribution is not an actual delta function. If I use mean squared loss, the optimal distribution will tend to the delta function. So I need to address this issue by using a different loss. I am thinking about using generative models that can sample its output. Maybe rejection sampling? This problem might be related to mode collapses and how to avoid them, so I would like to explore that as well. In addition, the model should not use too many parameters and I want to keep everything in low dimensional spaces. Finally, to test if my model works, I want to test on a small image dataset. In this setting, the delta function would be the salient feature or event in each image.","['Despite excellent progress in recent years, mode collapse remains a major\nunsolved problem in generative adversarial networks (GANs).In this paper, we\npresent spectral regularization for GANs (SR-GANs), a new and robust method for\ncombating the mode collapse problem in GANs. Theoretical analysis shows that\nthe optimal solution to the discriminator has a strong relationship to the\nspectral distributions of the weight matrix.Therefore, we monitor the spectral\ndistribution in the discriminator of spectral normalized GANs (SN-GANs), and\ndiscover a phenomenon which we refer to as spectral collapse, where a large\nnumber of singular values of the weight matrices drop dramatically when mode\ncollapse occurs. We show that there are strong evidence linking mode collapse\nto spectral collapse; and based on this link, we set out to tackle spectral\ncollapse as a surrogate of mode collapse. We have developed a spectral\nregularization method where we compensate the spectral distributions of the\nweight matrices to prevent them from collapsing, which in turn successfully\nprevents mode collapse in GANs. We provide theoretical explanations for why\nSR-GANs are more stable and can provide better performances than SN-GANs. We\nalso present extensive experimental results and analysis to show that SR-GANs\nnot only always outperform SN-GANs but also always succeed in combating mode\ncollapse where SN-GANs fail. The code is available at\nhttps://github.com/max-liu-112/SRGANs-Spectral-Regularization-GANs-.', 'Due to their flexibility and predictive performance, machine-learning based\nregression methods have become an important tool for predictive modeling and\nforecasting. However, most methods focus on estimating the conditional mean or\nspecific quantiles of the target quantity and do not provide the full\nconditional distribution, which contains uncertainty information that might be\ncrucial for decision making. In this article, we provide a general solution by\ntransforming a conditional distribution estimation problem into a constrained\nmulti-class classification problem, in which tools such as deep neural\nnetworks. We propose a novel joint binary cross-entropy loss function to\naccomplish this goal. We demonstrate its performance in various simulation\nstudies comparing to state-of-the-art competing methods. Additionally, our\nmethod shows improved accuracy in a probabilistic solar energy forecasting\nproblem.', 'Generative models have proven to be an outstanding tool for representing\nhigh-dimensional probability distributions and generating realistic-looking\nimages. An essential characteristic of generative models is their ability to\nproduce multi-modal outputs. However, while training, they are often\nsusceptible to mode collapse, that is models are limited in mapping input noise\nto only a few modes of the true data distribution. In this work, we draw\ninspiration from Determinantal Point Process (DPP) to propose an unsupervised\npenalty loss that alleviates mode collapse while producing higher quality\nsamples. DPP is an elegant probabilistic measure used to model negative\ncorrelations within a subset and hence quantify its diversity. We use DPP\nkernel to model the diversity in real data as well as in synthetic data. Then,\nwe devise an objective term that encourages generators to synthesize data with\nsimilar diversity to real data. In contrast to previous state-of-the-art\ngenerative models that tend to use additional trainable parameters or complex\ntraining paradigms, our method does not change the original training scheme.\nEmbedded in an adversarial training and variational autoencoder, our Generative\nDPP approach shows a consistent resistance to mode-collapse on a wide variety\nof synthetic data and natural image datasets including MNIST, CIFAR10, and\nCelebA, while outperforming state-of-the-art methods for data-efficiency,\ngeneration quality, and convergence-time whereas being 5.8x faster than its\nclosest competitor.', 'Despite its success, generative adversarial networks (GANs) still suffer from\nmode collapse, namely the generator can only map latent variables to a partial\nset of modes of the target distribution. In this paper, we analyze and try to\nregularize this issue with an independent and identically distributed (IID)\nsampling perspective and emphasize that holding the IID property for generation\nfor target distribution (i.e. real distribution) can naturally avoid mode\ncollapse. This is based on the basic IID assumption for real data in machine\nlearning. However, though the source samples $\\{\\mathbf{z}\\}$ obey IID, the\ngenerations $\\{G(\\mathbf{z})\\}$ may not necessarily be IID from the target\ndistribution. Based on this observation, we propose a necessary condition of\nIID generation and provide a new loss to encourage the closeness between the\ninverse source of real data and the Gaussian source in the latent space to\nregularize the generation to be IID from the target distribution. The logic is\nthat the inverse samples from target data should also be IID in the source\ndistribution. Experiments on both synthetic and real-world data show the\neffectiveness of our model.', 'In this paper, a shallow Ritz-type neural network for solving elliptic\nequations with delta function singular sources on an interface is developed.\nThere are three novel features in the present work; namely, (i) the delta\nfunction singularity is naturally removed, (ii) level set function is\nintroduced as a feature input, (iii) it is completely shallow, comprising only\none hidden layer. We first introduce the energy functional of the problem and\nthen transform the contribution of singular sources to a regular surface\nintegral along the interface. In such a way, the delta function singularity can\nbe naturally removed without introducing a discrete one that is commonly used\nin traditional regularization methods, such as the well-known immersed boundary\nmethod. The original problem is then reformulated as a minimization problem. We\npropose a shallow Ritz-type neural network with one hidden layer to approximate\nthe global minimizer of the energy functional. As a result, the network is\ntrained by minimizing the loss function that is a discrete version of the\nenergy. In addition, we include the level set function of the interface as a\nfeature input of the network and find that it significantly improves the\ntraining efficiency and accuracy. We perform a series of numerical tests to\nshow the accuracy of the present method and its capability for problems in\nirregular domains and higher dimensions.', 'We present an approach for continual learning (CL) that is based on fully\nprobabilistic (or generative) models of machine learning. In contrast to, e.g.,\nGANs that are ""generative"" in the sense that they can generate samples, fully\nprobabilistic models aim at modeling the data distribution directly.\nConsequently, they provide functionalities that are highly relevant for\ncontinual learning, such as density estimation (outlier detection) and sample\ngeneration. As a concrete realization of generative continual learning, we\npropose Gaussian Mixture Replay (GMR). GMR is a pseudo-rehearsal approach using\na Gaussian Mixture Model (GMM) instance for both generator and classifier\nfunctionalities. Relying on the MNIST, FashionMNIST and Devanagari benchmarks,\nwe first demonstrate unsupervised task boundary detection by GMM density\nestimation, which we also use to reject untypical generated samples. In\naddition, we show that GMR is capable of class-conditional sampling in the way\nof a cGAN. Lastly, we verify that GMR, despite its simple structure, achieves\nstate-of-the-art performance on common class-incremental learning problems at\nvery competitive time and memory complexity.', 'How to leverage cross-document interactions to improve ranking performance is\nan important topic in information retrieval (IR) research. However, this topic\nhas not been well-studied in the learning-to-rank setting and most of the\nexisting work still treats each document independently while scoring. The\nrecent development of deep learning shows strength in modeling complex\nrelationships across sequences and sets. It thus motivates us to study how to\nleverage cross-document interactions for learning-to-rank in the deep learning\nframework. In this paper, we formally define the permutation-equivariance\nrequirement for a scoring function that captures cross-document interactions.\nWe then propose a self-attention based document interaction network and show\nthat it satisfies the permutation-equivariant requirement, and can generate\nscores for document sets of varying sizes. Our proposed methods can\nautomatically learn to capture document interactions without any auxiliary\ninformation, and can scale across large document sets. We conduct experiments\non three ranking datasets: the benchmark Web30k, a Gmail search, and a Google\nDrive Quick Access dataset. Experimental results show that our proposed methods\nare both more effective and efficient than baselines.', ""Accurate noise modelling is important for training of deep learning\nreconstruction algorithms. While noise models are well known for traditional\nimaging techniques, the noise distribution of a novel sensor may be difficult\nto determine a priori. Therefore, we propose learning arbitrary noise\ndistributions. To do so, this paper proposes a fully connected neural network\nmodel to map samples from a uniform distribution to samples of any explicitly\nknown probability density function. During the training, the Jensen-Shannon\ndivergence between the distribution of the model's output and the target\ndistribution is minimized. We experimentally demonstrate that our model\nconverges towards the desired state. It provides an alternative to existing\nsampling methods such as inversion sampling, rejection sampling, Gaussian\nmixture models and Markov-Chain-Monte-Carlo. Our model has high sampling\nefficiency and is easily applied to any probability distribution, without the\nneed of further analytical or numerical calculations.""]","['Many engineering problems require the prediction of\nrealization-to-realization variability or a refined description of modeled\nquantities. In that case, it is necessary to sample elements from unknown\nhigh-dimensional spaces with possibly millions of degrees of freedom. While\nthere exist methods able to sample elements from probability density functions\n(PDF) with known shapes, several approximations need to be made when the\ndistribution is unknown. In this paper the sampling method, as well as the\ninference of the underlying distribution, are both handled with a data-driven\nmethod known as generative adversarial networks (GAN), which trains two\ncompeting neural networks to produce a network that can effectively generate\nsamples from the training set distribution. In practice, it is often necessary\nto draw samples from conditional distributions. When the conditional variables\nare continuous, only one (if any) data point corresponding to a particular\nvalue of a conditioning variable may be available, which is not sufficient to\nestimate the conditional distribution. This work handles this problem using an\na priori estimation of the conditional moments of a PDF. Two approaches,\nstochastic estimation, and an external neural network are compared here for\ncomputing these moments; however, any preferred method can be used. The\nalgorithm is demonstrated in the case of the deconvolution of a filtered\nturbulent flow field. It is shown that all the versions of the proposed\nalgorithm effectively sample the target conditional distribution with minimal\nimpact on the quality of the samples compared to state-of-the-art methods.\nAdditionally, the procedure can be used as a metric for the diversity of\nsamples generated by a conditional GAN (cGAN) conditioned with continuous\nvariables.', 'Extreme multi-label classification refers to supervised multi-label learning\ninvolving hundreds of thousands or even millions of labels. Datasets in extreme\nclassification exhibit fit to power-law distribution, i.e. a large fraction of\nlabels have very few positive instances in the data distribution. Most\nstate-of-the-art approaches for extreme multi-label classification attempt to\ncapture correlation among labels by embedding the label matrix to a\nlow-dimensional linear sub-space. However, in the presence of power-law\ndistributed extremely large and diverse label spaces, structural assumptions\nsuch as low rank can be easily violated.\n  In this work, we present DiSMEC, which is a large-scale distributed framework\nfor learning one-versus-rest linear classifiers coupled with explicit capacity\ncontrol to control model size. Unlike most state-of-the-art methods, DiSMEC\ndoes not make any low rank assumptions on the label matrix. Using double layer\nof parallelization, DiSMEC can learn classifiers for datasets consisting\nhundreds of thousands labels within few hours. The explicit capacity control\nmechanism filters out spurious parameters which keep the model compact in size,\nwithout losing prediction accuracy. We conduct extensive empirical evaluation\non publicly available real-world datasets consisting upto 670,000 labels. We\ncompare DiSMEC with recent state-of-the-art approaches, including - SLEEC which\nis a leading approach for learning sparse local embeddings, and FastXML which\nis a tree-based approach optimizing ranking based loss function. On some of the\ndatasets, DiSMEC can significantly boost prediction accuracies - 10% better\ncompared to SLECC and 15% better compared to FastXML, in absolute terms.', 'We focus on generative autoencoders, such as variational or adversarial\nautoencoders, which jointly learn a generative model alongside an inference\nmodel. Generative autoencoders are those which are trained to softly enforce a\nprior on the latent distribution learned by the inference model. We call the\ndistribution to which the inference model maps observed samples, the learned\nlatent distribution, which may not be consistent with the prior. We formulate a\nMarkov chain Monte Carlo (MCMC) sampling process, equivalent to iteratively\ndecoding and encoding, which allows us to sample from the learned latent\ndistribution. Since, the generative model learns to map from the learned latent\ndistribution, rather than the prior, we may use MCMC to improve the quality of\nsamples drawn from the generative model, especially when the learned latent\ndistribution is far from the prior. Using MCMC sampling, we are able to reveal\npreviously unseen differences between generative autoencoders trained either\nwith or without a denoising criterion.', 'Standard objective functions used during the training of neural-network-based\npredictive models do not consider clinical criteria, leading to models that are\nnot necessarily clinically acceptable. In this study, we look at this problem\nfrom the perspective of the forecasting of future glucose values for diabetic\npeople. In this study, we propose the coherent mean squared glycemic error\n(gcMSE) loss function. It penalizes the model during its training not only of\nthe prediction errors, but also on the predicted variation errors which is\nimportant in glucose prediction. Moreover, it makes possible to adjust the\nweighting of the different areas in the error space to better focus on\ndangerous regions. In order to use the loss function in practice, we propose an\nalgorithm that progressively improves the clinical acceptability of the model,\nso that we can achieve the best tradeoff possible between accuracy and given\nclinical criteria. We evaluate the approaches using two diabetes datasets, one\nhaving type-1 patients and the other type-2 patients. The results show that\nusing the gcMSE loss function, instead of a standard MSE loss function,\nimproves the clinical acceptability of the models. In particular, the\nimprovements are significant in the hypoglycemia region. We also show that this\nincreased clinical acceptability comes at the cost of a decrease in the average\naccuracy of the model. Finally, we show that this tradeoff between accuracy and\nclinical acceptability can be successfully addressed with the proposed\nalgorithm. For given clinical criteria, the algorithm can find the optimal\nsolution that maximizes the accuracy while at the same meeting the criteria.', 'The work investigates deep generative models, which allow us to use training\ndata from one domain to build a model for another domain. We propose the\nVariational Bi-domain Triplet Autoencoder (VBTA) that learns a joint\ndistribution of objects from different domains. We extend the VBTAs objective\nfunction by the relative constraints or triplets that sampled from the shared\nlatent space across domains. In other words, we combine the deep generative\nmodels with a metric learning ideas in order to improve the final objective\nwith the triplets information. The performance of the VBTA model is\ndemonstrated on different tasks: image-to-image translation, bi-directional\nimage generation and cross-lingual document classification.', 'The training loss function that enforces certain training sample distribution\npatterns plays a critical role in building a re-identification (ReID) system.\nBesides the basic requirement of discrimination, i.e., the features\ncorresponding to different identities should not be mixed, additional\nintra-class distribution constraints, such as features from the same identities\nshould be close to their centers, have been adopted to construct losses.\nDespite the advances of various new loss functions, it is still challenging to\nstrike the balance between the need of reducing the intra-class variation and\nallowing certain distribution freedom. In this paper, we propose a new loss\nbased on center predictivity, that is, a sample must be positioned in a\nlocation of the feature space such that from it we can roughly predict the\nlocation of the center of same-class samples. The prediction error is then\nregarded as a loss called Center Prediction Loss (CPL). We show that, without\nintroducing additional hyper-parameters, this new loss leads to a more flexible\nintra-class distribution constraint while ensuring the between-class samples\nare well-separated. Extensive experiments on various real-world ReID datasets\nshow that the proposed loss can achieve superior performance and can also be\ncomplementary to existing losses.', 'Given a task of predicting $Y$ from $X$, a loss function $L$, and a set of\nprobability distributions $\\Gamma$ on $(X,Y)$, what is the optimal decision\nrule minimizing the worst-case expected loss over $\\Gamma$? In this paper, we\naddress this question by introducing a generalization of the principle of\nmaximum entropy. Applying this principle to sets of distributions with marginal\non $X$ constrained to be the empirical marginal from the data, we develop a\ngeneral minimax approach for supervised learning problems. While for some loss\nfunctions such as squared-error and log loss, the minimax approach rederives\nwell-knwon regression models, for the 0-1 loss it results in a new linear\nclassifier which we call the maximum entropy machine. The maximum entropy\nmachine minimizes the worst-case 0-1 loss over the structured set of\ndistribution, and by our numerical experiments can outperform other well-known\nlinear classifiers such as SVM. We also prove a bound on the generalization\nworst-case error in the minimax approach.', 'Complex deep learning models now achieve state of the art performance for\nmany document retrieval tasks. The best models process the query or claim\njointly with the document. However for fast scalable search it is desirable to\nhave document embeddings which are independent of the claim. In this paper we\nshow that knowledge distillation can be used to encourage a model that\ngenerates claim independent document encodings to mimic the behavior of a more\ncomplex model which generates claim dependent encodings. We explore this\napproach in document retrieval for a fact extraction and verification task. We\nshow that by using the soft labels from a complex cross attention teacher\nmodel, the performance of claim independent student LSTM or CNN models is\nimproved across all the ranking metrics. The student models we use are 12x\nfaster in runtime and 20x smaller in number of parameters than the teacher', 'Generative adversarial networks (GANs) are very popular to generate realistic\nimages, but they often suffer from the training instability issues and the\nphenomenon of mode loss. In order to attain greater diversity in GAN\nsynthesized data, it is critical to solving the problem of mode loss. Our work\nexplores probabilistic approaches to GAN modelling that could allow us to\ntackle these issues. We present Prb-GANs, a new variation that uses dropout to\ncreate a distribution over the network parameters with the posterior learnt\nusing variational inference. We describe theoretically and validate\nexperimentally using simple and complex datasets the benefits of such an\napproach. We look into further improvements using the concept of uncertainty\nmeasures. Through a set of further modifications to the loss functions for each\nnetwork of the GAN, we are able to get results that show the improvement of GAN\nperformance. Our methods are extremely simple and require very little\nmodification to existing GAN architecture.', 'The choice of a loss function is an important factor when training neural\nnetworks for image restoration problems, such as single image super resolution.\nThe loss function should encourage natural and perceptually pleasing results. A\npopular choice for a loss is a pre-trained network, such as VGG, which is used\nas a feature extractor for computing the difference between restored and\nreference images. However, such an approach has multiple drawbacks: it is\ncomputationally expensive, requires regularization and hyper-parameter tuning,\nand involves a large network trained on an unrelated task. Furthermore, it has\nbeen observed that there is no single loss function that works best across all\napplications and across different datasets. In this work, we instead propose to\ntrain a set of loss functions that are application specific in nature. Our loss\nfunction comprises a series of discriminators that are trained to detect and\npenalize the presence of application-specific artifacts. We show that a single\nnatural image and corresponding distortions are sufficient to train our feature\nextractor that outperforms state-of-the-art loss functions in applications like\nsingle image super resolution, denoising, and JPEG artifact removal. Finally,\nwe conclude that an effective loss function does not have to be a good\npredictor of perceived image quality, but instead needs to be specialized in\nidentifying the distortions for a given restoration method.', 'Word embedding maps words into a low-dimensional continuous embedding space\nby exploiting the local word collocation patterns in a small context window. On\nthe other hand, topic modeling maps documents onto a low-dimensional topic\nspace, by utilizing the global word collocation patterns in the same document.\nThese two types of patterns are complementary. In this paper, we propose a\ngenerative topic embedding model to combine the two types of patterns. In our\nmodel, topics are represented by embedding vectors, and are shared across\ndocuments. The probability of each word is influenced by both its local context\nand its topic. A variational inference method yields the topic embeddings as\nwell as the topic mixing proportions for each document. Jointly they represent\nthe document in a low-dimensional continuous space. In two document\nclassification tasks, our method performs better than eight existing methods,\nwith fewer features. In addition, we illustrate with an example that our method\ncan generate coherent topics even based on only one document.', 'Proposed methods for prediction interval estimation so far focus on cases\nwhere input variables are numerical. In datasets with solely nominal input\nvariables, we observe records with the exact same input $x^u$, but different\nreal valued outputs due to the inherent noise in the system. Existing\nprediction interval estimation methods do not use representations that can\naccurately model such inherent noise in the case of nominal inputs. We propose\na new prediction interval estimation method tailored for this type of data,\nwhich is prevalent in biology and medicine. We call this method Distribution\nAdaptive Prediction Interval Estimation given Nominal inputs (DAPIEN) and has\nfour main phases. First, we select a distribution function that can best\nrepresent the inherent noise of the system for all unique inputs. Then we infer\nthe parameters $\\theta_i$ (e.g. $\\theta_i=[mean_i, variance_i]$) of the\nselected distribution function for all unique input vectors $x^u_i$ and\ngenerate a new corresponding training set using pairs of $x^u_i, \\theta_i$.\nIII). Then, we train a model to predict $\\theta$ given a new $x_u$. Finally, we\ncalculate the prediction interval for a new sample using the inverse of the\ncumulative distribution function once the parameters $\\theta$ is predicted by\nthe trained model. We compared DAPIEN to the commonly used Bootstrap method on\nthree synthetic datasets. Our results show that DAPIEN provides tighter\nprediction intervals while preserving the requested coverage when compared to\nBootstrap. This work can facilitate broader usage of regression methods in\nmedicine and biology where it is necessary to provide tight prediction\nintervals while preserving coverage when input variables are nominal.', 'Large-scale pretrained language models have led to dramatic improvements in\ntext generation. Impressive performance can be achieved by finetuning only on a\nsmall number of instances (few-shot setting). Nonetheless, almost all previous\nwork simply applies random sampling to select the few-shot training instances.\nLittle to no attention has been paid to the selection strategies and how they\nwould affect model performance. In this work, we present a study on training\ninstance selection in few-shot neural text generation. The selection decision\nis made based only on the unlabeled data so as to identify the most worthwhile\ndata points that should be annotated under some budget of labeling cost. Based\non the intuition that the few-shot training instances should be diverse and\nrepresentative of the entire data distribution, we propose a simple selection\nstrategy with K-means clustering. We show that even with the naive\nclustering-based approach, the generation models consistently outperform random\nsampling on three text generation tasks: data-to-text generation, document\nsummarization and question generation. We hope that this work will call for\nmore attention on this largely unexplored area.', 'A basic requirement for a mathematical model is often that its solution\n(output) shouldn\'t change much if the model\'s parameters (input) are perturbed.\nThis is important because the exact values of parameters may not be known and\none would like to avoid being mislead by an output obtained using incorrect\nvalues. Thus, it\'s rarely enough to address an application by formulating a\nmodel, solving the resulting optimization problem and presenting the solution\nas the answer. One would need to confirm that the model is suitable, i.e.,\n""good,"" and this can, at least in part, be achieved by considering a family of\noptimization problems constructed by perturbing parameters of concern. The\nresulting sensitivity analysis uncovers troubling situations with unstable\nsolutions, which we referred to as ""bad"" models, and indicates better model\nformulations. Embedding an actual problem of interest within a family of\nproblems is also a primary path to optimality conditions as well as\ncomputationally attractive, alternative problems, which under ideal\ncircumstances, and when properly tuned, may even furnish the minimum value of\nthe actual problem. The tuning of these alternative problems turns out to be\nintimately tied to finding multipliers in optimality conditions and thus\nemerges as a main component of several optimization algorithms. In fact, the\ntuning amounts to solving certain dual optimization problems. In this tutorial,\nwe\'ll discuss the opportunities and insights afforded by this broad\nperspective.', ""Neural networks are increasingly used to estimate parameters in quantitative\nMRI, in particular in magnetic resonance fingerprinting. Their advantages over\nthe gold standard non-linear least square fitting are their superior speed and\ntheir immunity to the non-convexity of many fitting problems. We find, however,\nthat in heterogeneous parameter spaces, i.e. in spaces in which the variance of\nthe estimated parameters varies considerably, good performance is hard to\nachieve and requires arduous tweaking of the loss function, hyper parameters,\nand the distribution of the training data in parameter space. Here, we address\nthese issues with a theoretically well-founded loss function: the Cram\\'er-Rao\nbound (CRB) provides a theoretical lower bound for the variance of an unbiased\nestimator and we propose to normalize the squared error with respective CRB.\nWith this normalization, we balance the contributions of hard-to-estimate and\nnot-so-hard-to-estimate parameters and areas in parameter space, and avoid a\ndominance of the former in the overall training loss. Further, the CRB-based\nloss function equals one for a maximally-efficient unbiased estimator, which we\nconsider the ideal estimator. Hence, the proposed CRB-based loss function\nprovides an absolute evaluation metric. We compare a network trained with the\nCRB-based loss with a network trained with the commonly used means squared\nerror loss and demonstrate the advantages of the former in numerical, phantom,\nand in vivo experiments."", 'Traditional Relational Topic Models provide a way to discover the hidden\ntopics from a document network. Many theoretical and practical tasks, such as\ndimensional reduction, document clustering, link prediction, benefit from this\nrevealed knowledge. However, existing relational topic models are based on an\nassumption that the number of hidden topics is known in advance, and this is\nimpractical in many real-world applications. Therefore, in order to relax this\nassumption, we propose a nonparametric relational topic model in this paper.\nInstead of using fixed-dimensional probability distributions in its generative\nmodel, we use stochastic processes. Specifically, a gamma process is assigned\nto each document, which represents the topic interest of this document.\nAlthough this method provides an elegant solution, it brings additional\nchallenges when mathematically modeling the inherent network structure of\ntypical document network, i.e., two spatially closer documents tend to have\nmore similar topics. Furthermore, we require that the topics are shared by all\nthe documents. In order to resolve these challenges, we use a subsampling\nstrategy to assign each document a different gamma process from the global\ngamma process, and the subsampling probabilities of documents are assigned with\na Markov Random Field constraint that inherits the document network structure.\nThrough the designed posterior inference algorithm, we can discover the hidden\ntopics and its number simultaneously. Experimental results on both synthetic\nand real-world network datasets demonstrate the capabilities of learning the\nhidden topics and, more importantly, the number of topics.', 'Despite their empirical success, most existing listwiselearning-to-rank (LTR)\nmodels are not built to be robust to errors in labeling or annotation,\ndistributional data shift, or adversarial data perturbations. To fill this gap,\nwe introduce a new listwise LTR model called Distributionally Robust\nMulti-output Regression Ranking (DRMRR). Different from existing methods, the\nscoring function of DRMRR was designed as a multivariate mapping from a feature\nvector to a vector of deviation scores, which captures local context\ninformation and cross-document interactions. DRMRR uses a Distributionally\nRobust Optimization (DRO) framework to minimize a multi-output loss function\nunder the most adverse distributions in the neighborhood of the empirical data\ndistribution defined by a Wasserstein ball. We show that this is equivalent to\na regularized regression problem with a matrix norm regularizer. Our\nexperiments were conducted on two real-world applications, medical document\nretrieval, and drug response prediction, showing that DRMRR notably outperforms\nstate-of-the-art LTR models. We also conducted a comprehensive analysis to\nassess the resilience of DRMRR against various types of noise: Gaussian noise,\nadversarial perturbations, and label poisoning. We show that DRMRR is not only\nable to achieve significantly better performance than other baselines, but it\ncan maintain a relatively stable performance as more noise is added to the\ndata.', 'Generative adversarial networks (GANs) are innovative techniques for learning\ngenerative models of complex data distributions from samples. Despite\nremarkable recent improvements in generating realistic images, one of their\nmajor shortcomings is the fact that in practice, they tend to produce samples\nwith little diversity, even when trained on diverse datasets. This phenomenon,\nknown as mode collapse, has been the main focus of several recent advances in\nGANs. Yet there is little understanding of why mode collapse happens and why\nexisting approaches are able to mitigate mode collapse. We propose a principled\napproach to handling mode collapse, which we call packing. The main idea is to\nmodify the discriminator to make decisions based on multiple samples from the\nsame class, either real or artificially generated. We borrow analysis tools\nfrom binary hypothesis testing---in particular the seminal result of Blackwell\n[Bla53]---to prove a fundamental connection between packing and mode collapse.\nWe show that packing naturally penalizes generators with mode collapse, thereby\nfavoring generator distributions with less mode collapse during the training\nprocess. Numerical experiments on benchmark datasets suggests that packing\nprovides significant improvements in practice as well.', 'In clinical trials and other applications, we often see regions of the\nfeature space that appear to exhibit interesting behaviour, but it is unclear\nwhether these observed phenomena are reflected at the population level.\nFocusing on a regression setting, we consider the subgroup selection challenge\nof identifying a region of the feature space on which the regression function\nexceeds a pre-determined threshold. We formulate the problem as one of\nconstrained optimisation, where we seek a low-complexity, data-dependent\nselection set on which, with a guaranteed probability, the regression function\nis uniformly at least as large as the threshold; subject to this constraint, we\nwould like the region to contain as much mass under the marginal feature\ndistribution as possible. This leads to a natural notion of regret, and our\nmain contribution is to determine the minimax optimal rate for this regret in\nboth the sample size and the Type I error probability. The rate involves a\ndelicate interplay between parameters that control the smoothness of the\nregression function, as well as exponents that quantify the extent to which the\noptimal selection set at the population level can be approximated by families\nof well-behaved subsets. Finally, we expand the scope of our previous results\nby illustrating how they may be generalised to a treatment and control setting,\nwhere interest lies in the heterogeneous treatment effect.', 'Latent topic models have been successfully applied as an unsupervised topic\ndiscovery technique in large document collections. With the proliferation of\nhypertext document collection such as the Internet, there has also been great\ninterest in extending these approaches to hypertext [6, 9]. These approaches\ntypically model links in an analogous fashion to how they model words - the\ndocument-link co-occurrence matrix is modeled in the same way that the\ndocument-word co-occurrence matrix is modeled in standard topic models. In this\npaper we present a probabilistic generative model for hypertext document\ncollections that explicitly models the generation of links. Specifically, links\nfrom a word w to a document d depend directly on how frequent the topic of w is\nin d, in addition to the in-degree of d. We show how to perform EM learning on\nthis model efficiently. By not modeling links as analogous to words, we end up\nusing far fewer free parameters and obtain better link prediction results.', 'Recently, the computer vision and machine learning community has been in\nfavor of feature extraction pipelines that rely on a coding step followed by a\nlinear classifier, due to their overall simplicity, well understood properties\nof linear classifiers, and their computational efficiency. In this paper we\npropose a novel view of this pipeline based on kernel methods and Nystrom\nsampling. In particular, we focus on the coding of a data point with a local\nrepresentation based on a dictionary with fewer elements than the number of\ndata points, and view it as an approximation to the actual function that would\ncompute pair-wise similarity to all data points (often too many to compute in\npractice), followed by a Nystrom sampling step to select a subset of all data\npoints.\n  Furthermore, since bounds are known on the approximation power of Nystrom\nsampling as a function of how many samples (i.e. dictionary size) we consider,\nwe can derive bounds on the approximation of the exact (but expensive to\ncompute) kernel matrix, and use it as a proxy to predict accuracy as a function\nof the dictionary size, which has been observed to increase but also to\nsaturate as we increase its size. This model may help explaining the positive\neffect of the codebook size and justifying the need to stack more layers (often\nreferred to as deep learning), as flat models empirically saturate as we add\nmore complexity.', 'At the present time, sequential item recommendation models are compared by\ncalculating metrics on a small item subset (target set) to speed up\ncomputation. The target set contains the relevant item and a set of negative\nitems that are sampled from the full item set. Two well-known strategies to\nsample negative items are uniform random sampling and sampling by popularity to\nbetter approximate the item frequency distribution in the dataset. Most\nrecently published papers on sequential item recommendation rely on sampling by\npopularity to compare the evaluated models. However, recent work has already\nshown that an evaluation with uniform random sampling may not be consistent\nwith the full ranking, that is, the model ranking obtained by evaluating a\nmetric using the full item set as target set, which raises the question whether\nthe ranking obtained by sampling by popularity is equal to the full ranking. In\nthis work, we re-evaluate current state-of-the-art sequential recommender\nmodels from the point of view, whether these sampling strategies have an impact\non the final ranking of the models. We therefore train four recently proposed\nsequential recommendation models on five widely known datasets. For each\ndataset and model, we employ three evaluation strategies. First, we compute the\nfull model ranking. Then we evaluate all models on a target set sampled by the\ntwo different sampling strategies, uniform random sampling and sampling by\npopularity with the commonly used target set size of 100, compute the model\nranking for each strategy and compare them with each other. Additionally, we\nvary the size of the sampled target set. Overall, we find that both sampling\nstrategies can produce inconsistent rankings compared with the full ranking of\nthe models. Furthermore, both sampling by popularity and uniform random\nsampling do not consistently produce the same ranking ...', 'A learned generative model often produces biased statistics relative to the\nunderlying data distribution. A standard technique to correct this bias is\nimportance sampling, where samples from the model are weighted by the\nlikelihood ratio under model and true distributions. When the likelihood ratio\nis unknown, it can be estimated by training a probabilistic classifier to\ndistinguish samples from the two distributions. We employ this likelihood-free\nimportance weighting method to correct for the bias in generative models. We\nfind that this technique consistently improves standard goodness-of-fit metrics\nfor evaluating the sample quality of state-of-the-art deep generative models,\nsuggesting reduced bias. Finally, we demonstrate its utility on representative\napplications in a) data augmentation for classification using generative\nadversarial networks, and b) model-based policy evaluation using off-policy\ndata.', ""In most machine learning tasks, we evaluate a model $M$ on a given data\npopulation $S$ by measuring a population-level metric $F(S;M)$. Examples of\nsuch evaluation metric $F$ include precision/recall for (binary) recognition,\nthe F1 score for multi-class classification, and the BLEU metric for language\ngeneration. On the other hand, the model $M$ is trained by optimizing a\nsample-level loss $G(S_t;M)$ at each learning step $t$, where $S_t$ is a subset\nof $S$ (a.k.a. the mini-batch). Popular choices of $G$ include cross-entropy\nloss, the Dice loss, and sentence-level BLEU scores. A fundamental assumption\nbehind this paradigm is that the mean value of the sample-level loss $G$, if\naveraged over all possible samples, should effectively represent the\npopulation-level metric $F$ of the task, such as, that $\\mathbb{E}[ G(S_t;M) ]\n\\approx F(S;M)$.\n  In this paper, we systematically investigate the above assumption in several\nNLP tasks. We show, both theoretically and experimentally, that some popular\ndesigns of the sample-level loss $G$ may be inconsistent with the true\npopulation-level metric $F$ of the task, so that models trained to optimize the\nformer can be substantially sub-optimal to the latter, a phenomenon we call it,\nSimpson's bias, due to its deep connections with the classic paradox known as\nSimpson's reversal paradox in statistics and social sciences."", 'Generative adversarial networks (GANs) provide state-of-the-art results in\nimage generation. However, despite being so powerful, they still remain very\nchallenging to train. This is in particular caused by their highly non-convex\noptimization space leading to a number of instabilities. Among them, mode\ncollapse stands out as one of the most daunting ones. This undesirable event\noccurs when the model can only fit a few modes of the data distribution, while\nignoring the majority of them. In this work, we combat mode collapse using\nsecond-order gradient information. To do so, we analyse the loss surface\nthrough its Hessian eigenvalues, and show that mode collapse is related to the\nconvergence towards sharp minima. In particular, we observe how the eigenvalues\nof the $G$ are directly correlated with the occurrence of mode collapse.\nFinally, motivated by these findings, we design a new optimization algorithm\ncalled nudged-Adam (NuGAN) that uses spectral information to overcome mode\ncollapse, leading to empirically more stable convergence properties.', 'Due to the increasing amount of data on the internet, finding a\nhighly-informative, low-dimensional representation for text is one of the main\nchallenges for efficient natural language processing tasks including text\nclassification. This representation should capture the semantic information of\nthe text while retaining their relevance level for document classification.\nThis approach maps the documents with similar topics to a similar space in\nvector space representation. To obtain representation for large text, we\npropose the utilization of deep Siamese neural networks. To embed document\nrelevance in topics in the distributed representation, we use a Siamese neural\nnetwork to jointly learn document representations. Our Siamese network consists\nof two sub-network of multi-layer perceptron. We examine our representation for\nthe text categorization task on BBC news dataset. The results show that the\nproposed representations outperform the conventional and state-of-the-art\nrepresentations in the text classification task on this dataset.', 'We analyse optimum reject strategies for prototype-based classifiers and\nreal-valued rejection measures, using the distance of a data point to the\nclosest prototype or probabilistic counterparts. We compare reject schemes with\nglobal thresholds, and local thresholds for the Voronoi cells of the\nclassifier. For the latter, we develop a polynomial-time algorithm to compute\noptimum thresholds based on a dynamic programming scheme, and we propose an\nintuitive linear time, memory efficient approximation thereof with competitive\naccuracy. Evaluating the performance in various benchmarks, we conclude that\nlocal reject options are beneficial in particular for simple prototype-based\nclassifiers, while the improvement is less pronounced for advanced models. For\nthe latter, an accuracy-reject curve which is comparable to support vector\nmachine classifiers with state of the art reject options can be reached.', 'Document network embedding aims at learning representations for a structured\ntext corpus i.e. when documents are linked to each other. Recent algorithms\nextend network embedding approaches by incorporating the text content\nassociated with the nodes in their formulations. In most cases, it is hard to\ninterpret the learned representations. Moreover, little importance is given to\nthe generalization to new documents that are not observed within the network.\nIn this paper, we propose an interpretable and inductive document network\nembedding method. We introduce a novel mechanism, the Topic-Word Attention\n(TWA), that generates document representations based on the interplay between\nword and topic representations. We train these word and topic vectors through\nour general model, Inductive Document Network Embedding (IDNE), by leveraging\nthe connections in the document network. Quantitative evaluations show that our\napproach achieves state-of-the-art performance on various networks and we\nqualitatively show that our model produces meaningful and interpretable\nrepresentations of the words, topics and documents.', 'This paper introduces an alternative approach to sampling from autoregressive\nmodels. Autoregressive models are typically sampled sequentially, according to\nthe transition dynamics defined by the model. Instead, we propose a sampling\nprocedure that initializes a sequence with white noise and follows a Markov\nchain defined by Langevin dynamics on the global log-likelihood of the\nsequence. This approach parallelizes the sampling process and generalizes to\nconditional sampling. Using an autoregressive model as a Bayesian prior, we can\nsteer the output of a generative model using a conditional likelihood or\nconstraints. We apply these techniques to autoregressive models in the visual\nand audio domains, with competitive results for audio source separation,\nsuper-resolution, and inpainting.', 'In this paper we study the problem of learning from multiple modal data for\npurpose of document classification. In this problem, each document is composed\ntwo different modals of data, i.e., an image and a text. Cross-modal factor\nanalysis (CFA) has been proposed to project the two different modals of data to\na shared data space, so that the classification of a image or a text can be\nperformed directly in this space. A disadvantage of CFA is that it has ignored\nthe supervision information. In this paper, we improve CFA by incorporating the\nsupervision information to represent and classify both image and text modals of\ndocuments. We project both image and text data to a shared data space by factor\nanalysis, and then train a class label predictor in the shared space to use the\nclass label information. The factor analysis parameter and the predictor\nparameter are learned jointly by solving one single objective function. With\nthis objective function, we minimize the distance between the projections of\nimage and text of the same document, and the classification error of the\nprojection measured by hinge loss function. The objective function is optimized\nby an alternate optimization strategy in an iterative algorithm. Experiments in\ntwo different multiple modal document data sets show the advantage of the\nproposed algorithm over other CFA methods.', 'The predictive quality of machine learning models is typically measured in\nterms of their (approximate) expected prediction accuracy or the so-called Area\nUnder the Curve (AUC). Minimizing the reciprocals of these measures are the\ngoals of supervised learning. However, when the models are constructed by the\nmeans of empirical risk minimization (ERM), surrogate functions such as the\nlogistic loss or hinge loss are optimized instead. In this work, we show that\nin the case of linear predictors, the expected error and the expected ranking\nloss can be effectively approximated by smooth functions whose closed form\nexpressions and those of their first (and second) order derivatives depend on\nthe first and second moments of the data distribution, which can be\nprecomputed. Hence, the complexity of an optimization algorithm applied to\nthese functions does not depend on the size of the training data. These\napproximation functions are derived under the assumption that the output of the\nlinear classifier for a given data set has an approximately normal\ndistribution. We argue that this assumption is significantly weaker than the\nGaussian assumption on the data itself and we support this claim by\ndemonstrating that our new approximation is quite accurate on data sets that\nare not necessarily Gaussian. We present computational results that show that\nour proposed approximations and related optimization algorithms can produce\nlinear classifiers with similar or better test accuracy or AUC, than those\nobtained using state-of-the-art approaches, in a fraction of the time.', ""BERT based ranking models have achieved superior performance on various\ninformation retrieval tasks. However, the large number of parameters and\ncomplex self-attention operation come at a significant latency overhead. To\nremedy this, recent works propose late-interaction architectures, which allow\npre-computation of intermediate document representations, thus reducing the\nruntime latency. Nonetheless, having solved the immediate latency issue, these\nmethods now introduce storage costs and network fetching latency, which limits\ntheir adoption in real-life production systems.\n  In this work, we propose the Succinct Document Representation (SDR) scheme\nthat computes highly compressed intermediate document representations,\nmitigating the storage/network issue. Our approach first reduces the dimension\nof token representations by encoding them using a novel autoencoder\narchitecture that uses the document's textual content in both the encoding and\ndecoding phases. After this token encoding step, we further reduce the size of\nentire document representations using a modern quantization technique.\n  Extensive evaluations on passage re-reranking on the MSMARCO dataset show\nthat compared to existing approaches using compressed document representations,\nour method is highly efficient, achieving 4x-11.6x better compression rates for\nthe same ranking quality."", 'We consider local planning in fixed-horizon MDPs with a generative model\nunder the assumption that the optimal value function lies close to the span of\na feature map. The generative model provides a local access to the MDP: The\nplanner can ask for random transitions from previously returned states and\narbitrary actions, and features are only accessible for states that are\nencountered in this process. As opposed to previous work (e.g. Lattimore et al.\n(2020)) where linear realizability of all policies was assumed, we consider the\nsignificantly relaxed assumption of a single linearly realizable\n(deterministic) policy. A recent lower bound by Weisz et al. (2020) established\nthat the related problem when the action-value function of the optimal policy\nis linearly realizable requires an exponential number of queries, either in $H$\n(the horizon of the MDP) or $d$ (the dimension of the feature mapping). Their\nconstruction crucially relies on having an exponentially large action set. In\ncontrast, in this work, we establish that poly$(H,d)$ planning is possible with\nstate value function realizability whenever the action set has a constant size.\nIn particular, we present the TensorPlan algorithm which uses\npoly$((dH/\\delta)^A)$ simulator queries to find a $\\delta$-optimal policy\nrelative to any deterministic policy for which the value function is linearly\nrealizable with some bounded parameter. This is the first algorithm to give a\npolynomial query complexity guarantee using only linear-realizability of a\nsingle competing value function. Whether the computation cost is similarly\nbounded remains an open question. We extend the upper bound to the\nnear-realizable case and to the infinite-horizon discounted setup. We also\npresent a lower bound in the infinite-horizon episodic setting: Planners that\nachieve constant suboptimality need exponentially many queries, either in $d$\nor the number of actions.', 'In neural networks, the loss function represents the core of the learning\nprocess that leads the optimizer to an approximation of the optimal convergence\nerror. Convolutional neural networks (CNN) use the loss function as a\nsupervisory signal to train a deep model and contribute significantly to\nachieving the state of the art in some fields of artificial vision.\nCross-entropy and Center loss functions are commonly used to increase the\ndiscriminating power of learned functions and increase the generalization\nperformance of the model. Center loss minimizes the class intra-class variance\nand at the same time penalizes the long distance between the deep features\ninside each class. However, the total error of the center loss will be heavily\ninfluenced by the majority of the instances and can lead to a freezing state in\nterms of intra-class variance. To address this, we introduce a new loss\nfunction called sigma squared reduction loss ($\\sigma^2$R loss), which is\nregulated by a sigmoid function to inflate/deflate the error per instance and\nthen continue to reduce the intra-class variance. Our loss has clear intuition\nand geometric interpretation, furthermore, we demonstrate by experiments the\neffectiveness of our proposal on several benchmark datasets showing the\nintra-class variance reduction and overcoming the results obtained with center\nloss and soft nearest neighbour functions.', ""Several interesting generative learning algorithms involve a complex\nprobability distribution over many random variables, involving intractable\nnormalization constants or latent variable normalization. Some of them may even\nnot have an analytic expression for the unnormalized probability function and\nno tractable approximation. This makes it difficult to estimate the quality of\nthese models, once they have been trained, or to monitor their quality (e.g.\nfor early stopping) while training. A previously proposed method is based on\nconstructing a non-parametric density estimator of the model's probability\nfunction from samples generated by the model. We revisit this idea, propose a\nmore efficient estimator, and prove that it provides a lower bound on the true\ntest log-likelihood, and an unbiased estimator as the number of generated\nsamples goes to infinity, although one that incorporates the effect of poor\nmixing. We further propose a biased variant of the estimator that can be used\nreliably with a finite number of samples for the purpose of model comparison."", 'Training supervised image synthesis models requires a critic to compare two\nimages: the ground truth to the result. Yet, this basic functionality remains\nan open problem. A popular line of approaches uses the L1 (mean absolute error)\nloss, either in the pixel or the feature space of pretrained deep networks.\nHowever, we observe that these losses tend to produce overly blurry and grey\nimages, and other techniques such as GANs need to be employed to fight these\nartifacts. In this work, we introduce an information theory based approach to\nmeasuring similarity between two images. We argue that a good reconstruction\nshould have high mutual information with the ground truth. This view enables\nlearning a lightweight critic to ""calibrate"" a feature space in a contrastive\nmanner, such that reconstructions of corresponding spatial patches are brought\ntogether, while other patches are repulsed. We show that our formulation\nimmediately boosts the perceptual realism of output images when used as a\ndrop-in replacement for the L1 loss, with or without an additional GAN loss.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'The Huber loss is a robust loss function used for a wide range of regression\ntasks. To utilize the Huber loss, a parameter that controls the transitions\nfrom a quadratic function to an absolute value function needs to be selected.\nWe believe the standard probabilistic interpretation that relates the Huber\nloss to the Huber density fails to provide adequate intuition for identifying\nthe transition point. As a result, a hyper-parameter search is often necessary\nto determine an appropriate value. In this work, we propose an alternative\nprobabilistic interpretation of the Huber loss, which relates minimizing the\nloss to minimizing an upper-bound on the Kullback-Leibler divergence between\nLaplace distributions, where one distribution represents the noise in the\nground-truth and the other represents the noise in the prediction. In addition,\nwe show that the parameters of the Laplace distributions are directly related\nto the transition point of the Huber loss. We demonstrate, through a toy\nproblem, that the optimal transition point of the Huber loss is closely related\nto the distribution of the noise in the ground-truth data. As a result, our\ninterpretation provides an intuitive way to identify well-suited\nhyper-parameters by approximating the amount of noise in the data, which we\ndemonstrate through a case study and experimentation on the Faster R-CNN and\nRetinaNet object detectors.', 'We look at a specific aspect of model interpretability: models often need to\nbe constrained in size for them to be considered interpretable. But smaller\nmodels also tend to have high bias. This suggests a trade-off between\ninterpretability and accuracy. Our work addresses this by: (a) showing that\nlearning a training distribution (often different from the test distribution)\ncan often increase accuracy of small models, and therefore may be used as a\nstrategy to compensate for small sizes, and (b) providing a model-agnostic\nalgorithm to learn such training distributions. We pose the distribution\nlearning problem as one of optimizing parameters for an Infinite Beta Mixture\nModel based on a Dirichlet Process, so that the held-out accuracy of a model\ntrained on a sample from this distribution is maximized. To make computation\ntractable, we project the training data onto one dimension: prediction\nuncertainty scores as provided by a highly accurate oracle model. A Bayesian\nOptimizer is used for learning the parameters. Empirical results using multiple\nreal world datasets, various oracles and interpretable models with different\nnotions of model sizes, are presented. We observe significant relative\nimprovements in the F1-score in most cases, occasionally seeing improvements\ngreater than 100% over baselines. Additionally we show that the proposed\nalgorithm provides the following benefits: (a) its a framework which allows for\nflexibility in implementation, (b) it can be used across feature spaces, e.g.,\nthe text classification accuracy of a Decision Tree using character n-grams is\nshown to improve when using a Gated Recurrent Unit as an oracle, which uses a\nsequence of characters as its input, (c) it can be used to train models that\nhave a non-differentiable training loss, e.g., Decision Trees, and (d)\nreasonable defaults exist for most parameters of the algorithm, which makes it\nconvenient to use.', 'In a regression setting we propose algorithms that reduce the dimensionality\nof the features while simultaneously maximizing a statistical measure of\ndependence known as distance correlation between the low-dimensional features\nand a response variable. This helps in solving the prediction problem with a\nlow-dimensional set of features. Our setting is different from subset-selection\nalgorithms where the problem is to choose the best subset of features for\nregression. Instead, we attempt to generate a new set of low-dimensional\nfeatures as in a feature-learning setting. We attempt to keep our proposed\napproach as model-free and our algorithm does not assume the application of any\nspecific regression model in conjunction with the low-dimensional features that\nit learns. The algorithm is iterative and is fomulated as a combination of the\nmajorization-minimization and concave-convex optimization procedures. We also\npresent spectral radius based convergence results for the proposed iterations.', 'Causal machine-learning is about predicting the net-effect (true-lift) of\ntreatments. Given the data of a treatment group and a control group, it is\nsimilar to a standard supervised-learning problem. Unfortunately, there is no\nsimilarly well-defined loss function due to the lack of point-wise true values\nin the data. Many advances in modern machine-learning are not directly\napplicable due to the absence of such loss function.\n  We propose a novel method to define a loss function in this context, which is\nequal to mean-square-error (MSE) in a standard regression problem. Our loss\nfunction is universally applicable, thus providing a general standard to\nevaluate the quality of any model/strategy that predicts the true-lift. We\ndemonstrate that despite its novel definition, one can still perform gradient\ndescent directly on this loss function to find the best fit. This leads to a\nnew way to train any parameter-based model, such as deep neural networks, to\nsolve causal machine-learning problems without going through the meta-learner\nstrategy.', 'Set prediction is about learning to predict a collection of unordered\nvariables with unknown interrelations. Training such models with set losses\nimposes the structure of a metric space over sets. We focus on stochastic and\nunderdefined cases, where an incorrectly chosen loss function leads to\nimplausible predictions. Example tasks include conditional point-cloud\nreconstruction and predicting future states of molecules. In this paper, we\npropose an alternative to training via set losses by viewing learning as\nconditional density estimation. Our learning framework fits deep energy-based\nmodels and approximates the intractable likelihood with gradient-guided\nsampling. Furthermore, we propose a stochastically augmented prediction\nalgorithm that enables multiple predictions, reflecting the possible variations\nin the target set. We empirically demonstrate on a variety of datasets the\ncapability to learn multi-modal densities and produce different plausible\npredictions. Our approach is competitive with previous set prediction models on\nstandard benchmarks. More importantly, it extends the family of addressable\ntasks beyond those that have unambiguous predictions.', 'The Maximum Mean Discrepancy (MMD) has been the state-of-the-art\nnonparametric test for tackling the two-sample problem. Its statistic is given\nby the difference in expectations of the witness function, a real-valued\nfunction defined as a weighted sum of kernel evaluations on a set of basis\npoints. Typically the kernel is optimized on a training set, and hypothesis\ntesting is performed on a separate test set to avoid overfitting (i.e., control\ntype-I error). That is, the test set is used to simultaneously estimate the\nexpectations and define the basis points, while the training set only serves to\nselect the kernel and is discarded. In this work, we propose to use the\ntraining data to also define the weights and the basis points for better data\nefficiency. We show that 1) the new test is consistent and has a\nwell-controlled type-I error; 2) the optimal witness function is given by a\nprecision-weighted mean in the reproducing kernel Hilbert space associated with\nthe kernel; and 3) the test power of the proposed test is comparable or exceeds\nthat of the MMD and other modern tests, as verified empirically on challenging\nsynthetic and real problems (e.g., Higgs data).', 'Deep learning has been immensely successful at a variety of tasks, ranging\nfrom classification to AI. Learning corresponds to fitting training data, which\nis implemented by descending a very high-dimensional loss function.\nUnderstanding under which conditions neural networks do not get stuck in poor\nminima of the loss, and how the landscape of that loss evolves as depth is\nincreased remains a challenge. Here we predict, and test empirically, an\nanalogy between this landscape and the energy landscape of repulsive ellipses.\nWe argue that in FC networks a phase transition delimits the over- and\nunder-parametrized regimes where fitting can or cannot be achieved. In the\nvicinity of this transition, properties of the curvature of the minima of the\nloss are critical. This transition shares direct similarities with the jamming\ntransition by which particles form a disordered solid as the density is\nincreased, which also occurs in certain classes of computational optimization\nand learning problems such as the perceptron. Our analysis gives a simple\nexplanation as to why poor minima of the loss cannot be encountered in the\noverparametrized regime, and puts forward the surprising result that the\nability of fully connected networks to fit random data is independent of their\ndepth. Our observations suggests that this independence also holds for real\ndata. We also study a quantity $\\Delta$ which characterizes how well\n($\\Delta<0$) or badly ($\\Delta>0$) a datum is learned. At the critical point it\nis power-law distributed, $P_+(\\Delta)\\sim\\Delta^\\theta$ for $\\Delta>0$ and\n$P_-(\\Delta)\\sim(-\\Delta)^{-\\gamma}$ for $\\Delta<0$, with $\\theta\\approx0.3$\nand $\\gamma\\approx0.2$. This observation suggests that near the transition the\nloss landscape has a hierarchical structure and that the learning dynamics is\nprone to avalanche-like dynamics, with abrupt changes in the set of patterns\nthat are learned.', 'Recent methods for knowledge grounded dialogs generate responses by\nincorporating information from an external textual document. These methods do\nnot require the exact document to be known during training and rely on the use\nof a retrieval system to fetch relevant documents from a large index. The\ndocuments used to generate the responses are modeled as latent variables whose\nprior probabilities need to be estimated. Models such as RAG and REALM,\nmarginalize the document probabilities over the documents retrieved from the\nindex to define the log likelihood loss function which is optimized end-to-end.\n  In this paper, we develop a variational approach to the above technique\nwherein, we instead maximize the Evidence Lower bound (ELBO). Using a\ncollection of three publicly available open-conversation datasets, we\ndemonstrate how the posterior distribution, that has information from the\nground-truth response, allows for a better approximation of the objective\nfunction during training. To overcome the challenges associated with sampling\nover a large knowledge collection, we develop an efficient approach to\napproximate the ELBO. To the best of our knowledge we are the first to apply\nvariational training for open-scale unsupervised knowledge grounded dialog\nsystems.', ""A common assumption in causal modeling posits that the data is generated by a\nset of independent mechanisms, and algorithms should aim to recover this\nstructure. Standard unsupervised learning, however, is often concerned with\ntraining a single model to capture the overall distribution or aspects thereof.\nInspired by clustering approaches, we consider mixtures of implicit generative\nmodels that ``disentangle'' the independent generative mechanisms underlying\nthe data. Relying on an additional set of discriminators, we propose a\ncompetitive training procedure in which the models only need to capture the\nportion of the data distribution from which they can produce realistic samples.\nAs a by-product, each model is simpler and faster to train. We empirically show\nthat our approach splits the training distribution in a sensible way and\nincreases the quality of the generated samples."", 'This work focuses on combining nonparametric topic models with Auto-Encoding\nVariational Bayes (AEVB). Specifically, we first propose iTM-VAE, where the\ntopics are treated as trainable parameters and the document-specific topic\nproportions are obtained by a stick-breaking construction. The inference of\niTM-VAE is modeled by neural networks such that it can be computed in a simple\nfeed-forward manner. We also describe how to introduce a hyper-prior into\niTM-VAE so as to model the uncertainty of the prior parameter. Actually, the\nhyper-prior technique is quite general and we show that it can be applied to\nother AEVB based models to alleviate the {\\it collapse-to-prior} problem\nelegantly. Moreover, we also propose HiTM-VAE, where the document-specific\ntopic distributions are generated in a hierarchical manner. HiTM-VAE is even\nmore flexible and can generate topic distributions with better variability.\nExperimental results on 20News and Reuters RCV1-V2 datasets show that the\nproposed models outperform the state-of-the-art baselines significantly. The\nadvantages of the hyper-prior technique and the hierarchical model construction\nare also confirmed by experiments.', 'In this work we study loss functions for learning and evaluating probability\ndistributions over large discrete domains. Unlike classification or regression\nwhere a wide variety of loss functions are used, in the distribution learning\nand density estimation literature, very few losses outside the dominant $log\\\nloss$ are applied. We aim to understand this fact, taking an axiomatic approach\nto the design of loss functions for learning distributions. We start by\nproposing a set of desirable criteria that any good loss function should\nsatisfy. Intuitively, these criteria require that the loss function faithfully\nevaluates a candidate distribution, both in expectation and when estimated on a\nfew samples. Interestingly, we observe that \\emph{no loss function} possesses\nall of these criteria. However, one can circumvent this issue by introducing a\nnatural restriction on the set of candidate distributions. Specifically, we\nrequire that candidates are $calibrated$ with respect to the target\ndistribution, i.e., they may contain less information than the target but\notherwise do not significantly distort the truth. We show that, after\nrestricting to this set of distributions, the log loss, along with a large\nvariety of other losses satisfy the desired criteria. These results pave the\nway for future investigations of distribution learning that look beyond the log\nloss, choosing a loss function based on application or domain need.', 'Developments in deep generative models have allowed for tractable learning of\nhigh-dimensional data distributions. While the employed learning procedures\ntypically assume that training data is drawn i.i.d. from the distribution of\ninterest, it may be desirable to model distinct distributions which are\nobserved sequentially, such as when different classes are encountered over\ntime. Although conditional variations of deep generative models permit multiple\ndistributions to be modeled by a single network in a disentangled fashion, they\nare susceptible to catastrophic forgetting when the distributions are\nencountered sequentially. In this paper, we adapt recent work in reducing\ncatastrophic forgetting to the task of training generative adversarial networks\non a sequence of distinct distributions, enabling continual generative\nmodeling.', 'In linear regression we wish to estimate the optimum linear least squares\npredictor for a distribution over $d$-dimensional input points and real-valued\nresponses, based on a small sample. Under standard random design analysis,\nwhere the sample is drawn i.i.d. from the input distribution, the least squares\nsolution for that sample can be viewed as the natural estimator of the optimum.\nUnfortunately, this estimator almost always incurs an undesirable bias coming\nfrom the randomness of the input points, which is a significant bottleneck in\nmodel averaging. In this paper we show that it is possible to draw a non-i.i.d.\nsample of input points such that, regardless of the response model, the least\nsquares solution is an unbiased estimator of the optimum. Moreover, this sample\ncan be produced efficiently by augmenting a previously drawn i.i.d. sample with\nan additional set of $d$ points, drawn jointly according to a certain\ndeterminantal point process constructed from the input distribution rescaled by\nthe squared volume spanned by the points. Motivated by this, we develop a\ntheoretical framework for studying volume-rescaled sampling, and in the process\nprove a number of new matrix expectation identities. We use them to show that\nfor any input distribution and $\\epsilon>0$ there is a random design consisting\nof $O(d\\log d+ d/\\epsilon)$ points from which an unbiased estimator can be\nconstructed whose expected square loss over the entire distribution is bounded\nby $1+\\epsilon$ times the loss of the optimum. We provide efficient algorithms\nfor generating such unbiased estimators in a number of practical settings and\nsupport our claims experimentally.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'The standard loss functions used in the literature on probabilistic\nprediction are the log loss function, the Brier loss function, and the\nspherical loss function; however, any computable proper loss function can be\nused for comparison of prediction algorithms. This note shows that the log loss\nfunction is most selective in that any prediction algorithm that is optimal for\na given data sequence (in the sense of the algorithmic theory of randomness)\nunder the log loss function will be optimal under any computable proper mixable\nloss function; on the other hand, there is a data sequence and a prediction\nalgorithm that is optimal for that sequence under either of the two other\nstandard loss functions but not under the log loss function.', 'In experimental design, we are given a large collection of vectors, each with\na hidden response value that we assume derives from an underlying linear model,\nand we wish to pick a small subset of the vectors such that querying the\ncorresponding responses will lead to a good estimator of the model. A classical\napproach in statistics is to assume the responses are linear, plus zero-mean\ni.i.d. Gaussian noise, in which case the goal is to provide an unbiased\nestimator with smallest mean squared error (A-optimal design). A related\napproach, more common in computer science, is to assume the responses are\narbitrary but fixed, in which case the goal is to estimate the least squares\nsolution using few responses, as quickly as possible, for worst-case inputs.\nDespite many attempts, characterizing the relationship between these two\napproaches has proven elusive. We address this by proposing a framework for\nexperimental design where the responses are produced by an arbitrary unknown\ndistribution. We show that there is an efficient randomized experimental design\nprocedure that achieves strong variance bounds for an unbiased estimator using\nfew responses in this general model. Nearly tight bounds for the classical\nA-optimality criterion, as well as improved bounds for worst-case responses,\nemerge as special cases of this result. In the process, we develop a new\nalgorithm for a joint sampling distribution called volume sampling, and we\npropose a new i.i.d. importance sampling method: inverse score sampling. A key\nnovelty of our analysis is in developing new expected error bounds for\nworst-case regression by controlling the tail behavior of i.i.d. sampling via\nthe jointness of volume sampling. Our result motivates a new minimax-optimality\ncriterion for experimental design which can be viewed as an extension of both\nA-optimal design and sampling for worst-case regression.', 'With the increase of information, document classification as one of the\nmethods of text mining, plays vital role in many management and organizing\ninformation. Document classification is the process of assigning a document to\none or more predefined category labels. Document classification includes\ndifferent parts such as text processing, term selection, term weighting and\nfinal classification. The accuracy of document classification is very\nimportant. Thus improvement in each part of classification should lead to\nbetter results and higher precision. Term weighting has a great impact on the\naccuracy of the classification. Most of the existing weighting methods exploit\nthe statistical information of terms in documents and do not consider semantic\nrelations between words. In this paper, an automated document classification\nsystem is presented that uses a novel term weighting method based on semantic\nrelations between terms. To evaluate the proposed method, three standard\nPersian corpuses are used. Experiment results show 2 to 4 percent improvement\nin classification accuracy compared with the best previous designed system for\nPersian documents.', 'Document clustering is a text mining technique used to provide better\ndocument search and browsing in digital libraries or online corpora. A lot of\nresearch has been done on biomedical document clustering that is based on using\nexisting ontology. But, associations and co-occurrences of the medical concepts\nare not well represented by using ontology. In this research, a vector\nrepresentation of concepts of diseases and similarity measurement between\nconcepts are proposed. They identify the closest concepts of diseases in the\ncontext of a corpus. Each document is represented by using the vector space\nmodel. A weight scheme is proposed to consider both local content and\nassociations between concepts. A Self-Organizing Map is used as document\nclustering algorithm. The vector projection and visualization features of SOM\nenable visualization and analysis of the clusters distributions and\nrelationships on the two dimensional space. The experimental results show that\nthe proposed document clustering framework generates meaningful clusters and\nfacilitate visualization of the clusters based on the concepts of diseases.', 'We consider the problem of engineering robust direct perception neural\nnetworks with output being regression. Such networks take high dimensional\ninput image data, and they produce affordances such as the curvature of the\nupcoming road segment or the distance to the front vehicle. Our proposal starts\nby allowing a neural network prediction to deviate from the label with\ntolerance $\\Delta$. The source of tolerance can be either contractual or from\nlimiting factors where two entities may label the same data with slightly\ndifferent numerical values. The tolerance motivates the use of a non-standard\nloss function where the loss is set to $0$ so long as the prediction-to-label\ndistance is less than $\\Delta$. We further extend the loss function and define\na new provably robust criterion that is parametric to the allowed output\ntolerance $\\Delta$, the layer index $\\tilde{l}$ where perturbation is\nconsidered, and the maximum perturbation amount $\\kappa$. During training, the\nrobust loss is computed by first propagating symbolic errors from the\n$\\tilde{l}$-th layer (with quantity bounded by $\\kappa$) to the output layer,\nfollowed by computing the overflow between the error bounds and the allowed\ntolerance. The overall concept is experimented in engineering a direct\nperception neural network for understanding the central position of the\nego-lane in pixel coordinates.', 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', 'Given a finite set of unknown distributions, or arms, that can be sampled, we\nconsider the problem of identifying the one with the largest mean using a\ndelta-correct algorithm (an adaptive, sequential algorithm that restricts the\nprobability of error to a specified delta) that has minimum sample complexity.\nLower bounds for delta-correct algorithms are well known. Delta-correct\nalgorithms that match the lower bound asymptotically as delta reduces to zero\nhave been previously developed when arm distributions are restricted to a\nsingle parameter exponential family. In this paper, we first observe a negative\nresult that some restrictions are essential, as otherwise under a delta-correct\nalgorithm, distributions with unbounded support would require an infinite\nnumber of samples in expectation. We then propose a delta-correct algorithm\nthat matches the lower bound as delta reduces to zero under the mild\nrestriction that a known bound on the expectation of a non-negative,\ncontinuous, increasing convex function (for example, the squared moment) of the\nunderlying random variables, exists. We also propose batch processing and\nidentify near-optimal batch sizes to substantially speed up the proposed\nalgorithm. The best-arm problem has many learning applications, including\nrecommendation systems and product selection. It is also a well studied classic\nproblem in the simulation community.', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', ""We introduce a method to stabilize Generative Adversarial Networks (GANs) by\ndefining the generator objective with respect to an unrolled optimization of\nthe discriminator. This allows training to be adjusted between using the\noptimal discriminator in the generator's objective, which is ideal but\ninfeasible in practice, and using the current value of the discriminator, which\nis often unstable and leads to poor solutions. We show how this technique\nsolves the common problem of mode collapse, stabilizes training of GANs with\ncomplex recurrent generators, and increases diversity and coverage of the data\ndistribution by the generator."", 'In few-shot learning, typically, the loss function which is applied at test\ntime is the one we are ultimately interested in minimising, such as the\nmean-squared-error loss for a regression problem. However, given that we have\nfew samples at test time, we argue that the loss function that we are\ninterested in minimising is not necessarily the loss function most suitable for\ncomputing gradients in a few-shot setting. We propose VIABLE, a generic\nmeta-learning extension that builds on existing meta-gradient-based methods by\nlearning a differentiable loss function, replacing the pre-defined inner-loop\nloss function in performing task-specific updates. We show that learning a loss\nfunction capable of leveraging relational information between samples reduces\nunderfitting, and significantly improves performance and sample efficiency on a\nsimple regression task. Furthermore, we show VIABLE is scalable by evaluating\non the Mini-Imagenet dataset.', 'In this paper, we explore text classification with extremely weak\nsupervision, i.e., only relying on the surface text of class names. This is a\nmore challenging setting than the seed-driven weak supervision, which allows a\nfew seed words per class. We opt to attack this problem from a representation\nlearning perspective -- ideal document representations should lead to nearly\nthe same results between clustering and the desired classification. In\nparticular, one can classify the same corpus differently (e.g., based on topics\nand locations), so document representations should be adaptive to the given\nclass names. We propose a novel framework X-Class to realize the adaptive\nrepresentations. Specifically, we first estimate class representations by\nincrementally adding the most similar word to each class until inconsistency\narises. Following a tailored mixture of class attention mechanisms, we obtain\nthe document representation via a weighted average of contextualized word\nrepresentations. With the prior of each document assigned to its nearest class,\nwe then cluster and align the documents to classes. Finally, we pick the most\nconfident documents from each cluster to train a text classifier. Extensive\nexperiments demonstrate that X-Class can rival and even outperform seed-driven\nweakly supervised methods on 7 benchmark datasets. Our dataset and code are\nreleased at https://github.com/ZihanWangKi/XClass/ .', 'Physics-informed neural networks (PINNs) are effective in solving inverse\nproblems based on differential and integral equations with sparse, noisy,\nunstructured, and multi-fidelity data. PINNs incorporate all available\ninformation into a loss function, thus recasting the original problem into an\noptimization problem. In this paper, we extend PINNs to parameter and function\ninference for integral equations such as nonlocal Poisson and nonlocal\nturbulence models, and we refer to them as nonlocal PINNs (nPINNs). The\ncontribution of the paper is three-fold. First, we propose a unified nonlocal\noperator, which converges to the classical Laplacian as one of the operator\nparameters, the nonlocal interaction radius $\\delta$ goes to zero, and to the\nfractional Laplacian as $\\delta$ goes to infinity. This universal operator\nforms a super-set of classical Laplacian and fractional Laplacian operators\nand, thus, has the potential to fit a broad spectrum of data sets. We provide\ntheoretical convergence rates with respect to $\\delta$ and verify them via\nnumerical experiments. Second, we use nPINNs to estimate the two parameters,\n$\\delta$ and $\\alpha$. The strong non-convexity of the loss function yielding\nmultiple (good) local minima reveals the occurrence of the operator mimicking\nphenomenon: different pairs of estimated parameters could produce multiple\nsolutions of comparable accuracy. Third, we propose another nonlocal operator\nwith spatially variable order $\\alpha(y)$, which is more suitable for modeling\nturbulent Couette flow. Our results show that nPINNs can jointly infer this\nfunction as well as $\\delta$. Also, these parameters exhibit a universal\nbehavior with respect to the Reynolds number, a finding that contributes to our\nunderstanding of nonlocal interactions in wall-bounded turbulence.', 'Machine learning has the potential to assist many communities in using the\nlarge datasets that are becoming more and more available. Unfortunately, much\nof that potential is not being realized because it would require sharing data\nin a way that compromises privacy. In order to overcome this hurdle, several\nmethods have been proposed that generate synthetic data while preserving the\nprivacy of the real data. In this paper we consider a key characteristic that\nsynthetic data should have in order to be useful for machine learning\nresearchers - the relative performance of two algorithms (trained and tested)\non the synthetic dataset should be the same as their relative performance (when\ntrained and tested) on the original dataset.', 'We propose a Label Propagation based algorithm for weakly supervised text\nclassification. We construct a graph where each document is represented by a\nnode and edge weights represent similarities among the documents. Additionally,\nwe discover underlying topics using Latent Dirichlet Allocation (LDA) and\nenrich the document graph by including the topics in the form of additional\nnodes. The edge weights between a topic and a text document represent level of\n""affinity"" between them. Our approach does not require document level\nlabelling, instead it expects manual labels only for topic nodes. This\nsignificantly minimizes the level of supervision needed as only a few topics\nare observed to be enough for achieving sufficiently high accuracy. The Label\nPropagation Algorithm is employed on this enriched graph to propagate labels\namong the nodes. Our approach combines the advantages of Label Propagation\n(through document-document similarities) and Topic Modelling (for minimal but\nsmart supervision). We demonstrate the effectiveness of our approach on various\ndatasets and compare with state-of-the-art weakly supervised text\nclassification approaches.', 'We introduce a novel type of text representation that preserves the 2D layout\nof a document. This is achieved by encoding each document page as a\ntwo-dimensional grid of characters. Based on this representation, we present a\ngeneric document understanding pipeline for structured documents. This pipeline\nmakes use of a fully convolutional encoder-decoder network that predicts a\nsegmentation mask and bounding boxes. We demonstrate its capabilities on an\ninformation extraction task from invoices and show that it significantly\noutperforms approaches based on sequential text or document images.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'In recent years online shopping has gained momentum and became an important\nvenue for customers wishing to save time and simplify their shopping process. A\nkey advantage of shopping online is the ability to read what other customers\nare saying about products of interest. In this work, we aim to maintain this\nadvantage in situations where extreme brevity is needed, for example, when\nshopping by voice. We suggest a novel task of extracting a single\nrepresentative helpful sentence from a set of reviews for a given product. The\nselected sentence should meet two conditions: first, it should be helpful for a\npurchase decision and second, the opinion it expresses should be supported by\nmultiple reviewers. This task is closely related to the task of Multi Document\nSummarization in the product reviews domain but differs in its objective and\nits level of conciseness. We collect a dataset in English of sentence\nhelpfulness scores via crowd-sourcing and demonstrate its reliability despite\nthe inherent subjectivity involved. Next, we describe a complete model that\nextracts representative helpful sentences with positive and negative sentiment\ntowards the product and demonstrate that it outperforms several baselines.', 'In this article I describe a research agenda for securing machine learning\nmodels against adversarial inputs at test time. This article does not present\nresults but instead shares some of my thoughts about where I think that the\nfield needs to go. Modern machine learning works very well on I.I.D. data: data\nfor which each example is drawn {\\em independently} and for which the\ndistribution generating each example is {\\em identical}. When these assumptions\nare relaxed, modern machine learning can perform very poorly. When machine\nlearning is used in contexts where security is a concern, it is desirable to\ndesign models that perform well even when the input is designed by a malicious\nadversary. So far most research in this direction has focused on an adversary\nwho violates the {\\em identical} assumption, and imposes some kind of\nrestricted worst-case distribution shift. I argue that machine learning\nsecurity researchers should also address the problem of relaxing the {\\em\nindependence} assumption and that current strategies designed for robustness to\ndistribution shift will not do so. I recommend {\\em dynamic models} that change\neach time they are run as a potential solution path to this problem, and show\nan example of a simple attack using correlated data that can be mitigated by a\nsimple dynamic defense. This is not intended as a real-world security measure,\nbut as a recommendation to explore this research direction and develop more\nrealistic defenses.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.', 'In this paper, we propose a novel {\\em $e$-exponentiated} transformation, $0\n\\le e<1$, for loss functions. When the transformation is applied to a convex\nloss function, the transformed loss function become more robust to outliers.\nUsing a novel generalization error bound, we have theoretically shown that the\ntransformed loss function has a tighter bound for datasets corrupted by\noutliers. Our empirical observation shows that the accuracy obtained using the\ntransformed loss function can be significantly better than the same obtained\nusing the original loss function and comparable to that obtained by some other\nstate of the art methods in the presence of label noise.', 'Extracting information from full documents is an important problem in many\ndomains, but most previous work focus on identifying relationships within a\nsentence or a paragraph. It is challenging to create a large-scale information\nextraction (IE) dataset at the document level since it requires an\nunderstanding of the whole document to annotate entities and their\ndocument-level relationships that usually span beyond sentences or even\nsections. In this paper, we introduce SciREX, a document level IE dataset that\nencompasses multiple IE tasks, including salient entity identification and\ndocument level $N$-ary relation identification from scientific articles. We\nannotate our dataset by integrating automatic and human annotations, leveraging\nexisting scientific knowledge resources. We develop a neural model as a strong\nbaseline that extends previous state-of-the-art IE models to document-level IE.\nAnalyzing the model performance shows a significant gap between human\nperformance and current baselines, inviting the community to use our dataset as\na challenge to develop document-level IE models. Our data and code are publicly\navailable at https://github.com/allenai/SciREX', 'We consider the problem of learning an unknown function $f_{\\star}$ on the\n$d$-dimensional sphere with respect to the square loss, given i.i.d. samples\n$\\{(y_i,{\\boldsymbol x}_i)\\}_{i\\le n}$ where ${\\boldsymbol x}_i$ is a feature\nvector uniformly distributed on the sphere and $y_i=f_{\\star}({\\boldsymbol\nx}_i)+\\varepsilon_i$. We study two popular classes of models that can be\nregarded as linearizations of two-layers neural networks around a random\ninitialization: the random features model of Rahimi-Recht (RF); the neural\ntangent kernel model of Jacot-Gabriel-Hongler (NT). Both these approaches can\nalso be regarded as randomized approximations of kernel ridge regression (with\nrespect to different kernels), and enjoy universal approximation properties\nwhen the number of neurons $N$ diverges, for a fixed dimension $d$.\n  We consider two specific regimes: the approximation-limited regime, in which\n$n=\\infty$ while $d$ and $N$ are large but finite; and the sample size-limited\nregime in which $N=\\infty$ while $d$ and $n$ are large but finite. In the first\nregime we prove that if $d^{\\ell + \\delta} \\le N\\le d^{\\ell+1-\\delta}$ for\nsmall $\\delta > 0$, then \\RF\\, effectively fits a degree-$\\ell$ polynomial in\nthe raw features, and \\NT\\, fits a degree-$(\\ell+1)$ polynomial. In the second\nregime, both RF and NT reduce to kernel methods with rotationally invariant\nkernels. We prove that, if the number of samples is $d^{\\ell + \\delta} \\le n\n\\le d^{\\ell +1-\\delta}$, then kernel methods can fit at most a a degree-$\\ell$\npolynomial in the raw features. This lower bound is achieved by kernel ridge\nregression. Optimal prediction error is achieved for vanishing ridge\nregularization.', 'In the negative perceptron problem we are given $n$ data points\n$({\\boldsymbol x}_i,y_i)$, where ${\\boldsymbol x}_i$ is a $d$-dimensional\nvector and $y_i\\in\\{+1,-1\\}$ is a binary label. The data are not linearly\nseparable and hence we content ourselves to find a linear classifier with the\nlargest possible \\emph{negative} margin. In other words, we want to find a unit\nnorm vector ${\\boldsymbol \\theta}$ that maximizes $\\min_{i\\le n}y_i\\langle\n{\\boldsymbol \\theta},{\\boldsymbol x}_i\\rangle$. This is a non-convex\noptimization problem (it is equivalent to finding a maximum norm vector in a\npolytope), and we study its typical properties under two random models for the\ndata.\n  We consider the proportional asymptotics in which $n,d\\to \\infty$ with\n$n/d\\to\\delta$, and prove upper and lower bounds on the maximum margin\n$\\kappa_{\\text{s}}(\\delta)$ or -- equivalently -- on its inverse function\n$\\delta_{\\text{s}}(\\kappa)$. In other words, $\\delta_{\\text{s}}(\\kappa)$ is the\noverparametrization threshold: for $n/d\\le\n\\delta_{\\text{s}}(\\kappa)-\\varepsilon$ a classifier achieving vanishing\ntraining error exists with high probability, while for $n/d\\ge\n\\delta_{\\text{s}}(\\kappa)+\\varepsilon$ it does not. Our bounds on\n$\\delta_{\\text{s}}(\\kappa)$ match to the leading order as $\\kappa\\to -\\infty$.\nWe then analyze a linear programming algorithm to find a solution, and\ncharacterize the corresponding threshold $\\delta_{\\text{lin}}(\\kappa)$. We\nobserve a gap between the interpolation threshold $\\delta_{\\text{s}}(\\kappa)$\nand the linear programming threshold $\\delta_{\\text{lin}}(\\kappa)$, raising the\nquestion of the behavior of other algorithms.', 'We study distributionally robust optimization (DRO) with Sinkhorn distance --\na variant of Wasserstein distance based on entropic regularization. We provide\nconvex programming dual reformulation for a general nominal distribution.\nCompared with Wasserstein DRO, it is computationally tractable for a larger\nclass of loss functions, and its worst-case distribution is more reasonable. We\npropose an efficient first-order algorithm with bisection search to solve the\ndual reformulation. We demonstrate that our proposed algorithm finds\n$\\delta$-optimal solution of the new DRO formulation with computation cost\n$\\tilde{O}(\\delta^{-3})$ and memory cost $\\tilde{O}(\\delta^{-2})$, and the\ncomputation cost further improves to $\\tilde{O}(\\delta^{-2})$ when the loss\nfunction is smooth. Finally, we provide various numerical examples using both\nsynthetic and real data to demonstrate its competitive performance and light\ncomputational speed.', ""Motivated by applications in crowdsourced entity resolution in database,\nsigned edge prediction in social networks and correlation clustering, Mazumdar\nand Saha [NIPS 2017] proposed an elegant theoretical model for studying\nclustering with a faulty oracle. In this model, given a set of $n$ items which\nbelong to $k$ unknown groups (or clusters), our goal is to recover the clusters\nby asking pairwise queries to an oracle. This oracle can answer the query that\n``do items $u$ and $v$ belong to the same cluster?''. However, the answer to\neach pairwise query errs with probability $\\varepsilon$, for some\n$\\varepsilon\\in(0,\\frac12)$. Mazumdar and Saha provided two algorithms under\nthis model: one algorithm is query-optimal while time-inefficient (i.e.,\nrunning in quasi-polynomial time), the other is time efficient (i.e., in\npolynomial time) while query-suboptimal. Larsen, Mitzenmacher and Tsourakakis\n[WWW 2020] then gave a new time-efficient algorithm for the special case of $2$\nclusters, which is query-optimal if the bias $\\delta:=1-2\\varepsilon$ of the\nmodel is large. It was left as an open question whether one can obtain a\nquery-optimal, time-efficient algorithm for the general case of $k$ clusters\nand other regimes of $\\delta$.\n  In this paper, we make progress on the above question and provide a\ntime-efficient algorithm with nearly-optimal query complexity (up to a factor\nof $O(\\log^2 n)$) for all constant $k$ and any $\\delta$ in the regime when\ninformation-theoretic recovery is possible. Our algorithm is built on a\nconnection to the stochastic block model."", 'We introduce a novel approach for scanned document representation to perform\nfield extraction. It allows the simultaneous encoding of the textual, visual\nand layout information in a 3-axis tensor used as an input to a segmentation\nmodel. We improve the recent Chargrid and Wordgrid \\cite{chargrid} models in\nseveral ways, first by taking into account the visual modality, then by\nboosting its robustness in regards to small datasets while keeping the\ninference time low. Our approach is tested on public and private document-image\ndatasets, showing higher performances compared to the recent state-of-the-art\nmethods.', ""In unconstrained optimisation on an Euclidean space, to prove convergence in\nGradient Descent processes (GD) $x_{n+1}=x_n-\\delta _n \\nabla f(x_n)$ it\nusually is required that the learning rates $\\delta _n$'s are bounded: $\\delta\n_n\\leq \\delta $ for some positive $\\delta $. Under this assumption, if the\nsequence $x_n$ converges to a critical point $z$, then with large values of $n$\nthe update will be small because $||x_{n+1}-x_n||\\lesssim ||\\nabla f(x_n)||$.\nThis may also force the sequence to converge to a bad minimum. If we can allow,\nat least theoretically, that the learning rates $\\delta _n$'s are not bounded,\nthen we may have better convergence to better minima.\n  A previous joint paper by the author showed convergence for the usual version\nof Backtracking GD under very general assumptions on the cost function $f$. In\nthis paper, we allow the learning rates $\\delta _n$ to be unbounded, in the\nsense that there is a function $h:(0,\\infty)\\rightarrow (0,\\infty )$ such that\n$\\lim _{t\\rightarrow 0}th(t)=0$ and $\\delta _n\\lesssim \\max \\{h(x_n),\\delta \\}$\nsatisfies Armijo's condition for all $n$, and prove convergence under the same\nassumptions as in the mentioned paper. It will be shown that this growth rate\nof $h$ is best possible if one wants convergence of the sequence $\\{x_n\\}$.\n  A specific way for choosing $\\delta _n$ in a discrete way connects to Two-way\nBacktracking GD defined in the mentioned paper. We provide some results which\neither improve or are implicitly contained in those in the mentioned paper and\nanother recent paper on avoidance of saddle points."", ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'This paper summarizes our entries to both subtasks of the first DialDoc\nshared task which focuses on the agent response prediction task in\ngoal-oriented document-grounded dialogs. The task is split into two subtasks:\npredicting a span in a document that grounds an agent turn and generating an\nagent response based on a dialog and grounding document. In the first subtask,\nwe restrict the set of valid spans to the ones defined in the dataset, use a\nbiaffine classifier to model spans, and finally use an ensemble of different\nmodels. For the second subtask, we use a cascaded model which grounds the\nresponse prediction on the predicted span instead of the full document. With\nthese approaches, we obtain significant improvements in both subtasks compared\nto the baseline.', 'We study the problem of testing identity against a given distribution with a\nfocus on the high confidence regime. More precisely, given samples from an\nunknown distribution $p$ over $n$ elements, an explicitly given distribution\n$q$, and parameters $0< \\epsilon, \\delta < 1$, we wish to distinguish, {\\em\nwith probability at least $1-\\delta$}, whether the distributions are identical\nversus $\\varepsilon$-far in total variation distance. Most prior work focused\non the case that $\\delta = \\Omega(1)$, for which the sample complexity of\nidentity testing is known to be $\\Theta(\\sqrt{n}/\\epsilon^2)$. Given such an\nalgorithm, one can achieve arbitrarily small values of $\\delta$ via black-box\namplification, which multiplies the required number of samples by\n$\\Theta(\\log(1/\\delta))$.\n  We show that black-box amplification is suboptimal for any $\\delta = o(1)$,\nand give a new identity tester that achieves the optimal sample complexity. Our\nnew upper and lower bounds show that the optimal sample complexity of identity\ntesting is \\[\n  \\Theta\\left( \\frac{1}{\\epsilon^2}\\left(\\sqrt{n \\log(1/\\delta)} +\n\\log(1/\\delta) \\right)\\right) \\] for any $n, \\varepsilon$, and $\\delta$. For\nthe special case of uniformity testing, where the given distribution is the\nuniform distribution $U_n$ over the domain, our new tester is surprisingly\nsimple: to test whether $p = U_n$ versus $d_{\\mathrm TV}(p, U_n) \\geq\n\\varepsilon$, we simply threshold $d_{\\mathrm TV}(\\widehat{p}, U_n)$, where\n$\\widehat{p}$ is the empirical probability distribution. The fact that this\nsimple ""plug-in"" estimator is sample-optimal is surprising, even in the\nconstant $\\delta$ case. Indeed, it was believed that such a tester would not\nattain sublinear sample complexity even for constant values of $\\varepsilon$\nand $\\delta$.', 'Over the past decades, researchers have been pushing the limits of Deep\nReinforcement Learning (DRL). Although DRL has attracted substantial interest\nfrom practitioners, many are blocked by having to search through a plethora of\navailable methodologies that are seemingly alike, while others are still\nbuilding RL agents from scratch based on classical theories. To address the\naforementioned gaps in adopting the latest DRL methods, I am particularly\ninterested in testing out if any of the recent technology developed by the\nleads in the field can be readily applied to a class of optimal trading\nproblems. Unsurprisingly, many prominent breakthroughs in DRL are investigated\nand tested on strategic games: from AlphaGo to AlphaStar and at about the same\ntime, OpenAI Five. Thus, in this writing, I want to show precisely how to use a\nDRL library that is initially built for games in a fundamental trading problem;\nmean reversion. And by introducing a framework that incorporates\neconomically-motivated function properties, I also demonstrate, through the\nlibrary, a highly-performant and convergent DRL solution to decision-making\nfinancial problems in general.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', 'We show that Bayes\' rule provides an effective mechanism for creating\ndocument translation models that can be learned from only parallel sentences\nand monolingual documents---a compelling benefit as parallel documents are not\nalways available. In our formulation, the posterior probability of a candidate\ntranslation is the product of the unconditional (prior) probability of the\ncandidate output document and the ""reverse translation probability"" of\ntranslating the candidate output back into the source language. Our proposed\nmodel uses a powerful autoregressive language model as the prior on target\nlanguage documents, but it assumes that each sentence is translated\nindependently from the target to the source language. Crucially, at test time,\nwhen a source document is observed, the document language model prior induces\ndependencies between the translations of the source sentences in the posterior.\nThe model\'s independence assumption not only enables efficient use of available\ndata, but it additionally admits a practical left-to-right beam-search\nalgorithm for carrying out inference. Experiments show that our model benefits\nfrom using cross-sentence context in the language model, and it outperforms\nexisting document translation approaches.', ""Neural networks with deep architectures have demonstrated significant\nperformance improvements in computer vision, speech recognition, and natural\nlanguage processing. The challenges in information retrieval (IR), however, are\ndifferent from these other application areas. A common form of IR involves\nranking of documents--or short passages--in response to keyword-based queries.\nEffective IR systems must deal with query-document vocabulary mismatch problem,\nby modeling relationships between different query and document terms and how\nthey indicate relevance. Models should also consider lexical matches when the\nquery contains rare terms--such as a person's name or a product model\nnumber--not seen during training, and to avoid retrieving semantically related\nbut irrelevant results. In many real-life IR tasks, the retrieval involves\nextremely large collections--such as the document index of a commercial Web\nsearch engine--containing billions of documents. Efficient IR methods should\ntake advantage of specialized IR data structures, such as inverted index, to\nefficiently retrieve from large collections. Given an information need, the IR\nsystem also mediates how much exposure an information artifact receives by\ndeciding whether it should be displayed, and where it should be positioned,\namong other results. Exposure-aware IR systems may optimize for additional\nobjectives, besides relevance, such as parity of exposure for retrieved items\nand content publishers. In this thesis, we present novel neural architectures\nand methods motivated by the specific needs and challenges of IR tasks."", 'Document categorization is a technique where the category of a document is\ndetermined. In this paper three well-known supervised learning techniques which\nare Support Vector Machine(SVM), Na\\""ive Bayes(NB) and Stochastic Gradient\nDescent(SGD) compared for Bengali document categorization. Besides classifier,\nclassification also depends on how feature is selected from dataset. For\nanalyzing those classifier performances on predicting a document against twelve\ncategories several feature selection techniques are also applied in this\narticle namely Chi square distribution, normalized TFIDF (term\nfrequency-inverse document frequency) with word analyzer. So, we attempt to\nexplore the efficiency of those three-classification algorithms by using two\ndifferent feature selection techniques in this article.', 'We derive the optimal $(0, \\delta)$-differentially private query-output\nindependent noise-adding mechanism for single real-valued query function under\na general cost-minimization framework. Under a mild technical condition, we\nshow that the optimal noise probability distribution is a uniform distribution\nwith a probability mass at the origin. We explicitly derive the optimal noise\ndistribution for general $\\ell^p$ cost functions, including $\\ell^1$ (for noise\nmagnitude) and $\\ell^2$ (for noise power) cost functions, and show that the\nprobability concentration on the origin occurs when $\\delta > \\frac{p}{p+1}$.\nOur result demonstrates an improvement over the existing Gaussian mechanisms by\na factor of two and three for $(0,\\delta)$-differential privacy in the high\nprivacy regime in the context of minimizing the noise magnitude and noise\npower, and the gain is more pronounced in the low privacy regime. Our result is\nconsistent with the existing result for $(0,\\delta)$-differential privacy in\nthe discrete setting, and identifies a probability concentration phenomenon in\nthe continuous setting.', 'We consider the problem of maximizing non-negative non-decreasing set\nfunctions. Although most of the recent work focus on exploiting submodularity,\nit turns out that several objectives we encounter in practice are not\nsubmodular. Nonetheless, often we leverage the greedy algorithms used in\nsubmodular functions to determine a solution to the non-submodular functions.\nHereafter, we propose to address the original problem by \\emph{approximating}\nthe non-submodular function and analyze the incurred error, as well as the\nperformance trade-offs. To quantify the approximation error, we introduce a\nnovel concept of $\\delta$-approximation of a function, which we used to define\nthe space of submodular functions that lie within an approximation error. We\nprovide necessary conditions on the existence of such $\\delta$-approximation\nfunctions, which might not be unique. Consequently, we characterize this\nsubspace which we refer to as \\emph{region of submodularity}. Furthermore,\nsubmodular functions are known to lead to different sub-optimality guarantees,\nso we generalize those dependencies upon a $\\delta$-approximation into the\nnotion of \\emph{greedy curvature}. Finally, we used this latter notion to\nsimplify some of the existing results and efficiently (i.e., linear complexity)\ndetermine tightened bounds on the sub-optimality guarantees using objective\nfunctions commonly used in practical setups and validate them using real data.', 'We study $(\\epsilon, \\delta)$-PAC best arm identification, where a\ndecision-maker must identify an $\\epsilon$-optimal arm with probability at\nleast $1 - \\delta$, while minimizing the number of arm pulls (samples). Most of\nthe work on this topic is in the sequential setting, where there is no\nconstraint on the time taken to identify such an arm; this allows the\ndecision-maker to pull one arm at a time. In this work, the decision-maker is\ngiven a deadline of $T$ rounds, where, on each round, it can adaptively choose\nwhich arms to pull and how many times to pull them; this distinguishes the\nnumber of decisions made (i.e., time or number of rounds) from the number of\nsamples acquired (cost). Such situations occur in clinical trials, where one\nmay need to identify a promising treatment under a deadline while minimizing\nthe number of test subjects, or in simulation-based studies run on the cloud,\nwhere we can elastically scale up or down the number of virtual machines to\nconduct as many experiments as we wish, but need to pay for the resource-time\nused. As the decision-maker can only make $T$ decisions, she may need to pull\nsome arms excessively relative to a sequential algorithm in order to perform\nwell on all possible problems. We formalize this added difficulty with two\nhardness results that indicate that unlike sequential settings, the ability to\nadapt to the problem difficulty is constrained by the finite deadline. We\npropose Elastic Batch Racing (EBR), a novel algorithm for this setting and\nbound its sample complexity, showing that EBR is optimal with respect to both\nhardness results. We present simulations evaluating EBR in this setting, where\nit outperforms baselines by several orders of magnitude.', 'We consider the stochastic bandit problem in the sublinear space setting,\nwhere one cannot record the win-loss record for all $K$ arms. We give an\nalgorithm using $O(1)$ words of space with regret \\[\n  \\sum_{i=1}^{K}\\frac{1}{\\Delta_i}\\log \\frac{\\Delta_i}{\\Delta}\\log T \\] where\n$\\Delta_i$ is the gap between the best arm and arm $i$ and $\\Delta$ is the gap\nbetween the best and the second-best arms. If the rewards are bounded away from\n$0$ and $1$, this is within an $O(\\log 1/\\Delta)$ factor of the optimum regret\npossible without space constraints.', 'This paper describes the short-term competition on the Components\nSegmentation Task of Document Photos that was prepared in the context of the\n16th International Conference on Document Analysis and Recognition (ICDAR\n2021). This competition aims to bring together researchers working in the field\nof identification document image processing and provides them a suitable\nbenchmark to compare their techniques on the component segmentation task of\ndocument images. Three challenge tasks were proposed entailing different\nsegmentation assignments to be performed on a provided dataset. The collected\ndata are from several types of Brazilian ID documents, whose personal\ninformation was conveniently replaced. There were 16 participants whose results\nobtained for some or all the three tasks show different rates for the adopted\nmetrics, like Dice Similarity Coefficient ranging from 0.06 to 0.99. Different\nDeep Learning models were applied by the entrants with diverse strategies to\nachieve the best results in each of the tasks. Obtained results show that the\ncurrently applied methods for solving one of the proposed tasks (document\nboundary detection) are already well established. However, for the other two\nchallenge tasks (text zone and handwritten sign detection) research and\ndevelopment of more robust approaches are still required to achieve acceptable\nresults.', 'We study the problem of identifying the best arm in a stochastic multi-armed\nbandit game. Given a set of $n$ arms indexed from $1$ to $n$, each arm $i$ is\nassociated with an unknown reward distribution supported on $[0,1]$ with mean\n$\\theta_i$ and variance $\\sigma_i^2$. Assume $\\theta_1 > \\theta_2 \\geq \\cdots\n\\geq\\theta_n$. We propose an adaptive algorithm which explores the gaps and\nvariances of the rewards of the arms and makes future decisions based on the\ngathered information using a novel approach called \\textit{grouped median\nelimination}. The proposed algorithm guarantees to output the best arm with\nprobability $(1-\\delta)$ and uses at most $O \\left(\\sum_{i = 1}^n\n\\left(\\frac{\\sigma_i^2}{\\Delta_i^2} + \\frac{1}{\\Delta_i}\\right)(\\ln \\delta^{-1}\n+ \\ln \\ln \\Delta_i^{-1})\\right)$ samples, where $\\Delta_i$ ($i \\geq 2$) denotes\nthe reward gap between arm $i$ and the best arm and we define $\\Delta_1 =\n\\Delta_2$. This achieves a significant advantage over the variance-independent\nalgorithms in some favorable scenarios and is the first result that removes the\nextra $\\ln n$ factor on the best arm compared with the state-of-the-art. We\nfurther show that $\\Omega \\left( \\sum_{i = 1}^n \\left(\n\\frac{\\sigma_i^2}{\\Delta_i^2} + \\frac{1}{\\Delta_i} \\right) \\ln \\delta^{-1}\n\\right)$ samples are necessary for an algorithm to achieve the same goal,\nthereby illustrating that our algorithm is optimal up to doubly logarithmic\nterms.', 'In this paper, we consider non-convex optimization problems under\n\\textit{unknown} yet safety-critical constraints. Such problems naturally arise\nin a variety of domains including robotics, manufacturing, and medical\nprocedures, where it is infeasible to know or identify all the constraints.\nTherefore, the parameter space should be explored in a conservative way to\nensure that none of the constraints are violated during the optimization\nprocess once we start from a safe initialization point. To this end, we develop\nan algorithm called Reliable Frank-Wolfe (Reliable-FW). Given a general\nnon-convex function and an unknown polytope constraint, Reliable-FW\nsimultaneously learns the landscape of the objective function and the boundary\nof the safety polytope. More precisely, by assuming that Reliable-FW has access\nto a (stochastic) gradient oracle of the objective function and a noisy\nfeasibility oracle of the safety polytope, it finds an $\\epsilon$-approximate\nfirst-order stationary point with the optimal ${\\mathcal{O}}({1}/{\\epsilon^2})$\ngradient oracle complexity (resp. $\\tilde{\\mathcal{O}}({1}/{\\epsilon^3})$ (also\noptimal) in the stochastic gradient setting), while ensuring the safety of all\nthe iterates. Rather surprisingly, Reliable-FW only makes\n$\\tilde{\\mathcal{O}}(({d^2}/{\\epsilon^2})\\log 1/\\delta)$ queries to the noisy\nfeasibility oracle (resp. $\\tilde{\\mathcal{O}}(({d^2}/{\\epsilon^4})\\log\n1/\\delta)$ in the stochastic gradient setting) where $d$ is the dimension and\n$\\delta$ is the reliability parameter, tightening the existing bounds even for\nsafe minimization of convex functions. We further specialize our results to the\ncase that the objective function is convex. A crucial component of our analysis\nis to introduce and apply a technique called geometric shrinkage in the context\nof safe optimization.', 'We consider the Max $K$-Armed Bandit problem, where a learning agent is faced\nwith several stochastic arms, each a source of i.i.d. rewards of unknown\ndistribution. At each time step the agent chooses an arm, and observes the\nreward of the obtained sample. Each sample is considered here as a separate\nitem with the reward designating its value, and the goal is to find an item\nwith the highest possible value. Our basic assumption is a known lower bound on\nthe {\\em tail function} of the reward distributions. Under the PAC framework,\nwe provide a lower bound on the sample complexity of any\n$(\\epsilon,\\delta)$-correct algorithm, and propose an algorithm that attains\nthis bound up to logarithmic factors. We analyze the robustness of the proposed\nalgorithm and in addition, we compare the performance of this algorithm to the\nvariant in which the arms are not distinguishable by the agent and are chosen\nrandomly at each stage. Interestingly, when the maximal rewards of the arms\nhappen to be similar, the latter approach may provide better performance.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am working on a project analyzing customer feedback for a new product launch, so I need a method that can effectively identify the topics and segments in each feedback. Since the feedback we received is very complicated and contains a lot of noise words that digress from main topics, I need a method that is able to remove noise and simplify these feedbacks. Using this method, I hope our system could provide insights into the most common issues customers are facing and help us improve the product accordingly. Some ideas I have right now to simplify feedback and remove noise are to leverage word order information and heuristics. The final resulting system must beat state-of-the-art systems in terms of topic segmentation.","['With the evolution of the cloud and customer centric culture, we inherently\naccumulate huge repositories of textual reviews, feedback, and support\ndata.This has driven enterprises to seek and research engagement patterns, user\nnetwork analysis, topic detections, etc.However, huge manual work is still\nnecessary to mine data to be able to mine actionable outcomes. In this paper,\nwe proposed and developed an innovative Semi-Supervised Learning approach by\nutilizing Deep Learning and Topic Modeling to have a better understanding of\nthe user voice.This approach combines a BERT-based multiclassification\nalgorithm through supervised learning combined with a novel Probabilistic and\nSemantic Hybrid Topic Inference (PSHTI) Model through unsupervised learning,\naiming at automating the process of better identifying the main topics or areas\nas well as the sub-topics from the textual feedback and support.There are three\nmajor break-through: 1. As the advancement of deep learning technology, there\nhave been tremendous innovations in the NLP field, yet the traditional topic\nmodeling as one of the NLP applications lag behind the tide of deep learning.\nIn the methodology and technical perspective, we adopt transfer learning to\nfine-tune a BERT-based multiclassification system to categorize the main topics\nand then utilize the novel PSHTI model to infer the sub-topics under the\npredicted main topics. 2. The traditional unsupervised learning-based topic\nmodels or clustering methods suffer from the difficulty of automatically\ngenerating a meaningful topic label, but our system enables mapping the top\nwords to the self-help issues by utilizing domain knowledge about the product\nthrough web-crawling. 3. This work provides a prominent showcase by leveraging\nthe state-of-the-art methodology in the real production to help shed light to\ndiscover user insights and drive business investment priorities.', 'This paper deals with the automatic analysis of conversations between a\ncustomer and an agent in a call centre of a customer care service. The purpose\nof the analysis is to hypothesize themes about problems and complaints\ndiscussed in the conversation. Themes are defined by the application\ndocumentation topics. A conversation may contain mentions that are irrelevant\nfor the application purpose and multiple themes whose mentions may be\ninterleaved portions of a conversation that cannot be well defined. Two methods\nare proposed for multiple theme hypothesization. One of them is based on a\ncosine similarity measure using a bag of features extracted from the entire\nconversation. The other method introduces the concept of thematic density\ndistributed around specific word positions in a conversation. In addition to\nautomatically selected words, word bi-grams with possible gaps between\nsuccessive words are also considered and selected. Experimental results show\nthat the results obtained with the proposed methods outperform the results\nobtained with support vector machines on the same data. Furthermore, using the\ntheme skeleton of a conversation from which thematic densities are derived, it\nwill be possible to extract components of an automatic conversation report to\nbe used for improving the service performance. Index Terms: multi-topic audio\ndocument classification, hu-man/human conversation analysis, speech analytics,\ndistance bigrams', 'Opinions in forums and social networks are released by millions of people due\nto the increasing number of users that use Web 2.0 platforms to opine about\nbrands and organizations. For enterprises or government agencies it is almost\nimpossible to track what people say producing a gap between user\nneeds/expectations and organizations actions. To bridge this gap we create\nViscovery, a platform for opinion summarization and trend tracking that is able\nto analyze a stream of opinions recovered from forums. To do this we use\ndynamic topic models, allowing to uncover the hidden structure of topics behind\nopinions, characterizing vocabulary dynamics. We extend dynamic topic models\nfor incremental learning, a key aspect needed in Viscovery for model updating\nin near-real time. In addition, we include in Viscovery sentiment analysis,\nallowing to separate positive/negative words for a specific topic at different\nlevels of granularity. Viscovery allows to visualize representative opinions\nand terms in each topic. At a coarse level of granularity, the dynamic of the\ntopics can be analyzed using a 2D topic embedding, suggesting longitudinal\ntopic merging or segmentation. In this paper we report our experience\ndeveloping this platform, sharing lessons learned and opportunities that arise\nfrom the use of sentiment analysis and topic modeling in real world\napplications.', 'A massive amount of reviews are generated daily from various platforms. It is\nimpossible for people to read through tons of reviews and to obtain useful\ninformation. Automatic summarizing customer reviews thus is important for\nidentifying and extracting the essential information to help users to obtain\nthe gist of the data. However, as customer reviews are typically short,\ninformal, and multifaceted, it is extremely challenging to generate topic-wise\nsummarization.While there are several studies aims to solve this issue, they\nare heuristic methods that are developed only utilizing customer reviews.\nUnlike existing method, we propose an effective new summarization method by\nanalyzing both reviews and summaries.To do that, we first segment reviews and\nsummaries into individual sentiments. As the sentiments are typically short, we\ncombine sentiments talking about the same aspect into a single document and\napply topic modeling method to identify hidden topics among customer reviews\nand summaries. Sentiment analysis is employed to distinguish positive and\nnegative opinions among each detected topic. A classifier is also introduced to\ndistinguish the writing pattern of summaries and that of customer reviews.\nFinally, sentiments are selected to generate the summarization based on their\ntopic relevance, sentiment analysis score and the writing pattern. To test our\nmethod, a new dataset comprising product reviews and summaries about 1028\nproducts are collected from Amazon and CNET. Experimental results show the\neffectiveness of our method compared with other methods.', 'Understanding and being able to react to customer feedback is the most\nfundamental task in providing good customer service. However, there are two\nmajor obstacles for international companies to automatically detect the meaning\nof customer feedback in a global multilingual environment. Firstly, there is no\nwidely acknowledged categorisation (classes) of meaning for customer feedback.\nSecondly, the applicability of one meaning categorisation, if it exists, to\ncustomer feedback in multiple languages is questionable. In this paper, we\nextracted representative real world samples of customer feedback from Microsoft\nOffice customers in multiple languages, English, Spanish and Japanese,and\nconcluded a five-class categorisation(comment, request, bug, complaint and\nmeaningless) for meaning classification that could be used across languages in\nthe realm of customer feedback analysis.', 'With the development of e-commerce, many products are now being sold\nworldwide, and manufacturers are eager to obtain a better understanding of\ncustomer behavior in various regions. To achieve this goal, most previous\nefforts have focused mainly on questionnaires, which are time-consuming and\ncostly. The tremendous volume of product reviews on e-commerce websites has\nseen a new trend emerge, whereby manufacturers attempt to understand user\npreferences by analyzing online reviews. Following this trend, this paper\naddresses the problem of studying customer behavior by exploiting recently\ndeveloped opinion mining techniques. This work is novel for three reasons.\nFirst, questionnaire-based investigation is automatically enabled by employing\nalgorithms for template-based question generation and opinion mining-based\nanswer extraction. Using this system, manufacturers are able to obtain reports\nof customer behavior featuring a much larger sample size, more direct\ninformation, a higher degree of automation, and a lower cost. Second,\ninternational customer behavior study is made easier by integrating tools for\nmultilingual opinion mining. Third, this is the first time an automatic\nquestionnaire investigation has been conducted to compare customer behavior in\nChina and America, where product reviews are written and read in Chinese and\nEnglish, respectively. Our study on digital cameras, smartphones, and tablet\ncomputers yields three findings. First, Chinese customers follow the Doctrine\nof the Mean, and often use euphemistic expressions, while American customers\nexpress their opinions more directly. Second, Chinese customers care more about\ngeneral feelings, while American customers pay more attention to product\ndetails. Third, Chinese customers focus on external features, while American\ncustomers care more about the internal features of products.', 'Topic evolution modeling has been researched for a long time and has gained\nconsiderable interest. A state-of-the-art method has been recently using word\nmodeling algorithms in combination with community detection mechanisms to\nachieve better results in a more effective way. We analyse results of this\napproach and discuss the two major challenges that this approach still faces.\nAlthough the topics that have resulted from the recent algorithm are good in\ngeneral, they are very noisy due to many topics that are very unimportant\nbecause of their size, words, or ambiguity. Additionally, the number of words\ndefining each topic is too large, making it difficult to analyse them in their\nunsorted state. In this paper, we propose approaches to tackle these challenges\nby adding topic filtering and network analysis metrics to define the importance\nof a topic. We test different combinations of these metrics to see which\ncombination yields the best results. Furthermore, we add word filtering and\nranking to each topic to identify the words with the highest novelty\nautomatically. We evaluate our enhancement methods in two ways: human\nqualitative evaluation and automatic quantitative evaluation. Moreover, we\ncreated two case studies to test the quality of the clusters and words. In the\nquantitative evaluation, we use the pairwise mutual information score to test\nthe coherency of topics. The quantitative evaluation also includes an analysis\nof execution times for each part of the program. The results of the\nexperimental evaluations show that the two evaluation methods agree on the\npositive feasibility of the algorithm. We then show possible extensions in the\nform of usability and future improvements to the algorithm.', 'Topic modeling is used for discovering latent semantic structure, usually\nreferred to as topics, in a large collection of documents. The most widely used\nmethods are Latent Dirichlet Allocation and Probabilistic Latent Semantic\nAnalysis. Despite their popularity they have several weaknesses. In order to\nachieve optimal results they often require the number of topics to be known,\ncustom stop-word lists, stemming, and lemmatization. Additionally these methods\nrely on bag-of-words representation of documents which ignore the ordering and\nsemantics of words. Distributed representations of documents and words have\ngained popularity due to their ability to capture semantics of words and\ndocuments. We present $\\texttt{top2vec}$, which leverages joint document and\nword semantic embedding to find $\\textit{topic vectors}$. This model does not\nrequire stop-word lists, stemming or lemmatization, and it automatically finds\nthe number of topics. The resulting topic vectors are jointly embedded with the\ndocument and word vectors with distance between them representing semantic\nsimilarity. Our experiments demonstrate that $\\texttt{top2vec}$ finds topics\nwhich are significantly more informative and representative of the corpus\ntrained on than probabilistic generative models.', 'The application and usage of opinion mining, especially for business\nintelligence, product recommendation, targeted marketing etc. have fascinated\nmany research attentions around the globe. Various research efforts attempted\nto mine opinions from customer reviews at different levels of granularity,\nincluding word-, sentence-, and document-level. However, development of a fully\nautomatic opinion mining and sentiment analysis system is still elusive. Though\nthe development of opinion mining and sentiment analysis systems are getting\nmomentum, most of them attempt to perform document-level sentiment analysis,\nclassifying a review document as positive, negative, or neutral. Such\ndocument-level opinion mining approaches fail to provide insight about users\nsentiment on individual features of a product or service. Therefore, it seems\nto be a great help for both customers and manufacturers, if the reviews could\nbe processed at a finer-grained level and presented in a summarized form\nthrough some visual means, highlighting individual features of a product and\nusers sentiment expressed over them. In this paper, the design of a unified\nopinion mining and sentiment analysis framework is presented at the\nintersection of both machine learning and natural language processing\napproaches. Also, design of a novel feature-level review summarization scheme\nis proposed to visualize mined features, opinions and their polarity values in\na comprehendible way.', ""A new opinion extraction method is proposed to summarize unstructured,\nuser-generated content (i.e., online customer reviews) in the fixed topic\ndomains. To differentiate the current approach from other opinion extraction\napproaches, which are often exposed to a sparsity problem and lack of sentiment\nscores, a confirmatory aspect-based opinion mining framework is introduced\nalong with its practical algorithm called DiSSBUS. In this procedure, 1) each\ncustomer review is disintegrated into a set of clauses; 2) each clause is\nsummarized to bi-terms-a topic word and an evaluation word-using a\npart-of-speech (POS) tagger; and 3) each bi-term is matched to a pre-specified\ntopic relevant to a specific domain. The proposed processes have two primary\nadvantages over existing methods: 1) they can decompose a single review into a\nset of bi-terms related to pre-specified topics in the domain of interest and,\ntherefore, 2) allow identification of the reviewer's opinions on the topics via\nevaluation words within the set of bi-terms. The proposed aspect-based opinion\nmining is applied to customer reviews of restaurants in Hawaii obtained from\nTripAdvisor, and the empirical findings validate the effectiveness of the\nmethod.\n  Keywords: Clause-based sentiment analysis, Customer review, Opinion mining,\nTopic modeling, User-generate-contents.""]","[""With the increasing number of merchandise on e-commerce platforms, users tend\nto refer to reviews of other shoppers to decide which product they should buy.\nHowever, with so many reviews of a product, users often have to spend lots of\ntime browsing through reviews talking about product attributes they do not care\nabout. We want to establish a system that can automatically summarize and\nanswer user's product specific questions.\n  In this study, we propose a novel seq2seq based text generation model to\ngenerate answers to user's question based on reviews posted by previous users.\nGiven a user question and/or target sentiment polarity, we extract aspects of\ninterest and generate an answer that summarizes previous relevant user reviews.\nSpecifically, our model performs attention between input reviews and target\naspects during encoding and is conditioned on both review rating and input\ncontext during decoding. We also incorporate a pre-trained auxiliary rating\nclassifier to improve model performance and accelerate convergence during\ntraining. Experiments using real-world e-commerce dataset show that our model\nachieves improvement in performance compared to previously introduced models."", ""In order to better engage with customers, retailers rely on extensive\ncustomer and product databases which allows them to better understand customer\nbehaviour and purchasing patterns. This has long been a challenging task as\ncustomer modelling is a multi-faceted, noisy and time-dependent problem. The\nmost common way to tackle this problem is indirectly through task-specific\nsupervised learning prediction problems, with relatively little literature on\nmodelling a customer by directly simulating their future transactions. In this\npaper we propose a method for generating realistic sequences of baskets that a\ngiven customer is likely to purchase over a period of time. Customer embedding\nrepresentations are learned using a Recurrent Neural Network (RNN) which takes\ninto account the entire sequence of transaction data. Given the customer state\nat a specific point in time, a Generative Adversarial Network (GAN) is trained\nto generate a plausible basket of products for the following week. The newly\ngenerated basket is then fed back into the RNN to update the customer's state.\nThe GAN is thus used in tandem with the RNN module in a pipeline alternating\nbetween basket generation and customer state updating steps. This allows for\nsampling over a distribution of a customer's future sequence of baskets, which\nthen can be used to gain insight into how to service the customer more\neffectively. The methodology is empirically shown to produce baskets that\nappear similar to real baskets and enjoy many common properties, including\nfrequencies of different product types, brands, and prices. Furthermore, the\ngenerated data is able to replicate most of the strongest sequential patterns\nthat exist between product types in the real data."", ""Text mining approaches are being used increasingly for business analytics. In\nparticular, such approaches are now central to understanding users' feedback\nregarding systems delivered via online application distribution platforms such\nas Google Play. In such settings, large volumes of reviews of potentially\nnumerous apps and systems means that it is infeasible to use manual mechanisms\nto extract insights and knowledge that could inform product improvement. In\nthis context of identifying software system improvement options, text mining\ntechniques are used to reveal the features that are mentioned most often as\nbeing in need of correction (e.g., GPS), and topics that are associated with\nfeatures perceived as being defective (e.g., inaccuracy of GPS). Other\napproaches may supplement such techniques to provide further insights for\nonline communities and solution providers. In this work we augment text mining\napproaches with social network analysis to demonstrate the utility of using\nmultiple techniques. Our outcomes suggest that text mining approaches may\nindeed be supplemented with other methods to deliver a broader range of\ninsights."", 'Commercial establishments like restaurants, service centres and retailers\nhave several sources of customer feedback about products and services, most of\nwhich need not be as structured as rated reviews provided by services like\nYelp, or Amazon, in terms of sentiment conveyed. For instance, Amazon provides\na fine-grained score on a numeric scale for product reviews. Some sources,\nhowever, like social media (Twitter, Facebook), mailing lists (Google Groups)\nand forums (Quora) contain text data that is much more voluminous, but\nunstructured and unlabelled. It might be in the best interests of a business\nestablishment to assess the general sentiment towards their brand on these\nplatforms as well. This text could be pipelined into a system with a built-in\nprediction model, with the objective of generating real-time graphs on opinion\nand sentiment trends. Although such tasks like the one described about have\nbeen explored with respect to document classification problems in the past, the\nimplementation described in this paper, by virtue of learning a continuous\nfunction rather than a discrete one, offers a lot more depth of insight as\ncompared to document classification approaches. This study aims to explore the\nvalidity of such a continuous function predicting model to quantify sentiment\nabout an entity, without the additional overhead of manual labelling, and\ncomputational preprocessing & feature extraction. This research project also\naims to design and implement a re-usable document regression pipeline as a\nframework, Rapid-Rate, that can be used to predict document scores in\nreal-time.', 'Customers post online reviews at any time. With the timestamp of online\nreviews, they can be regarded as a flow of information. With this\ncharacteristic, designers can capture the changes in customer feedback to help\nset up product improvement strategies. Here we propose an approach for\ncapturing changes of user expectation on product affordances based on the\nonline reviews for two generations of products. First, the approach uses a\nrule-based natural language processing method to automatically identify and\nstructure product affordances from review text. Then, inspired by the Kano\nmodel which classifies preferences of product attributes in five categories,\nconjoint analysis is used to quantitatively categorize the structured\naffordances. Finally, changes of user expectation can be found by applying the\nconjoint analysis on the online reviews posted for two successive generations\nof products. A case study based on the online reviews of Kindle e-readers\ndownloaded from amazon.com shows that designers can use our proposed approach\nto evaluate their product improvement strategies for previous products and\ndevelop new product improvement strategies for future products.', 'We address two challenges in topic models: (1) Context information around\nwords helps in determining their actual meaning, e.g., ""networks"" used in the\ncontexts ""artificial neural networks"" vs. ""biological neuron networks"".\nGenerative topic models infer topic-word distributions, taking no or only\nlittle context into account. Here, we extend a neural autoregressive topic\nmodel to exploit the full context information around words in a document in a\nlanguage modeling fashion. The proposed model is named as iDocNADE. (2) Due to\nthe small number of word occurrences (i.e., lack of context) in short text and\ndata sparsity in a corpus of few documents, the application of topic models is\nchallenging on such texts. Therefore, we propose a simple and efficient way of\nincorporating external knowledge into neural autoregressive topic models: we\nuse embeddings as a distributional prior. The proposed variants are named as\nDocNADEe and iDocNADEe.\n  We present novel neural autoregressive topic model variants that consistently\noutperform state-of-the-art generative topic models in terms of generalization,\ninterpretability (topic coherence) and applicability (retrieval and\nclassification) over 7 long-text and 8 short-text datasets from diverse\ndomains.', 'One of the biggest hurdles for customers when purchasing fashion online, is\nthe difficulty of finding products with the right fit. In order to provide a\nbetter online shopping experience, platforms need to find ways to recommend the\nright product sizes and the best fitting products to their customers. These\nrecommendation systems, however, require customer feedback in order to estimate\nthe most suitable sizing options. Such feedback is rare and often only\navailable as natural text. In this paper, we examine the extraction of product\nfit feedback from customer reviews using natural language processing\ntechniques. In particular, we compare traditional methods with more recent\ntransfer learning techniques for text classification, and analyze their\nresults. Our evaluation shows, that the transfer learning approach ULMFit is\nnot only comparatively fast to train, but also achieves highest accuracy on\nthis task. The integration of the extracted information with actual size\nrecommendation systems is left for future work.', ""Online product reviews are a valuable resource for product developers to\nimprove the design of their products. Yet, the potential value of customer\nfeedback to improve the sustainability performance of products is still to be\nexploited. The present paper investigates and analyzes Amazon product reviews\nto bring new light on the following question: ``What sustainable design\ninsights can be identified or interpreted from online product reviews?''. To do\nso, the top 100 reviews, evenly distributed by star ratings, for three product\ncategories (laptop, printer, cable) are collected, manually annotated, analyzed\nand interpreted. For each product category, the reviews of two similar products\n(one with environmental certification and one standard version) are compared\nand combined to come up with sustainable design solutions. In all, for the six\nproducts considered, between 12% and 20% of the reviews mentioned directly or\nindirectly aspects or attributes that could be exploited to improve the design\nof these products from a sustainability perspective. Concrete examples of\nsustainable design leads that could be elicited from product reviews are given\nand discussed. As such, this contribution provides a baseline for future work\nwilling to automate this process to gain further insights from online product\nreviews. Notably, the deployment of machine learning tools and the use of\nnatural language processing techniques to do so are discussed as promising\nlines for future research."", 'In this paper, we propose two automated text processing frameworks\nspecifically designed to analyze online reviews. The objective of the first\nframework is to summarize the reviews dataset by extracting essential sentence.\nThis is performed by converting sentences into numerical vectors and clustering\nthem using a community detection algorithm based on their similarity levels.\nAfterwards, a correlation score is measured for each sentence to determine its\nimportance level in each cluster and assign it as a tag for that community. The\nsecond framework is based on a question-answering neural network model trained\nto extract answers to multiple different questions. The collected answers are\neffectively clustered to find multiple distinct answers to a single question\nthat might be asked by a customer. The proposed frameworks are shown to be more\ncomprehensive than existing reviews processing solutions.', 'Topic modeling is commonly used to analyze and understand large document\ncollections. However, in practice, users want to focus on specific aspects or\n""targets"" rather than the entire corpus. For example, given a large collection\nof documents, users may want only a smaller subset which more closely aligns\nwith their interests, tasks, and domains. In particular, our paper focuses on\nlarge-scale document retrieval with high recall where any missed relevant\ndocuments can be critical. A simple keyword matching search is generally not\neffective nor efficient as 1) it is difficult to find a list of keyword queries\nthat can cover the documents of interest before exploring the dataset, 2) some\ndocuments may not contain the exact keywords of interest but may still be\nhighly relevant, and 3) some words have multiple meanings, which would result\nin irrelevant documents included in the retrieved subset. In this paper, we\npresent TopicSifter, a visual analytics system for interactive search space\nreduction. Our system utilizes targeted topic modeling based on nonnegative\nmatrix factorization and allows users to give relevance feedback in order to\nrefine their target and guide the topic modeling to the most relevant results.', ""Mobile apps are one of the building blocks of the mobile digital economy. A\ndifferentiating feature of mobile apps to traditional enterprise software is\nonline reviews, which are available on app marketplaces and represent a\nvaluable source of consumer feedback on the app. We create a supervised topic\nmodeling approach for app developers to use mobile reviews as useful sources of\nquality and customer feedback, thereby complementing traditional software\ntesting. The approach is based on a constrained matrix factorization that\nleverages the relationship between term frequency and a given response variable\nin addition to co-occurrences between terms to recover topics that are both\npredictive of consumer sentiment and useful for understanding the underlying\ntextual themes. The factorization is combined with ordinal regression to\nprovide guidance from online reviews on a single app's performance as well as\nsystematically compare different apps over time for benchmarking of features\nand consumer sentiment. We apply our approach using a dataset of over 100,000\nmobile reviews over several years for three of the most popular online travel\nagent apps from the iTunes and Google Play marketplaces."", 'E-commerce stores collect customer feedback to let sellers learn about\ncustomer concerns and enhance customer order experience. Because customer\nfeedback often contains redundant information, a concise summary of the\nfeedback can be generated to help sellers better understand the issues causing\ncustomer dissatisfaction. Previous state-of-the-art abstractive text\nsummarization models make two major types of factual errors when producing\nsummaries from customer feedback, which are wrong entity detection (WED) and\nincorrect product-defect description (IPD). In this work, we introduce a set of\nmethods to enhance the factual consistency of abstractive summarization on\ncustomer feedback. We augment the training data with artificially corrupted\nsummaries, and use them as counterparts of the target summaries. We add a\ncontrastive loss term into the training objective so that the model learns to\navoid certain factual errors. Evaluation results show that a large portion of\nWED and IPD errors are alleviated for BART and T5. Furthermore, our approaches\ndo not depend on the structure of the summarization model and thus are\ngeneralizable to any abstractive summarization systems.', ""Today's businesses face a high pressure to innovate in order to succeed in\nhighly competitive markets. Successful innovations, though, typically require\nthe identification and analysis of customer needs. While traditional,\nestablished need elicitation methods are time-proven and have demonstrated\ntheir capabilities to deliver valuable insights, they lack automation and\nscalability and, thus, are expensive and time-consuming. In this article, we\npropose an approach to automatically identify and quantify customer needs by\nutilizing a novel data source: Users voluntarily and publicly expose\ninformation about themselves via social media, as for instance Facebook or\nTwitter. These posts may contain valuable information about the needs, wants,\nand demands of their authors. We apply a Design Science Research (DSR)\nmethodology to add design knowledge and artifacts for the digitalization of\ninnovation processes, in particular to provide digital support for the\nelicitation of customer needs. We want to investigate whether automated,\nspeedy, and scalable need elicitation from social media is feasible. We\nconcentrate on Twitter as a data source and on e-mobility as an application\ndomain. In a first design cycle we conceive, implement and evaluate a method to\ndemonstrate the feasibility of identifying those social media posts that\nactually express customer needs. In a second cycle, we build on this artifact\nto additionally quantify the need information elicited, and prove its\nfeasibility. Third, we integrate both developed methods into an end-user\nsoftware artifact and test usability in an industrial use case. Thus, we add\nnew methods for need elicitation to the body of knowledge, and introduce\nconcrete tooling for innovation management in practice."", 'Auto dealerships receive thousands of calls daily from customers who are\ninterested in sales, service, vendors and jobseekers. With so many calls, it is\nvery important for auto dealers to understand the intent of these calls to\nprovide positive customer experiences that ensure customer satisfaction, deep\ncustomer engagement to boost sales and revenue, and optimum allocation of\nagents or customer service representatives across the business. In this paper,\nwe define the problem of customer phone call intent as a multi-class\nclassification problem stemming from the large database of recorded phone call\ntranscripts. To solve this problem, we develop a convolutional neural network\n(CNN)-based supervised learning model to classify the customer calls into four\nintent categories: sales, service, vendor and jobseeker. Experimental results\nshow that with the thrust of our scalable data labeling method to provide\nsufficient training data, the CNN-based predictive model performs very well on\nlong text classification according to the quantitative metrics of F1-Score,\nprecision, recall, and accuracy.', 'Topical Segmentation poses a great role in reducing search space of the\ntopics taught in a lecture video specially when the video metadata lacks topic\nwise segmentation information. This segmentation information eases user efforts\nof searching, locating and browsing a topic inside a lecture video. In this\nwork we propose an algorithm, that combines state-of-the art language model and\ndomain knowledge graph for automatically detecting different coherent topics\npresent inside a long lecture video. We use the language model on\nspeech-to-text transcription to capture the implicit meaning of the whole video\nwhile the knowledge graph provides us the domain specific dependencies between\ndifferent concepts of that subjects. Also leveraging the domain knowledge we\ncan capture the way instructor binds and connects different concepts while\nteaching, which helps us in achieving better segmentation accuracy. We tested\nour approach on NPTEL lecture videos and holistic evaluation shows that it out\nperforms the other methods described in the literature.', 'Social networking sites (SNSs) facilitate the sharing of ideas and\ninformation through different types of feedback including publishing posts,\nleaving comments and other type of reactions. However, some comments or\nfeedback on SNSs are inconsiderate and offensive, and sometimes this type of\nfeedback has a very negative effect on a target user. The phenomenon known as\nflaming goes hand-in-hand with this type of posting that can trigger almost\ninstantly on SNSs. Most popular users such as celebrities, politicians and news\nmedia are the major victims of the flaming behaviors and so detecting these\ntypes of events will be useful and appreciated. Flaming event can be monitored\nand identified by analyzing negative comments received on a post. Thus, our\nmain objective of this study is to identify a way to detect flaming events in\nSNS using a sentiment prediction method. We use a deep Neural Network (NN)\nmodel that can identity sentiments of variable length sentences and classifies\nthe sentiment of SNSs content (both comments and posts) to discover flaming\nevents. Our deep NN model uses Word2Vec and FastText word embedding methods as\nits training to explore which method is the most appropriate. The labeled\ndataset for training the deep NN is generated using an enhanced lexicon based\napproach. Our deep NN model classifies the sentiment of a sentence into five\nclasses: Very Positive, Positive, Neutral, Negative and Very Negative. To\ndetect flaming incidents, we focus only on the comments classified into the\nNegative and Very Negative classes. As a use-case, we try to explore the\nflaming phenomena in the news media domain and therefore we focused on news\nitems posted by three popular news media on Facebook (BBCNews, CNN and FoxNews)\nto train and test the model.', 'The most important goal of customer services is to keep the customer\nsatisfied. However, service resources are always limited and must be\nprioritized. Therefore, it is important to identify customers who potentially\nbecome unsatisfied and might lead to escalations. Today this prioritization of\ncustomers is often done manually. Data science on IoT data (esp. log data) for\nmachine health monitoring, as well as analytics on enterprise data for customer\nrelationship management (CRM) have mainly been researched and applied\nindependently. In this paper, we present a framework for a data-driven decision\nsupport system which combines IoT and enterprise data to model customer\nsentiment. Such decision support systems can help to prioritize customers and\nservice resources to effectively troubleshoot problems or even avoid them. The\nframework is applied in a real-world case study with a major medical device\nmanufacturer. This includes a fully automated and interpretable machine\nlearning pipeline designed to meet the requirements defined with domain experts\nand end users. The overall framework is currently deployed, learns and\nevaluates predictive models from terabytes of IoT and enterprise data to\nactively monitor the customer sentiment for a fleet of thousands of high-end\nmedical devices. Furthermore, we provide an anonymized industrial benchmark\ndataset for the research community.', ""Requirements elicitation has recently been complemented with crowd-based\ntechniques, which continuously involve large, heterogeneous groups of users who\nexpress their feedback through a variety of media. Crowd-based elicitation has\ngreat potential for engaging with (potential) users early on but also results\nin large sets of raw and unstructured feedback. Consolidating and analyzing\nthis feedback is a key challenge for turning it into sensible user\nrequirements. In this paper, we focus on topic modeling as a means to identify\ntopics within a large set of crowd-generated user stories and compare three\napproaches: (1) a traditional approach based on Latent Dirichlet Allocation,\n(2) a combination of word embeddings and principal component analysis, and (3)\na combination of word embeddings and Word Mover's Distance. We evaluate the\napproaches on a publicly available set of 2,966 user stories written and\ncategorized by crowd workers. We found that a combination of word embeddings\nand Word Mover's Distance is most promising. Depending on the word embeddings\nwe use in our approaches, we manage to cluster the user stories in two ways:\none that is closer to the original categorization and another that allows new\ninsights into the dataset, e.g. to find potentially new categories.\nUnfortunately, no measure exists to rate the quality of our results\nobjectively. Still, our findings provide a basis for future work towards\nanalyzing crowd-sourced user stories."", 'Topic segmentation and labeling is often considered a prerequisite for\nhigher-level conversation analysis and has been shown to be useful in many\nNatural Language Processing (NLP) applications. We present two new corpora of\nemail and blog conversations annotated with topics, and evaluate annotator\nreliability for the segmentation and labeling tasks in these asynchronous\nconversations. We propose a complete computational framework for topic\nsegmentation and labeling in asynchronous conversations. Our approach extends\nstate-of-the-art methods by considering a fine-grained structure of an\nasynchronous conversation, along with other conversational features by applying\nrecent graph-based methods for NLP. For topic segmentation, we propose two\nnovel unsupervised models that exploit the fine-grained conversational\nstructure, and a novel graph-theoretic supervised model that combines lexical,\nconversational and topic features. For topic labeling, we propose two novel\n(unsupervised) random walk models that respectively capture conversation\nspecific clues from two different sources: the leading sentences and the\nfine-grained conversational structure. Empirical evaluation shows that the\nsegmentation and the labeling performed by our best models beat the\nstate-of-the-art, and are highly correlated with human annotations.', 'Product recommendation systems are important for major movie studios during\nthe movie greenlight process and as part of machine learning personalization\npipelines. Collaborative Filtering (CF) models have proved to be effective at\npowering recommender systems for online streaming services with explicit\ncustomer feedback data. CF models do not perform well in scenarios in which\nfeedback data is not available, in cold start situations like new product\nlaunches, and situations with markedly different customer tiers (e.g., high\nfrequency customers vs. casual customers). Generative natural language models\nthat create useful theme-based representations of an underlying corpus of\ndocuments can be used to represent new product descriptions, like new movie\nplots. When combined with CF, they have shown to increase the performance in\ncold start situations. Outside of those cases though in which explicit customer\nfeedback is available, recommender engines must rely on binary purchase data,\nwhich materially degrades performance. Fortunately, purchase data can be\ncombined with product descriptions to generate meaningful representations of\nproducts and customer trajectories in a convenient product space in which\nproximity represents similarity. Learning to measure the distance between\npoints in this space can be accomplished with a deep neural network that trains\non customer histories and on dense vectorizations of product descriptions. We\ndeveloped a system based on Collaborative (Deep) Metric Learning (CML) to\npredict the purchase probabilities of new theatrical releases. We trained and\nevaluated the model using a large dataset of customer histories, and tested the\nmodel for a set of movies that were released outside of the training window.\nInitial experiments show gains relative to models that do not train on\ncollaborative preferences.', 'We address two challenges of probabilistic topic modelling in order to better\nestimate the probability of a word in a given context, i.e., P(word|context):\n(1) No Language Structure in Context: Probabilistic topic models ignore word\norder by summarizing a given context as a ""bag-of-word"" and consequently the\nsemantics of words in the context is lost. The LSTM-LM learns a vector-space\nrepresentation of each word by accounting for word order in local collocation\npatterns and models complex characteristics of language (e.g., syntax and\nsemantics), while the TM simultaneously learns a latent representation from the\nentire document and discovers the underlying thematic structure. We unite two\ncomplementary paradigms of learning the meaning of word occurrences by\ncombining a TM (e.g., DocNADE) and a LM in a unified probabilistic framework,\nnamed as ctx-DocNADE. (2) Limited Context and/or Smaller training corpus of\ndocuments: In settings with a small number of word occurrences (i.e., lack of\ncontext) in short text or data sparsity in a corpus of few documents, the\napplication of TMs is challenging. We address this challenge by incorporating\nexternal knowledge into neural autoregressive topic models via a language\nmodelling approach: we use word embeddings as input of a LSTM-LM with the aim\nto improve the word-topic mapping on a smaller and/or short-text corpus. The\nproposed DocNADE extension is named as ctx-DocNADEe.\n  We present novel neural autoregressive topic model variants coupled with\nneural LMs and embeddings priors that consistently outperform state-of-the-art\ngenerative TMs in terms of generalization (perplexity), interpretability (topic\ncoherence) and applicability (retrieval and classification) over 6 long-text\nand 8 short-text datasets from diverse domains.', 'Analyzing customer feedback is the best way to channelize the data into new\nmarketing strategies that benefit entrepreneurs as well as customers. Therefore\nan automated system which can analyze the customer behavior is in great demand.\nUsers may write feedbacks in any language, and hence mining appropriate\ninformation often becomes intractable. Especially in a traditional\nfeature-based supervised model, it is difficult to build a generic system as\none has to understand the concerned language for finding the relevant features.\nIn order to overcome this, we propose deep Convolutional Neural Network (CNN)\nand Recurrent Neural Network (RNN) based approaches that do not require\nhandcrafting of features. We evaluate these techniques for analyzing customer\nfeedback sentences in four languages, namely English, French, Japanese and\nSpanish. Our empirical analysis shows that our models perform well in all the\nfour languages on the setups of IJCNLP Shared Task on Customer Feedback\nAnalysis. Our model achieved the second rank in French, with an accuracy of\n71.75% and third ranks for all the other languages.', ""Improving the energy efficiency of mobile applications is a topic that has\ngained a lot of attention recently. It has been addressed in a number of ways\nsuch as identifying energy bugs and developing a catalog of energy patterns.\nPrevious work shows that users discuss the battery-related issues (energy\ninefficiency or energy consumption) of the apps in their reviews. However,\nthere is no work that addresses the automatic extraction of battery-related\nissues from users' feedback. In this paper, we report on a visualization tool\nthat is developed to empirically study machine learning algorithms and text\nfeatures to automatically identify the energy consumption specific reviews with\nthe highest accuracy. Other than the common machine learning algorithms, we\nutilize deep learning models with different word embeddings to compare the\nresults. Furthermore, to help the developers extract the main topics that are\ndiscussed in the reviews, two states of the art topic modeling algorithms are\napplied. The visualizations of the topics represent the keywords that are\nextracted for each topic along with a comparison with the results of string\nmatching. The developed web-browser based interactive visualization tool is a\nnovel framework developed with the intention of giving the app developers\ninsights about running time and accuracy of machine learning and deep learning\nmodels as well as extracted topics. The tool makes it easier for the developers\nto traverse through the extensive result set generated by the text\nclassification and topic modeling algorithms. The dynamic-data structure used\nfor the tool stores the baseline-results of the discussed approaches and is\nupdated when applied on new datasets. The tool is open-sourced to replicate the\nresearch results."", ""Business Processes, i.e., a set of coordinated tasks and activities to\nachieve a business goal, and their continuous improvements are key to the\noperation of any organization. In banking, business processes are increasingly\ndynamic as various technologies have made dynamic processes more prevalent. For\nexample, customer segmentation, i.e., the process of grouping related customers\nbased on common activities and behaviors, could be a data-driven and\nknowledge-intensive process. In this paper, we present an intelligent\ndata-driven pipeline composed of a set of processing elements to move\ncustomers' data from one system to another, transforming the data into the\ncontextualized data and knowledge along the way. The goal is to present a novel\nintelligent customer segmentation process which automates the feature\nengineering, i.e., the process of using (banking) domain knowledge to extract\nfeatures from raw data via data mining techniques, in the banking domain. We\nadopt a typical scenario for analyzing customer transaction records, to\nhighlight how the presented approach can significantly improve the quality of\nrisk-based customer segmentation in the absence of feature engineering.As\nresult, our proposed method is able to achieve accuracy of 91% compared to\nclassical approaches in terms of detecting, identifying and classifying\ntransaction to the right classification."", 'Understanding product attributes plays an important role in improving online\nshopping experience for customers and serves as an integral part for\nconstructing a product knowledge graph. Most existing methods focus on\nattribute extraction from text description or utilize visual information from\nproduct images such as shape and color. Compared to the inputs considered in\nprior works, a product image in fact contains more information, represented by\na rich mixture of words and visual clues with a layout carefully designed to\nimpress customers. This work proposes a more inclusive framework that fully\nutilizes these different modalities for attribute extraction. Inspired by\nrecent works in visual question answering, we use a transformer based sequence\nto sequence model to fuse representations of product text, Optical Character\nRecognition (OCR) tokens and visual objects detected in the product image. The\nframework is further extended with the capability to extract attribute value\nacross multiple product categories with a single model, by training the decoder\nto predict both product category and attribute value and conditioning its\noutput on product category. The model provides a unified attribute extraction\nsolution desirable at an e-commerce platform that offers numerous product\ncategories with a diverse body of product attributes. We evaluated the model on\ntwo product attributes, one with many possible values and one with a small set\nof possible values, over 14 product categories and found the model could\nachieve 15% gain on the Recall and 10% gain on the F1 score compared to\nexisting methods using text-only features.', 'Efficient and appropriate online customer service is essential to large\ne-commerce businesses. Existing solution recommendation methods for online\ncustomer service are unable to determine the best solutions at runtime, leading\nto poor satisfaction of end customers. This paper proposes a novel intelligent\nframework, called ICS-Assist, to recommend suitable customer service solutions\nfor service staff at runtime. Specifically, we develop a generalizable\ntwo-stage machine learning model to identify customer service scenarios and\ndetermine customer service solutions based on a scenario-solution mapping\ntable. We implement ICS-Assist and evaluate it using an over 6-month field\nstudy with Alibaba Group. In our experiment, over 12,000 customer service staff\nuse ICS-Assist to serve for over 230,000 cases per day on average. The\nexperimen-tal results show that ICS-Assist significantly outperforms the\ntraditional manual method, and improves the solution acceptance rate, the\nsolution coverage rate, the average service time, the customer satisfaction\nrate, and the business domain catering rate by up to 16%, 25%, 6%, 14% and 17%\nrespectively, compared to the state-of-the-art methods.', 'Accurate user interest modeling is important for news recommendation. Most\nexisting methods for news recommendation rely on implicit feedbacks like click\nfor inferring user interests and model training. However, click behaviors\nusually contain heavy noise, and cannot help infer complicated user interest\nsuch as dislike. Besides, the feed recommendation models trained solely on\nclick behaviors cannot optimize other objectives such as user engagement. In\nthis paper, we present a news feed recommendation method that can exploit\nvarious kinds of user feedbacks to enhance both user interest modeling and\nmodel training. We propose a unified user modeling framework to incorporate\nvarious explicit and implicit user feedbacks to infer both positive and\nnegative user interests. In addition, we propose a strong-to-weak attention\nnetwork that uses the representations of stronger feedbacks to distill positive\nand negative user interests from implicit weak feedbacks for accurate user\ninterest modeling. Besides, we propose a multi-feedback model training\nframework to learn an engagement-aware feed recommendation model. Extensive\nexperiments on a real-world dataset show that our approach can effectively\nimprove the model performance in terms of both news clicks and user engagement.', 'Understanding choices performed by online customers is a growing need in the\ntravel industry. In many practical situations, the only available information\nis the flight search query performed by the customer with no additional profile\nknowledge. In general, customer flight bookings are driven by prices, duration,\nnumber of connections, and so on. However, not all customers might assign the\nsame importance to each of those criteria. Here comes the need of grouping\ntogether all flight searches performed by the same kind of customer, that is\nhaving the same booking criteria. The effectiveness of some set of\nrecommendations, for a single cluster, can be measured in terms of the number\nof bookings historically performed. This effectiveness measure plays the role\nof a feedback, that is an external knowledge which can be recombined to\niteratively obtain a final segmentation. In this paper, we describe our Online\nTravel Agencies (OTA) flight search use case and highlight its specific\nfeatures. We address the flight search segmentation problem motivated above by\nproposing a novel algorithm called Split-or-Merge (S/M). This algorithm is a\nvariation of the Split-Merge-Evolve (SME) method. The SME method has already\nbeen introduced in the community as an iterative process updating a clustering\ngiven by the K-means algorithm by splitting and merging clusters subject to\nfeedback independent evaluations. No previous application of the SME method to\nthe real-word data is reported in literature to the best of our knowledge.\nHere, we provide experimental evaluations over real-world data to the SME and\nthe S/M methods. The impact on our domain-specific metrics obtained under the\nSME and the S/M methods suggests that feedback clustering techniques can be\nvery promising in the handling of the domain of OTA flight searches.', 'Current topic models often suffer from discovering topics not matching human\nintuition, unnatural switching of topics within documents and high\ncomputational demands. We address these concerns by proposing a topic model and\nan inference algorithm based on automatically identifying characteristic\nkeywords for topics. Keywords influence topic-assignments of nearby words. Our\nalgorithm learns (key)word-topic scores and it self-regulates the number of\ntopics. Inference is simple and easily parallelizable. Qualitative analysis\nyields comparable results to state-of-the-art models (eg. LDA), but with\ndifferent strengths and weaknesses. Quantitative analysis using 9 datasets\nshows gains in terms of classification accuracy, PMI score, computational\nperformance and consistency of topic assignments within documents, while most\noften using less topics.', 'Crowdsourced design feedback systems are emerging resources for getting large\namounts of feedback in a short period of time. Traditionally, the feedback\ncomes in the form of a declarative statement, which often contains positive or\nnegative sentiment. Prior research has shown that overly negative or positive\nsentiment can strongly influence the perceived usefulness and acceptance of\nfeedback and, subsequently, lead to ineffective design revisions. To enhance\nthe effectiveness of crowdsourced design feedback, we investigate a new\napproach for mitigating the effects of negative or positive feedback by\ncombining open-ended and thought-provoking questions with declarative feedback\nstatements. We conducted two user studies to assess the effects of\nquestion-based feedback on the sentiment and quality of design revisions in the\ncontext of graphic design. We found that crowdsourced question-based feedback\ncontains more neutral sentiment than statement-based feedback. Moreover, we\nprovide evidence that presenting feedback as questions followed by statements\nleads to better design revisions than question- or statement-based feedback\nalone.', 'Currently, while software engineers write code for various modules, quite\noften, various types of errors - coding, logic, semantic, and others (most of\nwhich are not caught by compilation and other tools) get introduced. Some of\nthese bugs might be found in the later stage of testing, and many times it is\nreported by customers on production code. Companies have to spend many\nresources, both money and time in finding and fixing the bugs which would have\nbeen avoided if coding was done right. Also, concealed flaws in software can\nlead to security vulnerabilities that potentially allow attackers to compromise\nsystems and applications. Interestingly, same or similar issues/bugs, which\nwere fixed in the past (although in different modules), tend to get introduced\nin production code again.\n  We developed a novel AI-based system which uses the deep representation of\nAbstract Syntax Tree (AST) created from the source code and also the active\nfeedback loop to identify and alert the potential bugs that could be caused at\nthe time of development itself i.e. as the developer is writing new code (logic\nand/or function). This tool integrated with IDE as a plugin would work in the\nbackground, point out existing similar functions/code-segments and any\nassociated bugs in those functions. The tool would enable the developer to\nincorporate suggestions right at the time of development, rather than waiting\nfor UT/QA/customer to raise a defect.\n  We assessed our tool on both open-source code and also on Cisco codebase for\nC and C++ programing language. Our results confirm that deep representation of\nsource code and the active feedback loop is an assuring approach for predicting\nsecurity and other vulnerabilities present in the code.', ""Customer Segmentation is the process of dividing customers into groups based\non common characteristics. An intelligent Customer Segmentation will not only\nenable an organization to effectively allocate marketing resources (e.g.,\nRecommender Systems in the Banking sector) but also it will enable identifying\nthe customer cohorts that are most likely to benefit from a specific policy\n(e.g., to discover diverse patient groups in the Health sector). While there\nhas been a significant improvement in approaches to Customer Segmentation, the\nmain challenge remains to be the understanding of the reasons behind the\nsegmentation need. This task is challenging as it is subjective and depends on\nthe goal of segmentation as well as the analyst's perspective. To address this\nchallenge, in this paper, we present an intelligent vector-based customer\nsegmentation approach. The proposed approach will leverage feature engineering\nto enable analysts to identify important features (from a pool of features such\nas demographics, geography, psychographics, behavioral, and more) and feed them\ninto a neural embedding framework named Customer2Vec. The Customer2Vec combines\nthe neural network classification and clustering methods as supervised and\nunsupervised learning techniques to embed the customer vector. We adopt a\ntypical scenario in the Banking Sector to highlight how Customer2Vec\nsignificantly improves the quality of the segmentation and detecting customer\nsimilarities."", 'Product feature recommendations are critical for online customers to purchase\nthe right products based on the right features. For a customer, selecting the\nproduct that has the best trade-off between price and functionality is a\ntime-consuming step in an online shopping experience, and customers can be\noverwhelmed by the available choices. However, determining the set of product\nfeatures that most differentiate a particular product is still an open question\nin online recommender systems. In this paper, we focus on using interpretable\nmachine learning methods to tackle this problem. First, we identify this unique\nproduct feature recommendation problem from a business perspective on a major\nUS e-commerce site. Second, we formulate the problem into a price-driven\nsupervised learning problem to discover the product features that could best\nexplain the price of a product in a given product category. We build machine\nlearning models with a model-agnostic method Shapley Values to understand the\nimportance of each feature, rank and recommend the most essential features.\nThird, we leverage human experts to evaluate its relevancy. The results show\nthat our method is superior to a strong baseline method based on customer\nbehavior and significantly boosts the coverage by 45%. Finally, our proposed\nmethod shows comparable conversion rate against the baseline in online A/B\ntests.', ""Decision makers need economical information to drive their decisions. The\nCompany Actualis SARL is specialized in the production and distribution of a\npress review about French regional economic actors. This economic review\nrepresents for a client a prospecting tool on partners and competitors. To\nreduce the overload of useless information, the company is moving towards a\ncustomized review for each customer. Three issues appear to achieve this goal.\nFirst, how to identify the elements in the text in order to extract objects\nthat match with the recommendation's criteria presented? Second, How to define\nthe structure of these objects, relationships and articles in order to provide\na source of knowledge usable by the extraction process to produce new knowledge\nfrom articles? The latter issue is the feedback on customer experience to\nidentify the quality of distributed information in real-time and to improve the\nrelevance of the recommendations. This paper presents a new type of\nrecommendation based on the semantic description of both articles and user\nprofile."", 'In marketing we are often confronted with a continuous stream of responses to\nmarketing messages. Such streaming data provide invaluable information\nregarding message effectiveness and segmentation. However, streaming data are\nhard to analyze using conventional methods: their high volume and the fact that\nthey are continuously augmented means that it takes considerable time to\nanalyze them. We propose a method for estimating a finite mixture of logistic\nregression models which can be used to cluster customers based on a continuous\nstream of responses. This method, which we coin oFMLR, allows segments to be\nidentified in data streams or extremely large static datasets. Contrary to\nblack box algorithms, oFMLR provides model estimates that are directly\ninterpretable. We first introduce oFMLR, explaining in passing general topics\nsuch as online estimation and the EM algorithm, making this paper a high level\noverview of possible methods of dealing with large data streams in marketing\npractice. Next, we discuss model convergence, identifiability, and relations to\nalternative, Bayesian, methods; we also identify more general issues that arise\nfrom dealing with continuously augmented data sets. Finally, we introduce the\noFMLR [R] package and evaluate the method by numerical simulation and by\nanalyzing a large customer clickstream dataset.', 'In recent years online shopping has gained momentum and became an important\nvenue for customers wishing to save time and simplify their shopping process. A\nkey advantage of shopping online is the ability to read what other customers\nare saying about products of interest. In this work, we aim to maintain this\nadvantage in situations where extreme brevity is needed, for example, when\nshopping by voice. We suggest a novel task of extracting a single\nrepresentative helpful sentence from a set of reviews for a given product. The\nselected sentence should meet two conditions: first, it should be helpful for a\npurchase decision and second, the opinion it expresses should be supported by\nmultiple reviewers. This task is closely related to the task of Multi Document\nSummarization in the product reviews domain but differs in its objective and\nits level of conciseness. We collect a dataset in English of sentence\nhelpfulness scores via crowd-sourcing and demonstrate its reliability despite\nthe inherent subjectivity involved. Next, we describe a complete model that\nextracts representative helpful sentences with positive and negative sentiment\ntowards the product and demonstrate that it outperforms several baselines.', ""In a wide variety of applications, including personalization, we want to\nmeasure the difference in outcome due to an intervention and thus have to deal\nwith counterfactual inference. The feedback from a customer in any of these\nsituations is only 'bandit feedback' - that is, a partial feedback based on\nwhether we chose to intervene or not. Typically randomized experiments are\ncarried out to understand whether an intervention is overall better than no\nintervention. Here we present a feature learning algorithm to learn from a\nrandomized experiment where the intervention in consideration is most effective\nand where it is least effective rather than only focusing on the overall\nimpact, thus adding a context to our learning mechanism and extract more\ninformation. From the randomized experiment, we learn the feature\nrepresentations which divide the population into subpopulations where we\nobserve statistically significant difference in average customer feedback\nbetween those who were subjected to the intervention and those who were not,\nwith a level of significance l, where l is a configurable parameter in our\nmodel. We use this information to derive the value of the intervention in\nconsideration for each instance in the population. With experiments, we show\nthat using this additional learning, in future interventions, the context for\neach instance could be leveraged to decide whether to intervene or not."", 'Distant and weak supervision allow to obtain large amounts of labeled\ntraining data quickly and cheaply, but these automatic annotations tend to\ncontain a high amount of errors. A popular technique to overcome the negative\neffects of these noisy labels is noise modelling where the underlying noise\nprocess is modelled. In this work, we study the quality of these estimated\nnoise models from the theoretical side by deriving the expected error of the\nnoise model. Apart from evaluating the theoretical results on commonly used\nsynthetic noise, we also publish NoisyNER, a new noisy label dataset from the\nNLP domain that was obtained through a realistic distant supervision technique.\nIt provides seven sets of labels with differing noise patterns to evaluate\ndifferent noise levels on the same instances. Parallel, clean labels are\navailable making it possible to study scenarios where a small amount of\ngold-standard data can be leveraged. Our theoretical results and the\ncorresponding experiments give insights into the factors that influence the\nnoise model estimation like the noise distribution and the sampling technique.', 'Teaching large classes remains a great challenge, primarily because it is\ndifficult to attend to all the student needs in a timely manner. Automatic text\nsummarization systems can be leveraged to summarize the student feedback,\nsubmitted immediately after each lecture, but it is left to be discovered what\nmakes a good summary for student responses. In this work we explore a new\nmethodology that effectively extracts summary phrases from the student\nresponses. Each phrase is tagged with the number of students who raise the\nissue. The phrases are evaluated along two dimensions: with respect to text\ncontent, they should be informative and well-formed, measured by the ROUGE\nmetric; additionally, they shall attend to the most pressing student needs,\nmeasured by a newly proposed metric. This work is enabled by a phrase-based\nannotation and highlighting scheme, which is new to the summarization task. The\nphrase-based framework allows us to summarize the student responses into a set\nof bullet points and present to the instructor promptly.', 'Topic models are in widespread use in natural language processing and beyond.\nHere, we propose a new framework for the evaluation of probabilistic topic\nmodeling algorithms based on synthetic corpora containing an unambiguously\ndefined ground truth topic structure. The major innovation of our approach is\nthe ability to quantify the agreement between the planted and inferred topic\nstructures by comparing the assigned topic labels at the level of the tokens.\nIn experiments, our approach yields novel insights about the relative strengths\nof topic models as corpus characteristics vary, and the first evidence of an\n""undetectable phase"" for topic models when the planted structure is weak. We\nalso establish the practical relevance of the insights gained for synthetic\ncorpora by predicting the performance of topic modeling algorithms in\nclassification tasks in real-world corpora.', ""Mixed-initiative systems allow users to interactively provide feedback to\npotentially improve system performance. Human feedback can correct model errors\nand update model parameters to dynamically adapt to changing data.\nAdditionally, many users desire the ability to have a greater level of control\nand fix perceived flaws in systems they rely on. However, how the ability to\nprovide feedback to autonomous systems influences user trust is a largely\nunexplored area of research. Our research investigates how the act of providing\nfeedback can affect user understanding of an intelligent system and its\naccuracy. We present a controlled experiment using a simulated object detection\nsystem with image data to study the effects of interactive feedback collection\non user impressions. The results show that providing human-in-the-loop feedback\nlowered both participants' trust in the system and their perception of system\naccuracy, regardless of whether the system accuracy improved in response to\ntheir feedback. These results highlight the importance of considering the\neffects of allowing end-user feedback on user trust when designing intelligent\nsystems."", ""Home appliance manufacturers strive to obtain feedback from users to improve\ntheir products and services to build a smart home system. To help manufacturers\ndevelop a smart home system, we design a federated learning (FL) system\nleveraging the reputation mechanism to assist home appliance manufacturers to\ntrain a machine learning model based on customers' data. Then, manufacturers\ncan predict customers' requirements and consumption behaviors in the future.\nThe working flow of the system includes two stages: in the first stage,\ncustomers train the initial model provided by the manufacturer using both the\nmobile phone and the mobile edge computing (MEC) server. Customers collect data\nfrom various home appliances using phones, and then they download and train the\ninitial model with their local data. After deriving local models, customers\nsign on their models and send them to the blockchain. In case customers or\nmanufacturers are malicious, we use the blockchain to replace the centralized\naggregator in the traditional FL system. Since records on the blockchain are\nuntampered, malicious customers or manufacturers' activities are traceable. In\nthe second stage, manufacturers select customers or organizations as miners for\ncalculating the averaged model using received models from customers. By the end\nof the crowdsourcing task, one of the miners, who is selected as the temporary\nleader, uploads the model to the blockchain. To protect customers' privacy and\nimprove the test accuracy, we enforce differential privacy on the extracted\nfeatures and propose a new normalization technique. We experimentally\ndemonstrate that our normalization technique outperforms batch normalization\nwhen features are under differential privacy protection. In addition, to\nattract more customers to participate in the crowdsourcing FL task, we design\nan incentive mechanism to award participants."", ""Research has been devoted in the past few years to relevance feedback as an\neffective solution to improve performance of information retrieval systems.\nRelevance feedback refers to an interactive process that helps to improve the\nretrieval performance. In this paper we propose the use of relevance feedback\nto improve document image retrieval System (DIRS) performance. This paper\ncompares a variety of strategies for positive and negative feedback. In\naddition, feature subspace is extracted and updated during the feedback process\nusing a Principal Component Analysis (PCA) technique and based on user's\nfeedback. That is, in addition to reducing the dimensionality of feature\nspaces, a proper subspace for each type of features is obtained in the feedback\nprocess to further improve the retrieval accuracy. Experiments show that using\nrelevance Feedback in DIR achieves better performance than common DIR."", ""As financial services (FS) companies have experienced drastic technology\ndriven changes, the availability of new data streams provides the opportunity\nfor more comprehensive customer understanding. We propose Dynamic Customer\nEmbeddings (DCE), a framework that leverages customers' digital activity and a\nwide range of financial context to learn dense representations of customers in\nthe FS industry. Our method examines customer actions and pageviews within a\nmobile or web digital session, the sequencing of the sessions themselves, and\nsnapshots of common financial features across our organization at the time of\nlogin. We test our customer embeddings using real world data in three\nprediction problems: 1) the intent of a customer in their next digital session,\n2) the probability of a customer calling the call centers after a session, and\n3) the probability of a digital session to be fraudulent. DCE showed\nperformance lift in all three downstream problems."", 'Obtaining large annotated datasets is critical for training successful\nmachine learning models and it is often a bottleneck in practice. Weak\nsupervision offers a promising alternative for producing labeled datasets\nwithout ground truth annotations by generating probabilistic labels using\nmultiple noisy heuristics. This process can scale to large datasets and has\ndemonstrated state of the art performance in diverse domains such as healthcare\nand e-commerce. One practical issue with learning from user-generated\nheuristics is that their creation requires creativity, foresight, and domain\nexpertise from those who hand-craft them, a process which can be tedious and\nsubjective. We develop the first framework for interactive weak supervision in\nwhich a method proposes heuristics and learns from user feedback given on each\nproposed heuristic. Our experiments demonstrate that only a small number of\nfeedback iterations are needed to train models that achieve highly competitive\ntest set performance without access to ground truth training labels. We conduct\nuser studies, which show that users are able to effectively provide feedback on\nheuristics and that test set results track the performance of simulated\noracles.', ""A promising approach to improve the robustness and exploration in\nReinforcement Learning is collecting human feedback and that way incorporating\nprior knowledge of the target environment. It is, however, often too expensive\nto obtain enough feedback of good quality. To mitigate the issue, we aim to\nrely on a group of multiple experts (and non-experts) with different skill\nlevels to generate enough feedback. Such feedback can therefore be inconsistent\nand infrequent. In this paper, we build upon prior work -- Advise, a Bayesian\napproach attempting to maximise the information gained from human feedback --\nextending the algorithm to accept feedback from this larger group of humans,\nthe trainers, while also estimating each trainer's reliability. We show how\naggregating feedback from multiple trainers improves the total feedback's\naccuracy and make the collection process easier in two ways. Firstly, this\napproach addresses the case of some of the trainers being adversarial.\nSecondly, having access to the information about each trainer reliability\nprovides a second layer of robustness and offers valuable information for\npeople managing the whole system to improve the overall trust in the system. It\noffers an actionable tool for improving the feedback collection process or\nmodifying the reward function design if needed. We empirically show that our\napproach can accurately learn the reliability of each trainer correctly and use\nit to maximise the information gained from the multiple trainers' feedback,\neven if some of the sources are adversarial."", ""Telecommunications operators (telcos) traditional sources of income, voice\nand SMS, are shrinking due to customers using over-the-top (OTT) applications\nsuch as WhatsApp or Viber. In this challenging environment it is critical for\ntelcos to maintain or grow their market share, by providing users with as good\nan experience as possible on their network.\n  But the task of extracting customer insights from the vast amounts of data\ncollected by telcos is growing in complexity and scale everey day. How can we\nmeasure and predict the quality of a user's experience on a telco network in\nreal-time? That is the problem that we address in this paper.\n  We present an approach to capture, in (near) real-time, the mobile customer\nexperience in order to assess which conditions lead the user to place a call to\na telco's customer care center. To this end, we follow a supervised learning\napproach for prediction and train our 'Restricted Random Forest' model using,\nas a proxy for bad experience, the observed customer transactions in the telco\ndata feed before the user places a call to a customer care center.\n  We evaluate our approach using a rich dataset provided by a major African\ntelecommunication's company and a novel big data architecture for both the\ntraining and scoring of predictive models. Our empirical study shows our\nsolution to be effective at predicting user experience by inferring if a\ncustomer will place a call based on his current context.\n  These promising results open new possibilities for improved customer service,\nwhich will help telcos to reduce churn rates and improve customer experience,\nboth factors that directly impact their revenue growth."", ""Data analytics software applications have become an integral part of the\ndecision-making process of analysts. Users of such a software face challenges\ndue to insufficient product and domain knowledge, and find themselves in need\nof help. To alleviate this, we propose a task-aware command recommendation\nsystem, to guide the user on what commands could be executed next. We rely on\ntopic modeling techniques to incorporate information about user's task into our\nmodels. We also present a help prediction model to detect if a user is in need\nof help, in which case the system proactively provides the aforementioned\ncommand recommendations. We leverage the log data of a web-based analytics\nsoftware to quantify the superior performance of our neural models, in\ncomparison to competitive baselines."", ""Motivated by online settings where users can provide explicit feedback about\nthe relevance of products that are sequentially presented to them, we look at\nthe recommendation process as a problem of dynamically optimizing this\nrelevance feedback. Such an algorithm optimizes the fine tradeoff between\npresenting the products that are most likely to be relevant, and learning the\npreferences of the user so that more relevant recommendations can be made in\nthe future.\n  We assume a standard predictive model inspired by collaborative filtering, in\nwhich a user is sampled from a distribution over a set of possible types. For\nevery product category, each type has an associated relevance feedback that is\nassumed to be binary: the category is either relevant or irrelevant. Assuming\nthat the user stays for each additional recommendation opportunity with\nprobability $\\beta$ independent of the past, the problem is to find a policy\nthat maximizes the expected number of recommendations that are deemed relevant\nin a session.\n  We analyze this problem and prove key structural properties of the optimal\npolicy. Based on these properties, we first present an algorithm that strikes a\nbalance between recursion and dynamic programming to compute this policy. We\nfurther propose and analyze two heuristic policies: a `farsighted' greedy\npolicy that attains at least $1-\\beta$ factor of the optimal payoff, and a\nnaive greedy policy that attains at least $\\frac{1-\\beta}{1+\\beta}$ factor of\nthe optimal payoff in the worst case. Extensive simulations show that these\nheuristics are very close to optimal in practice."", ""Recently, topic modeling has been widely used to discover the abstract topics\nin text corpora. Most of the existing topic models are based on the assumption\nof three-layer hierarchical Bayesian structure, i.e. each document is modeled\nas a probability distribution over topics, and each topic is a probability\ndistribution over words. However, the assumption is not optimal. Intuitively,\nit's more reasonable to assume that each topic is a probability distribution\nover concepts, and then each concept is a probability distribution over words,\ni.e. adding a latent concept layer between topic layer and word layer in\ntraditional three-layer assumption. In this paper, we verify the proposed\nassumption by incorporating the new assumption in two representative topic\nmodels, and obtain two novel topic models. Extensive experiments were conducted\namong the proposed models and corresponding baselines, and the results show\nthat the proposed models significantly outperform the baselines in terms of\ncase study and perplexity, which means the new assumption is more reasonable\nthan traditional one."", 'Existing document filtering systems learn user profiles based on user\nrelevance feedback on documents. In some cases, users may have prior knowledge\nabout what features are important. For example, a Spanish speaker may only want\nnews written in Spanish, and thus a relevant document should contain the\nfeature ""Language: Spanish""; a researcher focusing on HIV knows an article with\nthe medical subject ""Subject: AIDS"" is very likely to be relevant to him/her.\n  Semi-structured documents with rich metadata are increasingly prevalent on\nthe Internet. Motivated by the well-adopted faceted search interface in\ne-commerce, we study the exploitation of user prior knowledge on faceted\nfeatures for semi-structured document filtering. We envision two faceted\nfeedback mechanisms, and propose a novel user profile learning algorithm that\ncan incorporate user feedback on features. To evaluate the proposed work, we\nuse two data sets from the TREC filtering track, and conduct a user study on\nAmazon Mechanical Turk. Our experiment results show that user feedback on\nfaceted features is useful for filtering. The proposed user profile learning\nalgorithm can effectively learn from user feedback on both documents and\nfeatures, and performs better than several existing methods.', ""Manually labeled corpora are expensive to create and often not available for\nlow-resource languages or domains. Automatic labeling approaches are an\nalternative way to obtain labeled data in a quicker and cheaper way. However,\nthese labels often contain more errors which can deteriorate a classifier's\nperformance when trained on this data. We propose a noise layer that is added\nto a neural network architecture. This allows modeling the noise and train on a\ncombination of clean and noisy data. We show that in a low-resource NER task we\ncan improve performance by up to 35% by using additional, noisy data and\nhandling the noise."", 'Convincing a customer is always considered as a challenging task in every\nbusiness. But when it comes to online business, this task becomes even more\ndifficult. Online retailers try everything possible to gain the trust of the\ncustomer. One of the solutions is to provide an area for existing users to\nleave their comments. This service can effectively develop the trust of the\ncustomer however normally the customer comments about the product in their\nnative language using Roman script. If there are hundreds of comments this\nmakes difficulty even for the native customers to make a buying decision. This\nresearch proposes a system which extracts the comments posted in Roman Urdu,\ntranslate them, find their polarity and then gives us the rating of the\nproduct. This rating will help the native and non-native customers to make\nbuying decision efficiently from the comments posted in Roman Urdu.', 'Alternative recommender systems are critical for ecommerce companies. They\nguide customers to explore a massive product catalog and assist customers to\nfind the right products among an overwhelming number of options. However, it is\na non-trivial task to recommend alternative products that fit customer needs.\nIn this paper, we use both textual product information (e.g. product titles and\ndescriptions) and customer behavior data to recommend alternative products. Our\nresults show that the coverage of alternative products is significantly\nimproved in offline evaluations as well as recall and precision. The final A/B\ntest shows that our algorithm increases the conversion rate by 12 percent in a\nstatistically significant way. In order to better capture the semantic meaning\nof product information, we build a Siamese Network with Bidirectional LSTM to\nlearn product embeddings. In order to learn a similarity space that better\nmatches the preference of real customers, we use co-compared data from\nhistorical customer behavior as labels to train the network. In addition, we\nuse NMSLIB to accelerate the computationally expensive kNN computation for\nmillions of products so that the alternative recommendation is able to scale\nacross the entire catalog of a major ecommerce site.', 'Learning with noisy labels is a common challenge in supervised learning.\nExisting approaches often require practitioners to specify noise rates, i.e., a\nset of parameters controlling the severity of label noises in the problem, and\nthe specifications are either assumed to be given or estimated using additional\nsteps. In this work, we introduce a new family of loss functions that we name\nas peer loss functions, which enables learning from noisy labels and does not\nrequire a priori specification of the noise rates. Peer loss functions work\nwithin the standard empirical risk minimization (ERM) framework. We show that,\nunder mild conditions, performing ERM with peer loss functions on the noisy\ndataset leads to the optimal or a near-optimal classifier as if performing ERM\nover the clean training data, which we do not have access to. We pair our\nresults with an extensive set of experiments. Peer loss provides a way to\nsimplify model development when facing potentially noisy training labels, and\ncan be promoted as a robust candidate loss function in such situations.', 'Recently there has been significant activity in developing algorithms with\nprovable guarantees for topic modeling. In standard topic models, a topic (such\nas sports, business, or politics) is viewed as a probability distribution $\\vec\na_i$ over words, and a document is generated by first selecting a mixture $\\vec\nw$ over topics, and then generating words i.i.d. from the associated mixture\n$A{\\vec w}$. Given a large collection of such documents, the goal is to recover\nthe topic vectors and then to correctly classify new documents according to\ntheir topic mixture.\n  In this work we consider a broad generalization of this framework in which\nwords are no longer assumed to be drawn i.i.d. and instead a topic is a complex\ndistribution over sequences of paragraphs. Since one could not hope to even\nrepresent such a distribution in general (even if paragraphs are given using\nsome natural feature representation), we aim instead to directly learn a\ndocument classifier. That is, we aim to learn a predictor that given a new\ndocument, accurately predicts its topic mixture, without learning the\ndistributions explicitly. We present several natural conditions under which one\ncan do this efficiently and discuss issues such as noise tolerance and sample\ncomplexity in this model. More generally, our model can be viewed as a\ngeneralization of the multi-view or co-training setting in machine learning.', ""Capturing users' engagement is crucial for gathering feedback about the\nfeatures of a software product. In a market-driven context, current approaches\nto collect and analyze users' feedback are based on techniques leveraging\ninformation extracted from product reviews and social media. These approaches\nare hardly applicable in bespoke software development, or in contexts in which\none needs to gather information from specific users. In such cases, companies\nneed to resort to face-to-face interviews to get feedback on their products. In\nthis paper, we propose to utilize biometric data, in terms of physiological and\nvoice features, to complement interviews with information about the engagement\nof the user on the discussed product-relevant topics. We evaluate our approach\nby interviewing users while gathering their physiological data (i.e.,\nbiofeedback) using an Empatica E4 wristband, and capturing their voice through\nthe default audio-recorder of a common laptop. Our results show that we can\npredict users' engagement by training supervised machine learning algorithms on\nbiometric data (F1=0.72), and that voice features alone are sufficiently\neffective (F1=0.71). Our work contributes with one the first studies in\nrequirements engineering in which biometrics are used to identify emotions.\nThis is also the first study in software engineering that considers voice\nanalysis. The usage of voice features could be particularly helpful for\nemotion-aware requirements elicitation in remote communication, either\nperformed by human analysts or voice-based chatbots, and can also be exploited\nto support the analysis of meetings in software engineering research."", 'We present and study models of adversarial online learning where the feedback\nobserved by the learner is noisy, and the feedback is either full information\nfeedback or bandit feedback. Specifically, we consider binary losses xored with\nthe noise, which is a Bernoulli random variable. We consider both a constant\nnoise rate and a variable noise rate. Our main results are tight regret bounds\nfor learning with noise in the adversarial online learning model.', 'Users giving relevance feedback in exploratory search are often uncertain\nabout the correctness of their feedback, which may result in noisy or even\nerroneous feedback. Additionally, the search intent of the user may be volatile\nas the user is constantly learning and reformulating her search hypotheses\nduring the search. This may lead to a noticeable concept drift in the feedback.\nWe formulate a Bayesian regression model for predicting the accuracy of each\nindividual user feedback and thus find outliers in the feedback data set.\nAdditionally, we introduce a timeline interface that visualizes the feedback\nhistory to the user and gives her suggestions on which past feedback is likely\nin need of adjustment. This interface also allows the user to adjust the\nfeedback accuracy inferences made by the model. Simulation experiments\ndemonstrate that the performance of the new user model outperforms a simpler\nbaseline and that the performance approaches that of an oracle, given a small\namount of additional user interaction. A user study shows that the proposed\nmodelling technique, combined with the timeline interface, makes it easier for\nthe users to notice and correct mistakes in their feedback, and to discover new\nitems.', 'While buying a product from the e-commerce websites, customers generally have\na plethora of questions. From the perspective of both the e-commerce service\nprovider as well as the customers, there must be an effective question\nanswering system to provide immediate answers to the user queries. While\ncertain questions can only be answered after using the product, there are many\nquestions which can be answered from the product specification itself. Our work\ntakes a first step in this direction by finding out the relevant product\nspecifications, that can help answering the user questions. We propose an\napproach to automatically create a training dataset for this problem. We\nutilize recently proposed XLNet and BERT architectures for this problem and\nfind that they provide much better performance than the Siamese model,\npreviously applied for this problem. Our model gives a good performance even\nwhen trained on one vertical and tested across different verticals.', 'E-commerce websites use machine learned ranking models to serve shopping\nresults to customers. Typically, the websites log the customer search events,\nwhich include the query entered and the resulting engagement with the shopping\nresults, such as clicks and purchases. Each customer search event serves as\ninput training data for the models, and the individual customer engagement\nserves as a signal for customer preference. So a purchased shopping result, for\nexample, is perceived to be more important than one that is not. However, new\nor under-impressed products do not have enough customer engagement signals and\nend up at a disadvantage when being ranked alongside popular products. In this\npaper, we propose a novel method for data curation that aggregates all customer\nengagements within a day for the same query to use as input training data. This\naggregated customer engagement gives the models a complete picture of the\nrelative importance of shopping results. Training models on this aggregated\ndata leads to less reliance on behavioral features. This helps mitigate the\ncold start problem and boosted relevant new products to top search results. In\nthis paper, we present the offline and online analysis and results comparing\nthe individual and aggregated customer engagement models trained on e-commerce\ndata.', ""This paper addresses the problem of multiclass classification with corrupted\nor noisy bandit feedback. In this setting, the learner may not receive true\nfeedback. Instead, it receives feedback that has been flipped with some\nnon-zero probability. We propose a novel approach to deal with noisy bandit\nfeedback based on the unbiased estimator technique. We further offer a method\nthat can efficiently estimate the noise rates, thus providing an end-to-end\nframework. The proposed algorithm enjoys a mistake bound of the order of\n$O(\\sqrt{T})$ in the high noise case and of the order of\n$O(T^{\\nicefrac{2}{3}})$ in the worst case. We show our approach's\neffectiveness using extensive experiments on several benchmark datasets."", ""Recommendation is the task of improving customer experience through\npersonalized recommendation based on users' past feedback. In this paper, we\ninvestigate the most common scenario: the user-item (U-I) matrix of implicit\nfeedback. Even though many recommendation approaches are designed based on\nimplicit feedback, they attempt to project the U-I matrix into a low-rank\nlatent space, which is a strict restriction that rarely holds in practice. In\naddition, although misclassification costs from imbalanced classes are\nsignificantly different, few methods take the cost of classification error into\naccount. To address aforementioned issues, we propose a robust framework by\ndecomposing the U-I matrix into two components: (1) a low-rank matrix that\ncaptures the common preference, and (2) a sparse matrix that detects the\nuser-specific preference of individuals. A cost-sensitive learning model is\nembedded into the framework. Specifically, this model exploits different costs\nin the loss function for the observed and unobserved instances. We show that\nthe resulting non-smooth convex objective can be optimized efficiently by an\naccelerated projected gradient method with closed-form solutions. Morever, the\nproposed algorithm can be scaled up to large-sized datasets after a relaxation.\nThe theoretical result shows that even with a small fraction of 1's in the U-I\nmatrix $M\\in\\mathbb{R}^{n\\times m}$, the cost-sensitive error of the proposed\nmodel is upper bounded by $O(\\frac{\\alpha}{\\sqrt{mn}})$, where $\\alpha$ is a\nbias over imbalanced classes. Finally, empirical experiments are extensively\ncarried out to evaluate the effectiveness of our proposed algorithm.\nEncouraging experimental results show that our algorithm outperforms several\nstate-of-the-art algorithms on benchmark recommendation datasets."", 'We address online learning in complex auction settings, such as sponsored\nsearch auctions, where the value of the bidder is unknown to her, evolving in\nan arbitrary manner and observed only if the bidder wins an allocation. We\nleverage the structure of the utility of the bidder and the partial feedback\nthat bidders typically receive in auctions, in order to provide algorithms with\nregret rates against the best fixed bid in hindsight, that are exponentially\nfaster in convergence in terms of dependence on the action space, than what\nwould have been derived by applying a generic bandit algorithm and almost\nequivalent to what would have been achieved in the full information setting.\nOur results are enabled by analyzing a new online learning setting with\noutcome-based feedback, which generalizes learning with feedback graphs. We\nprovide an online learning algorithm for this setting, of independent interest,\nwith regret that grows only logarithmically with the number of actions and\nlinearly only in the number of potential outcomes (the latter being very small\nin most auction settings). Last but not least, we show that our algorithm\noutperforms the bandit approach experimentally and that this performance is\nrobust to dropping some of our theoretical assumptions or introducing noise in\nthe feedback that the bidder receives.', ""Intelligent assistants change the way people interact with computers and make\nit possible for people to search for products through conversations when they\nhave purchase needs. During the interactions, the system could ask questions on\ncertain aspects of the ideal products to clarify the users' needs. For example,\nprevious work proposed to ask users the exact characteristics of their ideal\nitems before showing results. However, users may not have clear ideas about\nwhat an ideal item looks like, especially when they have not seen any item. So\nit is more feasible to facilitate the conversational search by showing example\nitems and asking for feedback instead. In addition, when the users provide\nnegative feedback for the presented items, it is easier to collect their\ndetailed feedback on certain properties (aspect-value pairs) of the\nnon-relevant items. By breaking down the item-level negative feedback to\nfine-grained feedback on aspect-value pairs, more information is available to\nhelp clarify users' intents. So in this paper, we propose a conversational\nparadigm for product search driven by non-relevant items, based on which\nfine-grained feedback is collected and utilized to show better results in the\nnext iteration. We then propose an aspect-value likelihood model to incorporate\nboth positive and negative feedback on fine-grained aspect-value pairs of the\nnon-relevant items. Experimental results show that our model is significantly\nbetter than state-of-the-art product search baselines without using feedback\nand those baselines using item-level negative feedback."", ""Product search serves as an important entry point for online shopping. In\ncontrast to web search, the retrieved results in product search not only need\nto be relevant but also should satisfy customers' preferences in order to\nelicit purchases. Previous work has shown the efficacy of purchase history in\npersonalized product search. However, customers with little or no purchase\nhistory do not benefit from personalized product search. Furthermore,\npreferences extracted from a customer's purchase history are usually long-term\nand may not always align with her short-term interests. Hence, in this paper,\nwe leverage clicks within a query session, as implicit feedback, to represent\nusers' hidden intents, which further act as the basis for re-ranking subsequent\nresult pages for the query. It has been studied extensively to model user\npreference with implicit feedback in recommendation tasks. However, there has\nbeen little research on modeling users' short-term interest in product search.\nWe study whether short-term context could help promote users' ideal item in the\nfollowing result pages for a query. Furthermore, we propose an end-to-end\ncontext-aware embedding model which can capture long-term and short-term\ncontext dependencies. Our experimental results on the datasets collected from\nthe search log of a commercial product search engine show that short-term\ncontext leads to much better performance compared with long-term and no\ncontext. Our results also show that our proposed model is more effective than\nword-based context-aware models."", 'Non-technical losses (NTL) occur during the distribution of electricity in\npower grids and include, but are not limited to, electricity theft and faulty\nmeters. In emerging countries, they may range up to 40% of the total\nelectricity distributed. In order to detect NTLs, machine learning methods are\nused that learn irregular consumption patterns from customer data and\ninspection results. The Big Data paradigm followed in modern machine learning\nreflects the desire of deriving better conclusions from simply analyzing more\ndata, without the necessity of looking at theory and models. However, the\nsample of inspected customers may be biased, i.e. it does not represent the\npopulation of all customers. As a consequence, machine learning models trained\non these inspection results are biased as well and therefore lead to unreliable\npredictions of whether customers cause NTL or not. In machine learning, this\nissue is called covariate shift and has not been addressed in the literature on\nNTL detection yet. In this work, we present a novel framework for quantifying\nand visualizing covariate shift. We apply it to a commercial data set from\nBrazil that consists of 3.6M customers and 820K inspection results. We show\nthat some features have a stronger covariate shift than others, making\npredictions less reliable. In particular, previous inspections were focused on\ncertain neighborhoods or customer classes and that they were not sufficiently\nspread among the population of customers. This framework is about to be\ndeployed in a commercial product for NTL detection.', 'For a successful business, engaging in an effective campaign is a key task\nfor marketers. Most previous studies used various mathematical models to\nsegment customers without considering the correlation between customer\nsegmentation and a campaign. This work presents a conceptual model by studying\nthe significant campaign-dependent variables of customer targeting in customer\nsegmentation context. In this way, the processes of customer segmentation and\ntargeting thus can be linked and solved together. The outcomes of customer\nsegmentation of this study could be more meaningful and relevant for marketers.\nThis investigation applies a customer life time value (LTV) model to assess the\nfitness between targeted customer groups and marketing strategies. To integrate\ncustomer segmentation and customer targeting, this work uses the genetic\nalgorithm (GA) to determine the optimized marketing strategy. Later, we suggest\nusing C&RT (Classification and Regression Tree) in SPSS PASW Modeler as the\nreplacement to Genetic Algorithm technique to accomplish these results. We also\nsuggest using LOSSYCOUNTING and Counting Bloom Filter to dynamically design the\nright and up-to-date offer to the right customer.', 'This research proposes a new (old) metric for evaluating goodness of fit in\ntopic models, the coefficient of determination, or $R^2$. Within the context of\ntopic modeling, $R^2$ has the same interpretation that it does when used in a\nbroader class of statistical models. Reporting $R^2$ with topic models\naddresses two current problems in topic modeling: a lack of standard\ncross-contextual evaluation metrics for topic modeling and ease of\ncommunication with lay audiences. The author proposes that $R^2$ should be\nreported as a standard metric when constructing topic models.', ""A key challenge in Imitation Learning (IL) is that optimal state actions\ndemonstrations are difficult for the teacher to provide. For example in\nrobotics, providing kinesthetic demonstrations on a robotic manipulator\nrequires the teacher to control multiple degrees of freedom at once. The\ndifficulty of requiring optimal state action demonstrations limits the space of\nproblems where the teacher can provide quality feedback. As an alternative to\nstate action demonstrations, the teacher can provide corrective feedback such\nas their preferences or rewards. Prior work has created algorithms designed to\nlearn from specific types of noisy feedback, but across teachers and tasks\ndifferent forms of feedback may be required. Instead we propose that in order\nto learn from a diversity of scenarios we need to learn from a variety of\nfeedback. To learn from a variety of feedback we make the following insight:\nthe teacher's cost function is latent and we can model a stream of feedback as\na stream of loss functions. We then use any online learning algorithm to\nminimize the sum of these losses. With this insight we can learn from a\ndiversity of feedback that is weakly correlated with the teacher's true cost\nfunction. We unify prior work into a general corrective feedback meta-algorithm\nand show that regardless of feedback we can obtain the same regret bounds. We\ndemonstrate our approach by learning to perform a household navigation task on\na robotic racecar platform. Our results show that our approach can learn\nquickly from a variety of noisy feedback."", 'The main idea of this paper is to represent shopping items through vectors\nbecause these vectors act as the base for building em- beddings for customers\nand shopping carts. Also, these vectors are input to the mathematical models\nthat act as either a recommendation engine or help in targeting potential\ncustomers. We have used exponential family embeddings as the tool to construct\ntwo basic vectors - product embeddings and context vectors. Using the basic\nvectors, we build combined embeddings, trip embeddings and customer embeddings.\nCombined embeddings mix linguistic properties of product names with their\nshopping patterns. The customer embeddings establish an understand- ing of the\nbuying pattern of customers in a group and help in building customer profile.\nFor example a customer profile can represent customers frequently buying\npet-food. Identifying such profiles can help us bring out offers and discounts.\nSimilarly, trip embeddings are used to build trip profiles. People happen to\nbuy similar set of products in a trip and hence their trip embeddings can be\nused to predict the next product they would like to buy. This is a novel\ntechnique and the first of its kind to make recommendation using product, trip\nand customer embeddings.', 'Encoder-decoder based neural architectures serve as the basis of\nstate-of-the-art approaches in end-to-end open domain dialog systems. Since\nmost of such systems are trained with a maximum likelihood~(MLE) objective they\nsuffer from issues such as lack of generalizability and the generic response\nproblem, i.e., a system response that can be an answer to a large number of\nuser utterances, e.g., ""Maybe, I don\'t know."" Having explicit feedback on the\nrelevance and interestingness of a system response at each turn can be a useful\nsignal for mitigating such issues and improving system quality by selecting\nresponses from different approaches. Towards this goal, we present a system\nthat evaluates chatbot responses at each dialog turn for coherence and\nengagement. Our system provides explicit turn-level dialog quality feedback,\nwhich we show to be highly correlated with human evaluation. To show that\nincorporating this feedback in the neural response generation models improves\ndialog quality, we present two different and complementary mechanisms to\nincorporate explicit feedback into a neural response generation model:\nreranking and direct modification of the loss function during training. Our\nstudies show that a response generation model that incorporates these combined\nfeedback mechanisms produce more engaging and coherent responses in an\nopen-domain spoken dialog setting, significantly improving the response quality\nusing both automatic and human evaluation.', 'Realistic free-space optical communications systems suffer from turbulent\npropagation of light through the atmosphere and detector noise at the receiver,\nwhich can significantly degrade the optical mode quality of the received state,\nincrease cross-talk between modes, and correspondingly increase the symbol\nerror ratio (SER) of the system. In order to overcome these obstacles, we\ndevelop a state-of-the-art generative machine learning (GML) and convolutional\nneural network (CNN) system in combination, and demonstrate its efficacy in a\nfree-space optical (FSO) communications setting. The system corrects for the\ndistortion effects due to turbulence and reduces detector noise, resulting in\nsignificantly lowered SERs and cross-talk at the output of the receiver, while\nrequiring no feedback. This scheme is straightforward to scale, and may provide\na concrete and cost effective technique to establishing long range classical\nand quantum communication links in the near future.', 'In black-box optimization problems, we aim to maximize an unknown objective\nfunction, where the function is only accessible through feedbacks of an\nevaluation or simulation oracle. In real-life, the feedbacks of such oracles\nare often noisy and available after some unknown delay that may depend on the\ncomputation time of the oracle. Additionally, if the exact evaluations are\nexpensive but coarse approximations are available at a lower cost, the\nfeedbacks can have multi-fidelity. In order to address this problem, we propose\na generic extension of hierarchical optimistic tree search (HOO), called\nProCrastinated Tree Search (PCTS), that flexibly accommodates a delay and\nnoise-tolerant bandit algorithm. We provide a generic proof technique to\nquantify regret of PCTS under delayed, noisy, and multi-fidelity feedbacks.\nSpecifically, we derive regret bounds of PCTS enabled with delayed-UCB1 (DUCB1)\nand delayed-UCB-V (DUCBV) algorithms. Given a horizon $T$, PCTS retains the\nregret bound of non-delayed HOO for expected delay of $O(\\log T)$ and worsens\nby $O(T^{\\frac{1-\\alpha}{d+2}})$ for expected delays of $O(T^{1-\\alpha})$ for\n$\\alpha \\in (0,1]$. We experimentally validate on multiple synthetic functions\nand hyperparameter tuning problems that PCTS outperforms the state-of-the-art\nblack-box optimization methods for feedbacks with different noise levels,\ndelays, and fidelity.', 'Modeling and synthesizing image noise is an important aspect in many computer\nvision applications. The long-standing additive white Gaussian and\nheteroscedastic (signal-dependent) noise models widely used in the literature\nprovide only a coarse approximation of real sensor noise. This paper introduces\nNoise Flow, a powerful and accurate noise model based on recent normalizing\nflow architectures. Noise Flow combines well-established basic parametric noise\nmodels (e.g., signal-dependent noise) with the flexibility and expressiveness\nof normalizing flow networks. The result is a single, comprehensive, compact\nnoise model containing fewer than 2500 parameters yet able to represent\nmultiple cameras and gain factors. Noise Flow dramatically outperforms existing\nnoise models, with 0.42 nats/pixel improvement over the camera-calibrated noise\nlevel functions, which translates to 52% improvement in the likelihood of\nsampled noise. Noise Flow represents the first serious attempt to go beyond\nsimple parametric models to one that leverages the power of deep learning and\ndata-driven noise distributions.', 'Human feedback is widely used to train agents in many domains. However,\nprevious works rarely consider the uncertainty when humans provide feedback,\nespecially in cases that the optimal actions are not obvious to the trainers.\nFor example, the reward of a sub-optimal action can be stochastic and sometimes\nexceeds that of the optimal action, which is common in games or real-world.\nTrainers are likely to provide positive feedback to sub-optimal actions,\nnegative feedback to the optimal actions and even do not provide feedback in\nsome confusing situations. Existing works, which utilize the Expectation\nMaximization (EM) algorithm and treat the feedback model as hidden parameters,\ndo not consider uncertainties in the learning environment and human feedback.\nTo address this challenge, we introduce a novel feedback model that considers\nthe uncertainty of human feedback. However, this incurs intractable calculus in\nthe EM algorithm. To this end, we propose a novel approximate EM algorithm, in\nwhich we approximate the expectation step with the Gradient Descent method.\nExperimental results in both synthetic scenarios and two real-world scenarios\nwith human participants demonstrate the superior performance of our proposed\napproach.', ""Contextual dynamic pricing aims to set personalized prices based on\nsequential interactions with customers. At each time period, a customer who is\ninterested in purchasing a product comes to the platform. The customer's\nvaluation for the product is a linear function of contexts, including product\nand customer features, plus some random market noise. The seller does not\nobserve the customer's true valuation, but instead needs to learn the valuation\nby leveraging contextual information and historical binary purchase feedbacks.\nExisting models typically assume full or partial knowledge of the random noise\ndistribution. In this paper, we consider contextual dynamic pricing with\nunknown random noise in the valuation model. Our distribution-free pricing\npolicy learns both the contextual function and the market noise simultaneously.\nA key ingredient of our method is a novel perturbed linear bandit framework,\nwhere a modified linear upper confidence bound algorithm is proposed to balance\nthe exploration of market noise and the exploitation of the current knowledge\nfor better pricing. We establish the regret upper bound and a matching lower\nbound of our policy in the perturbed linear bandit framework and prove a\nsub-linear regret bound in the considered pricing problem. Finally, we\ndemonstrate the superior performance of our policy on simulations and a\nreal-life auto-loan dataset."", 'Query categorization is an essential part of query intent understanding in\ne-commerce search. A common query categorization task is to select the relevant\nfine-grained product categories in a product taxonomy. For frequent queries,\nrich customer behavior (e.g., click-through data) can be used to infer the\nrelevant product categories. However, for more rare queries, which cover a\nlarge volume of search traffic, relying solely on customer behavior may not\nsuffice due to the lack of this signal. To improve categorization of rare\nqueries, we adapt the Pseudo-Relevance Feedback (PRF) approach to utilize the\nlatent knowledge embedded in semantically or lexically similar product\ndocuments to enrich the representation of the more rare queries. To this end,\nwe propose a novel deep neural model named Attentive Pseudo Relevance Feedback\nNetwork (APRF-Net) to enhance the representation of rare queries for query\ncategorization. To demonstrate the effectiveness of our approach, we collect\nsearch queries from a large commercial search engine, and compare APRF-Net to\nstate-of-the-art deep learning models for text classification. Our results show\nthat the APRF-Net significantly improves query categorization by 5.9% on F1@1\nscore over the baselines, which increases to 8.2% improvement for the rare\n(tail) queries. The findings of this paper can be leveraged for further\nimprovements in search query representation and understanding.', ""Recommendation systems are often evaluated based on user's interactions that\nwere collected from an existing, already deployed recommendation system. In\nthis situation, users only provide feedback on the exposed items and they may\nnot leave feedback on other items since they have not been exposed to them by\nthe deployed system. As a result, the collected feedback dataset that is used\nto evaluate a new model is influenced by the deployed system, as a form of\nclosed loop feedback. In this paper, we show that the typical offline\nevaluation of recommender systems suffers from the so-called Simpson's paradox.\nSimpson's paradox is the name given to a phenomenon observed when a significant\ntrend appears in several different sub-populations of observational data but\ndisappears or is even reversed when these sub-populations are combined\ntogether. Our in-depth experiments based on stratified sampling reveal that a\nvery small minority of items that are frequently exposed by the deployed system\nplays a confounding factor in the offline evaluation of recommendation systems.\nIn addition, we propose a novel evaluation methodology that takes into account\nthe confounder, i.e the deployed system's characteristics. Using the relative\ncomparison of many recommendation models as in the typical offline evaluation\nof recommender systems, and based on the Kendall rank correlation coefficient,\nwe show that our proposed evaluation methodology exhibits statistically\nsignificant improvements of 14% and 40% on the examined open loop datasets\n(Yahoo! and Coat), respectively, in reflecting the true ranking of systems with\nan open loop (randomised) evaluation in comparison to the standard evaluation."", 'One of the goals of every business enterprise is to increase customer\nloyalty. The degree of customer loyalty is called customer quality which its\nforecasting will affect strategic marketing practices. The purpose of this\nstudy is to predict the quality of customers of large e-commerce social\nnetworks by big data algorithms and unsupervised learning. For this purpose, a\ngraph-based social network analysis framework was used for community detection\nin the Stanford Network Analysis Platform (SNAP). Then in the found\ncommunities, the quality of customers was predicted. The results showed that\nvarious visits with an impact of 37.13% can have the greatest impact on\ncustomer quality and the order of impact of other parameters were from highest\nto lowest: number of frequent customer visits (28.56%), role in social networks\n(28.37%), Indirect transactions (26.74%), activity days (25.62%) and customer\nsocial network size (25.06%).', ""We propose a general framework for interactively learning models, such as\n(binary or non-binary) classifiers, orderings/rankings of items, or clusterings\nof data points. Our framework is based on a generalization of Angluin's\nequivalence query model and Littlestone's online learning model: in each\niteration, the algorithm proposes a model, and the user either accepts it or\nreveals a specific mistake in the proposal. The feedback is correct only with\nprobability $p > 1/2$ (and adversarially incorrect with probability $1 - p$),\ni.e., the algorithm must be able to learn in the presence of arbitrary noise.\nThe algorithm's goal is to learn the ground truth model using few iterations.\n  Our general framework is based on a graph representation of the models and\nuser feedback. To be able to learn efficiently, it is sufficient that there be\na graph $G$ whose nodes are the models and (weighted) edges capture the user\nfeedback, with the property that if $s, s^*$ are the proposed and target\nmodels, respectively, then any (correct) user feedback $s'$ must lie on a\nshortest $s$-$s^*$ path in $G$. Under this one assumption, there is a natural\nalgorithm reminiscent of the Multiplicative Weights Update algorithm, which\nwill efficiently learn $s^*$ even in the presence of noise in the user's\nfeedback.\n  From this general result, we rederive with barely any extra effort classic\nresults on learning of classifiers and a recent result on interactive\nclustering; in addition, we easily obtain new interactive learning algorithms\nfor ordering/ranking."", 'In supervised learning, we typically leverage a fully labeled dataset to\ndesign methods for function estimation or prediction. In many practical\nsituations, we are able to obtain alternative feedback, possibly at a low cost.\nA broad goal is to understand the usefulness of, and to design algorithms to\nexploit, this alternative feedback. In this paper, we consider a\nsemi-supervised regression setting, where we obtain additional ordinal (or\ncomparison) information for the unlabeled samples. We consider ordinal feedback\nof varying qualities where we have either a perfect ordering of the samples, a\nnoisy ordering of the samples or noisy pairwise comparisons between the\nsamples. We provide a precise quantification of the usefulness of these types\nof ordinal feedback in both nonparametric and linear regression, showing that\nin many cases it is possible to accurately estimate an underlying function with\na very small labeled set, effectively \\emph{escaping the curse of\ndimensionality}. We also present lower bounds, that establish fundamental\nlimits for the task and show that our algorithms are optimal in a variety of\nsettings. Finally, we present extensive experiments on new datasets that\ndemonstrate the efficacy and practicality of our algorithms and investigate\ntheir robustness to various sources of noise and model misspecification.', ""We study the interplay between feedback and communication in a cooperative\nonline learning setting where a network of agents solves a task in which the\nlearners' feedback is determined by an arbitrary graph. We characterize regret\nin terms of the independence number of the strong product between the feedback\ngraph and the communication network. Our analysis recovers as special cases\nmany previously known bounds for distributed online learning with either expert\nor bandit feedback. A more detailed version of our results also captures the\ndependence of the regret on the delay caused by the time the information takes\nto traverse each graph. Experiments run on synthetic data show that the\nempirical behavior of our algorithm is consistent with the theoretical results."", 'We study a general problem of allocating limited resources to heterogeneous\ncustomers over time under model uncertainty. Each type of customer can be\nserviced using different actions, each of which stochastically consumes some\ncombination of resources, and returns different rewards for the resources\nconsumed. We consider a general model where the resource consumption\ndistribution associated with each (customer type, action)-combination is not\nknown, but is consistent and can be learned over time. In addition, the\nsequence of customer types to arrive over time is arbitrary and completely\nunknown.\n  We overcome both the challenges of model uncertainty and customer\nheterogeneity by judiciously synthesizing two algorithmic frameworks from the\nliterature: inventory balancing, which ""reserves"" a portion of each resource\nfor high-reward customer types which could later arrive, and online learning,\nwhich shows how to ""explore"" the resource consumption distributions of each\ncustomer type under different actions. We define an auxiliary problem, which\nallows for existing competitive ratio and regret bounds to be seamlessly\nintegrated. Furthermore, we show that the performance guarantee generated by\nour framework is tight, that is, we provide an information-theoretic lower\nbound which shows that both the loss from competitive ratio and the loss for\nregret are relevant in the combined problem.\n  Finally, we demonstrate the efficacy of our algorithms on a publicly\navailable hotel data set. Our framework is highly practical in that it requires\nno historical data (no fitted customer choice models, nor forecasting of\ncustomer arrival patterns) and can be used to initialize allocation strategies\nin fast-changing environments.', 'With high forward gain, a negative feedback system has the ability to perform\nthe inverse of a linear or non linear function that is in the feedback path.\nThis property of negative feedback systems has been widely used in analog\ncircuits to construct precise closed-loop functions. This paper describes how\nthe property of a negative feedback system to perform inverse of a function can\nbe used for training neural networks. This method does not require that the\ncost or activation functions be differentiable. Hence, it is able to learn a\nclass of non-differentiable functions as well where a gradient descent-based\nmethod fails. We also show that gradient descent emerges as a special case of\nthe proposed method. We have applied this method to the MNIST dataset and\nobtained results that shows the method is viable for neural network training.\nThis method, to the best of our knowledge, is novel in machine learning.', ""This work focuses on dynamic regret of online convex optimization that\ncompares the performance of online learning to a clairvoyant who knows the\nsequence of loss functions in advance and hence selects the minimizer of the\nloss function at each step. By assuming that the clairvoyant moves slowly\n(i.e., the minimizers change slowly), we present several improved\nvariation-based upper bounds of the dynamic regret under the true and noisy\ngradient feedback, which are {\\it optimal} in light of the presented lower\nbounds. The key to our analysis is to explore a regularity metric that measures\nthe temporal changes in the clairvoyant's minimizers, to which we refer as {\\it\npath variation}. Firstly, we present a general lower bound in terms of the path\nvariation, and then show that under full information or gradient feedback we\nare able to achieve an optimal dynamic regret. Secondly, we present a lower\nbound with noisy gradient feedback and then show that we can achieve optimal\ndynamic regrets under a stochastic gradient feedback and two-point bandit\nfeedback. Moreover, for a sequence of smooth loss functions that admit a small\nvariation in the gradients, our dynamic regret under the two-point bandit\nfeedback matches what is achieved with full information."", 'In e-commerce markets, on time delivery is of great importance to customer\nsatisfaction. In this paper, we present a Deep Reinforcement Learning (DRL)\napproach for deciding how and when orders should be batched and picked in a\nwarehouse to minimize the number of tardy orders. In particular, the technique\nfacilitates making decisions on whether an order should be picked individually\n(pick-by-order) or picked in a batch with other orders (pick-by-batch), and if\nso with which other orders. We approach the problem by formulating it as a\nsemi-Markov decision process and develop a vector-based state representation\nthat includes the characteristics of the warehouse system. This allows us to\ncreate a deep reinforcement learning solution that learns a strategy by\ninteracting with the environment and solve the problem with a proximal policy\noptimization algorithm. We evaluate the performance of the proposed DRL\napproach by comparing it with several batching and sequencing heuristics in\ndifferent problem settings. The results show that the DRL approach is able to\ndevelop a strategy that produces consistent, good solutions and performs better\nthan the proposed heuristics.', 'We study episodic reinforcement learning in Markov decision processes when\nthe agent receives additional feedback per step in the form of several\ntransition observations. Such additional observations are available in a range\nof tasks through extended sensors or prior knowledge about the environment\n(e.g., when certain actions yield similar outcome). We formalize this setting\nusing a feedback graph over state-action pairs and show that model-based\nalgorithms can leverage the additional feedback for more sample-efficient\nlearning. We give a regret bound that, ignoring logarithmic factors and\nlower-order terms, depends only on the size of the maximum acyclic subgraph of\nthe feedback graph, in contrast with a polynomial dependency on the number of\nstates and actions in the absence of a feedback graph. Finally, we highlight\nchallenges when leveraging a small dominating set of the feedback graph as\ncompared to the bandit setting and propose a new algorithm that can use\nknowledge of such a dominating set for more sample-efficient learning of a\nnear-optimal policy.', 'Relevance Feedback in Content-Based Image Retrieval is a method where the\nfeedback of the performance is being used to improve itself. Prior works use\nfeature re-weighting and classification techniques as the Relevance Feedback\nmethods. This paper shows a novel addition to the prior methods to further\nimprove the retrieval accuracy. In addition to all of these, the paper also\nshows a novel idea to even improve the 0-th iteration retrieval accuracy from\nthe information of Relevance Feedback.', ""We study a pricing setting where each customer is offered a contextualized\nprice based on customer and/or product features. Often only historical sales\ndata are available, so we observe whether a customer purchased a product at the\nprice prescribed rather than the customer's true valuation. Such observational\ndata are influenced by historical pricing policies, which introduce\ndifficulties in evaluating the effectiveness of future policies. The goal of\nthis paper is to formulate loss functions that can be used for evaluating\npricing policies directly from observational data, rather than going through an\nintermediate demand estimation stage, which may suffer from bias. To achieve\nthis, we adapt ideas from machine learning with corrupted labels, where we\nconsider each observed purchase decision as a known probabilistic\ntransformation of the customer's valuation. From this transformation, we derive\na class of unbiased loss functions. Within this class, we identify minimum\nvariance estimators and estimators robust to poor demand estimation.\nFurthermore, we show that for contextual pricing, estimators popular in the\noff-policy evaluation literature fall within this class of loss functions. We\noffer managerial insights into scenarios under which these estimators are\neffective."", 'To identify the most appropriate recommendation model for an e-commerce\nbusiness, a live evaluation should be performed on the shopping website to\nmeasure the influence of personalization in real-time. The aim of this paper is\nto introduce and justify two new metrics -- CTR NoRepeat and Click & Buy rate\n-- which stem from the standard metrics, Click-through(CTR) and Buy-through\nrate(BTR), respectively. The former variation tackles the issue of\noverestimation of clicks in the original CTR while the latter accounts for\nnoting purchases of products that have been previously clicked, in order to\nvalidate that the buy included in the metric is a result of customer\ninteractions. A significance test for independence of two means is conducted\nfor multiple datasets, between each of the new metrics and its respective\nparent to determine the novelty and necessity of the variants. The\nPearson-correlation coefficient is calculated to assess the strength of the\nlinear relationships and conclude on the predictability factor amongst the\naforementioned factors to investigate unknown connections between customer\nclicks and buys. Additionally, other metrics such as hits per customer, buyers\nper customer, clicks per customer etc. are introduced that help explain\nindicators of customer behavior on the e-commerce website in reference.', 'Through the method of Learning Feedback Linearization, we seek to learn a\nlinearizing controller to simplify the process of controlling a car to race\nautonomously. A soft actor-critic approach is used to learn a decoupling matrix\nand drift vector that effectively correct for errors in a hand-designed\nlinearizing controller. The result is an exactly linearizing controller that\ncan be used to enable the well-developed theory of linear systems to design\npath planning and tracking schemes that are easy to implement and significantly\nless computationally demanding. To demonstrate the method of feedback\nlinearization, it is first used to learn a simulated model whose exact\nstructure is known, but varied from the initial controller, so as to introduce\nerror. We further seek to apply this method to a system that introduces even\nmore error in the form of a gym environment specifically designed for modeling\nthe dynamics of car racing. To do so, we posit an extension to the method of\nlearning feedback linearization; a neural network that is trained using\nsupervised learning to convert the output of our linearizing controller to the\nrequired input for the racing environment. Our progress towards these goals is\nreported and the next steps in their accomplishment are discussed.', 'Personalized size and fit recommendations bear crucial significance for any\nfashion e-commerce platform. Predicting the correct fit drives customer\nsatisfaction and benefits the business by reducing costs incurred due to\nsize-related returns. Traditional collaborative filtering algorithms seek to\nmodel customer preferences based on their previous orders. A typical challenge\nfor such methods stems from extreme sparsity of customer-article orders. To\nalleviate this problem, we propose a deep learning based content-collaborative\nmethodology for personalized size and fit recommendation. Our proposed method\ncan ingest arbitrary customer and article data and can model multiple\nindividuals or intents behind a single account. The method optimizes a global\nset of parameters to learn population-level abstractions of size and fit\nrelevant information from observed customer-article interactions. It further\nemploys customer and article specific embedding variables to learn their\nproperties. Together with learned entity embeddings, the method maps additional\ncustomer and article attributes into a latent space to derive personalized\nrecommendations. Application of our method to two publicly available datasets\ndemonstrate an improvement over the state-of-the-art published results. On two\nproprietary datasets, one containing fit feedback from fashion experts and the\nother involving customer purchases, we further outperform comparable\nmethodologies, including a recent Bayesian approach for size recommendation.', 'In this paper, we relate the feedback capacity of parallel additive colored\nGaussian noise (ACGN) channels to a variant of the Kalman filter. By doing so,\nwe obtain lower bounds on the feedback capacity of such channels, as well as\nthe corresponding feedback (recursive) coding schemes, which are essentially\npower allocation policies with feedback, to achieve the bounds. The results are\nseen to reduce to existing lower bounds in the case of a single ACGN feedback\nchannel, whereas when it comes to parallel additive white Gaussian noise (AWGN)\nchannels with feedback, the recursive coding scheme reduces to a feedback\n""water-filling"" power allocation policy.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am working on a project where I input a sentence, my system will return related images based on my input sentence. I call this task image retrieval from sentences. This system needs to work with different languages. My current plan is to first create a synthetic dataset that consists of tuples of sentences in different languages and an image. I am trying to use image captioning or image annotation pre-trained models to create this synthetic dataset. After we have created this synthetic dataset, I will build a model that can learn the representations of sentences with respect to the image. I believe my work is in the intersection of information retrieval, learning sentence representation and data augmentation. In the end, I want my work to be integrated with a search engine that can be deployed in the cloud.","['The ability to describe images with natural language sentences is the\nhallmark for image and language understanding. Such a system has wide ranging\napplications such as annotating images and using natural sentences to search\nfor images.In this project we focus on the task of bidirectional image\nretrieval: such asystem is capable of retrieving an image based on a sentence\n(image search) andretrieve sentence based on an image query (image annotation).\nWe present asystem based on a global ranking objective function which uses a\ncombinationof convolutional neural networks (CNN) and multi layer perceptrons\n(MLP).It takes a pair of image and sentence and processes them in different\nchannels,finally embedding it into a common multimodal vector space. These\nembeddingsencode abstract semantic information about the two inputs and can be\ncomparedusing traditional information retrieval approaches. For each such pair,\nthe modelreturns a score which is interpretted as a similarity metric. If this\nscore is high,the image and sentence are likely to convey similar meaning, and\nif the score is low then they are likely not to.\n  The visual input is modeled via deep convolutional neural network. On\ntheother hand we explore three models for the textual module. The first one\nisbag of words with an MLP. The second one uses n-grams (bigram, trigrams,and a\ncombination of trigram & skip-grams) with an MLP. The third is morespecialized\ndeep network specific for modeling variable length sequences (SSE).We report\ncomparable performance to recent work in the field, even though ouroverall\nmodel is simpler. We also show that the training time choice of how wecan\ngenerate our negative samples has a significant impact on performance, and can\nbe used to specialize the bi-directional system in one particular task.', 'In this paper we propose a model to learn multimodal multilingual\nrepresentations for matching images and sentences in different languages, with\nthe aim of advancing multilingual versions of image search and image\nunderstanding. Our model learns a common representation for images and their\ndescriptions in two different languages (which need not be parallel) by\nconsidering the image as a pivot between two languages. We introduce a new\npairwise ranking loss function which can handle both symmetric and asymmetric\nsimilarity between the two modalities. We evaluate our models on\nimage-description ranking for German and English, and on semantic textual\nsimilarity of image descriptions in English. In both cases we achieve\nstate-of-the-art performance.', 'Cross-modal retrieval is an important functionality in modern search engines,\nas it increases the user experience by allowing queries and retrieved objects\nto pertain to different modalities. In this paper, we focus on the\nimage-sentence retrieval task, where the objective is to efficiently find\nrelevant images for a given sentence (image-retrieval) or the relevant\nsentences for a given image (sentence-retrieval). Computer vision literature\nreports the best results on the image-sentence matching task using deep neural\nnetworks equipped with attention and self-attention mechanisms. They evaluate\nthe matching performance on the retrieval task by performing sequential scans\nof the whole dataset. This method does not scale well with an increasing amount\nof images or captions. In this work, we explore different preprocessing\ntechniques to produce sparsified deep multi-modal features extracting them from\nstate-of-the-art deep-learning architectures for image-text matching. Our main\nobjective is to lay down the paths for efficient indexing of complex\nmulti-modal descriptions. We use the recently introduced TERN architecture as\nan image-sentence features extractor. It is designed for producing fixed-size\n1024-d vectors describing whole images and sentences, as well as\nvariable-length sets of 1024-d vectors describing the various building\ncomponents of the two modalities (image regions and sentence words\nrespectively). All these vectors are enforced by the TERN design to lie into\nthe same common space. Our experiments show interesting preliminary results on\nthe explored methods and suggest further experimentation in this important\nresearch direction.', ""Text-to-Image translation has been an active area of research in the recent\npast. The ability for a network to learn the meaning of a sentence and generate\nan accurate image that depicts the sentence shows ability of the model to think\nmore like humans. Popular methods on text to image translation make use of\nGenerative Adversarial Networks (GANs) to generate high quality images based on\ntext input, but the generated images don't always reflect the meaning of the\nsentence given to the model as input. We address this issue by using a\ncaptioning network to caption on generated images and exploit the distance\nbetween ground truth captions and generated captions to improve the network\nfurther. We show extensive comparisons between our method and existing methods."", 'Recent work in computer vision has yielded impressive results in\nautomatically describing images with natural language. Most of these systems\ngenerate captions in a sin- gle language, requiring multiple language-specific\nmodels to build a multilingual captioning system. We propose a very simple\ntechnique to build a single unified model across languages, using artificial\ntokens to control the language, making the captioning system more compact. We\nevaluate our approach on generating English and Japanese captions, and show\nthat a typical neural captioning architecture is capable of learning a single\nmodel that can switch between two different languages.', 'Our objective is video retrieval based on natural language queries. In\naddition, we consider the analogous problem of retrieving sentences or\ngenerating descriptions given an input video. Recent work has addressed the\nproblem by embedding visual and textual inputs into a common space where\nsemantic similarities correlate to distances. We also adopt the embedding\napproach, and make the following contributions: First, we utilize web image\nsearch in sentence embedding process to disambiguate fine-grained visual\nconcepts. Second, we propose embedding models for sentence, image, and video\ninputs whose parameters are learned simultaneously. Finally, we show how the\nproposed model can be applied to description generation. Overall, we observe a\nclear improvement over the state-of-the-art methods in the video and sentence\nretrieval tasks. In description generation, the performance level is comparable\nto the current state-of-the-art, although our embeddings were trained for the\nretrieval tasks.']","['This paper introduces a large-scale multimodal and multilingual dataset that\naims to facilitate research on grounding words to images in their contextual\nusage in language. The dataset consists of images selected to unambiguously\nillustrate concepts expressed in sentences from movie subtitles. The dataset is\na valuable resource as (i) the images are aligned to text fragments rather than\nwhole sentences; (ii) multiple images are possible for a text fragment and a\nsentence; (iii) the sentences are free-form and real-world like; (iv) the\nparallel texts are multilingual. We set up a fill-in-the-blank game for humans\nto evaluate the quality of the automatic image selection process of our\ndataset. We show the utility of the dataset on two automatic tasks: (i)\nfill-in-the-blank; (ii) lexical translation. Results of the human evaluation\nand automatic models demonstrate that images can be a useful complement to the\ntextual context. The dataset will benefit research on visual grounding of words\nespecially in the context of free-form sentences, and can be obtained from\nhttps://doi.org/10.5281/zenodo.5034604 under a Creative Commons licence.', 'Text contained in an image carries high-level semantics that can be exploited\nto achieve richer image understanding. In particular, the mere presence of text\nprovides strong guiding content that should be employed to tackle a diversity\nof computer vision tasks such as image retrieval, fine-grained classification,\nand visual question answering. In this paper, we address the problem of\nfine-grained classification and image retrieval by leveraging textual\ninformation along with visual cues to comprehend the existing intrinsic\nrelation between the two modalities. The novelty of the proposed model consists\nof the usage of a PHOC descriptor to construct a bag of textual words along\nwith a Fisher Vector Encoding that captures the morphology of text. This\napproach provides a stronger multimodal representation for this task and as our\nexperiments demonstrate, it achieves state-of-the-art results on two different\ntasks, fine-grained classification and image retrieval.', 'Image captioning has so far been explored mostly in English, as most\navailable datasets are in this language. However, the application of image\ncaptioning should not be restricted by language. Only few studies have been\nconducted for image captioning in a cross-lingual setting. Different from these\nworks that manually build a dataset for a target language, we aim to learn a\ncross-lingual captioning model fully from machine-translated sentences. To\nconquer the lack of fluency in the translated sentences, we propose in this\npaper a fluency-guided learning framework. The framework comprises a module to\nautomatically estimate the fluency of the sentences and another module to\nutilize the estimated fluency scores to effectively train an image captioning\nmodel for the target language. As experiments on two bilingual\n(English-Chinese) datasets show, our approach improves both fluency and\nrelevance of the generated captions in Chinese, but without using any manually\nwritten sentences from the target language.', 'Image Captioning is a task that requires models to acquire a multi-modal\nunderstanding of the world and to express this understanding in natural\nlanguage text. While the state-of-the-art for this task has rapidly improved in\nterms of n-gram metrics, these models tend to output the same generic captions\nfor similar images. In this work, we address this limitation and train a model\nthat generates more diverse and specific captions through an unsupervised\ntraining approach that incorporates a learning signal from an Image Retrieval\nmodel. We summarize previous results and improve the state-of-the-art on\ncaption diversity and novelty. We make our source code publicly available\nonline.', 'Recently, generative adversarial networks have gained a lot of popularity for\nimage generation tasks. However, such models are associated with complex\nlearning mechanisms and demand very large relevant datasets. This work borrows\nconcepts from image and video captioning models to form an image generative\nframework. The model is trained in a similar fashion as recurrent captioning\nmodel and uses the learned weights for image generation. This is done in an\ninverse direction, where the input is a caption and the output is an image. The\nvector representation of the sentence and frames are extracted from an\nencoder-decoder model which is initially trained on similar sentence and image\npairs. Our model conditions image generation on a natural language caption. We\nleverage a sequence-to-sequence model to generate synthetic captions that have\nthe same meaning for having a robust image generation. One key advantage of our\nmethod is that the traditional image captioning datasets can be used for\nsynthetic sentence paraphrases. Results indicate that images generated through\nmultiple captions are better at capturing the semantic meaning of the family of\ncaptions.', ""A big part of achieving Artificial General Intelligence(AGI) is to build a\nmachine that can see and listen like humans. Much work has focused on designing\nmodels for image classification, video classification, object detection, pose\nestimation, speech recognition, etc., and has achieved significant progress in\nrecent years thanks to deep learning. However, understanding the world is not\nenough. An AI agent also needs to know how to talk, especially how to\ncommunicate with a human. While perception (vision, for example) is more common\nacross animal species, the use of complicated language is unique to humans and\nis one of the most important aspects of intelligence.\n  In this thesis, we focus on generating textual output given visual input. In\nChapter 3, we focus on generating the referring expression, a text description\nfor an object in the image so that a receiver can infer which object is being\ndescribed. We use a comprehension machine to directly guide the generated\nreferring expressions to be more discriminative. In Chapter 4, we introduce a\nmethod that encourages discriminability in image caption generation. We show\nthat more discriminative captioning models generate more descriptive captions.\nIn Chapter 5, we study how training objectives and sampling methods affect the\nmodels' ability to generate diverse captions. We find that a popular captioning\ntraining strategy will be detrimental to the diversity of generated captions.\nIn Chapter 6, we propose a model that can control the length of generated\ncaptions. By changing the desired length, one can influence the style and\ndescriptiveness of the captions. Finally, in Chapter 7, we rank/generate\ninformative image tags according to their information utility. The proposed\nmethod better matches what humans think are the most important tags for the\nimages."", 'Automatically generating a natural language description of an image is a task\nclose to the heart of image understanding. In this paper, we present a\nmulti-model neural network method closely related to the human visual system\nthat automatically learns to describe the content of images. Our model consists\nof two sub-models: an object detection and localization model, which extract\nthe information of objects and their spatial relationship in images\nrespectively; Besides, a deep recurrent neural network (RNN) based on long\nshort-term memory (LSTM) units with attention mechanism for sentences\ngeneration. Each word of the description will be automatically aligned to\ndifferent objects of the input image when it is generated. This is similar to\nthe attention mechanism of the human visual system. Experimental results on the\nCOCO dataset showcase the merit of the proposed method, which outperforms\nprevious benchmark models.', ""Current image captioning systems perform at a merely descriptive level,\nessentially enumerating the objects in the scene and their relations. Humans,\non the contrary, interpret images by integrating several sources of prior\nknowledge of the world. In this work, we aim to take a step closer to producing\ncaptions that offer a plausible interpretation of the scene, by integrating\nsuch contextual information into the captioning pipeline. For this we focus on\nthe captioning of images used to illustrate news articles. We propose a novel\ncaptioning method that is able to leverage contextual information provided by\nthe text of news articles associated with an image. Our model is able to\nselectively draw information from the article guided by visual cues, and to\ndynamically extend the output dictionary to out-of-vocabulary named entities\nthat appear in the context source. Furthermore we introduce `GoodNews', the\nlargest news image captioning dataset in the literature and demonstrate\nstate-of-the-art results."", ""Named entity recognition (NER) is frequently addressed as a sequence\nclassification task where each input consists of one sentence of text. It is\nnevertheless clear that useful information for the task can often be found\noutside of the scope of a single-sentence context. Recently proposed\nself-attention models such as BERT can both efficiently capture long-distance\nrelationships in input as well as represent inputs consisting of several\nsentences, creating new opportunitites for approaches that incorporate\ncross-sentence information in natural language processing tasks. In this paper,\nwe present a systematic study exploring the use of cross-sentence information\nfor NER using BERT models in five languages. We find that adding context in the\nform of additional sentences to BERT input systematically increases NER\nperformance on all of the tested languages and models. Including multiple\nsentences in each input also allows us to study the predictions of the same\nsentences in different contexts. We propose a straightforward method,\nContextual Majority Voting (CMV), to combine different predictions for\nsentences and demonstrate this to further increase NER performance with BERT.\nOur approach does not require any changes to the underlying BERT architecture,\nrather relying on restructuring examples for training and prediction.\nEvaluation on established datasets, including the CoNLL'02 and CoNLL'03 NER\nbenchmarks, demonstrates that our proposed approach can improve on the\nstate-of-the-art NER results on English, Dutch, and Finnish, achieves the best\nreported BERT-based results on German, and is on par with performance reported\nwith other BERT-based approaches in Spanish. We release all methods implemented\nin this work under open licenses."", 'State-of-the-art approaches for image captioning require supervised training\ndata consisting of captions with paired image data. These methods are typically\nunable to use unsupervised data such as textual data with no corresponding\nimages, which is a much more abundant commodity. We here propose a novel way of\nusing such textual data by artificially generating missing visual information.\nWe evaluate this learning approach on a newly designed model that detects\nvisual concepts present in an image and feed them to a reviewer-decoder\narchitecture with an attention mechanism. Unlike previous approaches that\nencode visual concepts using word embeddings, we instead suggest using regional\nimage features which capture more intrinsic information. The main benefit of\nthis architecture is that it synthesizes meaningful thought vectors that\ncapture salient image properties and then applies a soft attentive decoder to\ndecode the thought vectors and generate image captions. We evaluate our model\non both Microsoft COCO and Flickr30K datasets and demonstrate that this model\ncombined with our semi-supervised learning method can largely improve\nperformance and help the model to generate more accurate and diverse captions.', 'Prevalent models based on artificial neural network (ANN) for sentence\nclassification often classify sentences in isolation without considering the\ncontext in which sentences appear. This hampers the traditional sentence\nclassification approaches to the problem of sequential sentence classification,\nwhere structured prediction is needed for better overall classification\nperformance. In this work, we present a hierarchical sequential labeling\nnetwork to make use of the contextual information within surrounding sentences\nto help classify the current sentence. Our model outperforms the\nstate-of-the-art results by 2%-3% on two benchmarking datasets for sequential\nsentence classification in medical scientific abstracts.', 'Image and sentence matching has made great progress recently, but it remains\nchallenging due to the large visual-semantic discrepancy. This mainly arises\nfrom that the representation of pixel-level image usually lacks of high-level\nsemantic information as in its matched sentence. In this work, we propose a\nsemantic-enhanced image and sentence matching model, which can improve the\nimage representation by learning semantic concepts and then organizing them in\na correct semantic order. Given an image, we first use a multi-regional\nmulti-label CNN to predict its semantic concepts, including objects,\nproperties, actions, etc. Then, considering that different orders of semantic\nconcepts lead to diverse semantic meanings, we use a context-gated sentence\ngeneration scheme for semantic order learning. It simultaneously uses the image\nglobal context containing concept relations as reference and the groundtruth\nsemantic order in the matched sentence as supervision. After obtaining the\nimproved image representation, we learn the sentence representation with a\nconventional LSTM, and then jointly perform image and sentence matching and\nsentence generation for model learning. Extensive experiments demonstrate the\neffectiveness of our learned semantic concepts and order, by achieving the\nstate-of-the-art results on two public benchmark datasets.', 'Building upon recent Deep Neural Network architectures, current approaches\nlying in the intersection of computer vision and natural language processing\nhave achieved unprecedented breakthroughs in tasks like automatic captioning or\nimage retrieval. Most of these learning methods, though, rely on large training\nsets of images associated with human annotations that specifically describe the\nvisual content. In this paper we propose to go a step further and explore the\nmore complex cases where textual descriptions are loosely related to the\nimages. We focus on the particular domain of News articles in which the textual\ncontent often expresses connotative and ambiguous relations that are only\nsuggested but not directly inferred from images. We introduce new deep learning\nmethods that address source detection, popularity prediction, article\nillustration and geolocation of articles. An adaptive CNN architecture is\nproposed, that shares most of the structure for all the tasks, and is suitable\nfor multitask and transfer learning. Deep Canonical Correlation Analysis is\ndeployed for article illustration, and a new loss function based on Great\nCircle Distance is proposed for geolocation. Furthermore, we present\nBreakingNews, a novel dataset with approximately 100K news articles including\nimages, text and captions, and enriched with heterogeneous meta-data (such as\nGPS coordinates and popularity metrics). We show this dataset to be appropriate\nto explore all aforementioned problems, for which we provide a baseline\nperformance using various Deep Learning architectures, and different\nrepresentations of the textual and visual features. We report very promising\nresults and bring to light several limitations of current state-of-the-art in\nthis kind of domain, which we hope will help spur progress in the field.', 'A widespread approach to processing spoken language is to first automatically\ntranscribe it into text. An alternative is to use an end-to-end approach:\nrecent works have proposed to learn semantic embeddings of spoken language from\nimages with spoken captions, without an intermediate transcription step. We\npropose to use multitask learning to exploit existing transcribed speech within\nthe end-to-end setting. We describe a three-task architecture which combines\nthe objectives of matching spoken captions with corresponding images, speech\nwith text, and text with images. We show that the addition of the speech/text\ntask leads to substantial performance improvements on image retrieval when\ncompared to training the speech/image task in isolation. We conjecture that\nthis is due to a strong inductive bias transcribed speech provides to the\nmodel, and offer supporting evidence for this.', 'In this paper, we propose a way of synthesizing realistic images directly\nwith natural language description, which has many useful applications, e.g.\nintelligent image manipulation. We attempt to accomplish such synthesis: given\na source image and a target text description, our model synthesizes images to\nmeet two requirements: 1) being realistic while matching the target text\ndescription; 2) maintaining other image features that are irrelevant to the\ntext description. The model should be able to disentangle the semantic\ninformation from the two modalities (image and text), and generate new images\nfrom the combined semantics. To achieve this, we proposed an end-to-end neural\narchitecture that leverages adversarial learning to automatically learn\nimplicit loss functions, which are optimized to fulfill the aforementioned two\nrequirements. We have evaluated our model by conducting experiments on\nCaltech-200 bird dataset and Oxford-102 flower dataset, and have demonstrated\nthat our model is capable of synthesizing realistic images that match the given\ndescriptions, while still maintain other features of original images.', 'We introduce a variety of models, trained on a supervised image captioning\ncorpus to predict the image features for a given caption, to perform sentence\nrepresentation grounding. We train a grounded sentence encoder that achieves\ngood performance on COCO caption and image retrieval and subsequently show that\nthis encoder can successfully be transferred to various NLP tasks, with\nimproved performance over text-only models. Lastly, we analyze the contribution\nof grounding, and show that word embeddings learned by this system outperform\nnon-grounded ones.', 'Content-based information retrieval is based on the information contained in\ndocuments rather than using metadata such as keywords. Most information\nretrieval methods are either based on text or image. In this paper, we\ninvestigate the usefulness of multimodal features for cross-lingual news search\nin various domains: politics, health, environment, sport, and finance. To this\nend, we consider five feature types for image and text and compare the\nperformance of the retrieval system using different combinations. Experimental\nresults show that retrieval results can be improved when considering both\nvisual and textual information. In addition, it is observed that among textual\nfeatures entity overlap outperforms word embeddings, while geolocation\nembeddings achieve better performance among visual features in the retrieval\ntask.', 'In this paper, we study the task of image retrieval, where the input query is\nspecified in the form of an image plus some text that describes desired\nmodifications to the input image. For example, we may present an image of the\nEiffel tower, and ask the system to find images which are visually similar but\nare modified in small ways, such as being taken at nighttime instead of during\nthe day. To tackle this task, we learn a similarity metric between a target\nimage and a source image plus source text, an embedding and composing function\nsuch that target image feature is close to the source image plus text\ncomposition feature. We propose a new way to combine image and text using such\nfunction that is designed for the retrieval task. We show this outperforms\nexisting approaches on 3 different datasets, namely Fashion-200k, MIT-States\nand a new synthetic dataset we create based on CLEVR. We also show that our\napproach can be used to classify input queries, in addition to image retrieval.', 'In this paper we explore the bi-directional mapping between images and their\nsentence-based descriptions. We propose learning this mapping using a recurrent\nneural network. Unlike previous approaches that map both sentences and images\nto a common embedding, we enable the generation of novel sentences given an\nimage. Using the same model, we can also reconstruct the visual features\nassociated with an image given its visual description. We use a novel recurrent\nvisual memory that automatically learns to remember long-term visual concepts\nto aid in both sentence generation and visual feature reconstruction. We\nevaluate our approach on several tasks. These include sentence generation,\nsentence retrieval and image retrieval. State-of-the-art results are shown for\nthe task of generating novel image descriptions. When compared to human\ngenerated captions, our automatically generated captions are preferred by\nhumans over $19.8\\%$ of the time. Results are better than or comparable to\nstate-of-the-art results on the image and sentence retrieval tasks for methods\nusing similar visual features.', 'Generating a novel textual description of an image is an interesting problem\nthat connects computer vision and natural language processing. In this paper,\nwe present a simple model that is able to generate descriptive sentences given\na sample image. This model has a strong focus on the syntax of the\ndescriptions. We train a purely bilinear model that learns a metric between an\nimage representation (generated from a previously trained Convolutional Neural\nNetwork) and phrases that are used to described them. The system is then able\nto infer phrases from a given image sample. Based on caption syntax statistics,\nwe propose a simple language model that can produce relevant descriptions for a\ngiven test image using the phrases inferred. Our approach, which is\nconsiderably simpler than state-of-the-art models, achieves comparable results\nin two popular datasets for the task: Flickr30k and the recently proposed\nMicrosoft COCO.', 'In this paper, we propose multimodal convolutional neural networks (m-CNNs)\nfor matching image and sentence. Our m-CNN provides an end-to-end framework\nwith convolutional architectures to exploit image representation, word\ncomposition, and the matching relations between the two modalities. More\nspecifically, it consists of one image CNN encoding the image content, and one\nmatching CNN learning the joint representation of image and sentence. The\nmatching CNN composes words to different semantic fragments and learns the\ninter-modal relations between image and the composed fragments at different\nlevels, thus fully exploit the matching relations between image and sentence.\nExperimental results on benchmark databases of bidirectional image and sentence\nretrieval demonstrate that the proposed m-CNNs can effectively capture the\ninformation necessary for image and sentence matching. Specifically, our\nproposed m-CNNs for bidirectional image and sentence retrieval on Flickr30K and\nMicrosoft COCO databases achieve the state-of-the-art performances.', 'Hypernymy, textual entailment, and image captioning can be seen as special\ncases of a single visual-semantic hierarchy over words, sentences, and images.\nIn this paper we advocate for explicitly modeling the partial order structure\nof this hierarchy. Towards this goal, we introduce a general method for\nlearning ordered representations, and show how it can be applied to a variety\nof tasks involving images and language. We show that the resulting\nrepresentations improve performance over current approaches for hypernym\nprediction and image-caption retrieval.', 'Image captioning is a research hotspot where encoder-decoder models combining\nconvolutional neural network (CNN) and long short-term memory (LSTM) achieve\npromising results. Despite significant progress, these models generate\nsentences differently from human cognitive styles. Existing models often\ngenerate a complete sentence from the first word to the end, without\nconsidering the influence of the following words on the whole sentence\ngeneration. In this paper, we explore the utilization of a human-like cognitive\nstyle, i.e., building overall cognition for the image to be described and the\nsentence to be constructed, for enhancing computer image understanding. This\npaper first proposes a Mutual-aid network structure with Bidirectional LSTMs\n(MaBi-LSTMs) for acquiring overall contextual information. In the training\nprocess, the forward and backward LSTMs encode the succeeding and preceding\nwords into their respective hidden states by simultaneously constructing the\nwhole sentence in a complementary manner. In the captioning process, the LSTM\nimplicitly utilizes the subsequent semantic information contained in its hidden\nstates. In fact, MaBi-LSTMs can generate two sentences in forward and backward\ndirections. To bridge the gap between cross-domain models and generate a\nsentence with higher quality, we further develop a cross-modal attention\nmechanism to retouch the two sentences by fusing their salient parts as well as\nthe salient areas of the image. Experimental results on the Microsoft COCO\ndataset show that the proposed model improves the performance of\nencoder-decoder models and achieves state-of-the-art results.', 'We present an approach to improve statistical machine translation of image\ndescriptions by multimodal pivots defined in visual space. The key idea is to\nperform image retrieval over a database of images that are captioned in the\ntarget language, and use the captions of the most similar images for\ncrosslingual reranking of translation outputs. Our approach does not depend on\nthe availability of large amounts of in-domain parallel data, but only relies\non available large datasets of monolingually captioned images, and on\nstate-of-the-art convolutional neural networks to compute image similarities.\nOur experimental evaluation shows improvements of 1 BLEU point over strong\nbaselines.', 'The traditional image captioning task uses generic reference captions to\nprovide textual information about images. Different user populations, however,\nwill care about different visual aspects of images. In this paper, we propose a\nnew task, Captioning with a Purpose (CapWAP). Our goal is to develop systems\nthat can be tailored to be useful for the information needs of an intended\npopulation, rather than merely provide generic information about an image. In\nthis task, we use question-answer (QA) pairs---a natural expression of\ninformation need---from users, instead of reference captions, for both training\nand post-inference evaluation. We show that it is possible to use reinforcement\nlearning to directly optimize for the intended information need, by rewarding\noutputs that allow a question answering model to provide correct answers to\nsampled user questions. We convert several visual question answering datasets\ninto CapWAP datasets, and demonstrate that under a variety of scenarios our\npurposeful captioning system learns to anticipate and fulfill specific\ninformation needs better than its generic counterparts, as measured by QA\nperformance on user questions from unseen images, when using the caption alone\nas context.', ""A conversational agent (chatbot) is a piece of software that is able to\ncommunicate with humans using natural language. Modeling conversation is an\nimportant task in natural language processing and artificial intelligence.\nWhile chatbots can be used for various tasks, in general they have to\nunderstand users' utterances and provide responses that are relevant to the\nproblem at hand.\n  In my work, I conduct an in-depth survey of recent literature, examining over\n70 publications related to chatbots published in the last 3 years. Then, I\nproceed to make the argument that the very nature of the general conversation\ndomain demands approaches that are different from current state-of-of-the-art\narchitectures. Based on several examples from the literature I show why current\nchatbot models fail to take into account enough priors when generating\nresponses and how this affects the quality of the conversation. In the case of\nchatbots, these priors can be outside sources of information that the\nconversation is conditioned on like the persona or mood of the conversers. In\naddition to presenting the reasons behind this problem, I propose several ideas\non how it could be remedied.\n  The next section focuses on adapting the very recent Transformer model to the\nchatbot domain, which is currently state-of-the-art in neural machine\ntranslation. I first present experiments with the vanilla model, using\nconversations extracted from the Cornell Movie-Dialog Corpus. Secondly, I\naugment the model with some of my ideas regarding the issues of encoder-decoder\narchitectures. More specifically, I feed additional features into the model\nlike mood or persona together with the raw conversation data. Finally, I\nconduct a detailed analysis of how the vanilla model performs on conversational\ndata by comparing it to previous chatbot models and how the additional features\naffect the quality of the generated responses."", 'Localizing moments in a longer video via natural language queries is a new,\nchallenging task at the intersection of language and video understanding.\nThough moment localization with natural language is similar to other language\nand vision tasks like natural language object retrieval in images, moment\nlocalization offers an interesting opportunity to model temporal dependencies\nand reasoning in text. We propose a new model that explicitly reasons about\ndifferent temporal segments in a video, and shows that temporal context is\nimportant for localizing phrases which include temporal language. To benchmark\nwhether our model, and other recent video localization models, can effectively\nreason about temporal language, we collect the novel TEMPOral reasoning in\nvideo and language (TEMPO) dataset. Our dataset consists of two parts: a\ndataset with real videos and template sentences (TEMPO - Template Language)\nwhich allows for controlled studies on temporal language, and a human language\ndataset which consists of temporal sentences annotated by humans (TEMPO - Human\nLanguage).', 'Image captioning models are becoming increasingly successful at describing\nthe content of images in restricted domains. However, if these models are to\nfunction in the wild - for example, as assistants for people with impaired\nvision - a much larger number and variety of visual concepts must be\nunderstood. To address this problem, we teach image captioning models new\nvisual concepts from labeled images and object detection datasets. Since image\nlabels and object classes can be interpreted as partial captions, we formulate\nthis problem as learning from partially-specified sequence data. We then\npropose a novel algorithm for training sequence models, such as recurrent\nneural networks, on partially-specified sequences which we represent using\nfinite state automata. In the context of image captioning, our method lifts the\nrestriction that previously required image captioning models to be trained on\npaired image-sentence corpora only, or otherwise required specialized model\narchitectures to take advantage of alternative data modalities. Applying our\napproach to an existing neural captioning model, we achieve state of the art\nresults on the novel object captioning task using the COCO dataset. We further\nshow that we can train a captioning model to describe new visual concepts from\nthe Open Images dataset while maintaining competitive COCO evaluation scores.', 'In this work we propose a simple and efficient framework for learning\nsentence representations from unlabelled data. Drawing inspiration from the\ndistributional hypothesis and recent work on learning sentence representations,\nwe reformulate the problem of predicting the context in which a sentence\nappears as a classification problem. Given a sentence and its context, a\nclassifier distinguishes context sentences from other contrastive sentences\nbased on their vector representations. This allows us to efficiently learn\ndifferent types of encoding functions, and we show that the model learns\nhigh-quality sentence representations. We demonstrate that our sentence\nrepresentations outperform state-of-the-art unsupervised and supervised\nrepresentation learning methods on several downstream NLP tasks that involve\nunderstanding sentence semantics while achieving an order of magnitude speedup\nin training time.', 'Learning distributed sentence representations remains an interesting problem\nin the field of Natural Language Processing (NLP). We want to learn a model\nthat approximates the conditional latent space over the representations of a\nlogical antecedent of the given statement. In our paper, we propose an approach\nto generating sentences, conditioned on an input sentence and a logical\ninference label. We do this by modeling the different possibilities for the\noutput sentence as a distribution over the latent representation, which we\ntrain using an adversarial objective. We evaluate the model using two\nstate-of-the-art models for the Recognizing Textual Entailment (RTE) task, and\nmeasure the BLEU scores against the actual sentences as a probe for the\ndiversity of sentences produced by our model. The experiment results show that,\ngiven our framework, we have clear ways to improve the quality and diversity of\ngenerated sentences.', 'Automatically creating the description of an image using any natural\nlanguages sentence like English is a very challenging task. It requires\nexpertise of both image processing as well as natural language processing. This\npaper discuss about different available models for image captioning task. We\nhave also discussed about how the advancement in the task of object recognition\nand machine translation has greatly improved the performance of image\ncaptioning model in recent years. In addition to that we have discussed how\nthis model can be implemented. In the end, we have also evaluated the\nperformance of model using standard evaluation matrices.', ""In this paper, we propose a new approach for retrieval of video segments\nusing natural language queries. Unlike most previous approaches such as\nconcept-based methods or rule-based structured models, the proposed method uses\nimage captioning model to construct sentential queries for visual information.\nIn detail, our approach exploits multiple captions generated by visual features\nin each image with `Densecap'. Then, the similarities between captions of\nadjacent images are calculated, which is used to track semantically similar\ncaptions over multiple frames. Besides introducing this novel idea of 'tracking\nby captioning', the proposed method is one of the first approaches that uses a\nlanguage generation model learned by neural networks to construct semantic\nquery describing the relations and properties of visual information. To\nevaluate the effectiveness of our approach, we have created a new evaluation\ndataset, which contains about 348 segments of scenes in 20 movie-trailers.\nThrough quantitative and qualitative evaluation, we show that our method is\neffective for retrieval of video segments using natural language queries."", ""Distributed representation plays an important role in deep learning based\nnatural language processing. However, the representation of a sentence often\nvaries in different tasks, which is usually learned from scratch and suffers\nfrom the limited amounts of training data. In this paper, we claim that a good\nsentence representation should be invariant and can benefit the various\nsubsequent tasks. To achieve this purpose, we propose a new scheme of\ninformation sharing for multi-task learning. More specifically, all tasks share\nthe same sentence representation and each task can select the task-specific\ninformation from the shared sentence representation with attention mechanism.\nThe query vector of each task's attention could be either static parameters or\ngenerated dynamically. We conduct extensive experiments on 16 different text\nclassification tasks, which demonstrate the benefits of our architecture."", 'Visual-semantic embedding aims to find a shared latent space where related\nvisual and textual instances are close to each other. Most current methods\nlearn injective embedding functions that map an instance to a single point in\nthe shared space. Unfortunately, injective embedding cannot effectively handle\npolysemous instances with multiple possible meanings; at best, it would find an\naverage representation of different meanings. This hinders its use in\nreal-world scenarios where individual instances and their cross-modal\nassociations are often ambiguous. In this work, we introduce Polysemous\nInstance Embedding Networks (PIE-Nets) that compute multiple and diverse\nrepresentations of an instance by combining global context with locally-guided\nfeatures via multi-head self-attention and residual learning. To learn\nvisual-semantic embedding, we tie-up two PIE-Nets and optimize them jointly in\nthe multiple instance learning framework. Most existing work on cross-modal\nretrieval focuses on image-text data. Here, we also tackle a more challenging\ncase of video-text retrieval. To facilitate further research in video-text\nretrieval, we release a new dataset of 50K video-sentence pairs collected from\nsocial media, dubbed MRW (my reaction when). We demonstrate our approach on\nboth image-text and video-text retrieval scenarios using MS-COCO, TGIF, and our\nnew MRW dataset.', 'Knowledge-intensive NLP tasks can benefit from linking natural language text\nwith facts from a Knowledge Graph (KG). Although facts themselves are\nlanguage-agnostic, the fact labels (i.e., language-specific representation of\nthe fact) in the KG are often present only in a few languages. This makes it\nchallenging to link KG facts to sentences in languages other than the limited\nset of languages. To address this problem, we introduce the task of\nMultilingual Fact Linking (MFL) where the goal is to link fact expressed in a\nsentence to corresponding fact in the KG, even when the fact label in the KG is\nnot available in the language of the sentence. To facilitate research in this\narea, we present a new evaluation dataset, IndicLink. This dataset contains\n11,293 linked WikiData facts and 6,429 sentences spanning English and six\nIndian languages. We propose a Retrieval+Generation model, ReFCoG, that can\nscale to millions of KG facts by combining Dual Encoder based retrieval with a\nSeq2Seq based generation model which is constrained to output only valid KG\nfacts. ReFCoG outperforms standard Retrieval+Re-ranking models by 10.7 pts in\nPrecision@1. In spite of this gain, the model achieves an overall score of\n52.1, showing ample scope for improvement in the task.ReFCoG code and IndicLink\ndata are available at https://github.com/SaiKeshav/mfl', ""A creative image-and-text generative AI system mimics humans' extraordinary\nabilities to provide users with diverse and comprehensive caption suggestions,\nas well as rich image creations. In this work, we demonstrate such an AI\ncreation system to produce both diverse captions and rich images. When users\nimagine an image and associate it with multiple captions, our system paints a\nrich image to reflect all captions faithfully. Likewise, when users upload an\nimage, our system depicts it with multiple diverse captions. We propose a\nunified multi-modal framework to achieve this goal. Specifically, our framework\njointly models image-and-text representations with a Transformer network, which\nsupports rich image creation by accepting multiple captions as input. We\nconsider the relations among input captions to encourage diversity in training\nand adopt a non-autoregressive decoding strategy to enable real-time inference.\nBased on these, our system supports both diverse captions and rich images\ngenerations. Our code is available online."", ""Discovering the logical sequence of events is one of the cornerstones in\nNatural Language Understanding. One approach to learn the sequence of events is\nto study the order of sentences in a coherent text. Sentence ordering can be\napplied in various tasks such as retrieval-based Question Answering, document\nsummarization, storytelling, text generation, and dialogue systems.\nFurthermore, we can learn to model text coherence by learning how to order a\nset of shuffled sentences. Previous research has relied on RNN, LSTM, and\nBiLSTM architecture for learning text language models. However, these networks\nhave performed poorly due to the lack of attention mechanisms. We propose an\nalgorithm for sentence ordering in a corpus of short stories. Our proposed\nmethod uses a language model based on Universal Transformers (UT) that captures\nsentences' dependencies by employing an attention mechanism. Our method\nimproves the previous state-of-the-art in terms of Perfect Match Ratio (PMR)\nscore in the ROCStories dataset, a corpus of nearly 100K short human-made\nstories. The proposed model includes three components: Sentence Encoder,\nLanguage Model, and Sentence Arrangement with Brute Force Search. The first\ncomponent generates sentence embeddings using SBERT-WK pre-trained model\nfine-tuned on the ROCStories data. Then a Universal Transformer network\ngenerates a sentence-level language model. For decoding, the network generates\na candidate sentence as the following sentence of the current sentence. We use\ncosine similarity as a scoring function to assign scores to the candidate\nembedding and the embeddings of other sentences in the shuffled set. Then a\nBrute Force Search is employed to maximize the sum of similarities between\npairs of consecutive sentences."", 'Generating a description of an image is called image captioning. Image\ncaptioning requires to recognize the important objects, their attributes and\ntheir relationships in an image. It also needs to generate syntactically and\nsemantically correct sentences. Deep learning-based techniques are capable of\nhandling the complexities and challenges of image captioning. In this survey\npaper, we aim to present a comprehensive review of existing deep learning-based\nimage captioning techniques. We discuss the foundation of the techniques to\nanalyze their performances, strengths and limitations. We also discuss the\ndatasets and the evaluation metrics popularly used in deep learning based\nautomatic image captioning.', 'Deep neural networks have achieved great successes on the image captioning\ntask. However, most of the existing models depend heavily on paired\nimage-sentence datasets, which are very expensive to acquire. In this paper, we\nmake the first attempt to train an image captioning model in an unsupervised\nmanner. Instead of relying on manually labeled image-sentence pairs, our\nproposed model merely requires an image set, a sentence corpus, and an existing\nvisual concept detector. The sentence corpus is used to teach the captioning\nmodel how to generate plausible sentences. Meanwhile, the knowledge in the\nvisual concept detector is distilled into the captioning model to guide the\nmodel to recognize the visual concepts in an image. In order to further\nencourage the generated captions to be semantically consistent with the image,\nthe image and caption are projected into a common latent space so that they can\nreconstruct each other. Given that the existing sentence corpora are mainly\ndesigned for linguistic research and are thus with little reference to image\ncontents, we crawl a large-scale image description corpus of two million\nnatural sentences to facilitate the unsupervised image captioning scenario.\nExperimental results show that our proposed model is able to produce quite\npromising results without any caption annotations.', 'Image captioning aims at automatically generating descriptions of an image in\nnatural language. This is a challenging problem in the field of artificial\nintelligence that has recently received significant attention in the computer\nvision and natural language processing. Among the existing approaches, visual\nretrieval based methods have been proven to be highly effective. These\napproaches search for similar images, then build a caption for the query image\nbased on the captions of the retrieved images. In this study, we present a\nmethod for visual retrieval based image captioning, in which we use a multi\ncriteria decision making algorithm to effectively combine several criteria with\nproportional impact weights to retrieve the most relevant caption for the query\nimage. The main idea of the proposed approach is to design a mechanism to\nretrieve more semantically relevant captions with the query image and then\nselecting the most appropriate caption by imitation of the human act based on a\nweighted multi-criteria decision making algorithm. Experiments conducted on MS\nCOCO benchmark dataset have shown that proposed method provides much more\neffective results in compare to the state-of-the-art models by using criteria\nwith proportional impact weights .', 'This paper discusses and demonstrates the outcomes from our experimentation\non Image Captioning. Image captioning is a much more involved task than image\nrecognition or classification, because of the additional challenge of\nrecognizing the interdependence between the objects/concepts in the image and\nthe creation of a succinct sentential narration. Experiments on several labeled\ndatasets show the accuracy of the model and the fluency of the language it\nlearns solely from image descriptions. As a toy application, we apply image\ncaptioning to create video captions, and we advance a few hypotheses on the\nchallenges we encountered.', 'We present an easy and efficient method to extend existing sentence embedding\nmodels to new languages. This allows to create multilingual versions from\npreviously monolingual models. The training is based on the idea that a\ntranslated sentence should be mapped to the same location in the vector space\nas the original sentence. We use the original (monolingual) model to generate\nsentence embeddings for the source language and then train a new system on\ntranslated sentences to mimic the original model. Compared to other methods for\ntraining multilingual sentence embeddings, this approach has several\nadvantages: It is easy to extend existing models with relatively few samples to\nnew languages, it is easier to ensure desired properties for the vector space,\nand the hardware requirements for training is lower. We demonstrate the\neffectiveness of our approach for 50+ languages from various language families.\nCode to extend sentence embeddings models to more than 400 languages is\npublicly available.', 'Sentiment analysis is a research topic focused on analysing data to extract\ninformation related to the sentiment that it causes. Applications of sentiment\nanalysis are wide, ranging from recommendation systems, and marketing to\ncustomer satisfaction. Recent approaches evaluate textual content using Machine\nLearning techniques that are trained over large corpora. However, as social\nmedia grown, other data types emerged in large quantities, such as images.\nSentiment analysis in images has shown to be a valuable complement to textual\ndata since it enables the inference of the underlying message polarity by\ncreating context and connections. Multimodal sentiment analysis approaches\nintend to leverage information of both textual and image content to perform an\nevaluation. Despite recent advances, current solutions still flounder in\ncombining both image and textual information to classify social media data,\nmainly due to subjectivity, inter-class homogeneity and fusion data\ndifferences. In this paper, we propose a method that combines both textual and\nimage individual sentiment analysis into a final fused classification based on\nAutoML, that performs a random search to find the best model. Our method\nachieved state-of-the-art performance in the B-T4SA dataset, with 95.19%\naccuracy.', ""In this paper, we propose a multimodal search engine that combines visual and\ntextual cues to retrieve items from a multimedia database aesthetically similar\nto the query. The goal of our engine is to enable intuitive retrieval of\nfashion merchandise such as clothes or furniture. Existing search engines treat\ntextual input only as an additional source of information about the query image\nand do not correspond to the real-life scenario where the user looks for 'the\nsame shirt but of denim'. Our novel method, dubbed DeepStyle, mitigates those\nshortcomings by using a joint neural network architecture to model contextual\ndependencies between features of different modalities. We prove the robustness\nof this approach on two different challenging datasets of fashion items and\nfurniture where our DeepStyle engine outperforms baseline methods by 18-21% on\nthe tested datasets. Our search engine is commercially deployed and available\nthrough a Web-based application."", 'The task of image-text matching aims to map representations from different\nmodalities into a common joint visual-textual embedding. However, the most\nwidely used datasets for this task, MSCOCO and Flickr30K, are actually image\ncaptioning datasets that offer a very limited set of relationships between\nimages and sentences in their ground-truth annotations. This limited ground\ntruth information forces us to use evaluation metrics based on binary\nrelevance: given a sentence query we consider only one image as relevant.\nHowever, many other relevant images or captions may be present in the dataset.\nIn this work, we propose two metrics that evaluate the degree of semantic\nrelevance of retrieved items, independently of their annotated binary\nrelevance. Additionally, we incorporate a novel strategy that uses an image\ncaptioning metric, CIDEr, to define a Semantic Adaptive Margin (SAM) to be\noptimized in a standard triplet loss. By incorporating our formulation to\nexisting models, a \\emph{large} improvement is obtained in scenarios where\navailable training data is limited. We also demonstrate that the performance on\nthe annotated image-caption pairs is maintained while improving on other\nnon-annotated relevant items when employing the full training set. Code with\nour metrics and adaptive margin formulation will be made public.', ""In this technical report I present my method for automatic synthetic dataset\ngeneration for object detection and demonstrate it on the video game League of\nLegends. This report furthermore serves as a handbook on how to automatically\ngenerate datasets and as an introduction on the dataset generation part of the\nLeagueAI framework. The LeagueAI framework is a software framework that\nprovides detailed information about the game League of Legends based on the\nsame input a human player would have, namely vision. The framework allows\nresearchers and enthusiasts to develop their own intelligent agents or to\nextract detailed information about the state of the game. A big problem of\nmachine vision applications usually is the laborious work of gathering large\namounts of hand labeled data. Thus, a crucial part of the vision pipeline of\nthe LeagueAI framework, the dataset generation, is presented in this report.\nThe method involves extracting image raw data from the game's 3D models and\ncombining them with the game background to create game-like synthetic images\nand to generate the corresponding labels automatically. In an experiment I\ncompared a model trained on synthetic data to a model trained on hand labeled\ndata and a model trained on a combined dataset. The model trained on the\nsynthetic data showed higher detection precision on more classes and more\nreliable tracking performance of the player character. The model trained on the\ncombined dataset did not perform better because of the different formats of the\nolder hand labeled dataset and the synthetic data."", ""Feature visualizations such as synthetic maximally activating images are a\nwidely used explanation method to better understand the information processing\nof convolutional neural networks (CNNs). At the same time, there are concerns\nthat these visualizations might not accurately represent CNNs' inner workings.\nHere, we measure how much extremely activating images help humans to predict\nCNN activations. Using a well-controlled psychophysical paradigm, we compare\nthe informativeness of synthetic images by Olah et al. (2017) with a simple\nbaseline visualization, namely exemplary natural images that also strongly\nactivate a specific feature map. Given either synthetic or natural reference\nimages, human participants choose which of two query images leads to strong\npositive activation. The experiments are designed to maximize participants'\nperformance, and are the first to probe intermediate instead of final layer\nrepresentations. We find that synthetic images indeed provide helpful\ninformation about feature map activations ($82\\pm4\\%$ accuracy; chance would be\n$50\\%$). However, natural images - originally intended as a baseline -\noutperform synthetic images by a wide margin ($92\\pm2\\%$). Additionally,\nparticipants are faster and more confident for natural images, whereas\nsubjective impressions about the interpretability of the feature visualizations\nare mixed. The higher informativeness of natural images holds across most\nlayers, for both expert and lay participants as well as for hand- and\nrandomly-picked feature visualizations. Even if only a single reference image\nis given, synthetic images provide less information than natural images\n($65\\pm5\\%$ vs. $73\\pm4\\%$). In summary, synthetic images from a popular\nfeature visualization method are significantly less informative for assessing\nCNN activations than natural images. We argue that visualization methods should\nimprove over this baseline."", 'In this paper, we use the framework of neural machine translation to learn\njoint sentence representations across six very different languages. Our aim is\nthat a representation which is independent of the language, is likely to\ncapture the underlying semantics. We define a new cross-lingual similarity\nmeasure, compare up to 1.4M sentence representations and study the\ncharacteristics of close sentences. We provide experimental evidence that\nsentences that are close in embedding space are indeed semantically highly\nrelated, but often have quite different structure and syntax. These relations\nalso hold when comparing sentences in different languages.', 'Recent literature shows that large-scale language modeling provides excellent\nreusable sentence representations with both recurrent and self-attentive\narchitectures. However, there has been less clarity on the commonalities and\ndifferences in the representational properties induced by the two\narchitectures. It also has been shown that visual information serves as one of\nthe means for grounding sentence representations. In this paper, we present a\nmeta-study assessing the representational quality of models where the training\nsignal is obtained from different modalities, in particular, language modeling,\nimage features prediction, and both textual and multimodal machine translation.\nWe evaluate textual and visual features of sentence representations obtained\nusing predominant approaches on image retrieval and semantic textual\nsimilarity. Our experiments reveal that on moderate-sized datasets, a sentence\ncounterpart in a target language or visual modality provides much stronger\ntraining signal for sentence representation than language modeling.\nImportantly, we observe that while the Transformer models achieve superior\nmachine translation quality, representations from the recurrent neural network\nbased models perform significantly better over tasks focused on semantic\nrelevance.', 'Textual network embeddings aim to learn a low-dimensional representation for\nevery node in the network so that both the structural and textual information\nfrom the networks can be well preserved in the representations. Traditionally,\nthe structural and textual embeddings were learned by models that rarely take\nthe mutual influences between them into account. In this paper, a deep neural\narchitecture is proposed to effectively fuse the two kinds of informations into\none representation. The novelties of the proposed architecture are manifested\nin the aspects of a newly defined objective function, the complementary\ninformation fusion method for structural and textual features, and the mutual\ngate mechanism for textual feature extraction. Experimental results show that\nthe proposed model outperforms the comparing methods on all three datasets.', 'In this paper, we describe our end-to-end content-based image retrieval\nsystem built upon Elasticsearch, a well-known and popular textual search\nengine. As far as we know, this is the first time such a system has been\nimplemented in eCommerce, and our efforts have turned out to be highly\nworthwhile. We end up with a novel and exciting visual search solution that is\nextremely easy to be deployed, distributed, scaled and monitored in a\ncost-friendly manner. Moreover, our platform is intrinsically flexible in\nsupporting multimodal searches, where visual and textual information can be\njointly leveraged in retrieval.\n  The core idea is to encode image feature vectors into a collection of string\ntokens in a way such that closer vectors will share more string tokens in\ncommon. By doing that, we can utilize Elasticsearch to efficiently retrieve\nsimilar images based on similarities within encoded sting tokens. As part of\nthe development, we propose a novel vector to string encoding method, which is\nshown to substantially outperform the previous ones in terms of both precision\nand latency.\n  First-hand experiences in implementing this Elasticsearch-based platform are\nextensively addressed, which should be valuable to practitioners also\ninterested in building visual search engine on top of Elasticsearch.', 'For general modeling methods applied to diverse languages, a natural question\nis: how well should we expect our models to work on languages with differing\ntypological profiles? In this work, we develop an evaluation framework for fair\ncross-linguistic comparison of language models, using translated text so that\nall models are asked to predict approximately the same information. We then\nconduct a study on 21 languages, demonstrating that in some languages, the\ntextual expression of the information is harder to predict with both $n$-gram\nand LSTM language models. We show complex inflectional morphology to be a cause\nof performance differences among languages.', 'In recent years, automatic generation of image descriptions (captions), that\nis, image captioning, has attracted a great deal of attention. In this paper,\nwe particularly consider generating Japanese captions for images. Since most\navailable caption datasets have been constructed for English language, there\nare few datasets for Japanese. To tackle this problem, we construct a\nlarge-scale Japanese image caption dataset based on images from MS-COCO, which\nis called STAIR Captions. STAIR Captions consists of 820,310 Japanese captions\nfor 164,062 images. In the experiment, we show that a neural network trained\nusing STAIR Captions can generate more natural and better Japanese captions,\ncompared to those generated using English-Japanese machine translation after\ngenerating English captions.', ""We present COSTRA 1.0, a dataset of complex sentence transformations. The\ndataset is intended for the study of sentence-level embeddings beyond simple\nword alternations or standard paraphrasing. This first version of the dataset\nis limited to sentences in Czech but the construction method is universal and\nwe plan to use it also for other languages. The dataset consist of 4,262 unique\nsentences with average length of 10 words, illustrating 15 types of\nmodifications such as simplification, generalization, or formal and informal\nlanguage variation. The hope is that with this dataset, we should be able to\ntest semantic properties of sentence embeddings and perhaps even to find some\ntopologically interesting 'skeleton' in the sentence embedding space. A\npreliminary analysis using LASER, multi-purpose multi-lingual sentence\nembeddings suggests that the LASER space does not exhibit the desired\nproperties."", ""The standard recurrent neural network language model (RNNLM) generates\nsentences one word at a time and does not work from an explicit global sentence\nrepresentation. In this work, we introduce and study an RNN-based variational\nautoencoder generative model that incorporates distributed latent\nrepresentations of entire sentences. This factorization allows it to explicitly\nmodel holistic properties of sentences such as style, topic, and high-level\nsyntactic features. Samples from the prior over these sentence representations\nremarkably produce diverse and well-formed sentences through simple\ndeterministic decoding. By examining paths through this latent space, we are\nable to generate coherent novel sentences that interpolate between known\nsentences. We present techniques for solving the difficult learning problem\npresented by this model, demonstrate its effectiveness in imputing missing\nwords, explore many interesting properties of the model's latent sentence\nspace, and present negative results on the use of the model in language\nmodeling."", 'Generative models have made significant progress in the tasks of modeling\ncomplex data distributions such as natural images. The introduction of\nGenerative Adversarial Networks (GANs) and auto-encoders lead to the\npossibility of training on big data sets in an unsupervised manner. However,\nfor many generative models it is not possible to specify what kind of image\nshould be generated and it is not possible to translate existing images into\nnew images of similar domains. Furthermore, models that can perform\nimage-to-image translation often need distinct models for each domain, making\nit hard to scale these systems to multiple domain image-to-image translation.\nWe introduce a model that can do both, controllable image generation and\nimage-to-image translation between multiple domains. We split our image\nrepresentation into two parts encoding unstructured and structured information\nrespectively. The latter is designed in a disentangled manner, so that\ndifferent parts encode different image characteristics. We train an encoder to\nencode images into these representations and use a small amount of labeled data\nto specify what kind of information should be encoded in the disentangled part.\nA generator is trained to generate images from these representations using the\ncharacteristics provided by the disentangled part of the representation.\nThrough this we can control what kind of images the generator generates,\ntranslate images between different domains, and even learn unknown\ndata-generating factors while only using one single model.', 'Few-Shot Learning is the challenge of training a model with only a small\namount of data. Many solutions to this problem use meta-learning algorithms,\ni.e. algorithms that learn to learn. By sampling few-shot tasks from a larger\ndataset, we can teach these algorithms to solve new, unseen tasks. This\ndocument reports my work on meta-learning algorithms for Few-Shot Computer\nVision. This work was done during my internship at Sicara, a French company\nbuilding image recognition solutions for businesses. It contains: 1. an\nextensive review of the state-of-the-art in few-shot computer vision; 2. a\nbenchmark of meta-learning algorithms for few-shot image classification; 3. the\nintroduction to a novel meta-learning algorithm for few-shot object detection,\nwhich is still in development.', ""There is a lot of research interest in encoding variable length sentences\ninto fixed length vectors, in a way that preserves the sentence meanings. Two\ncommon methods include representations based on averaging word vectors, and\nrepresentations based on the hidden states of recurrent neural networks such as\nLSTMs. The sentence vectors are used as features for subsequent machine\nlearning tasks or for pre-training in the context of deep learning. However,\nnot much is known about the properties that are encoded in these sentence\nrepresentations and about the language information they capture. We propose a\nframework that facilitates better understanding of the encoded representations.\nWe define prediction tasks around isolated aspects of sentence structure\n(namely sentence length, word content, and word order), and score\nrepresentations by the ability to train a classifier to solve each prediction\ntask when using the representation as input. We demonstrate the potential\ncontribution of the approach by analyzing different sentence representation\nmechanisms. The analysis sheds light on the relative strengths of different\nsentence embedding methods with respect to these low level prediction tasks,\nand on the effect of the encoded vector's dimensionality on the resulting\nrepresentations."", 'In order to operate autonomously, a robot should explore the environment and\nbuild a model of each of the surrounding objects. A common approach is to\ncarefully scan the whole workspace. This is time-consuming. It is also often\nimpossible to reach all the viewpoints required to acquire full knowledge about\nthe environment. Humans can perform shape completion of occluded objects by\nrelying on past experience. Therefore, we propose a method that generates\nimages of an object from various viewpoints using a single input RGB image. A\ndeep neural network is trained to imagine the object appearance from many\nviewpoints. We present the whole pipeline, which takes a single RGB image as\ninput and returns a sequence of RGB and depth images of the object. The method\nutilizes a CNN-based object detector to extract the object from the natural\nscene. Then, the proposed network generates a set of RGB and depth images. We\nshow the results both on a synthetic dataset and on real images.', 'We propose an approach to build a neural machine translation system with no\nsupervised resources (i.e., no parallel corpora) using multimodal embedded\nrepresentation over texts and images. Based on the assumption that text\ndocuments are often likely to be described with other multimedia information\n(e.g., images) somewhat related to the content, we try to indirectly estimate\nthe relevance between two languages. Using multimedia as the ""pivot"", we\nproject all modalities into one common hidden space where samples belonging to\nsimilar semantic concepts should come close to each other, whatever the\nobserved space of each sample is. This modality-agnostic representation is the\nkey to bridging the gap between different modalities. Putting a decoder on top\nof it, our network can flexibly draw the outputs from any input modality.\nNotably, in the testing phase, we need only source language texts as the input\nfor translation. In experiments, we tested our method on two benchmarks to show\nthat it can achieve reasonable translation performance. We compared and\ninvestigated several possible implementations and found that an end-to-end\nmodel that simultaneously optimized both rank loss in multimodal encoders and\ncross-entropy loss in decoders performed the best.', 'Pair-based metric learning has been widely adopted to learn sentence\nembedding in many NLP tasks such as semantic text similarity due to its\nefficiency in computation. Most existing works employed a sequence encoder\nmodel and utilized limited sentence pairs with a pair-based loss to learn\ndiscriminating sentence representation. However, it is known that the sentence\nrepresentation can be biased when the sampled sentence pairs deviate from the\ntrue distribution of all sentence pairs. In this paper, our theoretical\nanalysis shows that existing works severely suffered from a good pair sampling\nand instance weighting strategy. Instead of one time pair selection and\nlearning on equal weighted pairs, we propose a unified locality weighting and\nlearning framework to learn task-specific sentence embedding. Our model,\nSentPWNet, exploits the neighboring spatial distribution of each sentence as\nlocality weight to indicate the informative level of sentence pair. Such weight\nis updated along with pair-loss optimization in each round, ensuring the model\nkeep learning the most informative sentence pairs. Extensive experiments on\nfour public available datasets and a self-collected place search benchmark with\n1.4 million places clearly demonstrate that our model consistently outperforms\nexisting sentence embedding methods with comparable efficiency.', 'While most machine translation systems to date are trained on large parallel\ncorpora, humans learn language in a different way: by being grounded in an\nenvironment and interacting with other humans. In this work, we propose a\ncommunication game where two agents, native speakers of their own respective\nlanguages, jointly learn to solve a visual referential task. We find that the\nability to understand and translate a foreign language emerges as a means to\nachieve shared goals. The emergent translation is interactive and multimodal,\nand crucially does not require parallel corpora, but only monolingual,\nindependent text and corresponding images. Our proposed translation model\nachieves this by grounding the source and target languages into a shared visual\nmodality, and outperforms several baselines on both word-level and\nsentence-level translation tasks. Furthermore, we show that agents in a\nmultilingual community learn to translate better and faster than in a bilingual\ncommunication setting.', 'Image captioning systems have recently improved dramatically, but they still\ntend to produce captions that are insensitive to the communicative goals that\ncaptions should meet. To address this, we propose Issue-Sensitive Image\nCaptioning (ISIC). In ISIC, a captioning system is given a target image and an\nissue, which is a set of images partitioned in a way that specifies what\ninformation is relevant. The goal of the captioner is to produce a caption that\nresolves this issue. To model this task, we use an extension of the Rational\nSpeech Acts model of pragmatic language use. Our extension is built on top of\nstate-of-the-art pretrained neural image captioners and explicitly reasons\nabout issues in our sense. We establish experimentally that these models\ngenerate captions that are both highly descriptive and issue-sensitive, and we\nshow how ISIC can complement and enrich the related task of Visual Question\nAnswering.', 'Automatically generating natural language descriptions from an image is a\nchallenging problem in artificial intelligence that requires a good\nunderstanding of the visual and textual signals and the correlations between\nthem. The state-of-the-art methods in image captioning struggles to approach\nhuman level performance, especially when data is limited. In this paper, we\npropose to improve the performance of the state-of-the-art image captioning\nmodels by incorporating two sources of prior knowledge: (i) a conditional\nlatent topic attention, that uses a set of latent variables (topics) as an\nanchor to generate highly probable words and, (ii) a regularization technique\nthat exploits the inductive biases in syntactic and semantic structure of\ncaptions and improves the generalization of image captioning models. Our\nexperiments validate that our method produces more human interpretable captions\nand also leads to significant improvements on the MSCOCO dataset in both the\nfull and low data regimes.', 'Image captioning models have been able to generate grammatically correct and\nhuman understandable sentences. However most of the captions convey limited\ninformation as the model used is trained on datasets that do not caption all\npossible objects existing in everyday life. Due to this lack of prior\ninformation most of the captions are biased to only a few objects present in\nthe scene, hence limiting their usage in daily life. In this paper, we attempt\nto show the biased nature of the currently existing image captioning models and\npresent a new image captioning dataset, Egoshots, consisting of 978 real life\nimages with no captions. We further exploit the state of the art pre-trained\nimage captioning and object recognition networks to annotate our images and\nshow the limitations of existing works. Furthermore, in order to evaluate the\nquality of the generated captions, we propose a new image captioning metric,\nobject based Semantic Fidelity (SF). Existing image captioning metrics can\nevaluate a caption only in the presence of their corresponding annotations;\nhowever, SF allows evaluating captions generated for images without\nannotations, making it highly useful for real life generated captions.', 'Every person speaks or writes their own flavor of their native language,\ninfluenced by a number of factors: the content they tend to talk about, their\ngender, their social status, or their geographical origin.\n  When attempting to perform Machine Translation (MT), these variations have a\nsignificant effect on how the system should perform translation, but this is\nnot captured well by standard one-size-fits-all models.\n  In this paper, we propose a simple and parameter-efficient adaptation\ntechnique that only requires adapting the bias of the output softmax to each\nparticular user of the MT system, either directly or through a factored\napproximation.\n  Experiments on TED talks in three languages demonstrate improvements in\ntranslation accuracy, and better reflection of speaker traits in the target\ntext.', ""Dialogue systems attempt to facilitate conversations between humans and\ncomputers, for purposes as diverse as small talk to booking a vacation. We are\nhere inspired by the performance of the recurrent neural network-based model\nSequicity, which when conducting a dialogue uses a sequence-to-sequence\narchitecture to first produce a textual representation of what is going on in\nthe dialogue, and in a further step use this along with database findings to\nproduce a reply to the user. We here propose a dialogue system based on the\nTransformer architecture instead of Sequicity's RNN-based architecture, that\nworks similarly in an end-to-end, sequence-to-sequence fashion."", 'In this paper, we propose a multi-modal search engine for interior design\nthat combines visual and textual queries. The goal of our engine is to retrieve\ninterior objects, e.g. furniture or wall clocks, that share visual and\naesthetic similarities with the query. Our search engine allows the user to\ntake a photo of a room and retrieve with a high recall a list of items\nidentical or visually similar to those present in the photo. Additionally, it\nallows to return other items that aesthetically and stylistically fit well\ntogether. To achieve this goal, our system blends the results obtained using\ntextual and visual modalities. Thanks to this blending strategy, we increase\nthe average style similarity score of the retrieved items by 11%. Our work is\nimplemented as a Web-based application and it is planned to be opened to the\npublic.', ""Automatic image cropping techniques are commonly used to enhance the\naesthetic quality of an image; they do it by detecting the most beautiful or\nthe most salient parts of the image and removing the unwanted content to have a\nsmaller image that is more visually pleasing. In this thesis, I introduce an\nadditional dimension to the problem of cropping, semantics. I argue that image\ncropping can also enhance the image's relevancy for a given entity by using the\nsemantic information contained in the image. I call this problem, Semantic\nImage Cropping. To support my argument, I provide a new dataset containing 100\nimages with at least two different entities per image and four ground truth\ncroppings collected using Amazon Mechanical Turk. I use this dataset to show\nthat state-of-the-art cropping algorithms that only take into account\naesthetics do not perform well in the problem of semantic image cropping.\nAdditionally, I provide a new deep learning system that takes not just\naesthetics but also semantics into account to generate image croppings, and I\nevaluate its performance using my new semantic cropping dataset, showing that\nusing the semantic information of an image can help to produce better\ncroppings."", 'Sentence embedding methods offer a powerful approach for working with short\ntextual constructs or sequences of words. By representing sentences as dense\nnumerical vectors, many natural language processing (NLP) applications have\nimproved their performance. However, relatively little is understood about the\nlatent structure of sentence embeddings. Specifically, research has not\naddressed whether the length and structure of sentences impact the sentence\nembedding space and topology. This paper reports research on a set of\ncomprehensive clustering and network analyses targeting sentence and\nsub-sentence embedding spaces. Results show that one method generates the most\nclusterable embeddings. In general, the embeddings of span sub-sentences have\nbetter clustering properties than the original sentences. The results have\nimplications for future sentence embedding models and applications.', 'Natural language understanding involves reading between the lines with\nimplicit background knowledge. Current systems either rely on pre-trained\nlanguage models as the sole implicit source of world knowledge, or resort to\nexternal knowledge bases (KBs) to incorporate additional relevant knowledge. We\npropose an unsupervised framework based on self-talk as a novel alternative to\nmultiple-choice commonsense tasks. Inspired by inquiry-based discovery learning\n(Bruner, 1961), our approach inquires language models with a number of\ninformation seeking questions such as ""$\\textit{what is the definition of\n...}$"" to discover additional background knowledge. Empirical results\ndemonstrate that the self-talk procedure substantially improves the performance\nof zero-shot language model baselines on four out of six commonsense\nbenchmarks, and competes with models that obtain knowledge from external KBs.\nWhile our approach improves performance on several benchmarks, the self-talk\ninduced knowledge even when leading to correct answers is not always seen as\nuseful by human judges, raising interesting questions about the inner-workings\nof pre-trained language models for commonsense reasoning.', 'Data augmentation is an effective performance enhancement in neural machine\ntranslation (NMT) by generating additional bilingual data. In this paper, we\npropose a novel data augmentation enhancement strategy for neural machine\ntranslation. Different from existing data augmentation methods which simply\nchoose words with the same probability across different sentences for\nmodification, we set sentence-specific probability for word selection by\nconsidering their roles in sentence. We use dependency parse tree of input\nsentence as an effective clue to determine selecting probability for every\nwords in each sentence. Our proposed method is evaluated on WMT14\nEnglish-to-German dataset and IWSLT14 German-to-English dataset. The result of\nextensive experiments show our proposed syntax-aware data augmentation method\nmay effectively boost existing sentence-independent methods for significant\ntranslation performance improvement.', 'Artificial agents today can answer factual questions. But they fall short on\nquestions that require common sense reasoning. Perhaps this is because most\nexisting common sense databases rely on text to learn and represent knowledge.\nBut much of common sense knowledge is unwritten - partly because it tends not\nto be interesting enough to talk about, and partly because some common sense is\nunnatural to articulate in text. While unwritten, it is not unseen. In this\npaper we leverage semantic common sense knowledge learned from images - i.e.\nvisual common sense - in two textual tasks: fill-in-the-blank and visual\nparaphrasing. We propose to ""imagine"" the scene behind the text, and leverage\nvisual cues from the ""imagined"" scenes in addition to textual cues while\nanswering these questions. We imagine the scenes as a visual abstraction. Our\napproach outperforms a strong text-only baseline on these tasks. Our proposed\ntasks can serve as benchmarks to quantitatively evaluate progress in solving\ntasks that go ""beyond recognition"". Our code and datasets are publicly\navailable.', 'Everyone knows that thousand of words are represented by a single image. As a\nresult image search has become a very popular mechanism for the Web searchers.\nImage search means, the search results are produced by the search engine should\nbe a set of images along with their Web page Unified Resource Locator. Now Web\nsearcher can perform two types of image search, they are Text to Image and\nImage to Image search. In Text to Image search, search query should be a text.\nBased on the input text data system will generate a set of images along with\ntheir Web page URL as an output. On the other hand, in Image to Image search,\nsearch query should be an image and based on this image system will generate a\nset of images along with their Web page URL as an output. According to the\ncurrent scenarios, Text to Image search mechanism always not returns perfect\nresult. It matches the text data and then displays the corresponding images as\nan output, which is not always perfect. To resolve this problem, Web\nresearchers have introduced the Image to Image search mechanism. In this paper,\nwe have also proposed an alternate approach of Image to Image search mechanism\nusing Histogram.', 'When writing a summary, humans tend to choose content from one or two\nsentences and merge them into a single summary sentence. However, the\nmechanisms behind the selection of one or multiple source sentences remain\npoorly understood. Sentence fusion assumes multi-sentence input; yet sentence\nselection methods only work with single sentences and not combinations of them.\nThere is thus a crucial gap between sentence selection and fusion to support\nsummarizing by both compressing single sentences and fusing pairs. This paper\nattempts to bridge the gap by ranking sentence singletons and pairs together in\na unified space. Our proposed framework attempts to model human methodology by\nselecting either a single sentence or a pair of sentences, then compressing or\nfusing the sentence(s) to produce a summary sentence. We conduct extensive\nexperiments on both single- and multi-document summarization datasets and\nreport findings on sentence selection and abstraction.', 'Sentence Split and Rephrase aims to break down a complex sentence into\nseveral simple sentences with its meaning preserved. Previous studies tend to\naddress the issue by seq2seq learning from parallel sentence pairs, which takes\na complex sentence as input and sequentially generates a series of simple\nsentences. However, the conventional seq2seq learning has two limitations for\nthis task: (1) it does not take into account the facts stated in the long\nsentence; As a result, the generated simple sentences may miss or inaccurately\nstate the facts in the original sentence. (2) The order variance of the simple\nsentences to be generated may confuse the seq2seq model during training because\nthe simple sentences derived from the long source sentence could be in any\norder.\n  To overcome the challenges, we first propose the Fact-aware Sentence\nEncoding, which enables the model to learn facts from the long sentence and\nthus improves the precision of sentence split; then we introduce Permutation\nInvariant Training to alleviate the effects of order variance in seq2seq\nlearning for this task. Experiments on the WebSplit-v1.0 benchmark dataset show\nthat our approaches can largely improve the performance over the previous\nseq2seq learning approaches. Moreover, an extrinsic evaluation on oie-benchmark\nverifies the effectiveness of our approaches by an observation that splitting\nlong sentences with our state-of-the-art model as preprocessing is helpful for\nimproving OpenIE performance.', ""Prior work on training generative Visual Dialog models with reinforcement\nlearning(Das et al.) has explored a Qbot-Abot image-guessing game and shown\nthat this 'self-talk' approach can lead to improved performance at the\ndownstream dialog-conditioned image-guessing task. However, this improvement\nsaturates and starts degrading after a few rounds of interaction, and does not\nlead to a better Visual Dialog model. We find that this is due in part to\nrepeated interactions between Qbot and Abot during self-talk, which are not\ninformative with respect to the image. To improve this, we devise a simple\nauxiliary objective that incentivizes Qbot to ask diverse questions, thus\nreducing repetitions and in turn enabling Abot to explore a larger state space\nduring RL ie. be exposed to more visual concepts to talk about, and varied\nquestions to answer. We evaluate our approach via a host of automatic metrics\nand human studies, and demonstrate that it leads to better dialog, ie. dialog\nthat is more diverse (ie. less repetitive), consistent (ie. has fewer\nconflicting exchanges), fluent (ie. more human-like),and detailed, while still\nbeing comparably image-relevant as prior work and ablations."", 'Recently, the progress of learning-by-synthesis has proposed a training model\nfor synthetic images, which can effectively reduce the cost of human and\nmaterial resources. However, due to the different distribution of synthetic\nimages compared to real images, the desired performance cannot still be\nachieved. Real images consist of multiple forms of light orientation, while\nsynthetic images consist of a uniform light orientation. These features are\nconsidered to be characteristic of outdoor and indoor scenes, respectively. To\nsolve this problem, the previous method learned a model to improve the realism\nof the synthetic image. Different from the previous methods, this paper takes\nthe first step to purify real images. Through the style transfer task, the\ndistribution of outdoor real images is converted into indoor synthetic images,\nthereby reducing the influence of light. Therefore, this paper proposes a\nreal-time style transfer network that preserves image content information (eg,\ngaze direction, pupil center position) of an input image (real image) while\ninferring style information (eg, image color structure, semantic features) of\nstyle image (synthetic image). In addition, the network accelerates the\nconvergence speed of the model and adapts to multi-scale images. Experiments\nwere performed using mixed studies (qualitative and quantitative) methods to\ndemonstrate the possibility of purifying real images in complex directions.\nQualitatively, it compares the proposed method with the available methods in a\nseries of indoor and outdoor scenarios of the LPW dataset. In quantitative\nterms, it evaluates the purified image by training a gaze estimation model on\nthe cross data set. The results show a significant improvement over the\nbaseline method compared to the raw real image.', 'Unsupervised image captioning is a challenging task that aims at generating\ncaptions without the supervision of image-sentence pairs, but only with images\nand sentences drawn from different sources and object labels detected from the\nimages. In previous work, pseudo-captions, i.e., sentences that contain the\ndetected object labels, were assigned to a given image. The focus of the\nprevious work was on the alignment of input images and pseudo-captions at the\nsentence level. However, pseudo-captions contain many words that are irrelevant\nto a given image. In this work, we investigate the effect of removing\nmismatched words from image-sentence alignment to determine how they make this\ntask difficult. We propose a simple gating mechanism that is trained to align\nimage features with only the most reliable words in pseudo-captions: the\ndetected object labels. The experimental results show that our proposed method\noutperforms the previous methods without introducing complex sentence-level\nlearning objectives. Combined with the sentence-level alignment method of\nprevious work, our method further improves its performance. These results\nconfirm the importance of careful alignment in word-level details.', 'My overarching research goal is to provide robots with perceptional abilities\nthat allow interactions with humans in a human-like manner. To develop these\nperceptional abilities, I believe that it is useful to study the principles of\nthe human visual system. I use these principles to develop new computer vision\nalgorithms and validate their effectiveness in intelligent robotic systems. I\nam enthusiastic about this approach as it offers the dual benefit of uncovering\nprinciples inherent in the human visual system, as well as applying these\nprinciples to its artificial counterpart. Fig. 1 contains a depiction of my\nresearch.', 'Images of natural systems may represent patterns of network-like structure,\nwhich could reveal important information about the topological properties of\nthe underlying subject. However, the image itself does not automatically\nprovide a formal definition of a network in terms of sets of nodes and edges.\nInstead, this information should be suitably extracted from the raw image data.\nMotivated by this, we present a principled model to extract network topologies\nfrom images that is scalable and efficient. We map this goal into solving a\nrouting optimization problem where the solution is a network that minimizes an\nenergy function which can be interpreted in terms of an operational and\ninfrastructural cost. Our method relies on recent results from optimal\ntransport theory and is a principled alternative to standard image-processing\ntechniques that are based on heuristics. We test our model on real images of\nthe retinal vascular system, slime mold and river networks and compare with\nroutines combining image-processing techniques. Results are tested in terms of\na similarity measure related to the amount of information preserved in the\nextraction. We find that our model finds networks from retina vascular network\nimages that are more similar to hand-labeled ones, while also giving high\nperformance in extracting networks from images of rivers and slime mold for\nwhich there is no ground truth available. While there is no unique method that\nfits all the images the best, our approach performs consistently across\ndatasets, its algorithmic implementation is efficient and can be fully\nautomatized to be run on several datasets with little supervision.', 'Neural machine translation (NMT) systems are usually trained on a large\namount of bilingual sentence pairs and translate one sentence at a time,\nignoring inter-sentence information. This may make the translation of a\nsentence ambiguous or even inconsistent with the translations of neighboring\nsentences. In order to handle this issue, we propose an inter-sentence gate\nmodel that uses the same encoder to encode two adjacent sentences and controls\nthe amount of information flowing from the preceding sentence to the\ntranslation of the current sentence with an inter-sentence gate. In this way,\nour proposed model can capture the connection between sentences and fuse\nrecency from neighboring sentences into neural machine translation. On several\nNIST Chinese-English translation tasks, our experiments demonstrate that the\nproposed inter-sentence gate model achieves substantial improvements over the\nbaseline.', 'A desideratum of high-quality translation systems is that they preserve\nmeaning, in the sense that two sentences with different meanings should not\ntranslate to one and the same sentence in another language. However,\nstate-of-the-art systems often fail in this regard, particularly in cases where\nthe source and target languages partition the ""meaning space"" in different\nways. For instance, ""I cut my finger."" and ""I cut my finger off."" describe\ndifferent states of the world but are translated to French (by both Fairseq and\nGoogle Translate) as ""Je me suis coupe le doigt."", which is ambiguous as to\nwhether the finger is detached. More generally, translation systems are\ntypically many-to-one (non-injective) functions from source to target language,\nwhich in many cases results in important distinctions in meaning being lost in\ntranslation. Building on Bayesian models of informative utterance production,\nwe present a method to define a less ambiguous translation system in terms of\nan underlying pre-trained neural sequence-to-sequence model. This method\nincreases injectivity, resulting in greater preservation of meaning as measured\nby improvement in cycle-consistency, without impeding translation quality\n(measured by BLEU score).', 'Language recognition has been significantly advanced in recent years by means\nof modern machine learning methods such as deep learning and benchmarks with\nrich annotations. However, research is still limited in low-resource formal\nlanguages. This consists of a significant gap in describing the colloquial\nlanguage especially for low-resourced ones such as Persian. In order to target\nthis gap for low resource languages, we propose a ""Large Scale Colloquial\nPersian Dataset"" (LSCP). LSCP is hierarchically organized in a semantic\ntaxonomy that focuses on multi-task informal Persian language understanding as\na comprehensive problem. This encompasses the recognition of multiple semantic\naspects in the human-level sentences, which naturally captures from the\nreal-world sentences. We believe that further investigations and processing, as\nwell as the application of novel algorithms and methods, can strengthen\nenriching computerized understanding and processing of low resource languages.\nThe proposed corpus consists of 120M sentences resulted from 27M tweets\nannotated with parsing tree, part-of-speech tags, sentiment polarity and\ntranslation in five different languages.', ""Classifying the human emotion through facial expressions is a big topic in\nboth the Computer Vision and Deep learning fields. Human emotion can be\nclassified as one of the basic emotion types like being angry, happy or\ndimensional emotion with valence and arousal values. There are a lot of related\nchallenges in this topic, one of the most famous challenges is called the\n'Affect-in-the-wild Challenge'(Aff-Wild Challenge). It is the first challenge\non the estimation of valence and arousal in-the-wild. This project is an\nextension of the Aff-wild Challenge. Aff-wild database was created using images\nwith a mean resolution of 607*359, I and Dimitrios sought to find out the\nperformance of the model that is trained on a database that contains4K\nresolution in-the-wild images. Since there is no existing database to satisfy\nthe requirement, I built this database from scratch with help from Dimitrios\nand trained neural network models with different hyperparameters on this\ndatabase. I used network models likeVGG16, AlexNet, ResNet and also some\npre-trained models like Ima-geNet VGG. I compared the results of the different\nnetwork models alongside the results from the Aff-wild database to exploit the\noptimal model for my database."", 'A relation tuple consists of two entities and the relation between them, and\noften such tuples are found in unstructured text. There may be multiple\nrelation tuples present in a text and they may share one or both entities among\nthem. Extracting such relation tuples from a sentence is a difficult task and\nsharing of entities or overlapping entities among the tuples makes it more\nchallenging. Most prior work adopted a pipeline approach where entities were\nidentified first followed by finding the relations among them, thus missing the\ninteraction among the relation tuples in a sentence. In this paper, we propose\ntwo approaches to use encoder-decoder architecture for jointly extracting\nentities and relations. In the first approach, we propose a representation\nscheme for relation tuples which enables the decoder to generate one word at a\ntime like machine translation models and still finds all the tuples present in\na sentence with full entity names of different length and with overlapping\nentities. Next, we propose a pointer network-based decoding approach where an\nentire tuple is generated at every time step. Experiments on the publicly\navailable New York Times corpus show that our proposed approaches outperform\nprevious work and achieve significantly higher F1 scores.', 'The number of user reviews of tourist attractions, restaurants, mobile apps,\netc. is increasing for all languages; yet, research is lacking on how reviews\nin multiple languages should be aggregated and displayed. Speakers of different\nlanguages may have consistently different experiences, e.g., different\ninformation available in different languages at tourist attractions or\ndifferent user experiences with software due to\ninternationalization/localization choices. This paper assesses the similarity\nin the ratings given by speakers of different languages to London tourist\nattractions on TripAdvisor. The correlations between different languages are\ngenerally high, but some language pairs are more correlated than others. The\nresults question the common practice of computing average ratings from reviews\nin many languages.', 'Emotion stimulus extraction is a fine-grained subtask of emotion analysis\nthat focuses on identifying the description of the cause behind an emotion\nexpression from a text passage (e.g., in the sentence ""I am happy that I passed\nmy exam"" the phrase ""passed my exam"" corresponds to the stimulus.). Previous\nwork mainly focused on Mandarin and English, with no resources or models for\nGerman. We fill this research gap by developing a corpus of 2006 German news\nheadlines annotated with emotions and 811 instances with annotations of\nstimulus phrases. Given that such corpus creation efforts are time-consuming\nand expensive, we additionally work on an approach for projecting the existing\nEnglish GoodNewsEveryone (GNE) corpus to a machine-translated German version.\nWe compare the performance of a conditional random field (CRF) model (trained\nmonolingually on German and cross-lingually via projection) with a multilingual\nXLM-RoBERTa (XLM-R) model. Our results show that training with the German\ncorpus achieves higher F1 scores than projection. Experiments with XLM-R\noutperform their respective CRF counterparts.', 'Multilingual neural machine translation (NMT), which translates multiple\nlanguages using a single model, is of great practical importance due to its\nadvantages in simplifying the training process, reducing online maintenance\ncosts, and enhancing low-resource and zero-shot translation. Given there are\nthousands of languages in the world and some of them are very different, it is\nextremely burdensome to handle them all in a single model or use a separate\nmodel for each language pair. Therefore, given a fixed resource budget, e.g.,\nthe number of models, how to determine which languages should be supported by\none model is critical to multilingual NMT, which, unfortunately, has been\nignored by previous work. In this work, we develop a framework that clusters\nlanguages into different groups and trains one multilingual model for each\ncluster. We study two methods for language clustering: (1) using prior\nknowledge, where we cluster languages according to language family, and (2)\nusing language embedding, in which we represent each language by an embedding\nvector and cluster them in the embedding space. In particular, we obtain the\nembedding vectors of all the languages by training a universal neural machine\ntranslation model. Our experiments on 23 languages show that the first\nclustering method is simple and easy to understand but leading to suboptimal\ntranslation accuracy, while the second method sufficiently captures the\nrelationship among languages well and improves the translation accuracy for\nalmost all the languages over baseline methods', ""Recent advances in natural language processing (NLP) have the ability to\ntransform how classroom learning takes place. Combined with the increasing\nintegration of technology in today's classrooms, NLP systems leveraging\nquestion answering and dialog processing techniques can serve as private tutors\nor participants in classroom discussions to increase student engagement and\nlearning. To progress towards this goal, we use the classroom discourse\nframework of academically productive talk (APT) to learn strategies that make\nfor the best learning experience. In this paper, we introduce a new task,\ncalled future talk move prediction (FTMP): it consists of predicting the next\ntalk move -- an utterance strategy from APT -- given a conversation history\nwith its corresponding talk moves. We further introduce a neural network model\nfor this task, which outperforms multiple baselines by a large margin. Finally,\nwe compare our model's performance on FTMP to human performance and show\nseveral similarities between the two."", 'In recent years online shopping has gained momentum and became an important\nvenue for customers wishing to save time and simplify their shopping process. A\nkey advantage of shopping online is the ability to read what other customers\nare saying about products of interest. In this work, we aim to maintain this\nadvantage in situations where extreme brevity is needed, for example, when\nshopping by voice. We suggest a novel task of extracting a single\nrepresentative helpful sentence from a set of reviews for a given product. The\nselected sentence should meet two conditions: first, it should be helpful for a\npurchase decision and second, the opinion it expresses should be supported by\nmultiple reviewers. This task is closely related to the task of Multi Document\nSummarization in the product reviews domain but differs in its objective and\nits level of conciseness. We collect a dataset in English of sentence\nhelpfulness scores via crowd-sourcing and demonstrate its reliability despite\nthe inherent subjectivity involved. Next, we describe a complete model that\nextracts representative helpful sentences with positive and negative sentiment\ntowards the product and demonstrate that it outperforms several baselines.', 'Recently, there is a rising interest in perceiving image aesthetics. The\nexisting works deal with image aesthetics as a classification or regression\nproblem. To extend the cognition from rating to reasoning, a deeper\nunderstanding of aesthetics should be based on revealing why a high- or\nlow-aesthetic score should be assigned to an image. From such a point of view,\nwe propose a model referred to as Neural Aesthetic Image Reviewer, which can\nnot only give an aesthetic score for an image, but also generate a textual\ndescription explaining why the image leads to a plausible rating score.\nSpecifically, we propose two multi-task architectures based on shared\naesthetically semantic layers and task-specific embedding layers at a high\nlevel for performance improvement on different tasks. To facilitate researches\non this problem, we collect the AVA-Reviews dataset, which contains 52,118\nimages and 312,708 comments in total. Through multi-task learning, the proposed\nmodels can rate aesthetic images as well as produce comments in an end-to-end\nmanner. It is confirmed that the proposed models outperform the baselines\naccording to the performance evaluation on the AVA-Reviews dataset. Moreover,\nwe demonstrate experimentally that our model can generate textual reviews\nrelated to aesthetics, which are consistent with human perception.', 'Our world is open-ended, non-stationary, and constantly evolving; thus what\nwe talk about and how we talk about it change over time. This inherent dynamic\nnature of language contrasts with the current static language modelling\nparadigm, which trains and evaluates models on utterances from overlapping time\nperiods. Despite impressive recent progress, we demonstrate that Transformer-XL\nlanguage models perform worse in the realistic setup of predicting future\nutterances from beyond their training period, and that model performance\nbecomes increasingly worse with time. We find that, while increasing model size\nalone -- a key driver behind recent progress -- does not solve this problem,\nhaving models that continually update their knowledge with new information can\nindeed mitigate this performance degradation over time. Hence, given the\ncompilation of ever-larger language modelling datasets, combined with the\ngrowing list of language-model-based NLP applications that require up-to-date\nfactual knowledge about the world, we argue that now is the right time to\nrethink the static way in which we currently train and evaluate our language\nmodels, and develop adaptive language models that can remain up-to-date with\nrespect to our ever-changing and non-stationary world. We publicly release our\ndynamic, streaming language modelling benchmarks for WMT and arXiv to\nfacilitate language model evaluation that takes temporal dynamics into account.', 'Training a deep network to perform semantic segmentation requires large\namounts of labeled data. To alleviate the manual effort of annotating real\nimages, researchers have investigated the use of synthetic data, which can be\nlabeled automatically. Unfortunately, a network trained on synthetic data\nperforms relatively poorly on real images. While this can be addressed by\ndomain adaptation, existing methods all require having access to real images\nduring training. In this paper, we introduce a drastically different way to\nhandle synthetic images that does not require seeing any real images at\ntraining time. Our approach builds on the observation that foreground and\nbackground classes are not affected in the same manner by the domain shift, and\nthus should be treated differently. In particular, the former should be handled\nin a detection-based manner to better account for the fact that, while their\ntexture in synthetic images is not photo-realistic, their shape looks natural.\nOur experiments evidence the effectiveness of our approach on Cityscapes and\nCamVid with models trained on synthetic data only.', ""Classroom discussions in English Language Arts have a positive effect on\nstudents' reading, writing and reasoning skills. Although prior work has\nlargely focused on teacher talk and student-teacher interactions, we focus on\nthree theoretically-motivated aspects of high-quality student talk:\nargumentation, specificity, and knowledge domain. We introduce an annotation\nscheme, then show that the scheme can be used to produce reliable annotations\nand that the annotations are predictive of discussion quality. We also\nhighlight opportunities provided by our scheme for education and natural\nlanguage processing research."", 'In this paper, we pose the question: do people talk about women and men in\ndifferent ways? We introduce two datasets and a novel integration of approaches\nfor automatically inferring gender associations from language, discovering\ncoherent word clusters, and labeling the clusters for the semantic concepts\nthey represent. The datasets allow us to compare how people write about women\nand men in two different settings - one set draws from celebrity news and the\nother from student reviews of computer science professors. We demonstrate that\nthere are large-scale differences in the ways that people talk about women and\nmen and that these differences vary across domains. Human evaluations show that\nour methods significantly outperform strong baselines.', ""Convolututional Neural Networks have achieved state of the art in image\nclassification, object detection and other image related tasks. In this paper I\npresent another use of CNNs i.e. if given a set of images and then giving a\nsingle test image the network identifies that the test image is part of which\nimage from the images given before. This is a task somehow similar to measuring\nimage similarity and can be done using a simple CNN. Doing this task manually\nby looping can be quite a time consuming problem and won't be a generalizable\nsolution. The task is quite similar to doing object detection but for that lots\ntraining data should be given or in the case of sliding window it takes lot of\ntime and my algorithm can work with much fewer examples, is totally\nunsupervised and works much efficiently. Also, I explain that how unsupervised\nalgorithm like K-Means or supervised algorithm like K-NN are not good enough to\nperform this task. The basic idea is that image encodings are collected for\neach image from a CNN, when a test image comes it is replaced by a part of\noriginal image, the encoding is generated using the same network, the frobenius\nnorm is calculated and if it comes under a tolerance level then the test image\nis said to be the part of the original image."", 'Human-performed annotation of sentences in legal documents is an important\nprerequisite to many machine learning based systems supporting legal tasks.\nTypically, the annotation is done sequentially, sentence by sentence, which is\noften time consuming and, hence, expensive. In this paper, we introduce a\nproof-of-concept system for annotating sentences ""laterally."" The approach is\nbased on the observation that sentences that are similar in meaning often have\nthe same label in terms of a particular type system. We use this observation in\nallowing annotators to quickly view and annotate sentences that are\nsemantically similar to a given sentence, across an entire corpus of documents.\nHere, we present the interface of the system and empirically evaluate the\napproach. The experiments show that lateral annotation has the potential to\nmake the annotation process quicker and more consistent.', ""This paper attempts to connects the evolution of computer languages with the\nevolution of life, where the later has been dictated by \\emph{theory of\nevolution of species}, and tries to give supportive evidence that the new\nlanguages are more robust than the previous, carry-over the mixed features of\nolder languages, such that strong features gets added into them and weak\nfeatures of older languages gets removed. In addition, an analysis of most\nprominent programming languages is presented, emphasizing on how the features\nof existing languages have influenced the development of new programming\nlanguages. At the end, it suggests a set of experimental languages, which may\nrule the world of programming languages in the time of new multi-core\narchitectures.\n  Index terms- Programming languages' evolution, classifications of languages,\nfuture languages, scripting-languages."", 'To date, most state-of-the-art sequence modeling architectures use attention\nto build generative models for language based tasks. Some of these models use\nall the available sequence tokens to generate an attention distribution which\nresults in time complexity of $O(n^2)$. Alternatively, they utilize depthwise\nconvolutions with softmax normalized kernels of size $k$ acting as a\nlimited-window self-attention, resulting in time complexity of $O(k{\\cdot}n)$.\nIn this paper, we introduce Time-aware Large Kernel (TaLK) Convolutions, a\nnovel adaptive convolution operation that learns to predict the size of a\nsummation kernel instead of using a fixed-sized kernel matrix. This method\nyields a time complexity of $O(n)$, effectively making the sequence encoding\nprocess linear to the number of tokens. We evaluate the proposed method on\nlarge-scale standard machine translation, abstractive summarization and\nlanguage modeling datasets and show that TaLK Convolutions constitute an\nefficient improvement over other attention/convolution based approaches.', 'We propose a novel approach to synthesizing images that are effective for\ntraining object detectors. Starting from a small set of real images, our\nalgorithm estimates the rendering parameters required to synthesize similar\nimages given a coarse 3D model of the target object. These parameters can then\nbe reused to generate an unlimited number of training images of the object of\ninterest in arbitrary 3D poses, which can then be used to increase\nclassification performances.\n  A key insight of our approach is that the synthetically generated images\nshould be similar to real images, not in terms of image quality, but rather in\nterms of features used during the detector training. We show in the context of\ndrone, plane, and car detection that using such synthetically generated images\nyields significantly better performances than simply perturbing real images or\neven synthesizing images in such way that they look very realistic, as is often\ndone when only limited amounts of training data are available.', 'Automatic image captioning has recently approached human-level performance\ndue to the latest advances in computer vision and natural language\nunderstanding. However, most of the current models can only generate plain\nfactual descriptions about the content of a given image. However, for human\nbeings, image caption writing is quite flexible and diverse, where additional\nlanguage dimensions, such as emotion, humor and language styles, are often\nincorporated to produce diverse, emotional, or appealing captions. In\nparticular, we are interested in generating sentiment-conveying image\ndescriptions, which has received little attention. The main challenge is how to\neffectively inject sentiments into the generated captions without altering the\nsemantic matching between the visual content and the generated descriptions. In\nthis work, we propose two different models, which employ different schemes for\ninjecting sentiments into image captions. Compared with the few existing\napproaches, the proposed models are much simpler and yet more effective. The\nexperimental results show that our model outperform the state-of-the-art models\nin generating sentimental (i.e., sentiment-bearing) image captions. In\naddition, we can also easily manipulate the model by assigning different\nsentiments to the testing image to generate captions with the corresponding\nsentiments.', 'Recently we see a rising number of methods in the field of eXplainable\nArtificial Intelligence. To our surprise, their development is driven by model\ndevelopers rather than a study of needs for human end users. The analysis of\nneeds, if done, takes the form of an A/B test rather than a study of open\nquestions. To answer the question ""What would a human operator like to ask the\nML model?"" we propose a conversational system explaining decisions of the\npredictive model. In this experiment, we developed a chatbot called dr_ant to\ntalk about machine learning model trained to predict survival odds on Titanic.\nPeople can talk with dr_ant about different aspects of the model to understand\nthe rationale behind its predictions. Having collected a corpus of 1000+\ndialogues, we analyse the most common types of questions that users would like\nto ask. To our knowledge, it is the first study which uses a conversational\nsystem to collect the needs of human operators from the interactive and\niterative dialogue explorations of a predictive model.', 'My notes on Deep Learning for NLP.', 'In this dissertation, I present my work towards exploring temporal\ninformation for better video understanding. Specifically, I have worked on two\nproblems: action recognition and semantic segmentation. For action recognition,\nI have proposed a framework, termed hidden two-stream networks, to learn an\noptimal motion representation that does not require the computation of optical\nflow. My framework alleviates several challenges faced in video classification,\nsuch as learning motion representations, real-time inference, multi-framerate\nhandling, generalizability to unseen actions, etc. For semantic segmentation, I\nhave introduced a general framework that uses video prediction models to\nsynthesize new training samples. By scaling up the training dataset, my trained\nmodels are more accurate and robust than previous models even without\nmodifications to the network architectures or objective functions. I believe\nvideos have much more potential to be mined, and temporal information is one of\nthe most important cues for machines to perceive the visual world better.', 'There is a growing interest in investigating what neural NLP models learn\nabout language. A prominent open question is the question of whether or not it\nis necessary to model hierarchical structure. We present a linguistic\ninvestigation of a neural parser adding insights to this question. We look at\ntransitivity and agreement information of auxiliary verb constructions (AVCs)\nin comparison to finite main verbs (FMVs). This comparison is motivated by\ntheoretical work in dependency grammar and in particular the work of Tesni\\`ere\n(1959) where AVCs and FMVs are both instances of a nucleus, the basic unit of\nsyntax. An AVC is a dissociated nucleus, it consists of at least two words, and\nan FMV is its non-dissociated counterpart, consisting of exactly one word. We\nsuggest that the representation of AVCs and FMVs should capture similar\ninformation. We use diagnostic classifiers to probe agreement and transitivity\ninformation in vectors learned by a transition-based neural parser in four\ntypologically different languages. We find that the parser learns different\ninformation about AVCs and FMVs if only sequential models (BiLSTMs) are used in\nthe architecture but similar information when a recursive layer is used. We\nfind explanations for why this is the case by looking closely at how\ninformation is learned in the network and looking at what happens with\ndifferent dependency representations of AVCs. We conclude that there may be\nbenefits to using a recursive layer in dependency parsing and that we have not\nyet found the best way to integrate it in our parsers.', 'Current open-domain conversational models can easily be made to talk in\ninadequate ways. Online learning from conversational feedback given by the\nconversation partner is a promising avenue for a model to improve and adapt, so\nas to generate fewer of these safety failures. However, current\nstate-of-the-art models tend to react to feedback with defensive or oblivious\nresponses. This makes for an unpleasant experience and may discourage\nconversation partners from giving feedback in the future. This work proposes\nSaFeRDialogues, a task and dataset of graceful responses to conversational\nfeedback about safety failures. We collect a dataset of 10k dialogues\ndemonstrating safety failures, feedback signaling them, and a response\nacknowledging the feedback. We show how fine-tuning on this dataset results in\nconversations that human raters deem considerably more likely to lead to a\ncivil conversation, without sacrificing engagingness or general conversational\nability.', 'We argue that logical semantics might have faltered due to its failure in\ndistinguishing between two fundamentally very different types of concepts:\nontological concepts, that should be types in a strongly-typed ontology, and\nlogical concepts, that are predicates corresponding to properties of and\nrelations between objects of various ontological types. We will then show that\naccounting for these differences amounts to the integration of lexical and\ncompositional semantics in one coherent framework, and to an embedding in our\nlogical semantics of a strongly-typed ontology that reflects our commonsense\nview of the world and the way we talk about it in ordinary language. We will\nshow that in such a framework a number of challenges in natural language\nsemantics can be adequately and systematically treated.', 'In this paper, we are concerned with the detection of progressive dynamic\nsaliency from video sequences. More precisely, we are interested in saliency\nrelated to motion and likely to appear progressively over time. It can be\nrelevant to trigger alarms, to dedicate additional processing or to detect\nspecific events. Trajectories represent the best way to support progressive\ndynamic saliency detection. Accordingly, we will talk about trajectory\nsaliency. A trajectory will be qualified as salient if it deviates from normal\ntrajectories that share a common motion pattern related to a given context.\nFirst, we need a compact while discriminative representation of trajectories.\nWe adopt a (nearly) unsupervised learning-based approach. The latent code\nestimated by a recurrent auto-encoder provides the desired representation. In\naddition, we enforce consistency for normal (similar) trajectories through the\nauto-encoder loss function. The distance of the trajectory code to a prototype\ncode accounting for normality is the means to detect salient trajectories. We\nvalidate our trajectory saliency detection method on synthetic and real\ntrajectory datasets, and highlight the contributions of its different\ncomponents. We show that our method outperforms existing methods on several\nscenarios drawn from the publicly available dataset of pedestrian trajectories\nacquired in a railway station (Alahi 2014).', 'Our research topic is Human segmentation with static camera. This topic can\nbe divided into three sub-tasks, which are object detection, instance\nidentification and segmentation. These sub-tasks are three closely related\nsubjects. The development of each subject has great impact on the other two\nfields. In this literature review, we will first introduce the background of\nhuman segmentation and then talk about issues related to the above three fields\nas well as how they interact with each other.', ""Virtual Personal Assistants like Siri have great potential but such\ndevelopments hit the fundamental problem of how to make computational devices\nthat understand human speech. Natural language understanding is one of the more\ndisappointing failures of AI research and it seems there is something we\ncomputer scientists don't get about the nature of language. Of course\nphilosophers and linguists think quite differently about language and this\npaper describes how we have taken ideas from other disciplines and implemented\nthem. The background to the work is to take seriously the notion of language as\naction and look at what people actually do with language using the techniques\nof Conversation Analysis. The observation has been that human communication is\n(behind the scenes) about the management of social relations as well as the\n(foregrounded) passing of information. To claim this is one thing but to\nimplement it requires a mechanism. The mechanism described here is based on the\nnotion of language being intentional - we think intentionally, talk about them\nand recognise them in others - and cooperative in that we are compelled to help\nout. The way we are compelled points to a solution to the ever present problem\nof keeping the human on topic. The approach has led to a recent success in\nwhich we significantly improve user satisfaction independent of task\ncompletion. Talk Markup Language (TalkML) is a draft alternative to VoiceXML\nthat, we propose, greatly simplifies the scripting of interaction by providing\ndefault behaviours for no input and not recognised speech events."", ""How will my face look when I get older? Or, for a more challenging question:\nHow will my brain look when I get older? To answer this question one must\ndevise (and learn from data) a multivariate auto-regressive function which\ngiven an image and a desired target age generates an output image. While\ncollecting data for faces may be easier, collecting longitudinal brain data is\nnot trivial. We propose a deep learning-based method that learns to simulate\nsubject-specific brain ageing trajectories without relying on longitudinal\ndata. Our method synthesises images conditioned on two factors: age (a\ncontinuous variable), and status of Alzheimer's Disease (AD, an ordinal\nvariable). With an adversarial formulation we learn the joint distribution of\nbrain appearance, age and AD status, and define reconstruction losses to\naddress the challenging problem of preserving subject identity. We compare with\nseveral benchmarks using two widely used datasets. We evaluate the quality and\nrealism of synthesised images using ground-truth longitudinal data and a\npre-trained age predictor. We show that, despite the use of cross-sectional\ndata, our model learns patterns of gray matter atrophy in the middle temporal\ngyrus in patients with AD. To demonstrate generalisation ability, we train on\none dataset and evaluate predictions on the other. In conclusion, our model\nshows an ability to separate age, disease influence and anatomy using only 2D\ncross-sectional data that should be useful in large studies into\nneurodegenerative disease, that aim to combine several data sources. To\nfacilitate such future studies by the community at large our code is made\navailable at https://github.com/xiat0616/BrainAgeing."", 'Synthesizing images of a person in novel poses from a single image is a\nhighly ambiguous task. Most existing approaches require paired training images;\ni.e. images of the same person with the same clothing in different poses.\nHowever, obtaining sufficiently large datasets with paired data is challenging\nand costly. Previous methods that forego paired supervision lack realism. We\npropose a self-supervised framework named SPICE (Self-supervised Person Image\nCrEation) that closes the image quality gap with supervised methods. The key\ninsight enabling self-supervision is to exploit 3D information about the human\nbody in several ways. First, the 3D body shape must remain unchanged when\nreposing. Second, representing body pose in 3D enables reasoning about self\nocclusions. Third, 3D body parts that are visible before and after reposing,\nshould have similar appearance features. Once trained, SPICE takes an image of\na person and generates a new image of that person in a new target pose. SPICE\nachieves state-of-the-art performance on the DeepFashion dataset, improving the\nFID score from 29.9 to 7.8 compared with previous unsupervised methods, and\nwith performance similar to the state-of-the-art supervised method (6.4). SPICE\nalso generates temporally coherent videos given an input image and a sequence\nof poses, despite being trained on static images only.', 'Basic body-part names (BBPNs) were defined as body-part names in Swadesh\nbasic 200 words. Non-Mayan cognates of Mayan (MY) BBPNs were extensively\nsearched for, by comparing with non-MY vocabulary, including ca.1300 basic\nwords of 82 AN languages listed by Tryon (1985), etc. Thus found cognates (CGs)\nin non-MY are listed in Table 1, as classified by language groups to which most\nsimilar cognates (MSCs) of MY BBPNs belong. CGs of MY are classified to 23\nmutually unrelated CG-items, of which 17.5 CG-items have their MSCs in\nAustronesian (AN), giving its closest similarity score (CSS), CSS(AN) = 17.5,\nwhich consists of 10.33 MSCs in Formosan, 1.83 MSCs in Western\nMalayo-Polynesian (W.MP), 0.33 in Central MP, 0.0 in SHWNG, and 5.0 in Oceanic\n[i.e., CSS(FORM)= 10.33, CSS(W.MP) = 1.88, ..., CSS(OC)= 5.0]. These CSSs for\nlanguage (sub)groups are also listed in the underline portion of every section\nof (Section1 - Section 6) in Table 1. Chi-squar test (degree of freedom = 1)\nusing [Eq 1] and [Eqs.2] revealed that MSCs of MY BBPNs are distributed in\nFormosan in significantly higher frequency (P < 0.001) than in other subgroups\nof AN, as well as than in non-AN languages. MY is thus concluded to have been\nderived from Formosan of AN. Eskimo shows some BBPN similarities to FORM and\nMY.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am working on autonomous driving research and I need to minimize the risk of autonomous vehicles to the public, especially pedestrians. Therefore, I am trying to build a model that will inform my autonomous vehicles to avoid crowded streets where lots of people are walking. This model has to be deployed in real-time. My current idea is to use a computer vision model that can extract and count the number of pedestrians from surveillance footage videos. It needs to be a fast pedestrian detection model. How should I go about doing this? Should my model be able to process videos or simply just images ? If I use videos, the computational load would be higher but I could get a more accurate estimate of the walking speed of these pedestrians during a time interval. ","[""Detecting and predicting the behavior of pedestrians is extremely crucial for\nself-driving vehicles to plan and interact with them safely. Although there\nhave been several research works in this area, it is important to have fast and\nmemory efficient models such that it can operate in embedded hardware in these\nautonomous machines. In this work, we propose a novel architecture using\nspatial-temporal multi-tasking to do camera based pedestrian detection and\nintention prediction. Our approach significantly reduces the latency by being\nable to detect and predict all pedestrians' intention in a single shot manner\nwhile also being able to attain better accuracy by sharing features with\nrelevant object level information and interactions."", 'Studies of object detection and localization, particularly pedestrian\ndetection have received considerable attention in recent times due to its\nseveral prospective applications such as surveillance, driving assistance,\nautonomous cars, etc. Also, a significant trend of latest research studies in\nrelated problem areas is the use of sophisticated Deep Learning based\napproaches to improve the benchmark performance on various standard datasets. A\ntrade-off between the speed (number of video frames processed per second) and\ndetection accuracy has often been reported in the existing literature. In this\narticle, we present a new but simple deep learning based strategy for\npedestrian detection that improves this trade-off. Since training of similar\nmodels using publicly available sample datasets failed to improve the detection\nperformance to some significant extent, particularly for the instances of\npedestrians of smaller sizes, we have developed a new sample dataset consisting\nof more than 80K annotated pedestrian figures in videos recorded under varying\ntraffic conditions. Performance of the proposed model on the test samples of\nthe new dataset and two other existing datasets, namely Caltech Pedestrian\nDataset (CPD) and CityPerson Dataset (CD) have been obtained. Our proposed\nsystem shows nearly 16\\% improvement over the existing state-of-the-art result.', 'With the rise of self-driving vehicles comes the risk of accidents and the\nneed for higher safety, and protection for pedestrian detection in the\nfollowing scenarios: imminent crashes, thus the car should crash into an object\nand avoid the pedestrian, and in the case of road intersections, where it is\nimportant for the car to stop when pedestrians are crossing. Currently, a\nspecial topology of deep neural networks called Fused Deep Neural Network\n(F-DNN) is considered to be the state of the art in pedestrian detection, as it\nhas the lowest miss rate, yet it is very slow. Therefore, acceleration is\nneeded to speed up the performance. This project proposes two contributions to\naddress this problem, by using a deep neural network used for object detection,\ncalled Single Shot Multi-Box Detector (SSD). The first contribution is training\nand tuning the hyperparameters of SSD to improve pedestrian detection. The\nsecond contribution is a new FPGA design for accelerating the model on the\nAltera Arria 10 platform. The final system will be used in self-driving\nvehicles in real-time. Preliminary results of the improved SSD shows 3% higher\nmiss-rate than F-DNN on Caltech pedestrian detection benchmark, but 4x\nperformance improvement. The acceleration design is expected to achieve an\nadditional performance improvement significantly outweighing the minimal\ndifference in accuracy.', 'In this paper, we present a real-time robust multi-view pedestrian detection\nand tracking system for video surveillance using neural networks which can be\nused in dynamic environments. The proposed system consists of two phases:\nmulti-view pedestrian detection and tracking. First, pedestrian detection\nutilizes background subtraction to segment the foreground blob. An adaptive\nbackground subtraction method where each of the pixel of input image models as\na mixture of Gaussians and uses an on-line approximation to update the model\napplies to extract the foreground region. The Gaussian distributions are then\nevaluated to determine which are most likely to result from a background\nprocess. This method produces a steady, real-time tracker in outdoor\nenvironment that consistently deals with changes of lighting condition, and\nlong-term scene change. Second, the Tracking is performed at two phases:\npedestrian classification and tracking the individual subject. A sliding window\nis applied on foreground binary image to select an input window which is used\nfor selecting the input image patches from actually input frame. The neural\nnetworks is used for classification with PHOG features. Finally, a Kalman\nfilter is applied to calculate the subsequent step for tracking that aims at\nfinding the exact position of pedestrians in an input image. The experimental\nresult shows that the proposed approach yields promising performance on\nmulti-view pedestrian detection and tracking on different benchmark datasets.', 'Studies on microscopic pedestrian requires large amounts of trajectory data\nfrom real-world pedestrian crowds. Such data collection, if done manually,\nneeds tremendous effort and is very time consuming. Though many studies have\nasserted the possibility of automating this task using video cameras, we found\nthat only a few have demonstrated good performance in very crowded situations\nor from a top-angled view scene. This paper deals with tracking pedestrian\ncrowd under heavy occlusions from an angular scene. Our automated tracking\nsystem consists of two modules that perform sequentially. The first module\ndetects moving objects as blobs. The second module is a tracking system. We\nemploy probability distribution from the detection of each pedestrian and use\nBayesian update to track the next position. The result of such tracking is a\ndatabase of pedestrian trajectories over time and space. With certain prior\ninformation, we showed that the system can track a large number of people under\nocclusion and clutter scene.', 'Vehicle-to-Pedestrian (V2P) communication can significantly improve\npedestrian safety at a signalized intersection. It is unlikely that pedestrians\nwill carry a low latency communication enabled device and activate a pedestrian\nsafety application in their hand-held device all the time. Because of this\nlimitation, multiple traffic cameras at the signalized intersection can be used\nto accurately detect and locate pedestrians using deep learning and broadcast\nsafety alerts related to pedestrians to warn connected and automated vehicles\naround a signalized intersection. However, unavailability of high-performance\ncomputing infrastructure at the roadside and limited network bandwidth between\ntraffic cameras and the computing infrastructure limits the ability of\nreal-time data streaming and processing for pedestrian detection. In this\npaper, we develop an edge computing based real-time pedestrian detection\nstrategy combining pedestrian detection algorithm using deep learning and an\nefficient data communication approach to reduce bandwidth requirements while\nmaintaining a high object detection accuracy. We utilize a lossy compression\ntechnique on traffic camera data to determine the tradeoff between the\nreduction of the communication bandwidth requirements and a defined object\ndetection accuracy. The performance of the pedestrian-detection strategy is\nmeasured in terms of pedestrian classification accuracy with varying peak\nsignal-to-noise ratios. The analyses reveal that we detect pedestrians by\nmaintaining a defined detection accuracy with a peak signal-to-noise ratio\n(PSNR) 43 dB while reducing the communication bandwidth from 9.82 Mbits/sec to\n0.31 Mbits/sec, a 31x reduction.', 'Detecting pedestrian has been arguably addressed as a special topic beyond\ngeneral object detection. Although recent deep learning object detectors such\nas Fast/Faster R-CNN [1, 2] have shown excellent performance for general object\ndetection, they have limited success for detecting pedestrian, and previous\nleading pedestrian detectors were in general hybrid methods combining\nhand-crafted and deep convolutional features. In this paper, we investigate\nissues involving Faster R-CNN [2] for pedestrian detection. We discover that\nthe Region Proposal Network (RPN) in Faster R-CNN indeed performs well as a\nstand-alone pedestrian detector, but surprisingly, the downstream classifier\ndegrades the results. We argue that two reasons account for the unsatisfactory\naccuracy: (i) insufficient resolution of feature maps for handling small\ninstances, and (ii) lack of any bootstrapping strategy for mining hard negative\nexamples. Driven by these observations, we propose a very simple but effective\nbaseline for pedestrian detection, using an RPN followed by boosted forests on\nshared, high-resolution convolutional feature maps. We comprehensively evaluate\nthis method on several benchmarks (Caltech, INRIA, ETH, and KITTI), presenting\ncompetitive accuracy and good speed. Code will be made publicly available.', 'As camera quality improves and their deployment moves to areas with limited\nbandwidth, communication bottlenecks can impair real-time constraints of an ITS\napplication, such as video-based real-time pedestrian detection. Video\ncompression reduces the bandwidth requirement to transmit the video but\ndegrades the video quality. As the quality level of the video decreases, it\nresults in the corresponding decreases in the accuracy of the vision-based\npedestrian detection model. Furthermore, environmental conditions (e.g., rain\nand darkness) alter the compression ratio and can make maintaining a high\npedestrian detection accuracy more difficult. The objective of this study is to\ndevelop a real-time error-bounded lossy compression (EBLC) strategy to\ndynamically change the video compression level depending on different\nenvironmental conditions in order to maintain a high pedestrian detection\naccuracy. We conduct a case study to show the efficacy of our dynamic EBLC\nstrategy for real-time vision-based pedestrian detection under adverse\nenvironmental conditions. Our strategy selects the error tolerances dynamically\nfor lossy compression that can maintain a high detection accuracy across a\nrepresentative set of environmental conditions. Analyses reveal that our\nstrategy increases pedestrian detection accuracy up to 14% and reduces the\ncommunication bandwidth up to 14x for adverse environmental conditions compared\nto the same conditions but without our dynamic EBLC strategy. Our dynamic EBLC\nstrategy is independent of detection models and environmental conditions\nallowing other detection models and environmental conditions to be easily\nincorporated in our strategy.', 'Effective fusion of complementary information captured by multi-modal sensors\n(visible and infrared cameras) enables robust pedestrian detection under\nvarious surveillance situations (e.g. daytime and nighttime). In this paper, we\npresent a novel box-level segmentation supervised learning framework for\naccurate and real-time multispectral pedestrian detection by incorporating\nfeatures extracted in visible and infrared channels. Specifically, our method\ntakes pairs of aligned visible and infrared images with easily obtained\nbounding box annotations as input and estimates accurate prediction maps to\nhighlight the existence of pedestrians. It offers two major advantages over the\nexisting anchor box based multispectral detection methods. Firstly, it\novercomes the hyperparameter setting problem occurred during the training phase\nof anchor box based detectors and can obtain more accurate detection results,\nespecially for small and occluded pedestrian instances. Secondly, it is capable\nof generating accurate detection results using small-size input images, leading\nto improvement of computational efficiency for real-time autonomous driving\napplications. Experimental results on KAIST multispectral dataset show that our\nproposed method outperforms state-of-the-art approaches in terms of both\naccuracy and speed.', ""Pedestrians and drivers interact closely in a wide range of environments.\nAutonomous vehicles (AVs) correspondingly face the need to predict pedestrians'\nfuture trajectories in these same environments. Traditional model-based\nprediction methods have been limited to making predictions in highly structured\nscenes with signalized intersections, marked crosswalks, or curbs. Deep\nlearning methods have instead leveraged datasets to learn predictive features\nthat generalize across scenes, at the cost of model interpretability. This\npaper aims to achieve both widely applicable and interpretable predictions by\nproposing a risk-based attention mechanism to learn when pedestrians yield, and\na model of vehicle influence to learn how yielding affects motion. A novel\nprobabilistic method, Off the Sidewalk Predictions (OSP), uses these to achieve\naccurate predictions in both shared spaces and traditional scenes. Experiments\non urban datasets demonstrate that the realtime method achieves\nstate-of-the-art performance."", ""In this paper, we present an end-to-end future-prediction model that focuses\non pedestrian safety. Specifically, our model uses previous video frames,\nrecorded from the perspective of the vehicle, to predict if a pedestrian will\ncross in front of the vehicle. The long term goal of this work is to design a\nfully autonomous system that acts and reacts as a defensive human driver would\n--- predicting future events and reacting to mitigate risk. We focus on\npedestrian-vehicle interactions because of the high risk of harm to the\npedestrian if their actions are miss-predicted. Our end-to-end model consists\nof two stages: the first stage is an encoder/decoder network that learns to\npredict future video frames. The second stage is a deep spatio-temporal network\nthat utilizes the predicted frames of the first stage to predict the\npedestrian's future action. Our system achieves state-of-the-art accuracy on\npedestrian behavior prediction and future frames prediction on the Joint\nAttention for Autonomous Driving (JAAD) dataset."", ""Social distancing can reduce the infection rates in respiratory pandemics\nsuch as COVID-19. Traffic intersections are particularly suitable for\nmonitoring and evaluation of social distancing behavior in metropolises. We\npropose and evaluate a privacy-preserving social distancing analysis system\n(B-SDA), which uses bird's-eye view video recordings of pedestrians who cross\ntraffic intersections. We devise algorithms for video pre-processing, object\ndetection and tracking which are rooted in the known computer-vision and deep\nlearning techniques, but modified to address the problem of detecting very\nsmall objects/pedestrians captured by a highly elevated camera. We propose a\nmethod for incorporating pedestrian grouping for detection of social distancing\nviolations. B-SDA is used to compare pedestrian behavior based on pre-pandemic\nand pandemic videos in a major metropolitan area. The accomplished pedestrian\ndetection performance is $63.0\\%$ $AP_{50}$ and the tracking performance is\n$47.6\\%$ MOTA. The social distancing violation rate of $15.6\\%$ during the\npandemic is notably lower than $31.4\\%$ pre-pandemic baseline, indicating that\npedestrians followed CDC-prescribed social distancing recommendations. The\nproposed system is suitable for deployment in real-world applications."", 'Generative adversarial networks (GANs) have been promising for many computer\nvision problems due to their powerful capabilities to enhance the data for\ntraining and test. In this paper, we leveraged GANs and proposed a new\narchitecture with a cascaded Single Shot Detector (SSD) for pedestrian\ndetection at distance, which is yet a challenge due to the varied sizes of\npedestrians in videos at distance. To overcome the low-resolution issues in\npedestrian detection at distance, DCGAN is employed to improve the resolution\nfirst to reconstruct more discriminative features for a SSD to detect objects\nin images or videos. A crucial advantage of our method is that it learns a\nmulti-scale metric to distinguish multiple objects at different distances under\none image, while DCGAN serves as an encoder-decoder platform to generate parts\nof an image that contain better discriminative information. To measure the\neffectiveness of our proposed method, experiments were carried out on the\nCanadian Institute for Advanced Research (CIFAR) dataset, and it was\ndemonstrated that the proposed new architecture achieved a much better\ndetection rate, particularly on vehicles and pedestrians at distance, making it\nhighly suitable for smart cities applications that need to discover key objects\nor pedestrians at distance.', ""In recent years, road safety has attracted significant attention from\nresearchers and practitioners in the intelligent transport systems domain. As\none of the most common and vulnerable groups of road users, pedestrians cause\ngreat concerns due to their unpredictable behavior and movement, as subtle\nmisunderstandings in vehicle-pedestrian interaction can easily lead to risky\nsituations or collisions. Existing methods use either predefined\ncollision-based models or human-labeling approaches to estimate the\npedestrians' risks. These approaches are usually limited by their poor\ngeneralization ability and lack of consideration of interactions between the\nego vehicle and a pedestrian. This work tackles the listed problems by\nproposing a Pedestrian Risk Level Prediction system. The system consists of\nthree modules. Firstly, vehicle-perspective pedestrian data are collected.\nSince the data contains information regarding the movement of both the ego\nvehicle and pedestrian, it can simplify the prediction of spatiotemporal\nfeatures in an interaction-aware fashion. Using the long short-term memory\nmodel, the pedestrian trajectory prediction module predicts their\nspatiotemporal features in the subsequent five frames. As the predicted\ntrajectory follows certain interaction and risk patterns, a hybrid clustering\nand classification method is adopted to explore the risk patterns in the\nspatiotemporal features and train a risk level classifier using the learned\npatterns. Upon predicting the spatiotemporal features of pedestrians and\nidentifying the corresponding risk level, the risk patterns between the ego\nvehicle and pedestrians are determined. Experimental results verified the\ncapability of the PRLP system to predict the risk level of pedestrians, thus\nsupporting the collision risk assessment of intelligent vehicles and providing\nsafety warnings to both vehicles and pedestrians.""]","[""This paper presents a planning system for autonomous driving among many\npedestrians. A key ingredient of our approach is PORCA, a pedestrian motion\nprediction model that accounts for both a pedestrian's global navigation\nintention and local interactions with the vehicle and other pedestrians.\nUnfortunately, the autonomous vehicle does not know the pedestrian's intention\na priori and requires a planning algorithm that hedges against the uncertainty\nin pedestrian intentions. Our planning system combines a POMDP algorithm with\nthe pedestrian motion model and runs in near real time. Experiments show that\nit enables a robot vehicle to drive safely, efficiently, and smoothly among a\ncrowd with a density of nearly one person per square meter."", ""A major bottleneck of pedestrian detection lies on the sharp performance\ndeterioration in the presence of small-size pedestrians that are relatively far\nfrom the camera. Motivated by the observation that pedestrians of disparate\nspatial scales exhibit distinct visual appearances, we propose in this paper an\nactive pedestrian detector that explicitly operates over multiple-layer\nneuronal representations of the input still image. More specifically,\nconvolutional neural nets such as ResNet and faster R-CNNs are exploited to\nprovide a rich and discriminative hierarchy of feature representations as well\nas initial pedestrian proposals. Here each pedestrian observation of distinct\nsize could be best characterized in terms of the ResNet feature representation\nat a certain layer of the hierarchy; Meanwhile, initial pedestrian proposals\nare attained by faster R-CNNs techniques, i.e. region proposal network and\nfollow-up region of interesting pooling layer employed right after the specific\nResNet convolutional layer of interest, to produce joint predictions on the\nbounding-box proposals' locations and categories (i.e. pedestrian or not). This\nis engaged as input to our active detector where for each initial pedestrian\nproposal, a sequence of coordinate transformation actions is carried out to\ndetermine its proper x-y 2D location and layer of feature representation, or\neventually terminated as being background. Empirically our approach is\ndemonstrated to produce overall lower detection errors on widely-used\nbenchmarks, and it works particularly well with far-scale pedestrians. For\nexample, compared with 60.51% log-average miss rate of the state-of-the-art\nMS-CNN for far-scale pedestrians (those below 80 pixels in bounding-box height)\nof the Caltech benchmark, the miss rate of our approach is 41.85%, with a\nnotable reduction of 18.68%."", 'We develop predictive models of pedestrian dynamics by encoding the coupled\nnature of multi-pedestrian interaction using game theory, and deep\nlearning-based visual analysis to estimate person-specific behavior parameters.\nBuilding predictive models for multi-pedestrian interactions however, is very\nchallenging due to two reasons: (1) the dynamics of interaction are complex\ninterdependent processes, where the predicted behavior of one pedestrian can\naffect the actions taken by others and (2) dynamics are variable depending on\nan individuals physical characteristics (e.g., an older person may walk slowly\nwhile the younger person may walk faster). To address these challenges, we (1)\nutilize concepts from game theory to model the interdependent decision making\nprocess of multiple pedestrians and (2) use visual classifiers to learn a\nmapping from pedestrian appearance to behavior parameters. We evaluate our\nproposed model on several public multiple pedestrian interaction video\ndatasets. Results show that our strategic planning model explains human\ninteractions 25% better when compared to state-of-the-art methods.', ""As autonomous vehicles and autonomous racing rise in popularity, so does the\nneed for faster and more accurate detectors. While our naked eyes are able to\nextract contextual information almost instantly, even from far away, image\nresolution and computational resources limitations make detecting smaller\nobjects (that is, objects that occupy a small pixel area in the input image) a\ngenuinely challenging task for machines and a wide-open research field. This\nstudy explores how the popular YOLOv5 object detector can be modified to\nimprove its performance in detecting smaller objects, with a particular\napplication in autonomous racing. To achieve this, we investigate how replacing\ncertain structural elements of the model (as well as their connections and\nother parameters) can affect performance and inference time. In doing so, we\npropose a series of models at different scales, which we name `YOLO-Z', and\nwhich display an improvement of up to 6.9% in mAP when detecting smaller\nobjects at 50% IOU, at the cost of just a 3ms increase in inference time\ncompared to the original YOLOv5. Our objective is to inform future research on\nthe potential of adjusting a popular detector such as YOLOv5 to address\nspecific tasks and provide insights on how specific changes can impact small\nobject detection. Such findings, applied to the broader context of autonomous\nvehicles, could increase the amount of contextual information available to such\nsystems."", 'It is challenging for a mobile robot to navigate through human crowds.\nExisting approaches usually assume that pedestrians follow a predefined\ncollision avoidance strategy, like social force model (SFM) or optimal\nreciprocal collision avoidance (ORCA). However, their performances commonly\nneed to be further improved for practical applications, where pedestrians\nfollow multiple different collision avoidance strategies. In this paper, we\npropose a map-based deep reinforcement learning approach for crowd-aware robot\nnavigation with various pedestrians. We use the sensor map to represent the\nenvironmental information around the robot, including its shape and observable\nappearances of obstacles. We also introduce the pedestrian map that specifies\nthe movements of pedestrians around the robot. By applying both maps as inputs\nof the neural network, we show that a navigation policy can be trained to\nbetter interact with pedestrians following different collision avoidance\nstrategies. We evaluate our approach under multiple scenarios both in the\nsimulator and on an actual robot. The results show that our approach allows the\nrobot to successfully interact with various pedestrians and outperforms\ncompared methods in terms of the success rate.', 'Pedestrian detection is a popular research topic due to its paramount\nimportance for a number of applications, especially in the fields of\nautomotive, surveillance and robotics. Despite the significant improvements,\npedestrian detection is still an open challenge that calls for more and more\naccurate algorithms. In the last few years, deep learning and in particular\nconvolutional neural networks emerged as the state of the art in terms of\naccuracy for a number of computer vision tasks such as image classification,\nobject detection and segmentation, often outperforming the previous gold\nstandards by a large margin. In this paper, we propose a pedestrian detection\nsystem based on deep learning, adapting a general-purpose convolutional network\nto the task at hand. By thoroughly analyzing and optimizing each step of the\ndetection pipeline we propose an architecture that outperforms traditional\nmethods, achieving a task accuracy close to that of state-of-the-art\napproaches, while requiring a low computational time. Finally, we tested the\nsystem on an NVIDIA Jetson TK1, a 192-core platform that is envisioned to be a\nforerunner computational brain of future self-driving cars.', ""Pedestrian detection is a research hotspot and a difficult issue in the\ncomputer vision such as the Intelligent Surveillance System, the Intelligent\nTransport System, robotics, and automotive safety. However, the human body's\nposition, angle, and dress in a video scene are complicated and changeable,\nwhich have a great influence on the detection accuracy. In this paper, through\nthe analysis on the pros and cons of Census Transform Histogram (CENTRIST), a\nnovel feature is presented for human detection Ternary CENTRIST (T-CENTRIST).\nThe T-CENTRIST feature takes the relationship between each pixel and its\nneighborhood pixels into account. Meanwhile, it also considers the relevancy\namong these neighborhood pixels. Therefore, the proposed feature description\nmethod can reflect the silhouette of pedestrian more adequately and accurately\nthan that of CENTRIST. Second, we propose a fast pedestrian detection framework\nbased on T-CENTRIST in infrared image, which introduces the idea of extended\nblocks and the integral image. Finally, experimental results verify the\neffectiveness of the proposed pedestrian detection method."", 'A real-time Deep Learning based method for Pedestrian Detection (PD) is\napplied to the Human-Aware robot navigation problem. The pedestrian detector\ncombines the Aggregate Channel Features (ACF) detector with a deep\nConvolutional Neural Network (CNN) in order to obtain fast and accurate\nperformance. Our solution is firstly evaluated using a set of real images taken\nfrom onboard and offboard cameras and, then, it is validated in a typical robot\nnavigation environment with pedestrians (two distinct experiments are\nconducted). The results on both tests show that our pedestrian detector is\nrobust and fast enough to be used on robot navigation applications.', 'To obtain effective pedestrian detection results in surveillance video, there\nhave been many methods proposed to handle the problems from severe occlusion,\npose variation, clutter background, \\emph{etc}. Besides detection accuracy, a\nrobust surveillance video system should be stable to video quality degradation\nby network transmission, environment variation, etc. In this study, we conduct\nthe research on the robustness of pedestrian detection algorithms to video\nquality degradation. The main contribution of this work includes the following\nthree aspects. First, a large-scale Distorted Surveillance Video Data Set\n(DSurVD) is constructed from high-quality video sequences and their\ncorresponding distorted versions. Second, we design a method to evaluate\ndetection stability and a robustness measure called Robustness Quadrangle,\nwhich can be adopted to visualize detection accuracy of pedestrian detection\nalgorithms on high-quality video sequences and stability with video quality\ndegradation. Third, the robustness of seven existing pedestrian detection\nalgorithms is evaluated by the built DSurVD. Experimental results show that the\nrobustness can be further improved for existing pedestrian detection\nalgorithms. Additionally, we provide much in-depth discussion on how different\ndistortion types influence the performance of pedestrian detection algorithms,\nwhich is important to design effective pedestrian detection algorithms for\nsurveillance. The DSurVD data set can be download from BaiduYunDisk,\nhttps://pan.baidu.com/s/1I9Kqj8rmubOYu7bkBfkUpA, Password: lqmc', 'Pedestrian behavior prediction is one of the major challenges for intelligent\ndriving systems in urban environments. Pedestrians often exhibit a wide range\nof behaviors and adequate interpretations of those depend on various sources of\ninformation such as pedestrian appearance, states of other road users, the\nenvironment layout, etc. To address this problem, we propose a novel\nmulti-modal prediction algorithm that incorporates different sources of\ninformation captured from the environment to predict future crossing actions of\npedestrians. The proposed model benefits from a hybrid learning architecture\nconsisting of feedforward and recurrent networks for analyzing visual features\nof the environment and dynamics of the scene. Using the existing 2D pedestrian\nbehavior benchmarks and a newly annotated 3D driving dataset, we show that our\nproposed model achieves state-of-the-art performance in pedestrian crossing\nprediction.', 'Pedestrian detection plays an important role in many applications such as\nautonomous driving. We propose a method that explores semantic segmentation\nresults as self-attention cues to significantly improve the pedestrian\ndetection performance. Specifically, a multi-task network is designed to\njointly learn semantic segmentation and pedestrian detection from image\ndatasets with weak box-wise annotations. The semantic segmentation feature maps\nare concatenated with corresponding convolution features maps to provide more\ndiscriminative features for pedestrian detection and pedestrian classification.\nBy jointly learning segmentation and detection, our proposed pedestrian\nself-attention mechanism can effectively identify pedestrian regions and\nsuppress backgrounds. In addition, we propose to incorporate semantic attention\ninformation from multi-scale layers into deep convolution neural network to\nboost pedestrian detection. Experiment results show that the proposed method\nachieves the best detection performance with MR of 6.27% on Caltech dataset and\nobtain competitive performance on CityPersons dataset while maintaining high\ncomputational efficiency.', 'Detection of pedestrians in aerial imagery captured by drones has many\napplications including intersection monitoring, patrolling, and surveillance,\nto name a few. However, the problem is involved due to continuouslychanging\ncamera viewpoint and object appearance as well as the need for lightweight\nalgorithms to run on on-board embedded systems. To address this issue, the\npaper proposes a framework for pedestrian detection in videos based on the YOLO\nobject detection network [6] while having a high throughput of more than 5 FPS\non the Jetson TX2 embedded board. The framework exploits deep learning for\nrobust operation and uses a pre-trained model without the need for any\nadditional training which makes it flexible to apply on different setups with\nminimum amount of tuning. The method achieves ~81 mAP when applied on a sample\nvideo from the Embedded Real-Time Inference (ERTI) Challenge where pedestrians\nare monitored by a UAV.', 'Pedestrian detection is an important component for safety of autonomous\nvehicles, as well as for traffic and street surveillance. There are extensive\nbenchmarks on this topic and it has been shown to be a challenging problem when\napplied on real use-case scenarios. In purely image-based pedestrian detection\napproaches, the state-of-the-art results have been achieved with convolutional\nneural networks (CNN) and surprisingly few detection frameworks have been built\nupon multi-cue approaches. In this work, we develop a new pedestrian detector\nfor autonomous vehicles that exploits LiDAR data, in addition to visual\ninformation. In the proposed approach, LiDAR data is utilized to generate\nregion proposals by processing the three dimensional point cloud that it\nprovides. These candidate regions are then further processed by a\nstate-of-the-art CNN classifier that we have fine-tuned for pedestrian\ndetection. We have extensively evaluated the proposed detection process on the\nKITTI dataset. The experimental results show that the proposed LiDAR space\nclustering approach provides a very efficient way of generating region\nproposals leading to higher recall rates and fewer misses for pedestrian\ndetection. This indicates that LiDAR data can provide auxiliary information for\nCNN-based approaches.', 'With the unprecedented shift towards automated urban environments in recent\nyears, a new paradigm is required to study pedestrian behaviour. Studying\npedestrian behaviour in futuristic scenarios requires modern data sources that\nconsider both the Automated Vehicle (AV) and pedestrian perspectives. Current\nopen datasets on AVs predominantly fail to account for the latter, as they do\nnot include an adequate number of events and associated details that involve\npedestrian and vehicle interactions. To address this issue, we propose using\nVirtual Reality (VR) data as a complementary resource to current datasets,\nwhich can be designed to measure pedestrian behaviour under specific\nconditions. In this research, we focus on the context-aware pedestrian\ntrajectory prediction framework for automated vehicles at mid-block\nunsignalized crossings. For this purpose, we develop a novel multi-input\nnetwork of Long Short-Term Memory (LSTM) and fully connected dense layers. In\naddition to past trajectories, the proposed framework incorporates pedestrian\nhead orientations and distance to the upcoming vehicles as sequential input\ndata. By merging the sequential data with contextual information of the\nenvironment, we train a model to predict the future pedestrian trajectory. Our\nresults show that the prediction error and overfitting to the training data are\nreduced by considering contextual information in the model. To analyze the\napplication of the methods to real AV data, the proposed framework is trained\nand applied to pedestrian trajectory extracted from an open-access video\ndataset. Finally, by implementing a game theory-based model interpretability\nmethod, we provide detailed insights and propose recommendations to improve the\ncurrent automated vehicle sensing systems from a pedestrian-oriented point of\nview.', 'Action and intention recognition of pedestrians in urban settings are\nchallenging problems for Advanced Driver Assistance Systems as well as future\nautonomous vehicles to maintain smooth and safe traffic. This work investigates\na number of feature extraction methods in combination with several machine\nlearning algorithms to build knowledge on how to automatically detect the\naction and intention of pedestrians in urban traffic. We focus on the motion\nand head orientation to predict whether the pedestrian is about to cross the\nstreet or not. The work is based on the Joint Attention for Autonomous Driving\n(JAAD) dataset, which contains 346 videoclips of various traffic scenarios\ncaptured with cameras mounted in the windshield of a car. An accuracy of 72%\nfor head orientation estimation and 85% for motion detection is obtained in our\nexperiments.', ""One of the major challenges for autonomous vehicles in urban environments is\nto understand and predict other road users' actions, in particular, pedestrians\nat the point of crossing. The common approach to solving this problem is to use\nthe motion history of the agents to predict their future trajectories. However,\npedestrians exhibit highly variable actions most of which cannot be understood\nwithout visual observation of the pedestrians themselves and their\nsurroundings. To this end, we propose a solution for the problem of pedestrian\naction anticipation at the point of crossing. Our approach uses a novel stacked\nRNN architecture in which information collected from various sources, both\nscene dynamics and visual features, is gradually fused into the network at\ndifferent levels of processing. We show, via extensive empirical evaluations,\nthat the proposed algorithm achieves a higher prediction accuracy compared to\nalternative recurrent network architectures. We conduct experiments to\ninvestigate the impact of the length of observation, time to event and types of\nfeatures on the performance of the proposed method. Finally, we demonstrate how\ndifferent data fusion strategies impact prediction accuracy."", 'Pedestrians in videos have a wide range of appearances such as body poses,\nocclusions, and complex backgrounds, and there exists the proposal shift\nproblem in pedestrian detection that causes the loss of body parts such as head\nand legs. To address it, we propose part-level convolutional neural networks\n(CNN) for pedestrian detection using saliency and boundary box alignment in\nthis paper. The proposed network consists of two sub-networks: detection and\nalignment. We use saliency in the detection sub-network to remove false\npositives such as lamp posts and trees. We adopt bounding box alignment on\ndetection proposals in the alignment sub-network to address the proposal shift\nproblem. First, we combine FCN and CAM to extract deep features for pedestrian\ndetection. Then, we perform part-level CNN to recall the lost body parts.\nExperimental results on various datasets demonstrate that the proposed method\nremarkably improves accuracy in pedestrian detection and outperforms existing\nstate-of-the-arts in terms of log average miss rate at false position per image\n(FPPI).', 'Predicting the collective motion of a group of pedestrians (a crowd) under\nthe vehicle influence is essential for the development of autonomous vehicles\nto deal with mixed urban scenarios where interpersonal interaction and\nvehicle-crowd interaction (VCI) are significant. This usually requires a model\nthat can describe individual pedestrian motion under the influence of nearby\npedestrians and the vehicle. This study proposed two pedestrian trajectory\ndatasets, CITR dataset and DUT dataset, so that the pedestrian motion models\ncan be further calibrated and verified, especially when vehicle influence on\npedestrians plays an important role. CITR dataset consists of experimentally\ndesigned fundamental VCI scenarios (front, back, and lateral VCIs) and provides\nunique ID for each pedestrian, which is suitable for exploring a specific\naspect of VCI. DUT dataset gives two ordinary and natural VCI scenarios in\ncrowded university campus, which can be used for more general purpose VCI\nexploration. The trajectories of pedestrians, as well as vehicles, were\nextracted by processing video frames that come from a down-facing camera\nmounted on a hovering drone as the recording equipment. The final trajectories\nof pedestrians and vehicles were refined by Kalman filters with linear\npoint-mass model and nonlinear bicycle model, respectively, in which\nxy-velocity of pedestrians and longitudinal speed and orientation of vehicles\nwere estimated. The statistics of the velocity magnitude distribution\ndemonstrated the validity of the proposed dataset. In total, there are\napproximate 340 pedestrian trajectories in CITR dataset and 1793 pedestrian\ntrajectories in DUT dataset. The dataset is available at GitHub.', ""Vehicle-pedestrian interaction (VPI) is one of the most challenging tasks for\nautomated driving systems. The design of driving strategies for such systems\nusually starts with verifying VPI in simulation. This work proposed an improved\nframework for the study of VPI in uncontrolled pedestrian crossing scenarios.\nThe framework admits the mutual effect between the pedestrian and the vehicle.\nA multi-state social force based pedestrian motion model was designed to\ndescribe the microscopic motion of the pedestrian crossing behavior. The\npedestrian model considers major interaction factors such as the accepted gap\nof the pedestrian's decision on when to start crossing, the desired speed of\nthe pedestrian, and the effect of the vehicle on the pedestrian while the\npedestrian is crossing the road. Vehicle driving strategies focus on the\nlongitudinal motion control, for which the feedback obstacle avoidance control\nand the model predictive control were tested and compared in the framework. The\nsimulation results verified that the proposed framework can generate a variety\nof VPI scenarios, consisting of either the pedestrian yielding to the vehicle\nor the vehicle yielding to the pedestrian. The framework can be easily extended\nto apply different approaches to the VPI problems."", 'We present a novel, realtime algorithm to compute the trajectory of each\npedestrian in moderately dense crowd scenes. Our formulation is based on an\nadaptive particle filtering scheme that uses a multi-agent motion model based\non velocity-obstacles, and takes into account local interactions as well as\nphysical and personal constraints of each pedestrian. Our method dynamically\nchanges the number of particles allocated to each pedestrian based on different\nconfidence metrics. Additionally, we use a new high-definition crowd video\ndataset, which is used to evaluate the performance of different pedestrian\ntracking algorithms. This dataset consists of videos of indoor and outdoor\nscenes, recorded at different locations with 30-80 pedestrians. We highlight\nthe performance benefits of our algorithm over prior techniques using this\ndataset. In practice, our algorithm can compute trajectories of tens of\npedestrians on a multi-core desktop CPU at interactive rates (27-30 frames per\nsecond). To the best of our knowledge, our approach is 4-5 times faster than\nprior methods, which provide similar accuracy.', 'To deal with many pedestrian data, automatic data collection is needed. This\npaper describes how to automate the microscopic pedestrian flow data collection\nfrom video files. The study is restricted only to pedestrians without\nconsidering vehicular - pedestrian interaction. Pedestrian tracking system\nconsists of three sub-systems, which calculates the image processing, object\ntracking and traffic flow variables. The system receives input of stacks of\nimages and parameters. The first sub-system performs Image Processing analysis\nwhile the second sub-system carries out the tracking of pedestrians by matching\nthe features and tracing the pedestrian numbers frame by frame. The last\nsub-system deals with a NTXY database to calculate the pedestrian traffic-flow\ncharacteristic such as flow rate, speed and area module. Comparison with manual\ndata collection method confirmed that the procedures described have significant\npotential to automate the data collection of both microscopic and macroscopic\npedestrian flow variables.', 'Various methods to automate traffic data collection have recently been\ndeveloped by many researchers. A macroscopic data collection through image\nprocessing has been proposed. For microscopic traffic flow data, such as\nindividual speed and time or distance headway, tracking of individual movement\nis needed. The tracking algorithms for pedestrian or vehicle have been\ndeveloped to trace the movement of one or two pedestrians based on sign\npattern, and feature detection. No research has been done to track many\npedestrians or vehicles at once. This paper describes a new and fast algorithm\nto track the movement of many individual vehicles or pedestrians', ""Pedestrians and vehicles often share the road in complex inner city traffic.\nThis leads to interactions between the vehicle and pedestrians, with each\naffecting the other's motion. In order to create robust methods to reason about\npedestrian behavior and to design interfaces of communication between\nself-driving cars and pedestrians we need to better understand such\ninteractions. In this paper, we present a data-driven approach to implicitly\nmodel pedestrians' interactions with vehicles, to better predict pedestrian\nbehavior. We propose a LSTM model that takes as input the past trajectories of\nthe pedestrian and ego-vehicle, and pedestrian head orientation, and predicts\nthe future positions of the pedestrian. Our experiments based on a real-world,\ninner city dataset captured with vehicle mounted cameras, show that the usage\nof such cues improve pedestrian prediction when compared to a baseline that\npurely uses the past trajectory of the pedestrian."", 'Risk is traditionally described as the expected likelihood of an undesirable\noutcome, such as collisions for autonomous vehicles. Accurately predicting risk\nor potentially risky situations is critical for the safe operation of\nautonomous vehicles. In our previous work, we showed that risk could be\ncharacterized by two components: 1) the probability of an undesirable outcome\nand 2) an estimate of how undesirable the outcome is (loss). This paper is an\nextension to our previous work. In this paper, using our trained deep\nreinforcement learning model for navigating around crowds, we developed a\nrisk-based decision-making framework for the autonomous vehicle that integrates\nthe high-level risk-based path planning with the reinforcement learning-based\nlow-level control. We evaluated our method in a high-fidelity simulation such\nas CARLA. This work can improve safety by allowing an autonomous vehicle to one\nday avoid and react to risky situations.', 'Predicting the future trajectories of pedestrians is a challenging problem\nthat has a range of application, from crowd surveillance to autonomous driving.\nIn literature, methods to approach pedestrian trajectory prediction have\nevolved, transitioning from physics-based models to data-driven models based on\nrecurrent neural networks. In this work, we propose a new approach to\npedestrian trajectory prediction, with the introduction of a novel 2D\nconvolutional model. This new model outperforms recurrent models, and it\nachieves state-of-the-art results on the ETH and TrajNet datasets. We also\npresent an effective system to represent pedestrian positions and powerful data\naugmentation techniques, such as the addition of Gaussian noise and the use of\nrandom rotations, which can be applied to any model. As an additional\nexploratory analysis, we present experimental results on the inclusion of\noccupancy methods to model social information, which empirically show that\nthese methods are ineffective in capturing social interaction.', ""Autonomous vehicles (AVs) must share space with pedestrians, both in\ncarriageway cases such as cars at pedestrian crossings and off-carriageway\ncases such as delivery vehicles navigating through crowds on pedestrianized\nhigh-streets. Unlike static obstacles, pedestrians are active agents with\ncomplex, interactive motions. Planning AV actions in the presence of\npedestrians thus requires modelling of their probable future behaviour as well\nas detecting and tracking them. This narrative review article is Part II of a\npair, together surveying the current technology stack involved in this process,\norganising recent research into a hierarchical taxonomy ranging from low-level\nimage detection to high-level psychological models, from the perspective of an\nAV designer. This self-contained Part II covers the higher levels of this\nstack, consisting of models of pedestrian behaviour, from prediction of\nindividual pedestrians' likely destinations and paths, to game-theoretic models\nof interactions between pedestrians and autonomous vehicles. This survey\nclearly shows that, although there are good models for optimal walking\nbehaviour, high-level psychological and social modelling of pedestrian\nbehaviour still remains an open research question that requires many conceptual\nissues to be clarified. Early work has been done on descriptive and qualitative\nmodels of behaviour, but much work is still needed to translate them into\nquantitative algorithms for practical AV control."", 'We present a Pedestrian Dominance Model (PDM) to identify the dominance\ncharacteristics of pedestrians for robot navigation. Through a perception study\non a simulated dataset of pedestrians, PDM models the perceived dominance\nlevels of pedestrians with varying motion behaviors corresponding to\ntrajectory, speed, and personal space. At runtime, we use PDM to identify the\ndominance levels of pedestrians to facilitate socially-aware navigation for the\nrobots. PDM can predict dominance levels from trajectories with ~85% accuracy.\nPrior studies in psychology literature indicate that when interacting with\nhumans, people are more comfortable around people that exhibit complementary\nmovement behaviors. Our algorithm leverages this by enabling the robots to\nexhibit complementing responses to pedestrian dominance. We also present an\napplication of PDM for generating dominance-based collision-avoidance behaviors\nin the navigation of autonomous vehicles among pedestrians. We demonstrate the\nbenefits of our algorithm for robots navigating among tens of pedestrians in\nsimulated environments.', 'As autonomous vehicles become an every-day reality, high-accuracy pedestrian\ndetection is of paramount practical importance. Pedestrian detection is a\nhighly researched topic with mature methods, but most datasets focus on common\nscenes of people engaged in typical walking poses on sidewalks. But performance\nis most crucial for dangerous scenarios, such as children playing in the street\nor people using bicycles/skateboards in unexpected ways. Such ""in-the-tail""\ndata is notoriously hard to observe, making both training and testing\ndifficult. To analyze this problem, we have collected a novel annotated dataset\nof dangerous scenarios called the Precarious Pedestrian dataset. Even given a\ndedicated collection effort, it is relatively small by contemporary standards\n(around 1000 images). To allow for large-scale data-driven learning, we explore\nthe use of synthetic data generated by a game engine. A significant challenge\nis selected the right ""priors"" or parameters for synthesis: we would like\nrealistic data with poses and object configurations that mimic true Precarious\nPedestrians. Inspired by Generative Adversarial Networks (GANs), we generate a\nmassive amount of synthetic data and train a discriminative classifier to\nselect a realistic subset, which we deem the Adversarial Imposters. We\ndemonstrate that this simple pipeline allows one to synthesize realistic\ntraining data by making use of rendering/animation engines within a GAN\nframework. Interestingly, we also demonstrate that such data can be used to\nrank algorithms, suggesting that Adversarial Imposters can also be used for\n""in-the-tail"" validation at test-time, a notoriously difficult challenge for\nreal-world deployment.', 'Pedestrian detection in a crowd is a challenging task due to a high number of\nmutually-occluding human instances, which brings ambiguity and optimization\ndifficulties to the current IoU-based ground truth assignment procedure in\nclassical object detection methods. In this paper, we develop a unique\nperspective of pedestrian detection as a variational inference problem. We\nformulate a novel and efficient algorithm for pedestrian detection by modeling\nthe dense proposals as a latent variable while proposing a customized Auto\nEncoding Variational Bayes (AEVB) algorithm. Through the optimization of our\nproposed algorithm, a classical detector can be fashioned into a variational\npedestrian detector. Experiments conducted on CrowdHuman and CityPersons\ndatasets show that the proposed algorithm serves as an efficient solution to\nhandle the dense pedestrian detection problem for the case of single-stage\ndetectors. Our method can also be flexibly applied to two-stage detectors,\nachieving notable performance enhancement.', 'We present a novel, real-time algorithm to track the trajectory of each\npedestrian in moderately dense crowded scenes. Our formulation is based on an\nadaptive particle-filtering scheme that uses a combination of various\nmulti-agent heterogeneous pedestrian simulation models. We automatically\ncompute the optimal parameters for each of these different models based on\nprior tracked data and use the best model as motion prior for our\nparticle-filter based tracking algorithm. We also use our ""mixture of motion\nmodels"" for adaptive particle selection and accelerate the performance of the\nonline tracking algorithm. The motion model parameter estimation is formulated\nas an optimization problem, and we use an approach that solves this\ncombinatorial optimization problem in a model independent manner and hence\nscalable to any multi-agent pedestrian motion model. We evaluate the\nperformance of our approach on different crowd video datasets and highlight the\nimprovement in accuracy over homogeneous motion models and a baseline\nmean-shift based tracker. In practice, our formulation can compute trajectories\nof tens of pedestrians on a multi-core desktop CPU in in real time and offer\nhigher accuracy as compared to prior real time pedestrian tracking algorithms.', ""To ensure pedestrian friendly streets in the era of automated vehicles,\nreassessment of current policies, practices, design, rules and regulations of\nurban areas is of importance. This study investigates pedestrian crossing\nbehaviour, as an important element of urban dynamics that is expected to be\naffected by the presence of automated vehicles. For this purpose, an\ninterpretable machine learning framework is proposed to explore factors\naffecting pedestrians' wait time before crossing mid-block crosswalks in the\npresence of automated vehicles. To collect rich behavioural data, we developed\na dynamic and immersive virtual reality experiment, with 180 participants from\na heterogeneous population in 4 different locations in the Greater Toronto Area\n(GTA). Pedestrian wait time behaviour is then analyzed using a data-driven Cox\nProportional Hazards (CPH) model, in which the linear combination of the\ncovariates is replaced by a flexible non-linear deep neural network. The\nproposed model achieved a 5% improvement in goodness of fit, but more\nimportantly, enabled us to incorporate a richer set of covariates. A game\ntheoretic based interpretability method is used to understand the contribution\nof different covariates to the time pedestrians wait before crossing. Results\nshow that the presence of automated vehicles on roads, wider lane widths, high\ndensity on roads, limited sight distance, and lack of walking habits are the\nmain contributing factors to longer wait times. Our study suggested that, to\nmove towards pedestrian-friendly urban areas, national level educational\nprograms for children, enhanced safety measures for seniors, promotion of\nactive modes of transportation, and revised traffic rules and regulations\nshould be considered."", 'In mixed traffic scenarios, a certain number of pedestrians might coexist in\na small area while interacting with vehicles. In this situation, every\npedestrian must simultaneously react to the surrounding pedestrians and\nvehicles. Analytical modeling of such collective pedestrian motion can benefit\nintelligent transportation practices like shared space design and urban\nautonomous driving. This work proposed the sub-goal social force model (SG-SFM)\nto describe the collective pedestrian motion under vehicle influence. The\nproposed model introduced a new design of vehicle influence on pedestrian\nmotion, which was smoothly combined with the influence of surrounding\npedestrians using the sub-goal concept. This model aims to describe generalized\npedestrian motion, i.e., it is applicable to various vehicle-pedestrian\ninteraction patterns. The generalization was verified by both quantitative and\nqualitative evaluation. The quantitative evaluation was conducted to reproduce\npedestrian motion in three different datasets, HBS, CITR, and DUT. It also\ncompared two different ways of calibrating the model parameters. The\nqualitative evaluation examined the simulation of collective pedestrian motion\nin a series of fundamental vehicle-pedestrian interaction scenarios. The above\nevaluation results demonstrated the effectiveness of the proposed model.', ""In order to operate safely on the road, autonomous vehicles need not only to\nbe able to identify objects in front of them, but also to be able to estimate\nthe risk level of the object in front of the vehicle automatically. It is\nobvious that different objects have different levels of danger to autonomous\nvehicles. An evaluation system is needed to automatically determine the danger\nlevel of the object for the autonomous vehicle. It would be too subjective and\nincomplete if the system were completely defined by humans. Based on this, we\npropose a framework based on nonparametric Bayesian learning method -- a sticky\nhierarchical Dirichlet process hidden Markov model(sticky HDP-HMM), and\ndiscover the relationship between driving scenarios and driving styles. We use\nthe analysis of driving styles of autonomous vehicles to reflect the risk\nlevels of driving scenarios to the vehicles. In this framework, we firstly use\nsticky HDP-HMM to extract driving styles from the dataset and get different\nclusters, then an evaluation system is proposed to evaluate and rank the\nurgency levels of the clusters. Finally, we map the driving scenarios to the\nranking results and thus get clusters of driving scenarios in different risk\nlevels. More importantly, we find the relationship between driving scenarios\nand driving styles. The experiment shows that our framework can cluster and\nrank driving styles of different urgency levels and find the relationship\nbetween driving scenarios and driving styles and the conclusions also fit\npeople's common sense when driving. Furthermore, this framework can be used for\nautonomous vehicles to estimate risk levels of driving scenarios and help them\nmake precise and safe decisions."", 'Prediction of pedestrian behavior is critical for fully autonomous vehicles\nto drive in busy city streets safely and efficiently. The future autonomous\ncars need to fit into mixed conditions with not only technical but also social\ncapabilities. As more algorithms and datasets have been developed to predict\npedestrian behaviors, these efforts lack the benchmark labels and the\ncapability to estimate the temporal-dynamic intent changes of the pedestrians,\nprovide explanations of the interaction scenes, and support algorithms with\nsocial intelligence. This paper proposes and shares another benchmark dataset\ncalled the IUPUI-CSRC Pedestrian Situated Intent (PSI) data with two innovative\nlabels besides comprehensive computer vision labels. The first novel label is\nthe dynamic intent changes for the pedestrians to cross in front of the\nego-vehicle, achieved from 24 drivers with diverse backgrounds. The second one\nis the text-based explanations of the driver reasoning process when estimating\npedestrian intents and predicting their behaviors during the interaction\nperiod. These innovative labels can enable several computer vision tasks,\nincluding pedestrian intent/behavior prediction, vehicle-pedestrian interaction\nsegmentation, and video-to-language mapping for explainable algorithms. The\nreleased dataset can fundamentally improve the development of pedestrian\nbehavior prediction models and develop socially intelligent autonomous cars to\ninteract with pedestrians efficiently. The dataset has been evaluated with\ndifferent tasks and is released to the public to access.', 'Pedestrian detection methods have been significantly improved with the\ndevelopment of deep convolutional neural networks. Nevertheless, detecting\nsmall-scaled pedestrians and occluded pedestrians remains a challenging\nproblem. In this paper, we propose a pedestrian detection method with a\ncouple-network to simultaneously address these two issues. One of the\nsub-networks, the gated multi-layer feature extraction sub-network, aims to\nadaptively generate discriminative features for pedestrian candidates in order\nto robustly detect pedestrians with large variations on scales. The second\nsub-network targets in handling the occlusion problem of pedestrian detection\nby using deformable regional RoI-pooling. We investigate two different gate\nunits for the gated sub-network, namely, the channel-wise gate unit and the\nspatio-wise gate unit, which can enhance the representation ability of the\nregional convolutional features among the channel dimensions or across the\nspatial domain, repetitively. Ablation studies have validated the effectiveness\nof both the proposed gated multi-layer feature extraction sub-network and the\ndeformable occlusion handling sub-network. With the coupled framework, our\nproposed pedestrian detector achieves state-of-the-art results on the Caltech\nand the CityPersons pedestrian detection benchmarks.', 'In this paper, we present an efficient pedestrian detection system, designed\nby fusion of multiple deep neural network (DNN) systems. Pedestrian candidates\nare first generated by a single shot convolutional multi-box detector at\ndifferent locations with various scales and aspect ratios. The candidate\ngenerator is designed to provide the majority of ground truth pedestrian\nannotations at the cost of a large number of false positives. Then, a\nclassification system using the idea of ensemble learning is deployed to\nimprove the detection accuracy. The classification system further classifies\nthe generated candidates based on opinions of multiple deep verification\nnetworks and a fusion network which utilizes a novel soft-rejection fusion\nmethod to adjust the confidence in the detection results. To improve the\ntraining of the deep verification networks, a novel soft-label method is\ndevised to assign floating point labels to the generated pedestrian candidates.\nA deep context aggregation semantic segmentation network also provides\npixel-level classification of the scene and its results are softly fused with\nthe detection results by the single shot detector. Our pedestrian detector\ncompared favorably to state-of-art methods on all popular pedestrian detection\ndatasets. For example, our fused DNN has better detection accuracy on the\nCaltech Pedestrian dataset than all previous state of art methods, while also\nbeing the fastest. We significantly improved the log-average miss rate on the\nCaltech pedestrian dataset to 7.67% and achieved the new state-of-the-art.', 'Autonomous vehicles (AVs) must share space with pedestrians, both in\ncarriageway cases such as cars at pedestrian crossings and off-carriageway\ncases such as delivery vehicles navigating through crowds on pedestrianized\nhigh-streets. Unlike static obstacles, pedestrians are active agents with\ncomplex, interactive motions. Planning AV actions in the presence of\npedestrians thus requires modelling of their probable future behaviour as well\nas detecting and tracking them. This narrative review article is Part I of a\npair, together surveying the current technology stack involved in this process,\norganising recent research into a hierarchical taxonomy ranging from low-level\nimage detection to high-level psychology models, from the perspective of an AV\ndesigner. This self-contained Part I covers the lower levels of this stack,\nfrom sensing, through detection and recognition, up to tracking of pedestrians.\nTechnologies at these levels are found to be mature and available as\nfoundations for use in high-level systems, such as behaviour modelling,\nprediction and interaction control.', 'The main essence of this paper is to investigate the performance of RetinaNet\nbased object detectors on pedestrian detection. Pedestrian detection is an\nimportant research topic as it provides a baseline for general object detection\nand has a great number of practical applications like autonomous car, robotics\nand Security camera. Though extensive research has made huge progress in\npedestrian detection, there are still many issues and open for more research\nand improvement. Recent deep learning based methods have shown state-of-the-art\nperformance in computer vision tasks such as image classification, object\ndetection, and segmentation. Wider pedestrian detection challenge aims at\nfinding improve solutions for pedestrian detection problem. In this paper, We\npropose a pedestrian detection system based on RetinaNet. Our solution has\nscored 0.4061 mAP. The code is available at\nhttps://github.com/miltonbd/ECCV_2018_pedestrian_detection_challenege.', ""Anticipating the intentions of vulnerable road users (VRUs) such as\npedestrians and cyclists is critical for performing safe and comfortable\ndriving maneuvers. This is the case for human driving and, thus, should be\ntaken into account by systems providing any level of driving assistance, from\nadvanced driver assistant systems (ADAS) to fully autonomous vehicles (AVs). In\nthis paper, we show how the latest advances on monocular vision-based human\npose estimation, i.e. those relying on deep Convolutional Neural Networks\n(CNNs), enable to recognize the intentions of such VRUs. In the case of\ncyclists, we assume that they follow traffic rules to indicate future maneuvers\nwith arm signals. In the case of pedestrians, no indications can be assumed.\nInstead, we hypothesize that the walking pattern of a pedestrian allows to\ndetermine if he/she has the intention of crossing the road in the path of the\nego-vehicle, so that the ego-vehicle must maneuver accordingly (e.g. slowing\ndown or stopping). In this paper, we show how the same methodology can be used\nfor recognizing pedestrians and cyclists' intentions. For pedestrians, we\nperform experiments on the JAAD dataset. For cyclists, we did not found an\nanalogous dataset, thus, we created our own one by acquiring and annotating\nvideos which we share with the research community. Overall, the proposed\npipeline provides new state-of-the-art results on the intention recognition of\nVRUs."", 'As autonomous vehicles (AVs) inch closer to reality, a central requirement\nfor acceptance will be earning the trust of humans in everyday driving\nsituations. In particular, the interaction between AVs and pedestrians is of\nhigh importance, as every human is a pedestrian at some point of the day. This\npaper considers the interaction of a pedestrian and an autonomous vehicle at a\nmid-block, unsignalized intersection where there is ambiguity over when the\npedestrian should cross and when and how the vehicle should yield. By modeling\npedestrian behavior through the concept of gap acceptance, the authors show\nthat a hybrid controller with just four distinct modes allows an autonomous\nvehicle to successfully interact with a pedestrian across a continuous spectrum\nof possible crosswalk entry behaviors. The controller is validated through\nextensive simulation and compared to an alternate POMDP solution and\nexperimental results are provided on a research vehicle for a virtual\npedestrian.', ""Pedestrian trajectory prediction is essential for collision avoidance in\nautonomous driving and robot navigation. However, predicting a pedestrian's\ntrajectory in crowded environments is non-trivial as it is influenced by other\npedestrians' motion and static structures that are present in the scene. Such\nhuman-human and human-space interactions lead to non-linearities in the\ntrajectories. In this paper, we present a new spatio-temporal graph based Long\nShort-Term Memory (LSTM) network for predicting pedestrian trajectory in\ncrowded environments, which takes into account the interaction with static\n(physical objects) and dynamic (other pedestrians) elements in the scene. Our\nresults are based on two widely-used datasets to demonstrate that the proposed\nmethod outperforms the state-of-the-art approaches in human trajectory\nprediction. In particular, our method leads to a reduction in Average\nDisplacement Error (ADE) and Final Displacement Error (FDE) of up to 55% and\n61% respectively over state-of-the-art approaches."", ""Social acceptance is a major hurdle for autonomous vehicle technology,\ncentral to which is ensuring both passengers and nearby pedestrians feel safe.\nThis idea of `feeling safe' and perceived safety is highly subjective and\nrooted in human intuition. As such, traditional analytical approaches to\nautonomous navigation often fail to cater for the social expectations of\nindividuals. Therefore, this paper proposes an approach to capture the\ncomplexity of social expectations and integrate this complexity into a\n3-layered Contextual Speed Controller. The layers were; the legal road speed\nlimit, the socially acceptable speed given the number of nearby pedestrians,\nand the socially acceptable speed based on proximity to nearby pedestrians. An\nimplementation of this layered approach was tested in areas of both low and\nhigh vehicle-pedestrian interactions. From the experiments conducted, the lower\ntwo layers were seen working in tandem to modulate the vehicle speed to\nappropriate levels that mimicked conservative human driver behaviour. In\nsummary, this work quantified the relationship between pedestrian context and\nsocially acceptable vehicle speeds, allowing for more perceivably safe\nautonomous driving. Furthermore, the need for different driving schemes for\nnavigating different road environments was identified."", 'Pedestrian trajectory prediction is a prominent research track that has\nadvanced towards modelling of crowd social and contextual interactions, with\nextensive usage of Long Short-Term Memory (LSTM) for temporal representation of\nwalking trajectories.\n  Existing approaches use virtual neighborhoods as a fixed grid for pooling\nsocial states of pedestrians with tuning process that controls how social\ninteractions are being captured. This entails performance customization to\nspecific scenes but lowers the generalization capability of the approaches. In\nour work, we deploy \\textit{Grid-LSTM}, a recent extension of LSTM, which\noperates over multidimensional feature inputs. We present a new perspective to\ninteraction modeling by proposing that pedestrian neighborhoods can become\nadaptive in design. We use \\textit{Grid-LSTM} as an encoder to learn about\npotential future neighborhoods and their influence on pedestrian motion given\nthe visual and the spatial boundaries. Our model outperforms state-of-the-art\napproaches that collate resembling features over several publicly-tested\nsurveillance videos. The experiment results clearly illustrate the\ngeneralization of our approach across datasets that varies in scene features\nand crowd dynamics.', 'Thermal images are mainly used to detect the presence of people at night or\nin bad lighting conditions, but perform poorly at daytime. To solve this\nproblem, most state-of-the-art techniques employ a fusion network that uses\nfeatures from paired thermal and color images. Instead, we propose to augment\nthermal images with their saliency maps, to serve as an attention mechanism for\nthe pedestrian detector especially during daytime. We investigate how such an\napproach results in improved performance for pedestrian detection using only\nthermal images, eliminating the need for paired color images. For our\nexperiments, we train the Faster R-CNN for pedestrian detection and report the\nadded effect of saliency maps generated using static and deep methods (PiCA-Net\nand R3-Net). Our best performing model results in an absolute reduction of miss\nrate by 13.4% and 19.4% over the baseline in day and night images respectively.\nWe also annotate and release pixel level masks of pedestrians on a subset of\nthe KAIST Multispectral Pedestrian Detection dataset, which is a first publicly\navailable dataset for salient pedestrian detection.', ""Currently, studying the vehicle-human interactive behavior in the emergency\nneeds a large amount of datasets in the actual emergent situations that are\nalmost unavailable. Existing public data sources on autonomous vehicles (AVs)\nmainly focus either on the normal driving scenarios or on emergency situations\nwithout human involvement. To fill this gap and facilitate related research,\nthis paper provides a new yet convenient way to extract the interactive\nbehavior data (i.e., the trajectories of vehicles and humans) from actual\naccident videos that were captured by both the surveillance cameras and driving\nrecorders. The main challenge for data extraction from real-time accident video\nlies in the fact that the recording cameras are un-calibrated and the angles of\nsurveillance are unknown. The approach proposed in this paper employs image\nprocessing to obtain a new perspective which is different from the original\nvideo's perspective. Meanwhile, we manually detect and mark object feature\npoints in each image frame. In order to acquire a gradient of reference ratios,\na geometric model is implemented in the analysis of reference pixel value, and\nthe feature points are then scaled to the object trajectory based on the\ngradient of ratios. The generated trajectories not only restore the object\nmovements completely but also reflect changes in vehicle velocity and rotation\nbased on the feature points distributions."", ""In order to be globally deployed, autonomous cars must guarantee the safety\nof pedestrians. This is the reason why forecasting pedestrians' intentions\nsufficiently in advance is one of the most critical and challenging tasks for\nautonomous vehicles. This work tries to solve this problem by jointly\npredicting the intention and visual states of pedestrians. In terms of visual\nstates, whereas previous work focused on x-y coordinates, we will also predict\nthe size and indeed the whole bounding box of the pedestrian. The method is a\nrecurrent neural network in a multi-task learning approach. It has one head\nthat predicts the intention of the pedestrian for each one of its future\nposition and another one predicting the visual states of the pedestrian.\nExperiments on the JAAD dataset show the superiority of the performance of our\nmethod compared to previous works for intention prediction. Also, although its\nsimple architecture (more than 2 times faster), the performance of the bounding\nbox prediction is comparable to the ones yielded by much more complex\narchitectures. Our code is available online."", ""Autonomous landing of Unmanned Aerial Vehicles (UAVs) in crowded scenarios is\ncrucial for successful deployment of UAVs in populated areas, particularly in\nemergency landing situations where the highest priority is to avoid hurting\npeople. In this work, a new visual-based algorithm for identifying Safe Landing\nZones (SLZ) in crowded scenarios is proposed, considering a camera mounted on\nan UAV, where the people in the scene move with unknown dynamics. To do so, a\ndensity map is generated for each image frame using a Deep Neural Network, from\nwhere a binary occupancy map is obtained aiming to overestimate the people's\nlocation for security reasons. Then, the occupancy map is projected to the\nhead's plane, and the SLZ candidates are obtained as circular regions in the\nhead's plane with a minimum security radius. Finally, to keep track of the SLZ\ncandidates, a multiple instance tracking algorithm is implemented using Kalman\nFilters along with the Hungarian algorithm for data association. Several\nscenarios were studied to prove the validity of the proposed strategy,\nincluding public datasets and real uncontrolled scenarios with people moving in\npublic squares, taken from an UAV in flight. The study showed promising results\nin the search of preventing the UAV from hurting people during emergency\nlanding."", 'Pedestrian detection has achieved significant progress with the availability\nof existing benchmark datasets. However, there is a gap in the diversity and\ndensity between real world requirements and current pedestrian detection\nbenchmarks: 1) most of existing datasets are taken from a vehicle driving\nthrough the regular traffic scenario, usually leading to insufficient\ndiversity; 2) crowd scenarios with highly occluded pedestrians are still under\nrepresented, resulting in low density. To narrow this gap and facilitate future\npedestrian detection research, we introduce a large and diverse dataset named\nWiderPerson for dense pedestrian detection in the wild. This dataset involves\nfive types of annotations in a wide range of scenarios, no longer limited to\nthe traffic scenario. There are a total of $13,382$ images with $399,786$\nannotations, i.e., $29.87$ annotations per image, which means this dataset\ncontains dense pedestrians with various kinds of occlusions. Hence, pedestrians\nin the proposed dataset are extremely challenging due to large variations in\nthe scenario and occlusion, which is suitable to evaluate pedestrian detectors\nin the wild. We introduce an improved Faster R-CNN and the vanilla RetinaNet to\nserve as baselines for the new pedestrian detection benchmark. Several\nexperiments are conducted on previous datasets including Caltech-USA and\nCityPersons to analyze the generalization capabilities of the proposed dataset\nand we achieve state-of-the-art performances on these previous datasets without\nbells and whistles. Finally, we analyze common failure cases and find the\nclassification ability of pedestrian detector needs to be improved to reduce\nfalse alarm and miss detection rates. The proposed dataset is available at\nhttp://www.cbsr.ia.ac.cn/users/sfzhang/WiderPerson', 'Pedestrian detection is used in many vision based applications ranging from\nvideo surveillance to autonomous driving. Despite achieving high performance,\nit is still largely unknown how well existing detectors generalize to unseen\ndata. This is important because a practical detector should be ready to use in\nvarious scenarios in applications. To this end, we conduct a comprehensive\nstudy in this paper, using a general principle of direct cross-dataset\nevaluation. Through this study, we find that existing state-of-the-art\npedestrian detectors, though perform quite well when trained and tested on the\nsame dataset, generalize poorly in cross dataset evaluation. We demonstrate\nthat there are two reasons for this trend. Firstly, their designs (e.g. anchor\nsettings) may be biased towards popular benchmarks in the traditional\nsingle-dataset training and test pipeline, but as a result largely limit their\ngeneralization capability. Secondly, the training source is generally not dense\nin pedestrians and diverse in scenarios. Under direct cross-dataset evaluation,\nsurprisingly, we find that a general purpose object detector, without\npedestrian-tailored adaptation in design, generalizes much better compared to\nexisting state-of-the-art pedestrian detectors. Furthermore, we illustrate that\ndiverse and dense datasets, collected by crawling the web, serve to be an\nefficient source of pre-training for pedestrian detection. Accordingly, we\npropose a progressive training pipeline and find that it works well for\nautonomous-driving oriented pedestrian detection. Consequently, the study\nconducted in this paper suggests that more emphasis should be put on\ncross-dataset evaluation for the future design of generalizable pedestrian\ndetectors. Code and models can be accessed at\nhttps://github.com/hasanirtiza/Pedestron.', 'In this work, we consider the problem of pedestrian detection in natural\nscenes. Intuitively, instances of pedestrians with different spatial scales may\nexhibit dramatically different features. Thus, large variance in instance\nscales, which results in undesirable large intra-category variance in features,\nmay severely hurt the performance of modern object instance detection methods.\nWe argue that this issue can be substantially alleviated by the\ndivide-and-conquer philosophy. Taking pedestrian detection as an example, we\nillustrate how we can leverage this philosophy to develop a Scale-Aware Fast\nR-CNN (SAF R-CNN) framework. The model introduces multiple built-in\nsub-networks which detect pedestrians with scales from disjoint ranges. Outputs\nfrom all the sub-networks are then adaptively combined to generate the final\ndetection results that are shown to be robust to large variance in instance\nscales, via a gate function defined over the sizes of object proposals.\nExtensive evaluations on several challenging pedestrian detection datasets well\ndemonstrate the effectiveness of the proposed SAF R-CNN. Particularly, our\nmethod achieves state-of-the-art performance on Caltech, INRIA, and ETH, and\nobtains competitive results on KITTI.', 'Smart video sensors for applications related to surveillance and security are\nIOT-based as they use Internet for various purposes. Such applications include\ncrowd behaviour monitoring and advanced decision support systems operating and\ntransmitting information over internet. The analysis of crowd and pedestrian\nbehaviour is an important task for smart IoT cameras and in particular video\nprocessing. In order to provide related behavioural models, simulation and\ntracking approaches have been considered in the literature. In both cases\nground truth is essential to train deep models and provide a meaningful\nquantitative evaluation. We propose a framework for crowd simulation and\nautomatic data generation and annotation that supports multiple cameras and\nmultiple targets. The proposed approach is based on synthetically generated\nhuman agents, augmented frames and compositing techniques combined with path\nfinding and planning methods. A number of popular crowd and pedestrian data\nsets were used to validate the model, and scenarios related to annotation and\nsimulation were considered.', 'Pedestrian detection in crowded scenes is a challenging problem, because\nocclusion happens frequently among different pedestrians. In this paper, we\npropose an effective and efficient detection network to hunt pedestrians in\ncrowd scenes. The proposed method, namely PedHunter, introduces strong\nocclusion handling ability to existing region-based detection networks without\nbringing extra computations in the inference stage. Specifically, we design a\nmask-guided module to leverage the head information to enhance the feature\nrepresentation learning of the backbone network. Moreover, we develop a strict\nclassification criterion by improving the quality of positive samples during\ntraining to eliminate common false positives of pedestrian detection in crowded\nscenes. Besides, we present an occlusion-simulated data augmentation to enrich\nthe pattern and quantity of occlusion samples to improve the occlusion\nrobustness. As a consequent, we achieve state-of-the-art results on three\npedestrian detection datasets including CityPersons, Caltech-USA and\nCrowdHuman. To facilitate further studies on the occluded pedestrian detection\nin surveillance scenes, we release a new pedestrian dataset, called SUR-PED,\nwith a total of over 162k high-quality manually labeled instances in 10k\nimages. The proposed dataset, source codes and trained models will be released.', 'Pedestrian trajectory prediction for surveillance video is one of the\nimportant research topics in the field of computer vision and a key technology\nof intelligent surveillance systems. Social relationship among pedestrians is a\nkey factor influencing pedestrian walking patterns but was mostly ignored in\nthe literature. Pedestrians with different social relationships play different\nroles in the motion decision of target pedestrian. Motivated by this idea, we\npropose a Social Relationship Attention LSTM (SRA-LSTM) model to predict future\ntrajectories. We design a social relationship encoder to obtain the\nrepresentation of their social relationship through the relative position\nbetween each pair of pedestrians. Afterwards, the social relationship feature\nand latent movements are adopted to acquire the social relationship attention\nof this pair of pedestrians. Social interaction modeling is achieved by\nutilizing social relationship attention to aggregate movement information from\nneighbor pedestrians. Experimental results on two public walking pedestrian\nvideo datasets (ETH and UCY), our model achieves superior performance compared\nwith state-of-the-art methods. Contrast experiments with other attention\nmethods also demonstrate the effectiveness of social relationship attention.', 'Monitoring vehicle flows in cities is crucial to improve the urban\nenvironment and quality of life of citizens. Images are the best sensing\nmodality to perceive and assess the flow of vehicles in large areas. Current\ntechnologies for vehicle counting in images hinge on large quantities of\nannotated data, preventing their scalability to city-scale as new cameras are\nadded to the system. This is a recurrent problem when dealing with physical\nsystems and a key research area in Machine Learning and AI. We propose and\ndiscuss a new methodology to design image-based vehicle density estimators with\nfew labeled data via multiple camera domain adaptations.', 'Autonomous valet parking is a specific application for autonomous vehicles.\nIn this task, vehicles need to navigate in narrow, crowded and GPS-denied\nparking lots. Accurate localization ability is of great importance. Traditional\nvisual-based methods suffer from tracking lost due to texture-less regions,\nrepeated structures, and appearance changes. In this paper, we exploit robust\nsemantic features to build the map and localize vehicles in parking lots.\nSemantic features contain guide signs, parking lines, speed bumps, etc, which\ntypically appear in parking lots. Compared with traditional features, these\nsemantic features are long-term stable and robust to the perspective and\nillumination change. We adopt four surround-view cameras to increase the\nperception range. Assisting by an IMU (Inertial Measurement Unit) and wheel\nencoders, the proposed system generates a global visual semantic map. This map\nis further used to localize vehicles at the centimeter level. We analyze the\naccuracy and recall of our system and compare it against other methods in real\nexperiments. Furthermore, we demonstrate the practicability of the proposed\nsystem by the autonomous parking application.', 'Deep learning-based methods for video pedestrian detection and tracking\nrequire large volumes of training data to achieve good performance. However,\ndata acquisition in crowded public environments raises data privacy concerns --\nwe are not allowed to simply record and store data without the explicit consent\nof all participants. Furthermore, the annotation of such data for computer\nvision applications usually requires a substantial amount of manual effort,\nespecially in the video domain. Labeling instances of pedestrians in highly\ncrowded scenarios can be challenging even for human annotators and may\nintroduce errors in the training data. In this paper, we study how we can\nadvance different aspects of multi-person tracking using solely synthetic data.\nTo this end, we generate MOTSynth, a large, highly diverse synthetic dataset\nfor object detection and tracking using a rendering game engine. Our\nexperiments show that MOTSynth can be used as a replacement for real data on\ntasks such as pedestrian detection, re-identification, segmentation, and\ntracking.', 'We developed a machine vision system to automatically capture the dynamics of\npedestrians under four different traffic scenarios. By considering the overhead\nview of each pedestrian as a digital object, the system processes the image\nsequences to track the pedestrians. Considering the perspective effect of the\ncamera lens and the projected area of the hallway at the top-view scene, the\ndistance of each tracked object from its original position to its current\nposition is approximated every video frame. Using the approximated distance and\nthe video frame rate (30 frames per second), the respective velocity and\nacceleration of each tracked object are later derived. The quantified motion\ncharacteristics of the pedestrians are displayed by the system through\n2-dimensional graphs of the kinematics of motion. The system also outputs video\nimages of the pedestrians with superimposed markers for tracking. These visual\nmarkers were used to visually describe and quantify the behavior of the\npedestrians under different traffic scenarios.', 'In applications such as autonomous driving, it is important to understand,\ninfer, and anticipate the intention and future behavior of pedestrians. This\nability allows vehicles to avoid collisions and improve ride safety and\nquality. This paper proposes a biomechanically inspired recurrent neural\nnetwork (Bio-LSTM) that can predict the location and 3D articulated body pose\nof pedestrians in a global coordinate frame, given 3D poses and locations\nestimated in prior frames with inaccuracy. The proposed network is able to\npredict poses and global locations for multiple pedestrians simultaneously, for\npedestrians up to 45 meters from the cameras (urban intersection scale). The\noutputs of the proposed network are full-body 3D meshes represented in Skinned\nMulti-Person Linear (SMPL) model parameters. The proposed approach relies on a\nnovel objective function that incorporates the periodicity of human walking\n(gait), the mirror symmetry of the human body, and the change of ground\nreaction forces in a human gait cycle. This paper presents prediction results\non the PedX dataset, a large-scale, in-the-wild data set collected at real\nurban intersections with heavy pedestrian traffic. Results show that the\nproposed network can successfully learn the characteristics of pedestrian gait\nand produce accurate and consistent 3D pose predictions.', 'Accurate pedestrian orientation estimation of autonomous driving helps the\nego vehicle obtain the intentions of pedestrians in the related environment,\nwhich are the base of safety measures such as collision avoidance and\nprewarning. However, because of relatively small sizes and high-level\ndeformation of pedestrians, common pedestrian orientation estimation models\nfail to extract sufficient and comprehensive information from them, thus having\ntheir performance restricted, especially monocular ones which fail to obtain\ndepth information of objects and related environment. In this paper, a novel\nmonocular pedestrian orientation estimation model, called FFNet, is proposed.\nApart from camera captures, the model adds the 2D and 3D dimensions of\npedestrians as two other inputs according to the logic relationship between\norientation and them. The 2D and 3D dimensions of pedestrians are determined\nfrom the camera captures and further utilized through two feedforward links\nconnected to the orientation estimator. The feedforward links strengthen the\nlogicality and interpretability of the network structure of the proposed model.\nExperiments show that the proposed model has at least 1.72% AOS increase than\nmost state-of-the-art models after identical training processes. The model also\nhas competitive results in orientation estimation evaluation on KITTI dataset.', 'In urban or crowded environments, humans rely on eye contact for fast and\nefficient communication with nearby people. Autonomous agents also need to\ndetect eye contact to interact with pedestrians and safely navigate around\nthem. In this paper, we focus on eye contact detection in the wild, i.e.,\nreal-world scenarios for autonomous vehicles with no control over the\nenvironment or the distance of pedestrians. We introduce a model that leverages\nsemantic keypoints to detect eye contact and show that this high-level\nrepresentation (i) achieves state-of-the-art results on the publicly-available\ndataset JAAD, and (ii) conveys better generalization properties than leveraging\nraw images in an end-to-end network. To study domain adaptation, we create\nLOOK: a large-scale dataset for eye contact detection in the wild, which\nfocuses on diverse and unconstrained scenarios for real-world generalization.\nThe source code and the LOOK dataset are publicly shared towards an open\nscience mission.', 'Interactions between vehicles and pedestrians have always been a major\nproblem in traffic safety. Experienced human drivers are able to analyze the\nenvironment and choose driving strategies that will help them avoid crashes.\nWhat is not yet clear, however, is how automated vehicles will interact with\npedestrians. This paper proposes a new method for evaluating the safety and\nfeasibility of the driving strategy of automated vehicles when encountering\nunsignalized crossings. MobilEye sensors installed on buses in Ann Arbor,\nMichigan, collected data on 2,973 valid crossing events. A stochastic\ninteraction model was then created using a multivariate Gaussian mixture model.\nThis model allowed us to simulate the movements of pedestrians reacting to an\noncoming vehicle when approaching unsignalized crossings, and to evaluate the\npassing strategies of automated vehicles. A simulation was then conducted to\ndemonstrate the evaluation procedure.', ""Studies predict that demand for autonomous vehicles will increase tenfold\nbetween 2019 and 2026. However, recent high-profile accidents have\nsignificantly impacted consumer confidence in this technology. The cause for\nmany of these accidents can be traced back to the inability of these vehicles\nto correctly sense the impending danger. In response, manufacturers have been\nimproving the already extensive on-vehicle sensor packages to ensure that the\nsystem always has access to the data necessary to ensure safe navigation.\nHowever, these sensor packages only provide a view from the vehicle's\nperspective and, as a result, autonomous vehicles still require frequent human\nintervention to ensure safety.\n  To address this issue, I developed a system, called Horus, that combines\non-vehicle and infrastructure-based sensors to provide a more complete view of\nthe environment, including areas not visible from the vehicle. I built a\nsmall-scale experimental testbed as a proof of concept. My measurements of the\nimpact of sensor failures showed that even short outages (1 second) at slow\nspeeds (25 km/hr scaled velocity) prevents vehicles that rely on on-vehicle\nsensors from navigating properly. My experiments also showed that Horus\ndramatically improves driving safety and that the sensor fusion algorithm\nselected plays a significant role in the quality of the navigation. With just a\npair of infrastructure sensors, Horus could tolerate sensors that fail 40% of\nthe time and still navigate safely. These results are a promising first step\ntowards safer autonomous vehicles."", 'Automated vehicles require a comprehensive understanding of traffic\nsituations to ensure safe and anticipatory driving. In this context, the\nprediction of pedestrians is particularly challenging as pedestrian behavior\ncan be influenced by multiple factors. In this paper, we thoroughly analyze the\nrequirements on pedestrian behavior prediction for automated driving via a\nsystem-level approach. To this end we investigate real-world pedestrian-vehicle\ninteractions with human drivers. Based on human driving behavior we then derive\nappropriate reaction patterns of an automated vehicle and determine\nrequirements for the prediction of pedestrians. This includes a novel metric\ntailored to measure prediction performance from a system-level perspective. The\nproposed metric is evaluated on a large-scale dataset comprising thousands of\nreal-world pedestrian-vehicle interactions. We furthermore conduct an ablation\nstudy to evaluate the importance of different contextual cues and compare these\nresults to ones obtained using established performance metrics for pedestrian\nprediction. Our results highlight the importance of a system-level approach to\npedestrian behavior prediction.', ""State-of-the-art pedestrian detection models have achieved great success in\nmany benchmarks. However, these models require lots of annotation information\nand the labeling process usually takes much time and efforts. In this paper, we\npropose a method to generate labeled pedestrian data and adapt them to support\nthe training of pedestrian detectors. The proposed framework is built on the\nGenerative Adversarial Network (GAN) with multiple discriminators, trying to\nsynthesize realistic pedestrians and learn the background context\nsimultaneously. To handle the pedestrians of different sizes, we adopt the\nSpatial Pyramid Pooling (SPP) layer in the discriminator. We conduct\nexperiments on two benchmarks. The results show that our framework can smoothly\nsynthesize pedestrians on background images of variations and different levels\nof details. To quantitatively evaluate our approach, we add the generated\nsamples into training data of the baseline pedestrian detectors and show the\nsynthetic images are able to improve the detectors' performance."", 'One of the major challenges that autonomous cars are facing today is driving\nin urban environments. To make it a reality, autonomous vehicles require the\nability to communicate with other road users and understand their intentions.\nSuch interactions are essential between the vehicles and pedestrians as the\nmost vulnerable road users. Understanding pedestrian behavior, however, is not\nintuitive and depends on various factors such as demographics of the\npedestrians, traffic dynamics, environmental conditions, etc. In this paper, we\nidentify these factors by surveying pedestrian behavior studies, both the\nclassical works on pedestrian-driver interaction and the modern ones that\ninvolve autonomous vehicles. To this end, we will discuss various methods of\nstudying pedestrian behavior, and analyze how the factors identified in the\nliterature are interrelated. We will also review the practical applications\naimed at solving the interaction problem including design approaches for\nautonomous vehicles that communicate with pedestrians and visual perception and\nreasoning algorithms tailored to understanding pedestrian intention. Based on\nour findings, we will discuss the open problems and propose future research\ndirections.', 'While human and group analysis have become an important area in last decades,\nsome current and relevant applications involve to estimate future motion of\npedestrians in real video sequences. This paper presents a method to provide\nmotion estimation of real pedestrians in next seconds, using crowd simulation.\nOur method is based on Physics and heuristics and use BioCrowds as crowd\nsimulation methodology to estimate future positions of people in video\nsequences. Results show that our method for estimation works well even for\ncomplex videos where events can happen. The maximum achieved average error is\n$2.72$cm when estimating the future motion of 32 pedestrians with more than 2\nseconds in advance. This paper discusses this and other results.', 'Understanding how people view and interact with autonomous vehicles is\nimportant to guide future directions of research. One such way of aiding\nunderstanding is through simulations of virtual environments involving people\nand autonomous vehicles. We present a simulation model that incorporates people\nand autonomous vehicles in a shared urban space. The model is able to simulate\nmany thousands of people and vehicles in real-time. This is achieved by use of\nGPU hardware, and through a novel linear program solver optimized for large\nnumbers of problems on the GPU. The model is up to 30 times faster than the\nequivalent multi-core CPU model.', 'This manuscript presents an efficient approach to map pedestrian surveillance\nfootage to an aerial view for global assessment of features. The analysis of\nthe footages relies on low level computer vision and enable real-time\nsurveillance. While we neglect object tracking, we introduce cumulative grids\non top view scene flow visualization to highlight situations of interest in the\nfootage. Our approach is tested on multiview footage both from RGB cameras and,\nfor the first time in the field, on RGB-D-sensors.', 'Autonomous surveillance unmanned aerial vehicles (UAVs) are deployed to\nobserve the streets of the city for any suspicious activities. This paper\nutilizes surveillance UAVs for the purpose of detecting the presence of a fire\nin the streets. An extensive database is collected from UAV surveillance\ndrones. With the aid of artificial intelligence (AI), fire stations can swiftly\nidentify the presence of a fire emerging in the neighborhood. Spatio-temporal\nsplit learning is applied to this scenario to preserve privacy and globally\ntrain a fire classification model. Fires are hazardous natural disasters that\ncan spread very quickly. Swift identification of fire is required to deploy\nfirefighters to the scene. In order to do this, strong communication between\nthe UAV and the central server where the deep learning process occurs is\nrequired. Improving communication resilience is integral to enhancing a safe\nexperience on the roads. Therefore, this paper explores the adequate number of\nclients and data ratios for split learning in this UAV setting, as well as the\nrequired network infrastructure.', 'If autonomous vehicles are to be widely accepted, we need to ensure their\nsafe operation. For this reason, verification and validation (V&V) approaches\nmust be developed that are suitable for this domain. Model checking is a formal\ntechnique which allows us to exhaustively explore the paths of an abstract\nmodel of a system. Using a probabilistic model checker such as PRISM, we may\ndetermine properties such as the expected time for a mission, or the\nprobability that a specific mission failure occurs. However, model checking of\ncomplex systems is difficult due to the loss of information during abstraction.\nThis is especially so when considering systems such as autonomous vehicles\nwhich are subject to external influences. An alternative solution is the use of\nMonte Carlo simulation to explore the results of a continuous-time model of the\nsystem. The main disadvantage of this approach is that the approach is not\nexhaustive as not all executions of the system are analysed. We are therefore\ninterested in developing a framework for formal verification of autonomous\nvehicles, using Monte Carlo simulation to inform and validate our symbolic\nmodels during the initial stages of development. In this paper, we present a\ncontinuous-time model of a quadrotor unmanned aircraft undertaking an\nautonomous mission. We employ this model in Monte Carlo simulation to obtain\nspecific mission properties which will inform the symbolic models employed in\nformal verification.', 'Building trajectories for biped robot walking is a complex task considering\nall degrees of freedom (DOFs) commonly bound within the mechanical structure. A\ntypical problem for such robots is the instability produced by violent\ntransitions between walking phases in particular when a swinging leg impacts\nthe surface. Although extensive research on novel efficient walking algorithms\nhas been conducted, falls commonly appear as the walking speed increases or as\nthe terrain condition changes. This paper presents a polynomial trajectory\ngeneration algorithm (PTA) to implement the walking on biped robots following\nthe cubic Hermitian polynomial interpolation between initial and final\nconditions. The proposed algorithm allows smooth transitions between walking\nphases, significantly reducing the possibility of falling. The algorithm has\nbeen successfully tested by generating walking trajectories under different\nterrain conditions on a biped robot of 10 DOFs. PTA has shown to be simple and\nsuitable to generate real time walking trajectories, despite reduced computing\nresources of a commercial embedded microcontroller. Experimental evidence and\ncomparisons to other state-of-the-art methods demonstrates a better performance\nof the proposed method in generating walking trajectories under different\nground conditions.', ""Despite growing attention in autonomy, there are still many open problems,\nincluding how autonomous vehicles will interact and communicate with other\nagents, such as human drivers and pedestrians. Unlike most approaches that\nfocus on pedestrian detection and planning for collision avoidance, this paper\nconsiders modeling the interaction between human drivers and pedestrians and\nhow it might influence map estimation, as a proxy for detection. We take a\nmapping inspired approach and incorporate people as sensors into mapping\nframeworks. By taking advantage of other agents' actions, we demonstrate how we\ncan impute portions of the map that would otherwise be occluded. We evaluate\nour framework in human driving experiments and on real-world data, using\noccupancy grids and landmark-based mapping approaches. Our approach\nsignificantly improves overall environment awareness and out-performs standard\nmapping techniques."", 'Paper-by-paper results make it easy to miss the forest for the trees.We\nanalyse the remarkable progress of the last decade by discussing the main ideas\nexplored in the 40+ detectors currently present in the Caltech pedestrian\ndetection benchmark. We observe that there exist three families of approaches,\nall currently reaching similar detection quality. Based on our analysis, we\nstudy the complementarity of the most promising ideas by combining multiple\npublished strategies. This new decision forest detector achieves the current\nbest known performance on the challenging Caltech-USA dataset.', 'Machine learning algorithms have found several applications in the field of\nrobotics and control systems. The control systems community has started to show\ninterest towards several machine learning algorithms from the sub-domains such\nas supervised learning, imitation learning and reinforcement learning to\nachieve autonomous control and intelligent decision making. Amongst many\ncomplex control problems, stable bipedal walking has been the most challenging\nproblem. In this paper, we present an architecture to design and simulate a\nplanar bipedal walking robot(BWR) using a realistic robotics simulator, Gazebo.\nThe robot demonstrates successful walking behaviour by learning through several\nof its trial and errors, without any prior knowledge of itself or the world\ndynamics. The autonomous walking of the BWR is achieved using reinforcement\nlearning algorithm called Deep Deterministic Policy Gradient(DDPG). DDPG is one\nof the algorithms for learning controls in continuous action spaces. After\ntraining the model in simulation, it was observed that, with a proper shaped\nreward function, the robot achieved faster walking or even rendered a running\ngait with an average speed of 0.83 m/s. The gait pattern of the bipedal walker\nwas compared with the actual human walking pattern. The results show that the\nbipedal walking pattern had similar characteristics to that of a human walking\npattern. The video presenting our experiment is available at\nhttps://goo.gl/NHXKqR.', 'Typical methods for pedestrian detection focus on either tackling mutual\nocclusions between crowded pedestrians, or dealing with the various scales of\npedestrians. Detecting pedestrians with substantial appearance diversities such\nas different pedestrian silhouettes, different viewpoints or different\ndressing, remains a crucial challenge. Instead of learning each of these\ndiverse pedestrian appearance features individually as most existing methods\ndo, we propose to perform contrastive learning to guide the feature learning in\nsuch a way that the semantic distance between pedestrians with different\nappearances in the learned feature space is minimized to eliminate the\nappearance diversities, whilst the distance between pedestrians and background\nis maximized. To facilitate the efficiency and effectiveness of contrastive\nlearning, we construct an exemplar dictionary with representative pedestrian\nappearances as prior knowledge to construct effective contrastive training\npairs and thus guide contrastive learning. Besides, the constructed exemplar\ndictionary is further leveraged to evaluate the quality of pedestrian proposals\nduring inference by measuring the semantic distance between the proposal and\nthe exemplar dictionary. Extensive experiments on both daytime and nighttime\npedestrian detection validate the effectiveness of the proposed method.', 'Identity documents recognition is an important sub-field of document\nanalysis, which deals with tasks of robust document detection, type\nidentification, text fields recognition, as well as identity fraud prevention\nand document authenticity validation given photos, scans, or video frames of an\nidentity document capture. Significant amount of research has been published on\nthis topic in recent years, however a chief difficulty for such research is\nscarcity of datasets, due to the subject matter being protected by security\nrequirements. A few datasets of identity documents which are available lack\ndiversity of document types, capturing conditions, or variability of document\nfield values. In addition, the published datasets were typically designed only\nfor a subset of document recognition problems, not for a complex identity\ndocument analysis. In this paper, we present a dataset MIDV-2020 which consists\nof 1000 video clips, 2000 scanned images, and 1000 photos of 1000 unique mock\nidentity documents, each with unique text field values and unique artificially\ngenerated faces, with rich annotation. For the presented benchmark dataset\nbaselines are provided for such tasks as document location and identification,\ntext fields recognition, and face detection. With 72409 annotated images in\ntotal, to the date of publication the proposed dataset is the largest publicly\navailable identity documents dataset with variable artificially generated data,\nand we believe that it will prove invaluable for advancement of the field of\ndocument analysis and recognition. The dataset is available for download at\nftp://smartengines.com/midv-2020 and http://l3i-share.univ-lr.fr .', 'My overarching research goal is to provide robots with perceptional abilities\nthat allow interactions with humans in a human-like manner. To develop these\nperceptional abilities, I believe that it is useful to study the principles of\nthe human visual system. I use these principles to develop new computer vision\nalgorithms and validate their effectiveness in intelligent robotic systems. I\nam enthusiastic about this approach as it offers the dual benefit of uncovering\nprinciples inherent in the human visual system, as well as applying these\nprinciples to its artificial counterpart. Fig. 1 contains a depiction of my\nresearch.', 'Multi-pedestrian tracking in aerial imagery has several applications such as\nlarge-scale event monitoring, disaster management, search-and-rescue missions,\nand as input into predictive crowd dynamic models. Due to the challenges such\nas the large number and the tiny size of the pedestrians (e.g., 4 x 4 pixels)\nwith their similar appearances as well as different scales and atmospheric\nconditions of the images with their extremely low frame rates (e.g., 2 fps),\ncurrent state-of-the-art algorithms including the deep learning-based ones are\nunable to perform well. In this paper, we propose AerialMPTNet, a novel\napproach for multi-pedestrian tracking in geo-referenced aerial imagery by\nfusing appearance features from a Siamese Neural Network, movement predictions\nfrom a Long Short-Term Memory, and pedestrian interconnections from a GraphCNN.\nIn addition, to address the lack of diverse aerial pedestrian tracking\ndatasets, we introduce the Aerial Multi-Pedestrian Tracking (AerialMPT) dataset\nconsisting of 307 frames and 44,740 pedestrians annotated. We believe that\nAerialMPT is the largest and most diverse dataset to this date and will be\nreleased publicly. We evaluate AerialMPTNet on AerialMPT and KIT AIS, and\nbenchmark with several state-of-the-art tracking methods. Results indicate that\nAerialMPTNet significantly outperforms other methods on accuracy and\ntime-efficiency.', ""Convolututional Neural Networks have achieved state of the art in image\nclassification, object detection and other image related tasks. In this paper I\npresent another use of CNNs i.e. if given a set of images and then giving a\nsingle test image the network identifies that the test image is part of which\nimage from the images given before. This is a task somehow similar to measuring\nimage similarity and can be done using a simple CNN. Doing this task manually\nby looping can be quite a time consuming problem and won't be a generalizable\nsolution. The task is quite similar to doing object detection but for that lots\ntraining data should be given or in the case of sliding window it takes lot of\ntime and my algorithm can work with much fewer examples, is totally\nunsupervised and works much efficiently. Also, I explain that how unsupervised\nalgorithm like K-Means or supervised algorithm like K-NN are not good enough to\nperform this task. The basic idea is that image encodings are collected for\neach image from a CNN, when a test image comes it is replaced by a part of\noriginal image, the encoding is generated using the same network, the frobenius\nnorm is calculated and if it comes under a tolerance level then the test image\nis said to be the part of the original image."", 'Deep convolutional neural networks continue to advance the state-of-the-art\nin many domains as they grow bigger and more complex. It has been observed that\nmany of the parameters of a large network are redundant, allowing for the\npossibility of learning a smaller network that mimics the outputs of the large\nnetwork through a process called Knowledge Distillation. We show, however, that\nstandard Knowledge Distillation is not effective for learning small models for\nthe task of pedestrian detection. To improve this process, we introduce a\nhigher-dimensional hint layer to increase information flow. We also estimate\nthe variance in the outputs of the large network and propose a loss function to\nincorporate this uncertainty. Finally, we attempt to boost the complexity of\nthe small network without increasing its size by using as input hand-designed\nfeatures that have been demonstrated to be effective for pedestrian detection.\nWe succeed in training a model that contains $400\\times$ fewer parameters than\nthe large network while outperforming AlexNet on the Caltech Pedestrian\nDataset.', 'One of the objectives of understanding pedestrian behavior is to predict the\neffect of proposed changes in the design or evaluation of pedestrian\nfacilities. We want to know the impact to the user of the facilities, as the\ndesign of the facilities change. That impact was traditionally evaluated by\nlevel of service standards. Another design criterion to measure the impact of\ndesign change is measured by the pedestrian flow performance index. This paper\ndescribes the determination of pedestrian flow performance based video tracking\nor any microscopic pedestrian simulation models. Most of pedestrian researches\nhave been done on a macroscopic level, which is an aggregation of all\npedestrian movement in pedestrian areas into flow, average speed and area\nmodule. Macroscopic level, however, does not consider the interaction between\npedestrians. It is also not well suited for prediction of pedestrian flow\nperformance in pedestrian areas or in buildings with some obstruction, that\nreduces the effective width of the walkways. On the other hand, the microscopic\nlevel has a more general usage and considers detail in the design. More\nefficient pedestrian flow can even be reached with less space. Those results\nhave rejected the linearity assumption of space and flow in the macroscopic\nlevel.', 'Deep learning-based computer vision is usually data-hungry. Many researchers\nattempt to augment datasets with synthesized data to improve model robustness.\nHowever, the augmentation of popular pedestrian datasets, such as Caltech and\nCitypersons, can be extremely challenging because real pedestrians are commonly\nin low quality. Due to the factors like occlusions, blurs, and low-resolution,\nit is significantly difficult for existing augmentation approaches, which\ngenerally synthesize data using 3D engines or generative adversarial networks\n(GANs), to generate realistic-looking pedestrians. Alternatively, to access\nmuch more natural-looking pedestrians, we propose to augment pedestrian\ndetection datasets by transforming real pedestrians from the same dataset into\ndifferent shapes. Accordingly, we propose the Shape Transformation-based\nDataset Augmentation (STDA) framework. The proposed framework is composed of\ntwo subsequent modules, i.e. the shape-guided deformation and the environment\nadaptation. In the first module, we introduce a shape-guided warping field to\nhelp deform the shape of a real pedestrian into a different shape. Then, in the\nsecond stage, we propose an environment-aware blending map to better adapt the\ndeformed pedestrians into surrounding environments, obtaining more\nrealistic-looking pedestrians and more beneficial augmentation results for\npedestrian detection. Extensive empirical studies on different pedestrian\ndetection benchmarks show that the proposed STDA framework consistently\nproduces much better augmentation results than other pedestrian synthesis\napproaches using low-quality pedestrians. By augmenting the original datasets,\nour proposed framework also improves the baseline pedestrian detector by up to\n38% on the evaluated benchmarks, achieving state-of-the-art performance.', ""Early detection of psychological distress is key to effective treatment.\nAutomatic detection of distress, such as depression, is an active area of\nresearch. Current approaches utilise vocal, facial, and bodily modalities. Of\nthese, the bodily modality is the least investigated, partially due to the\ndifficulty in extracting bodily representations from videos, and partially due\nto the lack of viable datasets. Existing body modality approaches use automatic\ncategorization of expressions to represent body language as a series of\nspecific expressions, much like words within natural language. In this\ndissertation I present a new type of feature, within the body modality, that\nrepresents meta information of gestures, such as speed, and use it to predict a\nnon-clinical depression label. This differs to existing work by representing\noverall behaviour as a small set of aggregated meta features derived from a\nperson's movement. In my method I extract pose estimation from videos, detect\ngestures within body parts, extract meta information from individual gestures,\nand finally aggregate these features to generate a small feature vector for use\nin prediction tasks. I introduce a new dataset of 65 video recordings of\ninterviews with self-evaluated distress, personality, and demographic labels.\nThis dataset enables the development of features utilising the whole body in\ndistress detection tasks. I evaluate my newly introduced meta-features for\npredicting depression, anxiety, perceived stress, somatic stress, five standard\npersonality measures, and gender. A linear regression based classifier using\nthese features achieves a 82.70% F1 score for predicting depression within my\nnovel dataset."", 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'For the simulation of pedestrians a method is introduced to find routing\nalternatives from any origin position to a given destination area in a given\ngeometry composed of walking areas and obstacles. The method includes a\nparameter which sets a threshold for the approximate minimum size of obstacles\nto generate routing alternatives. The resulting data structure for navigation\nis constructed such that it does not introduce artifacts to the movement of\nsimulated pedestrians and that locally pedestrians prefer to walk on the\nshortest path. The generated set of routes can be used with iterating static or\ndynamic assignment methods.', 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.', 'Computation of document image quality metrics often depends upon the\navailability of a ground truth image corresponding to the document. This limits\nthe applicability of quality metrics in applications such as hyperparameter\noptimization of image processing algorithms that operate on-the-fly on unseen\ndocuments. This work proposes the use of surrogate models to learn the behavior\nof a given document quality metric on existing datasets where ground truth\nimages are available. The trained surrogate model can later be used to predict\nthe metric value on previously unseen document images without requiring access\nto ground truth images. The surrogate model is empirically evaluated on the\nDocument Image Binarization Competition (DIBCO) and the Handwritten Document\nImage Binarization Competition (H-DIBCO) datasets.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : In archaeology, a lot of old photos are produced by polarization cameras, and I am trying to reconstruct an accurate 3D model of an object using images from these polarization cameras. However, I have difficulties determining the surface of the object because of the features of the polarization camera. I need a model that can analyze the angle of polarization of reflected light in captured images. I think both geometric and photometric cues extracted from input color polarization images should be utilized to complete the analysis. I also need a model that can optimize photometric rendering errors and handle different materials given an initial standard geometric reconstruction. I will test the models on synthetic and real data and use them as an inspiration for my model.","[""Reconstructing an object's high-quality 3D shape with inherent spectral\nreflectance property, beyond typical device-dependent RGB albedos, opens the\ndoor to applications requiring a high-fidelity 3D model in terms of both\ngeometry and photometry. In this paper, we propose a novel Multi-View Inverse\nRendering (MVIR) method called Spectral MVIR for jointly reconstructing the 3D\nshape and the spectral reflectance for each point of object surfaces from\nmulti-view images captured using a standard RGB camera and low-cost lighting\nequipment such as an LED bulb or an LED projector. Our main contributions are\ntwofold: (i) We present a rendering model that considers both geometric and\nphotometric principles in the image formation by explicitly considering camera\nspectral sensitivity, light's spectral power distribution, and light source\npositions. (ii) Based on the derived model, we build a cost-optimization MVIR\nframework for the joint reconstruction of the 3D shape and the per-vertex\nspectral reflectance while estimating the light source positions and the\nshadows. Different from most existing spectral-3D acquisition methods, our\nmethod does not require expensive special equipment and cumbersome geometric\ncalibration. Experimental results using both synthetic and real-world data\ndemonstrate that our Spectral MVIR can acquire a high-quality 3D model with\naccurate spectral reflectance property."", 'This paper makes a first attempt to bring the Shape from Polarization (SfP)\nproblem to the realm of deep learning. The previous state-of-the-art methods\nfor SfP have been purely physics-based. We see value in these principled\nmodels, and blend these physical models as priors into a neural network\narchitecture. This proposed approach achieves results that exceed the previous\nstate-of-the-art on a challenging dataset we introduce. This dataset consists\nof polarization images taken over a range of object textures, paints, and\nlighting conditions. We report that our proposed method achieves the lowest\ntest error on each tested condition in our dataset, showing the value of\nblending data-driven and physics-driven approaches.', 'We present a novel method for efficient acquisition of shape and spatially\nvarying reflectance of 3D objects using polarization cues. Unlike previous\nworks that have exploited polarization to estimate material or object\nappearance under certain constraints (known shape or multiview acquisition), we\nlift such restrictions by coupling polarization imaging with deep learning to\nachieve high quality estimate of 3D object shape (surface normals and depth)\nand SVBRDF using single-view polarization imaging under frontal flash\nillumination. In addition to acquired polarization images, we provide our deep\nnetwork with strong novel cues related to shape and reflectance, in the form of\na normalized Stokes map and an estimate of diffuse color. We additionally\ndescribe modifications to network architecture and training loss which provide\nfurther qualitative improvements. We demonstrate our approach to achieve\nsuperior results compared to recent works employing deep learning in\nconjunction with flash illumination.', 'Semantic Segmentation (SS) is promising for outdoor scene perception in\nsafety-critical applications like autonomous vehicles, assisted navigation and\nso on. However, traditional SS is primarily based on RGB images, which limits\nthe reliability of SS in complex outdoor scenes, where RGB images lack\nnecessary information dimensions to fully perceive unconstrained environments.\nAs preliminary investigation, we examine SS in an unexpected obstacle detection\nscenario, which demonstrates the necessity of multimodal fusion. Thereby, in\nthis work, we present EAFNet, an Efficient Attention-bridged Fusion Network to\nexploit complementary information coming from different optical sensors.\nSpecifically, we incorporate polarization sensing to obtain supplementary\ninformation, considering its optical characteristics for robust representation\nof diverse materials. By using a single-shot polarization sensor, we build the\nfirst RGB-P dataset which consists of 394 annotated pixel-aligned\nRGB-Polarization images. A comprehensive variety of experiments shows the\neffectiveness of EAFNet to fuse polarization and RGB information, as well as\nthe flexibility to be adapted to other sensor combination scenarios.', 'Autonomous robotics is critically affected by the robustness of its scene\nunderstanding algorithms. We propose a two-axis pipeline based on polarization\nindices to analyze dynamic urban scenes. As robots evolve in unknown\nenvironments, they are prone to encountering specular obstacles. Usually,\nspecular phenomena are rarely taken into account by algorithms which causes\nmisinterpretations and erroneous estimates. By exploiting all the light\nproperties, systems can greatly increase their robustness to events. In\naddition to the conventional photometric characteristics, we propose to include\npolarization sensing.\n  We demonstrate in this paper that the contribution of polarization\nmeasurement increases both the performances of segmentation and the quality of\ndepth estimation. Our polarimetry-based approaches are compared here with other\nstate-of-the-art RGB-centric methods showing interest of using polarization\nimaging.', ""This paper describes a multimodal vision sensor that integrates three types\nof cameras, including a stereo camera, a polarization camera and a panoramic\ncamera. Each sensor provides a specific dimension of information: the stereo\ncamera measures depth per pixel, the polarization obtains the degree of\npolarization, and the panoramic camera captures a 360-degree landscape. Data\nfusion and advanced environment perception could be built upon the combination\nof sensors. Designed especially for autonomous driving, this vision sensor is\nshipped with a robust semantic segmentation network. In addition, we\ndemonstrate how cross-modal enhancement could be achieved by registering the\ncolor image and the polarization image. An example of water hazard detection is\ngiven. To prove the multimodal vision sensor's compatibility with different\ndevices, a brief runtime performance analysis is carried out."", 'The stomatopod (mantis shrimp) visual system has recently provided a\nblueprint for the design of paradigm-shifting polarization and multispectral\nimaging sensors, enabling solutions to challenging medical and remote sensing\nproblems. However, these bioinspired sensors lack the high dynamic range (HDR)\nand asynchronous polarization vision capabilities of the stomatopod visual\nsystem, limiting temporal resolution to \\~12 ms and dynamic range to \\~ 72 dB.\nHere we present a novel stomatopod-inspired polarization camera which mimics\nthe sustained and transient biological visual pathways to save power and sample\ndata beyond the maximum Nyquist frame rate. This bio-inspired sensor\nsimultaneously captures both synchronous intensity frames and asynchronous\npolarization brightness change information with sub-millisecond latencies over\na million-fold range of illumination. Our PDAVIS camera is comprised of 346x260\npixels, organized in 2-by-2 macropixels, which filter the incoming light with\nfour linear polarization filters offset by 45 degrees. Polarization information\nis reconstructed using both low cost and latency event-based algorithms and\nmore accurate but slower deep neural networks. Our sensor is used to image HDR\npolarization scenes which vary at high speeds and to observe dynamical\nproperties of single collagen fibers in bovine tendon under rapid cyclical\nloads', 'This paper tackles the problem of estimating 3D body shape of clothed humans\nfrom single polarized 2D images, i.e. polarization images. Polarization images\nare known to be able to capture polarized reflected lights that preserve rich\ngeometric cues of an object, which has motivated its recent applications in\nreconstructing surface normal of the objects of interest. Inspired by the\nrecent advances in human shape estimation from single color images, in this\npaper, we attempt at estimating human body shapes by leveraging the geometric\ncues from single polarization images. A dedicated two-stage deep learning\napproach, SfP, is proposed: given a polarization image, stage one aims at\ninferring the fined-detailed body surface normal; stage two gears to\nreconstruct the 3D body shape of clothing details. Empirical evaluations on a\nsynthetic dataset (SURREAL) as well as a real-world dataset (PHSPD) demonstrate\nthe qualitative and quantitative performance of our approach in estimating\nhuman poses and shapes. This indicates polarization camera is a promising\nalternative to the more conventional color or depth imaging for human shape\nestimation. Further, normal maps inferred from polarization imaging play a\nsignificant role in accurately recovering the body shapes of clothed people.', 'Photometric stereo provides an important method for high-fidelity 3D\nreconstruction based on multiple intensity images captured under different\nillumination directions. In this paper, we present a complete framework,\nincluding a multilight source illumination and acquisition hardware system and\na two-stage convolutional neural network (CNN) architecture, to construct\ninter- and intraframe representations for accurate normal estimation of\nnon-Lambertian objects. We experimentally investigate numerous network design\nalternatives for identifying the optimal scheme to deploy inter- and intraframe\nfeature extraction modules for the photometric stereo problem. Moreover, we\npropose utilizing the easily obtained object mask to eliminate adverse\ninterference from invalid background regions in intraframe spatial\nconvolutions, thus effectively improving the accuracy of normal estimation for\nsurfaces made of dark materials or with cast shadows. Experimental results\ndemonstrate that the proposed masked two-stage photometric stereo CNN model\n(MT-PS-CNN) performs favourably against state-of-the-art photometric stereo\ntechniques in terms of both accuracy and efficiency. In addition, the proposed\nmethod is capable of predicting accurate and rich surface normal details for\nnon-Lambertian objects of complex geometry and performs stably given inputs\ncaptured in both sparse and dense lighting distributions.', 'This paper addresses the problem of simultaneous 3D reconstruction and\nmaterial recognition and segmentation. Enabling robots to recognise different\nmaterials (concrete, metal etc.) in a scene is important for many tasks, e.g.\nrobotic interventions in nuclear decommissioning. Previous work on 3D semantic\nreconstruction has predominantly focused on recognition of everyday domestic\nobjects (tables, chairs etc.), whereas previous work on material recognition\nhas largely been confined to single 2D images without any 3D reconstruction.\nMeanwhile, most 3D semantic reconstruction methods rely on computationally\nexpensive post-processing, using Fully-Connected Conditional Random Fields\n(CRFs), to achieve consistent segmentations. In contrast, we propose a deep\nlearning method which performs 3D reconstruction while simultaneously\nrecognising different types of materials and labelling them at the pixel level.\nUnlike previous methods, we propose a fully end-to-end approach, which does not\nrequire hand-crafted features or CRF post-processing. Instead, we use only\nlearned features, and the CRF segmentation constraints are incorporated inside\nthe fully end-to-end learned system. We present the results of experiments, in\nwhich we trained our system to perform real-time 3D semantic reconstruction for\n23 different materials in a real-world application. The run-time performance of\nthe system can be boosted to around 10Hz, using a conventional GPU, which is\nenough to achieve real-time semantic reconstruction using a 30fps RGB-D camera.\nTo the best of our knowledge, this work is the first real-time end-to-end\nsystem for simultaneous 3D reconstruction and material recognition.', 'Photometric Stereo methods seek to reconstruct the 3d shape of an object from\nmotionless images obtained with varying illumination. Most existing methods\nsolve a restricted problem where the physical reflectance model, such as\nLambertian reflectance, is known in advance. In contrast, we do not restrict\nourselves to a specific reflectance model. Instead, we offer a method that\nworks on a wide variety of reflectances. Our approach uses a simple yet\nuncommonly used property of the problem - the sought after normals are points\non a unit hemisphere. We present a novel embedding method that maps pixels to\nnormals on the unit hemisphere. Our experiments demonstrate that this approach\noutperforms existing manifold learning methods for the task of hemisphere\nembedding. We further show successful reconstructions of objects from a wide\nvariety of reflectances including smooth, rough, diffuse and specular surfaces,\neven in the presence of significant attached shadows. Finally, we empirically\nprove that under these challenging settings we obtain more accurate shape\nreconstructions than existing methods.', 'In this paper we present ADOP, a novel point-based, differentiable neural\nrendering pipeline. Like other neural renderers, our system takes as input\ncalibrated camera images and a proxy geometry of the scene, in our case a point\ncloud. To generate a novel view, the point cloud is rasterized with learned\nfeature vectors as colors and a deep neural network fills the remaining holes\nand shades each output pixel. The rasterizer renders points as one-pixel\nsplats, which makes it very fast and allows us to compute gradients with\nrespect to all relevant input parameters efficiently. Furthermore, our pipeline\ncontains a fully differentiable physically-based photometric camera model,\nincluding exposure, white balance, and a camera response function. Following\nthe idea of inverse rendering, we use our renderer to refine its input in order\nto reduce inconsistencies and optimize the quality of its output. In\nparticular, we can optimize structural parameters like the camera pose, lens\ndistortions, point positions and features, and a neural environment map, but\nalso photometric parameters like camera response function, vignetting, and\nper-image exposure and white balance. Because our pipeline includes photometric\nparameters, e.g.~exposure and camera response function, our system can smoothly\nhandle input images with varying exposure and white balance, and generates\nhigh-dynamic range output. We show that due to the improved input, we can\nachieve high render quality, also for difficult input, e.g. with imperfect\ncamera calibrations, inaccurate proxy geometry, or varying exposure. As a\nresult, a simpler and thus faster deep neural network is sufficient for\nreconstruction. In combination with the fast point rasterization, ADOP achieves\nreal-time rendering rates even for models with well over 100M points.\nhttps://github.com/darglein/ADOP', 'We present a new data-driven approach with physics-based priors to\nscene-level normal estimation from a single polarization image. Existing shape\nfrom polarization (SfP) works mainly focus on estimating the normal of a single\nobject rather than complex scenes in the wild. A key barrier to high-quality\nscene-level SfP is the lack of real-world SfP data in complex scenes. Hence, we\ncontribute the first real-world scene-level SfP dataset with paired input\npolarization images and ground-truth normal maps. Then we propose a\nlearning-based framework with a multi-head self-attention module and viewing\nencoding, which is designed to handle increasing polarization ambiguities\ncaused by complex materials and non-orthographic projection in scene-level SfP.\nOur trained model can be generalized to far-field outdoor scenes as the\nrelationship between polarized light and surface normals is not affected by\ndistance. Experimental results demonstrate that our approach significantly\noutperforms existing SfP models on two datasets. Our dataset and source code\nwill be publicly available at https://github.com/ChenyangLEI/sfp-wild']","['This paper presents an uncalibrated deep neural network framework for the\nphotometric stereo problem. For training models to solve the problem, existing\nneural network-based methods either require exact light directions or\nground-truth surface normals of the object or both. However, in practice, it is\nchallenging to procure both of this information precisely, which restricts the\nbroader adoption of photometric stereo algorithms for vision application. To\nbypass this difficulty, we propose an uncalibrated neural inverse rendering\napproach to this problem. Our method first estimates the light directions from\nthe input images and then optimizes an image reconstruction loss to calculate\nthe surface normals, bidirectional reflectance distribution function value, and\ndepth. Additionally, our formulation explicitly models the concave and convex\nparts of a complex surface to consider the effects of interreflections in the\nimage formation process. Extensive evaluation of the proposed method on the\nchallenging subjects generally shows comparable or better results than the\nsupervised and classical approaches.', 'In current practice, scene survey is carried out by workers using total\nstations. The method has high accuracy, but it incurs high costs if continuous\nmonitoring is needed. Techniques based on photogrammetry, with the relatively\ncheaper digital cameras, have gained wide applications in many fields. Besides\npoint measurement, photogrammetry can also create a three-dimensional (3D)\nmodel of the scene. Accurate 3D model reconstruction depends on high quality\nimages. Degraded images will result in large errors in the reconstructed 3D\nmodel. In this paper, we propose a method that can be used to improve the\nvisibility of the images, and eventually reduce the errors of the 3D scene\nmodel. The idea is inspired by image dehazing. Each original image is first\ntransformed into multiple exposure images by means of gamma-correction\noperations and adaptive histogram equalization. The transformed images are\nanalyzed by the computation of the local binary patterns. The image is then\nenhanced, with each pixel generated from the set of transformed image pixels\nweighted by a function of the local pattern feature and image saturation.\nPerformance evaluation has been performed on benchmark image dehazing datasets.\nExperimentations have been carried out on outdoor and indoor surveys. Our\nanalysis finds that the method works on different types of degradation that\nexist in both outdoor and indoor images. When fed into the photogrammetry\nsoftware, the enhanced images can reconstruct 3D scene models with\nsub-millimeter mean errors.', 'Shape from Polarization (SfP) estimates surface normals using photos captured\nat different polarizer rotations. Fundamentally, the SfP model assumes that\nlight is reflected either diffusely or specularly. However, this model is not\nvalid for many real-world surfaces exhibiting a mixture of diffuse and specular\nproperties. To address this challenge, previous methods have used a sequential\nsolution: first, use an existing algorithm to separate the scene into diffuse\nand specular components, then apply the appropriate SfP model. In this paper,\nwe propose a new method that jointly uses viewpoint and polarization data to\nholistically separate diffuse and specular components, recover refractive\nindex, and ultimately recover 3D shape. By involving the physics of\npolarization in the separation process, we demonstrate competitive results with\na benchmark method, while recovering additional information (e.g. refractive\nindex).', 'We present a novel framework to learn to convert the perpixel photometric\ninformation at each view into spatially distinctive and view-invariant\nlow-level features, which can be plugged into existing multi-view stereo\npipeline for enhanced 3D reconstruction. Both the illumination conditions\nduring acquisition and the subsequent per-pixel feature transform can be\njointly optimized in a differentiable fashion. Our framework automatically\nadapts to and makes efficient use of the geometric information available in\ndifferent forms of input data. High-quality 3D reconstructions of a variety of\nchallenging objects are demonstrated on the data captured with an illumination\nmultiplexing device, as well as a point light. Our results compare favorably\nwith state-of-the-art techniques.', 'The emergence of the single-chip polarized color sensor now allows for\nsimultaneously capturing chromatic and polarimetric information of the scene on\na monochromatic image plane. However, unlike the usual camera with an embedded\ndemosaicing method, the latest polarized color camera is not delivered with an\nin-built demosaicing tool. For demosaicing, the users have to down-sample the\ncaptured images or to use traditional interpolation techniques. Neither of them\ncan perform well since the polarization and color are interdependent.\nTherefore, joint chromatic and polarimetric demosaicing is the key to obtaining\nhigh-quality polarized color images. In this paper, we propose a joint\nchromatic and polarimetric demosaicing model to address this challenging\nproblem. Instead of mechanically demosaicing for the multi-channel polarized\ncolor image, we further present a sparse representation-based optimization\nstrategy that utilizes chromatic information and polarimetric information to\njointly optimize the model. To avoid the interaction between color and\npolarization during demosaicing, we separately construct the corresponding\ndictionaries. We also build an optical data acquisition system to collect a\ndataset, which contains various sources of polarization, such as illumination,\nreflectance and birefringence. Results of both qualitative and quantitative\nexperiments have shown that our method is capable of faithfully recovering full\nRGB information of four polarization angles for each pixel from a single mosaic\ninput image. Moreover, the proposed method can perform well not only on the\nsynthetic data but the real captured data.', 'Neural implicit surfaces have become an important technique for multi-view 3D\nreconstruction but their accuracy remains limited. In this paper, we argue that\nthis comes from the difficulty to learn and render high frequency textures with\nneural networks. We thus propose to add to the standard neural rendering\noptimization a direct photo-consistency term across the different views.\nIntuitively, we optimize the implicit geometry so that it warps views on each\nother in a consistent way. We demonstrate that two elements are key to the\nsuccess of such an approach: (i) warping entire patches, using the predicted\noccupancy and normals of the 3D points along each ray, and measuring their\nsimilarity with a robust structural similarity (SSIM); (ii) handling visibility\nand occlusion in such a way that incorrect warps are not given too much\nimportance while encouraging a reconstruction as complete as possible. We\nevaluate our approach, dubbed NeuralWarp, on the standard DTU and EPFL\nbenchmarks and show it outperforms state of the art unsupervised implicit\nsurfaces reconstructions by over 20% on both datasets.', 'Since specular reflection often exists in the real captured images and causes\ndeviation between the recorded color and intrinsic color, specular reflection\nseparation can bring advantages to multiple applications that require\nconsistent object surface appearance. However, due to the color of an object is\nsignificantly influenced by the color of the illumination, the existing\nresearches still suffer from the near-duplicate challenge, that is, the\nseparation becomes unstable when the illumination color is close to the surface\ncolor. In this paper, we derive a polarization guided model to incorporate the\npolarization information into a designed iteration optimization separation\nstrategy to separate the specular reflection. Based on the analysis of\npolarization, we propose a polarization guided model to generate a polarization\nchromaticity image, which is able to reveal the geometrical profile of the\ninput image in complex scenarios, such as diversity of illumination. The\npolarization chromaticity image can accurately cluster the pixels with similar\ndiffuse color. We further use the specular separation of all these clusters as\nan implicit prior to ensure that the diffuse components will not be mistakenly\nseparated as the specular components. With the polarization guided model, we\nreformulate the specular reflection separation into a unified optimization\nfunction which can be solved by the ADMM strategy. The specular reflection will\nbe detected and separated jointly by RGB and polarimetric information. Both\nqualitative and quantitative experimental results have shown that our method\ncan faithfully separate the specular reflection, especially in some challenging\nscenarios.', ""Stereo cameras are a popular choice for obstacle avoidance for outdoor\nlighweight, low-cost robotics applications. However, they are unable to sense\nthin and reflective objects well. Currently, many algorithms are tuned to\nperform well on indoor scenes like the Middlebury dataset. When navigating\noutdoors, reflective objects, like windows and glass, and thin obstacles, like\nwires, are not well handled by most stereo disparity algorithms. Reflections,\nrepeating patterns and objects parallel to the cameras' baseline causes\nmismatches between image pairs which leads to bad disparity estimates. Thin\nobstacles are difficult for many sliding window based disparity methods to\ndetect because they do not take up large portions of the pixels in the sliding\nwindow. We use a trinocular camera setup and micropolarizer camera capable of\ndetecting reflective objects to overcome these issues. We present a\nhierarchical disparity algorithm that reduces noise, separately identify wires\nusing semantic object triangulation in three images, and use information about\nthe polarization of light to estimate the disparity of reflective objects. We\nevaluate our approach on outdoor data that we collected. Our method contained\nan average of 9.27% of bad pixels compared to a typical stereo algorithm's\n18.4% of bad pixels in scenes containing reflective objects. Our trinocular and\nsemantic wire disparity methods detected 53% of wire pixels, whereas a typical\ntwo camera stereo algorithm detected 5%."", 'Several approaches to image stitching use different constraints to estimate\nthe motion model between image pairs. These constraints can be roughly divided\ninto two categories: geometric constraints and photometric constraints. In this\npaper, geometric and photometric constraints are combined to improve the\nalignment quality, which is based on the observation that these two kinds of\nconstraints are complementary. On the one hand, geometric constraints (e.g.,\npoint and line correspondences) are usually spatially biased and are\ninsufficient in some extreme scenes, while photometric constraints are always\nevenly and densely distributed. On the other hand, photometric constraints are\nsensitive to displacements and are not suitable for images with large\nparallaxes, while geometric constraints are usually imposed by feature matching\nand are more robust to handle parallaxes. The proposed method therefore\ncombines them together in an efficient mesh-based image warping framework. It\nachieves better alignment quality than methods only with geometric constraints,\nand can handle larger parallax than photometric-constraint-based method.\nExperimental results on various images illustrate that the proposed method\noutperforms representative state-of-the-art image stitching methods reported in\nthe literature.', 'This paper focuses on a new problem of estimating human pose and shape from\nsingle polarization images. Polarization camera is known to be able to capture\nthe polarization of reflected lights that preserves rich geometric cues of an\nobject surface. Inspired by the recent applications in surface normal\nreconstruction from polarization images, in this paper, we attempt to estimate\nhuman pose and shape from single polarization images by leveraging the\npolarization-induced geometric cues. A dedicated two-stage pipeline is\nproposed: given a single polarization image, stage one (Polar2Normal) focuses\non the fine detailed human body surface normal estimation; stage two\n(Polar2Shape) then reconstructs clothed human shape from the polarization image\nand the estimated surface normal. To empirically validate our approach, a\ndedicated dataset (PHSPD) is constructed, consisting of over 500K frames with\naccurate pose and parametric shape annotations. Empirical evaluations on this\nreal-world dataset as well as a synthetic dataset, SURREAL, demonstrate the\neffectiveness of our approach. It suggests polarization camera as a promising\nalternative to the more conventional RGB camera for human pose and shape\nestimation.', 'Inspired by the recent advance of image-based object reconstruction using\ndeep learning, we present an active reconstruction model using a guided view\nplanner. We aim to reconstruct a 3D model using images observed from a planned\nsequence of informative and discriminative views. But where are such\ninformative and discriminative views around an object? To address this we\npropose a unified model for view planning and object reconstruction, which is\nutilized to learn a guided information acquisition model and to aggregate\ninformation from a sequence of images for reconstruction. Experiments show that\nour model (1) increases our reconstruction accuracy with an increasing number\nof views (2) and generally predicts a more informative sequence of views for\nobject reconstruction compared to other alternative methods.', 'The problem of obtaining dense reconstruction of an object in a natural\nsequence of images has been long studied in computer vision. Classically this\nproblem has been solved through the application of bundle adjustment (BA). More\nrecently, excellent results have been attained through the application of\nphotometric bundle adjustment (PBA) methods -- which directly minimize the\nphotometric error across frames. A fundamental drawback to BA & PBA, however,\nis: (i) their reliance on having to view all points on the object, and (ii) for\nthe object surface to be well textured. To circumvent these limitations we\npropose semantic PBA which incorporates a 3D object prior, obtained through\ndeep learning, within the photometric bundle adjustment problem. We demonstrate\nstate of the art performance in comparison to leading methods for object\nreconstruction across numerous natural sequences.', 'Polarization images are known to be able to capture polarized reflected\nlights that preserve rich geometric cues of an object, which has motivated its\nrecent applications in reconstructing detailed surface normal of the objects of\ninterest. Meanwhile, inspired by the recent breakthroughs in human shape\nestimation from a single color image, we attempt to investigate the new\nquestion of whether the geometric cues from polarization camera could be\nleveraged in estimating detailed human body shapes. This has led to the\ncuration of Polarization Human Shape and Pose Dataset (PHSPD), our home-grown\npolarization image dataset of various human shapes and poses.', ""Current self-driving car systems operate well in sunny weather but struggle\nin adverse conditions. One of the most commonly encountered adverse conditions\ninvolves water on the road caused by rain, sleet, melting snow or flooding.\nWhile some advances have been made in using conventional RGB camera and LIDAR\ntechnology for detecting water hazards, other sources of information such as\npolarization offer a promising and potentially superior approach to this\nproblem in terms of performance and cost. In this paper, we present a novel\nstereo-polarization system for detecting and tracking water hazards based on\npolarization and color variation of reflected light, with consideration of the\neffect of polarized light from sky as function of reflection and azimuth\nangles. To evaluate this system, we present a new large `water on road'\ndatasets spanning approximately 2 km of driving in various on-road and off-road\nconditions and demonstrate for the first time reliable water detection and\ntracking over a wide range of realistic car driving water conditions using\npolarized vision as the primary sensing modality. Our system successfully\ndetects water hazards up to more than 100m. Finally, we discuss several\ninteresting challenges and propose future research directions for further\nimproving robust autonomous car perception in hazardous wet conditions using\npolarization sensors."", 'Wide field-of-view (FOV) cameras, which capture a larger scene area than\nnarrow FOV cameras, are used in many applications including 3D reconstruction,\nautonomous driving, and video surveillance. However, wide-angle images contain\ndistortions that violate the assumptions underlying pinhole camera models,\nresulting in object distortion, difficulties in estimating scene distance,\narea, and direction, and preventing the use of off-the-shelf deep models\ntrained on undistorted images for downstream computer vision tasks. Image\nrectification, which aims to correct these distortions, can solve these\nproblems. In this paper, we comprehensively survey progress in wide-angle image\nrectification from transformation models to rectification methods.\nSpecifically, we first present a detailed description and discussion of the\ncamera models used in different approaches. Then, we summarize several\ndistortion models including radial distortion and projection distortion. Next,\nwe review both traditional geometry-based image rectification methods and deep\nlearning-based methods, where the former formulate distortion parameter\nestimation as an optimization problem and the latter treat it as a regression\nproblem by leveraging the power of deep neural networks. We evaluate the\nperformance of state-of-the-art methods on public datasets and show that\nalthough both kinds of methods can achieve good results, these methods only\nwork well for specific camera models and distortion types. We also provide a\nstrong baseline model and carry out an empirical study of different distortion\nmodels on synthetic datasets and real-world wide-angle images. Finally, we\ndiscuss several potential research directions that are expected to further\nadvance this area in the future.', 'In this paper, we present a novel deep method to reconstruct a point cloud of\nan object from a single still image. Prior arts in the field struggle to\nreconstruct an accurate and scalable 3D model due to either the inefficient and\nexpensive 3D representations, the dependency between the output and number of\nmodel parameters or the lack of a suitable computing operation. We propose to\novercome these by deforming a random point cloud to the object shape through\ntwo steps: feature blending and deformation. In the first step, the global and\npoint-specific shape features extracted from a 2D object image are blended with\nthe encoded feature of a randomly generated point cloud, and then this mixture\nis sent to the deformation step to produce the final representative point set\nof the object. In the deformation process, we introduce a new layer termed as\nGraphX that considers the inter-relationship between points like common graph\nconvolutions but operates on unordered sets. Moreover, with a simple trick, the\nproposed model can generate an arbitrary-sized point cloud, which is the first\ndeep method to do so. Extensive experiments verify that we outperform existing\nmodels and halve the state-of-the-art distance score in single image 3D\nreconstruction.', ""Simultaneous reconstruction of geometry and reflectance properties in\nuncontrolled environments remains a challenging problem. In this paper, we\npropose an efficient method to reconstruct the scene's 3D geometry and\nreflectance from multi-view photography using conventional hand-held cameras.\nOur method automatically builds a virtual scene in a differentiable rendering\nsystem that roughly matches the real world's scene parameters, optimized by\nminimizing photometric objectives alternatingly and stochastically. With the\noptimal scene parameters evaluated, photo-realistic novel views for various\nviewing angles and distances can then be generated by our approach. We present\nthe results of captured scenes with complex geometry and various reflection\ntypes. Our method also shows superior performance compared to state-of-the-art\nalternatives in novel view synthesis visually and quantitatively."", 'This paper presents a method of passive non-line-of-sight (NLOS) imaging\nusing polarization cues. A key observation is that the oblique light has a\ndifferent polarimetric signal. It turns out this effect is due to the\npolarization axis rotation, a phenomena which can be used to better condition\nthe light transport matrix for non-line-of-sight imaging. Our analysis and\nresults show that the use of a polarizer in front of the camera is not only a\nseparate technique, but it can be seen as an enhancement technique for more\nadvanced forms of passive NLOS imaging. For example, this paper shows that\npolarization can enhance passive NLOS imaging both with and without occluders.\nIn all tested cases, despite the light attenuation from polarization optics,\nrecovery of the occluded images is improved.', 'We present a novel 3D face reconstruction technique that leverages sparse\nphotometric stereo (PS) and latest advances on face registration/modeling from\na single image. We observe that 3D morphable faces approach provides a\nreasonable geometry proxy for light position calibration. Specifically, we\ndevelop a robust optimization technique that can calibrate per-pixel lighting\ndirection and illumination at a very high precision without assuming uniform\nsurface albedos. Next, we apply semantic segmentation on input images and the\ngeometry proxy to refine hairy vs. bare skin regions using tailored filters.\nExperiments on synthetic and real data show that by using a very small set of\nimages, our technique is able to reconstruct fine geometric details such as\nwrinkles, eyebrows, whelks, pores, etc, comparable to and sometimes surpassing\nmovie quality productions.', 'Photometric stereo is a method that seeks to reconstruct the normal vectors\nof an object from a set of images of the object illuminated under different\nlight sources. While effective in some situations, classical photometric stereo\nrelies on a diffuse surface model that cannot handle objects with complex\nreflectance patterns, and it is sensitive to non-idealities in the images. In\nthis work, we propose a novel approach to photometric stereo that relies on\ndictionary learning to produce robust normal vector reconstructions.\nSpecifically, we develop two formulations for applying dictionary learning to\nphotometric stereo. We propose a model that applies dictionary learning to\nregularize and reconstruct the normal vectors from the images under the classic\nLambertian reflectance model. We then generalize this model to explicitly model\nnon-Lambertian objects. We investigate both approaches through extensive\nexperimentation on synthetic and real benchmark datasets and observe\nstate-of-the-art performance compared to existing robust photometric stereo\nmethods.', 'We present a method to capture both 3D shape and spatially varying\nreflectance with a multi-view photometric stereo (MVPS) technique that works\nfor general isotropic materials. Our algorithm is suitable for perspective\ncameras and nearby point light sources. Our data capture setup is simple, which\nconsists of only a digital camera, some LED lights, and an optional automatic\nturntable. From a single viewpoint, we use a set of photometric stereo images\nto identify surface points with the same distance to the camera. We collect\nthis information from multiple viewpoints and combine it with\nstructure-from-motion to obtain a precise reconstruction of the complete 3D\nshape. The spatially varying isotropic bidirectional reflectance distribution\nfunction (BRDF) is captured by simultaneously inferring a set of basis BRDFs\nand their mixing weights at each surface point. In experiments, we demonstrate\nour algorithm with two different setups: a studio setup for highest precision\nand a desktop setup for best usability. According to our experiments, under the\nstudio setting, the captured shapes are accurate to 0.5 millimeters and the\ncaptured reflectance has a relative root-mean-square error (RMSE) of 9%. We\nalso quantitatively evaluate state-of-the-art MVPS on a newly collected\nbenchmark dataset, which is publicly available for inspiring future research.', 'This paper is concerned with polarimetric dense map reconstruction based on a\npolarization camera with the help of relative depth information as a prior. In\ngeneral, polarization imaging is able to reveal information about surface\nnormal such as azimuth and zenith angles, which can support the development of\nsolutions to the problem of dense reconstruction, especially in texture-poor\nregions. However, polarimetric shape cues are ambiguous due to two types of\npolarized reflection (specular/diffuse). Although methods have been proposed to\naddress this issue, they either are offline and therefore not practical in\nrobotics applications, or use incomplete polarimetric cues, leading to\nsub-optimal performance. In this paper, we propose an online reconstruction\nmethod that uses full polarimetric cues available from the polarization camera.\nWith our online method, we can propagate sparse depth values both along and\nperpendicular to iso-depth contours. Through comprehensive experiments on\nchallenging image sequences, we demonstrate that our method is able to\nsignificantly improve the accuracy of the depthmap as well as increase its\ndensity, specially in regions of poor texture.', 'The recent development of the on-chip micro-polarizer technology has made it\npossible to acquire four spatially aligned and temporally synchronized\npolarization images with the same ease of operation as a conventional camera.\nIn this paper, we investigate the use of this sensor technology in\nhigh-dynamic-range (HDR) imaging. Specifically, observing that natural light\ncan be attenuated differently by varying the orientation of the polarization\nfilter, we treat the multiple images captured by the polarization camera as a\nset captured under different exposure times. In our approach, we first study\nthe relationship among polarizer orientation, degree and angle of polarization\nof light to the exposure time of a pixel in the polarization image.\nSubsequently, we propose a deep snapshot HDR reconstruction framework to\nrecover an HDR image using the polarization images. A polarized HDR dataset is\ncreated to train and evaluate our approach. We demonstrate that our approach\nperforms favorably against state-of-the-art HDR reconstruction algorithms.', 'In this paper, we address the problem of 3D object mesh reconstruction from\nRGB videos. Our approach combines the best of multi-view geometric and\ndata-driven methods for 3D reconstruction by optimizing object meshes for\nmulti-view photometric consistency while constraining mesh deformations with a\nshape prior. We pose this as a piecewise image alignment problem for each mesh\nface projection. Our approach allows us to update shape parameters from the\nphotometric error without any depth or mask information. Moreover, we show how\nto avoid a degeneracy of zero photometric gradients via rasterizing from a\nvirtual viewpoint. We demonstrate 3D object mesh reconstruction results from\nboth synthetic and real-world videos with our photometric mesh optimization,\nwhich is unachievable with either na\\""ive mesh generation networks or\ntraditional pipelines of surface reconstruction without heavy manual\npost-processing.', 'Pixel-wise operations between polarimetric images are important for\nprocessing polarization information. For the lack of such operations, the\npolarization information cannot be fully utilized in convolutional neural\nnetwork(CNN). In this paper, a novel end-to-end CNN framework for polarization\nvision tasks is proposed, which enables the networks to take full advantage of\npolarimetric images. The framework consists of two sub-networks: a\npolarization-parameter-constructing network (PPCN) and a task network. PPCN\nimplements pixel-wise operations between images in the CNN form with 1x1\nconvolution kernels. It takes raw polarimetric images as input, and outputs\npolarization-parametric images to task network so as to complete a vison task.\nBy training together, the PPCN can learn to provide the most suitable\npolarization-parametric images for the task network and the dataset. Taking\nfaster R-CNN as task network, the experimental results show that compared with\nexisting methods, the proposed framework achieves much higher\nmean-average-precision (mAP) in object detection task', 'This paper addresses the problem of photometric stereo for non-Lambertian\nsurfaces. Existing approaches often adopt simplified reflectance models to make\nthe problem more tractable, but this greatly hinders their applications on\nreal-world objects. In this paper, we propose a deep fully convolutional\nnetwork, called PS-FCN, that takes an arbitrary number of images of a static\nobject captured under different light directions with a fixed camera as input,\nand predicts a normal map of the object in a fast feed-forward pass. Unlike the\nrecently proposed learning based method, PS-FCN does not require a pre-defined\nset of light directions during training and testing, and can handle multiple\nimages and light directions in an order-agnostic manner. Although we train\nPS-FCN on synthetic data, it can generalize well on real datasets. We further\nshow that PS-FCN can be easily extended to handle the problem of uncalibrated\nphotometric stereo.Extensive experiments on public real datasets show that\nPS-FCN outperforms existing approaches in calibrated photometric stereo, and\npromising results are achieved in uncalibrated scenario, clearly demonstrating\nits effectiveness.', 'We introduce a novel learning-based method to reconstruct the high-quality\ngeometry and complex, spatially-varying BRDF of an arbitrary object from a\nsparse set of only six images captured by wide-baseline cameras under\ncollocated point lighting. We first estimate per-view depth maps using a deep\nmulti-view stereo network; these depth maps are used to coarsely align the\ndifferent views. We propose a novel multi-view reflectance estimation network\narchitecture that is trained to pool features from these coarsely aligned\nimages and predict per-view spatially-varying diffuse albedo, surface normals,\nspecular roughness and specular albedo. We do this by jointly optimizing the\nlatent space of our multi-view reflectance network to minimize the photometric\nerror between images rendered with our predictions and the input images. While\nprevious state-of-the-art methods fail on such sparse acquisition setups, we\ndemonstrate, via extensive experiments on synthetic and real data, that our\nmethod produces high-quality reconstructions that can be used to render\nphotorealistic images.', 'Reconstruction of geometric structures from images using supervised learning\nsuffers from limited available amount of accurate data. One type of such data\nis accurate real-world RGB-D images. A major challenge in acquiring such ground\ntruth data is the accurate alignment between RGB images and the point cloud\nmeasured by a depth scanner. To overcome this difficulty, we consider a\ndifferential optimization method that aligns a colored point cloud with a given\ncolor image through iterative geometric and color matching. In the proposed\nframework, the optimization minimizes the photometric difference between the\ncolors of the point cloud and the corresponding colors of the image pixels.\nUnlike other methods that try to reduce this photometric error, we analyze the\ncomputation of the gradient on the image plane and propose a different direct\nscheme. We assume that the colors produced by the geometric scanner camera and\nthe color camera sensor are different and therefore characterized by different\nchromatic acquisition properties. Under these multimodal conditions, we find\nthe transformation between the camera image and the point cloud colors. We\nalternately optimize for aligning the position of the point cloud and matching\nthe different color spaces. The alignments produced by the proposed method are\ndemonstrated on both synthetic data with quantitative evaluation and real\nscenes with qualitative results.', 'Recently, 3D face reconstruction from a single image has achieved great\nsuccess with the help of deep learning and shape prior knowledge, but they\noften fail to produce accurate geometry details. On the other hand, photometric\nstereo methods can recover reliable geometry details, but require dense inputs\nand need to solve a complex optimization problem. In this paper, we present a\nlightweight strategy that only requires sparse inputs or even a single image to\nrecover high-fidelity face shapes with images captured under near-field lights.\nTo this end, we construct a dataset containing 84 different subjects with 29\nexpressions under 3 different lights. Data augmentation is applied to enrich\nthe data in terms of diversity in identity, lighting, expression, etc. With\nthis constructed dataset, we propose a novel neural network specially designed\nfor photometric stereo based 3D face reconstruction. Extensive experiments and\ncomparisons demonstrate that our method can generate high-quality\nreconstruction results with one to three facial images captured under\nnear-field lights. Our full framework is available at\nhttps://github.com/Juyong/FacePSNet.', ""This paper proposes an artificial neural network to determine orientation\nusing polarized skylight. This neural network has specific dilated convolution,\nwhich can extract light intensity information of different polarization\ndirections. Then, the degree of polarization (DOP) and angle of polarization\n(AOP) are directly extracted in the network. In addition, the exponential\nfunction encoding of orientation is designed as the network output, which can\nbetter reflect the insect's encoding of polarization information, and improve\nthe accuracy of orientation determination. Finally, training and testing were\nconducted on a public polarized skylight navigation dataset, and the\nexperimental results proved the stability and effectiveness of the network."", ""We present a method for estimating neural scenes representations of objects\ngiven only a single image. The core of our method is the estimation of a\ngeometric scaffold for the object and its use as a guide for the reconstruction\nof the underlying radiance field. Our formulation is based on a generative\nprocess that first maps a latent code to a voxelized shape, and then renders it\nto an image, with the object appearance being controlled by a second latent\ncode. During inference, we optimize both the latent codes and the networks to\nfit a test image of a new object. The explicit disentanglement of shape and\nappearance allows our model to be fine-tuned given a single image. We can then\nrender new views in a geometrically consistent manner and they represent\nfaithfully the input object. Additionally, our method is able to generalize to\nimages outside of the training domain (more realistic renderings and even real\nphotographs). Finally, the inferred geometric scaffold is itself an accurate\nestimate of the object's 3D shape. We demonstrate in several experiments the\neffectiveness of our approach in both synthetic and real images."", 'We propose a method to refine geometry of 3D meshes from a consumer level\ndepth camera, e.g. Kinect, by exploiting shading cues captured from an infrared\n(IR) camera. A major benefit to using an IR camera instead of an RGB camera is\nthat the IR images captured are narrow band images that filter out most\nundesired ambient light, which makes our system robust against natural indoor\nillumination. Moreover, for many natural objects with colorful textures in the\nvisible spectrum, the subjects appear to have a uniform albedo in the IR\nspectrum. Based on our analyses on the IR projector light of the Kinect, we\ndefine a near light source IR shading model that describes the captured\nintensity as a function of surface normals, albedo, lighting direction, and\ndistance between light source and surface points. To resolve the ambiguity in\nour model between the normals and distances, we utilize an initial 3D mesh from\nthe Kinect fusion and multi-view information to reliably estimate surface\ndetails that were not captured and reconstructed by the Kinect fusion. Our\napproach directly operates on the mesh model for geometry refinement. We ran\nexperiments on our algorithm for geometries captured by both the Kinect I and\nKinect II, as the depth acquisition in Kinect I is based on a structured-light\ntechnique and that of the Kinect II is based on a time-of-flight (ToF)\ntechnology. The effectiveness of our approach is demonstrated through several\nchallenging real-world examples. We have also performed a user study to\nevaluate the quality of the mesh models before and after our refinements.', 'While the problem of estimating shapes and diffuse reflectances of human\nfaces from images has been extensively studied, there is relatively less work\ndone on recovering the specular albedo. This paper presents a lightweight\nsolution for inferring photorealistic facial reflectance and geometry. Our\nsystem processes video streams from two views of a subject, and outputs two\nreflectance maps for diffuse and specular albedos, as well as a vector map of\nsurface normals. A model-based optimization approach is used, consisting of the\nthree stages of multi-view face model fitting, facial reflectance inference and\nfacial geometry refinement. Our approach is based on a novel formulation built\nupon the 3D morphable model (3DMM) for representing 3D textured faces in\nconjunction with the Blinn-Phong reflection model. It has the advantage of\nrequiring only a simple setup with two video streams, and is able to exploit\nthe interaction between the diffuse and specular reflections across multiple\nviews as well as time frames. As a result, the method is able to reliably\nrecover high-fidelity facial reflectance and geometry, which facilitates\nvarious applications such as generating photorealistic facial images under new\nviewpoints or illumination conditions.', 'Event cameras have a lot of advantages over traditional cameras, such as low\nlatency, high temporal resolution, and high dynamic range. However, since the\noutputs of event cameras are the sequences of asynchronous events overtime\nrather than actual intensity images, existing algorithms could not be directly\napplied. Therefore, it is demanding to generate intensity images from events\nfor other tasks. In this paper, we unlock the potential of event camera-based\nconditional generative adversarial networks to create images/videos from an\nadjustable portion of the event data stream. The stacks of space-time\ncoordinates of events are used as inputs and the network is trained to\nreproduce images based on the spatio-temporal intensity changes. The usefulness\nof event cameras to generate high dynamic range(HDR) images even in extreme\nillumination conditions and also non blurred images under rapid motion is also\nshown.In addition, the possibility of generating very high frame rate videos is\ndemonstrated, theoretically up to 1 million frames per second (FPS) since the\ntemporal resolution of event cameras are about 1{\\mu}s. Proposed methods are\nevaluated by comparing the results with the intensity images captured on the\nsame pixel grid-line of events using online available real datasets and\nsynthetic datasets produced by the event camera simulator.', 'Direct methods have shown promise on visual odometry and SLAM, leading to\ngreater accuracy and robustness over feature-based methods. However, offline\n3-d reconstruction from internet images has not yet benefited from a joint,\nphotometric optimization over dense geometry and camera parameters. Issues such\nas the lack of brightness constancy, and the sheer volume of data, make this a\nmore challenging task. This work presents a framework for jointly optimizing\nmillions of scene points and hundreds of camera poses and intrinsics, using a\nphotometric cost that is invariant to local lighting changes. The improvement\nin metric reconstruction accuracy that it confers over feature-based bundle\nadjustment is demonstrated on the large-scale Tanks & Temples benchmark. We\nfurther demonstrate qualitative reconstruction improvements on an internet\nphoto collection, with challenging diversity in lighting and camera intrinsics.', 'Our previous work has demonstrated that Rayleigh model, which is widely used\nin polarized skylight navigation to describe skylight polarization patterns,\ndoes not contain three-dimensional (3D) attitude information [1]. However, it\nis still necessary to further explore whether the skylight polarization\npatterns contain 3D attitude information. So, in this paper, a social spider\noptimization (SSO) method is proposed to estimate three Euler angles, which\nconsiders the difference of each pixel among polarization images based on\ntemplate matching (TM) to make full use of the captured polarization\ninformation. In addition, to explore this problem, we not only use angle of\npolarization (AOP) and degree of polarization (DOP) information, but also the\nlight intensity (LI) information. So, a sky model is established, which\ncombines Berry model and Hosek model to fully describe AOP, DOP, and LI\ninformation in the sky, and considers the influence of four neutral points,\nground albedo, atmospheric turbidity, and wavelength. The results of simulation\nshow that the SSO algorithm can estimate 3D attitude and the established sky\nmodel contains 3D attitude information. However, when there are measurement\nnoise or model error, the accuracy of 3D attitude estimation drops\nsignificantly. Especially in field experiment, it is very difficult to estimate\n3D attitude. Finally, the results are discussed in detail.', 'Polarized light microscopy provides high contrast to birefringent specimen\nand is widely used as a diagnostic tool in pathology. However, polarization\nmicroscopy systems typically operate by analyzing images collected from two or\nmore light paths in different states of polarization, which lead to relatively\ncomplex optical designs, high system costs or experienced technicians being\nrequired. Here, we present a deep learning-based holographic polarization\nmicroscope that is capable of obtaining quantitative birefringence retardance\nand orientation information of specimen from a phase recovered hologram, while\nonly requiring the addition of one polarizer/analyzer pair to an existing\nholographic imaging system. Using a deep neural network, the reconstructed\nholographic images from a single state of polarization can be transformed into\nimages equivalent to those captured using a single-shot computational polarized\nlight microscope (SCPLM). Our analysis shows that a trained deep neural network\ncan extract the birefringence information using both the sample specific\nmorphological features as well as the holographic amplitude and phase\ndistribution. To demonstrate the efficacy of this method, we tested it by\nimaging various birefringent samples including e.g., monosodium urate (MSU) and\ntriamcinolone acetonide (TCA) crystals. Our method achieves similar results to\nSCPLM both qualitatively and quantitatively, and due to its simpler optical\ndesign and significantly larger field-of-view, this method has the potential to\nexpand the access to polarization microscopy and its use for medical diagnosis\nin resource limited settings.', 'A division-of-focal-plane or microgrid image polarimeter enables us to\nacquire a set of polarization images in one shot. Since the polarimeter\nconsists of an image sensor equipped with a monochrome or color polarization\nfilter array (MPFA or CPFA), the demosaicking process to interpolate missing\npixel values plays a crucial role in obtaining high-quality polarization\nimages. In this paper, we propose a novel MPFA demosaicking method based on\nedge-aware residual interpolation (EARI) and also extend it to CPFA\ndemosaicking. The key of EARI is a new edge detector for generating an\neffective guide image used to interpolate the missing pixel values. We also\npresent a newly constructed full color-polarization image dataset captured\nusing a 3-CCD camera and a rotating polarizer. Using the dataset, we\nexperimentally demonstrate that our EARI-based method outperforms existing\nmethods in MPFA and CPFA demosaicking.', 'In 3D reconstruction, the recovery of the calibration parameters of the\ncameras is paramount since it provides metric information about the observed\nscene, e.g., measures of angles and ratios of distances. Autocalibration\nenables the estimation of the camera parameters without using a calibration\ndevice, but by enforcing simple constraints on the camera parameters. In the\nabsence of information about the internal camera parameters such as the focal\nlength and the principal point, the knowledge of the camera pixel shape is\nusually the only available constraint. Given a projective reconstruction of a\nrigid scene, we address the problem of the autocalibration of a minimal set of\ncameras with known pixel shape and otherwise arbitrarily varying intrinsic and\nextrinsic parameters. We propose an algorithm that only requires 5 cameras (the\ntheoretical minimum), thus halving the number of cameras required by previous\nalgorithms based on the same constraint. To this purpose, we introduce as our\nbasic geometric tool the six-line conic variety (SLCV), consisting in the set\nof planes intersecting six given lines of 3D space in points of a conic. We\nshow that the set of solutions of the Euclidean upgrading problem for three\ncameras with known pixel shape can be parameterized in a computationally\nefficient way. This parameterization is then used to solve autocalibration from\nfive or more cameras, reducing the three-dimensional search space to a\ntwo-dimensional one. We provide experiments with real images showing the good\nperformance of the technique.', 'The reconstruction of dense 3D models of face geometry and appearance from a\nsingle image is highly challenging and ill-posed. To constrain the problem,\nmany approaches rely on strong priors, such as parametric face models learned\nfrom limited 3D scan data. However, prior models restrict generalization of the\ntrue diversity in facial geometry, skin reflectance and illumination. To\nalleviate this problem, we present the first approach that jointly learns 1) a\nregressor for face shape, expression, reflectance and illumination on the basis\nof 2) a concurrently learned parametric face model. Our multi-level face model\ncombines the advantage of 3D Morphable Models for regularization with the\nout-of-space generalization of a learned corrective space. We train end-to-end\non in-the-wild images without dense annotations by fusing a convolutional\nencoder with a differentiable expert-designed renderer and a self-supervised\ntraining loss, both defined at multiple detail levels. Our approach compares\nfavorably to the state-of-the-art in terms of reconstruction quality, better\ngeneralizes to real world faces, and runs at over 250 Hz.', 'We present a new color photometric stereo (CPS) method that recovers high\nquality, detailed 3D face geometry in a single shot. Our system uses three\nuncalibrated near point lights of different colors and a single camera. For\nrobust self-calibration of the light sources, we use 3D morphable model (3DMM)\nand semantic segmentation of facial parts. We address the spectral ambiguity\nproblem by incorporating albedo consensus, albedo similarity, and proxy prior\ninto a unified framework. We avoid the need for spatial constancy of albedo;\ninstead, we use a new measure for albedo similarity that is based on the albedo\nnorm profile. Experiments show that our new approach produces state-of-the-art\nresults from single image with high-fidelity geometry that includes details\nsuch as wrinkles.', ""Removing perspective distortion from hand held camera captured document\nimages is one of the primitive tasks in document analysis, but unfortunately,\nno such method exists that can reliably remove the perspective distortion from\ndocument images automatically. In this paper, we propose a convolutional neural\nnetwork based method for recovering homography from hand-held camera captured\ndocuments.\n  Our proposed method works independent of document's underlying content and is\ntrained end-to-end in a fully automatic way. Specifically, this paper makes\nfollowing three contributions: Firstly, we introduce a large scale synthetic\ndataset for recovering homography from documents images captured under\ndifferent geometric and photometric transformations; secondly, we show that a\ngeneric convolutional neural network based architecture can be successfully\nused for regressing the corners positions of documents captured under wild\nsettings; thirdly, we show that L1 loss can be reliably used for corners\nregression. Our proposed method gives state-of-the-art performance on the\ntested datasets, and has potential to become an integral part of document\nanalysis pipeline."", 'The reflections caused by common semi-reflectors, such as glass windows, can\nimpact the performance of computer vision algorithms. State-of-the-art methods\ncan remove reflections on synthetic data and in controlled scenarios. However,\nthey are based on strong assumptions and do not generalize well to real-world\nimages. Contrary to a common misconception, real-world images are challenging\neven when polarization information is used. We present a deep learning approach\nto separate the reflected and the transmitted components of the recorded\nirradiance, which explicitly uses the polarization properties of light. To\ntrain it, we introduce an accurate synthetic data generation pipeline, which\nsimulates realistic reflections, including those generated by curved and\nnon-ideal surfaces, non-static scenes, and high-dynamic-range scenes.', 'Scientific imaging techniques such as optical and electron microscopy and\ncomputed tomography (CT) scanning are used to study the 3D structure of an\nobject through 2D observations. These observations are related to the original\n3D object through orthogonal integral projections. For common 3D reconstruction\nalgorithms, computational efficiency requires the modeling of the 3D structures\nto take place in Fourier space by applying the Fourier slice theorem. At\npresent, it is unclear how to differentiate through the projection operator,\nand hence current learning algorithms can not rely on gradient based methods to\noptimize 3D structure models. In this paper we show how back-propagation\nthrough the projection operator in Fourier space can be achieved. We\ndemonstrate the validity of the approach with experiments on 3D reconstruction\nof proteins. We further extend our approach to learning probabilistic models of\n3D objects. This allows us to predict regions of low sampling rates or estimate\nnoise. A higher sample efficiency can be reached by utilizing the learned\nuncertainties of the 3D structure as an unsupervised estimate of the model fit.\nFinally, we demonstrate how the reconstruction algorithm can be extended with\nan amortized inference scheme on unknown attributes such as object pose.\nThrough empirical studies we show that joint inference of the 3D structure and\nthe object pose becomes more difficult when the ground truth object contains\nmore symmetries. Due to the presence of for instance (approximate) rotational\nsymmetries, the pose estimation can easily get stuck in local optima,\ninhibiting a fine-grained high-quality estimate of the 3D structure.', 'Visual perception is regularly used by humans and robots for navigation. By\neither implicitly or explicitly mapping the environment, ego-motion can be\ndetermined and a path of actions can be planned. The process of mapping and\nnavigation are delicately intertwined; therefore, improving one can often lead\nto an improvement of the other. Both processes are sensitive to the interior\norientation parameters of the camera system and mathematically modelling these\nsystematic errors can often improve the precision and accuracy of the overall\nsolution. This paper presents an automatic camera calibration method suitable\nfor any lens, without having prior knowledge about the sensor. Statistical\ninference is performed to map the environment and localize the camera\nsimultaneously. K-nearest neighbour regression is used to model the geometric\ndistortions of the images. A normal-angle lens Nikon camera and wide-angle lens\nGoPro camera were calibrated using the proposed method, as well as the\nconventional bundle adjustment with self-calibration method (for comparison).\nResults showed that the mapping error was reduced from an average of 14.9 mm to\n1.2 mm (i.e. a 92% improvement) and 66.6 mm to 1.5 mm (i.e. a 98% improvement)\nusing the proposed method for the Nikon and GoPro cameras, respectively. In\ncontrast, the conventional approach achieved an average 3D error of 0.9 mm\n(i.e. 94% improvement) and 3.3 mm (i.e. 95% improvement) for the Nikon and\nGoPro cameras, respectively. Thus, the proposed method performs well\nirrespective of the lens/sensor used: it yields results that are comparable to\nthe conventional approach for normal-angle lens cameras, and it has the\nadditional benefit of improving calibration results for wide-angle lens\ncameras.', 'This paper shows that accurate underwater 3D shape reconstruction is possible\nusing a single camera, observing a target through a refractive interface. We\nprovide unified reconstruction techniques for a variety of scenarios such as\nsingle static camera and moving refractive interface, single moving camera and\nstatic refractive interface, and single moving camera and moving refractive\ninterface. In our basic setup, we assume that the refractive interface is\nplanar, and simultaneously estimate the unknown transformations of the planar\ninterface and the camera, and the unknown target shape using bundle adjustment.\nWe also extend it to relax the planarity assumption, which enables us to use\nwaves of the refractive interface for the reconstruction task. Experiments with\nreal data show the superiority of our method to existing methods.', 'Recently, a lot of attention has been focused on the incorporation of 3D data\ninto face analysis and its applications. Despite providing a more accurate\nrepresentation of the face, 3D facial images are more complex to acquire than\n2D pictures. As a consequence, great effort has been invested in developing\nsystems that reconstruct 3D faces from an uncalibrated 2D image. However, the\n3D-from-2D face reconstruction problem is ill-posed, thus prior knowledge is\nneeded to restrict the solutions space. In this work, we review 3D face\nreconstruction methods proposed in the last decade, focusing on those that only\nuse 2D pictures captured under uncontrolled conditions. We present a\nclassification of the proposed methods based on the technique used to add prior\nknowledge, considering three main strategies, namely, statistical model\nfitting, photometry, and deep learning, and reviewing each of them separately.\nIn addition, given the relevance of statistical 3D facial models as prior\nknowledge, we explain the construction procedure and provide a list of the most\npopular publicly available 3D facial models. After the exhaustive study of\n3D-from-2D face reconstruction approaches, we observe that the deep learning\nstrategy is rapidly growing since the last few years, becoming the standard\nchoice in replacement of the widespread statistical model fitting. Unlike the\nother two strategies, photometry-based methods have decreased in number due to\nthe need for strong underlying assumptions that limit the quality of their\nreconstructions compared to statistical model fitting and deep learning\nmethods. The review also identifies current challenges and suggests avenues for\nfuture research.', ""Three-dimensional reconstruction of objects from shading information is a\nchallenging task in computer vision. As most of the approaches facing the\nPhotometric Stereo problem use simplified far-field assumptions, real-world\nscenarios have essentially more complex physical effects that need to be\nhandled for accurately reconstructing the 3D shape. An increasing number of\nmethods have been proposed to address the problem when point light sources are\nassumed to be nearby the target object. The proximity of the light sources\ncomplicates the modeling of the image formation as the light behaviour requires\nnon-linear parameterisation to describe its propagation and attenuation.\n  To understand the capability of the approaches dealing with this near-field\nscenario, the literature till now has used synthetically rendered photometric\nimages or minimal and very customised real-world data. In order to fill the gap\nin evaluating near-field photometric stereo methods, we introduce LUCES the\nfirst real-world 'dataset for near-fieLd point light soUrCe photomEtric Stereo'\nof 14 objects of a varying of materials. A device counting 52 LEDs has been\ndesigned to lit each object positioned 10 to 30 centimeters away from the\ncamera. Together with the raw images, in order to evaluate the 3D\nreconstructions, the dataset includes both normal and depth maps for comparing\ndifferent features of the retrieved 3D geometry. Furthermore, we evaluate the\nperformance of the latest near-field Photometric Stereo algorithms on the\nproposed dataset to assess the SOTA method with respect to actual close range\neffects and object materials."", ""The ubiquity of smartphone cameras has led to more and more documents being\ncaptured by cameras rather than scanned. Unlike flatbed scanners, photographed\ndocuments are often folded and crumpled, resulting in large local variance in\ntext structure. The problem of document rectification is fundamental to the\nOptical Character Recognition (OCR) process on documents, and its ability to\novercome geometric distortions significantly affects recognition accuracy.\nDespite the great progress in recent OCR systems, most still rely on a\npre-process that ensures the text lines are straight and axis aligned. Recent\nworks have tackled the problem of rectifying document images taken in-the-wild\nusing various supervision signals and alignment means. However, they focused on\nglobal features that can be extracted from the document's boundaries, ignoring\nvarious signals that could be obtained from the document's content.\n  We present CREASE: Content Aware Rectification using Angle Supervision, the\nfirst learned method for document rectification that relies on the document's\ncontent, the location of the words and specifically their orientation, as hints\nto assist in the rectification process. We utilize a novel pixel-wise angle\nregression approach and a curvature estimation side-task for optimizing our\nrectification model. Our method surpasses previous approaches in terms of OCR\naccuracy, geometric error and visual similarity."", 'Vehicles, search and rescue personnel, and endoscopes use flash lights to\nlocate, identify, and view objects in their surroundings. Here we show the\nfirst steps of how all these tasks can be done around corners with consumer\ncameras. Recent techniques for NLOS imaging using consumer cameras have not\nbeen able to both localize and identify the hidden object. We introduce a\nmethod that couples traditional geometric understanding and data-driven\ntechniques. To avoid the limitation of large dataset gathering, we train the\ndata-driven models on rendered samples to computationally recover the hidden\nscene on real data. The method has three independent operating modes: 1) a\nregression output to localize a hidden object in 2D, 2) an identification\noutput to identify the object type or pose, and 3) a generative network to\nreconstruct the hidden scene from a new viewpoint. The method is able to\nlocalize 12cm wide hidden objects in 2D with 1.7cm accuracy. The method also\nidentifies the hidden object class with 87.7% accuracy (compared to 33.3%\nrandom accuracy). This paper also provides an analysis on the distribution of\ninformation that encodes the occluded object in the accessible scene. We show\nthat, unlike previously thought, the area that extends beyond the corner is\nessential for accurate object localization and identification.', 'Nowadays, as cameras are rapidly adopted in our daily routine, images of\ndocuments are becoming both abundant and prevalent. Unlike natural images that\ncapture physical objects, document-images contain a significant amount of text\nwith critical semantics and complicated layouts. In this work, we devise a\ngeneric unsupervised technique to learn multimodal affinities between textual\nentities in a document-image, considering their visual style, the content of\ntheir underlying text and their geometric context within the image. We then use\nthese learned affinities to automatically cluster the textual entities in the\nimage into different semantic groups. The core of our approach is a deep\noptimization scheme dedicated for an image provided by the user that detects\nand leverages reliable pairwise connections in the multimodal representation of\nthe textual elements in order to properly learn the affinities. We show that\nour technique can operate on highly varying images spanning a wide range of\ndocuments and demonstrate its applicability for various editing operations\nmanipulating the content, appearance and geometry of the image.', 'Airborne LiDAR (Light Detection and Ranging) data is widely applied in\nbuilding reconstruction, with studies reporting success in typical buildings.\nHowever, the reconstruction of curved buildings remains an open research\nproblem. To this end, we propose a new framework for curved building\nreconstruction via assembling and deforming geometric primitives. The input\nLiDAR point cloud are first converted into contours where individual buildings\nare identified. After recognizing geometric units (primitives) from building\ncontours, we get initial models by matching basic geometric primitives to these\nprimitives. To polish assembly models, we employ a warping field for model\nrefinements. Specifically, an embedded deformation (ED) graph is constructed\nvia downsampling the initial model. Then, the point-to-model displacements are\nminimized by adjusting node parameters in the ED graph based on our objective\nfunction. The presented framework is validated on several highly curved\nbuildings collected by various LiDAR in different cities. The experimental\nresults, as well as accuracy comparison, demonstrate the advantage and\neffectiveness of our method. {The new insight attributes to an efficient\nreconstruction manner.} Moreover, we prove that the primitive-based framework\nsignificantly reduces the data storage to 10-20 percent of classical mesh\nmodels.', 'We present a novel formulation to removing reflection from polarized images\nin the wild. We first identify the misalignment issues of existing reflection\nremoval datasets where the collected reflection-free images are not perfectly\naligned with input mixed images due to glass refraction. Then we build a new\ndataset with more than 100 types of glass in which obtained transmission images\nare perfectly aligned with input mixed images. Second, capitalizing on the\nspecial relationship between reflection and polarized light, we propose a\npolarized reflection removal model with a two-stage architecture. In addition,\nwe design a novel perceptual NCC loss that can improve the performance of\nreflection removal and general image decomposition tasks. We conduct extensive\nexperiments, and results suggest that our model outperforms state-of-the-art\nmethods on reflection removal.', 'Event cameras, i.e., the Dynamic and Active-pixel Vision Sensor (DAVIS) ones,\ncapture the intensity changes in the scene and generates a stream of events in\nan asynchronous fashion. The output rate of such cameras can reach up to 10\nmillion events per second in high dynamic environments. DAVIS cameras use novel\nvision sensors that mimic human eyes. Their attractive attributes, such as high\noutput rate, High Dynamic Range (HDR), and high pixel bandwidth, make them an\nideal solution for applications that require high-frequency tracking. Moreover,\napplications that operate in challenging lighting scenarios can exploit the\nhigh HDR of event cameras, i.e., 140 dB compared to 60 dB of traditional\ncameras. In this paper, a novel asynchronous corner tracking method is proposed\nthat uses both events and intensity images captured by a DAVIS camera. The\nHarris algorithm is used to extract features, i.e., frame-corners from\nkeyframes, i.e., intensity images. Afterward, a matching algorithm is used to\nextract event-corners from the stream of events. Events are solely used to\nperform asynchronous tracking until the next keyframe is captured. Neighboring\nevents, within a window size of 5x5 pixels around the event-corner, are used to\ncalculate the velocity and direction of extracted event-corners by fitting the\n2D planar using a randomized Hough transform algorithm. Experimental evaluation\nshowed that our approach is able to update the location of the extracted\ncorners up to 100 times during the blind time of traditional cameras, i.e.,\nbetween two consecutive intensity images.', 'Understanding 3D scenes from a single image is fundamental to a wide variety\nof tasks, such as for robotics, motion planning, or augmented reality. Existing\nworks in 3D perception from a single RGB image tend to focus on geometric\nreconstruction only, or geometric reconstruction with semantic segmentation or\ninstance segmentation. Inspired by 2D panoptic segmentation, we propose to\nunify the tasks of geometric reconstruction, 3D semantic segmentation, and 3D\ninstance segmentation into the task of panoptic 3D scene reconstruction - from\na single RGB image, predicting the complete geometric reconstruction of the\nscene in the camera frustum of the image, along with semantic and instance\nsegmentations. We thus propose a new approach for holistic 3D scene\nunderstanding from a single RGB image which learns to lift and propagate 2D\nfeatures from an input image to a 3D volumetric scene representation. We\ndemonstrate that this holistic view of joint scene reconstruction, semantic,\nand instance segmentation is beneficial over treating the tasks independently,\nthus outperforming alternative approaches.', 'This paper primarily focuses on figuring out the best array of cameras, or\nvisual sensors, so that such a placement enables the maximum utilization of\nthese visual sensors. Maximizing the utilization of these cameras can convert\nto another problem that is simpler for the formulation, that is, maximizing the\ntotal coverage with these cameras. To solve the problem, the coverage problem\nis first defined subject to the capabilities and limits of cameras. Then, poses\nof cameras are analyzed for the best arrangement.', 'We propose a simple and efficient method for exploiting synthetic images when\ntraining a Deep Network to predict a 3D pose from an image. The ability of\nusing synthetic images for training a Deep Network is extremely valuable as it\nis easy to create a virtually infinite training set made of such images, while\ncapturing and annotating real images can be very cumbersome. However, synthetic\nimages do not resemble real images exactly, and using them for training can\nresult in suboptimal performance. It was recently shown that for exemplar-based\napproaches, it is possible to learn a mapping from the exemplar representations\nof real images to the exemplar representations of synthetic images. In this\npaper, we show that this approach is more general, and that a network can also\nbe applied after the mapping to infer a 3D pose: At run time, given a real\nimage of the target object, we first compute the features for the image, map\nthem to the feature space of synthetic images, and finally use the resulting\nfeatures as input to another network which predicts the 3D pose. Since this\nnetwork can be trained very effectively by using synthetic images, it performs\nvery well in practice, and inference is faster and more accurate than with an\nexemplar-based approach. We demonstrate our approach on the LINEMOD dataset for\n3D object pose estimation from color images, and the NYU dataset for 3D hand\npose estimation from depth maps. We show that it allows us to outperform the\nstate-of-the-art on both datasets.', 'We present an unsupervised learning framework for decomposing images into\nlayers of automatically discovered object models. Contrary to recent approaches\nthat model image layers with autoencoder networks, we represent them as\nexplicit transformations of a small set of prototypical images. Our model has\nthree main components: (i) a set of object prototypes in the form of learnable\nimages with a transparency channel, which we refer to as sprites; (ii)\ndifferentiable parametric functions predicting occlusions and transformation\nparameters necessary to instantiate the sprites in a given image; (iii) a\nlayered image formation model with occlusion for compositing these instances\ninto complete images including background. By jointly learning the sprites and\nocclusion/transformation predictors to reconstruct images, our approach not\nonly yields accurate layered image decompositions, but also identifies object\ncategories and instance parameters. We first validate our approach by providing\nresults on par with the state of the art on standard multi-object synthetic\nbenchmarks (Tetrominoes, Multi-dSprites, CLEVR6). We then demonstrate the\napplicability of our model to real images in tasks that include clustering\n(SVHN, GTSRB), cosegmentation (Weizmann Horse) and object discovery from\nunfiltered social network images. To the best of our knowledge, our approach is\nthe first layered image decomposition algorithm that learns an explicit and\nshared concept of object type, and is robust enough to be applied to real\nimages.', 'Event cameras are bio-inspired sensors that differ from conventional frame\ncameras: Instead of capturing images at a fixed rate, they asynchronously\nmeasure per-pixel brightness changes, and output a stream of events that encode\nthe time, location and sign of the brightness changes. Event cameras offer\nattractive properties compared to traditional cameras: high temporal resolution\n(in the order of microseconds), very high dynamic range (140 dB vs. 60 dB), low\npower consumption, and high pixel bandwidth (on the order of kHz) resulting in\nreduced motion blur. Hence, event cameras have a large potential for robotics\nand computer vision in challenging scenarios for traditional cameras, such as\nlow-latency, high speed, and high dynamic range. However, novel methods are\nrequired to process the unconventional output of these sensors in order to\nunlock their potential. This paper provides a comprehensive overview of the\nemerging field of event-based vision, with a focus on the applications and the\nalgorithms developed to unlock the outstanding properties of event cameras. We\npresent event cameras from their working principle, the actual sensors that are\navailable and the tasks that they have been used for, from low-level vision\n(feature detection and tracking, optic flow, etc.) to high-level vision\n(reconstruction, segmentation, recognition). We also discuss the techniques\ndeveloped to process events, including learning-based techniques, as well as\nspecialized processors for these novel sensors, such as spiking neural\nnetworks. Additionally, we highlight the challenges that remain to be tackled\nand the opportunities that lie ahead in the search for a more efficient,\nbio-inspired way for machines to perceive and interact with the world.', 'Moving object detection is important in computer vision. Event-based cameras\nare bio-inspired cameras that work by mimicking the working of the human eye.\nThese cameras have multiple advantages over conventional frame-based cameras,\nlike reduced latency, HDR, reduced motion blur during high motion, low power\nconsumption, etc. In spite of these advantages, event-based cameras are\nnoise-sensitive and have low resolution. Moreover, the task of moving object\ndetection in these cameras is difficult, as event-based sensors lack useful\nvisual features like texture and color. In this paper, we investigate the\napplication of the k-means clustering technique in detecting moving objects in\nevent-based data.', 'This work presents a flexible system to reconstruct 3D models of objects\ncaptured with an RGB-D sensor. A major advantage of the method is that our\nreconstruction pipeline allows the user to acquire a full 3D model of the\nobject. This is achieved by acquiring several partial 3D models in different\nsessions that are automatically merged together to reconstruct a full model. In\naddition, the 3D models acquired by our system can be directly used by\nstate-of-the-art object instance recognition and object tracking modules,\nproviding object-perception capabilities for different applications, such as\nhuman-object interaction analysis or robot grasping. The system does not impose\nconstraints in the appearance of objects (textured, untextured) nor in the\nmodelling setup (moving camera with static object or a turn-table setup). The\nproposed reconstruction system has been used to model a large number of objects\nresulting in metrically accurate and visually appealing 3D models.', 'While a great variety of 3D cameras have been introduced in recent years,\nmost publicly available datasets for object recognition and pose estimation\nfocus on one single camera. In this work, we present a dataset of 32 scenes\nthat have been captured by 7 different 3D cameras, totaling 49,294 frames. This\nallows evaluating the sensitivity of pose estimation algorithms to the\nspecifics of the used camera and the development of more robust algorithms that\nare more independent of the camera model. Vice versa, our dataset enables\nresearchers to perform a quantitative comparison of the data from several\ndifferent cameras and depth sensing technologies and evaluate their algorithms\nbefore selecting a camera for their specific task. The scenes in our dataset\ncontain 20 different objects from the common benchmark YCB object and model set\n[1], [2]. We provide full ground truth 6DoF poses for each object, per-pixel\nsegmentation, 2D and 3D bounding boxes and a measure of the amount of occlusion\nof each object. We have also performed an initial evaluation of the cameras\nusing our dataset on a state-of-the-art object recognition and pose estimation\nsystem [3].', 'Event cameras or neuromorphic cameras mimic the human perception system as\nthey measure the per-pixel intensity change rather than the actual intensity\nlevel. In contrast to traditional cameras, such cameras capture new information\nabout the scene at MHz frequency in the form of sparse events. The high\ntemporal resolution comes at the cost of losing the familiar per-pixel\nintensity information. In this work we propose a variational model that\naccurately models the behaviour of event cameras, enabling reconstruction of\nintensity images with arbitrary frame rate in real-time. Our method is\nformulated on a per-event-basis, where we explicitly incorporate information\nabout the asynchronous nature of events via an event manifold induced by the\nrelative timestamps of events. In our experiments we verify that solving the\nvariational model on the manifold produces high-quality images without\nexplicitly estimating optical flow.', 'In this paper, we propose a novel technique to reconstruct 3D surface of an\nunderwater object using stereo images. Reconstructing the 3D surface of an\nunderwater object is really a challenging task due to degraded quality of\nunderwater images. There are various reason of quality degradation of\nunderwater images i.e., non-uniform illumination of light on the surface of\nobjects, scattering and absorption effects. Floating particles present in\nunderwater produces Gaussian noise on the captured underwater images which\ndegrades the quality of images. The degraded underwater images are preprocessed\nby applying homomorphic, wavelet denoising and anisotropic filtering\nsequentially. The uncalibrated rectification technique is applied to\npreprocessed images to rectify the left and right images. The rectified left\nand right image lies on a common plane. To find the correspondence points in a\nleft and right images, we have applied dense stereo matching technique i.e.,\ngraph cut method. Finally, we estimate the depth of images using triangulation\ntechnique. The experimental result shows that the proposed method reconstruct\n3D surface of underwater objects accurately using captured underwater stereo\nimages.', 'Non-Rigid Structure from Motion (NRSfM) refers to the problem of\nreconstructing cameras and the 3D point cloud of a non-rigid object from an\nensemble of images with 2D correspondences. Current NRSfM algorithms are\nlimited from two perspectives: (i) the number of images, and (ii) the type of\nshape variability they can handle. These difficulties stem from the inherent\nconflict between the condition of the system and the degrees of freedom needing\nto be modeled -- which has hampered its practical utility for many applications\nwithin vision. In this paper we propose a novel hierarchical sparse coding\nmodel for NRSFM which can overcome (i) and (ii) to such an extent, that NRSFM\ncan be applied to problems in vision previously thought too ill posed. Our\napproach is realized in practice as the training of an unsupervised deep neural\nnetwork (DNN) auto-encoder with a unique architecture that is able to\ndisentangle pose from 3D structure. Using modern deep learning computational\nplatforms allows us to solve NRSfM problems at an unprecedented scale and shape\ncomplexity. Our approach has no 3D supervision, relying solely on 2D point\ncorrespondences. Further, our approach is also able to handle missing/occluded\n2D points without the need for matrix completion. Extensive experiments\ndemonstrate the impressive performance of our approach where we exhibit\nsuperior precision and robustness against all available state-of-the-art works\nin some instances by an order of magnitude. We further propose a new quality\nmeasure (based on the network weights) which circumvents the need for 3D\nground-truth to ascertain the confidence we have in the reconstructability. We\nbelieve our work to be a significant advance over state-of-the-art in NRSFM.', 'With the availability of more powerful computers, iterative reconstruction\nalgorithms are the subject of an ongoing work in the design of more efficient\nreconstruction algorithms for X-ray computed tomography. In this work, we show\nhow two analytical reconstruction algorithms can be improved by correcting the\ncorresponding reconstructions using a randomized iterative reconstruction\nalgorithm. The combined analytical reconstruction followed by randomized\niterative reconstruction can also be viewed as a reconstruction algorithm\nwhich, in the experiments we have conducted, uses up to $35\\%$ less projection\nangles as compared to the analytical reconstruction algorithms and produces the\nsame results in terms of quality of reconstruction, without increasing the\nexecution time significantly.', 'Monocular 3D facial shape reconstruction from a single 2D facial image has\nbeen an active research area due to its wide applications. Inspired by the\nsuccess of deep neural networks (DNN), we propose a DNN-based approach for\nEnd-to-End 3D FAce Reconstruction (UH-E2FAR) from a single 2D image. Different\nfrom recent works that reconstruct and refine the 3D face in an iterative\nmanner using both an RGB image and an initial 3D facial shape rendering, our\nDNN model is end-to-end, and thus the complicated 3D rendering process can be\navoided. Moreover, we integrate in the DNN architecture two components, namely\na multi-task loss function and a fusion convolutional neural network (CNN) to\nimprove facial expression reconstruction. With the multi-task loss function, 3D\nface reconstruction is divided into neutral 3D facial shape reconstruction and\nexpressive 3D facial shape reconstruction. The neutral 3D facial shape is\nclass-specific. Therefore, higher layer features are useful. In comparison, the\nexpressive 3D facial shape favors lower or intermediate layer features. With\nthe fusion-CNN, features from different intermediate layers are fused and\ntransformed for predicting the 3D expressive facial shape. Through extensive\nexperiments, we demonstrate the superiority of our end-to-end framework in\nimproving the accuracy of 3D face reconstruction.', 'We present a practical approach to capturing ear-to-ear face models\ncomprising both 3D meshes and intrinsic textures (i.e. diffuse and specular\nalbedo). Our approach is a hybrid of geometric and photometric methods and\nrequires no geometric calibration. Photometric measurements made in a\nlightstage are used to estimate view dependent high resolution normal maps. We\novercome the problem of having a single photometric viewpoint by capturing in\nmultiple poses. We use uncalibrated multiview stereo to estimate a coarse base\nmesh to which the photometric views are registered. We propose a novel approach\nto robustly stitching surface normal and intrinsic texture data into a\nseamless, complete and highly detailed face model. The resulting relightable\nmodels provide photorealistic renderings in any view.', 'A basic problem in computer vision is to understand the structure of a\nreal-world scene given several images of it. Here we study several theoretical\naspects of the intra multi-view geometry of calibrated cameras when all that\nthey can reliably recognize is each other. With the proliferation of wearable\ncameras, autonomous vehicles and drones, the geometry of these multiple cameras\nis a timely and relevant problem to study.', ""We study the problem of object detection over scanned images of scientific\ndocuments. We consider images that contain objects of varying aspect ratios and\nsizes and range from coarse elements such as tables and figures to fine\nelements such as equations and section headers. We find that current object\ndetectors fail to produce properly localized region proposals over such page\nobjects. We revisit the original R-CNN model and present a method for\ngenerating fine-grained proposals over document elements. We also present a\nregion embedding model that uses the convolutional maps of a proposal's\nneighbors as context to produce an embedding for each proposal. This region\nembedding is able to capture the semantic relationships between a target region\nand its surrounding context. Our end-to-end model produces an embedding for\neach proposal, then classifies each proposal by using a multi-head attention\nmodel that attends to the most important neighbors of a proposal. To evaluate\nour model, we collect and annotate a dataset of publications from heterogeneous\njournals. We show that our model, referred to as Attentive-RCNN, yields a 17%\nmAP improvement compared to standard object detection models."", 'Event cameras sense intensity changes and have many advantages over\nconventional cameras. To take advantage of event cameras, some methods have\nbeen proposed to reconstruct intensity images from event streams. However, the\noutputs are still in low resolution (LR), noisy, and unrealistic. The\nlow-quality outputs stem broader applications of event cameras, where high\nspatial resolution (HR) is needed as well as high temporal resolution, dynamic\nrange, and no motion blur. We consider the problem of reconstructing and\nsuper-resolving intensity images from LR events, when no ground truth (GT) HR\nimages and down-sampling kernels are available. To tackle the challenges, we\npropose a novel end-to-end pipeline that reconstructs LR images from event\nstreams, enhances the image qualities and upsamples the enhanced images, called\nEventSR. For the absence of real GT images, our method is primarily\nunsupervised, deploying adversarial learning. To train EventSR, we create an\nopen dataset including both real-world and simulated scenes. The use of both\ndatasets boosts up the network performance, and the network architectures and\nvarious loss functions in each phase help improve the image qualities. The\nwhole pipeline is trained in three phases. While each phase is mainly for one\nof the three tasks, the networks in earlier phases are fine-tuned by respective\nloss functions in an end-to-end manner. Experimental results show that EventSR\nreconstructs high-quality SR images from events for both simulated and\nreal-world data.', 'Division-of-focal-plane (DoFP) polarization imaging technical recently has\nbeen applied in many fields. However, the images captured by such sensors\ncannot be used directly because they suffer from instantaneous field-of-view\nerrors and low resolution problem. This paper builds a fast DoFP demosaicing\nsystem with proposed progressive polarization demosaicing convolutional neural\nnetwork (PPDN), which is specifically designed for edge-side GPU devices like\nNavidia Jetson TX2. The proposed network consists of two parts: reconstruction\nstage and refining stage. The former recovers four polarization channels from a\nsingle DoFP image. The latter fine-tune the four channels to obtain more\naccurate polarization information. PPDN can be implemented in another version:\nPPDN-L (large), for the platforms of high computing resources. Experiments show\nthat PPDN can compete with the best existing methods with fewer parameters and\nfaster inference speed and meet the real-time demands of imaging system.', 'An innovative 3-D radar imaging technique is developed for fast and efficient\nidentification and characterization of radar backscattering components of\ncomplex objects, when the collected scattered field is made of\npolarization-diverse measurements. In this context, all the polarimetric\ninformation seems irretrievably mixed. A direct model, derived from a simple\nbut original extension of the widespread ""multiple scattering model"" leads to a\nhigh dimensional linear inverse problem. It is solved by a fast dedicated\nimaging algorithm that performs to determine at a time three huge 3-D scatterer\nmaps which correspond to HH, VV and HV polarizations at emission and reception.\nIt is applied successfully to various mock-ups and data sets collected from an\naccurate and dedicated 3D spherical experimental layout that provides\nconcentric polarization-diverse RCS measurements.', 'Despite significant progress on current state-of-the-art image generation\nmodels, synthesis of document images containing multiple and complex object\nlayouts is a challenging task. This paper presents a novel approach, called\nDocSynth, to automatically synthesize document images based on a given layout.\nIn this work, given a spatial layout (bounding boxes with object categories) as\na reference by the user, our proposed DocSynth model learns to generate a set\nof realistic document images consistent with the defined layout. Also, this\nframework has been adapted to this work as a superior baseline model for\ncreating synthetic document image datasets for augmenting real data during\ntraining for document layout analysis tasks. Different sets of learning\nobjectives have been also used to improve the model performance.\nQuantitatively, we also compare the generated results of our model with real\ndata using standard evaluation metrics. The results highlight that our model\ncan successfully generate realistic and diverse document images with multiple\nobjects. We also present a comprehensive qualitative analysis summary of the\ndifferent scopes of synthetic image generation tasks. Lastly, to our knowledge\nthis is the first work of its kind.', '3D face reconstruction is a challenging problem but also an important task in\nthe field of computer vision and graphics. Recently, many researchers put\nattention to the problem and a large number of articles have been published.\nSingle image reconstruction is one of the branches of 3D face reconstruction,\nwhich has a lot of applications in our life. This paper is a review of the\nrecent literature on 3D face reconstruction from a single image.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'We introduce a novel learning method for 3D pose estimation from color\nimages. While acquiring annotations for color images is a difficult task, our\napproach circumvents this problem by learning a mapping from paired color and\ndepth images captured with an RGB-D camera. We jointly learn the pose from\nsynthetic depth images that are easy to generate, and learn to align these\nsynthetic depth images with the real depth images. We show our approach for the\ntask of 3D hand pose estimation and 3D object pose estimation, both from color\nimages only. Our method achieves performances comparable to state-of-the-art\nmethods on popular benchmark datasets, without requiring any annotations for\nthe color images.', 'We propose a novel approach to synthesizing images that are effective for\ntraining object detectors. Starting from a small set of real images, our\nalgorithm estimates the rendering parameters required to synthesize similar\nimages given a coarse 3D model of the target object. These parameters can then\nbe reused to generate an unlimited number of training images of the object of\ninterest in arbitrary 3D poses, which can then be used to increase\nclassification performances.\n  A key insight of our approach is that the synthetically generated images\nshould be similar to real images, not in terms of image quality, but rather in\nterms of features used during the detector training. We show in the context of\ndrone, plane, and car detection that using such synthetically generated images\nyields significantly better performances than simply perturbing real images or\neven synthesizing images in such way that they look very realistic, as is often\ndone when only limited amounts of training data are available.', 'Cross view feature fusion is the key to address the occlusion problem in\nhuman pose estimation. The current fusion methods need to train a separate\nmodel for every pair of cameras making them difficult to scale. In this work,\nwe introduce MetaFuse, a pre-trained fusion model learned from a large number\nof cameras in the Panoptic dataset. The model can be efficiently adapted or\nfinetuned for a new pair of cameras using a small number of labeled images. The\nstrong adaptation power of MetaFuse is due in large part to the proposed\nfactorization of the original fusion model into two parts (1) a generic fusion\nmodel shared by all cameras, and (2) lightweight camera-dependent\ntransformations. Furthermore, the generic model is learned from many cameras by\na meta-learning style algorithm to maximize its adaptation capability to\nvarious camera poses. We observe in experiments that MetaFuse finetuned on the\npublic datasets outperforms the state-of-the-arts by a large margin which\nvalidates its value in practice.', 'The distributed representation of correlated multi-view images is an\nimportant problem that arise in vision sensor networks. This paper concentrates\non the joint reconstruction problem where the distributively compressed\ncorrelated images are jointly decoded in order to improve the reconstruction\nquality of all the compressed images. We consider a scenario where the images\ncaptured at different viewpoints are encoded independently using common coding\nsolutions (e.g., JPEG, H.264 intra) with a balanced rate distribution among\ndifferent cameras. A central decoder first estimates the underlying correlation\nmodel from the independently compressed images which will be used for the joint\nsignal recovery. The joint reconstruction is then cast as a constrained convex\noptimization problem that reconstructs total-variation (TV) smooth images that\ncomply with the estimated correlation model. At the same time, we add\nconstraints that force the reconstructed images to be consistent with their\ncompressed versions. We show by experiments that the proposed joint\nreconstruction scheme outperforms independent reconstruction in terms of image\nquality, for a given target bit rate. In addition, the decoding performance of\nour proposed algorithm compares advantageously to state-of-the-art distributed\ncoding schemes based on disparity learning and on the DISCOVER.', 'Imaging polarimetry allows more information to be extracted from a scene than\nconventional intensity or colour imaging. However, a major challenge of imaging\npolarimetry is image degradation due to noise. This paper investigates the\nmitigation of noise through denoising algorithms and compares existing\ndenoising algorithms with a new method, based on BM3D. This algorithm, PBM3D,\ngives visual quality superior to the state of the art across all images and\nnoise standard deviations tested. We show that denoising polarization images\nusing PBM3D allows the degree of polarization to be more accurately calculated\nby comparing it to spectroscopy methods.', 'This paper presents an explore-and-classify framework for structured\narchitectural reconstruction from an aerial image. Starting from a potentially\nimperfect building reconstruction by an existing algorithm, our approach 1)\nexplores the space of building models by modifying the reconstruction via\nheuristic actions; 2) learns to classify the correctness of building models\nwhile generating classification labels based on the ground-truth, and 3)\nrepeat. At test time, we iterate exploration and classification, seeking for a\nresult with the best classification score. We evaluate the approach using\ninitial reconstructions by two baselines and two state-of-the-art\nreconstruction algorithms. Qualitative and quantitative evaluations demonstrate\nthat our approach consistently improves the reconstruction quality from every\ninitial reconstruction.', 'Event cameras are novel sensors that report brightness changes in the form of\na stream of asynchronous ""events"" instead of intensity frames. They offer\nsignificant advantages with respect to conventional cameras: high temporal\nresolution, high dynamic range, and no motion blur. While the stream of events\nencodes in principle the complete visual signal, the reconstruction of an\nintensity image from a stream of events is an ill-posed problem in practice.\nExisting reconstruction approaches are based on hand-crafted priors and strong\nassumptions about the imaging process as well as the statistics of natural\nimages. In this work we propose to learn to reconstruct intensity images from\nevent streams directly from data instead of relying on any hand-crafted priors.\nWe propose a novel recurrent network to reconstruct videos from a stream of\nevents, and train it on a large amount of simulated event data. During training\nwe propose to use a perceptual loss to encourage reconstructions to follow\nnatural image statistics. We further extend our approach to synthesize color\nimages from color event streams. Our network surpasses state-of-the-art\nreconstruction methods by a large margin in terms of image quality (> 20%),\nwhile comfortably running in real-time. We show that the network is able to\nsynthesize high framerate videos (> 5,000 frames per second) of high-speed\nphenomena (e.g. a bullet hitting an object) and is able to provide high dynamic\nrange reconstructions in challenging lighting conditions. We also demonstrate\nthe effectiveness of our reconstructions as an intermediate representation for\nevent data. We show that off-the-shelf computer vision algorithms can be\napplied to our reconstructions for tasks such as object classification and\nvisual-inertial odometry and that this strategy consistently outperforms\nalgorithms that were specifically designed for event data.', 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.', 'Novel vision sensors such as thermal, hyperspectral, polarization, and event\ncameras provide information that is not available from conventional intensity\ncameras. An obstacle to using these sensors with current powerful deep neural\nnetworks is the lack of large labeled training datasets. This paper proposes a\nNetwork Grafting Algorithm (NGA), where a new front end network driven by\nunconventional visual inputs replaces the front end network of a pretrained\ndeep network that processes intensity frames. The self-supervised training uses\nonly synchronously-recorded intensity frames and novel sensor data to maximize\nfeature similarity between the pretrained network and the grafted network. We\nshow that the enhanced grafted network reaches competitive average precision\n(AP50) scores to the pretrained network on an object detection task using\nthermal and event camera datasets, with no increase in inference costs.\nParticularly, the grafted network driven by thermal frames showed a relative\nimprovement of 49.11% over the use of intensity frames. The grafted front end\nhas only 5--8% of the total parameters and can be trained in a few hours on a\nsingle GPU equivalent to 5% of the time that would be needed to train the\nentire object detector from labeled data. NGA allows new vision sensors to\ncapitalize on previously pretrained powerful deep models, saving on training\ncost and widening a range of applications for novel sensors.', 'Cosmic polarization rotation, which may result from parity-violating new\nphysics or the presence of primordial magnetic fields, converts $E$-mode\npolarization of the cosmic microwave background (CMB) into $B$-mode\npolarization. Anisotropic cosmic polarization rotation leads to statistical\nanisotropy in CMB polarization and can be reconstructed with quadratic\nestimator techniques similar to those designed for gravitational lensing of the\nCMB. At the sensitivity of upcoming CMB surveys, lensing-induced $B$-mode\npolarization will act as a limiting factor in the search for anisotropic cosmic\npolarization rotation, meaning that an analysis which incorporates some form of\ndelensing will be required to improve constraints on the effect with future\nsurveys. In this paper we extend the ResUNet-CMB convolutional neural network\nto reconstruct anisotropic cosmic polarization rotation in the presence of\ngravitational lensing and patchy reionization, and we show that the network\nsimultaneously reconstructs all three effects with variance that is lower than\nthat from the standard quadratic estimator nearly matching the performance of\nan iterative reconstruction method.', 'Polarization of the mammalian embryo at the right developmental time is\ncritical for its development to term and would be valuable in assessing the\npotential of human embryos. However, tracking polarization requires invasive\nfluorescence staining, impermissible in the in vitro fertilization clinic.\nHere, we report the use of artificial intelligence to detect polarization from\nunstained time-lapse movies of mouse embryos. We assembled a dataset of\nbright-field movie frames from 8-cell-stage embryos, side-by-side with\ncorresponding images of fluorescent markers of cell polarization. We then used\nan ensemble learning model to detect whether any bright-field frame showed an\nembryo before or after onset of polarization. Our resulting model has an\naccuracy of 85% for detecting polarization, significantly outperforming human\nvolunteers trained on the same data (61% accuracy). We discovered that our\nself-learning model focuses upon the angle between cells as one known cue for\ncompaction, which precedes polarization, but it outperforms the use of this cue\nalone. By compressing three-dimensional time-lapsed image data into\ntwo-dimensions, we are able to reduce data to an easily manageable size for\ndeep learning processing. In conclusion, we describe a method for detecting a\nkey developmental feature of embryo development that avoids clinically\nimpermissible fluorescence staining.', 'For applications such as airport border control, biometric technologies that\ncan process many capture subjects quickly, efficiently, with weak supervision,\nand with minimal discomfort are desirable. Facial recognition is particularly\nappealing because it is minimally invasive yet offers relatively good\nrecognition performance. Unfortunately, the combination of weak supervision and\nminimal invasiveness makes even highly accurate facial recognition systems\nsusceptible to spoofing via presentation attacks. Thus, there is great demand\nfor an effective and low cost system capable of rejecting such attacks.To this\nend we introduce PARAPH -- a novel hardware extension that exploits different\nmeasurements of light polarization to yield an image space in which\npresentation media are readily discernible from Bona Fide facial\ncharacteristics. The PARAPH system is inexpensive with an added cost of less\nthan 10 US dollars. The system makes two polarization measurements in rapid\nsuccession, allowing them to be approximately pixel-aligned, with a frame rate\nlimited by the camera, not the system. There are no moving parts above the\nmolecular level, due to the efficient use of twisted nematic liquid crystals.\nWe present evaluation images using three presentation attack media next to an\nactual face -- high quality photos on glossy and matte paper and a video of the\nface on an LCD. In each case, the actual face in the image generated by PARAPH\nis structurally discernible from the presentations, which appear either as\nnoise (print attacks) or saturated images (replay attacks).']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am doing theoretical experiments on whether deep learning models can understand quantitative logic reasoning. To continue my experiment, I am looking for a deep learning model that can perform deductive and quantitative inference tasks. Deductive tasks include several premises and a conclusion, and the model needs to identify whether the conclusion is reasonable based on the premise; and quantitative inference problem needs the model to interpret real-life problems into math problems and then solve it. I hope the single model can handle both deductive, quantitative inference tasks and a task that is a combination of the 2 subtasks. Additionally, I need another system that is able to efficiently take answers from humans, extract the key part and compare human answers to the answers generated by the model.","['Deep learning is very effective at jointly learning feature representations\nand classification models, especially when dealing with high dimensional input\npatterns. Probabilistic logic reasoning, on the other hand, is capable to take\nconsistent and robust decisions in complex environments. The integration of\ndeep learning and logic reasoning is still an open-research problem and it is\nconsidered to be the key for the development of real intelligent agents. This\npaper presents Deep Logic Models, which are deep graphical models integrating\ndeep learning and logic reasoning both for learning and inference. Deep Logic\nModels create an end-to-end differentiable architecture, where deep learners\nare embedded into a network implementing a continuous relaxation of the logic\nknowledge. The learning process allows to jointly learn the weights of the deep\nlearners and the meta-parameters controlling the high-level reasoning. The\nexperimental results show that the proposed methodology overtakes the\nlimitations of the other approaches that have been proposed to bridge deep\nlearning and reasoning.', 'Recent years have witnessed the great success of deep neural networks in many\nresearch areas. The fundamental idea behind the design of most neural networks\nis to learn similarity patterns from data for prediction and inference, which\nlacks the ability of logical reasoning. However, the concrete ability of\nlogical reasoning is critical to many theoretical and practical problems. In\nthis paper, we propose Neural Logic Network (NLN), which is a dynamic neural\narchitecture that builds the computational graph according to input logical\nexpressions. It learns basic logical operations as neural modules, and conducts\npropositional logical reasoning through the network for inference. Experiments\non simulated data show that NLN achieves significant performance on solving\nlogical equations. Further experiments on real-world data show that NLN\nsignificantly outperforms state-of-the-art models on collaborative filtering\nand personalized recommendation tasks.', 'The ability to conduct logical reasoning is a fundamental aspect of\nintelligent human behavior, and thus an important problem along the way to\nhuman-level artificial intelligence. Traditionally, logic-based symbolic\nmethods from the field of knowledge representation and reasoning have been used\nto equip agents with capabilities that resemble human logical reasoning\nqualities. More recently, however, there has been an increasing interest in\nusing machine learning rather than logic-based symbolic formalisms to tackle\nthese tasks. In this paper, we employ state-of-the-art methods for training\ndeep neural networks to devise a novel model that is able to learn how to\neffectively perform logical reasoning in the form of basic ontology reasoning.\nThis is an important and at the same time very natural logical reasoning task,\nwhich is why the presented approach is applicable to a plethora of important\nreal-world problems. We present the outcomes of several experiments, which show\nthat our model is able to learn to perform highly accurate ontology reasoning\non very large, diverse, and challenging benchmarks. Furthermore, it turned out\nthat the suggested approach suffers much less from different obstacles that\nprohibit logic-based symbolic reasoning, and, at the same time, is surprisingly\nplausible from a biological point of view.', 'Analytical reasoning is an essential and challenging task that requires a\nsystem to analyze a scenario involving a set of particular circumstances and\nperform reasoning over it to make conclusions. In this paper, we study the\nchallenge of analytical reasoning of text and introduce a new dataset\nconsisting of questions from the Law School Admission Test from 1991 to 2016.\nWe analyze what knowledge understanding and reasoning abilities are required to\ndo well on this task. Furthermore, to address this reasoning challenge, we\ndesign two different baselines: (1) a Transformer-based method which leverages\nthe state-of-the-art pre-trained language models and (2) Analytical Reasoning\nMachine (ARM), a logical-level reasoning framework extracting symbolic\nknowledge (e.g, participants, facts, logical functions) to deduce legitimate\nsolutions. In our experiments, we find that the Transformer-based models\nstruggle to solve this task as their performance is close to random guess and\nARM achieves better performance by leveraging symbolic knowledge and\ninterpretable reasoning steps. Results show that both methods still lag far\nbehind human performance, which leave further space for future research.', 'An important aspect of artificial intelligence (AI) is the ability to reason\nin a step-by-step ""algorithmic"" manner that can be inspected and verified for\nits correctness. This is especially important in the domain of question\nanswering (QA). We argue that the challenge of algorithmic reasoning in QA can\nbe effectively tackled with a ""systems"" approach to AI which features a hybrid\nuse of symbolic and sub-symbolic methods including deep neural networks.\nAdditionally, we argue that while neural network models with end-to-end\ntraining pipelines perform well in narrow applications such as image\nclassification and language modelling, they cannot, on their own, successfully\nperform algorithmic reasoning, especially if the task spans multiple domains.\nWe discuss a few notable exceptions and point out how they are still limited\nwhen the QA problem is widened to include other intelligence-requiring tasks.\nHowever, deep learning, and machine learning in general, do play important\nroles as components in the reasoning process. We propose an approach to\nalgorithm reasoning for QA, Deep Algorithmic Question Answering (DAQA), based\non three desirable properties: interpretability, generalizability, and\nrobustness which such an AI system should possess, and conclude that they are\nbest achieved with a combination of hybrid and compositional AI.']","['Higher-level cognition includes logical reasoning and the ability of question\nanswering with common sense. The RatioLog project addresses the problem of\nrational reasoning in deep question answering by methods from automated\ndeduction and cognitive computing. In a first phase, we combine techniques from\ninformation retrieval and machine learning to find appropriate answer\ncandidates from the huge amount of text in the German version of the free\nencyclopedia ""Wikipedia"". In a second phase, an automated theorem prover tries\nto verify the answer candidates on the basis of their logical representations.\nIn a third phase - because the knowledge may be incomplete and inconsistent -,\nwe consider extensions of logical reasoning to improve the results. In this\ncontext, we work toward the application of techniques from human reasoning: We\nemploy defeasible reasoning to compare the answers w.r.t. specificity, deontic\nlogic, normative reasoning, and model construction. Moreover, we use integrated\ncase-based reasoning and machine learning techniques on the basis of the\nsemantic structure of the questions and answer candidates to learn giving the\nright answers.', 'Recently, low-dimensional vector space representations of knowledge graphs\n(KGs) have been applied to find answers to conjunctive queries (CQs) over\nincomplete KGs. However, the current methods only focus on inductive reasoning,\ni.e. answering CQs by predicting facts based on patterns learned from the data,\nand lack the ability of deductive reasoning by applying external domain\nknowledge. Such (expert or commonsense) domain knowledge is an invaluable\nresource which can be used to advance machine intelligence. To address this\nshortcoming, we introduce a neural-symbolic method for ontology-mediated CQ\nanswering over incomplete KGs that operates in the embedding space. More\nspecifically, we propose various data augmentation strategies to generate\ntraining queries using query-rewriting based methods and then exploit a novel\nloss function for training the model. The experimental results demonstrate the\neffectiveness of our training strategies and the new loss function, i.e., our\nmethod significantly outperforms the baseline in the settings that require both\ninductive and deductive reasoning.', ""People readily generalize knowledge to novel domains and stimuli. We present\na theory, instantiated in a computational model, based on the idea that\ncross-domain generalization in humans is a case of analogical inference over\nstructured (i.e., symbolic) relational representations. The model is an\nextension of the LISA and DORA models of relational inference and learning. The\nresulting model learns both the content and format (i.e., structure) of\nrelational representations from non-relational inputs without supervision, when\naugmented with the capacity for reinforcement learning, leverages these\nrepresentations to learn individual domains, and then generalizes to new\ndomains on the first exposure (i.e., zero-shot learning) via analogical\ninference. We demonstrate the capacity of the model to learn structured\nrelational representations from a variety of simple visual stimuli, and to\nperform cross-domain generalization between video games (Breakout and Pong) and\nbetween several psychological tasks. We demonstrate that the model's trajectory\nclosely mirrors the trajectory of children as they learn about relations,\naccounting for phenomena from the literature on the development of children's\nreasoning and analogy making. The model's ability to generalize between domains\ndemonstrates the flexibility afforded by representing domains in terms of their\nunderlying relational structure, rather than simply in terms of the statistical\nrelations between their inputs and outputs."", 'Recent years have witnessed the success of deep neural networks in many\nresearch areas. The fundamental idea behind the design of most neural networks\nis to learn similarity patterns from data for prediction and inference, which\nlacks the ability of cognitive reasoning. However, the concrete ability of\nreasoning is critical to many theoretical and practical problems. On the other\nhand, traditional symbolic reasoning methods do well in making logical\ninference, but they are mostly hard rule-based reasoning, which limits their\ngeneralization ability to different tasks since difference tasks may require\ndifferent rules. Both reasoning and generalization ability are important for\nprediction tasks such as recommender systems, where reasoning provides strong\nconnection between user history and target items for accurate prediction, and\ngeneralization helps the model to draw a robust user portrait over noisy\ninputs.\n  In this paper, we propose Logic-Integrated Neural Network (LINN) to integrate\nthe power of deep learning and logic reasoning. LINN is a dynamic neural\narchitecture that builds the computational graph according to input logical\nexpressions. It learns basic logical operations such as AND, OR, NOT as neural\nmodules, and conducts propositional logical reasoning through the network for\ninference. Experiments on theoretical task show that LINN achieves significant\nperformance on solving logical equations and variables. Furthermore, we test\nour approach on the practical task of recommendation by formulating the task\ninto a logical inference problem. Experiments show that LINN significantly\noutperforms state-of-the-art recommendation models in Top-K recommendation,\nwhich verifies the potential of LINN in practice.', ""Whether neural networks can learn abstract reasoning or whether they merely\nrely on superficial statistics is a topic of recent debate. Here, we propose a\ndataset and challenge designed to probe abstract reasoning, inspired by a\nwell-known human IQ test. To succeed at this challenge, models must cope with\nvarious generalisation `regimes' in which the training and test data differ in\nclearly-defined ways. We show that popular models such as ResNets perform\npoorly, even when the training and test sets differ only minimally, and we\npresent a novel architecture, with a structure designed to encourage reasoning,\nthat does significantly better. When we vary the way in which the test\nquestions and training data differ, we find that our model is notably\nproficient at certain forms of generalisation, but notably weak at others. We\nfurther show that the model's ability to generalise improves markedly if it is\ntrained to predict symbolic explanations for its answers. Altogether, we\nintroduce and explore ways to both measure and induce stronger abstract\nreasoning in neural networks. Our freely-available dataset should motivate\nfurther progress in this direction."", 'we introduce deductive association networks(DANs), a network that performs\ndeductive reasoning. To have high-dimensional thinking, combining various\naxioms and putting the results back into another axiom is necessary to produce\nnew relationships and results. For example, it would be given two propositions:\n""Socrates is a man."" and ""All men are mortals."" and two propositions could be\nused to infer the new proposition, ""Therefore Socrates is mortal."". To\nevaluate, we used MNIST Dataset, a handwritten numerical image dataset, to\napply it to the group theory and show the results of performing deductive\nlearning.', 'We consider three modern roles for logic in artificial intelligence, which\nare based on the theory of tractable Boolean circuits: (1) logic as a basis for\ncomputation, (2) logic for learning from a combination of data and knowledge,\nand (3) logic for reasoning about the behavior of machine learning systems.', 'Deep learning has been shown to achieve impressive results in several tasks\nwhere a large amount of training data is available. However, deep learning\nsolely focuses on the accuracy of the predictions, neglecting the reasoning\nprocess leading to a decision, which is a major issue in life-critical\napplications. Probabilistic logic reasoning allows to exploit both statistical\nregularities and specific domain expertise to perform reasoning under\nuncertainty, but its scalability and brittle integration with the layers\nprocessing the sensory data have greatly limited its applications. For these\nreasons, combining deep architectures and probabilistic logic reasoning is a\nfundamental goal towards the development of intelligent agents operating in\ncomplex environments. This paper presents Relational Neural Machines, a novel\nframework allowing to jointly train the parameters of the learners and of a\nFirst--Order Logic based reasoner. A Relational Neural Machine is able to\nrecover both classical learning from supervised data in case of pure\nsub-symbolic learning, and Markov Logic Networks in case of pure symbolic\nreasoning, while allowing to jointly train and perform inference in hybrid\nlearning tasks. Proper algorithmic solutions are devised to make learning and\ninference tractable in large-scale problems. The experiments show promising\nresults in different relational tasks.', 'Numerical reasoning is often important to accurately understand the world.\nRecently, several format-specific datasets have been proposed, such as\nnumerical reasoning in the settings of Natural Language Inference (NLI),\nReading Comprehension (RC), and Question Answering (QA). Several\nformat-specific models and architectures in response to those datasets have\nalso been proposed. However, there exists a strong need for a benchmark which\ncan evaluate the abilities of models, in performing question format independent\nnumerical reasoning, as (i) the numerical reasoning capabilities we want to\nteach are not controlled by question formats, (ii) for numerical reasoning\ntechnology to have the best possible application, it must be able to process\nlanguage and reason in a way that is not exclusive to a single format, task,\ndataset or domain. In pursuit of this goal, we introduce NUMBERGAME, a\nmultifaceted benchmark to evaluate model performance across numerical reasoning\ntasks of eight diverse formats. We add four existing question types in our\ncompilation. Two of the new types we add are about questions that require\nexternal numerical knowledge, commonsense knowledge and domain knowledge. For\nbuilding a more practical numerical reasoning system, NUMBERGAME demands four\ncapabilities beyond numerical reasoning: (i) detecting question format directly\nfrom data (ii) finding intermediate common format to which every format can be\nconverted (iii) incorporating commonsense knowledge (iv) handling data\nimbalance across formats. We build several baselines, including a new model\nbased on knowledge hunting using a cheatsheet. However, all baselines perform\npoorly in contrast to the human baselines, indicating the hardness of our\nbenchmark. Our work takes forward the recent progress in generic system\ndevelopment, demonstrating the scope of these under-explored tasks.', 'Combining machine learning with logic-based expert systems in order to get\nthe best of both worlds are becoming increasingly popular. However, to what\nextent machine learning can already learn to reason over rule-based knowledge\nis still an open problem. In this paper, we explore how symbolic logic, defined\nas logic programs at a character level, is learned to be represented in a\nhigh-dimensional vector space using RNN-based iterative neural networks to\nperform reasoning. We create a new dataset that defines 12 classes of logic\nprograms exemplifying increased level of complexity of logical reasoning and\ntrain the networks in an end-to-end fashion to learn whether a logic program\nentails a given query. We analyse how learning the inference algorithm gives\nrise to representations of atoms, literals and rules within logic programs and\nevaluate against increasing lengths of predicate and constant symbols as well\nas increasing steps of multi-hop reasoning.', 'Recursive neural network models and their accompanying vector representations\nfor words have seen success in an array of increasingly semantically\nsophisticated tasks, but almost nothing is known about their ability to\naccurately capture the aspects of linguistic meaning that are necessary for\ninterpretation or reasoning. To evaluate this, I train a recursive model on a\nnew corpus of constructed examples of logical reasoning in short sentences,\nlike the inference of ""some animal walks"" from ""some dog walks"" or ""some cat\nwalks,"" given that dogs and cats are animals. This model learns representations\nthat generalize well to new types of reasoning pattern in all but a few cases,\na result which is promising for the ability of learned representation models to\ncapture logical reasoning.', 'Deep learning models are widely used for various industrial and scientific\napplications. Even though these models have achieved considerable success in\nrecent years, there exists a lack of understanding of the rationale behind\ndecisions made by such systems in the machine learning community. This problem\nof interpretability is further aggravated by the increasing complexity of such\nmodels. This paper utilizes concepts from machine learning, quantum computation\nand quantum field theory to demonstrate how a many valued quantum logic system\nnaturally arises in a specific class of generative deep learning models called\nConvolutional Deep Belief Networks. It provides a robust theoretical framework\nfor constructing deep learning models equipped with the interpretability of\nmany valued quantum logic systems without compromising their computing\nefficiency.', 'Language-guided robots must be able to both ask humans questions and\nunderstand answers. Much existing work focuses only on the latter. In this\npaper, we go beyond instruction following and introduce a two-agent task where\none agent navigates and asks questions that a second, guiding agent answers.\nInspired by theory of mind, we propose the Recursive Mental Model (RMM). The\nnavigating agent models the guiding agent to simulate answers given candidate\ngenerated questions. The guiding agent in turn models the navigating agent to\nsimulate navigation steps it would take to generate answers. We use the\nprogress agents make towards the goal as a reinforcement learning reward signal\nto directly inform not only navigation actions, but also both question and\nanswer generation. We demonstrate that RMM enables better generalization to\nnovel environments. Interlocutor modelling may be a way forward for human-agent\ndialogue where robots need to both ask and answer questions.', 'Large-scale knowledge graphs provide structured representations of human\nknowledge. However, as it is impossible to collect all knowledge, knowledge\ngraphs are usually incomplete. Reasoning based on existing facts paves a way to\ndiscover missing facts. In this paper, we study the problem of learning logical\nrules for reasoning on knowledge graphs for completing missing factual\ntriplets. Learning logical rules equips a model with strong interpretability as\nwell as the ability to generalize to similar tasks. We propose a model able to\nfully use training data which also considers multi-target scenarios. In\naddition, considering the deficiency in evaluating the performance of models\nand the quality of mined rules, we further propose two novel indicators to help\nwith the problem. Experimental results empirically demonstrate that our model\noutperforms state-of-the-art methods on five benchmark datasets. The results\nalso prove the effectiveness of the indicators.', 'Semantic Web knowledge representation standards, and in particular RDF and\nOWL, often come endowed with a formal semantics which is considered to be of\nfundamental importance for the field. Reasoning, i.e., the drawing of logical\ninferences from knowledge expressed in such standards, is traditionally based\non logical deductive methods and algorithms which can be proven to be sound and\ncomplete and terminating, i.e. correct in a very strong sense. For various\nreasons, though, in particular, the scalability issues arising from the\never-increasing amounts of Semantic Web data available and the inability of\ndeductive algorithms to deal with noise in the data, it has been argued that\nalternative means of reasoning should be investigated which bear high promise\nfor high scalability and better robustness. From this perspective, deductive\nalgorithms can be considered the gold standard regarding correctness against\nwhich alternative methods need to be tested. In this paper, we show that it is\npossible to train a Deep Learning system on RDF knowledge graphs, such that it\nis able to perform reasoning over new RDF knowledge graphs, with high precision\nand recall compared to the deductive gold standard.', 'Synthesizing user-intended programs from a small number of input-output\nexamples is a challenging problem with several important applications like\nspreadsheet manipulation, data wrangling and code refactoring. Existing\nsynthesis systems either completely rely on deductive logic techniques that are\nextensively hand-engineered or on purely statistical models that need massive\namounts of data, and in general fail to provide real-time synthesis on\nchallenging benchmarks. In this work, we propose Neural Guided Deductive Search\n(NGDS), a hybrid synthesis technique that combines the best of both symbolic\nlogic techniques and statistical models. Thus, it produces programs that\nsatisfy the provided specifications by construction and generalize well on\nunseen examples, similar to data-driven systems. Our technique effectively\nutilizes the deductive search framework to reduce the learning problem of the\nneural component to a simple supervised learning setup. Further, this allows us\nto both train on sparingly available real-world data and still leverage\npowerful recurrent neural network encoders. We demonstrate the effectiveness of\nour method by evaluating on real-world customer scenarios by synthesizing\naccurate programs with up to 12x speed-up compared to state-of-the-art systems.', 'While perception tasks such as visual object recognition and text\nunderstanding play an important role in human intelligence, the subsequent\ntasks that involve inference, reasoning and planning require an even higher\nlevel of intelligence. The past few years have seen major advances in many\nperception tasks using deep learning models. For higher-level inference,\nhowever, probabilistic graphical models with their Bayesian nature are still\nmore powerful and flexible. To achieve integrated intelligence that involves\nboth perception and inference, it is naturally desirable to tightly integrate\ndeep learning and Bayesian models within a principled probabilistic framework,\nwhich we call Bayesian deep learning. In this unified framework, the perception\nof text or images using deep learning can boost the performance of higher-level\ninference and in return, the feedback from the inference process is able to\nenhance the perception of text or images. This paper proposes a general\nframework for Bayesian deep learning and reviews its recent applications on\nrecommender systems, topic models, and control. In this paper, we also discuss\nthe relationship and differences between Bayesian deep learning and other\nrelated topics like Bayesian treatment of neural networks.', 'Algorithms of inference in a computer system oriented to input and semantic\nprocessing of text information are presented. Such inference is necessary for\nlogical questions when the direct comparison of objects from a question and\ndatabase can not give a result. The following classes of problems are\nconsidered: a check of hypotheses for persons and non-typical actions, the\ndetermination of persons and circumstances for non-typical actions, planning\nactions, the determination of event cause and state of persons. To form an\nanswer both deduction and plausible reasoning are used. As a knowledge domain\nunder consideration is social behavior of persons, plausible reasoning is based\non laws of social psychology. Proposed algorithms of inference and plausible\nreasoning can be realized in computer systems closely connected with text\nprocessing (criminology, operation of business, medicine, document systems).', 'Quantitative reasoning is a higher-order reasoning skill that any intelligent\nnatural language understanding system can reasonably be expected to handle. We\npresent EQUATE (Evaluating Quantitative Understanding Aptitude in Textual\nEntailment), a new framework for quantitative reasoning in textual entailment.\nWe benchmark the performance of 9 published NLI models on EQUATE, and find that\non average, state-of-the-art methods do not achieve an absolute improvement\nover a majority-class baseline, suggesting that they do not implicitly learn to\nreason with quantities. We establish a new baseline Q-REAS that manipulates\nquantities symbolically. In comparison to the best performing NLI model, it\nachieves success on numerical reasoning tests (+24.2%), but has limited verbal\nreasoning capabilities (-8.1%). We hope our evaluation framework will support\nthe development of models of quantitative reasoning in language understanding.', ""In complex inferential tasks like question answering, machine learning models\nmust confront two challenges: the need to implement a compositional reasoning\nprocess, and, in many applications, the need for this reasoning process to be\ninterpretable to assist users in both development and prediction. Existing\nmodels designed to produce interpretable traces of their decision-making\nprocess typically require these traces to be supervised at training time. In\nthis paper, we present a novel neural modular approach that performs\ncompositional reasoning by automatically inducing a desired sub-task\ndecomposition without relying on strong supervision. Our model allows linking\ndifferent reasoning tasks though shared modules that handle common routines\nacross tasks. Experiments show that the model is more interpretable to human\nevaluators compared to other state-of-the-art models: users can better\nunderstand the model's underlying reasoning procedure and predict when it will\nsucceed or fail based on observing its intermediate outputs."", 'Even with impressive advances in automated formal methods, certain problems\nin system verification and synthesis remain challenging. Examples include the\nverification of quantitative properties of software involving constraints on\ntiming and energy consumption, and the automatic synthesis of systems from\nspecifications. The major challenges include environment modeling,\nincompleteness in specifications, and the complexity of underlying decision\nproblems.\n  This position paper proposes sciduction, an approach to tackle these\nchallenges by integrating inductive inference, deductive reasoning, and\nstructure hypotheses. Deductive reasoning, which leads from general rules or\nconcepts to conclusions about specific problem instances, includes techniques\nsuch as logical inference and constraint solving. Inductive inference, which\ngeneralizes from specific instances to yield a concept, includes algorithmic\nlearning from examples. Structure hypotheses are used to define the class of\nartifacts, such as invariants or program fragments, generated during\nverification or synthesis. Sciduction constrains inductive and deductive\nreasoning using structure hypotheses, and actively combines inductive and\ndeductive reasoning: for instance, deductive techniques generate examples for\nlearning, and inductive reasoning is used to guide the deductive engines.\n  We illustrate this approach with three applications: (i) timing analysis of\nsoftware; (ii) synthesis of loop-free programs, and (iii) controller synthesis\nfor hybrid systems. Some future applications are also discussed.', 'Exact structured inference with neural network scoring functions is\ncomputationally challenging but several methods have been proposed for\napproximating inference. One approach is to perform gradient descent with\nrespect to the output structure directly (Belanger and McCallum, 2016). Another\napproach, proposed recently, is to train a neural network (an ""inference\nnetwork"") to perform inference (Tu and Gimpel, 2018). In this paper, we compare\nthese two families of inference methods on three sequence labeling datasets. We\nchoose sequence labeling because it permits us to use exact inference as a\nbenchmark in terms of speed, accuracy, and search error. Across datasets, we\ndemonstrate that inference networks achieve a better speed/accuracy/search\nerror trade-off than gradient descent, while also being faster than exact\ninference at similar accuracy levels. We find further benefit by combining\ninference networks and gradient descent, using the former to provide a warm\nstart for the latter.', 'Many machine learning applications require the ability to learn from and\nreason about noisy multi-relational data. To address this, several effective\nrepresentations have been developed that provide both a language for expressing\nthe structural regularities of a domain, and principled support for\nprobabilistic inference. In addition to these two aspects, however, many\napplications also involve a third aspect-the need to reason about\nsimilarities-which has not been directly supported in existing frameworks. This\npaper introduces probabilistic similarity logic (PSL), a general-purpose\nframework for joint reasoning about similarity in relational domains that\nincorporates probabilistic reasoning about similarities and relational\nstructure in a principled way. PSL can integrate any existing domain-specific\nsimilarity measures and also supports reasoning about similarities between sets\nof entities. We provide efficient inference and learning techniques for PSL and\ndemonstrate its effectiveness both in common relational tasks and in settings\nthat require reasoning about similarity.', 'Automated reasoning about uncertain knowledge has many applications. One\ndifficulty when developing such systems is the lack of a completely\nsatisfactory integration of logic and probability. We address this problem\ndirectly. Expressive languages like higher-order logic are ideally suited for\nrepresenting and reasoning about structured knowledge. Uncertain knowledge can\nbe modeled by using graded probabilities rather than binary truth-values. The\nmain technical problem studied in this paper is the following: Given a set of\nsentences, each having some probability of being true, what probability should\nbe ascribed to other (query) sentences? A natural wish-list, among others, is\nthat the probability distribution (i) is consistent with the knowledge base,\n(ii) allows for a consistent inference procedure and in particular (iii)\nreduces to deductive logic in the limit of probabilities being 0 and 1, (iv)\nallows (Bayesian) inductive reasoning and (v) learning in the limit and in\nparticular (vi) allows confirmation of universally quantified\nhypotheses/sentences. We translate this wish-list into technical requirements\nfor a prior probability and show that probabilities satisfying all our criteria\nexist. We also give explicit constructions and several general\ncharacterizations of probabilities that satisfy some or all of the criteria and\nvarious (counter) examples. We also derive necessary and sufficient conditions\nfor extending beliefs about finitely many sentences to suitable probabilities\nover all sentences, and in particular least dogmatic or least biased ones. We\nconclude with a brief outlook on how the developed theory might be used and\napproximated in autonomous reasoning agents. Our theory is a step towards a\nglobally consistent and empirically satisfactory unification of probability and\nlogic.', 'The representation of knowledge based on first-order logic captures the\nrichness of natural language and supports multiple probabilistic inference\nmodels. Although symbolic representation enables quantitative reasoning with\nstatistical probability, it is difficult to utilize with machine learning\nmodels as they perform numerical operations. In contrast, knowledge embedding\n(i.e., high-dimensional and continuous vectors) is a feasible approach to\ncomplex reasoning that can not only retain the semantic information of\nknowledge but also establish the quantifiable relationship among them. In this\npaper, we propose recursive neural knowledge network (RNKN), which combines\nmedical knowledge based on first-order logic with recursive neural network for\nmulti-disease diagnosis. After RNKN is efficiently trained from manually\nannotated Chinese Electronic Medical Records (CEMRs), diagnosis-oriented\nknowledge embeddings and weight matrixes are learned. Experimental results\nverify that the diagnostic accuracy of RNKN is superior to that of some\nclassical machine learning models and Markov logic network (MLN). The results\nalso demonstrate that the more explicit the evidence extracted from CEMRs is,\nthe better is the performance achieved. RNKN gradually exhibits the\ninterpretation of knowledge embeddings as the number of training epochs\nincreases.', 'Large knowledge bases (KBs) are useful in many tasks, but it is unclear how\nto integrate this sort of knowledge into ""deep"" gradient-based learning\nsystems. To address this problem, we describe a probabilistic deductive\ndatabase, called TensorLog, in which reasoning uses a differentiable process.\nIn TensorLog, each clause in a logical theory is first converted into certain\ntype of factor graph. Then, for each type of query to the factor graph, the\nmessage-passing steps required to perform belief propagation (BP) are\n""unrolled"" into a function, which is differentiable. We show that these\nfunctions can be composed recursively to perform inference in non-trivial\nlogical theories containing multiple interrelated clauses and predicates. Both\ncompilation and inference in TensorLog are efficient: compilation is linear in\ntheory size and proof depth, and inference is linear in database size and the\nnumber of message-passing steps used in BP. We also present experimental\nresults with TensorLog and discuss its relationship to other first-order\nprobabilistic logics.', 'In this work, we present a novel approach to ontology reasoning that is based\non deep learning rather than logic-based formal reasoning. To this end, we\nintroduce a new model for statistical relational learning that is built upon\ndeep recursive neural networks, and give experimental evidence that it can\neasily compete with, or even outperform, existing logic-based reasoners on the\ntask of ontology reasoning. More precisely, we compared our implemented system\nwith one of the best logic-based ontology reasoners at present, RDFox, on a\nnumber of large standard benchmark datasets, and found that our system attained\nhigh reasoning quality, while being up to two orders of magnitude faster.', ""We propose a framework for inferring the latent attitudes or preferences of\nusers by performing probabilistic first-order logical reasoning over the social\nnetwork graph. Our method answers questions about Twitter users like {\\em Does\nthis user like sushi?} or {\\em Is this user a New York Knicks fan?} by building\na probabilistic model that reasons over user attributes (the user's location or\ngender) and the social network (the user's friends and spouse), via inferences\nlike homophily (I am more likely to like sushi if spouse or friends like sushi,\nI am more likely to like the Knicks if I live in New York). The algorithm uses\ndistant supervision, semi-supervised data harvesting and vector space models to\nextract user attributes (e.g. spouse, education, location) and preferences\n(likes and dislikes) from text. The extracted propositions are then fed into a\nprobabilistic reasoner (we investigate both Markov Logic and Probabilistic Soft\nLogic). Our experiments show that probabilistic logical reasoning significantly\nimproves the performance on attribute and relation extraction, and also\nachieves an F-score of 0.791 at predicting a users likes or dislikes,\nsignificantly better than two strong baselines."", 'Probabilistic inference procedures are usually coded painstakingly from\nscratch, for each target model and each inference algorithm. We reduce this\neffort by generating inference procedures from models automatically. We make\nthis code generation modular by decomposing inference algorithms into reusable\nprogram-to-program transformations. These transformations perform exact\ninference as well as generate probabilistic programs that compute expectations,\ndensities, and MCMC samples. The resulting inference procedures are about as\naccurate and fast as other probabilistic programming systems on real-world\nproblems.', ""Search engines can quickly response a hyperlink list according to query\nkeywords. However, when a query is complex, developers need to repeatedly\nrefine the search keywords and open a large number of web pages to find and\nsummarize answers. Many research works of question and answering (Q and A)\nsystem attempt to assist search engines by providing simple, accurate and\nunderstandable answers. However, without original semantic contexts, these\nanswers lack explainability, making them difficult for users to trust and\nadopt. In this paper, a brain-inspired search engine assistant named\nDeveloperBot based on knowledge graph is proposed, which aligns to the\ncognitive process of human and has the capacity to answer complex queries with\nexplainability. Specifically, DeveloperBot firstly constructs a multi-layer\nquery graph by splitting a complex multi-constraint query into several ordered\nconstraints. Then it models the constraint reasoning process as subgraph search\nprocess inspired by the spreading activation model of cognitive science. In the\nend, novel features of the subgraph will be extracted for decision-making. The\ncorresponding reasoning subgraph and answer confidence will be derived as\nexplanations. The results of the decision-making demonstrate that DeveloperBot\ncan estimate the answers and answer confidences with high accuracy. We\nimplement a prototype and conduct a user study to evaluate whether and how the\ndirect answers and the explanations provided by DeveloperBot can assist\ndevelopers' information needs."", ""Real-world tasks often exhibit a compositional structure that contains a\nsequence of simpler sub-tasks. For instance, opening a door requires reaching,\ngrasping, rotating, and pulling the door knob. Such compositional tasks require\nan agent to reason about the sub-task at hand while orchestrating global\nbehavior accordingly. This can be cast as an online task inference problem,\nwhere the current task identity, represented by a context variable, is\nestimated from the agent's past experiences with probabilistic inference.\nPrevious approaches have employed simple latent distributions, e.g., Gaussian,\nto model a single context for the entire task. However, this formulation lacks\nthe expressiveness to capture the composition and transition of the sub-tasks.\nWe propose a variational inference framework OCEAN to perform online task\ninference for compositional tasks. OCEAN models global and local context\nvariables in a joint latent space, where the global variables represent a\nmixture of sub-tasks required for the task, while the local variables capture\nthe transitions between the sub-tasks. Our framework supports flexible latent\ndistributions based on prior knowledge of the task structure and can be trained\nin an unsupervised manner. Experimental results show that OCEAN provides more\neffective task inference with sequential context adaptation and thus leads to a\nperformance boost on complex, multi-stage tasks."", 'The dynamic, real-time, and accurate inference of model parameters from\nempirical data is of great importance in many scientific and engineering\ndisciplines that use computational models (such as a digital twin) for the\nanalysis and prediction of complex physical processes. However, fast and\naccurate inference for processes with large and high dimensional datasets\ncannot easily be achieved with state-of-the-art methods under noisy real-world\nconditions. The primary reason is that the inference of model parameters with\ntraditional techniques based on optimisation or sampling often suffers from\ncomputational and statistical challenges, resulting in a trade-off between\naccuracy and deployment time. In this paper, we propose a novel framework for\ninference of model parameters based on reinforcement learning. The contribution\nof the paper is twofold: 1) We reformulate the inference problem as a tracking\nproblem with the objective of learning a policy that forces the response of the\nphysics-based model to follow the observations; 2) We propose the constrained\nLyapunov-based actor-critic (CLAC) algorithm to enable the robust and accurate\ninference of physics-based model parameters in real time under noisy real-world\nconditions. The proposed methodology is demonstrated and evaluated on two\nmodel-based diagnostics test cases utilizing two different physics-based models\nof turbofan engines. The performance of the methodology is compared to that of\ntwo alternative approaches: a state update method (unscented Kalman filter) and\na supervised end-to-end mapping with deep neural networks. The experimental\nresults demonstrate that the proposed methodology outperforms all other tested\nmethods in terms of speed and robustness, with high inference accuracy.', 'We present the Power Law Graph Transformer, a transformer model with well\ndefined deductive and inductive tasks for prediction and representation\nlearning. The deductive task learns the dataset level (global) and instance\nlevel (local) graph structures in terms of learnable power law distribution\nparameters. The inductive task outputs the prediction probabilities using the\ndeductive task output, similar to a transductive model. We trained our model\nwith Turkish-English and Portuguese-English datasets from TED talk transcripts\nfor machine translation and compared the model performance and characteristics\nto a transformer model with scaled dot product attention trained on the same\nexperimental setup. We report BLEU scores of $17.79$ and $28.33$ on the\nTurkish-English and Portuguese-English translation tasks with our model,\nrespectively. We also show how a duality between a quantization set and\nN-dimensional manifold representation can be leveraged to transform between\nlocal and global deductive-inductive outputs using successive application of\nlinear and non-linear transformations end-to-end.', 'The questions in a crowdsourcing task typically exhibit varying degrees of\ndifficulty and subjectivity. Their joint effects give rise to the variation in\nresponses to the same question by different crowd-workers. This variation is\nlow when the question is easy to answer and objective, and high when it is\ndifficult and subjective. Unfortunately, current quality control methods for\ncrowdsourcing consider only the question difficulty to account for the\nvariation. As a result,these methods cannot distinguish workers personal\npreferences for different correct answers of a partially subjective question\nfrom their ability/expertise to avoid objectively wrong answers for that\nquestion. To address this issue, we present a probabilistic model which (i)\nexplicitly encodes question difficulty as a model parameter and (ii) implicitly\nencodes question subjectivity via latent preference factors for crowd-workers.\nWe show that question subjectivity induces grouping of crowd-workers, revealed\nthrough clustering of their latent preferences. Moreover, we develop a\nquantitative measure of the subjectivity of a question. Experiments show that\nour model(1) improves the performance of both quality control for crowd-sourced\nanswers and next answer prediction for crowd-workers,and (2) can potentially\nprovide coherent rankings of questions in terms of their difficulty and\nsubjectivity, so that task providers can refine their designs of the\ncrowdsourcing tasks, e.g. by removing highly subjective questions or\ninappropriately difficult questions.', 'Our interest in this paper is in meeting a rapidly growing industrial demand\nfor information extraction from images of documents such as invoices, bills,\nreceipts etc. In practice users are able to provide a very small number of\nexample images labeled with the information that needs to be extracted. We\nadopt a novel two-level neuro-deductive, approach where (a) we use pre-trained\ndeep neural networks to populate a relational database with facts about each\ndocument-image; and (b) we use a form of deductive reasoning, related to\nmeta-interpretive learning of transition systems to learn extraction programs:\nGiven task-specific transitions defined using the entities and relations\nidentified by the neural detectors and a small number of instances (usually 1,\nsometimes 2) of images and the desired outputs, a resource-bounded\nmeta-interpreter constructs proofs for the instance(s) via logical deduction; a\nset of logic programs that extract each desired entity is easily synthesized\nfrom such proofs. In most cases a single training example together with a\nnoisy-clone of itself suffices to learn a program-set that generalizes well on\ntest documents, at which time the value of each entity is determined by a\nmajority vote across its program-set. We demonstrate our two-level\nneuro-deductive approach on publicly available datasets (""Patent"" and ""Doctor\'s\nBills"") and also describe its use in a real-life industrial problem.', 'The Visual Dialogue task requires an agent to engage in a conversation about\nan image with a human. It represents an extension of the Visual Question\nAnswering task in that the agent needs to answer a question about an image, but\nit needs to do so in light of the previous dialogue that has taken place. The\nkey challenge in Visual Dialogue is thus maintaining a consistent, and natural\ndialogue while continuing to answer questions correctly. We present a novel\napproach that combines Reinforcement Learning and Generative Adversarial\nNetworks (GANs) to generate more human-like responses to questions. The GAN\nhelps overcome the relative paucity of training data, and the tendency of the\ntypical MLE-based approach to generate overly terse answers. Critically, the\nGAN is tightly integrated into the attention mechanism that generates\nhuman-interpretable reasons for each answer. This means that the discriminative\nmodel of the GAN has the task of assessing whether a candidate answer is\ngenerated by a human or not, given the provided reason. This is significant\nbecause it drives the generative model to produce high quality answers that are\nwell supported by the associated reasoning. The method also generates the\nstate-of-the-art results on the primary benchmark.', ""Answering open-ended questions is an essential capability for any intelligent\nagent. One of the most interesting recent open-ended question answering\nchallenges is Visual Question Answering (VQA) which attempts to evaluate a\nsystem's visual understanding through its answers to natural language questions\nabout images. There exist many approaches to VQA, the majority of which do not\nexhibit deeper semantic understanding of the candidate answers they produce. We\nstudy the importance of generating plausible answers to a given question by\nintroducing the novel task of `Answer Proposal': for a given open-ended\nquestion, a system should generate a ranked list of candidate answers informed\nby the semantics of the question. We experiment with various models including a\nneural generative model as well as a semantic graph matching one. We provide\nboth intrinsic and extrinsic evaluations for the task of Answer Proposal,\nshowing that our best model learns to propose plausible answers with a high\nrecall and performs competitively with some other solutions to VQA."", 'The black-box nature of neural models has motivated a line of research that\naims to generate natural language rationales to explain why a model made\ncertain predictions. Such rationale generation models, to date, have been\ntrained on dataset-specific crowdsourced rationales, but this approach is\ncostly and is not generalizable to new tasks and domains. In this paper, we\ninvestigate the extent to which neural models can reason about natural language\nrationales that explain model predictions, relying only on distant supervision\nwith no additional annotation cost for human-written rationales. We investigate\nmultiple ways to automatically generate rationales using pre-trained language\nmodels, neural knowledge models, and distant supervision from related tasks,\nand train generative models capable of composing explanatory rationales for\nunseen instances. We demonstrate our approach on the defeasible inference task,\na nonmonotonic reasoning task in which an inference may be strengthened or\nweakened when new information (an update) is introduced. Our model shows\npromises at generating post-hoc rationales explaining why an inference is more\nor less likely given the additional information, however, it mostly generates\ntrivial rationales reflecting the fundamental limitations of neural language\nmodels. Conversely, the more realistic setup of jointly predicting the update\nor its type and generating rationale is more challenging, suggesting an\nimportant future direction.', 'This paper makes a first step towards a logic of learning from experiments.\nFor this, we investigate formal frameworks for modeling the interaction of\ncausal and (qualitative) epistemic reasoning. Crucial for our approach is the\nidea that the notion of an intervention can be used as a formal expression of a\n(real or hypothetical) experiment. In a first step we extend the well-known\ncausal models with a simple Hintikka-style representation of the epistemic\nstate of an agent. In the resulting setting, one can talk not only about the\nknowledge of an agent about the values of variables and how interventions\naffect them, but also about knowledge update. The resulting logic can model\nreasoning about thought experiments. However, it is unable to account for\nlearning from experiments, which is clearly brought out by the fact that it\nvalidates the no learning principle for interventions. Therefore, in a second\nstep, we implement a more complex notion of knowledge that allows an agent to\nobserve (measure) certain variables when an experiment is carried out. This\nextended system does allow for learning from experiments. For all the proposed\nlogical systems, we provide a sound and complete axiomatization.', 'Reasoning and inference are central to human and artificial intelligence.\nModeling inference in human language is very challenging. With the availability\nof large annotated data (Bowman et al., 2015), it has recently become feasible\nto train neural network based inference models, which have shown to be very\neffective. In this paper, we present a new state-of-the-art result, achieving\nthe accuracy of 88.6% on the Stanford Natural Language Inference Dataset.\nUnlike the previous top models that use very complicated network architectures,\nwe first demonstrate that carefully designing sequential inference models based\non chain LSTMs can outperform all previous models. Based on this, we further\nshow that by explicitly considering recursive architectures in both local\ninference modeling and inference composition, we achieve additional\nimprovement. Particularly, incorporating syntactic parsing information\ncontributes to our best result---it further improves the performance even when\nadded to the already very strong model.', 'The tension between deduction and induction is perhaps the most fundamental\nissue in areas such as philosophy, cognition and artificial intelligence (AI).\nThe deduction camp concerns itself with questions about the expressiveness of\nformal languages for capturing knowledge about the world, together with proof\nsystems for reasoning from such knowledge bases. The learning camp attempts to\ngeneralize from examples about partial descriptions about the world. In AI,\nhistorically, these camps have loosely divided the development of the field,\nbut advances in cross-over areas such as statistical relational learning,\nneuro-symbolic systems, and high-level control have illustrated that the\ndichotomy is not very constructive, and perhaps even ill-formed. In this\narticle, we survey work that provides further evidence for the connections\nbetween logic and learning. Our narrative is structured in terms of three\nstrands: logic versus learning, machine learning for logic, and logic for\nmachine learning, but naturally, there is considerable overlap. We place an\nemphasis on the following ""sore"" point: there is a common misconception that\nlogic is for discrete properties, whereas probability theory and machine\nlearning, more generally, is for continuous properties. We report on results\nthat challenge this view on the limitations of logic, and expose the role that\nlogic can play for learning in infinite domains.', ""We fine-tune GPT-3 to answer long-form questions using a text-based\nweb-browsing environment, which allows the model to search and navigate the\nweb. By setting up the task so that it can be performed by humans, we are able\nto train models on the task using imitation learning, and then optimize answer\nquality with human feedback. To make human evaluation of factual accuracy\neasier, models must collect references while browsing in support of their\nanswers. We train and evaluate our models on ELI5, a dataset of questions asked\nby Reddit users. Our best model is obtained by fine-tuning GPT-3 using behavior\ncloning, and then performing rejection sampling against a reward model trained\nto predict human preferences. This model's answers are preferred by humans 56%\nof the time to those of our human demonstrators, and 69% of the time to the\nhighest-voted answer from Reddit."", 'Inference and decision making under uncertainty are key processes in every\nautonomous system and numerous robotic problems. In recent years, the\nsimilarities between inference and decision making triggered much work, from\ndeveloping unified computational frameworks to pondering about the duality\nbetween the two. In spite of these efforts, inference and control, as well as\ninference and belief space planning (BSP) are still treated as two separate\nprocesses. In this paper we propose a paradigm shift, a novel approach which\ndeviates from conventional Bayesian inference and utilizes the similarities\nbetween inference and BSP. We make the key observation that inference can be\nefficiently updated using predictions made during the decision making stage,\neven in light of inconsistent data association between the two. We developed a\ntwo staged process that implements our novel approach and updates inference\nusing calculations from the precursory planning phase. Using autonomous\nnavigation in an unknown environment along with iSAM2 efficient methodologies\nas a test case, we benchmarked our novel approach against standard Bayesian\ninference, both with synthetic and real-world data (KITTI dataset). Results\nindicate that not only our approach improves running time by at least a factor\nof two while providing the same estimation accuracy, but it also alleviates the\ncomputational burden of state dimensionality and loop closures.', 'Logic reasoning is a significant ability of human intelligence and also an\nimportant task in artificial intelligence. The existing logic reasoning\nmethods, quite often, need to design some reasoning patterns beforehand. This\nhas led to an interesting question: can logic reasoning patterns be directly\nlearned from given data? The problem is termed as a data concept logic. In this\nstudy, a learning logic task from images, called a LiLi task, first is\nproposed. This task is to learn and reason the logic relation from images,\nwithout presetting any reasoning patterns. As a preliminary exploration, we\ndesign six LiLi data sets (Bitwise And, Bitwise Or, Bitwise Xor, Addition,\nSubtraction and Multiplication), in which each image is embedded with a n-digit\nnumber. It is worth noting that a learning model beforehand does not know the\nmeaning of the n-digit numbers embedded in images and the relation between the\ninput images and the output image. In order to tackle the task, in this work we\nuse many typical neural network models and produce fruitful results. However,\nthese models have the poor performances on the difficult logic task. For\nfurthermore addressing this task, a novel network framework called a divide and\nconquer model by adding some label information is designed, achieving a high\ntesting accuracy.', 'We explore a quantitative interpretation of 2-dimensional intuitionistic type\ntheory (ITT) in which the identity type is interpreted as a ""type of\ndifferences"". We show that a fragment of ITT, that we call difference type\ntheory (dTT), yields a general logical framework to talk about quantitative\nproperties of programs like approximate equivalence and metric preservation. To\ndemonstrate this fact, we show that dTT can be used to capture compositional\nreasoning in presence of errors, since any program can be associated with a\n""derivative"" relating errors in input with errors in output. Moreover, after\nrelating the semantics of dTT to the standard weak factorization systems\nsemantics of ITT, we describe the interpretation of dTT in some quantitative\nmodels developed for approximate program transformations, incremental\ncomputing, program differentiation and differential privacy.', 'Many intellectual endeavors require mathematical problem solving, but this\nskill remains beyond the capabilities of computers. To measure this ability in\nmachine learning models, we introduce MATH, a new dataset of 12,500 challenging\ncompetition mathematics problems. Each problem in MATH has a full step-by-step\nsolution which can be used to teach models to generate answer derivations and\nexplanations. To facilitate future research and increase accuracy on MATH, we\nalso contribute a large auxiliary pretraining dataset which helps teach models\nthe fundamentals of mathematics. Even though we are able to increase accuracy\non MATH, our results show that accuracy remains relatively low, even with\nenormous Transformer models. Moreover, we find that simply increasing budgets\nand model parameter counts will be impractical for achieving strong\nmathematical reasoning if scaling trends continue. While scaling Transformers\nis automatically solving most other text-based tasks, scaling is not currently\nsolving MATH. To have more traction on mathematical problem solving we will\nlikely need new algorithmic advancements from the broader research community.', 'This paper presents a novel method to generate answers for non-extraction\nmachine reading comprehension (MRC) tasks whose answers cannot be simply\nextracted as one span from the given passages. Using a pointer network-style\nextractive decoder for such type of MRC may result in unsatisfactory\nperformance when the ground-truth answers are given by human annotators or\nhighly re-paraphrased from parts of the passages. On the other hand, using\ngenerative decoder cannot well guarantee the resulted answers with well-formed\nsyntax and semantics when encountering long sentences. Therefore, to alleviate\nthe obvious drawbacks of both sides, we propose an answer making-up method from\nextracted multi-spans that are learned by our model as highly confident\n$n$-gram candidates in the given passage. That is, the returned answers are\ncomposed of discontinuous multi-spans but not just one consecutive span in the\ngiven passages anymore. The proposed method is simple but effective: empirical\nexperiments on MS MARCO show that the proposed method has a better performance\non accurately generating long answers, and substantially outperforms two\ncompetitive typical one-span and Seq2Seq baseline decoders.', 'The analysis of practical probabilistic models on the computer demands a\nconvenient representation for the available knowledge and an efficient\nalgorithm to perform inference. An appealing representation is the influence\ndiagram, a network that makes explicit the random variables in a model and\ntheir probabilistic dependencies. Recent advances have developed solution\nprocedures based on the influence diagram. In this paper, we examine the\nfundamental properties that underlie those techniques, and the information\nabout the probabilistic structure that is available in the influence diagram\nrepresentation. The influence diagram is a convenient representation for\ncomputer processing while also being clear and non-mathematical. It displays\nprobabilistic dependence precisely, in a way that is intuitive for decision\nmakers and experts to understand and communicate. As a result, the same\ninfluence diagram can be used to build, assess and analyze a model,\nfacilitating changes in the formulation and feedback from sensitivity analysis.\nThe goal in this paper is to determine arbitrary conditional probability\ndistributions from a given probabilistic model. Given qualitative information\nabout the dependence of the random variables in the model we can, for a\nspecific conditional expression, specify precisely what quantitative\ninformation we need to be able to determine the desired conditional probability\ndistribution. It is also shown how we can find that probability distribution by\nperforming operations locally, that is, over subspaces of the joint\ndistribution. In this way, we can exploit the conditional independence present\nin the model to avoid having to construct or manipulate the full joint\ndistribution. These results are extended to include maximal processing when the\ninformation available is incomplete, and optimal decision making in an\nuncertain environment. Influence diagrams as a computer-aided modeling tool\nwere developed by Miller, Merkofer, and Howard [5] and extended by Howard and\nMatheson [2]. Good descriptions of how to use them in modeling are in Owen [7]\nand Howard and Matheson [2]. The notion of solving a decision problem through\ninfluence diagrams was examined by Olmsted [6] and such an algorithm was\ndeveloped by Shachter [8]. The latter paper also shows how influence diagrams\ncan be used to perform a variety of sensitivity analyses. This paper extends\nthose results by developing a theory of the properties of the diagram that are\nused by the algorithm, and the information needed to solve arbitrary\nprobability inference problems. Section 2 develops the notation and the\nframework for the paper and the relationship between influence diagrams and\njoint probability distributions. The general probabilistic inference problem is\nposed in Section 3. In Section 4 the transformations on the diagram are\ndeveloped and then put together into a solution procedure in Section 5. In\nSection 6, this procedure is used to calculate the information requirement to\nsolve an inference problem and the maximal processing that can be performed\nwith incomplete information. Section 7 contains a summary of results.', 'Self/other distinction and self-recognition are important skills for\ninteracting with the world, as it allows humans to differentiate own actions\nfrom others and be self-aware. However, only a selected group of animals,\nmainly high order mammals such as humans, has passed the mirror test, a\nbehavioural experiment proposed to assess self-recognition abilities. In this\npaper, we describe self-recognition as a process that is built on top of body\nperception unconscious mechanisms. We present an algorithm that enables a robot\nto perform non-appearance self-recognition on a mirror and distinguish its\nsimple actions from other entities, by answering the following question: am I\ngenerating these sensations? The algorithm combines active inference, a\ntheoretical model of perception and action in the brain, with neural network\nlearning. The robot learns the relation between its actions and its body with\nthe effect produced in the visual field and its body sensors. The prediction\nerror generated between the models and the real observations during the\ninteraction is used to infer the body configuration through free energy\nminimization and to accumulate evidence for recognizing its body. Experimental\nresults on a humanoid robot show the reliability of the algorithm for different\ninitial conditions, such as mirror recognition in any perspective, robot-robot\ndistinction and human-robot differentiation.', 'The use of deductive techniques, such as theorem provers, has several\nadvantages in safety verification of hybrid sys- tems; however,\nstate-of-the-art theorem provers require ex- tensive manual intervention.\nFurthermore, there is often a gap between the type of assistance that a theorem\nprover requires to make progress on a proof task and the assis- tance that a\nsystem designer is able to provide. This paper presents an extension to\nKeYmaera, a deductive verification tool for differential dynamic logic; the new\ntechnique allows local reasoning using system designer intuition about per-\nformance within particular modes as part of a proof task. Our approach allows\nthe theorem prover to leverage for- ward invariants, discovered using numerical\ntechniques, as part of a proof of safety. We introduce a new inference rule\ninto the proof calculus of KeYmaera, the forward invariant cut rule, and we\npresent a methodology to discover useful forward invariants, which are then\nused with the new cut rule to complete verification tasks. We demonstrate how\nour new approach can be used to complete verification tasks that lie out of the\nreach of existing deductive approaches us- ing several examples, including one\ninvolving an automotive powertrain control system.', ""Most software verification tools can be classified into one of a number of\nestablished families, each of which has their own focus and strengths. For\nexample, concrete counterexample generation in model checking, invariant\ninference in abstract interpretation and completeness via annotation for\ndeductive verification. This creates a significant and fundamental usability\nproblem as users may have to learn and use one technique to find potential\nproblems but then need an entirely different one to show that they have been\nfixed. This paper presents a single, unified algorithm kIkI, which strictly\ngeneralises abstract interpretation, bounded model checking and k-induction.\nThis not only combines the strengths of these techniques but allows them to\ninteract and reinforce each other, giving a `single-tool' approach to\nverification."", 'Conversational and task-oriented dialogue systems aim to interact with the\nuser using natural responses through multi-modal interfaces, such as text or\nspeech. These desired responses are in the form of full-length natural answers\ngenerated over facts retrieved from a knowledge source. While the task of\ngenerating natural answers to questions from an answer span has been widely\nstudied, there has been little research on natural sentence generation over\nspoken content. We propose a novel system to generate full length natural\nlanguage answers from spoken questions and factoid answers. The spoken sequence\nis compactly represented as a confusion network extracted from a pre-trained\nAutomatic Speech Recognizer. This is the first attempt towards generating\nfull-length natural answers from a graph input(confusion network) to the best\nof our knowledge. We release a large-scale dataset of 259,788 samples of spoken\nquestions, their factoid answers and corresponding full-length textual answers.\nFollowing our proposed approach, we achieve comparable performance with best\nASR hypothesis.', ""Objective: We investigate whether deep learning techniques for natural\nlanguage processing (NLP) can be used efficiently for patient phenotyping.\nPatient phenotyping is a classification task for determining whether a patient\nhas a medical condition, and is a crucial part of secondary analysis of\nhealthcare data. We assess the performance of deep learning algorithms and\ncompare them with classical NLP approaches.\n  Materials and Methods: We compare convolutional neural networks (CNNs),\nn-gram models, and approaches based on cTAKES that extract pre-defined medical\nconcepts from clinical notes and use them to predict patient phenotypes. The\nperformance is tested on 10 different phenotyping tasks using 1,610 discharge\nsummaries extracted from the MIMIC-III database.\n  Results: CNNs outperform other phenotyping algorithms in all 10 tasks. The\naverage F1-score of our model is 76 (PPV of 83, and sensitivity of 71) with our\nmodel having an F1-score up to 37 points higher than alternative approaches. We\nadditionally assess the interpretability of our model by presenting a method\nthat extracts the most salient phrases for a particular prediction.\n  Conclusion: We show that NLP methods based on deep learning improve the\nperformance of patient phenotyping. Our CNN-based algorithm automatically\nlearns the phrases associated with each patient phenotype. As such, it reduces\nthe annotation complexity for clinical domain experts, who are normally\nrequired to develop task-specific annotation rules and identify relevant\nphrases. Our method performs well in terms of both performance and\ninterpretability, which indicates that deep learning is an effective approach\nto patient phenotyping based on clinicians' notes."", ""Question answering is an important and difficult task in the natural language\nprocessing domain, because many basic natural language processing tasks can be\ncast into a question answering task. Several deep neural network architectures\nhave been developed recently, which employ memory and inference components to\nmemorize and reason over text information, and generate answers to questions.\nHowever, a major drawback of many such models is that they are capable of only\ngenerating single-word answers. In addition, they require large amount of\ntraining data to generate accurate answers. In this paper, we introduce the\nLong-Term Memory Network (LTMN), which incorporates both an external memory\nmodule and a Long Short-Term Memory (LSTM) module to comprehend the input data\nand generate multi-word answers. The LTMN model can be trained end-to-end using\nback-propagation and requires minimal supervision. We test our model on two\nsynthetic data sets (based on Facebook's bAbI data set) and the real-world\nStanford question answering data set, and show that it can achieve\nstate-of-the-art performance."", 'Tackling Natural Language Inference with a logic-based method is becoming\nless and less common. While this might have been counterintuitive several\ndecades ago, nowadays it seems pretty obvious. The main reasons for such a\nconception are that (a) logic-based methods are usually brittle when it comes\nto processing wide-coverage texts, and (b) instead of automatically learning\nfrom data, they require much of manual effort for development. We make a step\ntowards to overcome such shortcomings by modeling learning from data as\nabduction: reversing a theorem-proving procedure to abduce semantic relations\nthat serve as the best explanation for the gold label of an inference problem.\nIn other words, instead of proving sentence-level inference relations with the\nhelp of lexical relations, the lexical relations are proved taking into account\nthe sentence-level inference relations. We implement the learning method in a\ntableau theorem prover for natural language and show that it improves the\nperformance of the theorem prover on the SICK dataset by 1.4% while still\nmaintaining high precision (>94%). The obtained results are competitive with\nthe state of the art among logic-based systems.', 'Natural Language Inference (NLI) is the task of determining the semantic\nrelationship between a premise and a hypothesis. In this paper, we focus on the\n{\\em generation} of hypotheses from premises in a multimodal setting, to\ngenerate a sentence (hypothesis) given an image and/or its description\n(premise) as the input. The main goals of this paper are (a) to investigate\nwhether it is reasonable to frame NLI as a generation task; and (b) to consider\nthe degree to which grounding textual premises in visual information is\nbeneficial to generation. We compare different neural architectures, showing\nthrough automatic and human evaluation that entailments can indeed be generated\nsuccessfully. We also show that multimodal models outperform unimodal models in\nthis task, albeit marginally.', ""As the interest in Artificial Intelligence continues to grow it is becoming\nmore and more important to investigate formalization and tools that allow us to\nexploit logic to reason about the world. In particular, given the increasing\nnumber of multi-agents systems that could benefit from techniques of automated\nreasoning, exploring new ways to define not only the world's status but also\nthe agents' information is constantly growing in importance. This type of\nreasoning, i.e., about agents' perception of the world and also about agents'\nknowledge of her and others' knowledge, is referred to as epistemic reasoning.\n  In our work we will try to formalize this concept, expressed through\nepistemic logic, for dynamic domains. In particular we will attempt to define a\nnew action-based language for multi-agent epistemic planning and to implement\nan epistemic planner based on it. This solver should provide a tool flexible\nenough to be able to reason on different domains, e.g., economy, security,\njustice and politics, where reasoning about others' beliefs could lead to\nwinning strategies or help in changing a group of agents' view of the world."", 'Owing to the advancement of deep learning, artificial systems are now rival\nto humans in several pattern recognition tasks, such as visual recognition of\nobject categories. However, this is only the case with the tasks for which\ncorrect answers exist independent of human perception. There is another type of\ntasks for which what to predict is human perception itself, in which there are\noften individual differences. Then, there are no longer single ""correct""\nanswers to predict, which makes evaluation of artificial systems difficult. In\nthis paper, focusing on pairwise ranking tasks sensitive to individual\ndifferences, we propose an evaluation method. Given a ranking result for\nmultiple item pairs that is generated by an artificial system, our method\nquantifies the probability that the same ranking result will be generated by\nhumans, and judges if it is distinguishable from human-generated results. We\nintroduce a probabilistic model of human ranking behavior, and present an\nefficient computation method for the judgment. To estimate model parameters\naccurately from small-size samples, we present a method that uses confidence\nscores given by annotators for ranking each item pair. Taking as an example a\ntask of ranking image pairs according to material attributes of objects, we\ndemonstrate how the proposed method works.', 'This paper tackles the goal of conclusion-supplement answer generation for\nnon-factoid questions, which is a critical issue in the field of Natural\nLanguage Processing (NLP) and Artificial Intelligence (AI), as users often\nrequire supplementary information before accepting a conclusion. The current\nencoder-decoder framework, however, has difficulty generating such answers,\nsince it may become confused when it tries to learn several different long\nanswers to the same non-factoid question. Our solution, called an ensemble\nnetwork, goes beyond single short sentences and fuses logically connected\nconclusion statements and supplementary statements. It extracts the context\nfrom the conclusion decoder\'s output sequence and uses it to create\nsupplementary decoder states on the basis of an attention mechanism. It also\nassesses the closeness of the question encoder\'s output sequence and the\nseparate outputs of the conclusion and supplement decoders as well as their\ncombination. As a result, it generates answers that match the questions and\nhave natural-sounding supplementary sequences in line with the context\nexpressed by the conclusion sequence. Evaluations conducted on datasets\nincluding ""Love Advice"" and ""Arts & Humanities"" categories indicate that our\nmodel outputs much more accurate results than the tested baseline models do.', 'The most important problems for society are describable only in vague terms,\ndependent on subjective positions, and missing highly relevant data. This\nthesis is intended to revive and further develop the view that giving\nnon-trivial, rigorous deductive arguments concerning such problems -without\neliminating the complications of vagueness, subjectivity, and uncertainty- is,\nthough very difficult, not problematic in principle, does not require the\ninvention of new logics -classical first-order logic will do- and is something\nthat more mathematically-inclined people should be pursuing. The framework of\ninterpreted formal proofs is presented for formalizing and criticizing rigorous\ndeductive arguments about vague, subjective, and uncertain issues, and its\nadequacy is supported largely by a number of major examples. This thesis also\ndocuments progress towards a web system for collaboratively authoring and\ncriticizing such arguments, which is the ultimate goal of this project.', 'The dynamics of belief and knowledge is one of the major components of any\nautonomous system that should be able to incorporate new pieces of information.\nIn order to apply the rationality result of belief dynamics theory to various\npractical problems, it should be generalized in two respects: first it should\nallow a certain part of belief to be declared as immutable; and second, the\nbelief state need not be deductively closed. Such a generalization of belief\ndynamics, referred to as base dynamics, is presented in this paper, along with\nthe concept of a generalized revision algorithm for knowledge bases (Horn or\nHorn logic with stratified negation). We show that knowledge base dynamics has\nan interesting connection with kernel change via hitting set and abduction. In\nthis paper, we show how techniques from disjunctive logic programming can be\nused for efficient (deductive) database updates. The key idea is to transform\nthe given database together with the update request into a disjunctive\n(datalog) logic program and apply disjunctive techniques (such as minimal model\nreasoning) to solve the original update problem. The approach extends and\nintegrates standard techniques for efficient query answering and integrity\nchecking. The generation of a hitting set is carried out through a hyper\ntableaux calculus and magic set that is focused on the goal of minimality.', 'Visual Question Answering is a multi-modal task that aims to measure\nhigh-level visual understanding. Contemporary VQA models are restrictive in the\nsense that answers are obtained via classification over a limited vocabulary\n(in the case of open-ended VQA), or via classification over a set of\nmultiple-choice-type answers. In this work, we present a completely generative\nformulation where a multi-word answer is generated for a visual query. To take\nthis a step forward, we introduce a new task: ViQAR (Visual Question Answering\nand Reasoning), wherein a model must generate the complete answer and a\nrationale that seeks to justify the generated answer. We propose an end-to-end\narchitecture to solve this task and describe how to evaluate it. We show that\nour model generates strong answers and rationales through qualitative and\nquantitative evaluation, as well as through a human Turing Test.', 'For a computational system to be intelligent, it should be able to perform,\nat least, basic deductions. Nonetheless, since deductions are, in some sense,\nequivalent to tautologies, it seems that they do not provide new information.\nThe present article proposes a measure the degree of semantic informativity of\nvalid deductions in a dynamic setting. Concepts of coherency and relevancy,\ndisplayed in terms of insertions and deletions on databases, are used to define\nsemantic informativity. In this way, the article shows that a solution to the\nproblem about the informativity of deductions provides a heuristic principle to\nimprove the deductive power of computational systems.', 'We explore the problem of step-wise explaining how to solve constraint\nsatisfaction problems, with a use case on logic grid puzzles. More\nspecifically, we study the problem of explaining the inference steps that one\ncan take during propagation, in a way that is easy to interpret for a person.\nThereby, we aim to give the constraint solver explainable agency, which can\nhelp in building trust in the solver by being able to understand and even learn\nfrom the explanations. The main challenge is that of finding a sequence of\nsimple explanations, where each explanation should aim to be as cognitively\neasy as possible for a human to verify and understand. This contrasts with the\narbitrary combination of facts and constraints that the solver may use when\npropagating. We propose the use of a cost function to quantify how simple an\nindividual explanation of an inference step is, and identify the\nexplanation-production problem of finding the best sequence of explanations of\na CSP. Our approach is agnostic of the underlying constraint propagation\nmechanisms, and can provide explanations even for inference steps resulting\nfrom combinations of constraints. In case multiple constraints are involved, we\nalso develop a mechanism that allows to break the most difficult steps up and\nthus gives the user the ability to zoom in on specific parts of the\nexplanation. Our proposed algorithm iteratively constructs the explanation\nsequence by using an optimistic estimate of the cost function to guide the\nsearch for the best explanation at each step. Our experiments on logic grid\npuzzles show the feasibility of the approach in terms of the quality of the\nindividual explanations and the resulting explanation sequences obtained.', '(l) I have enough evidence to render the sentence S probable. (la) So,\nrelative to what I know, it is rational of me to believe S. (2) Now that I have\nmore evidence, S may no longer be probable. (2a) So now, relative to what I\nknow, it is not rational of me to believe S. These seem a perfectly ordinary,\ncommon sense, pair of situations. Generally and vaguely, I take them to embody\nwhat I shall call probabilistic inference. This form of inference is clearly\nnon-monotonic. Relatively few people have taken this form of inference, based\non high probability, to serve as a foundation for non-monotonic logic or for a\nlogical or defeasible inference. There are exceptions: Jane Nutter [16] thinks\nthat sometimes probability has something to do with non-monotonic reasoning.\nJudea Pearl [ 17] has recently been exploring the possibility. There are any\nnumber of people whom one might call probability enthusiasts who feel that\nprobability provides all the answers by itself, with no need of help from\nlogic. Cheeseman [1], Henrion [5] and others think it useful to look at a\ndistribution of probabilities over a whole algebra of statements, to update\nthat distribution in the light of new evidence, and to use the latest updated\ndistribution of probability over the algebra as a basis for planning and\ndecision making. A slightly weaker form of this approach is captured by Nilsson\n[15], where one assumes certain probabilities for certain statements, and\ninfers the probabilities, or constraints on the probabilities of other\nstatement. None of this corresponds to what I call probabilistic inference. All\nof the inference that is taking place, either in Bayesian updating, or in\nprobabilistic logic, is strictly deductive. Deductive inference, particularly\nthat concerned with the distribution of classical probabilities or chances, is\nof great importance. But this is not to say that there is no important role for\nwhat earlier logicians have called ""ampliative"" or ""inductive"" or ""scientific""\ninference, in which the conclusion goes beyond the premises, asserts more than\ndo the premises. This depends on what David Israel [6] has called ""real rules\nof inference"". It is characteristic of any such logic or inference procedure\nthat it can go wrong: that statements accepted at one point may be rejected at\na later point. Research underlying the results reported here has been partially\nsupported by the Signals Warfare Center of the United States Army.', 'We present a unified logical framework for representing and reasoning about\nboth probability quantitative and qualitative preferences in probability answer\nset programming, called probability answer set optimization programs. The\nproposed framework is vital to allow defining probability quantitative\npreferences over the possible outcomes of qualitative preferences. We show the\napplication of probability answer set optimization programs to a variant of the\nwell-known nurse restoring problem, called the nurse restoring with probability\npreferences problem. To the best of our knowledge, this development is the\nfirst to consider a logical framework for reasoning about probability\nquantitative preferences, in general, and reasoning about both probability\nquantitative and qualitative preferences in particular.', 'We present a unified logical framework for representing and reasoning about\nboth quantitative and qualitative preferences in fuzzy answer set programming,\ncalled fuzzy answer set optimization programs. The proposed framework is vital\nto allow defining quantitative preferences over the possible outcomes of\nqualitative preferences. We show the application of fuzzy answer set\noptimization programs to the course scheduling with fuzzy preferences problem.\nTo the best of our knowledge, this development is the first to consider a\nlogical framework for reasoning about quantitative preferences, in general, and\nreasoning about both quantitative and qualitative preferences in particular.', 'Deep inference is a proof theoretic methodology that generalizes the standard\nnotion of inference of the sequent calculus, whereby inference rules become\napplicable at any depth inside logical expressions. Deep inference provides\nmore freedom in the design of deductive systems for different logics and a rich\ncombinatoric analysis of proofs. In particular, construction of exponentially\nshorter analytic proofs becomes possible, however with the cost of a greater\nnondeterminism than in the sequent calculus. In this paper, we show that the\nnondeterminism in proof search can be reduced without losing the shorter proofs\nand without sacrificing proof theoretic cleanliness. For this, we exploit an\ninteraction and depth scheme in the logical expressions. We demonstrate our\nmethod on deep inference systems for multiplicative linear logic and classical\nlogic, discuss its proof complexity and its relation to focusing, and present\nimplementations.', 'Recently, FO(C), the integration of C-Log with classical logic, was\nintroduced as a knowledge representation language. Up to this point, no systems\nexist that perform inference on FO(C), and very little is known about\nproperties of inference in FO(C). In this paper, we study both of the above\nproblems. We define normal forms for FO(C), one of which corresponds to FO(ID).\nWe define transformations between these normal forms, and show that, using\nthese transformations, several inference tasks for FO(C) can be reduced to\ninference tasks for FO(ID), for which solvers exist. We implemented a prototype\nof this transformation, and thus present the first system to perform inference\nin FO(C). We also provide results about the complexity of reasoning in FO(C).', 'We propose a benchmark to measure whether a language model is truthful in\ngenerating answers to questions. The benchmark comprises 817 questions that\nspan 38 categories, including health, law, finance and politics. We crafted\nquestions that some humans would answer falsely due to a false belief or\nmisconception. To perform well, models must avoid generating false answers\nlearned from imitating human texts. We tested GPT-3, GPT-Neo/J, GPT-2 and a\nT5-based model. The best model was truthful on 58% of questions, while human\nperformance was 94%. Models generated many false answers that mimic popular\nmisconceptions and have the potential to deceive humans. The largest models\nwere generally the least truthful. This contrasts with other NLP tasks, where\nperformance improves with model size. However, this result is expected if false\nanswers are learned from the training distribution. We suggest that scaling up\nmodels alone is less promising for improving truthfulness than fine-tuning\nusing training objectives other than imitation of text from the web.', 'Arriving at the complete probabilistic knowledge of a domain, i.e., learning\nhow all variables interact, is indeed a demanding task. In reality, settings\noften arise for which an individual merely possesses partial knowledge of the\ndomain, and yet, is expected to give adequate answers to a variety of posed\nqueries. That is, although precise answers to some queries, in principle,\ncannot be achieved, a range of plausible answers is attainable for each query\ngiven the available partial knowledge. In this paper, we propose the\nMulti-Context Model (MCM), a new graphical model to represent the state of\npartial knowledge as to a domain. MCM is a middle ground between Probabilistic\nLogic, Bayesian Logic, and Probabilistic Graphical Models. For this model we\ndiscuss: (i) the dynamics of constructing a contradiction-free MCM, i.e., to\nform partial beliefs regarding a domain in a gradual and probabilistically\nconsistent way, and (ii) how to perform inference, i.e., to evaluate a\nprobability of interest involving some variables of the domain.', ""It is well known that reinforcement learning can be cast as inference in an\nappropriate probabilistic model. However, this commonly involves introducing a\ndistribution over agent trajectories with probabilities proportional to\nexponentiated rewards. In this work, we formulate reinforcement learning as\nBayesian inference without resorting to rewards, and show that rewards are\nderived from agent's preferences, rather than the other way around. We argue\nthat agent preferences should be specified stochastically rather than\ndeterministically. Reinforcement learning via inference with stochastic\npreferences naturally describes agent behaviors, does not require introducing\nrewards and exponential weighing of trajectories, and allows to reason about\nagents using the solid foundation of Bayesian statistics. Stochastic\nconditioning, a probabilistic programming paradigm for conditioning models on\ndistributions rather than values, is the formalism behind agents with\nprobabilistic preferences. We demonstrate realization of our approach on case\nstudies using both a two-agent coordinate game and a single agent acting in a\nnoisy environment, showing that despite superficial differences, both cases can\nbe modeled and reasoned about based on the same principles."", 'Many algorithms and applications involve repeatedly solving variations of the\nsame inference problem; for example we may want to introduce new evidence to\nthe model or perform updates to conditional dependencies. The goal of adaptive\ninference is to take advantage of what is preserved in the model and perform\ninference more rapidly than from scratch. In this paper, we describe techniques\nfor adaptive inference on general graphs that support marginal computation and\nupdates to the conditional probabilities and dependencies in logarithmic time.\nWe give experimental results for an implementation of our algorithm, and\ndemonstrate its potential performance benefit in the study of protein\nstructure.', 'Planning is useful. It lets people take actions that have desirable long-term\nconsequences. But, planning is hard. It requires thinking about consequences,\nwhich consumes limited computational and cognitive resources. Thus, people\nshould plan their actions, but they should also be smart about how they deploy\nresources used for planning their actions. Put another way, people should also\n""plan their plans"". Here, we formulate this aspect of planning as a\nmeta-reasoning problem and formalize it in terms of a recursive Bellman\nobjective that incorporates both task rewards and information-theoretic\nplanning costs. Our account makes quantitative predictions about how people\nshould plan and meta-plan as a function of the overall structure of a task,\nwhich we test in two experiments with human participants. We find that people\'s\nreaction times reflect a planned use of information processing, consistent with\nour account. This formulation of planning to plan provides new insight into the\nfunction of hierarchical planning, state abstraction, and cognitive control in\nboth humans and machines.', 'This paper concerns the relation between process algebra and Hoare logic. We\ninvestigate the question whether and how a Hoare logic can be used for\nreasoning about how data change in the course of a process when reasoning\nequationally about that process. We introduce an extension of ACP (Algebra of\nCommunicating Processes) with features that are relevant to processes in which\ndata are involved, present a Hoare logic for the processes considered in this\nprocess algebra, and discuss the use of this Hoare logic as a complement to\npure equational reasoning with the equational axioms of the process algebra.', 'Continuous queries over data streams may suffer from blocking operations\nand/or unbound wait, which may delay answers until some relevant input arrives\nthrough the data stream. These delays may turn answers, when they arrive,\nobsolete to users who sometimes have to make decisions with no help whatsoever.\nTherefore, it can be useful to provide hypothetical answers - ""given the\ncurrent information, it is possible that X will become true at time t"" -\ninstead of no information at all.\n  In this paper we present a semantics for queries and corresponding answers\nthat covers such hypothetical answers, together with an online algorithm for\nupdating the set of facts that are consistent with the currently available\ninformation.', ""Bilateralists hold that the meanings of the connectives are determined by\nrules of inference for their use in deductive reasoning with asserted and\ndenied formulas. This paper presents two bilateral connectives comparable to\nPrior's tonk, for which, unlike for tonk, there are reduction steps for the\nremoval of maximal formulas arising from introducing and eliminating formulas\nwith those connectives as main operators. Adding either of them to bilateral\nclassical logic results in an incoherent system. One way around this problem is\nto count formulas as maximal that are the conclusion of reductio and major\npremise of an elimination rule and to require their removability from\ndeductions. The main part of the paper consists in a proof of a normalisation\ntheorem for bilateral logic. The closing sections address philosophical\nconcerns whether the proof provides a satisfactory solution to the problem at\nhand and confronts bilateralists with the dilemma that a bilateral notion of\nstability sits uneasily with the core bilateral thesis.\n  The Note corrects an error in one of the reduction steps in the paper."", 'Learning representations of words in a continuous space is perhaps the most\nfundamental task in NLP, however words interact in ways much richer than vector\ndot product similarity can provide. Many relationships between words can be\nexpressed set-theoretically, for example, adjective-noun compounds (eg. ""red\ncars""$\\subseteq$""cars"") and homographs (eg. ""tongue""$\\cap$""body"" should be\nsimilar to ""mouth"", while ""tongue""$\\cap$""language"" should be similar to\n""dialect"") have natural set-theoretic interpretations. Box embeddings are a\nnovel region-based representation which provide the capability to perform these\nset-theoretic operations. In this work, we provide a fuzzy-set interpretation\nof box embeddings, and learn box representations of words using a set-theoretic\ntraining objective. We demonstrate improved performance on various word\nsimilarity tasks, particularly on less common words, and perform a quantitative\nand qualitative analysis exploring the additional unique expressivity provided\nby Word2Box.', 'In this paper, we propose generative probabilistic models for label\naggregation. We use Gibbs sampling and a novel variational inference algorithm\nto perform the posterior inference. Empirical results show that our methods\nconsistently outperform state-of-the-art methods.', 'Approximate probabilistic inference algorithms are central to many fields.\nExamples include sequential Monte Carlo inference in robotics, variational\ninference in machine learning, and Markov chain Monte Carlo inference in\nstatistics. A key problem faced by practitioners is measuring the accuracy of\nan approximate inference algorithm on a specific data set. This paper\nintroduces the auxiliary inference divergence estimator (AIDE), an algorithm\nfor measuring the accuracy of approximate inference algorithms. AIDE is based\non the observation that inference algorithms can be treated as probabilistic\nmodels and the random variables used within the inference algorithm can be\nviewed as auxiliary variables. This view leads to a new estimator for the\nsymmetric KL divergence between the approximating distributions of two\ninference algorithms. The paper illustrates application of AIDE to algorithms\nfor inference in regression, hidden Markov, and Dirichlet process mixture\nmodels. The experiments show that AIDE captures the qualitative behavior of a\nbroad class of inference algorithms and can detect failure modes of inference\nalgorithms that are missed by standard heuristics.', 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', 'It is suggested that an AI inference system should reflect an inference\npolicy that is tailored to the domain of problems to which it is applied -- and\nfurthermore that an inference policy need not conform to any general theory of\nrational inference or induction. We note, for instance, that Bayesian reasoning\nabout the probabilistic characteristics of an inference domain may result in\nthe specification of an nonBayesian procedure for reasoning within the\ninference domain. In this paper, the idea of an inference policy is explored in\nsome detail. To support this exploration, the characteristics of some standard\nand nonstandard inference policies are examined.', 'Inter-domain Gaussian processes (GPs) allow for high flexibility and low\ncomputational cost when performing approximate inference in GP models. They are\nparticularly suitable for modeling data exhibiting global structure but are\nlimited to stationary covariance functions and thus fail to model\nnon-stationary data effectively. We propose Inter-domain Deep Gaussian\nProcesses, an extension of inter-domain shallow GPs that combines the\nadvantages of inter-domain and deep Gaussian processes (DGPs), and demonstrate\nhow to leverage existing approximate inference methods to perform simple and\nscalable approximate inference using inter-domain features in DGPs. We assess\nthe performance of our method on a range of regression tasks and demonstrate\nthat it outperforms inter-domain shallow GPs and conventional DGPs on\nchallenging large-scale real-world datasets exhibiting both global structure as\nwell as a high-degree of non-stationarity.', 'We propose a general framework for modelling and solving deductive games,\nwhere one player selects a secret code and the other player strives to discover\nthis code using a minimal number of allowed experiments that reveal some\npartial information about the code. The framework is implemented in a software\ntool Cobra, and its functionality is demonstrated by producing new results\nabout existing deductive games.', 'Lifted probabilistic inference algorithms exploit regularities in the\nstructure of graphical models to perform inference more efficiently. More\nspecifically, they identify groups of interchangeable variables and perform\ninference once per group, as opposed to once per variable. The groups are\ndefined by means of constraints, so the flexibility of the grouping is\ndetermined by the expressivity of the constraint language. Existing approaches\nfor exact lifted inference use specific languages for (in)equality constraints,\nwhich often have limited expressivity. In this article, we decouple lifted\ninference from the constraint language. We define operators for lifted\ninference in terms of relational algebra operators, so that they operate on the\nsemantic level (the constraints extension) rather than on the syntactic level,\nmaking them language-independent. As a result, lifted inference can be\nperformed using more powerful constraint languages, which provide more\nopportunities for lifting. We empirically demonstrate that this can improve\ninference efficiency by orders of magnitude, allowing exact inference where\nuntil now only approximate inference was feasible.', 'We study the model-checking problem for a quantitative extension of the modal\nmu-calculus on a class of hybrid systems. Qualitative model checking has been\nproved decidable and implemented for several classes of systems, but this is\nnot the case for quantitative questions that arise naturally in this context.\nRecently, quantitative formalisms that subsume classical temporal logics and\nallow the measurement of interesting quantitative phenomena were introduced. We\nshow how a powerful quantitative logic, the quantitative mu-calculus, can be\nmodel checked with arbitrary precision on initialised linear hybrid systems. To\nthis end, we develop new techniques for the discretisation of continuous state\nspaces based on a special class of strategies in model-checking games and\npresent a reduction to a class of counter parity games.', ""Substructural logics naturally support a quantitative interpretation of\nformulas, as they are seen as consumable resources. Distances are the\nquantitative counterpart of equivalence relations: they measure how much two\nobjects are similar, rather than just saying whether they are equivalent or\nnot. Hence, they provide the natural choice for modelling equality in a\nsubstructural setting. In this paper, we develop this idea, using the\ncategorical language of Lawvere's doctrines. We work in a minimal fragment of\nLinear Logic enriched by graded modalities, which are needed to write a\nresource sensitive substitution rule for equality, enabling its quantitative\ninterpretation as a distance. We introduce both a deductive calculus and the\nnotion of Lipschitz doctrine to give it a sound and complete categorical\nsemantics. The study of 2-categorical properties of Lipschitz doctrines\nprovides us with a universal construction, which generates examples based for\ninstance on metric spaces and quantitative realisability. Finally, we show how\nto smoothly extend our results to richer substructural logics, up to full\nLinear Logic with quantifiers."", 'Probabilistic logic programs are logic programs in which some of the facts\nare annotated with probabilities. This paper investigates how classical\ninference and learning tasks known from the graphical model community can be\ntackled for probabilistic logic programs. Several such tasks such as computing\nthe marginals given evidence and learning from (partial) interpretations have\nnot really been addressed for probabilistic logic programs before.\n  The first contribution of this paper is a suite of efficient algorithms for\nvarious inference tasks. It is based on a conversion of the program and the\nqueries and evidence to a weighted Boolean formula. This allows us to reduce\nthe inference tasks to well-studied tasks such as weighted model counting,\nwhich can be solved using state-of-the-art methods known from the graphical\nmodel and knowledge compilation literature. The second contribution is an\nalgorithm for parameter estimation in the learning from interpretations\nsetting. The algorithm employs Expectation Maximization, and is built on top of\nthe developed inference algorithms.\n  The proposed approach is experimentally evaluated. The results show that the\ninference algorithms improve upon the state-of-the-art in probabilistic logic\nprogramming and that it is indeed possible to learn the parameters of a\nprobabilistic logic program from interpretations.', 'The language of epistemic specifications and epistemic logic programs extends\ndisjunctive logic programs under the stable model semantics with modal\nconstructs called subjective literals. Using subjective literals, it is\npossible to check whether a regular literal is true in every or some stable\nmodels of the program, those models, in this context also called \\emph{belief\nsets}, being collected in a set called world view. This allows for\nrepresenting, within the language, whether some proposition should be\nunderstood accordingly to the open or the closed world assumption. Several\nattempts for capturing the intuitions underlying the language by means of a\nformal semantics were given, resulting in a multitude of proposals that makes\nit difficult to understand the current state of the art. In this paper, we\nprovide an overview of the inception of the field and the knowledge\nrepresentation and reasoning tasks it is suitable for. We also provide a\ndetailed analysis of properties of proposed semantics, and an outlook of\nchallenges to be tackled by future research in the area. Under consideration in\nTheory and Practice of Logic Programming (TPLP)', 'Model counting is the problem of computing the number of models that satisfy\na given propositional theory. It has recently been applied to solving inference\ntasks in probabilistic logic programming, where the goal is to compute the\nprobability of given queries being true provided a set of mutually independent\nrandom variables, a model (a logic program) and some evidence. The core of\nsolving this inference task involves translating the logic program to a\npropositional theory and using a model counter. In this paper, we show that for\nsome problems that involve inductive definitions like reachability in a graph,\nthe translation of logic programs to SAT can be expensive for the purpose of\nsolving inference tasks. For such problems, direct implementation of stable\nmodel semantics allows for more efficient solving. We present two\nimplementation techniques, based on unfounded set detection, that extend a\npropositional model counter to a stable model counter. Our experiments show\nthat for particular problems, our approach can outperform a state-of-the-art\nprobabilistic logic programming solver by several orders of magnitude in terms\nof running time and space requirements, and can solve instances of\nsignificantly larger sizes on which the current solver runs out of time or\nmemory.', 'A modal logic that is strong enough to fully characterize the behavior of a\nsystem is called expressive. Recently, with the growing diversity of systems to\nbe reasoned about (probabilistic, cyber-physical, etc.), the focus shifted to\nquantitative settings which resulted in a number of expressivity results for\nquantitative logics and behavioral metrics. Each of these quantitative\nexpressivity results uses a tailor-made argument; distilling the essence of\nthese arguments is non-trivial, yet important to support the design of\nexpressive modal logics for new quantitative settings. In this paper, we\npresent the first categorical framework for deriving quantitative expressivity\nresults, based on the new notion of approximating family. A key ingredient is\nthe codensity lifting -- a uniform observation-centric construction of various\nbisimilarity-like notions such as bisimulation metrics. We show that several\nrecent quantitative expressivity results (e.g. by K\\""onig et al. and by\nFijalkow et al.) are accommodated in our framework; a new expressivity result\nis derived, too, for what we call bisimulation uniformity.', 'We present a formal logic for quantitative reasoning about security\nproperties of network protocols. The system allows us to derive concrete\nsecurity bounds that can be used to choose key lengths and other security\nparameters. We provide axioms for reasoning about digital signatures and random\nnonces, with security properties based on the concrete security of signature\nschemes and pseudorandom number generators (PRG). The formal logic supports\nfirst-order reasoning and reasoning about protocol invariants, taking concrete\nsecurity bounds into account. Proofs constructed in our logic also provide\nconventional asymptotic security guarantees because of the way that concrete\nbounds accumulate in proofs. As an illustrative example, we use the formal\nlogic to prove an authentication property with concrete bounds of a\nsignature-based challenge-response protocol.', 'This paper introduces a new technique for quantifying the approximation error\nof a broad class of probabilistic inference programs, including ones based on\nboth variational and Monte Carlo approaches. The key idea is to derive a\nsubjective bound on the symmetrized KL divergence between the distribution\nachieved by an approximate inference program and its true target distribution.\nThe bound\'s validity (and subjectivity) rests on the accuracy of two auxiliary\nprobabilistic programs: (i) a ""reference"" inference program that defines a gold\nstandard of accuracy and (ii) a ""meta-inference"" program that answers the\nquestion ""what internal random choices did the original approximate inference\nprogram probably make given that it produced a particular result?"" The paper\nincludes empirical results on inference problems drawn from linear regression,\nDirichlet process mixture modeling, HMMs, and Bayesian networks. The\nexperiments show that the technique is robust to the quality of the reference\ninference program and that it can detect implementation bugs that are not\napparent from predictive performance.', 'In this paper we study the uses and the semantics of non-monotonic negation\nin probabilistic deductive data bases. Based on the stable semantics for\nclassical logic programming, we introduce the notion of stable formula,\nfunctions. We show that stable formula, functions are minimal fixpoints of\noperators associated with probabilistic deductive databases with negation.\nFurthermore, since a. probabilistic deductive database may not necessarily have\na stable formula function, we provide a stable class semantics for such\ndatabases. Finally, we demonstrate that the proposed semantics can handle\ndefault reasoning naturally in the context of probabilistic deduction.', 'This work presents how persistent predicates have been included in the\nin-memory deductive system DES by relying on external SQL database management\nsystems. We introduce how persistence is supported from a user-point of view\nand the possible applications the system opens up, as the deductive expressive\npower is projected to relational databases. Also, we describe how it is\npossible to intermix computations of the deductive engine and the external\ndatabase, explaining its implementation and some optimizations. Finally, a\nperformance analysis is undertaken, comparing the system with current\nrelational database systems.', 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.', 'Bayesian inference on structured models typically relies on the ability to\ninfer posterior distributions of underlying hidden variables. However,\ninference in implicit models or complex posterior distributions is hard. A\npopular tool for learning implicit models are generative adversarial networks\n(GANs) which learn parameters of generators by fooling discriminators.\nTypically, GANs are considered to be models themselves and are not understood\nin the context of inference. Current techniques rely on inefficient global\ndiscrimination of joint distributions to perform learning, or only consider\ndiscriminating a single output variable. We overcome these limitations by\ntreating GANs as a basis for likelihood-free inference in generative models and\ngeneralize them to Bayesian posterior inference over factor graphs. We propose\nlocal learning rules based on message passing minimizing a global divergence\ncriterion involving cooperating local adversaries used to sidestep explicit\nlikelihood evaluations. This allows us to compose models and yields a unified\ninference and learning framework for adversarial learning. Our framework treats\nmodel specification and inference separately and facilitates richly structured\nmodels within the family of Directed Acyclic Graphs, including components such\nas intractable likelihoods, non-differentiable models, simulators and generally\ncumbersome models. A key result of our treatment is the insight that Bayesian\ninference on structured models can be performed only with sampling and\ndiscrimination when using nonparametric variational families, without access to\nexplicit distributions. As a side-result, we discuss the link to likelihood\nmaximization. These approaches hold promise to be useful in the toolbox of\nprobabilistic modelers and enrich the gamut of current probabilistic\nprogramming applications.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I need a literature review of the clinician system for a project I am planning to do. I am specifically interested in the application of reinforcement learning (RL) in the clinician system and identify the benefits of RL for the healthcare research community. I hope the literature review can contain detailed exploration of the findings from the literature, including an explanation of the clinician system, how intensive care data can be utilized, and how reinforcement learning can help to make treatment recommendations. I hope the literature review also compares the differences of different RL algorithms used in the clinician system field, and finds out the one that is most efficient under large amounts of data.","['The rapid increase in the percentage of chronic disease patients along with\nthe recent pandemic pose immediate threats on healthcare expenditure and\nelevate causes of death. This calls for transforming healthcare systems away\nfrom one-on-one patient treatment into intelligent health systems, to improve\nservices, access and scalability, while reducing costs. Reinforcement Learning\n(RL) has witnessed an intrinsic breakthrough in solving a variety of complex\nproblems for diverse applications and services. Thus, we conduct in this paper\na comprehensive survey of the recent models and techniques of RL that have been\ndeveloped/used for supporting Intelligent-healthcare (I-health) systems. This\npaper can guide the readers to deeply understand the state-of-the-art regarding\nthe use of RL in the context of I-health. Specifically, we first present an\noverview for the I-health systems challenges, architecture, and how RL can\nbenefit these systems. We then review the background and mathematical modeling\nof different RL, Deep RL (DRL), and multi-agent RL models. After that, we\nprovide a deep literature review for the applications of RL in I-health\nsystems. In particular, three main areas have been tackled, i.e., edge\nintelligence, smart core network, and dynamic treatment regimes. Finally, we\nhighlight emerging challenges and outline future research directions in driving\nthe future success of RL in I-health systems, which opens the door for\nexploring some interesting and unsolved problems.', 'Dynamic treatment recommendation systems based on large-scale electronic\nhealth records (EHRs) become a key to successfully improve practical clinical\noutcomes. Prior relevant studies recommend treatments either use supervised\nlearning (e.g. matching the indicator signal which denotes doctor\nprescriptions), or reinforcement learning (e.g. maximizing evaluation signal\nwhich indicates cumulative reward from survival rates). However, none of these\nstudies have considered to combine the benefits of supervised learning and\nreinforcement learning. In this paper, we propose Supervised Reinforcement\nLearning with Recurrent Neural Network (SRL-RNN), which fuses them into a\nsynergistic learning framework. Specifically, SRL-RNN applies an off-policy\nactor-critic framework to handle complex relations among multiple medications,\ndiseases and individual characteristics. The ""actor"" in the framework is\nadjusted by both the indicator signal and evaluation signal to ensure effective\nprescription and low mortality. RNN is further utilized to solve the\nPartially-Observed Markov Decision Process (POMDP) problem due to the lack of\nfully observed states in real world applications. Experiments on the publicly\nreal-world dataset, i.e., MIMIC-3, illustrate that our model can reduce the\nestimated mortality, while providing promising accuracy in matching doctors\'\nprescriptions.', ""Although reinforcement learning (RL) has tremendous success in many fields,\napplying RL to real-world settings such as healthcare is challenging when the\nreward is hard to specify and no exploration is allowed. In this work, we focus\non recovering clinicians' rewards in treating patients. We incorporate the\nwhat-if reasoning to explain the clinician's treatments based on their\npotential future outcomes. We use generalized additive models (GAMs) - a class\nof accurate, interpretable models - to recover the reward. In both simulation\nand a real-world hospital dataset, we show our model outperforms baselines.\nFinally, our model's explanations match several clinical guidelines when\ntreating patients while we found the commonly-used linear model often\ncontradicts them."", 'Reinforcement Learning (RL) has recently been applied to sequential\nestimation and prediction problems identifying and developing hypothetical\ntreatment strategies for septic patients, with a particular focus on offline\nlearning with observational data. In practice, successful RL relies on\ninformative latent states derived from sequential observations to develop\noptimal treatment strategies. To date, how best to construct such states in a\nhealthcare setting is an open question. In this paper, we perform an empirical\nstudy of several information encoding architectures using data from septic\npatients in the MIMIC-III dataset to form representations of a patient state.\nWe evaluate the impact of representation dimension, correlations with\nestablished acuity scores, and the treatment policies derived from them. We\nfind that sequentially formed state representations facilitate effective policy\nlearning in batch settings, validating a more thoughtful approach to\nrepresentation learning that remains faithful to the sequential and partial\nnature of healthcare data.', 'Many healthcare decisions involve navigating through a multitude of treatment\noptions in a sequential and iterative manner to find an optimal treatment\npathway with the goal of an optimal patient outcome. Such optimization problems\nmay be amenable to reinforcement learning. A reinforcement learning agent could\nbe trained to provide treatment recommendations for physicians, acting as a\ndecision support tool. However, a number of difficulties arise when using RL\nbeyond benchmark environments, such as specifying the reward function, choosing\nan appropriate state representation and evaluating the learned policy.', ""As a subfield of machine learning, reinforcement learning (RL) aims at\nempowering one's capabilities in behavioural decision making by using\ninteraction experience with the world and an evaluative feedback. Unlike\ntraditional supervised learning methods that usually rely on one-shot,\nexhaustive and supervised reward signals, RL tackles with sequential decision\nmaking problems with sampled, evaluative and delayed feedback simultaneously.\nSuch distinctive features make RL technique a suitable candidate for developing\npowerful solutions in a variety of healthcare domains, where diagnosing\ndecisions or treatment regimes are usually characterized by a prolonged and\nsequential procedure. This survey discusses the broad applications of RL\ntechniques in healthcare domains, in order to provide the research community\nwith systematic understanding of theoretical foundations, enabling methods and\ntechniques, existing challenges, and new insights of this emerging paradigm. By\nfirst briefly examining theoretical foundations and key techniques in RL\nresearch from efficient and representational directions, we then provide an\noverview of RL applications in healthcare domains ranging from dynamic\ntreatment regimes in chronic diseases and critical care, automated medical\ndiagnosis from both unstructured and structured clinical data, as well as many\nother control or scheduling domains that have infiltrated many aspects of a\nhealthcare system. Finally, we summarize the challenges and open issues in\ncurrent research, and point out some potential solutions and directions for\nfuture research.""]","['Our aim is to establish a framework where reinforcement learning (RL) of\noptimizing interventions retrospectively allows us a regulatory compliant\npathway to prospective clinical testing of the learned policies in a clinical\ndeployment. We focus on infections in intensive care units which are one of the\nmajor causes of death and difficult to treat because of the complex and opaque\npatient dynamics, and the clinically debated, highly-divergent set of\nintervention policies required by each individual patient, yet intensive care\nunits are naturally data rich. In our work, we build on RL approaches in\nhealthcare (""AI Clinicians""), and learn off-policy continuous dosing policy of\npharmaceuticals for sepsis treatment using historical intensive care data under\npartially observable MDPs (POMDPs). POMPDs capture uncertainty in patient state\nbetter by taking in all historical information, yielding an efficient\nrepresentation, which we investigate through ablations. We compensate for the\nlack of exploration in our retrospective data by evaluating each encountered\nstate with a best-first tree search. We mitigate state distributional shift by\noptimizing our policy in the vicinity of the clinicians\' compound policy.\nCrucially, we evaluate our model recommendations using not only conventional\npolicy evaluations but a novel framework that incorporates human experts: a\nmodel-agnostic pre-clinical evaluation method to estimate the accuracy and\nuncertainty of clinician\'s decisions versus our system recommendations when\nconfronted with the same individual patient history (""shadow mode"").', 'Off-policy reinforcement learning enables near-optimal policy from suboptimal\nexperience, thereby provisions opportunity for artificial intelligence\napplications in healthcare. Previous works have mainly framed patient-clinician\ninteractions as Markov decision processes, while true physiological states are\nnot necessarily fully observable from clinical data. We capture this situation\nwith partially observable Markov decision process, in which an agent optimises\nits actions in a belief represented as a distribution of patient states\ninferred from individual history trajectories. A Gaussian mixture model is\nfitted for the observed data. Moreover, we take into account the fact that\nnuance in pharmaceutical dosage could presumably result in significantly\ndifferent effect by modelling a continuous policy through a Gaussian\napproximator directly in the policy space, i.e. the actor. To address the\nchallenge of infinite number of possible belief states which renders exact\nvalue iteration intractable, we evaluate and plan for only every encountered\nbelief, through heuristic search tree by tightly maintaining lower and upper\nbounds of the true value of belief. We further resort to function\napproximations to update value bounds estimation, i.e. the critic, so that the\ntree search can be improved through more compact bounds at the fringe nodes\nthat will be back-propagated to the root. Both actor and critic parameters are\nlearned via gradient-based approaches. Our proposed policy trained from real\nintensive care unit data is capable of dictating dosing on vasopressors and\nintravenous fluids for sepsis patients that lead to the best patient outcomes.', ""This paper presents the first deep reinforcement learning (DRL) framework to\nestimate the optimal Dynamic Treatment Regimes from observational medical data.\nThis framework is more flexible and adaptive for high dimensional action and\nstate spaces than existing reinforcement learning methods to model real-life\ncomplexity in heterogeneous disease progression and treatment choices, with the\ngoal of providing doctor and patients the data-driven personalized decision\nrecommendations. The proposed DRL framework comprises (i) a supervised learning\nstep to predict the most possible expert actions, and (ii) a deep reinforcement\nlearning step to estimate the long-term value function of Dynamic Treatment\nRegimes. Both steps depend on deep neural networks.\n  As a key motivational example, we have implemented the proposed framework on\na data set from the Center for International Bone Marrow Transplant Research\n(CIBMTR) registry database, focusing on the sequence of prevention and\ntreatments for acute and chronic graft versus host disease after\ntransplantation. In the experimental results, we have demonstrated promising\naccuracy in predicting human experts' decisions, as well as the high expected\nreward function in the DRL-based dynamic treatment regimes."", ""A large set of the explainable Artificial Intelligence (XAI) literature is\nemerging on feature relevance techniques to explain a deep neural network (DNN)\noutput or explaining models that ingest image source data. However, assessing\nhow XAI techniques can help understand models beyond classification tasks, e.g.\nfor reinforcement learning (RL), has not been extensively studied. We review\nrecent works in the direction to attain Explainable Reinforcement Learning\n(XRL), a relatively new subfield of Explainable Artificial Intelligence,\nintended to be used in general public applications, with diverse audiences,\nrequiring ethical, responsible and trustable algorithms. In critical situations\nwhere it is essential to justify and explain the agent's behaviour, better\nexplainability and interpretability of RL models could help gain scientific\ninsight on the inner workings of what is still considered a black box. We\nevaluate mainly studies directly linking explainability to RL, and split these\ninto two categories according to the way the explanations are generated:\ntransparent algorithms and post-hoc explainaility. We also review the most\nprominent XAI works from the lenses of how they could potentially enlighten the\nfurther deployment of the latest advances in RL, in the demanding present and\nfuture of everyday problems."", 'This article surveys reinforcement learning approaches in social robotics.\nReinforcement learning is a framework for decision-making problems in which an\nagent interacts through trial-and-error with its environment to discover an\noptimal behavior. Since interaction is a key component in both reinforcement\nlearning and social robotics, it can be a well-suited approach for real-world\ninteractions with physically embodied social robots. The scope of the paper is\nfocused particularly on studies that include social physical robots and\nreal-world human-robot interactions with users. We present a thorough analysis\nof reinforcement learning approaches in social robotics. In addition to a\nsurvey, we categorize existent reinforcement learning approaches based on the\nused method and the design of the reward mechanisms. Moreover, since\ncommunication capability is a prominent feature of social robots, we discuss\nand group the papers based on the communication medium used for reward\nformulation. Considering the importance of designing the reward function, we\nalso provide a categorization of the papers based on the nature of the reward.\nThis categorization includes three major themes: interactive reinforcement\nlearning, intrinsically motivated methods, and task performance-driven methods.\nThe benefits and challenges of reinforcement learning in social robotics,\nevaluation methods of the papers regarding whether or not they use subjective\nand algorithmic measures, a discussion in the view of real-world reinforcement\nlearning challenges and proposed solutions, the points that remain to be\nexplored, including the approaches that have thus far received less attention\nis also given in the paper. Thus, this paper aims to become a starting point\nfor researchers interested in using and applying reinforcement learning methods\nin this particular research field.', 'Healthcare data suffers from both noise and lack of ground truth. The cost of\ndata increases as it is cleaned and annotated in healthcare. Unlike other data\nsets, medical data annotation, which is critical to accurate ground truth,\nrequires medical domain expertise for a better patient outcome. In this work,\nwe report on the use of reinforcement learning to mimic the decision making\nprocess of annotators for medical events, to automate annotation and labelling.\nThe reinforcement agent learns to annotate alarm data based on annotations done\nby an expert. Our method shows promising results on medical alarm data sets. We\ntrained DQN and A2C agents using the data from monitoring devices annotated by\nan expert. Initial results from these RL agents learning the expert annotation\nbehavior are promising. The A2C agent performs better in terms of learning the\nsparse events in a given state, thereby choosing more right actions compared to\nDQN agent. To the best of our knowledge, this is the first reinforcement\nlearning application for the automation of medical events annotation, which has\nfar-reaching practical use.', 'Reinforcement learning (RL) has the potential to significantly improve\nclinical decision making. However, treatment policies learned via RL from\nobservational data are sensitive to subtle choices in study design. We\nhighlight a simple approach, trajectory inspection, to bring clinicians into an\niterative design process for model-based RL studies. We identify where the\nmodel recommends unexpectedly aggressive treatments or expects surprisingly\npositive outcomes from its recommendations. Then, we examine clinical\ntrajectories simulated with the learned model and policy alongside the actual\nhospital course. Applying this approach to recent work on RL for sepsis\nmanagement, we uncover a model bias towards discharge, a preference for high\nvasopressor doses that may be linked to small sample sizes, and clinically\nimplausible expectations of discharge without weaning off vasopressors. We hope\nthat iterations of detecting and addressing the issues unearthed by our method\nwill result in RL policies that inspire more confidence in deployment.', 'Much attention has been devoted recently to the development of machine\nlearning algorithms with the goal of improving treatment policies in\nhealthcare. Reinforcement learning (RL) is a sub-field within machine learning\nthat is concerned with learning how to make sequences of decisions so as to\noptimize long-term effects. Already, RL algorithms have been proposed to\nidentify decision-making strategies for mechanical ventilation, sepsis\nmanagement and treatment of schizophrenia. However, before implementing\ntreatment policies learned by black-box algorithms in high-stakes clinical\ndecision problems, special care must be taken in the evaluation of these\npolicies.\n  In this document, our goal is to expose some of the subtleties associated\nwith evaluating RL algorithms in healthcare. We aim to provide a conceptual\nstarting point for clinical and computational researchers to ask the right\nquestions when designing and evaluating algorithms for new ways of treating\npatients. In the following, we describe how choices about how to summarize a\nhistory, variance of statistical estimators, and confounders in more ad-hoc\nmeasures can result in unreliable, even misleading estimates of the quality of\na treatment policy. We also provide suggestions for mitigating these\neffects---for while there is much promise for mining observational health data\nto uncover better treatment policies, evaluation must be performed\nthoughtfully.', 'From 2017 to 2018 the number of scientific publications found via PubMed\nsearch using the keyword ""Machine Learning"" increased by 46% (4,317 to 6,307).\nThe results of studies involving machine learning, artificial intelligence\n(AI), and big data have captured the attention of healthcare practitioners,\nhealthcare managers, and the public at a time when Western medicine grapples\nwith unmitigated cost increases and public demands for accountability. The\ncomplexity involved in healthcare applications of machine learning and the size\nof the associated data sets has afforded many researchers an uncontested\nopportunity to satisfy these demands with relatively little oversight. In a\nrecent Nature Medicine article, ""The Artificial Intelligence Clinician learns\noptimal treatment strategies for sepsis in intensive care,"" Komorowski and his\ncoauthors propose methods to train an artificial intelligence clinician to\ntreat sepsis patients with vasopressors and IV fluids. In this post, we will\nclosely examine the claims laid out in this paper. In particular, we will study\nthe individual treatment profiles suggested by their AI Clinician to gain\ninsight into how their AI Clinician intends to treat patients on an individual\nlevel.', 'Many real-world systems problems require reasoning about the long term\nconsequences of actions taken to configure and manage the system. These\nproblems with delayed and often sequentially aggregated reward, are often\ninherently reinforcement learning problems and present the opportunity to\nleverage the recent substantial advances in deep reinforcement learning.\nHowever, in some cases, it is not clear why deep reinforcement learning is a\ngood fit for the problem. Sometimes, it does not perform better than the\nstate-of-the-art solutions. And in other cases, random search or greedy\nalgorithms could outperform deep reinforcement learning. In this paper, we\nreview, discuss, and evaluate the recent trends of using deep reinforcement\nlearning in system optimization. We propose a set of essential metrics to guide\nfuture works in evaluating the efficacy of using deep reinforcement learning in\nsystem optimization. Our evaluation includes challenges, the types of problems,\ntheir formulation in the deep reinforcement learning setting, embedding, the\nmodel used, efficiency, and robustness. We conclude with a discussion on open\nchallenges and potential directions for pushing further the integration of\nreinforcement learning in system optimization.', ""A long-term goal of reinforcement learning agents is to be able to perform\ntasks in complex real-world scenarios. The use of external information is one\nway of scaling agents to more complex problems. However, there is a general\nlack of collaboration or interoperability between different approaches using\nexternal information. In this work, while reviewing externally-influenced\nmethods, we propose a conceptual framework and taxonomy for assisted\nreinforcement learning, aimed at fostering collaboration by classifying and\ncomparing various methods that use external information in the learning\nprocess. The proposed taxonomy details the relationship between the external\ninformation source and the learner agent, highlighting the process of\ninformation decomposition, structure, retention, and how it can be used to\ninfluence agent learning. As well as reviewing state-of-the-art methods, we\nidentify current streams of reinforcement learning that use external\ninformation in order to improve the agent's performance and its decision-making\nprocess. These include heuristic reinforcement learning, interactive\nreinforcement learning, learning from demonstration, transfer learning, and\nlearning from multiple sources, among others. These streams of reinforcement\nlearning operate with the shared objective of scaffolding the learner agent.\nLastly, we discuss further possibilities for future work in the field of\nassisted reinforcement learning systems."", ""While reinforcement learning (RL) has proven to be the approach of choice for\ntackling many complex problems, it remains challenging to develop and deploy RL\nagents in real-life scenarios successfully. This paper presents pH-RL\n(personalization in e-Health with RL) a general RL architecture for\npersonalization to bring RL to health practice. pH-RL allows for various levels\nof personalization in health applications and allows for online and batch\nlearning. Furthermore, we provide a general-purpose implementation framework\nthat can be integrated with various healthcare applications. We describe a\nstep-by-step guideline for the successful deployment of RL policies in a mobile\napplication. We implemented our open-source RL architecture and integrated it\nwith the MoodBuster mobile application for mental health to provide messages to\nincrease daily adherence to the online therapeutic modules. We then performed a\ncomprehensive study with human participants over a sustained period. Our\nexperimental results show that the developed policies learn to select\nappropriate actions consistently using only a few days' worth of data.\nFurthermore, we empirically demonstrate the stability of the learned policies\nduring the study."", ""In mobile health (mHealth) smart devices deliver behavioral treatments\nrepeatedly over time to a user with the goal of helping the user adopt and\nmaintain healthy behaviors. Reinforcement learning appears ideal for learning\nhow to optimally make these sequential treatment decisions. However,\nsignificant challenges must be overcome before reinforcement learning can be\neffectively deployed in a mobile healthcare setting. In this work we are\nconcerned with the following challenges: 1) individuals who are in the same\ncontext can exhibit differential response to treatments 2) only a limited\namount of data is available for learning on any one individual, and 3)\nnon-stationary responses to treatment. To address these challenges we\ngeneralize Thompson-Sampling bandit algorithms to develop IntelligentPooling.\nIntelligentPooling learns personalized treatment policies thus addressing\nchallenge one. To address the second challenge, IntelligentPooling updates each\nuser's degree of personalization while making use of available data on other\nusers to speed up learning. Lastly, IntelligentPooling allows responsivity to\nvary as a function of a user's time since beginning treatment, thus addressing\nchallenge three. We show that IntelligentPooling achieves an average of 26%\nlower regret than state-of-the-art. We demonstrate the promise of this approach\nand its ability to learn from even a small group of users in a live clinical\ntrial."", 'In this article, we aim to provide a literature review of different\nformulations and approaches to continual reinforcement learning (RL), also\nknown as lifelong or non-stationary RL. We begin by discussing our perspective\non why RL is a natural fit for studying continual learning. We then provide a\ntaxonomy of different continual RL formulations by mathematically\ncharacterizing two key properties of non-stationarity, namely, the scope and\ndriver non-stationarity. This offers a unified view of various formulations.\nNext, we review and present a taxonomy of continual RL approaches. We go on to\ndiscuss evaluation of continual RL agents, providing an overview of benchmarks\nused in the literature and important metrics for understanding agent\nperformance. Finally, we highlight open problems and challenges in bridging the\ngap between the current state of continual RL and findings in neuroscience.\nWhile still in its early days, the study of continual RL has the promise to\ndevelop better incremental reinforcement learners that can function in\nincreasingly realistic applications where non-stationarity plays a vital role.\nThese include applications such as those in the fields of healthcare,\neducation, logistics, and robotics.', 'Deep reinforcement learning augments the reinforcement learning framework and\nutilizes the powerful representation of deep neural networks. Recent works have\ndemonstrated the remarkable successes of deep reinforcement learning in various\ndomains including finance, medicine, healthcare, video games, robotics, and\ncomputer vision. In this work, we provide a detailed review of recent and\nstate-of-the-art research advances of deep reinforcement learning in computer\nvision. We start with comprehending the theories of deep learning,\nreinforcement learning, and deep reinforcement learning. We then propose a\ncategorization of deep reinforcement learning methodologies and discuss their\nadvantages and limitations. In particular, we divide deep reinforcement\nlearning into seven main categories according to their applications in computer\nvision, i.e. (i)landmark localization (ii) object detection; (iii) object\ntracking; (iv) registration on both 2D image and 3D image volumetric data (v)\nimage segmentation; (vi) videos analysis; and (vii) other applications. Each of\nthese categories is further analyzed with reinforcement learning techniques,\nnetwork design, and performance. Moreover, we provide a comprehensive analysis\nof the existing publicly available datasets and examine source code\navailability. Finally, we present some open issues and discuss future research\ndirections on deep reinforcement learning in computer vision', ""The increasing availability of large collections of electronic health record\n(EHR) data and unprecedented technical advances in deep learning (DL) have\nsparked a surge of research interest in developing DL based clinical decision\nsupport systems for diagnosis, prognosis, and treatment. Despite the\nrecognition of the value of deep learning in healthcare, impediments to further\nadoption in real healthcare settings remain due to the black-box nature of DL.\nTherefore, there is an emerging need for interpretable DL, which allows end\nusers to evaluate the model decision making to know whether to accept or reject\npredictions and recommendations before an action is taken. In this review, we\nfocus on the interpretability of the DL models in healthcare. We start by\nintroducing the methods for interpretability in depth and comprehensively as a\nmethodological reference for future researchers or clinical practitioners in\nthis field. Besides the methods' details, we also include a discussion of\nadvantages and disadvantages of these methods and which scenarios each of them\nis suitable for, so that interested readers can know how to compare and choose\namong them for use. Moreover, we discuss how these methods, originally\ndeveloped for solving general-domain problems, have been adapted and applied to\nhealthcare problems and how they can help physicians better understand these\ndata-driven technologies. Overall, we hope this survey can help researchers and\npractitioners in both artificial intelligence (AI) and clinical fields\nunderstand what methods we have for enhancing the interpretability of their DL\nmodels and choose the optimal one accordingly."", 'Reinforcement learning (RL) has been widely applied to game-playing and\nsurpassed the best human-level performance in many domains, yet there are few\nuse-cases in industrial or commercial settings. We introduce OR-Gym, an\nopen-source library for developing reinforcement learning algorithms to address\noperations research problems. In this paper, we apply reinforcement learning to\nthe knapsack, multi-dimensional bin packing, multi-echelon supply chain, and\nmulti-period asset allocation model problems, as well as benchmark the RL\nsolutions against MILP and heuristic models. These problems are used in\nlogistics, finance, engineering, and are common in many business operation\nsettings. We develop environments based on prototypical models in the\nliterature and implement various optimization and heuristic models in order to\nbenchmark the RL results. By re-framing a series of classic optimization\nproblems as RL tasks, we seek to provide a new tool for the operations research\ncommunity, while also opening those in the RL community to many of the problems\nand challenges in the OR field.', 'Bayesian methods for machine learning have been widely investigated, yielding\nprincipled methods for incorporating prior information into inference\nalgorithms. In this survey, we provide an in-depth review of the role of\nBayesian methods for the reinforcement learning (RL) paradigm. The major\nincentives for incorporating Bayesian reasoning in RL are: 1) it provides an\nelegant approach to action-selection (exploration/exploitation) as a function\nof the uncertainty in learning; and 2) it provides a machinery to incorporate\nprior knowledge into the algorithms. We first discuss models and methods for\nBayesian inference in the simple single-step Bandit model. We then review the\nextensive recent literature on Bayesian methods for model-based RL, where prior\ninformation can be expressed on the parameters of the Markov model. We also\npresent Bayesian methods for model-free RL, where priors are expressed over the\nvalue function or policy class. The objective of the paper is to provide a\ncomprehensive survey on Bayesian RL algorithms and their theoretical and\nempirical properties.', ""Rising costs in recent years of developing new drugs and treatments have led\nto extensive research in optimization techniques in biomolecular design.\nCurrently, the most widely used approach in biomolecular design is directed\nevolution, which is a greedy hill-climbing algorithm that simulates biological\nevolution. In this paper, we propose a new benchmark of applying reinforcement\nlearning to RNA sequence design, in which the objective function is defined to\nbe the free energy in the sequence's secondary structure. In addition to\nexperimenting with the vanilla implementations of each reinforcement learning\nalgorithm from standard libraries, we analyze variants of each algorithm in\nwhich we modify the algorithm's reward function and tune the model's\nhyperparameters. We show results of the ablation analysis that we do for these\nalgorithms, as well as graphs indicating the algorithm's performance across\nbatches and its ability to search the possible space of RNA sequences. We find\nthat our DQN algorithm performs by far the best in this setting, contrasting\nwith, in which PPO performs the best among all tested algorithms. Our results\nshould be of interest to those in the biomolecular design community and should\nserve as a baseline for future experiments involving machine learning in\nmolecule design."", ""Recent efforts in bioinformatics have achieved tremendous progress in the\nmachine reading of biomedical literature, and the assembly of the extracted\nbiochemical interactions into large-scale models such as protein signaling\npathways. However, batch machine reading of literature at today's scale (PubMed\nalone indexes over 1 million papers per year) is unfeasible due to both cost\nand processing overhead. In this work, we introduce a focused reading approach\nto guide the machine reading of biomedical literature towards what literature\nshould be read to answer a biomedical query as efficiently as possible. We\nintroduce a family of algorithms for focused reading, including an intuitive,\nstrong baseline, and a second approach which uses a reinforcement learning (RL)\nframework that learns when to explore (widen the search) or exploit (narrow\nit). We demonstrate that the RL approach is capable of answering more queries\nthan the baseline, while being more efficient, i.e., reading fewer documents."", 'We present Catalyst.RL, an open-source PyTorch framework for reproducible and\nsample efficient reinforcement learning (RL) research. Main features of\nCatalyst.RL include large-scale asynchronous distributed training, efficient\nimplementations of various RL algorithms and auxiliary tricks, such as n-step\nreturns, value distributions, hyperbolic reinforcement learning, etc. To\ndemonstrate the effectiveness of Catalyst.RL, we applied it to a physics-based\nreinforcement learning challenge ""NeurIPS 2019: Learn to Move -- Walk Around""\nwith the objective to build a locomotion controller for a human musculoskeletal\nmodel. The environment is computationally expensive, has a high-dimensional\ncontinuous action space and is stochastic. Our team took the 2nd place,\ncapitalizing on the ability of Catalyst.RL to train high-quality and\nsample-efficient RL agents in only a few hours of training time. The\nimplementation along with experiments is open-sourced so results can be\nreproduced and novel ideas tried out.', 'Designing optimal controllers continues to be challenging as systems are\nbecoming complex and are inherently nonlinear. The principal advantage of\nreinforcement learning (RL) is its ability to learn from the interaction with\nthe environment and provide optimal control strategy. In this paper, RL is\nexplored in the context of control of the benchmark cartpole dynamical system\nwith no prior knowledge of the dynamics. RL algorithms such as\ntemporal-difference, policy gradient actor-critic, and value function\napproximation are compared in this context with the standard LQR solution.\nFurther, we propose a novel approach to integrate RL and swing-up controllers.', 'Sequential decision making, commonly formalized as Markov Decision Process\n(MDP) optimization, is a important challenge in artificial intelligence. Two\nkey approaches to this problem are reinforcement learning (RL) and planning.\nThis paper presents a survey of the integration of both fields, better known as\nmodel-based reinforcement learning. Model-based RL has two main steps. First,\nwe systematically cover approaches to dynamics model learning, including\nchallenges like dealing with stochasticity, uncertainty, partial observability,\nand temporal abstraction. Second, we present a systematic categorization of\nplanning-learning integration, including aspects like: where to start planning,\nwhat budgets to allocate to planning and real data collection, how to plan, and\nhow to integrate planning in the learning and acting loop. After these two\nsections, we also discuss implicit model-based RL as an end-to-end alternative\nfor model learning and planning, and we cover the potential benefits of\nmodel-based RL. Along the way, the survey also draws connections to several\nrelated RL fields, like hierarchical RL and transfer learning. Altogether, the\nsurvey presents a broad conceptual overview of the combination of planning and\nlearning for MDP optimization.', 'Offline reinforcement learning (RL) defines the task of learning from a fixed\nbatch of data. Due to errors in value estimation from out-of-distribution\nactions, most offline RL algorithms take the approach of constraining or\nregularizing the policy with the actions contained in the dataset. Built on\npre-existing RL algorithms, modifications to make an RL algorithm work offline\ncomes at the cost of additional complexity. Offline RL algorithms introduce new\nhyperparameters and often leverage secondary components such as generative\nmodels, while adjusting the underlying RL algorithm. In this paper we aim to\nmake a deep RL algorithm work while making minimal changes. We find that we can\nmatch the performance of state-of-the-art offline RL algorithms by simply\nadding a behavior cloning term to the policy update of an online RL algorithm\nand normalizing the data. The resulting algorithm is a simple to implement and\ntune baseline, while more than halving the overall run time by removing the\nadditional computational overhead of previous methods.', 'With the development of deep representation learning, the domain of\nreinforcement learning (RL) has become a powerful learning framework now\ncapable of learning complex policies in high dimensional environments. This\nreview summarises deep reinforcement learning (DRL) algorithms and provides a\ntaxonomy of automated driving tasks where (D)RL methods have been employed,\nwhile addressing key computational challenges in real world deployment of\nautonomous driving agents. It also delineates adjacent domains such as behavior\ncloning, imitation learning, inverse reinforcement learning that are related\nbut are not classical RL algorithms. The role of simulators in training agents,\nmethods to validate, test and robustify existing solutions in RL are discussed.', 'We discuss deep reinforcement learning in an overview style. We draw a big\npicture, filled with details. We discuss six core elements, six important\nmechanisms, and twelve applications, focusing on contemporary work, and in\nhistorical contexts. We start with background of artificial intelligence,\nmachine learning, deep learning, and reinforcement learning (RL), with\nresources. Next we discuss RL core elements, including value function, policy,\nreward, model, exploration vs. exploitation, and representation. Then we\ndiscuss important mechanisms for RL, including attention and memory,\nunsupervised learning, hierarchical RL, multi-agent RL, relational RL, and\nlearning to learn. After that, we discuss RL applications, including games,\nrobotics, natural language processing (NLP), computer vision, finance, business\nmanagement, healthcare, education, energy, transportation, computer systems,\nand, science, engineering, and art. Finally we summarize briefly, discuss\nchallenges and opportunities, and close with an epilogue.', 'With AlphaGo defeats top human players, reinforcement learning(RL) algorithms\nhave gradually become the code-base of building stronger artificial\nintelligence(AI). The RL algorithm design firstly needs to adapt to the\nspecific environment, so the designed environment guides the rapid and profound\ndevelopment of RL algorithms. However, the existing environments, which can be\ndivided into real world games and customized toy environments, have obvious\nshortcomings. For real world games, it is designed for human entertainment, and\ntoo much difficult for most of RL researchers. For customized toy environments,\nthere is no widely accepted unified evaluation standard for all RL algorithms.\nTherefore, we introduce the first virtual user-friendly environment framework\nfor RL. In this framework, the environment can be easily configured to realize\nall kinds of RL tasks in the mainstream research. Then all the mainstream\nstate-of-the-art(SOTA) RL algorithms can be conveniently evaluated and\ncompared. Therefore, our contributions mainly includes the following aspects:\n1.single configured environment for all classification of SOTA RL algorithms;\n2.combined environment of more than one classification RL algorithms; 3.the\nevaluation standard for all kinds of RL algorithms. With all these efforts, a\npossibility for breeding an AI with capability of general competency in a\nvariety of tasks is provided, and maybe it will open up a new chapter for AI.', 'Reinforcement learning (RL) is a popular paradigm for addressing sequential\ndecision tasks in which the agent has only limited environmental feedback.\nDespite many advances over the past three decades, learning in many domains\nstill requires a large amount of interaction with the environment, which can be\nprohibitively expensive in realistic scenarios. To address this problem,\ntransfer learning has been applied to reinforcement learning such that\nexperience gained in one task can be leveraged when starting to learn the next,\nharder task. More recently, several lines of research have explored how tasks,\nor data samples themselves, can be sequenced into a curriculum for the purpose\nof learning a problem that may otherwise be too difficult to learn from\nscratch. In this article, we present a framework for curriculum learning (CL)\nin reinforcement learning, and use it to survey and classify existing CL\nmethods in terms of their assumptions, capabilities, and goals. Finally, we use\nour framework to find open problems and suggest directions for future RL\ncurriculum learning research.', 'Objectives-Geriatric clinical care is a multidisciplinary assessment designed\nto evaluate older patients (age 65 years and above) functional ability,\nphysical health, and cognitive wellbeing. The majority of these patients suffer\nfrom multiple chronic conditions and require special attention. Recently,\nhospitals utilize various artificial intelligence (AI) systems to improve care\nfor elderly patients. The purpose of this systematic literature review is to\nunderstand the current use of AI systems, particularly machine learning (ML),\nin geriatric clinical care for chronic diseases. Materials and Methods-We\nrestricted our search to eight databases, namely PubMed, WorldCat, MEDLINE,\nProQuest, ScienceDirect, SpringerLink, Wiley, and ERIC, to analyze research\narticles published in English between January 2010 and June 2019. We focused on\nstudies that used ML algorithms in the care of geriatrics patients with chronic\nconditions. Results-We identified 35 eligible studies and classified in three\ngroups-psychological disorder (n=22), eye diseases (n=6), and others (n=7).\nThis review identified the lack of standardized ML evaluation metrics and the\nneed for data governance specific to health care applications. Conclusion- More\nstudies and ML standardization tailored to health care applications are\nrequired to confirm whether ML could aid in improving geriatric clinical care.', 'Reinforcement Learning (RL) has achieved state-of-the-art results in domains\nsuch as robotics and games. We build on this previous work by applying RL\nalgorithms to a selection of canonical online stochastic optimization problems\nwith a range of practical applications: Bin Packing, Newsvendor, and Vehicle\nRouting. While there is a nascent literature that applies RL to these problems,\nthere are no commonly accepted benchmarks which can be used to compare proposed\napproaches rigorously in terms of performance, scale, or generalizability. This\npaper aims to fill that gap. For each problem we apply both standard approaches\nas well as newer RL algorithms and analyze results. In each case, the\nperformance of the trained RL policy is competitive with or superior to the\ncorresponding baselines, while not requiring much in the way of domain\nknowledge. This highlights the potential of RL in real-world dynamic resource\nallocation problems.', 'In this chapter, we will give comprehensive examples of applying RL in\noptimizing the physical layer of wireless communications by defining different\nclass of problems and the possible solutions to handle them. In Section 9.2, we\npresent all the basic theory needed to address a RL problem, i.e. Markov\ndecision process (MDP), Partially observable Markov decision process (POMDP),\nbut also two very important and widely used algorithms for RL, i.e. the\nQ-learning and SARSA algorithms. We also introduce the deep reinforcement\nlearning (DRL) paradigm and the section ends with an introduction to the\nmulti-armed bandits (MAB) framework. Section 9.3 focuses on some toy examples\nto illustrate how the basic concepts of RL are employed in communication\nsystems. We present applications extracted from literature with simplified\nsystem models using similar notation as in Section 9.2 of this Chapter. In\nSection 9.3, we also focus on modeling RL problems, i.e. how action and state\nspaces and rewards are chosen. The Chapter is concluded in Section 9.4 with a\nprospective thought on RL trends and it ends with a review of a broader state\nof the art in Section 9.5.', 'The integration of deep learning to reinforcement learning (RL) has enabled\nRL to perform efficiently in high-dimensional environments. Deep RL methods\nhave been applied to solve many complex real-world problems in recent years.\nHowever, development of a deep RL-based system is challenging because of\nvarious issues such as the selection of a suitable deep RL algorithm, its\nnetwork configuration, training time, training methods, and so on. This paper\nproposes a comprehensive software framework that not only plays a vital role in\ndesigning a connect-the-dots deep RL architecture but also provides a guideline\nto develop a realistic RL application in a short time span. We have designed\nand developed a deep RL-based software framework that strictly ensures\nflexibility, robustness, and scalability. By inheriting the proposed\narchitecture, software managers can foresee any challenges when designing a\ndeep RL-based system. As a result, they can expedite the design process and\nactively control every stage of software development, which is especially\ncritical in agile development environments. To enforce generalization, the\nproposed architecture does not depend on a specific RL algorithm, a network\nconfiguration, the number of agents, or the type of agents. Using our\nframework, software developers can develop and integrate new RL algorithms or\nnew types of agents, and can flexibly change network configuration or the\nnumber of agents.', ""The recent developments in the field of biomedicine have made large volumes\nof biomedical literature available to the medical practitioners. Due to the\nlarge size and lack of efficient searching strategies, medical practitioners\nstruggle to obtain necessary information available in the biomedical\nliterature. Moreover, the most sophisticated search engines of age are not\nintelligent enough to interpret the clinicians' questions. These facts reflect\nthe urgent need of an information retrieval system that accepts the queries\nfrom medical practitioners' in natural language and returns the answers quickly\nand efficiently. In this paper, we present an implementation of a machine\nintelligence based CLINIcal Question Answering system (CLINIQA) to answer\nmedical practitioner's questions. The system was rigorously evaluated on\ndifferent text mining algorithms and the best components for the system were\nselected. The system makes use of Unified Medical Language System for semantic\nanalysis of both questions and medical documents. In addition, the system\nemploys supervised machine learning algorithms for classification of the\ndocuments, identifying the focus of the question and answer selection.\nEffective domain-specific heuristics are designed for answer ranking. The\nperformance evaluation on hundred clinical questions shows the effectiveness of\nour approach."", 'Deep reinforcement learning is the combination of reinforcement learning (RL)\nand deep learning. This field of research has been able to solve a wide range\nof complex decision-making tasks that were previously out of reach for a\nmachine. Thus, deep RL opens up many new applications in domains such as\nhealthcare, robotics, smart grids, finance, and many more. This manuscript\nprovides an introduction to deep reinforcement learning models, algorithms and\ntechniques. Particular focus is on the aspects related to generalization and\nhow deep RL can be used for practical applications. We assume the reader is\nfamiliar with basic machine learning concepts.', 'In recent years deep reinforcement learning (RL) systems have attained\nsuperhuman performance in a number of challenging task domains. However, a\nmajor limitation of such applications is their demand for massive amounts of\ntraining data. A critical present objective is thus to develop deep RL methods\nthat can adapt rapidly to new tasks. In the present work we introduce a novel\napproach to this challenge, which we refer to as deep meta-reinforcement\nlearning. Previous work has shown that recurrent networks can support\nmeta-learning in a fully supervised context. We extend this approach to the RL\nsetting. What emerges is a system that is trained using one RL algorithm, but\nwhose recurrent dynamics implement a second, quite separate RL procedure. This\nsecond, learned RL algorithm can differ from the original one in arbitrary\nways. Importantly, because it is learned, it is configured to exploit structure\nin the training domain. We unpack these points in a series of seven\nproof-of-concept experiments, each of which examines a key aspect of deep\nmeta-RL. We consider prospects for extending and scaling up the approach, and\nalso point out some potentially important implications for neuroscience.', 'While current benchmark reinforcement learning (RL) tasks have been useful to\ndrive progress in the field, they are in many ways poor substitutes for\nlearning with real-world data. By testing increasingly complex RL algorithms on\nlow-complexity simulation environments, we often end up with brittle RL\npolicies that generalize poorly beyond the very specific domain. To combat\nthis, we propose three new families of benchmark RL domains that contain some\nof the complexity of the natural world, while still supporting fast and\nextensive data acquisition. The proposed domains also permit a characterization\nof generalization through fair train/test separation, and easy comparison and\nreplication of results. Through this work, we challenge the RL research\ncommunity to develop more robust algorithms that meet high standards of\nevaluation.', ""Sepsis is the leading cause of mortality in the ICU. It is challenging to\nmanage because individual patients respond differently to treatment. Thus,\ntailoring treatment to the individual patient is essential for the best\noutcomes. In this paper, we take steps toward this goal by applying a\nmixture-of-experts framework to personalize sepsis treatment. The mixture model\nselectively alternates between neighbor-based (kernel) and deep reinforcement\nlearning (DRL) experts depending on patient's current history. On a large\nretrospective cohort, this mixture-based approach outperforms physician, kernel\nonly, and DRL-only experts."", 'Building upon prior research that highlighted the need for standardizing\nenvironments for building control research, and inspired by recently introduced\nchallenges for real life reinforcement learning control, here we propose a\nnon-exhaustive set of nine real world challenges for reinforcement learning\ncontrol in grid-interactive buildings. We argue that research in this area\nshould be expressed in this framework in addition to providing a standardized\nenvironment for repeatability. Advanced controllers such as model predictive\ncontrol and reinforcement learning (RL) control have both advantages and\ndisadvantages that prevent them from being implemented in real world problems.\nComparisons between the two are rare, and often biased. By focusing on the\nchallenges, we can investigate the performance of the controllers under a\nvariety of situations and generate a fair comparison. As a demonstration, we\nimplement the offline learning challenge in CityLearn and study the impact of\ndifferent levels of domain knowledge and complexity of RL algorithms. We show\nthat the sequence of operations utilized in a rule based controller (RBC) used\nfor offline training affects the performance of the RL agents when evaluated on\na set of four energy flexibility metrics. Longer offline learning from an\noptimized RBC leads to improved performance in the long run. RL agents that\nlearn from a simplified RBC risk poorer performance as the offline learning\nperiod increases. We also observe no impact on performance from information\nsharing amongst agents. We call for a more interdisciplinary effort of the\nresearch community to address the real world challenges, and unlock the\npotential of grid-interactive building', 'Clinicians face several significant barriers to search and synthesize\naccurate, succinct, updated, and trustworthy medical information from several\nliterature sources during the practice of medicine and patient care. In this\ntalk, we will be presenting our research behind the development of a Focused\nClinical Search Service, powered by a Healthcare Knowledge Graph, to interpret\nthe query intent behind clinical search queries and retrieve relevant medical\nsnippets from a diverse corpus of medical literature.', 'As a promising tool to navigate in the vast chemical space, artificial\nintelligence (AI) is leveraged for drug design. From the year 2017 to 2021, the\nnumber of applications of several recent AI models (i.e. graph neural network\n(GNN), recurrent neural network (RNN), variation autoencoder (VAE), generative\nadversarial network (GAN), flow and reinforcement learning (RL)) in drug design\nincreases significantly. Many relevant literature reviews exist. However, none\nof them provides an in-depth summary of many applications of the recent AI\nmodels in drug design. To complement the existing literature, this survey\nincludes the theoretical development of the previously mentioned AI models and\ndetailed summaries of 42 recent applications of AI in drug design. Concretely,\n13 of them leverage GNN for molecular property prediction and 29 of them use RL\nand/or deep generative models for molecule generation and optimization. In most\ncases, the focus of the summary is the models, their variants, and\nmodifications for specific tasks in drug design. Moreover, 60 additional\napplications of AI in molecule generation and optimization are briefly\nsummarized in a table. Finally, this survey provides a holistic discussion of\nthe abundant applications so that the tasks, potential solutions, and\nchallenges in AI-based drug design become evident.', 'Reinforcement learning is a learning paradigm for solving sequential\ndecision-making problems. Recent years have witnessed remarkable progress in\nreinforcement learning upon the fast development of deep neural networks. Along\nwith the promising prospects of reinforcement learning in numerous domains such\nas robotics and game-playing, transfer learning has arisen to tackle various\nchallenges faced by reinforcement learning, by transferring knowledge from\nexternal expertise to facilitate the efficiency and effectiveness of the\nlearning process. In this survey, we systematically investigate the recent\nprogress of transfer learning approaches in the context of deep reinforcement\nlearning. Specifically, we provide a framework for categorizing the\nstate-of-the-art transfer learning approaches, under which we analyze their\ngoals, methodologies, compatible reinforcement learning backbones, and\npractical applications. We also draw connections between transfer learning and\nother relevant topics from the reinforcement learning perspective and explore\ntheir potential challenges that await future research progress.', ""Much work aims to explain a model's prediction on a static input. We consider\nexplanations in a temporal setting where a stateful dynamical model produces a\nsequence of risk estimates given an input at each time step. When the estimated\nrisk increases, the goal of the explanation is to attribute the increase to a\nfew relevant inputs from the past. While our formal setup and techniques are\ngeneral, we carry out an in-depth case study in a clinical setting. The goal\nhere is to alert a clinician when a patient's risk of deterioration rises. The\nclinician then has to decide whether to intervene and adjust the treatment.\nGiven a potentially long sequence of new events since she last saw the patient,\na concise explanation helps her to quickly triage the alert. We develop methods\nto lift static attribution techniques to the dynamical setting, where we\nidentify and address challenges specific to dynamics. We then experimentally\nassess the utility of different explanations of clinical alerts through expert\nevaluation."", ""A consequence of the fragmented and siloed healthcare landscape is that\npatient care (and data) is split along multitude of different facilities and\ncomputer systems and enabling interoperability between these systems is hard.\nThe lack interoperability not only hinders continuity of care and burdens\nproviders, but also hinders effective application of Machine Learning (ML)\nalgorithms. Thus, most current ML algorithms, designed to understand patient\ncare and facilitate clinical decision-support, are trained on limited datasets.\nThis approach is analogous to the Newtonian paradigm of Reductionism in which a\nsystem is broken down into elementary components and a description of the whole\nis formed by understanding those components individually. A key limitation of\nthe reductionist approach is that it ignores the component-component\ninteractions and dynamics within the system which are often of prime\nsignificance in understanding the overall behaviour of complex adaptive systems\n(CAS). Healthcare is a CAS.\n  Though the application of ML on health data have shown incremental\nimprovements for clinical decision support, ML has a much a broader potential\nto restructure care delivery as a whole and maximize care value. However, this\nML potential remains largely untapped: primarily due to functional limitations\nof Electronic Health Records (EHR) and the inability to see the healthcare\nsystem as a whole. This viewpoint (i) articulates the healthcare as a complex\nsystem which has a biological and an organizational perspective, (ii) motivates\nwith examples, the need of a system's approach when addressing healthcare\nchallenges via ML and, (iii) emphasizes to unleash EHR functionality - while\nduly respecting all ethical and legal concerns - to reap full benefits of ML."", 'We give an overview of recent exciting achievements of deep reinforcement\nlearning (RL). We discuss six core elements, six important mechanisms, and\ntwelve applications. We start with background of machine learning, deep\nlearning and reinforcement learning. Next we discuss core RL elements,\nincluding value function, in particular, Deep Q-Network (DQN), policy, reward,\nmodel, planning, and exploration. After that, we discuss important mechanisms\nfor RL, including attention and memory, unsupervised learning, transfer\nlearning, multi-agent RL, hierarchical RL, and learning to learn. Then we\ndiscuss various applications of RL, including games, in particular, AlphaGo,\nrobotics, natural language processing, including dialogue systems, machine\ntranslation, and text generation, computer vision, neural architecture design,\nbusiness management, finance, healthcare, Industry 4.0, smart grid, intelligent\ntransportation systems, and computer systems. We mention topics not reviewed\nyet, and list a collection of RL resources. After presenting a brief summary,\nwe close with discussions.\n  Please see Deep Reinforcement Learning, arXiv:1810.06339, for a significant\nupdate.', 'Catastrophic forgetting remains a severe hindrance to the broad application\nof artificial neural networks (ANNs), however, it continues to be a poorly\nunderstood phenomenon. Despite the extensive amount of work on catastrophic\nforgetting, we argue that it is still unclear how exactly the phenomenon should\nbe quantified, and, moreover, to what degree all of the choices we make when\ndesigning learning systems affect the amount of catastrophic forgetting. We use\nvarious testbeds from the reinforcement learning and supervised learning\nliterature to (1) provide evidence that the choice of which modern\ngradient-based optimization algorithm is used to train an ANN has a significant\nimpact on the amount of catastrophic forgetting and show that-surprisingly-in\nmany instances classical algorithms such as vanilla SGD experience less\ncatastrophic forgetting than the more modern algorithms such as Adam. We\nempirically compare four different existing metrics for quantifying\ncatastrophic forgetting and (2) show that the degree to which the learning\nsystems experience catastrophic forgetting is sufficiently sensitive to the\nmetric used that a change from one principled metric to another is enough to\nchange the conclusions of a study dramatically. Our results suggest that a much\nmore rigorous experimental methodology is required when looking at catastrophic\nforgetting. Based on our results, we recommend inter-task forgetting in\nsupervised learning must be measured with both retention and relearning metrics\nconcurrently, and intra-task forgetting in reinforcement learning must-at the\nvery least-be measured with pairwise interference.', 'Deep reinforcement learning is poised to revolutionise the field of AI and\nrepresents a step towards building autonomous systems with a higher level\nunderstanding of the visual world. Currently, deep learning is enabling\nreinforcement learning to scale to problems that were previously intractable,\nsuch as learning to play video games directly from pixels. Deep reinforcement\nlearning algorithms are also applied to robotics, allowing control policies for\nrobots to be learned directly from camera inputs in the real world. In this\nsurvey, we begin with an introduction to the general field of reinforcement\nlearning, then progress to the main streams of value-based and policy-based\nmethods. Our survey will cover central algorithms in deep reinforcement\nlearning, including the deep $Q$-network, trust region policy optimisation, and\nasynchronous advantage actor-critic. In parallel, we highlight the unique\nadvantages of deep neural networks, focusing on visual understanding via\nreinforcement learning. To conclude, we describe several current areas of\nresearch within the field.', 'Motivated by recent advancements in Deep Reinforcement Learning (RL), we have\ndeveloped an RL agent to manage the operation of storage devices in a household\nand is designed to maximize demand-side cost savings. The proposed technique is\ndata-driven, and the RL agent learns from scratch how to efficiently use the\nenergy storage device given variable tariff structures. In most of the studies,\nthe RL agent is considered as a black box, and how the agent has learned is\noften ignored. We explain the learning progression of the RL agent, and the\nstrategies it follows based on the capacity of the storage device.', ""Reinforcement learning (RL) combines a control problem with statistical\nestimation: The system dynamics are not known to the agent, but can be learned\nthrough experience. A recent line of research casts `RL as inference' and\nsuggests a particular framework to generalize the RL problem as probabilistic\ninference. Our paper surfaces a key shortcoming in that approach, and clarifies\nthe sense in which RL can be coherently cast as an inference problem. In\nparticular, an RL agent must consider the effects of its actions upon future\nrewards and observations: The exploration-exploitation tradeoff. In all but the\nmost simple settings, the resulting inference is computationally intractable so\nthat practical RL algorithms must resort to approximation. We demonstrate that\nthe popular `RL as inference' approximation can perform poorly in even very\nbasic problems. However, we show that with a small modification the framework\ndoes yield algorithms that can provably perform well, and we show that the\nresulting algorithm is equivalent to the recently proposed K-learning, which we\nfurther connect with Thompson sampling."", 'We start with a brief introduction to reinforcement learning (RL), about its\nsuccessful stories, basics, an example, issues, the ICML 2019 Workshop on RL\nfor Real Life, how to use it, study material and an outlook. Then we discuss a\nselection of RL applications, including recommender systems, computer systems,\nenergy, finance, healthcare, robotics, and transportation.', 'Rapid online adaptation to changing tasks is an important problem in machine\nlearning and, recently, a focus of meta-reinforcement learning. However,\nreinforcement learning (RL) algorithms struggle in POMDP environments because\nthe state of the system, essential in a RL framework, is not always visible.\nAdditionally, hand-designed meta-RL architectures may not include suitable\ncomputational structures for specific learning problems. The evolution of\nonline learning mechanisms, on the contrary, has the ability to incorporate\nlearning strategies into an agent that can (i) evolve memory when required and\n(ii) optimize adaptation speed to specific online learning problems. In this\npaper, we exploit the highly adaptive nature of neuromodulated neural networks\nto evolve a controller that uses the latent space of an autoencoder in a POMDP.\nThe analysis of the evolved networks reveals the ability of the proposed\nalgorithm to acquire inborn knowledge in a variety of aspects such as the\ndetection of cues that reveal implicit rewards, and the ability to evolve\nlocation neurons that help with navigation. The integration of inborn knowledge\nand online plasticity enabled fast adaptation and better performance in\ncomparison to some non-evolutionary meta-reinforcement learning algorithms. The\nalgorithm proved also to succeed in the 3D gaming environment Malmo Minecraft.', ""With the continuous development of machine learning technology, major\ne-commerce platforms have launched recommendation systems based on it to serve\na large number of customers with different needs more efficiently. Compared\nwith traditional supervised learning, reinforcement learning can better capture\nthe user's state transition in the decision-making process, and consider a\nseries of user actions, not just the static characteristics of the user at a\ncertain moment. In theory, it will have a long-term perspective, producing a\nmore effective recommendation. The special requirements of reinforcement\nlearning for data make it need to rely on an offline virtual system for\ntraining. Our project mainly establishes a virtual user environment for offline\ntraining. At the same time, we tried to improve a reinforcement learning\nalgorithm based on bi-clustering to expand the action space and recommended\npath space of the recommendation agent."", ""This paper surveys the field of deep multiagent reinforcement learning. The\ncombination of deep neural networks with reinforcement learning has gained\nincreased traction in recent years and is slowly shifting the focus from\nsingle-agent to multiagent environments. Dealing with multiple agents is\ninherently more complex as (a) the future rewards depend on multiple players'\njoint actions and (b) the computational complexity increases. We present the\nmost common multiagent problem representations and their main challenges, and\nidentify five research areas that address one or more of these challenges:\ncentralised training and decentralised execution, opponent modelling,\ncommunication, efficient coordination, and reward shaping. We find that many\ncomputational studies rely on unrealistic assumptions or are not generalisable\nto other settings; they struggle to overcome the curse of dimensionality or\nnonstationarity. Approaches from psychology and sociology capture promising\nrelevant behaviours, such as communication and coordination, to help agents\nachieve better performance in multiagent settings. We suggest that, for\nmultiagent reinforcement learning to be successful, future research should\naddress these challenges with an interdisciplinary approach to open up new\npossibilities in multiagent reinforcement learning."", 'In constrained reinforcement learning (RL), a learning agent seeks to not\nonly optimize the overall reward but also satisfy the additional safety,\ndiversity, or budget constraints. Consequently, existing constrained RL\nsolutions require several new algorithmic ingredients that are notably\ndifferent from standard RL. On the other hand, reward-free RL is independently\ndeveloped in the unconstrained literature, which learns the transition dynamics\nwithout using the reward information, and thus naturally capable of addressing\nRL with multiple objectives under the common dynamics. This paper bridges\nreward-free RL and constrained RL. Particularly, we propose a simple\nmeta-algorithm such that given any reward-free RL oracle, the approachability\nand constrained RL problems can be directly solved with negligible overheads in\nsample complexity. Utilizing the existing reward-free RL solvers, our framework\nprovides sharp sample complexity results for constrained RL in the tabular MDP\nsetting, matching the best existing results up to a factor of horizon\ndependence; our framework directly extends to a setting of tabular two-player\nMarkov games, and gives a new result for constrained RL with linear function\napproximation.', 'In recent years some researchers have explored the use of reinforcement\nlearning (RL) algorithms as key components in the solution of various natural\nlanguage processing tasks. For instance, some of these algorithms leveraging\ndeep neural learning have found their way into conversational systems. This\npaper reviews the state of the art of RL methods for their possible use for\ndifferent problems of natural language processing, focusing primarily on\nconversational systems, mainly due to their growing relevance. We provide\ndetailed descriptions of the problems as well as discussions of why RL is\nwell-suited to solve them. Also, we analyze the advantages and limitations of\nthese methods. Finally, we elaborate on promising research directions in\nnatural language processing that might benefit from reinforcement learning.', ""With increasing and extensive use of electronic health records, clinicians\nare often under time pressure when they need to retrieve important information\nefficiently among large amounts of patients' health records in clinics. While a\nsearch function can be a useful alternative to browsing through a patient's\nrecord, it is cumbersome for clinicians to search repeatedly for the same or\nsimilar information on similar patients. Under such circumstances, there is a\ncritical need to build effective recommender systems that can generate accurate\nsearch term recommendations for clinicians. In this manuscript, we developed a\nhybrid collaborative filtering model using patients' encounter and search term\ninformation to recommend the next search terms for clinicians to retrieve\nimportant information fast in clinics. For each patient, the model will\nrecommend terms that either have high co-occurrence frequencies with his/her\nmost recent ICD codes or are highly relevant to the most recent search terms on\nthis patient. We have conducted comprehensive experiments to evaluate the\nproposed model, and the experimental results demonstrate that our model can\noutperform all the state-of-the-art baseline methods for top-N search term\nrecommendation on different datasets."", 'Deep reinforcement learning is revolutionizing the artificial intelligence\nfield. Currently, it serves as a good starting point for constructing\nintelligent autonomous systems which offer a better knowledge of the visual\nworld. It is possible to scale deep reinforcement learning with the use of deep\nlearning and do amazing tasks such as use of pixels in playing video games. In\nthis paper, key concepts of deep reinforcement learning including reward\nfunction, differences between reinforcement learning and supervised learning\nand models for implementation of reinforcement are discussed. Key challenges\nrelated to the implementation of reinforcement learning in conversational AI\ndomain are identified as well as discussed in detail. Various conversational\nmodels which are based on deep reinforcement learning (as well as deep\nlearning) are also discussed. In summary, this paper discusses key aspects of\ndeep reinforcement learning which are crucial for designing an efficient\nconversational AI.', 'This article is an introduction to machine learning for financial\nforecasting, planning and analysis (FP\\&A). Machine learning appears well\nsuited to support FP\\&A with the highly automated extraction of information\nfrom large amounts of data. However, because most traditional machine learning\ntechniques focus on forecasting (prediction), we discuss the particular care\nthat must be taken to avoid the pitfalls of using them for planning and\nresource allocation (causal inference). While the naive application of machine\nlearning usually fails in this context, the recently developed double machine\nlearning framework can address causal questions of interest. We review the\ncurrent literature on machine learning in FP\\&A and illustrate in a simulation\nstudy how machine learning can be used for both forecasting and planning. We\nalso investigate how forecasting and planning improve as the number of data\npoints increases.', 'Automatic machine learning (AutoML) is an area of research aimed at\nautomating machine learning (ML) activities that currently require human\nexperts. One of the most challenging tasks in this field is the automatic\ngeneration of end-to-end ML pipelines: combining multiple types of ML\nalgorithms into a single architecture used for end-to-end analysis of\npreviously-unseen data. This task has two challenging aspects: the first is the\nneed to explore a large search space of algorithms and pipeline architectures.\nThe second challenge is the computational cost of training and evaluating\nmultiple pipelines. In this study we present DeepLine, a reinforcement learning\nbased approach for automatic pipeline generation. Our proposed approach\nutilizes an efficient representation of the search space and leverages past\nknowledge gained from previously-analyzed datasets to make the problem more\ntractable. Additionally, we propose a novel hierarchical-actions algorithm that\nserves as a plugin, mediating the environment-agent interaction in deep\nreinforcement learning problems. The plugin significantly speeds up the\ntraining process of our model. Evaluation on 56 datasets shows that DeepLine\noutperforms state-of-the-art approaches both in accuracy and in computational\ncost.', 'To achieve general artificial intelligence, reinforcement learning (RL)\nagents should learn not only to optimize returns for one specific task but also\nto constantly build more complex skills and scaffold their knowledge about the\nworld, without forgetting what has already been learned. In this paper, we\ndiscuss the desired characteristics of environments that can support the\ntraining and evaluation of lifelong reinforcement learning agents, review\nexisting environments from this perspective, and propose recommendations for\ndevising suitable environments in the future.', 'Reinforcement learning (RL) algorithms find applications in inventory\ncontrol, recommender systems, vehicular traffic management, cloud computing and\nrobotics. The real-world complications of many tasks arising in these domains\nmakes them difficult to solve with the basic assumptions underlying classical\nRL algorithms. RL agents in these applications often need to react and adapt to\nchanging operating conditions. A significant part of research on single-agent\nRL techniques focuses on developing algorithms when the underlying assumption\nof stationary environment model is relaxed. This paper provides a survey of RL\nmethods developed for handling dynamically varying environment models. The goal\nof methods not limited by the stationarity assumption is to help autonomous\nagents adapt to varying operating conditions. This is possible either by\nminimizing the rewards lost during learning by RL agent or by finding a\nsuitable policy for the RL agent which leads to efficient operation of the\nunderlying system. A representative collection of these algorithms is discussed\nin detail in this work along with their categorization and their relative\nmerits and demerits. Additionally we also review works which are tailored to\napplication domains. Finally, we discuss future enhancements for this field.', 'An expert classification system having statistical information about the\nprior probabilities of the different classes should be able to use this\nknowledge to reduce the amount of additional information that it must collect,\ne.g., through questions, in order to make a correct classification. This paper\nexamines how best to use such prior information and additional\ninformation-collection opportunities to reduce uncertainty about the class to\nwhich a case belongs, thus minimizing the average cost or effort required to\ncorrectly classify new cases.', 'Recent years have witnessed significant advances in reinforcement learning\n(RL), which has registered great success in solving various sequential\ndecision-making problems in machine learning. Most of the successful RL\napplications, e.g., the games of Go and Poker, robotics, and autonomous\ndriving, involve the participation of more than one single agent, which\nnaturally fall into the realm of multi-agent RL (MARL), a domain with a\nrelatively long history, and has recently re-emerged due to advances in\nsingle-agent RL techniques. Though empirically successful, theoretical\nfoundations for MARL are relatively lacking in the literature. In this chapter,\nwe provide a selective overview of MARL, with focus on algorithms backed by\ntheoretical analysis. More specifically, we review the theoretical results of\nMARL algorithms mainly within two representative frameworks, Markov/stochastic\ngames and extensive-form games, in accordance with the types of tasks they\naddress, i.e., fully cooperative, fully competitive, and a mix of the two. We\nalso introduce several significant but challenging applications of these\nalgorithms. Orthogonal to the existing reviews on MARL, we highlight several\nnew angles and taxonomies of MARL theory, including learning in extensive-form\ngames, decentralized MARL with networked agents, MARL in the mean-field regime,\n(non-)convergence of policy-based methods for learning in games, etc. Some of\nthe new angles extrapolate from our own research endeavors and interests. Our\noverall goal with this chapter is, beyond providing an assessment of the\ncurrent state of the field on the mark, to identify fruitful future research\ndirections on theoretical studies of MARL. We expect this chapter to serve as\ncontinuing stimulus for researchers interested in working on this exciting\nwhile challenging topic.', 'Causal inference methods are widely applied in the fields of medicine,\npolicy, and economics. Central to these applications is the estimation of\ntreatment effects to make decisions. Current methods make binary yes-or-no\ndecisions based on the treatment effect of a single outcome dimension. These\nmethods are unable to capture continuous space treatment policies with a\nmeasure of intensity. They also lack the capacity to consider the complexity of\ntreatment such as matching candidate treatments with the subject. We propose to\nformulate the effectiveness of treatment as a parametrizable model, expanding\nto a multitude of treatment intensities and complexities through the continuous\npolicy treatment function, and the likelihood of matching. Our proposal to\ndecompose treatment effect functions into effectiveness factors presents a\nframework to model a rich space of actions using causal inference. We utilize\ndeep learning to optimize the desired holistic metric space instead of\npredicting single-dimensional treatment counterfactual. This approach employs a\npopulation-wide effectiveness measure and significantly improves the overall\neffectiveness of the model. The performance of our algorithms is. demonstrated\nwith experiments. When using generic continuous space treatments and matching\narchitecture, we observe a 41% improvement upon prior art with\ncost-effectiveness and 68% improvement upon a similar method in the average\ntreatment effect. The algorithms capture subtle variations in treatment space,\nstructures the efficient optimizations techniques, and opens up the arena for\nmany applications.', 'In the artificial intelligence field, learning often corresponds to changing\nthe parameters of a parameterized function. A learning rule is an algorithm or\nmathematical expression that specifies precisely how the parameters should be\nchanged. When creating an artificial intelligence system, we must make two\ndecisions: what representation should be used (i.e., what parameterized\nfunction should be used) and what learning rule should be used to search\nthrough the resulting set of representable functions. Using most learning\nrules, these two decisions are coupled in a subtle (and often unintentional)\nway. That is, using the same learning rule with two different representations\nthat can represent the same sets of functions can result in two different\noutcomes. After arguing that this coupling is undesirable, particularly when\nusing artificial neural networks, we present a method for partially decoupling\nthese two decisions for a broad class of learning rules that span unsupervised\nlearning, reinforcement learning, and supervised learning.', 'Hierarchical reinforcement learning has demonstrated significant success at\nsolving difficult reinforcement learning (RL) tasks. Previous works have\nmotivated the use of hierarchy by appealing to a number of intuitive benefits,\nincluding learning over temporally extended transitions, exploring over\ntemporally extended periods, and training and exploring in a more semantically\nmeaningful action space, among others. However, in fully observed, Markovian\nsettings, it is not immediately clear why hierarchical RL should provide\nbenefits over standard ""shallow"" RL architectures. In this work, we isolate and\nevaluate the claimed benefits of hierarchical RL on a suite of tasks\nencompassing locomotion, navigation, and manipulation. Surprisingly, we find\nthat most of the observed benefits of hierarchy can be attributed to improved\nexploration, as opposed to easier policy learning or imposed hierarchical\nstructures. Given this insight, we present exploration techniques inspired by\nhierarchy that achieve performance competitive with hierarchical RL while at\nthe same time being much simpler to use and implement.', ""Prior work on training generative Visual Dialog models with reinforcement\nlearning(Das et al.) has explored a Qbot-Abot image-guessing game and shown\nthat this 'self-talk' approach can lead to improved performance at the\ndownstream dialog-conditioned image-guessing task. However, this improvement\nsaturates and starts degrading after a few rounds of interaction, and does not\nlead to a better Visual Dialog model. We find that this is due in part to\nrepeated interactions between Qbot and Abot during self-talk, which are not\ninformative with respect to the image. To improve this, we devise a simple\nauxiliary objective that incentivizes Qbot to ask diverse questions, thus\nreducing repetitions and in turn enabling Abot to explore a larger state space\nduring RL ie. be exposed to more visual concepts to talk about, and varied\nquestions to answer. We evaluate our approach via a host of automatic metrics\nand human studies, and demonstrate that it leads to better dialog, ie. dialog\nthat is more diverse (ie. less repetitive), consistent (ie. has fewer\nconflicting exchanges), fluent (ie. more human-like),and detailed, while still\nbeing comparably image-relevant as prior work and ablations."", 'Reinforcement learning is one of the core components in designing an\nartificial intelligent system emphasizing real-time response. Reinforcement\nlearning influences the system to take actions within an arbitrary environment\neither having previous knowledge about the environment model or not. In this\npaper, we present a comprehensive study on Reinforcement Learning focusing on\nvarious dimensions including challenges, the recent development of different\nstate-of-the-art techniques, and future directions. The fundamental objective\nof this paper is to provide a framework for the presentation of available\nmethods of reinforcement learning that is informative enough and simple to\nfollow for the new researchers and academics in this domain considering the\nlatest concerns. First, we illustrated the core techniques of reinforcement\nlearning in an easily understandable and comparable way. Finally, we analyzed\nand depicted the recent developments in reinforcement learning approaches. My\nanalysis pointed out that most of the models focused on tuning policy values\nrather than tuning other things in a particular state of reasoning.', ""The increasing incorporation of Artificial Intelligence in the form of\nautomated systems into decision-making procedures highlights not only the\nimportance of decision theory for automated systems but also the need for these\ndecision procedures to be explainable to the people involved in them.\nTraditional realist accounts of explanation, wherein explanation is a relation\nthat holds (or does not hold) eternally between an explanans and an\nexplanandum, are not adequate to account for the notion of explanation required\nfor artificial decision procedures. We offer an alternative account of\nexplanation as used in the context of automated decision-making that makes\nexplanation an epistemic phenomenon, and one that is dependent on context. This\naccount of explanation better accounts for the way that we talk about, and use,\nexplanations and derived concepts, such as `explanatory power', and also allows\nus to differentiate between reasons or causes on the one hand, which do not\nneed to have an epistemic aspect, and explanations on the other, which do have\nsuch an aspect. Against this theoretical backdrop we then review existing\napproaches to explanation in Artificial Intelligence and Machine Learning, and\nsuggest desiderata which truly explainable decision systems should fulfill."", 'Our world is open-ended, non-stationary, and constantly evolving; thus what\nwe talk about and how we talk about it change over time. This inherent dynamic\nnature of language contrasts with the current static language modelling\nparadigm, which trains and evaluates models on utterances from overlapping time\nperiods. Despite impressive recent progress, we demonstrate that Transformer-XL\nlanguage models perform worse in the realistic setup of predicting future\nutterances from beyond their training period, and that model performance\nbecomes increasingly worse with time. We find that, while increasing model size\nalone -- a key driver behind recent progress -- does not solve this problem,\nhaving models that continually update their knowledge with new information can\nindeed mitigate this performance degradation over time. Hence, given the\ncompilation of ever-larger language modelling datasets, combined with the\ngrowing list of language-model-based NLP applications that require up-to-date\nfactual knowledge about the world, we argue that now is the right time to\nrethink the static way in which we currently train and evaluate our language\nmodels, and develop adaptive language models that can remain up-to-date with\nrespect to our ever-changing and non-stationary world. We publicly release our\ndynamic, streaming language modelling benchmarks for WMT and arXiv to\nfacilitate language model evaluation that takes temporal dynamics into account.', 'Brain science and artificial intelligence have made great progress toward the\nunderstanding and engineering of the human mind. The progress has accelerated\nsignificantly since the turn of the century thanks to new methods for probing\nthe brain (both structure and function), and rapid development in deep learning\nresearch. However, despite these new developments, there are still many open\nquestions, such as how to understand the brain at the system level, and various\nrobustness issues and limitations of deep learning. In this informal essay, I\nwill talk about some of the concepts that are central to brain science and\nartificial intelligence, such as information and memory, and discuss how a\ndifferent view on these concepts can help us move forward, beyond current\nlimits of our understanding in these fields.', ""This paper provides novel formal methods and empirical demonstrations of a\nparticular safety concern in reinforcement learning (RL)-based recommendation\nalgorithms. We call this safety concern `user tampering' -- a phenomenon\nwhereby an RL-based recommender system might manipulate a media user's opinions\nvia its recommendations as part of a policy to increase long-term user\nengagement. We then apply techniques from causal modelling to analyse the\nleading approaches in the literature for implementing scalable RL-based\nrecommenders, and we observe that the current approaches permit user tampering.\nAdditionally, we review the existing mitigation strategies for reward tampering\nproblems and show that they do not transfer well to the user tampering\nphenomenon found in the recommendation context. Furthermore, we provide a\nsimulation study of a media RL-based recommendation problem constrained to the\nrecommendation of political content. We show that a Q-learning algorithm\nconsistently learns to exploit its opportunities to polarise simulated users\nwith its early recommendations in order to have more consistent success with\nlater recommendations catering to that polarisation. This latter contribution\ncalls for urgency in designing safer RL-based recommenders; the former suggests\nthat creating such safe recommenders will require a fundamental shift in design\naway from the approaches we have seen in the recent literature."", 'The potential presented by Artificial Intelligence (AI) for healthcare has\nlong been recognised by the technical community. More recently, this potential\nhas been recognised by policymakers, resulting in considerable public and\nprivate investment in the development of AI for healthcare across the globe.\nDespite this, excepting limited success stories, real-world implementation of\nAI systems into front-line healthcare has been limited. There are numerous\nreasons for this, but a main contributory factor is the lack of internationally\naccepted, or formalised, regulatory standards to assess AI safety and impact\nand effectiveness. This is a well-recognised problem with numerous ongoing\nresearch and policy projects to overcome it. Our intention here is to\ncontribute to this problem-solving effort by seeking to set out a minimally\nviable framework for evaluating the safety, acceptability and efficacy of AI\nsystems for healthcare. We do this by conducting a systematic search across\nScopus, PubMed and Google Scholar to identify all the relevant literature\npublished between January 1970 and November 2020 related to the evaluation of:\noutput performance; efficacy; and real-world use of AI systems, and\nsynthesising the key themes according to the stages of evaluation: pre-clinical\n(theoretical phase); exploratory phase; definitive phase; and post-market\nsurveillance phase (monitoring). The result is a framework to guide AI system\ndevelopers, policymakers, and regulators through a sufficient evaluation of an\nAI system designed for use in healthcare.', 'No comprehensive review of Bayesian networks (BNs) in healthcare has been\npublished in the past, making it difficult to organize the research\ncontributions in the present and identify challenges and neglected areas that\nneed to be addressed in the future. This unique and novel scoping review of BNs\nin healthcare provides an analytical framework for comprehensively\ncharacterizing the domain and its current state. The review shows that: (1) BNs\nin healthcare are not used to their full potential; (2) a generic BN\ndevelopment process is lacking; (3) limitations exists in the way BNs in\nhealthcare are presented in the literature, which impacts understanding,\nconsensus towards systematic methodologies, practice and adoption of BNs; and\n(4) a gap exists between having an accurate BN and a useful BN that impacts\nclinical practice. This review empowers researchers and clinicians with an\nanalytical framework and findings that will enable understanding of the need to\naddress the problems of restricted aims of BNs, ad hoc BN development methods,\nand the lack of BN adoption in practice. To map the way forward, the paper\nproposes future research directions and makes recommendations regarding BN\ndevelopment methods and adoption in practice.', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', ""Question answering (QA) systems provide a way of querying the information\navailable in various formats including, but not limited to, unstructured and\nstructured data in natural languages. It constitutes a considerable part of\nconversational artificial intelligence (AI) which has led to the introduction\nof a special research topic on Conversational Question Answering (CQA), wherein\na system is required to understand the given context and then engages in\nmulti-turn QA to satisfy the user's information needs. Whilst the focus of most\nof the existing research work is subjected to single-turn QA, the field of\nmulti-turn QA has recently grasped attention and prominence owing to the\navailability of large-scale, multi-turn QA datasets and the development of\npre-trained language models. With a good amount of models and research papers\nadding to the literature every year recently, there is a dire need of arranging\nand presenting the related work in a unified manner to streamline future\nresearch. This survey, therefore, is an effort to present a comprehensive\nreview of the state-of-the-art research trends of CQA primarily based on\nreviewed papers from 2016-2021. Our findings show that there has been a trend\nshift from single-turn to multi-turn QA which empowers the field of\nConversational AI from different perspectives. This survey is intended to\nprovide an epitome for the research community with the hope of laying a strong\nfoundation for the field of CQA."", 'We consider the problem of providing users of deep Reinforcement Learning\n(RL) based systems with a better understanding of when their output can be\ntrusted. We offer an explainable artificial intelligence (XAI) framework that\nprovides a three-fold explanation: a graphical depiction of the systems\ngeneralization and performance in the current game state, how well the agent\nwould play in semantically similar environments, and a narrative explanation of\nwhat the graphical information implies. We created a user-interface for our XAI\nframework and evaluated its efficacy via a human-user experiment. The results\ndemonstrate a statistically significant increase in user trust and acceptance\nof the AI system with explanation, versus the AI system without explanation.', 'P-values are widely used in both the social and natural sciences to quantify\nthe statistical significance of observed results. The recent surge of big data\nresearch has made the p-value an even more popular tool to test the\nsignificance of a study. However, substantial literature has been produced\ncritiquing how p-values are used and understood. In this paper we review this\nrecent critical literature, much of which is routed in the life sciences, and\nconsider its implications for social scientific research. We provide a coherent\npicture of what the main criticisms are, and draw together and disambiguate\ncommon themes. In particular, we explain how the False Discovery Rate is\ncalculated, and how this differs from a p-value. We also make explicit the\nBayesian nature of many recent criticisms, a dimension that is often\nunderplayed or ignored. We conclude by identifying practical steps to help\nremediate some of the concerns identified. We recommend that (i) far lower\nsignificance levels are used, such as $0.01$ or $0.001$, and (ii) p-values are\ninterpreted contextually, and situated within both the findings of the\nindividual study and the broader field of inquiry (through, for example,\nmeta-analyses).', ""Campbell-Goodhart's law relates to the causal inference error whereby\ndecision-making agents aim to influence variables which are correlated to their\ngoal objective but do not reliably cause it. This is a well known error in\nEconomics and Political Science but not widely labelled in Artificial\nIntelligence research. Through a simple example, we show how off-the-shelf deep\nReinforcement Learning (RL) algorithms are not necessarily immune to this\ncognitive error. The off-policy learning method is tricked, whilst the\non-policy method is not. The practical implication is that naive application of\nRL to complex real life problems can result in the same types of policy errors\nthat humans make. Great care should be taken around understanding the causal\nmodel that underpins a solution derived from Reinforcement Learning."", 'In this paper we introduce DeepCrawl, a fully-playable Roguelike prototype\nfor iOS and Android in which all agents are controlled by policy networks\ntrained using Deep Reinforcement Learning (DRL). Our aim is to understand\nwhether recent advances in DRL can be used to develop convincing behavioral\nmodels for non-player characters in videogames. We begin with an analysis of\nrequirements that such an AI system should satisfy in order to be practically\napplicable in video game development, and identify the elements of the DRL\nmodel used in the DeepCrawl prototype. The successes and limitations of\nDeepCrawl are documented through a series of playability tests performed on the\nfinal game. We believe that the techniques we propose offer insight into\ninnovative new avenues for the development of behaviors for non-player\ncharacters in video games, as they offer the potential to overcome critical\nissues with', 'In this paper, we study how to learn an appropriate lane changing strategy\nfor autonomous vehicles by using deep reinforcement learning. We show that the\nreward of the system should consider the overall traffic efficiency instead of\nthe travel efficiency of an individual vehicle. In summary, cooperation leads\nto a more harmonic and efficient traffic system rather than competition', 'A new design of an AI assistant that has become increasingly popular is that\nof an ""aggregated assistant"" -- realized as an orchestrated composition of\nseveral individual skills or agents that can each perform atomic tasks. In this\npaper, we will talk about the role of planning in the automated composition of\nsuch assistants and explore how concepts in automated planning can help to\nestablish transparency of the inner workings of the assistant to the end-user.', 'Moores Paradox is a test case for any formal theory of belief. In Knowledge\nand Belief, Hintikka developed a multimodal logic for statements that express\nsentences containing the epistemic notions of knowledge and belief. His account\npurports to offer an explanation of the paradox. In this paper I argue that\nHintikkas interpretation of one of the doxastic operators is philosophically\nproblematic and leads to an unnecessarily strong logical system. I offer a\nweaker alternative that captures in a more accurate way our logical intuitions\nabout the notion of belief without sacrificing the possibility of providing an\nexplanation for problematic cases such as Moores Paradox.', 'This paper presents a plausible reasoning system to illustrate some broad\nissues in knowledge representation: dualities between different reasoning\nforms, the difficulty of unifying complementary reasoning styles, and the\napproximate nature of plausible reasoning. These issues have a common\nunderlying theme: there should be an underlying belief calculus of which the\nmany different reasoning forms are special cases, sometimes approximate. The\nsystem presented allows reasoning about defaults, likelihood, necessity and\npossibility in a manner similar to the earlier work of Adams. The system is\nbased on the belief calculus of subjective Bayesian probability which itself is\nbased on a few simple assumptions about how belief should be manipulated.\nApproximations, semantics, consistency and consequence results are presented\nfor the system. While this puts these often discussed plausible reasoning forms\non a probabilistic footing, useful application to practical problems remains an\nissue.', 'This paper makes a first step towards a logic of learning from experiments.\nFor this, we investigate formal frameworks for modeling the interaction of\ncausal and (qualitative) epistemic reasoning. Crucial for our approach is the\nidea that the notion of an intervention can be used as a formal expression of a\n(real or hypothetical) experiment. In a first step we extend the well-known\ncausal models with a simple Hintikka-style representation of the epistemic\nstate of an agent. In the resulting setting, one can talk not only about the\nknowledge of an agent about the values of variables and how interventions\naffect them, but also about knowledge update. The resulting logic can model\nreasoning about thought experiments. However, it is unable to account for\nlearning from experiments, which is clearly brought out by the fact that it\nvalidates the no learning principle for interventions. Therefore, in a second\nstep, we implement a more complex notion of knowledge that allows an agent to\nobserve (measure) certain variables when an experiment is carried out. This\nextended system does allow for learning from experiments. For all the proposed\nlogical systems, we provide a sound and complete axiomatization.', 'Visualization tools for supervised learning have allowed users to interpret,\nintrospect, and gain intuition for the successes and failures of their models.\nWhile reinforcement learning practitioners ask many of the same questions,\nexisting tools are not applicable to the RL setting. In this work, we describe\nour initial attempt at constructing a prototype of these ideas, through\nidentifying possible features that such a system should encapsulate. Our design\nis motivated by envisioning the system to be a platform on which to experiment\nwith interpretable reinforcement learning.', 'Knowledge graphs in manufacturing and production aim to make production lines\nmore efficient and flexible with higher quality output. This makes knowledge\ngraphs attractive for companies to reach Industry 4.0 goals. However, existing\nresearch in the field is quite preliminary, and more research effort on\nanalyzing how knowledge graphs can be applied in the field of manufacturing and\nproduction is needed. Therefore, we have conducted a systematic literature\nreview as an attempt to characterize the state-of-the-art in this field, i.e.,\nby identifying exiting research and by identifying gaps and opportunities for\nfurther research. To do that, we have focused on finding the primary studies in\nthe existing literature, which were classified and analyzed according to four\ncriteria: bibliometric key facts, research type facets, knowledge graph\ncharacteristics, and application scenarios. Besides, an evaluation of the\nprimary studies has also been carried out to gain deeper insights in terms of\nmethodology, empirical evidence, and relevance. As a result, we can offer a\ncomplete picture of the domain, which includes such interesting aspects as the\nfact that knowledge fusion is currently the main use case for knowledge graphs,\nthat empirical research and industrial application are still missing to a large\nextent, that graph embeddings are not fully exploited, and that technical\nliterature is fast-growing but seems to be still far from its peak.', 'Systematic literature reviews play a vital role in identifying the best\navailable evidence for health and social care policy. The resources required to\nproduce systematic reviews can be significant, and a key to the success of any\nreview is the search strategy used to identify relevant literature. However,\nthe methods used to construct search strategies can be complex, time consuming,\nresource intensive and error prone. In this review, we examine the state of the\nart in resolving complex structured information needs, focusing primarily on\nthe healthcare context. We analyse the literature to identify key challenges\nand issues and explore appropriate solutions and workarounds. From this\nanalysis we propose a way forward to facilitate trust and transparency and to\naid explainability, reproducibility and replicability through a set of key\ndesign principles for tools to support the development of search strategies in\nsystematic literature reviews.', 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', 'Two interpretations about syllogistic statements are described in this paper.\nOne is the so-called set-based interpretation, which assumes that quantified\nstatements and syllogisms talk about quantity-relationships between sets. The\nother one, the so-called conditional interpretation, assumes that quantified\npropositions talk about conditional propositions and how strong are the links\nbetween the antecedent and the consequent. Both interpretations are compared\nattending to three different questions (existential import, singular statements\nand non-proportional quantifiers) from the point of view of their impact on the\nfurther development of this type of reasoning.', 'Currently, knowledge discovery in databases is an essential step to identify\nvalid, novel and useful patterns for decision making. There are many real-world\nscenarios, such as bankruptcy prediction, option pricing or medical diagnosis,\nwhere the classification models to be learned need to fulfil restrictions of\nmonotonicity (i.e. the target class label should not decrease when input\nattributes values increase). For instance, it is rational to assume that a\nhigher debt ratio of a company should never result in a lower level of\nbankruptcy risk. Consequently, there is a growing interest from the data mining\nresearch community concerning monotonic predictive models. This paper aims to\npresent an overview about the literature in the field, analyzing existing\ntechniques and proposing a taxonomy of the algorithms based on the type of\nmodel generated. For each method, we review the quality metrics considered in\nthe evaluation and the different data sets and monotonic problems used in the\nanalysis. In this way, this paper serves as an overview of the research about\nmonotonic classification in specialized literature and can be used as a\nfunctional guide of the field.', 'We evaluate four computational models of explanation in Bayesian networks by\ncomparing model predictions to human judgments. In two experiments, we present\nhuman participants with causal structures for which the models make divergent\npredictions and either solicit the best explanation for an observed event\n(Experiment 1) or have participants rate provided explanations for an observed\nevent (Experiment 2). Across two versions of two causal structures and across\nboth experiments, we find that the Causal Explanation Tree and Most Relevant\nExplanation models provide better fits to human data than either Most Probable\nExplanation or Explanation Tree models. We identify strengths and shortcomings\nof these models and what they can reveal about human explanation. We conclude\nby suggesting the value of pursuing computational and psychological\ninvestigations of explanation in parallel.', ""Human input has enabled autonomous systems to improve their capabilities and\nachieve complex behaviors that are otherwise challenging to generate\nautomatically. Recent work focuses on how robots can use such input - like\ndemonstrations or corrections - to learn intended objectives. These techniques\nassume that the human's desired objective already exists within the robot's\nhypothesis space. In reality, this assumption is often inaccurate: there will\nalways be situations where the person might care about aspects of the task that\nthe robot does not know about. Without this knowledge, the robot cannot infer\nthe correct objective. Hence, when the robot's hypothesis space is\nmisspecified, even methods that keep track of uncertainty over the objective\nfail because they reason about which hypothesis might be correct, and not\nwhether any of the hypotheses are correct. In this paper, we posit that the\nrobot should reason explicitly about how well it can explain human inputs given\nits hypothesis space and use that situational confidence to inform how it\nshould incorporate human input. We demonstrate our method on a 7\ndegree-of-freedom robot manipulator in learning from two important types of\nhuman input: demonstrations of manipulation tasks, and physical corrections\nduring the robot's task execution."", ""There is knowledge. There is belief. And there is tacit agreement.' 'We may\ntalk about objects. We may talk about attributes of the objects. Or we may talk\nboth about objects and their attributes.' This work inspects tacit agreements\non assumptions about the relation between objects and their attributes, and\nstudies a way of expressing them, presenting as the result what we term gradual\nlogic in which the sense of truth gradually shifts. It extends classical logic\ninstances with a new logical connective capturing the object-attribute\nrelation. A formal semantics is presented. Decidability is proved. Para-\nconsistent/epistemic/conditional/intensional/description/combined logics are\ncompared."", 'Multi-agent reinforcement learning offers a way to study how communication\ncould emerge in communities of agents needing to solve specific problems. In\nthis paper, we study the emergence of communication in the negotiation\nenvironment, a semi-cooperative model of agent interaction. We introduce two\ncommunication protocols -- one grounded in the semantics of the game, and one\nwhich is \\textit{a priori} ungrounded and is a form of cheap talk. We show that\nself-interested agents can use the pre-grounded communication channel to\nnegotiate fairly, but are unable to effectively use the ungrounded channel.\nHowever, prosocial agents do learn to use cheap talk to find an optimal\nnegotiating strategy, suggesting that cooperation is necessary for language to\nemerge. We also study communication behaviour in a setting where one agent\ninteracts with agents in a community with different levels of prosociality and\nshow how agent identifiability can aid negotiation.', 'In the era of intelligent biohybrid neurotechnologies for brain repair, new\nfanciful terms are appearing in the scientific dictionary to define what has so\nfar been unimaginable. As the emerging neurotechnologies are becoming\nincreasingly polyhedral and sophisticated, should we talk about evolution and\nrank the intelligence of these devices?', 'We argue that logical semantics might have faltered due to its failure in\ndistinguishing between two fundamentally very different types of concepts:\nontological concepts, that should be types in a strongly-typed ontology, and\nlogical concepts, that are predicates corresponding to properties of and\nrelations between objects of various ontological types. We will then show that\naccounting for these differences amounts to the integration of lexical and\ncompositional semantics in one coherent framework, and to an embedding in our\nlogical semantics of a strongly-typed ontology that reflects our commonsense\nview of the world and the way we talk about it in ordinary language. We will\nshow that in such a framework a number of challenges in natural language\nsemantics can be adequately and systematically treated.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I want to build a healthcare decision support system that will recommend treatment plans for patients. In order to make the prediction process transparent and interpretable, I believe I could use reinforcement learning. Given the patient’s past medical history, I will spawn many processes, in each process an artificial or digital twin for the patient is spawned to mimic the patient’s health trajectories. In each process, the patient will receive many treatments, prescriptions, surgeries, and will experience a rest and recovery period, and eventually some form of complications or disease will terminate the patient, and thus terminate the process as well. I want to model this entire environment dynamic using reinforcement learning, conditioned on the patient's past medical record. I will model each dynamic in the environment as some form of numeric optimization problems where the human body strives to optimize its ability to recover and detect possible illness. I need the entire framework to be effective in casting optimization problems, and efficiently solve numeric planning problems involving obstacles. I will use this modeling approach to determine the best course of treatment for the patient to maximize the patient’s life expectancy.","[""Although reinforcement learning (RL) has tremendous success in many fields,\napplying RL to real-world settings such as healthcare is challenging when the\nreward is hard to specify and no exploration is allowed. In this work, we focus\non recovering clinicians' rewards in treating patients. We incorporate the\nwhat-if reasoning to explain the clinician's treatments based on their\npotential future outcomes. We use generalized additive models (GAMs) - a class\nof accurate, interpretable models - to recover the reward. In both simulation\nand a real-world hospital dataset, we show our model outperforms baselines.\nFinally, our model's explanations match several clinical guidelines when\ntreating patients while we found the commonly-used linear model often\ncontradicts them."", 'The rapid increase in the percentage of chronic disease patients along with\nthe recent pandemic pose immediate threats on healthcare expenditure and\nelevate causes of death. This calls for transforming healthcare systems away\nfrom one-on-one patient treatment into intelligent health systems, to improve\nservices, access and scalability, while reducing costs. Reinforcement Learning\n(RL) has witnessed an intrinsic breakthrough in solving a variety of complex\nproblems for diverse applications and services. Thus, we conduct in this paper\na comprehensive survey of the recent models and techniques of RL that have been\ndeveloped/used for supporting Intelligent-healthcare (I-health) systems. This\npaper can guide the readers to deeply understand the state-of-the-art regarding\nthe use of RL in the context of I-health. Specifically, we first present an\noverview for the I-health systems challenges, architecture, and how RL can\nbenefit these systems. We then review the background and mathematical modeling\nof different RL, Deep RL (DRL), and multi-agent RL models. After that, we\nprovide a deep literature review for the applications of RL in I-health\nsystems. In particular, three main areas have been tackled, i.e., edge\nintelligence, smart core network, and dynamic treatment regimes. Finally, we\nhighlight emerging challenges and outline future research directions in driving\nthe future success of RL in I-health systems, which opens the door for\nexploring some interesting and unsolved problems.', 'Dynamic treatment recommendation systems based on large-scale electronic\nhealth records (EHRs) become a key to successfully improve practical clinical\noutcomes. Prior relevant studies recommend treatments either use supervised\nlearning (e.g. matching the indicator signal which denotes doctor\nprescriptions), or reinforcement learning (e.g. maximizing evaluation signal\nwhich indicates cumulative reward from survival rates). However, none of these\nstudies have considered to combine the benefits of supervised learning and\nreinforcement learning. In this paper, we propose Supervised Reinforcement\nLearning with Recurrent Neural Network (SRL-RNN), which fuses them into a\nsynergistic learning framework. Specifically, SRL-RNN applies an off-policy\nactor-critic framework to handle complex relations among multiple medications,\ndiseases and individual characteristics. The ""actor"" in the framework is\nadjusted by both the indicator signal and evaluation signal to ensure effective\nprescription and low mortality. RNN is further utilized to solve the\nPartially-Observed Markov Decision Process (POMDP) problem due to the lack of\nfully observed states in real world applications. Experiments on the publicly\nreal-world dataset, i.e., MIMIC-3, illustrate that our model can reduce the\nestimated mortality, while providing promising accuracy in matching doctors\'\nprescriptions.', 'Clinical decision making regarding treatments based on personal\ncharacteristics leads to effective health improvements. Machine learning (ML)\nhas been the primary concern of diagnosis support according to comprehensive\npatient information. However, the remaining prominent issue is the development\nof objective treatment processes in clinical situations. This study proposes a\nnovel framework to plan treatment processes in a data-driven manner. A key\npoint of the framework is the evaluation of the ""actionability"" for personal\nhealth improvements by using a surrogate Bayesian model in addition to a\nhigh-performance nonlinear ML model. We first evaluated the framework from the\nviewpoint of its methodology using a synthetic dataset. Subsequently, the\nframework was applied to an actual health checkup dataset comprising data from\n3,132 participants, to improve systolic blood pressure values at the individual\nlevel. We confirmed that the computed treatment processes are actionable and\nconsistent with clinical knowledge for lowering blood pressure. These results\ndemonstrate that our framework could contribute toward decision making in the\nmedical field, providing clinicians with deeper insights.', 'Off-policy reinforcement learning enables near-optimal policy from suboptimal\nexperience, thereby provisions opportunity for artificial intelligence\napplications in healthcare. Previous works have mainly framed patient-clinician\ninteractions as Markov decision processes, while true physiological states are\nnot necessarily fully observable from clinical data. We capture this situation\nwith partially observable Markov decision process, in which an agent optimises\nits actions in a belief represented as a distribution of patient states\ninferred from individual history trajectories. A Gaussian mixture model is\nfitted for the observed data. Moreover, we take into account the fact that\nnuance in pharmaceutical dosage could presumably result in significantly\ndifferent effect by modelling a continuous policy through a Gaussian\napproximator directly in the policy space, i.e. the actor. To address the\nchallenge of infinite number of possible belief states which renders exact\nvalue iteration intractable, we evaluate and plan for only every encountered\nbelief, through heuristic search tree by tightly maintaining lower and upper\nbounds of the true value of belief. We further resort to function\napproximations to update value bounds estimation, i.e. the critic, so that the\ntree search can be improved through more compact bounds at the fringe nodes\nthat will be back-propagated to the root. Both actor and critic parameters are\nlearned via gradient-based approaches. Our proposed policy trained from real\nintensive care unit data is capable of dictating dosing on vasopressors and\nintravenous fluids for sepsis patients that lead to the best patient outcomes.', ""Most of the existing medicine recommendation systems that are mainly based on\nelectronic medical records (EMRs) are significantly assisting doctors to make\nbetter clinical decisions benefiting both patients and caregivers. Even though\nthe growth of EMRs is at a lighting fast speed in the era of big data, content\nlimitations in EMRs restrain the existed recommendation systems to reflect\nrelevant medical facts, such as drug-drug interactions. Many medical knowledge\ngraphs that contain drug-related information, such as DrugBank, may give hope\nfor the recommendation systems. However, the direct use of these knowledge\ngraphs in the systems suffers from robustness caused by the incompleteness of\nthe graphs. To address these challenges, we stand on recent advances in graph\nembedding learning techniques and propose a novel framework, called Safe\nMedicine Recommendation (SMR), in this paper. Specifically, SMR first\nconstructs a high-quality heterogeneous graph by bridging EMRs (MIMIC-III) and\nmedical knowledge graphs (ICD-9 ontology and DrugBank). Then, SMR jointly\nembeds diseases, medicines, patients, and their corresponding relations into a\nshared lower dimensional space. Finally, SMR uses the embeddings to decompose\nthe medicine recommendation into a link prediction process while considering\nthe patient's diagnoses and adverse drug reactions. To our best knowledge, SMR\nis the first to learn embeddings of a patient-disease-medicine graph for\nmedicine recommendation in the world. Extensive experiments on real datasets\nare conducted to evaluate the effectiveness of proposed framework.""]","['In the modern healthcare system, rapidly expanding costs/complexity, the\ngrowing myriad of treatment options, and exploding information streams that\noften do not effectively reach the front lines hinder the ability to choose\noptimal treatment decisions over time. The goal in this paper is to develop a\ngeneral purpose (non-disease-specific) computational/artificial intelligence\n(AI) framework to address these challenges. This serves two potential\nfunctions: 1) a simulation environment for exploring various healthcare\npolicies, payment methodologies, etc., and 2) the basis for clinical artificial\nintelligence - an AI that can think like a doctor. This approach combines\nMarkov decision processes and dynamic decision networks to learn from clinical\ndata and develop complex plans via simulation of alternative sequential\ndecision paths while capturing the sometimes conflicting, sometimes synergistic\ninteractions of various components in the healthcare system. It can operate in\npartially observable environments (in the case of missing observations or data)\nby maintaining belief states about patient health status and functions as an\nonline agent that plans and re-plans. This framework was evaluated using real\npatient data from an electronic health record. Such an AI framework easily\noutperforms the current treatment-as-usual (TAU) case-rate/fee-for-service\nmodels of healthcare (Cost per Unit Change: $189 vs. $497) while obtaining a\n30-35% increase in patient outcomes. Tweaking certain model parameters further\nenhances this advantage, obtaining roughly 50% more improvement for roughly\nhalf the costs. Given careful design and problem formulation, an AI simulation\nframework can approximate optimal decisions even in complex and uncertain\nenvironments. Future work is described that outlines potential lines of\nresearch and integration of machine learning algorithms for personalized\nmedicine.', ""This paper presents the first deep reinforcement learning (DRL) framework to\nestimate the optimal Dynamic Treatment Regimes from observational medical data.\nThis framework is more flexible and adaptive for high dimensional action and\nstate spaces than existing reinforcement learning methods to model real-life\ncomplexity in heterogeneous disease progression and treatment choices, with the\ngoal of providing doctor and patients the data-driven personalized decision\nrecommendations. The proposed DRL framework comprises (i) a supervised learning\nstep to predict the most possible expert actions, and (ii) a deep reinforcement\nlearning step to estimate the long-term value function of Dynamic Treatment\nRegimes. Both steps depend on deep neural networks.\n  As a key motivational example, we have implemented the proposed framework on\na data set from the Center for International Bone Marrow Transplant Research\n(CIBMTR) registry database, focusing on the sequence of prevention and\ntreatments for acute and chronic graft versus host disease after\ntransplantation. In the experimental results, we have demonstrated promising\naccuracy in predicting human experts' decisions, as well as the high expected\nreward function in the DRL-based dynamic treatment regimes."", 'Reinforcement Learning (RL) has recently been applied to sequential\nestimation and prediction problems identifying and developing hypothetical\ntreatment strategies for septic patients, with a particular focus on offline\nlearning with observational data. In practice, successful RL relies on\ninformative latent states derived from sequential observations to develop\noptimal treatment strategies. To date, how best to construct such states in a\nhealthcare setting is an open question. In this paper, we perform an empirical\nstudy of several information encoding architectures using data from septic\npatients in the MIMIC-III dataset to form representations of a patient state.\nWe evaluate the impact of representation dimension, correlations with\nestablished acuity scores, and the treatment policies derived from them. We\nfind that sequentially formed state representations facilitate effective policy\nlearning in batch settings, validating a more thoughtful approach to\nrepresentation learning that remains faithful to the sequential and partial\nnature of healthcare data.', 'Personalized predictive medicine necessitates the modeling of patient illness\nand care processes, which inherently have long-term temporal dependencies.\nHealthcare observations, recorded in electronic medical records, are episodic\nand irregular in time. We introduce DeepCare, an end-to-end deep dynamic neural\nnetwork that reads medical records, stores previous illness history, infers\ncurrent illness states and predicts future medical outcomes. At the data level,\nDeepCare represents care episodes as vectors in space, models patient health\nstate trajectories through explicit memory of historical records. Built on Long\nShort-Term Memory (LSTM), DeepCare introduces time parameterizations to handle\nirregular timed events by moderating the forgetting and consolidation of memory\ncells. DeepCare also incorporates medical interventions that change the course\nof illness and shape future medical risk. Moving up to the health state level,\nhistorical and present health states are then aggregated through multiscale\ntemporal pooling, before passing through a neural network that estimates future\noutcomes. We demonstrate the efficacy of DeepCare for disease progression\nmodeling, intervention recommendation, and future risk prediction. On two\nimportant cohorts with heavy social and economic burden -- diabetes and mental\nhealth -- the results show improved modeling and risk prediction accuracy.', ""As a subfield of machine learning, reinforcement learning (RL) aims at\nempowering one's capabilities in behavioural decision making by using\ninteraction experience with the world and an evaluative feedback. Unlike\ntraditional supervised learning methods that usually rely on one-shot,\nexhaustive and supervised reward signals, RL tackles with sequential decision\nmaking problems with sampled, evaluative and delayed feedback simultaneously.\nSuch distinctive features make RL technique a suitable candidate for developing\npowerful solutions in a variety of healthcare domains, where diagnosing\ndecisions or treatment regimes are usually characterized by a prolonged and\nsequential procedure. This survey discusses the broad applications of RL\ntechniques in healthcare domains, in order to provide the research community\nwith systematic understanding of theoretical foundations, enabling methods and\ntechniques, existing challenges, and new insights of this emerging paradigm. By\nfirst briefly examining theoretical foundations and key techniques in RL\nresearch from efficient and representational directions, we then provide an\noverview of RL applications in healthcare domains ranging from dynamic\ntreatment regimes in chronic diseases and critical care, automated medical\ndiagnosis from both unstructured and structured clinical data, as well as many\nother control or scheduling domains that have infiltrated many aspects of a\nhealthcare system. Finally, we summarize the challenges and open issues in\ncurrent research, and point out some potential solutions and directions for\nfuture research."", ""Management of chronic diseases such as heart failure, diabetes, and chronic\nobstructive pulmonary disease (COPD) is a major problem in health care. A\nstandard approach that the medical community has devised to manage widely\nprevalent chronic diseases such as chronic heart failure (CHF) is to have a\ncommittee of experts develop guidelines that all physicians should follow.\nThese guidelines typically consist of a series of complex rules that make\nrecommendations based on a patient's information. Due to their complexity,\noften the guidelines are either ignored or not complied with at all, which can\nresult in poor medical practices. It is not even clear whether it is humanly\npossible to follow these guidelines due to their length and complexity. In the\ncase of CHF management, the guidelines run nearly 80 pages. In this paper we\ndescribe a physician-advisory system for CHF management that codes the entire\nset of clinical practice guidelines for CHF using answer set programming. Our\napproach is based on developing reasoning templates (that we call knowledge\npatterns) and using these patterns to systemically code the clinical guidelines\nfor CHF as ASP rules. Use of the knowledge patterns greatly facilitates the\ndevelopment of our system. Given a patient's medical information, our system\ngenerates a recommendation for treatment just as a human physician would, using\nthe guidelines. Our system will work even in the presence of incomplete\ninformation. Our work makes two contributions: (i) it shows that highly complex\nguidelines can be successfully coded as ASP rules, and (ii) it develops a\nseries of knowledge patterns that facilitate the coding of knowledge expressed\nin a natural language and that can be used for other application domains. This\npaper is under consideration for acceptance in TPLP."", 'Objective: Electronic medical records (EMRs) contain an amount of medical\nknowledge which can be used for clinical decision support (CDS). Our objective\nis a general system that can extract and represent these knowledge contained in\nEMRs to support three CDS tasks: test recommendation, initial diagnosis, and\ntreatment plan recommendation, with the given condition of one patient.\nMethods: We extracted four kinds of medical entities from records and\nconstructed an EMR-based medical knowledge network (EMKN), in which nodes are\nentities and edges reflect their co-occurrence in a single record. Three\nbipartite subgraphs (bi-graphs) were extracted from the EMKN to support each\ntask. One part of the bi-graph was the given condition (e.g., symptoms), and\nthe other was the condition to be inferred (e.g., diseases). Each bi-graph was\nregarded as a Markov random field to support the inference. Three lazy energy\nfunctions and one parameter-based energy function were proposed, as well as two\nknowledge representation learning-based energy functions, which can provide a\ndistributed representation of medical entities. Three measures were utilized\nfor performance evaluation. Results: On the initial diagnosis task, 80.11% of\nthe test records identified at least one correct disease from top 10\ncandidates. Test and treatment recommendation results were 87.88% and 92.55%,\nrespectively. These results altogether indicate that the proposed system\noutperformed the baseline methods. The distributed representation of medical\nentities does reflect similarity relationships in regards to knowledge level.\nConclusion: Combining EMKN and MRF is an effective approach for general medical\nknowledge representation and inference. Different tasks, however, require\ndesigning their energy functions individually.', 'There is a clear need to involve patients in medical decisions. However,\ncognitive psychological research has highlighted the cognitive limitations of\nhumans with respect to 1. Probabilistic assessment of the patient state and of\npotential outcomes of various decisions, 2. Elicitation of the patient utility\nfunction, and 3. Integration of the probabilistic knowledge and of patient\npreferences to determine the optimal strategy. Therefore, without adequate\ncomputational support, current shared decision models have severe ethical\ndeficiencies. An informed consent model unfairly transfers the responsibility\nto a patient who does not have the necessary knowledge, nor the integration\ncapability. A paternalistic model endows with exaggerated power a physician who\nmight not be aware of the patient preferences, is prone to multiple cognitive\nbiases, and whose computational integration capability is bounded. Recent\nprogress in Artificial Intelligence suggests adding a third agent: a computer,\nin all deliberative medical decisions: Non emergency medical decisions in which\nmore than one alternative exists, the patient preferences can be elicited, the\ntherapeutic alternatives might be influenced by these preferences, medical\nknowledge exists regarding the likelihood of the decision outcomes, and there\nis sufficient decision time. Ethical physicians should exploit computational\ndecision support technologies, neither making the decisions solely on their\nown, nor shirking their duty and shifting the responsibility to patients in the\nname of informed consent. The resulting three way (patient, care provider,\ncomputer) human machine model that we suggest emphasizes the patient\npreferences, the physician knowledge, and the computational integration of both\naspects, does not diminish the physician role, but rather brings out the best\nin human and machine.', 'In this document, we explore in more detail our published work (Komorowski,\nCeli, Badawi, Gordon, & Faisal, 2018) for the benefit of the AI in Healthcare\nresearch community. In the above paper, we developed the AI Clinician system,\nwhich demonstrated how reinforcement learning could be used to make useful\nrecommendations towards optimal treatment decisions from intensive care data.\nSince publication a number of authors have reviewed our work (e.g. Abbasi,\n2018; Bos, Azoulay, & Martin-Loeches, 2019; Saria, 2018). Given the difference\nof our framework to previous work, the fact that we are bridging two very\ndifferent academic communities (intensive care and machine learning) and that\nour work has impact on a number of other areas with more traditional\ncomputer-based approaches (biosignal processing and control, biomedical\nengineering), we are providing here additional details on our recent\npublication.', 'Sepsis is a leading cause of mortality in intensive care units and costs\nhospitals billions annually. Treating a septic patient is highly challenging,\nbecause individual patients respond very differently to medical interventions\nand there is no universally agreed-upon treatment for sepsis. In this work, we\npropose an approach to deduce treatment policies for septic patients by using\ncontinuous state-space models and deep reinforcement learning. Our model learns\nclinically interpretable treatment policies, similar in important aspects to\nthe treatment policies of physicians. The learned policies could be used to aid\nintensive care clinicians in medical decision making and improve the likelihood\nof patient survival.', ""Sepsis is a leading cause of mortality in intensive care units (ICUs) and\ncosts hospitals billions annually. Treating a septic patient is highly\nchallenging, because individual patients respond very differently to medical\ninterventions and there is no universally agreed-upon treatment for sepsis.\nUnderstanding more about a patient's physiological state at a given time could\nhold the key to effective treatment policies. In this work, we propose a new\napproach to deduce optimal treatment policies for septic patients by using\ncontinuous state-space models and deep reinforcement learning. Learning\ntreatment policies over continuous spaces is important, because we retain more\nof the patient's physiological information. Our model is able to learn\nclinically interpretable treatment policies, similar in important aspects to\nthe treatment policies of physicians. Evaluating our algorithm on past ICU\npatient data, we find that our model could reduce patient mortality in the\nhospital by up to 3.6% over observed clinical policies, from a baseline\nmortality of 13.7%. The learned treatment policies could be used to aid\nintensive care clinicians in medical decision making and improve the likelihood\nof patient survival."", 'Finding an effective medical treatment often requires a search by trial and\nerror. Making this search more efficient by minimizing the number of\nunnecessary trials could lower both costs and patient suffering. We formalize\nthis problem as learning a policy for finding a near-optimal treatment in a\nminimum number of trials using a causal inference framework. We give a\nmodel-based dynamic programming algorithm which learns from observational data\nwhile being robust to unmeasured confounding. To reduce time complexity, we\nsuggest a greedy algorithm which bounds the near-optimality constraint. The\nmethods are evaluated on synthetic and real-world healthcare data and compared\nto model-free reinforcement learning. We find that our methods compare\nfavorably to the model-free baseline while offering a more transparent\ntrade-off between search time and treatment efficacy.', 'Decision makers, such as doctors and judges, make crucial decisions such as\nrecommending treatments to patients, and granting bails to defendants on a\ndaily basis. Such decisions typically involve weighting the potential benefits\nof taking an action against the costs involved. In this work, we aim to\nautomate this task of learning \\emph{cost-effective, interpretable and\nactionable treatment regimes}. We formulate this as a problem of learning a\ndecision list -- a sequence of if-then-else rules -- which maps characteristics\nof subjects (eg., diagnostic test results of patients) to treatments. We\npropose a novel objective to construct a decision list which maximizes outcomes\nfor the population, and minimizes overall costs. We model the problem of\nlearning such a list as a Markov Decision Process (MDP) and employ a variant of\nthe Upper Confidence Bound for Trees (UCT) strategy which leverages customized\nchecks for pruning the search space effectively. Experimental results on real\nworld observational data capturing judicial bail decisions and treatment\nrecommendations for asthma patients demonstrate the effectiveness of our\napproach.', ""Clinical decision support tools rooted in machine learning and optimization\ncan provide significant value to healthcare providers, including through better\nmanagement of intensive care units. In particular, it is important that the\npatient discharge task addresses the nuanced trade-off between decreasing a\npatient's length of stay (and associated hospitalization costs) and the risk of\nreadmission or even death following the discharge decision. This work\nintroduces an end-to-end general framework for capturing this trade-off to\nrecommend optimal discharge timing decisions given a patient's electronic\nhealth records. A data-driven approach is used to derive a parsimonious,\ndiscrete state space representation that captures a patient's physiological\ncondition. Based on this model and a given cost function, an infinite-horizon\ndiscounted Markov decision process is formulated and solved numerically to\ncompute an optimal discharge policy, whose value is assessed using off-policy\nevaluation strategies. Extensive numerical experiments are performed to\nvalidate the proposed framework using real-life intensive care unit patient\ndata."", ""In mobile health (mHealth) smart devices deliver behavioral treatments\nrepeatedly over time to a user with the goal of helping the user adopt and\nmaintain healthy behaviors. Reinforcement learning appears ideal for learning\nhow to optimally make these sequential treatment decisions. However,\nsignificant challenges must be overcome before reinforcement learning can be\neffectively deployed in a mobile healthcare setting. In this work we are\nconcerned with the following challenges: 1) individuals who are in the same\ncontext can exhibit differential response to treatments 2) only a limited\namount of data is available for learning on any one individual, and 3)\nnon-stationary responses to treatment. To address these challenges we\ngeneralize Thompson-Sampling bandit algorithms to develop IntelligentPooling.\nIntelligentPooling learns personalized treatment policies thus addressing\nchallenge one. To address the second challenge, IntelligentPooling updates each\nuser's degree of personalization while making use of available data on other\nusers to speed up learning. Lastly, IntelligentPooling allows responsivity to\nvary as a function of a user's time since beginning treatment, thus addressing\nchallenge three. We show that IntelligentPooling achieves an average of 26%\nlower regret than state-of-the-art. We demonstrate the promise of this approach\nand its ability to learn from even a small group of users in a live clinical\ntrial."", 'In both the fields of computer science and medicine there is very strong\ninterest in developing personalized treatment policies for patients who have\nvariable responses to treatments. In particular, I aim to find an optimal\npersonalized treatment policy which is a non-deterministic function of the\npatient specific covariate data that maximizes the expected survival time or\nclinical outcome. I developed an algorithmic framework to solve multistage\ndecision problem with a varying number of stages that are subject to censoring\nin which the ""rewards"" are expected survival times. In specific, I developed a\nnovel Q-learning algorithm that dynamically adjusts for these parameters.\nFurthermore, I found finite upper bounds on the generalized error of the\ntreatment paths constructed by this algorithm. I have also shown that when the\noptimal Q-function is an element of the approximation space, the anticipated\nsurvival times for the treatment regime constructed by the algorithm will\nconverge to the optimal treatment path. I demonstrated the performance of the\nproposed algorithmic framework via simulation studies and through the analysis\nof chronic depression data and a hypothetical clinical trial. The censored\nQ-learning algorithm I developed is more effective than the state of the art\nclinical decision support systems and is able to operate in environments when\nmany covariate parameters may be unobtainable or censored.', ""Automated medical prognosis has gained interest as artificial intelligence\nevolves and the potential for computer-aided medicine becomes evident.\nNevertheless, it is challenging to design an effective system that, given a\npatient's medical history, is able to predict probable future conditions.\nPrevious works, mostly carried out over private datasets, have tackled the\nproblem by using artificial neural network architectures that cannot deal with\nlow-cardinality datasets, or by means of non-generalizable inference\napproaches. We introduce a Deep Learning architecture whose design results from\nan intensive experimental process. The final architecture is based on two\nparallel Minimal Gated Recurrent Unit networks working in bi-directional\nmanner, which was extensively tested with the open-access Mimic-III dataset.\nOur results demonstrate significant improvements in automated medical\nprognosis, as measured with Recall@k. We summarize our experience as a set of\nrelevant insights for the design of Deep Learning architectures. Our work\nimproves the performance of computer-aided medicine and can serve as a guide in\ndesigning artificial neural networks used in prediction tasks."", 'Our aim is to establish a framework where reinforcement learning (RL) of\noptimizing interventions retrospectively allows us a regulatory compliant\npathway to prospective clinical testing of the learned policies in a clinical\ndeployment. We focus on infections in intensive care units which are one of the\nmajor causes of death and difficult to treat because of the complex and opaque\npatient dynamics, and the clinically debated, highly-divergent set of\nintervention policies required by each individual patient, yet intensive care\nunits are naturally data rich. In our work, we build on RL approaches in\nhealthcare (""AI Clinicians""), and learn off-policy continuous dosing policy of\npharmaceuticals for sepsis treatment using historical intensive care data under\npartially observable MDPs (POMDPs). POMPDs capture uncertainty in patient state\nbetter by taking in all historical information, yielding an efficient\nrepresentation, which we investigate through ablations. We compensate for the\nlack of exploration in our retrospective data by evaluating each encountered\nstate with a best-first tree search. We mitigate state distributional shift by\noptimizing our policy in the vicinity of the clinicians\' compound policy.\nCrucially, we evaluate our model recommendations using not only conventional\npolicy evaluations but a novel framework that incorporates human experts: a\nmodel-agnostic pre-clinical evaluation method to estimate the accuracy and\nuncertainty of clinician\'s decisions versus our system recommendations when\nconfronted with the same individual patient history (""shadow mode"").', 'Modern healthcare is ripe for disruption by AI. A game changer would be\nautomatic understanding the latent processes from electronic medical records,\nwhich are being collected for billions of people worldwide. However, these\nhealthcare processes are complicated by the interaction between at least three\ndynamic components: the illness which involves multiple diseases, the care\nwhich involves multiple treatments, and the recording practice which is biased\nand erroneous. Existing methods are inadequate in capturing the dynamic\nstructure of care. We propose Resset, an end-to-end recurrent model that reads\nmedical record and predicts future risk. The model adopts the algebraic view in\nthat discrete medical objects are embedded into continuous vectors lying in the\nsame space. We formulate the problem as modeling sequences of sets, a novel\nsetting that have rarely, if not, been addressed. Within Resset, the bag of\ndiseases recorded at each clinic visit is modeled as function of sets. The same\nhold for the bag of treatments. The interaction between the disease bag and the\ntreatment bag at a visit is modeled in several, one of which as residual of\ndiseases minus the treatments. Finally, the health trajectory, which is a\nsequence of visits, is modeled using a recurrent neural network. We report\nresults on over a hundred thousand hospital visits by patients suffered from\ntwo costly chronic diseases -- diabetes and mental health. Resset shows\npromises in multiple predictive tasks such as readmission prediction,\ntreatments recommendation and diseases progression.', ""While reinforcement learning (RL) has proven to be the approach of choice for\ntackling many complex problems, it remains challenging to develop and deploy RL\nagents in real-life scenarios successfully. This paper presents pH-RL\n(personalization in e-Health with RL) a general RL architecture for\npersonalization to bring RL to health practice. pH-RL allows for various levels\nof personalization in health applications and allows for online and batch\nlearning. Furthermore, we provide a general-purpose implementation framework\nthat can be integrated with various healthcare applications. We describe a\nstep-by-step guideline for the successful deployment of RL policies in a mobile\napplication. We implemented our open-source RL architecture and integrated it\nwith the MoodBuster mobile application for mental health to provide messages to\nincrease daily adherence to the online therapeutic modules. We then performed a\ncomprehensive study with human participants over a sustained period. Our\nexperimental results show that the developed policies learn to select\nappropriate actions consistently using only a few days' worth of data.\nFurthermore, we empirically demonstrate the stability of the learned policies\nduring the study."", ""Choosing the best treatment-plan for each individual patient requires\naccurate forecasts of their outcome trajectories as a function of the\ntreatment, over time. While large observational data sets constitute rich\nsources of information to learn from, they also contain biases as treatments\nare rarely assigned randomly in practice. To provide accurate and unbiased\nforecasts, we introduce the Disentangled Counterfactual Recurrent Network\n(DCRN), a novel sequence-to-sequence architecture that estimates treatment\noutcomes over time by learning representations of patient histories that are\ndisentangled into three separate latent factors: a treatment factor,\ninfluencing only treatment selection; an outcome factor, influencing only the\noutcome; and a confounding factor, influencing both. With an architecture that\nis completely inspired by the causal structure of treatment influence over\ntime, we advance forecast accuracy and disease understanding, as our\narchitecture allows for practitioners to infer which patient features influence\nwhich part in a patient's trajectory, contrasting other approaches in this\ndomain. We demonstrate that DCRN outperforms current state-of-the-art methods\nin forecasting treatment responses, on both real and simulated data."", 'Objectives: Electronic health records (EHRs) are only a first step in\ncapturing and utilizing health-related data - the challenge is turning that\ndata into useful information. Furthermore, EHRs are increasingly likely to\ninclude data relating to patient outcomes, functionality such as clinical\ndecision support, and genetic information as well, and, as such, can be seen as\nrepositories of increasingly valuable information about patients\' health\nconditions and responses to treatment over time. Methods: We describe a case\nstudy of 423 patients treated by Centerstone within Tennessee and Indiana in\nwhich we utilized electronic health record data to generate predictive\nalgorithms of individual patient treatment response. Multiple models were\nconstructed using predictor variables derived from clinical, financial and\ngeographic data. Results: For the 423 patients, 101 deteriorated, 223 improved\nand in 99 there was no change in clinical condition. Based on modeling of\nvarious clinical indicators at baseline, the highest accuracy in predicting\nindividual patient response ranged from 70-72% within the models tested. In\nterms of individual predictors, the Centerstone Assessment of Recovery Level -\nAdult (CARLA) baseline score was most significant in predicting outcome over\ntime (odds ratio 4.1 + 2.27). Other variables with consistently significant\nimpact on outcome included payer, diagnostic category, location and provision\nof case management services. Conclusions: This approach represents a promising\navenue toward reducing the current gap between research and practice across\nhealthcare, developing data-driven clinical decision support based on\nreal-world populations, and serving as a component of embedded clinical\nartificial intelligences that ""learn"" over time.', 'Many healthcare decisions involve navigating through a multitude of treatment\noptions in a sequential and iterative manner to find an optimal treatment\npathway with the goal of an optimal patient outcome. Such optimization problems\nmay be amenable to reinforcement learning. A reinforcement learning agent could\nbe trained to provide treatment recommendations for physicians, acting as a\ndecision support tool. However, a number of difficulties arise when using RL\nbeyond benchmark environments, such as specifying the reward function, choosing\nan appropriate state representation and evaluating the learned policy.', ""Automatically reasoning with conflicting generic clinical guidelines is a\nburning issue in patient-centric medical reasoning where patient-specific\nconditions and goals need to be taken into account. It is even more challenging\nin the presence of preferences such as patient's wishes and clinician's\npriorities over goals. We advance a structured argumentation formalism for\nreasoning with conflicting clinical guidelines, patient-specific information\nand preferences. Our formalism integrates assumption-based reasoning and\ngoal-driven selection among reasoning outcomes. Specifically, we assume\napplicability of guideline recommendations concerning the generic goal of\npatient well-being, resolve conflicts among recommendations using patient's\nconditions and preferences, and then consider prioritised patient-centered\ngoals to yield non-conflicting, goal-maximising and preference-respecting\nrecommendations. We rely on the state-of-the-art Transition-based Medical\nRecommendation model for representing guideline recommendations and augment it\nwith context given by the patient's conditions, goals, as well as preferences\nover recommendations and goals. We establish desirable properties of our\napproach in terms of sensitivity to recommendation conflicts and patient\ncontext."", 'Much attention has been devoted recently to the development of machine\nlearning algorithms with the goal of improving treatment policies in\nhealthcare. Reinforcement learning (RL) is a sub-field within machine learning\nthat is concerned with learning how to make sequences of decisions so as to\noptimize long-term effects. Already, RL algorithms have been proposed to\nidentify decision-making strategies for mechanical ventilation, sepsis\nmanagement and treatment of schizophrenia. However, before implementing\ntreatment policies learned by black-box algorithms in high-stakes clinical\ndecision problems, special care must be taken in the evaluation of these\npolicies.\n  In this document, our goal is to expose some of the subtleties associated\nwith evaluating RL algorithms in healthcare. We aim to provide a conceptual\nstarting point for clinical and computational researchers to ask the right\nquestions when designing and evaluating algorithms for new ways of treating\npatients. In the following, we describe how choices about how to summarize a\nhistory, variance of statistical estimators, and confounders in more ad-hoc\nmeasures can result in unreliable, even misleading estimates of the quality of\na treatment policy. We also provide suggestions for mitigating these\neffects---for while there is much promise for mining observational health data\nto uncover better treatment policies, evaluation must be performed\nthoughtfully.', 'We consider a reinforcement learning (RL) setting in which the agent\ninteracts with a sequence of episodic MDPs. At the start of each episode the\nagent has access to some side-information or context that determines the\ndynamics of the MDP for that episode. Our setting is motivated by applications\nin healthcare where baseline measurements of a patient at the start of a\ntreatment episode form the context that may provide information about how the\npatient might respond to treatment decisions. We propose algorithms for\nlearning in such Contextual Markov Decision Processes (CMDPs) under an\nassumption that the unobserved MDP parameters vary smoothly with the observed\ncontext. We also give lower and upper PAC bounds under the smoothness\nassumption. Because our lower bound has an exponential dependence on the\ndimension, we consider a tractable linear setting where the context is used to\ncreate linear combinations of a finite set of MDPs. For the linear setting, we\ngive a PAC learning algorithm based on KWIK learning techniques.', 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'Identifying when to give treatments to patients and how to select among\nmultiple treatments over time are important medical problems with a few\nexisting solutions. In this paper, we introduce the Counterfactual Recurrent\nNetwork (CRN), a novel sequence-to-sequence model that leverages the\nincreasingly available patient observational data to estimate treatment effects\nover time and answer such medical questions. To handle the bias from\ntime-varying confounders, covariates affecting the treatment assignment policy\nin the observational data, CRN uses domain adversarial training to build\nbalancing representations of the patient history. At each timestep, CRN\nconstructs a treatment invariant representation which removes the association\nbetween patient history and treatment assignments and thus can be reliably used\nfor making counterfactual predictions. On a simulated model of tumour growth,\nwith varying degree of time-dependent confounding, we show how our model\nachieves lower error in estimating counterfactuals and in choosing the correct\ntreatment and timing of treatment than current state-of-the-art methods.', 'Medical head CT-scan imaging has been successfully combined with deep\nlearning for medical diagnostics of head diseases and lesions[1]. State of the\nart classification models and algorithms for this task usually are based on 3d\nconvolution layers for volumetric data on a supervised learning setting (1\ninput volume, 1 prediction per patient) or 2d convolution layers in a\nsupervised setting (1 input image, 1 prediction per image). However a very\ncommon scenario in developing countries is to have the volume metadata lost due\nmultiple reasons for example formatting conversion in images (for example\n.dicom to jpg), in this scenario the doctor analyses the collection of images\nand then emits a single diagnostic for the patient (with possibly an unfixed\nand variable number of images per patient) , this prevents it from being\npossible to use state of the art 3d models, but also is not possible to convert\nit to a supervised problem in a (1 image,1 diagnostic) setting because\ndifferent angles or positions of the images for a single patient may not\ncontain the disease or lesion. In this study we propose a solution for this\nscenario by combining 2d convolutional[2] models with sequence models which\ngenerate a prediction only after all images have been processed by the model\nfor a given patient \\(i\\), this creates a multi-image to single-diagnostic\nsetting \\(y^i=f(x_1,x_2,..,x_n)\\) where \\(n\\) may be different between\npatients. The experimental results demonstrate that it is possible to get a\nmulti-image to single diagnostic model which mimics human doctor diagnostic\nprocess: evaluate the collection of patient images and then use important\ninformation in memory to decide a single diagnostic for the patient.', 'Healthcare data suffers from both noise and lack of ground truth. The cost of\ndata increases as it is cleaned and annotated in healthcare. Unlike other data\nsets, medical data annotation, which is critical to accurate ground truth,\nrequires medical domain expertise for a better patient outcome. In this work,\nwe report on the use of reinforcement learning to mimic the decision making\nprocess of annotators for medical events, to automate annotation and labelling.\nThe reinforcement agent learns to annotate alarm data based on annotations done\nby an expert. Our method shows promising results on medical alarm data sets. We\ntrained DQN and A2C agents using the data from monitoring devices annotated by\nan expert. Initial results from these RL agents learning the expert annotation\nbehavior are promising. The A2C agent performs better in terms of learning the\nsparse events in a given state, thereby choosing more right actions compared to\nDQN agent. To the best of our knowledge, this is the first reinforcement\nlearning application for the automation of medical events annotation, which has\nfar-reaching practical use.', 'Recent advances in personalized recommendation have sparked great interest in\nthe exploitation of rich structured information provided by knowledge graphs.\nUnlike most existing approaches that only focus on leveraging knowledge graphs\nfor more accurate recommendation, we perform explicit reasoning with knowledge\nfor decision making so that the recommendations are generated and supported by\nan interpretable causal inference procedure. To this end, we propose a method\ncalled Policy-Guided Path Reasoning (PGPR), which couples recommendation and\ninterpretability by providing actual paths in a knowledge graph. Our\ncontributions include four aspects. We first highlight the significance of\nincorporating knowledge graphs into recommendation to formally define and\ninterpret the reasoning process. Second, we propose a reinforcement learning\n(RL) approach featuring an innovative soft reward strategy, user-conditional\naction pruning and a multi-hop scoring function. Third, we design a\npolicy-guided graph search algorithm to efficiently and effectively sample\nreasoning paths for recommendation. Finally, we extensively evaluate our method\non several large-scale real-world benchmark datasets, obtaining favorable\nresults compared with state-of-the-art methods.', 'Randomized controlled trials typically analyze the effectiveness of\ntreatments with the goal of making treatment recommendations for patient\nsubgroups. With the advance of electronic health records, a great variety of\ndata has been collected in clinical practice, enabling the evaluation of\ntreatments and treatment policies based on observational data. In this paper,\nwe focus on learning individualized treatment rules (ITRs) to derive a\ntreatment policy that is expected to generate a better outcome for an\nindividual patient. In our framework, we cast ITRs learning as a contextual\nbandit problem and minimize the expected risk of the treatment policy. We\nconduct experiments with the proposed framework both in a simulation study and\nbased on a real-world dataset. In the latter case, we apply our proposed method\nto learn the optimal ITRs for the administration of intravenous (IV) fluids and\nvasopressors (VP). Based on various offline evaluation methods, we could show\nthat the policy derived in our framework demonstrates better performance\ncompared to both the physicians and other baselines, including a simple\ntreatment prediction approach. As a long-term goal, our derived policy might\neventually lead to better clinical guidelines for the administration of IV and\nVP.', 'Evaluating the clinical similarities between pairwise patients is a\nfundamental problem in healthcare informatics. A proper patient similarity\nmeasure enables various downstream applications, such as cohort study and\ntreatment comparative effectiveness research. One major carrier for conducting\npatient similarity research is Electronic Health Records(EHRs), which are\nusually heterogeneous, longitudinal, and sparse. Though existing studies on\nlearning patient similarity from EHRs have shown being useful in solving real\nclinical problems, their applicability is limited due to the lack of medical\ninterpretations. Moreover, most previous methods assume a vector-based\nrepresentation for patients, which typically requires aggregation of medical\nevents over a certain time period. As a consequence, temporal information will\nbe lost. In this paper, we propose a patient similarity evaluation framework\nbased on the temporal matching of longitudinal patient EHRs. Two efficient\nmethods are presented, unsupervised and supervised, both of which preserve the\ntemporal properties in EHRs. The supervised scheme takes a convolutional neural\nnetwork architecture and learns an optimal representation of patient clinical\nrecords with medical concept embedding. The empirical results on real-world\nclinical data demonstrate substantial improvement over the baselines. We make\nour code and sample data available for further study.', ""Low attendance levels in medical appointments have been associated with poor\nhealth outcomes and efficiency problems for service providers. To address this\nproblem, healthcare managers could aim at improving attendance levels or\nminimizing the operational impact of no-shows by adapting resource allocation\npolicies. However, given the uncertainty of patient behaviour, generating\nrelevant information regarding no-show probabilities could support the\ndecision-making process for both approaches. In this context many researchers\nhave used multiple regression models to identify patient and appointment\ncharacteristics than can be used as good predictors for no-show probabilities.\nThis work develops a Decision Support System (DSS) to support the\nimplementation of strategies to encourage attendance, for a preventive care\nprogram targeted at underserved communities in Bogot\\'a, Colombia. Our\ncontribution to literature is threefold. Firstly, we assess the effectiveness\nof different machine learning approaches to improve the accuracy of regression\nmodels. In particular, Random Forest and Neural Networks are used to model the\nproblem accounting for non-linearity and variable interactions. Secondly, we\npropose a novel use of Layer-wise Relevance Propagation in order to improve the\nexplainability of neural network predictions and obtain insights from the\nmodelling step. Thirdly, we identify variables explaining no-show probabilities\nin a developing context and study its policy implications and potential for\nimproving healthcare access. In addition to quantifying relationships reported\nin previous studies, we find that income and neighbourhood crime statistics\naffect no-show probabilities. Our results will support patient prioritization\nin a pilot behavioural intervention and will inform appointment planning\ndecisions."", 'Machine learning has successfully framed many sequential decision making\nproblems as either supervised prediction, or optimal decision-making policy\nidentification via reinforcement learning. In data-constrained offline\nsettings, both approaches may fail as they assume fully optimal behavior or\nrely on exploring alternatives that may not exist. We introduce an inherently\ndifferent approach that identifies possible ""dead-ends"" of a state space. We\nfocus on the condition of patients in the intensive care unit, where a ""medical\ndead-end"" indicates that a patient will expire, regardless of all potential\nfuture treatment sequences. We postulate ""treatment security"" as avoiding\ntreatments with probability proportional to their chance of leading to\ndead-ends, present a formal proof, and frame discovery as an RL problem. We\nthen train three independent deep neural models for automated state\nconstruction, dead-end discovery and confirmation. Our empirical results\ndiscover that dead-ends exist in real clinical data among septic patients, and\nfurther reveal gaps between secure treatments and those that were administered.', 'Objective: To combine medical knowledge and medical data to interpretably\npredict the risk of disease. Methods: We formulated the disease prediction task\nas a random walk along a knowledge graph (KG). Specifically, we build a KG to\nrecord relationships between diseases and risk factors according to validated\nmedical knowledge. Then, a mathematical object walks along the KG. It starts\nwalking at a patient entity, which connects the KG based on the patient current\ndiseases or risk factors and stops at a disease entity, which represents the\npredicted disease. The trajectory generated by the object represents an\ninterpretable disease progression path of the given patient. The dynamics of\nthe object are controlled by a policy-based reinforcement learning (RL) module,\nwhich is trained by electronic health records (EHRs). Experiments: We utilized\ntwo real-world EHR datasets to evaluate the performance of our model. In the\ndisease prediction task, our model achieves 0.743 and 0.639 in terms of macro\narea under the curve (AUC) in predicting 53 circulation system diseases in the\ntwo datasets, respectively. This performance is comparable to the commonly used\nmachine learning (ML) models in medical research. In qualitative analysis, our\nclinical collaborator reviewed the disease progression paths generated by our\nmodel and advocated their interpretability and reliability. Conclusion:\nExperimental results validate the proposed model in interpretably evaluating\nand optimizing disease prediction. Significance: Our work contributes to\nleveraging the potential of medical knowledge and medical data jointly for\ninterpretable prediction tasks.', ""Opioids are the preferred medications for the treatment of pain in the\nintensive care unit. While undertreatment leads to unrelieved pain and poor\nclinical outcomes, excessive use of opioids puts patients at risk of\nexperiencing multiple adverse effects. In this work, we present a sequential\ndecision making framework for opioid dosing based on deep reinforcement\nlearning. It provides real-time clinically interpretable dosing\nrecommendations, personalized according to each patient's evolving pain and\nphysiological condition. We focus on morphine, one of the most commonly\nprescribed opioids. To train and evaluate the model, we used retrospective data\nfrom the publicly available MIMIC-3 database. Our results demonstrate that\nreinforcement learning may be used to aid decision making in the intensive care\nsetting by providing personalized pain management interventions."", 'A treatment regime is a function that maps individual patient information to\na recommended treatment, hence explicitly incorporating the heterogeneity in\nneed for treatment across individuals. Patient responses are dichotomous and\ncan be predicted through an unknown relationship that depends on the patient\ninformation and the selected treatment. The goal is to find the treatments that\nlead to the best patient responses on average. Each experiment is expensive,\nforcing us to learn the most from each experiment. We adopt a Bayesian approach\nboth to incorporate possible prior information and to update our treatment\nregime continuously as information accrues, with the potential to allow smaller\nyet more informative trials and for patients to receive better treatment. By\nformulating the problem as contextual bandits, we introduce a knowledge\ngradient policy to guide the treatment assignment by maximizing the expected\nvalue of information, for which an approximation method is used to overcome\ncomputational challenges. We provide a detailed study on how to make sequential\nmedical decisions under uncertainty to reduce health care costs on a real world\nknee replacement dataset. We use clustering and LASSO to deal with the\nintrinsic sparsity in health datasets. We show experimentally that even though\nthe problem is sparse, through careful selection of physicians (versus picking\nthem at random), we can significantly improve the success rates.', 'We propose to use boosted regression trees as a way to compute\nhuman-interpretable solutions to reinforcement learning problems. Boosting\ncombines several regression trees to improve their accuracy without\nsignificantly reducing their inherent interpretability. Prior work has focused\nindependently on reinforcement learning and on interpretable machine learning,\nbut there has been little progress in interpretable reinforcement learning. Our\nexperimental results show that boosted regression trees compute solutions that\nare both interpretable and match the quality of leading reinforcement learning\nmethods.', 'Doctors often rely on their past experience in order to diagnose patients.\nFor a doctor with enough experience, almost every patient would have\nsimilarities to key cases seen in the past, and each new patient could be\nviewed as a mixture of these key past cases. Because doctors often tend to\nreason this way, an efficient computationally aided diagnostic tool that thinks\nin the same way might be helpful in locating key past cases of interest that\ncould assist with diagnosis. This article develops a novel mathematical model\nto mimic the type of logical thinking that physicians use when considering past\ncases. The proposed model can also provide physicians with explanations that\nwould be similar to the way they would naturally reason about cases. The\nproposed method is designed to yield predictive accuracy, computational\nefficiency, and insight into medical data; the key element is the insight into\nmedical data, in some sense we are automating a complicated process that\nphysicians might perform manually. We finally implemented the result of this\nwork on two publicly available healthcare datasets, for heart disease\nprediction and breast cancer prediction.', 'Cardiovascular diseases and heart failures in particular are the main cause\nof non-communicable disease mortality in the world. Constant patient monitoring\nenables better medical treatment as it allows practitioners to react on time\nand provide the appropriate treatment. Telemedicine can provide constant remote\nmonitoring so patients can stay in their homes, only requiring medical sensing\nequipment and network connections. A limiting factor for telemedical centers is\nthe amount of patients that can be monitored simultaneously. We aim to increase\nthis amount by implementing a decision support system. This paper investigates\na machine learning model to estimate a risk score based on patient vital\nparameters that allows sorting all cases every day to help practitioners focus\ntheir limited capacities on the most severe cases. The model we propose reaches\nan AUCROC of 0.84, whereas the baseline rule-based model reaches an AUCROC of\n0.73. Our results indicate that the usage of deep learning to improve the\nefficiency of telemedical centers is feasible. This way more patients could\nbenefit from better health-care through remote monitoring.', ""During the first wave of COVID-19, hospitals were overwhelmed with the high\nnumber of admitted patients. An accurate prediction of the most likely\nindividual disease progression can improve the planning of limited resources\nand finding the optimal treatment for patients. However, when dealing with a\nnewly emerging disease such as COVID-19, the impact of patient- and\ndisease-specific factors (e.g. body weight or known co-morbidities) on the\nimmediate course of disease is by and large unknown. In the case of COVID-19,\nthe need for intensive care unit (ICU) admission of pneumonia patients is often\ndetermined only by acute indicators such as vital signs (e.g. breathing rate,\nblood oxygen levels), whereas statistical analysis and decision support systems\nthat integrate all of the available data could enable an earlier prognosis. To\nthis end, we propose a holistic graph-based approach combining both imaging and\nnon-imaging information. Specifically, we introduce a multimodal similarity\nmetric to build a population graph for clustering patients and an image-based\nend-to-end Graph Attention Network to process this graph and predict the\nCOVID-19 patient outcomes: admission to ICU, need for ventilation and\nmortality. Additionally, the network segments chest CT images as an auxiliary\ntask and extracts image features and radiomics for feature fusion with the\navailable metadata. Results on a dataset collected in Klinikum rechts der Isar\nin Munich, Germany show that our approach outperforms single modality and\nnon-graph baselines. Moreover, our clustering and graph attention allow for\nincreased understanding of the patient relationships within the population\ngraph and provide insight into the network's decision-making process."", ""For mental disorders, patients' underlying mental states are non-observed\nlatent constructs which have to be inferred from observed multi-domain\nmeasurements such as diagnostic symptoms and patient functioning scores.\nAdditionally, substantial heterogeneity in the disease diagnosis between\npatients needs to be addressed for optimizing individualized treatment policy\nin order to achieve precision medicine. To address these challenges, we propose\nan integrated learning framework that can simultaneously learn patients'\nunderlying mental states and recommend optimal treatments for each individual.\nThis learning framework is based on the measurement theory in psychiatry for\nmodeling multiple disease diagnostic measures as arising from the underlying\ncauses (true mental states). It allows incorporation of the multivariate pre-\nand post-treatment outcomes as well as biological measures while preserving the\ninvariant structure for representing patients' latent mental states. A\nmulti-layer neural network is used to allow complex treatment effect\nheterogeneity. Optimal treatment policy can be inferred for future patients by\ncomparing their potential mental states under different treatments given the\nobserved multi-domain pre-treatment measurements. Experiments on simulated data\nand a real-world clinical trial data show that the learned treatment polices\ncompare favorably to alternative methods on heterogeneous treatment effects,\nand have broad utilities which lead to better patient outcomes on multiple\ndomains."", ""Adverse drug events (ADEs) are a serious health problem that can be\nlife-threatening. While a lot of studies have been performed on detect\ncorrelation between a drug and an AE, limited studies have been conducted on\npersonalized ADE risk prediction. Among treatment alternatives, avoiding the\ndrug that has high likelihood of causing severe AE can help physicians to\nprovide safer treatment to patients. Existing work on personalized ADE risk\nprediction uses the information obtained in the current medical visit. However,\non the other hand, medical history reveals each patient's unique\ncharacteristics and comprehensive medical information. The goal of this study\nis to assess personalized ADE risks that a target drug may induce on a target\npatient, based on patient medical history recorded in claims codes, which\nprovide information about diagnosis, drugs taken, related medical supplies\nbesides billing information. We developed a HTNNR model (Hierarchical\nTime-aware Neural Network for ADE Risk) that capture characteristics of claim\ncodes and their relationship. The empirical evaluation show that the proposed\nHTNNR model substantially outperforms the comparison methods, especially for\nrare drugs."", 'The huge wealth of data in the health domain can be exploited to create\nmodels that predict development of health states over time. Temporal learning\nalgorithms are well suited to learn relationships between health states and\nmake predictions about their future developments. However, these algorithms:\n(1) either focus on learning one generic model for all patients, providing\ngeneral insights but often with limited predictive performance, or (2) learn\nindividualized models from which it is hard to derive generic concepts. In this\npaper, we present a middle ground, namely parameterized dynamical systems\nmodels that are generated from data using a Genetic Programming (GP) framework.\nA fitness function suitable for the health domain is exploited. An evaluation\nof the approach in the mental health domain shows that performance of the model\ngenerated by the GP is on par with a dynamical systems model developed based on\ndomain knowledge, significantly outperforms a generic Long Term Short Term\nMemory (LSTM) model and in some cases also outperforms an individualized LSTM\nmodel.', ""Driven by the need to capture users' evolving interests and optimize their\nlong-term experiences, more and more recommender systems have started to model\nrecommendation as a Markov decision process and employ reinforcement learning\nto address the problem. Shouldn't research on the fairness of recommender\nsystems follow the same trend from static evaluation and one-shot intervention\nto dynamic monitoring and non-stop control? In this paper, we portray the\nrecent developments in recommender systems first and then discuss how fairness\ncould be baked into the reinforcement learning techniques for recommendation.\nMoreover, we argue that in order to make further progress in recommendation\nfairness, we may want to consider multi-agent (game-theoretic) optimization,\nmulti-objective (Pareto) optimization, and simulation-based optimization, in\nthe general framework of stochastic games."", 'Accurate and explainable health event predictions are becoming crucial for\nhealthcare providers to develop care plans for patients. The availability of\nelectronic health records (EHR) has enabled machine learning advances in\nproviding these predictions. However, many deep learning based methods are not\nsatisfactory in solving several key challenges: 1) effectively utilizing\ndisease domain knowledge; 2) collaboratively learning representations of\npatients and diseases; and 3) incorporating unstructured text. To address these\nissues, we propose a collaborative graph learning model to explore\npatient-disease interactions and medical domain knowledge. Our solution is able\nto capture structural features of both patients and diseases. The proposed\nmodel also utilizes unstructured text data by employing an attention regulation\nstrategy and then integrates attentive text features into a sequential learning\nprocess. We conduct extensive experiments on two important healthcare problems\nto show the competitive prediction performance of the proposed method compared\nwith various state-of-the-art models. We also confirm the effectiveness of\nlearned representations and model interpretability by a set of ablation and\ncase studies.', ""Patient similarity assessment, which identifies patients similar to a given\npatient, can help improve medical care. The assessment can be performed using\nElectronic Medical Records (EMRs). Patient similarity measurement requires\nconverting heterogeneous EMRs into comparable formats to calculate their\ndistance. While versatile document representation learning methods have been\ndeveloped in recent years, it is still unclear how complex EMR data should be\nprocessed to create the most useful patient representations. This study\npresents a new data representation method for EMRs that takes the information\nin clinical narratives into account. To address the limitations of previous\napproaches in handling complex parts of EMR data, an unsupervised method is\nproposed for building a patient representation, which integrates unstructured\ndata with structured data extracted from patients' EMRs. In order to model the\nextracted data, we employed a tree structure that captures the temporal\nrelations of multiple medical events from EMR. We processed clinical notes to\nextract symptoms, signs, and diseases using different tools such as medspaCy,\nMetaMap, and scispaCy and mapped entities to the Unified Medical Language\nSystem (UMLS). After creating a tree data structure, we utilized two novel\nrelabeling methods for the non-leaf nodes of the tree to capture two temporal\naspects of the extracted events. By traversing the tree, we generated a\nsequence that could create an embedding vector for each patient. The\ncomprehensive evaluation of the proposed method for patient similarity and\nmortality prediction tasks demonstrated that our proposed model leads to lower\nmean squared error (MSE), higher precision, and normalized discounted\ncumulative gain (NDCG) relative to baselines."", ""As part of a plan launched by the Ministry of Health of Brazil to increase\nthe availability of linear accelerators for radiotherapy treatment for the\nwhole country, for which Varian Medical Systems company has won the bidding, a\ntechnical cooperation agreement was signed inviting Brazilian Scientific and\nTechnological Institutions to participate in a technology transfer program. As\na result, jointly, the Eldorado Research Institute and the Center for\nBiomedical Engineering of the University of Campinas presents in this work, the\nconcepts behind of a proposed rule engine to aid in the evaluation and\ndecision-making in radiotherapy treatment planning. Normally, the determination\nof the radiation dose for a given patient is a complex and intensive procedure,\nwhich requires a lot of domain knowledge and subjective experience from the\noncologists' team. In order to help them in this complex task, and\nadditionally, provide an auxiliary tool for less experienced oncologists, it is\npresented a project conception of a software system that will make use of a\nhybrid data-oriented approach. The proposed rule engine will apply both\ninference mechanism and expression evaluation to verify and accredit the\nquality of an external beam radiation treatment plan by considering, at first,\nthe 3D-conformal radiotherapy (3DCRT) technique."", 'Accurate and transparent prediction of cancer survival times on the level of\nindividual patients can inform and improve patient care and treatment\npractices. In this paper, we design a model that concurrently learns to\naccurately predict patient-specific survival distributions and to explain its\npredictions in terms of patient attributes such as clinical tests or\nassessments. Our model is flexible and based on a recurrent network, can handle\nvarious modalities of data including temporal measurements, and yet constructs\nand uses simple explanations in the form of patient- and time-specific linear\nregression. For analysis, we use two publicly available datasets and show that\nour networks outperform a number of baselines in prediction while providing a\nway to inspect the reasons behind each prediction.', 'How to leverage cross-document interactions to improve ranking performance is\nan important topic in information retrieval (IR) research. However, this topic\nhas not been well-studied in the learning-to-rank setting and most of the\nexisting work still treats each document independently while scoring. The\nrecent development of deep learning shows strength in modeling complex\nrelationships across sequences and sets. It thus motivates us to study how to\nleverage cross-document interactions for learning-to-rank in the deep learning\nframework. In this paper, we formally define the permutation-equivariance\nrequirement for a scoring function that captures cross-document interactions.\nWe then propose a self-attention based document interaction network and show\nthat it satisfies the permutation-equivariant requirement, and can generate\nscores for document sets of varying sizes. Our proposed methods can\nautomatically learn to capture document interactions without any auxiliary\ninformation, and can scale across large document sets. We conduct experiments\non three ranking datasets: the benchmark Web30k, a Gmail search, and a Google\nDrive Quick Access dataset. Experimental results show that our proposed methods\nare both more effective and efficient than baselines.', ""Machine learning is used in medicine to support physicians in examination,\ndiagnosis, and predicting outcomes. One of the most dynamic area is the usage\nof patient generated health data from intensive care units. The goal of this\npaper is to demonstrate how we advance cross-patient ML model development by\ncombining the patient's demographics data with their physiological data. We\nused a population of patients undergoing Carotid Enderarterectomy (CEA), where\nwe studied differences in model performance and explainability when trained for\nall patients and one patient at a time. The results show that patients'\ndemographics has a large impact on the performance and explainability and thus\ntrustworthiness. We conclude that we can increase trust in ML models in a\ncross-patient context, by careful selection of models and patients based on\ntheir demographics and the surgical procedure."", 'In this paper we describe COLIN, a forward-chaining heuristic search planner,\ncapable of reasoning with COntinuous LINear numeric change, in addition to the\nfull temporal semantics of PDDL. Through this work we make two advances to the\nstate-of-the-art in terms of expressive reasoning capabilities of planners: the\nhandling of continuous linear change, and the handling of duration-dependent\neffects in combination with duration inequalities, both of which require\ntightly coupled temporal and numeric reasoning during planning. COLIN combines\nFF-style forward chaining search, with the use of a Linear Program (LP) to\ncheck the consistency of the interacting temporal and numeric constraints at\neach state. The LP is used to compute bounds on the values of variables in each\nstate, reducing the range of actions that need to be considered for\napplication. In addition, we develop an extension of the Temporal Relaxed\nPlanning Graph heuristic of CRIKEY3, to support reasoning directly with\ncontinuous change. We extend the range of task variables considered to be\nsuitable candidates for specifying the gradient of the continuous numeric\nchange effected by an action. Finally, we explore the potential for employing\nmixed integer programming as a tool for optimising the timestamps of the\nactions in the plan, once a solution has been found. To support this, we\nfurther contribute a selection of extended benchmark domains that include\ncontinuous numeric effects. We present results for COLIN that demonstrate its\nscalability on a range of benchmarks, and compare to existing state-of-the-art\nplanners.', ""With increasing and extensive use of electronic health records, clinicians\nare often under time pressure when they need to retrieve important information\nefficiently among large amounts of patients' health records in clinics. While a\nsearch function can be a useful alternative to browsing through a patient's\nrecord, it is cumbersome for clinicians to search repeatedly for the same or\nsimilar information on similar patients. Under such circumstances, there is a\ncritical need to build effective recommender systems that can generate accurate\nsearch term recommendations for clinicians. In this manuscript, we developed a\nhybrid collaborative filtering model using patients' encounter and search term\ninformation to recommend the next search terms for clinicians to retrieve\nimportant information fast in clinics. For each patient, the model will\nrecommend terms that either have high co-occurrence frequencies with his/her\nmost recent ICD codes or are highly relevant to the most recent search terms on\nthis patient. We have conducted comprehensive experiments to evaluate the\nproposed model, and the experimental results demonstrate that our model can\noutperform all the state-of-the-art baseline methods for top-N search term\nrecommendation on different datasets."", 'With the successful adoption of machine learning on electronic health records\n(EHRs), numerous computational models have been deployed to address a variety\nof clinical problems. However, due to the heterogeneity of EHRs, models trained\non different patient groups suffer from poor generalizability. How to mitigate\ndomain shifts between the source patient group where the model is built upon\nand the target one where the model will be deployed becomes a critical issue.\nIn this paper, we propose a data augmentation method to facilitate domain\nadaptation, which leverages knowledge from the source patient group when\ntraining model on the target one. Specifically, adversarially generated samples\nare used during domain adaptation to fill the generalization gap between the\ntwo patient groups. The proposed method is evaluated by a case study on\ndifferent predictive modeling tasks on MIMIC-III EHR dataset. Results confirm\nthe effectiveness of our method and the generality on different tasks.', 'Management of chronic diseases such as heart failure (HF) is a major public\nhealth problem. A standard approach to managing chronic diseases by medical\ncommunity is to have a committee of experts develop guidelines that all\nphysicians should follow. Due to their complexity, these guidelines are\ndifficult to implement and are adopted slowly by the medical community at\nlarge. We have developed a physician advisory system that codes the entire set\nof clinical practice guidelines for managing HF using answer set\nprogramming(ASP). In this paper we show how abductive reasoning can be deployed\nto find missing symptoms and conditions that the patient must exhibit in order\nfor a treatment prescribed by a physician to work effectively. Thus, if a\nphysician does not make an appropriate recommendation or makes a non-adherent\nrecommendation, our system will advise the physician about symptoms and\nconditions that must be in effect for that recommendation to apply. It is under\nconsideration for acceptance in TPLP.', ""In clinical practice, physicians make a series of treatment decisions over\nthe course of a patient's disease based on his/her baseline and evolving\ncharacteristics. A dynamic treatment regime is a set of sequential decision\nrules that operationalizes this process. Each rule corresponds to a decision\npoint and dictates the next treatment action based on the accrued information.\nUsing existing data, a key goal is estimating the optimal regime, that, if\nfollowed by the patient population, would yield the most favorable outcome on\naverage. Q- and A-learning are two main approaches for this purpose. We provide\na detailed account of these methods, study their performance, and illustrate\nthem using data from a depression study."", 'Medical AI has tremendous potential to advance healthcare by supporting the\nevidence-based practice of medicine, personalizing patient treatment, reducing\ncosts, and improving provider and patient experience. We argue that unlocking\nthis potential requires a systematic way to measure the performance of medical\nAI models on large-scale heterogeneous data. To meet this need, we are building\nMedPerf, an open framework for benchmarking machine learning in the medical\ndomain. MedPerf will enable federated evaluation in which models are securely\ndistributed to different facilities for evaluation, thereby empowering\nhealthcare organizations to assess and verify the performance of AI models in\nan efficient and human-supervised process, while prioritizing privacy. We\ndescribe the current challenges healthcare and AI communities face, the need\nfor an open platform, the design philosophy of MedPerf, its current\nimplementation status, and our roadmap. We call for researchers and\norganizations to join us in creating the MedPerf open benchmarking platform.', 'A main research goal in various studies is to use an observational data set\nand provide a new set of counterfactual guidelines that can yield causal\nimprovements. Dynamic Treatment Regimes (DTRs) are widely studied to formalize\nthis process. However, available methods in finding optimal DTRs often rely on\nassumptions that are violated in real-world applications (e.g., medical\ndecision-making or public policy), especially when (a) the existence of\nunobserved confounders cannot be ignored, and (b) the unobserved confounders\nare time-varying (e.g., affected by previous actions). When such assumptions\nare violated, one often faces ambiguity regarding the underlying causal model.\nThis ambiguity is inevitable, since the dynamics of unobserved confounders and\ntheir causal impact on the observed part of the data cannot be understood from\nthe observed data. Motivated by a case study of finding superior treatment\nregimes for patients who underwent transplantation in our partner hospital and\nfaced a medical condition known as New Onset Diabetes After Transplantation\n(NODAT), we extend DTRs to a new class termed Ambiguous Dynamic Treatment\nRegimes (ADTRs), in which the causal impact of treatment regimes is evaluated\nbased on a ""cloud"" of causal models. We then connect ADTRs to Ambiguous\nPartially Observable Mark Decision Processes (APOMDPs) and develop\nReinforcement Learning methods, which enable using the observed data to\nefficiently learn an optimal treatment regime. We establish theoretical results\nfor these learning methods, including (weak) consistency and asymptotic\nnormality. We further evaluate the performance of these learning methods both\nin our case study and in simulation experiments.', 'We propose a general formulation for continuous treatment recommendation\nproblems in settings with clinical survival data, which we call the Deep\nSurvival Dose Response Function (DeepSDRF). That is, we consider the problem of\nlearning the conditional average dose response (CADR) function solely from\nhistorical data in which observed factors (confounders) affect both observed\ntreatment and time-to-event outcomes. The estimated treatment effect from\nDeepSDRF enables us to develop recommender algorithms with the correction for\nselection bias. We compared two recommender approaches based on random search\nand reinforcement learning and found similar performance in terms of patient\noutcome. We tested the DeepSDRF and the corresponding recommender on extensive\nsimulation studies and the eICU Research Institute (eRI) database. To the best\nof our knowledge, this is the first time that causal models are used to address\nthe continuous treatment effect with observational data in a medical context.', 'Today, treatment effect estimation at the individual level is a vital problem\nin many areas of science and business. For example, in marketing, estimates of\nthe treatment effect are used to select the most efficient promo-mechanics; in\nmedicine, individual treatment effects are used to determine the optimal dose\nof medication for each patient and so on. At the same time, the question on\nchoosing the best method, i.e., the method that ensures the smallest predictive\nerror (for instance, RMSE) or the highest total (average) value of the effect,\nremains open. Accordingly, in this paper we compare the effectiveness of\nmachine learning methods for estimation of individual treatment effects. The\ncomparison is performed on the Criteo Uplift Modeling Dataset. In this paper we\nshow that the combination of the Logistic Regression method and the Difference\nScore method as well as Uplift Random Forest method provide the best\ncorrectness of Individual Treatment Effect prediction on the top 30\\%\nobservations of the test dataset.', ""Temporal planning often involves numeric effects that are directly\nproportional to their action's duration. These include continuous effects,\nwhere a numeric variable is subjected to a rate of change while the action is\nbeing executed, and discrete duration-dependent effects, where the variable is\nupdated instantaneously but the magnitude of such change is computed from the\naction's duration. When these effects are linear, state--of--the--art temporal\nplanners often make use of Linear Programming to ensure that these numeric\nupdates are consistent with the chosen start times and durations of the plan's\nactions. This is typically done for each evaluated state as part of the search\nprocess. This exhaustive approach is not scalable to solve real-world problems\nthat require long plans, because the linear program's size becomes larger and\nslower to solve. In this work we propose techniques that minimise this overhead\nby computing these checks more selectively and formulating linear programs that\nhave a smaller footprint. The effectiveness of these techniques is demonstrated\non domains that use a mix of discrete and continuous effects, which is typical\nof real-world planning problems. The resultant planner also outperforms most\nstate-of-the-art temporal-numeric and hybrid planners, in terms of both\ncoverage and scalability."", 'Clinical trials play important roles in drug development but often suffer\nfrom expensive, inaccurate and insufficient patient recruitment. The\navailability of massive electronic health records (EHR) data and trial\neligibility criteria (EC) bring a new opportunity to data driven patient\nrecruitment. One key task named patient-trial matching is to find qualified\npatients for clinical trials given structured EHR and unstructured EC text\n(both inclusion and exclusion criteria). How to match complex EC text with\nlongitudinal patient EHRs? How to embed many-to-many relationships between\npatients and trials? How to explicitly handle the difference between inclusion\nand exclusion criteria? In this paper, we proposed CrOss-Modal PseudO-SiamEse\nnetwork (COMPOSE) to address these challenges for patient-trial matching. One\npath of the network encodes EC using convolutional highway network. The other\npath processes EHR with multi-granularity memory network that encodes\nstructured patient records into multiple levels based on medical ontology.\nUsing the EC embedding as query, COMPOSE performs attentional record alignment\nand thus enables dynamic patient-trial matching. COMPOSE also introduces a\ncomposite loss term to maximize the similarity between patient records and\ninclusion criteria while minimize the similarity to the exclusion criteria.\nExperiment results show COMPOSE can reach 98.0% AUC on patient-criteria\nmatching and 83.7% accuracy on patient-trial matching, which leads 24.3%\nimprovement over the best baseline on real-world patient-trial matching tasks.', 'In realistic scenarios, multivariate timeseries evolve over case-by-case\ntime-scales. This is particularly clear in medicine, where the rate of clinical\nevents varies by ward, patient, and application. Increasingly complex models\nhave been shown to effectively predict patient outcomes, but have failed to\nadapt granularity to these inherent temporal resolutions. As such, we introduce\na novel, more realistic, approach to generating patient outcome predictions at\nan adaptive rate based on uncertainty accumulation in Bayesian recurrent\nmodels. We use a Recurrent Neural Network (RNN) and a Bayesian embedding layer\nwith a new aggregation method to demonstrate adaptive prediction timing. Our\nmodel predicts more frequently when events are dense or the model is certain of\nevent latent representations, and less frequently when readings are sparse or\nthe model is uncertain. At 48 hours after patient admission, our model achieves\nequal performance compared to its static-windowed counterparts, while\ngenerating patient- and event-specific prediction timings that lead to improved\npredictive performance over the crucial first 12 hours of the patient stay.', ""Sepsis is the leading cause of mortality in the ICU. It is challenging to\nmanage because individual patients respond differently to treatment. Thus,\ntailoring treatment to the individual patient is essential for the best\noutcomes. In this paper, we take steps toward this goal by applying a\nmixture-of-experts framework to personalize sepsis treatment. The mixture model\nselectively alternates between neighbor-based (kernel) and deep reinforcement\nlearning (DRL) experts depending on patient's current history. On a large\nretrospective cohort, this mixture-based approach outperforms physician, kernel\nonly, and DRL-only experts."", 'When observational data is available from practical studies and a directed\ncyclic graph for how various variables affect each other is known based on\nsubstantive understanding of the process, we consider a problem in which a\ncontrol plan of a treatment variable is conducted in order to bring a response\nvariable close to a target value with variation reduction. We formulate an\noptimal control plan concerning a certain treatment variable through path\ncoefficients in the framework of linear nonrecursive structural equation\nmodels. Based on the formulation, we clarify the properties of causal effects\nwhen conducting a control plan. The results enable us to evaluate the effect of\na control plan on the variance from observational data.', 'Given recent deep learning results that demonstrate the ability to\neffectively optimize high-dimensional non-convex functions with gradient\ndescent optimization on GPUs, we ask in this paper whether symbolic gradient\noptimization tools such as Tensorflow can be effective for planning in hybrid\n(mixed discrete and continuous) nonlinear domains with high dimensional state\nand action spaces? To this end, we demonstrate that hybrid planning with\nTensorflow and RMSProp gradient descent is competitive with mixed integer\nlinear program (MILP) based optimization on piecewise linear planning domains\n(where we can compute optimal solutions) and substantially outperforms\nstate-of-the-art interior point methods for nonlinear planning domains.\nFurthermore, we remark that Tensorflow is highly scalable, converging to a\nstrong plan on a large-scale concurrent domain with a total of 576,000\ncontinuous action parameters distributed over a horizon of 96 time steps and\n100 parallel instances in only 4 minutes. We provide a number of insights that\nclarify such strong performance including observations that despite long\nhorizons, RMSProp avoids both the vanishing and exploding gradient problems.\nTogether these results suggest a new frontier for highly scalable planning in\nnonlinear hybrid domains by leveraging GPUs and the power of recent advances in\ngradient descent with highly optimized toolkits like Tensorflow.', ""Personalized medicine, a paradigm of medicine tailored to a patient's\ncharacteristics, is an increasingly attractive field in health care. An\nimportant goal of personalized medicine is to identify a subgroup of patients,\nbased on baseline covariates, that benefits more from the targeted treatment\nthan other comparative treatments. Most of the current subgroup identification\nmethods only focus on obtaining a subgroup with an enhanced treatment effect\nwithout paying attention to subgroup size. Yet, a clinically meaningful\nsubgroup learning approach should identify the maximum number of patients who\ncan benefit from the better treatment. In this paper, we present an optimal\nsubgroup selection rule (SSR) that maximizes the number of selected patients,\nand in the meantime, achieves the pre-specified clinically meaningful mean\noutcome, such as the average treatment effect. We derive two equivalent\ntheoretical forms of the optimal SSR based on the contrast function that\ndescribes the treatment-covariates interaction in the outcome. We further\npropose a ConstrAined PolIcy Tree seArch aLgorithm (CAPITAL) to find the\noptimal SSR within the interpretable decision tree class. The proposed method\nis flexible to handle multiple constraints that penalize the inclusion of\npatients with negative treatment effects, and to address time to event data\nusing the restricted mean survival time as the clinically interesting mean\noutcome. Extensive simulations, comparison studies, and real data applications\nare conducted to demonstrate the validity and utility of our method."", 'Widespread adoption of electronic health records (EHRs) has fueled the\ndevelopment of using machine learning to build prediction models for various\nclinical outcomes. This process is often constrained by having a relatively\nsmall number of patient records for training the model. We demonstrate that\nusing patient representation schemes inspired from techniques in natural\nlanguage processing can increase the accuracy of clinical prediction models by\ntransferring information learned from the entire patient population to the task\nof training a specific model, where only a subset of the population is\nrelevant. Such patient representation schemes enable a 3.5% mean improvement in\nAUROC on five prediction tasks compared to standard baselines, with the average\nimprovement rising to 19% when only a small number of patient records are\navailable for training the clinical prediction model.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'The introduction of data analytics into medicine has changed the nature of\npatient treatment. In this, patients are asked to disclose personal information\nsuch as genetic markers, lifestyle habits, and clinical history. This data is\nthen used by statistical models to predict personalized treatments. However,\ndue to privacy concerns, patients often desire to withhold sensitive\ninformation. This self-censorship can impede proper diagnosis and treatment,\nwhich may lead to serious health complications and even death over time. In\nthis paper, we present privacy distillation, a mechanism which allows patients\nto control the type and amount of information they wish to disclose to the\nhealthcare providers for use in statistical models. Meanwhile, it retains the\naccuracy of models that have access to all patient data under a sufficient but\nnot full set of privacy-relevant information. We validate privacy distillation\nusing a corpus of patients prescribed to warfarin for a personalized dosage. We\nuse a deep neural network to implement privacy distillation for training and\nmaking dose predictions. We find that privacy distillation with sufficient\nprivacy-relevant information i) retains accuracy almost as good as having all\npatient data (only 3\\% worse), and ii) is effective at preventing errors that\nintroduce health-related risks (only 3.9\\% worse under- or over-prescriptions).', ""Tens of millions of women suffer from infertility worldwide each year. In\nvitro fertilization (IVF) is the best choice for many such patients. However,\nIVF is expensive, time-consuming, and both physically and emotionally\ndemanding. The first question that a patient usually asks before the IVF is how\nlikely she will conceive, given her basic medical examination information. This\npaper proposes three approaches to predict the cumulative pregnancy rate after\nmultiple oocyte pickup cycles. Experiments on 11,190 patients showed that first\nclustering the patients into different groups and then building a support\nvector machine model for each group can achieve the best overall performance.\nOur model could be a quick and economic approach for reliably estimating the\ncumulative pregnancy rate for a patient, given only her basic medical\nexamination information, well before starting the actual IVF procedure. The\npredictions can help the patient make optimal decisions on whether to use her\nown oocyte or donor oocyte, how many oocyte pickup cycles she may need, whether\nto use embryo frozen, etc. They will also reduce the patient's cost and time to\npregnancy, and improve her quality of life."", 'Patient similarity analysis is important in health care applications. It\ntakes patient information such as their electronic medical records and genetic\ndata as input and computes the pairwise similarity between patients. Procedures\nof typical a patient similarity study can be divided into several steps\nincluding data integration, similarity measurement, and neighborhood\nidentification. And according to an analysis of patient similarity, doctors can\neasily find the most suitable treatments. There are many methods to analyze the\nsimilarity such as cluster analysis. And during machine learning become more\nand more popular, Using neural networks such as CNN is a new hot topic. This\nreview summarizes representative methods used in each step and discusses\napplications of patient similarity networks especially in the context of\nprecision medicine.', 'With the increase of information, document classification as one of the\nmethods of text mining, plays vital role in many management and organizing\ninformation. Document classification is the process of assigning a document to\none or more predefined category labels. Document classification includes\ndifferent parts such as text processing, term selection, term weighting and\nfinal classification. The accuracy of document classification is very\nimportant. Thus improvement in each part of classification should lead to\nbetter results and higher precision. Term weighting has a great impact on the\naccuracy of the classification. Most of the existing weighting methods exploit\nthe statistical information of terms in documents and do not consider semantic\nrelations between words. In this paper, an automated document classification\nsystem is presented that uses a novel term weighting method based on semantic\nrelations between terms. To evaluate the proposed method, three standard\nPersian corpuses are used. Experiment results show 2 to 4 percent improvement\nin classification accuracy compared with the best previous designed system for\nPersian documents.', ""Alzheimer's disease (AD) is a degenerative brain disease impairing a person's\nability to perform day to day activities. The clinical manifestations of\nAlzheimer's disease are characterized by heterogeneity in age, disease span,\nprogression rate, impairment of memory and cognitive abilities. Due to these\nvariabilities, personalized care and treatment planning, as well as patient\ncounseling about their individual progression is limited. Recent developments\nin machine learning to detect hidden patterns in complex, multi-dimensional\ndatasets provides significant opportunities to address this critical need. In\nthis work, we use unsupervised and supervised machine learning approaches for\nsubtype identification and prediction. We apply machine learning methods to the\nextensive clinical observations available at the Alzheimer's Disease\nNeuroimaging Initiative (ADNI) data set to identify patient subtypes and to\npredict disease progression. Our analysis depicts the progression space for the\nAlzheimer's disease into low, moderate and high disease progression zones. The\nproposed work will enable early detection and characterization of distinct\ndisease subtypes based on clinical heterogeneity. We anticipate that our models\nwill enable patient counseling, clinical trial design, and ultimately\nindividualized clinical care."", 'Avoiding collisions is one of the vital tasks for systems of autonomous\nmobile agents. We focus on the problem of finding continuous coordinated paths\nfor multiple mobile disc agents in a 2-d environment with polygonal obstacles.\nThe problem is PSPACE-hard, with the state space growing exponentially in the\nnumber of agents. Therefore, the state of the art methods include mainly\nreactive techniques and sampling-based iterative algorithms.\n  We compare the performance of a widely-used reactive method ORCA with three\nvariants of a popular planning algorithm RRT* applied to multi-agent path\nplanning and find that an algorithm combining reactive collision avoidance and\nRRT* planning, which we call ORCA-RRT* can be used to solve instances that are\nout of the reach of either of the techniques. We experimentally show that: 1)\nthe reactive part of the algorithm can efficiently solve many multi-agent path\nfinding problems involving large number of agents, for which RRT* algorithm is\noften unable to find a solution in limited time and 2) the planning component\nof the algorithm is able to solve many instances containing local minima, where\nreactive techniques typically fail.', 'Traditional medicine typically applies one-size-fits-all treatment for the\nentire patient population whereas precision medicine develops tailored\ntreatment schemes for different patient subgroups. The fact that some factors\nmay be more significant for a specific patient subgroup motivates clinicians\nand medical researchers to develop new approaches to subgroup detection and\nanalysis, which is an effective strategy to personalize treatment. In this\nstudy, we propose a novel patient subgroup detection method, called Supervised\nBiclustring (SUBIC) using convex optimization and apply our approach to detect\npatient subgroups and prioritize risk factors for hypertension (HTN) in a\nvulnerable demographic subgroup (African-American). Our approach not only finds\npatient subgroups with guidance of a clinically relevant target variable but\nalso identifies and prioritizes risk factors by pursuing sparsity of the input\nvariables and encouraging similarity among the input variables and between the\ninput and target variables', ""Current AI approaches have frequently been used to help personalize many\naspects of medical experiences and tailor them to a specific individuals'\nneeds. However, while such systems consider medically-relevant information,\nthey ignore socially-relevant information about how this diagnosis should be\ncommunicated and discussed with the patient. The lack of this capability may\nlead to mis-communication, resulting in serious implications, such as patients\nopting out of the best treatment. Consider a case in which the same treatment\nis proposed to two different individuals. The manner in which this treatment is\nmediated to each should be different, depending on the individual patient's\nhistory, knowledge, and mental state. While it is clear that this communication\nshould be conveyed via a human medical expert and not a software-based system,\nhumans are not always capable of considering all of the relevant aspects and\ntraversing all available information. We pose the challenge of creating\nIntelligent Agents (IAs) to assist medical service providers (MSPs) and\nconsumers in establishing a more personalized human-to-human dialogue.\nPersonalizing conversations will enable patients and MSPs to reach a solution\nthat is best for their particular situation, such that a relation of trust can\nbe built and commitment to the outcome of the interaction is assured. We\npropose a four-part conceptual framework for personalized social interactions,\nexpand on which techniques are available within current AI research and discuss\nwhat has yet to be achieved."", 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', ""Electronic health records (EHRs) are longitudinal records of a patient's\ninteractions with healthcare systems. A patient's EHR data is organized as a\nthree-level hierarchy from top to bottom: patient journey - all the experiences\nof diagnoses and treatments over a period of time; individual visit - a set of\nmedical codes in a particular visit; and medical code - a specific record in\nthe form of medical codes. As EHRs begin to amass in millions, the potential\nbenefits, which these data might hold for medical research and medical outcome\nprediction, are staggering - including, for example, predicting future\nadmissions to hospitals, diagnosing illnesses or determining the efficacy of\nmedical treatments. Each of these analytics tasks requires a domain knowledge\nextraction method to transform the hierarchical patient journey into a vector\nrepresentation for further prediction procedure. The representations should\nembed a sequence of visits and a set of medical codes with a specific\ntimestamp, which are crucial to any downstream prediction tasks. Hence,\nexpressively powerful representations are appealing to boost learning\nperformance. To this end, we propose a novel self-attention mechanism that\ncaptures the contextual dependency and temporal relationships within a\npatient's healthcare journey. An end-to-end bidirectional temporal encoder\nnetwork (BiteNet) then learns representations of the patient's journeys, based\nsolely on the proposed attention mechanism. We have evaluated the effectiveness\nof our methods on two supervised prediction and two unsupervised clustering\ntasks with a real-world EHR dataset. The empirical results demonstrate the\nproposed BiteNet model produces higher-quality representations than\nstate-of-the-art baseline methods."", 'In several medical decision-making problems, such as antibiotic prescription,\nlaboratory testing can provide precise indications for how a patient will\nrespond to different treatment options. This enables us to ""fully observe"" all\npotential treatment outcomes, but while present in historical data, these\nresults are infeasible to produce in real-time at the point of the initial\ntreatment decision. Moreover, treatment policies in these settings often need\nto trade off between multiple competing objectives, such as effectiveness of\ntreatment and harmful side effects. We present, compare, and evaluate three\napproaches for learning individualized treatment policies in this setting:\nFirst, we consider two indirect approaches, which use predictive models of\ntreatment response to construct policies optimal for different trade-offs\nbetween objectives. Second, we consider a direct approach that constructs such\na set of policies without intermediate models of outcomes. Using a medical\ndataset of Urinary Tract Infection (UTI) patients, we show that all approaches\nlearn policies that achieve strictly better performance on all outcomes than\nclinicians, while also trading off between different objectives. We demonstrate\nadditional benefits of the direct approach, including flexibly incorporating\nother goals such as deferral to physicians on simple cases.', ""This paper presents a probabilistic model for reasoning about the state of a\nsystem as it changes over time, both due to exogenous and endogenous\ninfluences. Our target domain is a class of medical prediction problems that\nare neither so urgent as to preclude careful diagnosis nor progress so slowly\nas to allow arbitrary testing and treatment options. In these domains there is\ntypically enough time to gather information about the patient's state and\nconsider alternative diagnoses and treatments, but the temporal interaction\nbetween the timing of tests, treatments, and the course of the disease must\nalso be considered. Our approach is to elicit a qualitative structural model of\nthe patient from a human expert---the model identifies important attributes,\nthe way in which exogenous changes affect attribute values, and the way in\nwhich the patient's condition changes endogenously. We then elicit\nprobabilistic information to capture the expert's uncertainty about the effects\nof tests and treatments and the nature and timing of endogenous state changes.\nThis paper describes the model in the context of a problem in treating vehicle\naccident trauma, and suggests a method for solving the model based on the\ntechnique of sequential imputation. A complementary goal of this work is to\nunderstand and synthesize a disparate collection of research efforts all using\nthe name ?probabilistic temporal reasoning.? This paper analyzes related work\nand points out essential differences between our proposed model and other\napproaches in the literature."", ""Clinical event sequences consist of thousands of clinical events that\nrepresent records of patient care in time. Developing accurate prediction\nmodels for such sequences is of a great importance for defining representations\nof a patient state and for improving patient care. One important challenge of\nlearning a good predictive model of clinical sequences is patient-specific\nvariability. Based on underlying clinical complications, each patient's\nsequence may consist of different sets of clinical events. However,\npopulation-based models learned from such sequences may not accurately predict\npatient-specific dynamics of event sequences. To address the problem, we\ndevelop a new adaptive event sequence prediction framework that learns to\nadjust its prediction for individual patients through an online model update."", ""A crucial step within secondary analysis of electronic health records (EHRs)\nis to identify the patient cohort under investigation. While EHRs contain\nmedical billing codes that aim to represent the conditions and treatments\npatients may have, much of the information is only present in the patient\nnotes. Therefore, it is critical to develop robust algorithms to infer\npatients' conditions and treatments from their written notes. In this paper, we\nintroduce a dataset for patient phenotyping, a task that is defined as the\nidentification of whether a patient has a given medical condition (also\nreferred to as clinical indication or phenotype) based on their patient note.\nNursing Progress Notes and Discharge Summaries from the Intensive Care Unit of\na large tertiary care hospital were manually annotated for the presence of\nseveral high-context phenotypes relevant to treatment and risk of\nre-hospitalization. This dataset contains 1102 Discharge Summaries and 1000\nNursing Progress Notes. Each Discharge Summary and Progress Note has been\nannotated by at least two expert human annotators (one clinical researcher and\none resident physician). Annotated phenotypes include treatment non-adherence,\nchronic pain, advanced/metastatic cancer, as well as 10 other phenotypes. This\ndataset can be utilized for academic and industrial research in medicine and\ncomputer science, particularly within the field of medical natural language\nprocessing."", 'Contemporary patient surveillance systems have streamlined central\nsurveillance into the electronic health record interface. They are able to\nprocess the sheer volume of patient data by adopting machine learning\napproaches. However, these systems are not suitable for implementation in many\nhospitals, mostly in developing countries, with limited human, financial, and\ntechnological resources. Through conducting thorough research on intensive care\nfacilities, we designed a novel central patient monitoring system and in this\npaper, we describe the working prototype of our system. The proposed prototype\ncomprises of inexpensive peripherals and simplistic user interface. Our central\npatient monitoring system implements Kernel-based On-line Anomaly Detection\n(KOAD) algorithm for emergency event signaling. By evaluating continuous\npatient data, we show that the system is able to detect critical events in\nreal-time reliably and has low false alarm rate.', 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', 'The main question to address in this paper is to recommend optimal signal\ntiming plans in real time under incidents by incorporating domain knowledge\ndeveloped with the traffic signal timing plans tuned for possible incidents,\nand learning from historical data of both traffic and implemented signals\ntiming. The effectiveness of traffic incident management is often limited by\nthe late response time and excessive workload of traffic operators. This paper\nproposes a novel decision-making framework that learns from both data and\ndomain knowledge to real-time recommend contingency signal plans that\naccommodate non-recurrent traffic, with the outputs from real-time traffic\nprediction at least 30 minutes in advance. Specifically, considering the rare\noccurrences of engagement of contingency signal plans for incidents, we propose\nto decompose the end-to-end recommendation task into two hierarchical models:\nreal-time traffic prediction and plan association. We learn the connections\nbetween the two models through metric learning, which reinforces partial-order\npreferences observed from historical signal engagement records. We demonstrate\nthe effectiveness of our approach by testing this framework on the traffic\nnetwork in Cranberry Township in 2019. Results show that our recommendation\nsystem has a precision score of 96.75% and recall of 87.5% on the testing plan,\nand make recommendation of an average of 22.5 minutes lead time ahead of Waze\nalerts. The results suggest that our framework is capable of giving traffic\noperators a significant time window to access the conditions and respond\nappropriately.', ""More than 144 000 Australians were diagnosed with cancer in 2019. The\nmajority will first present to their GP symptomatically, even for cancer for\nwhich screening programs exist. Diagnosing cancer in primary care is\nchallenging due to the non-specific nature of cancer symptoms and its low\nprevalence. Understanding the epidemiology of cancer symptoms and patterns of\npresentation in patient's medical history from primary care data could be\nimportant to improve earlier detection and cancer outcomes. As past medical\ndata about a patient can be incomplete, irregular or missing, this creates\nadditional challenges when attempting to use the patient's history for any new\ndiagnosis. Our research aims to investigate the opportunities in a patient's\npathology history available to a GP, initially focused on the results within\nthe frequently ordered full blood count to determine relevance to a future\nhigh-risk cancer prognosis, and treatment outcome. We investigated how past\npathology test results can lead to deriving features that can be used to\npredict cancer outcomes, with emphasis on patients at risk of not surviving the\ncancer within 2-year period. This initial work focuses on patients with lung\ncancer, although the methodology can be applied to other types of cancer and\nother data within the medical record. Our findings indicate that even in cases\nof incomplete or obscure patient history, hematological measures can be useful\nin generating features relevant for predicting cancer risk and survival. The\nresults strongly indicate to add the use of pathology test data for potential\nhigh-risk cancer diagnosis, and the utilize additional pathology metrics or\nother primary care datasets even more for similar purposes."", 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'Practitioners in medicine, business, political science, and other fields are\nincreasingly aware that decisions should be personalized to each patient,\ncustomer, or voter. A given treatment (e.g. a drug or advertisement) should be\nadministered only to those who will respond most positively, and certainly not\nto those who will be harmed by it. Individual-level treatment effects can be\nestimated with tools adapted from machine learning, but different models can\nyield contradictory estimates. Unlike risk prediction models, however,\ntreatment effect models cannot be easily evaluated against each other using a\nheld-out test set because the true treatment effect itself is never directly\nobserved. Besides outcome prediction accuracy, several metrics that can\nleverage held-out data to evaluate treatment effects models have been proposed,\nbut they are not widely used. We provide a didactic framework that elucidates\nthe relationships between the different approaches and compare them all using a\nvariety of simulations of both randomized and observational data. Our results\nshow that researchers estimating heterogenous treatment effects need not limit\nthemselves to a single model-fitting algorithm. Instead of relying on a single\nmethod, multiple models fit by a diverse set of algorithms should be evaluated\nagainst each other using an objective function learned from the validation set.\nThe model minimizing that objective should be used for estimating the\nindividual treatment effect for future individuals.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', ""In recent years, people from all over the world are suffering from one of the\nmost severe diseases in history, known as Coronavirus disease 2019, COVID-19\nfor short. When the virus reaches the lungs, it has a higher probability to\ncause lung pneumonia and sepsis. X-ray image is a powerful tool in identifying\nthe typical features of the infection for COVID-19 patients. The radiologists\nand pathologists observe that ground-glass opacity appears in the chest X-ray\nfor infected patient \\cite{cozzi2021ground}, and it could be used as one of the\ncriteria during the diagnosis process. In the past few years, deep learning has\nproven to be one of the most powerful methods in the field of image\nclassification. Due to significant differences in Chest X-Ray between normal\nand infected people \\cite{rousan2020chest}, deep models could be used to\nidentify the presence of the disease given a patient's Chest X-Ray. Many deep\nmodels are complex, and it evolves with lots of input parameters. Designers\nsometimes struggle with the tuning process for deep models, especially when\nthey build up the model from scratch. Genetic Algorithm, inspired by the\nbiological evolution process, plays a key role in solving such complex\nproblems. In this paper, I proposed a genetic-based approach to optimize the\nConvolutional Neural Network(CNN) for the Chest X-Ray classification task."", 'Critically ill patients in regular wards are vulnerable to unanticipated\nadverse events which require prompt transfer to the intensive care unit (ICU).\nTo allow for accurate prognosis of deteriorating patients, we develop a novel\ncontinuous-time probabilistic model for a monitored patient\'s temporal sequence\nof physiological data. Our model captures ""informatively sampled"" patient\nepisodes: the clinicians\' decisions on when to observe a hospitalized patient\'s\nvital signs and lab tests over time are represented by a marked Hawkes process,\nwith intensity parameters that are modulated by the patient\'s latent clinical\nstates, and with observable physiological data (mark process) modeled as a\nswitching multi-task Gaussian process. In addition, our model captures\n""informatively censored"" patient episodes by representing the patient\'s latent\nclinical states as an absorbing semi-Markov jump process. The model parameters\nare learned from offline patient episodes in the electronic health records via\nan EM-based algorithm. Experiments conducted on a cohort of patients admitted\nto a major medical center over a 3-year period show that risk prognosis based\non our model significantly outperforms the currently deployed medical risk\nscores and other baseline machine learning algorithms.', ""To address potential gaps noted in patient monitoring in the hospital, a\nnovel patient behavior detection system using mmWave radar and deep convolution\nneural network (CNN), which supports the simultaneous recognition of multiple\npatients' behaviors in real-time, is proposed. In this study, we use an mmWave\nradar to track multiple patients and detect the scattering point cloud of each\none. For each patient, the Doppler pattern of the point cloud over a time\nperiod is collected as the behavior signature. A three-layer CNN model is\ncreated to classify the behavior for each patient. The tracking and point\nclouds detection algorithm was also implemented on an mmWave radar hardware\nplatform with an embedded graphics processing unit (GPU) board to collect\nDoppler pattern and run the CNN model. A training dataset of six types of\nbehavior were collected, over a long duration, to train the model using Adam\noptimizer with an objective to minimize cross-entropy loss function. Lastly,\nthe system was tested for real-time operation and obtained a very good\ninference accuracy when predicting each patient's behavior in a two-patient\nscenario."", 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : As a computer vision researcher, I need a system that does 3D face reconstruction from images from a single camera. My current idea is to leverage the highly correlated spectral and spatial information in 3D visual face models, which are treated as tensors. I want to introduce a tensor completion model to generate a vast amount of possible face completions from all angles and poses. Given these generated images related to a single facial image, the system should use semantic segmentation to identify different facial parts. From these facial parts, the system should discover geometry and global structure of the face. In the end, the system should also determine the best subset of generated images that can be used to optimize the likelihood of the 3D facial reconstruction, conditioned on learned geometry and global structure of the face. The system should outperform the current state-of-the-art facial reconstruction models.","['3D face reconstruction is a fundamental Computer Vision problem of\nextraordinary difficulty. Current systems often assume the availability of\nmultiple facial images (sometimes from the same subject) as input, and must\naddress a number of methodological challenges such as establishing dense\ncorrespondences across large facial poses, expressions, and non-uniform\nillumination. In general these methods require complex and inefficient\npipelines for model building and fitting. In this work, we propose to address\nmany of these limitations by training a Convolutional Neural Network (CNN) on\nan appropriate dataset consisting of 2D images and 3D facial models or scans.\nOur CNN works with just a single 2D facial image, does not require accurate\nalignment nor establishes dense correspondence between images, works for\narbitrary facial poses and expressions, and can be used to reconstruct the\nwhole 3D facial geometry (including the non-visible parts of the face)\nbypassing the construction (during training) and fitting (during testing) of a\n3D Morphable Model. We achieve this via a simple CNN architecture that performs\ndirect regression of a volumetric representation of the 3D facial geometry from\na single 2D image. We also demonstrate how the related task of facial landmark\nlocalization can be incorporated into the proposed framework and help improve\nreconstruction quality, especially for the cases of large poses and facial\nexpressions. Testing code will be made available online, along with pre-trained\nmodels http://aaronsplace.co.uk/papers/jackson2017recon', 'Traditional methods for image-based 3D face reconstruction and facial motion\nretargeting fit a 3D morphable model (3DMM) to the face, which has limited\nmodeling capacity and fail to generalize well to in-the-wild data. Use of\ndeformation transfer or multilinear tensor as a personalized 3DMM for\nblendshape interpolation does not address the fact that facial expressions\nresult in different local and global skin deformations in different persons.\nMoreover, existing methods learn a single albedo per user which is not enough\nto capture the expression-specific skin reflectance variations. We propose an\nend-to-end framework that jointly learns a personalized face model per user and\nper-frame facial motion parameters from a large corpus of in-the-wild videos of\nuser expressions. Specifically, we learn user-specific expression blendshapes\nand dynamic (expression-specific) albedo maps by predicting personalized\ncorrections on top of a 3DMM prior. We introduce novel constraints to ensure\nthat the corrected blendshapes retain their semantic meanings and the\nreconstructed geometry is disentangled from the albedo. Experimental results\nshow that our personalization accurately captures fine-grained facial dynamics\nin a wide range of conditions and efficiently decouples the learned face model\nfrom facial motion, resulting in more accurate face reconstruction and facial\nmotion retargeting compared to state-of-the-art methods.', 'In this paper, we present a large-scale detailed 3D face dataset, FaceScape,\nand the corresponding benchmark to evaluate single-view facial 3D\nreconstruction. By training on FaceScape data, a novel algorithm is proposed to\npredict elaborate riggable 3D face models from a single image input. FaceScape\ndataset provides 18,760 textured 3D faces, captured from 938 subjects and each\nwith 20 specific expressions. The 3D models contain the pore-level facial\ngeometry that is also processed to be topologically uniformed. These fine 3D\nfacial models can be represented as a 3D morphable model for rough shapes and\ndisplacement maps for detailed geometry. Taking advantage of the large-scale\nand high-accuracy dataset, a novel algorithm is further proposed to learn the\nexpression-specific dynamic details using a deep neural network. The learned\nrelationship serves as the foundation of our 3D face prediction system from a\nsingle image input. Different than the previous methods, our predicted 3D\nmodels are riggable with highly detailed geometry under different expressions.\nWe also use FaceScape data to generate the in-the-wild and in-the-lab benchmark\nto evaluate recent methods of single-view face reconstruction. The accuracy is\nreported and analyzed on the dimensions of camera pose and focal length, which\nprovides a faithful and comprehensive evaluation and reveals new challenges.\nThe unprecedented dataset, benchmark, and code have been released to the public\nfor research purpose.', 'Monocular 3D facial shape reconstruction from a single 2D facial image has\nbeen an active research area due to its wide applications. Inspired by the\nsuccess of deep neural networks (DNN), we propose a DNN-based approach for\nEnd-to-End 3D FAce Reconstruction (UH-E2FAR) from a single 2D image. Different\nfrom recent works that reconstruct and refine the 3D face in an iterative\nmanner using both an RGB image and an initial 3D facial shape rendering, our\nDNN model is end-to-end, and thus the complicated 3D rendering process can be\navoided. Moreover, we integrate in the DNN architecture two components, namely\na multi-task loss function and a fusion convolutional neural network (CNN) to\nimprove facial expression reconstruction. With the multi-task loss function, 3D\nface reconstruction is divided into neutral 3D facial shape reconstruction and\nexpressive 3D facial shape reconstruction. The neutral 3D facial shape is\nclass-specific. Therefore, higher layer features are useful. In comparison, the\nexpressive 3D facial shape favors lower or intermediate layer features. With\nthe fusion-CNN, features from different intermediate layers are fused and\ntransformed for predicting the 3D expressive facial shape. Through extensive\nexperiments, we demonstrate the superiority of our end-to-end framework in\nimproving the accuracy of 3D face reconstruction.', 'Deep learning methods have brought many breakthroughs to computer vision,\nespecially in 2D face recognition. However, the bottleneck of deep learning\nbased 3D face recognition is that it is difficult to collect millions of 3D\nfaces, whether for industry or academia. In view of this situation, there are\nmany methods to generate more 3D faces from existing 3D faces through 3D face\ndata augmentation, which are used to train deep 3D face recognition models.\nHowever, to the best of our knowledge, there is no method to generate 3D faces\nfrom 2D face images for training deep 3D face recognition models. This letter\nfocuses on the role of reconstructed 3D facial surfaces in 3D face\nidentification and proposes a framework of 2D-aided deep 3D face\nidentification. In particular, we propose to reconstruct millions of 3D face\nscans from a large scale 2D face database (i.e.VGGFace2), using a deep learning\nbased 3D face reconstruction method (i.e.ExpNet). Then, we adopt a two-phase\ntraining approach: In the first phase, we use millions of face images to\npre-train the deep convolutional neural network (DCNN), and in the second\nphase, we use normal component images (NCI) of reconstructed 3D face scans to\ntrain the DCNN. Extensive experimental results illustrate that the proposed\napproach can greatly improve the rank-1 score of 3D face identification on the\nFRGC v2.0, the Bosphorus, and the BU-3DFE 3D face databases, compared to the\nmodel trained by 2D face images. Finally, our proposed approach achieves\nstate-of-the-art rank-1 scores on the FRGC v2.0 (97.6%), Bosphorus (98.4%), and\nBU-3DFE (98.8%) databases. The experimental results show that the reconstructed\n3D facial surfaces are useful and our 2D-aided deep 3D face identification\nframework is meaningful, facing the scarcity of 3D faces.', 'We present a novel 3D face reconstruction technique that leverages sparse\nphotometric stereo (PS) and latest advances on face registration/modeling from\na single image. We observe that 3D morphable faces approach provides a\nreasonable geometry proxy for light position calibration. Specifically, we\ndevelop a robust optimization technique that can calibrate per-pixel lighting\ndirection and illumination at a very high precision without assuming uniform\nsurface albedos. Next, we apply semantic segmentation on input images and the\ngeometry proxy to refine hairy vs. bare skin regions using tailored filters.\nExperiments on synthetic and real data show that by using a very small set of\nimages, our technique is able to reconstruct fine geometric details such as\nwrinkles, eyebrows, whelks, pores, etc, comparable to and sometimes surpassing\nmovie quality productions.', 'Reconstructing 3D facial geometry from a single RGB image has recently\ninstigated wide research interest. However, it is still an ill-posed problem\nand most methods rely on prior models hence undermining the accuracy of the\nrecovered 3D faces. In this paper, we exploit the Epipolar Plane Images (EPI)\nobtained from light field cameras and learn CNN models that recover horizontal\nand vertical 3D facial curves from the respective horizontal and vertical EPIs.\nOur 3D face reconstruction network (FaceLFnet) comprises a densely connected\narchitecture to learn accurate 3D facial curves from low resolution EPIs. To\ntrain the proposed FaceLFnets from scratch, we synthesize photo-realistic light\nfield images from 3D facial scans. The curve by curve 3D face estimation\napproach allows the networks to learn from only 14K images of 80 identities,\nwhich still comprises over 11 Million EPIs/curves. The estimated facial curves\nare merged into a single pointcloud to which a surface is fitted to get the\nfinal 3D face. Our method is model-free, requires only a few training samples\nto learn FaceLFnet and can reconstruct 3D faces with high accuracy from single\nlight field images under varying poses, expressions and lighting conditions.\nComparison on the BU-3DFE and BU-4DFE datasets show that our method reduces\nreconstruction errors by over 20% compared to recent state of the art.', 'Face alignment and 3D face reconstruction are traditionally accomplished as\nseparated tasks. By exploring the strong correlation between 2D landmarks and\n3D shapes, in contrast, we propose a joint face alignment and 3D face\nreconstruction method to simultaneously solve these two problems for 2D face\nimages of arbitrary poses and expressions. This method, based on a summation\nmodel of 3D faces and cascaded regression in 2D and 3D shape spaces,\niteratively and alternately applies two cascaded regressors, one for updating\n2D landmarks and the other for 3D shape. The 3D shape and the landmarks are\ncorrelated via a 3D-to-2D mapping matrix, which is updated in each iteration to\nrefine the location and visibility of 2D landmarks. Unlike existing methods,\nthe proposed method can fully automatically generate both\npose-and-expression-normalized (PEN) and expressive 3D faces and localize both\nvisible and invisible 2D landmarks. Based on the PEN 3D faces, we devise a\nmethod to enhance face recognition accuracy across poses and expressions. Both\nlinear and nonlinear implementations of the proposed method are presented and\nevaluated in this paper. Extensive experiments show that the proposed method\ncan achieve the state-of-the-art accuracy in both face alignment and 3D face\nreconstruction, and benefit face recognition owing to its reconstructed PEN 3D\nface.']","['Recently, 3D face reconstruction from a single image has achieved great\nsuccess with the help of deep learning and shape prior knowledge, but they\noften fail to produce accurate geometry details. On the other hand, photometric\nstereo methods can recover reliable geometry details, but require dense inputs\nand need to solve a complex optimization problem. In this paper, we present a\nlightweight strategy that only requires sparse inputs or even a single image to\nrecover high-fidelity face shapes with images captured under near-field lights.\nTo this end, we construct a dataset containing 84 different subjects with 29\nexpressions under 3 different lights. Data augmentation is applied to enrich\nthe data in terms of diversity in identity, lighting, expression, etc. With\nthis constructed dataset, we propose a novel neural network specially designed\nfor photometric stereo based 3D face reconstruction. Extensive experiments and\ncomparisons demonstrate that our method can generate high-quality\nreconstruction results with one to three facial images captured under\nnear-field lights. Our full framework is available at\nhttps://github.com/Juyong/FacePSNet.', 'In this work, we present a new method for 3D face reconstruction from\nsparse-view RGB images. Unlike previous methods which are built upon 3D\nmorphable models (3DMMs) with limited details, we leverage an implicit\nrepresentation to encode rich geometric features. Our overall pipeline consists\nof two major components, including a geometry network, which learns a\ndeformable neural signed distance function (SDF) as the 3D face representation,\nand a rendering network, which learns to render on-surface points of the neural\nSDF to match the input images via self-supervised optimization. To handle\nin-the-wild sparse-view input of the same target with different expressions at\ntest time, we propose residual latent code to effectively expand the shape\nspace of the learned implicit face representation as well as a novel\nview-switch loss to enforce consistency among different views. Our experimental\nresults on several benchmark datasets demonstrate that our approach outperforms\nalternative baselines and achieves superior face reconstruction results\ncompared to state-of-the-art methods.', 'We present a minimalistic but effective neural network that computes dense\nfacial correspondences in highly unconstrained RGB images. Our network learns a\nper-pixel flow and a matchability mask between 2D input photographs of a person\nand the projection of a textured 3D face model. To train such a network, we\ngenerate a massive dataset of synthetic faces with dense labels using\nrenderings of a morphable face model with variations in pose, expressions,\nlighting, and occlusions. We found that a training refinement using real\nphotographs is required to drastically improve the ability to handle real\nimages. When combined with a facial detection and 3D face fitting step, we show\nthat our approach outperforms the state-of-the-art face alignment methods in\nterms of accuracy and speed. By directly estimating dense correspondences, we\ndo not rely on the full visibility of sparse facial landmarks and are not\nlimited to the model space of regression-based approaches. We also assess our\nmethod on video frames and demonstrate successful per-frame processing under\nextreme pose variations, occlusions, and lighting conditions. Compared to\nexisting 3D facial tracking techniques, our fitting does not rely on previous\nframes or frontal facial initialization and is robust to imperfect face\ndetections.', 'Recovery of a 3D head model including the complete face and hair regions is\nstill a challenging problem in computer vision and graphics. In this paper, we\nconsider this problem using only a few multi-view portrait images as input.\nPrevious multi-view stereo methods that have been based, either on optimization\nstrategies or deep learning techniques, suffer from low-frequency geometric\nstructures such as unclear head structures and inaccurate reconstruction in\nhair regions. To tackle this problem, we propose a prior-guided implicit neural\nrendering network. Specifically, we model the head geometry with a learnable\nsigned distance field (SDF) and optimize it via an implicit differentiable\nrenderer with the guidance of some human head priors, including the facial\nprior knowledge, head semantic segmentation information and 2D hair orientation\nmaps. The utilization of these priors can improve the reconstruction accuracy\nand robustness, leading to a high-quality integrated 3D head model. Extensive\nablation studies and comparisons with state-of-the-art methods demonstrate that\nour method can generate high-fidelity 3D head geometries with the guidance of\nthese priors.', 'Reconstructing the detailed geometric structure of a face from a given image\nis a key to many computer vision and graphics applications, such as motion\ncapture and reenactment. The reconstruction task is challenging as human faces\nvary extensively when considering expressions, poses, textures, and intrinsic\ngeometries. While many approaches tackle this complexity by using additional\ndata to reconstruct the face of a single subject, extracting facial surface\nfrom a single image remains a difficult problem. As a result, single-image\nbased methods can usually provide only a rough estimate of the facial geometry.\nIn contrast, we propose to leverage the power of convolutional neural networks\nto produce a highly detailed face reconstruction from a single image. For this\npurpose, we introduce an end-to-end CNN framework which derives the shape in a\ncoarse-to-fine fashion. The proposed architecture is composed of two main\nblocks, a network that recovers the coarse facial geometry (CoarseNet),\nfollowed by a CNN that refines the facial features of that geometry (FineNet).\nThe proposed networks are connected by a novel layer which renders a depth\nimage given a mesh in 3D. Unlike object recognition and detection problems,\nthere are no suitable datasets for training CNNs to perform face geometry\nreconstruction. Therefore, our training regime begins with a supervised phase,\nbased on synthetic images, followed by an unsupervised phase that uses only\nunconstrained facial images. The accuracy and robustness of the proposed model\nis demonstrated by both qualitative and quantitative evaluation tests.', 'We introduce the concept of unconstrained real-time 3D facial performance\ncapture through explicit semantic segmentation in the RGB input. To ensure\nrobustness, cutting edge supervised learning approaches rely on large training\ndatasets of face images captured in the wild. While impressive tracking quality\nhas been demonstrated for faces that are largely visible, any occlusion due to\nhair, accessories, or hand-to-face gestures would result in significant visual\nartifacts and loss of tracking accuracy. The modeling of occlusions has been\nmostly avoided due to its immense space of appearance variability. To address\nthis curse of high dimensionality, we perform tracking in unconstrained images\nassuming non-face regions can be fully masked out. Along with recent\nbreakthroughs in deep learning, we demonstrate that pixel-level facial\nsegmentation is possible in real-time by repurposing convolutional neural\nnetworks designed originally for general semantic segmentation. We develop an\nefficient architecture based on a two-stream deconvolution network with\ncomplementary characteristics, and introduce carefully designed training\nsamples and data augmentation strategies for improved segmentation accuracy and\nrobustness. We adopt a state-of-the-art regression-based facial tracking\nframework with segmented face images as training, and demonstrate accurate and\nuninterrupted facial performance capture in the presence of extreme occlusion\nand even side views. Furthermore, the resulting segmentation can be directly\nused to composite partial 3D face models on the input images and enable\nseamless facial manipulation tasks, such as virtual make-up or face\nreplacement.', 'State-of-the-art face super-resolution methods employ deep convolutional\nneural networks to learn a mapping between low- and high- resolution facial\npatterns by exploring local appearance knowledge. However, most of these\nmethods do not well exploit facial structures and identity information, and\nstruggle to deal with facial images that exhibit large pose variations. In this\npaper, we propose a novel face super-resolution method that explicitly\nincorporates 3D facial priors which grasp the sharp facial structures. Our work\nis the first to explore 3D morphable knowledge based on the fusion of\nparametric descriptions of face attributes (e.g., identity, facial expression,\ntexture, illumination, and face pose). Furthermore, the priors can easily be\nincorporated into any network and are extremely efficient in improving the\nperformance and accelerating the convergence speed. Firstly, a 3D face\nrendering branch is set up to obtain 3D priors of salient facial structures and\nidentity knowledge. Secondly, the Spatial Attention Module is used to better\nexploit this hierarchical information (i.e., intensity similarity, 3D facial\nstructure, and identity content) for the super-resolution problem. Extensive\nexperiments demonstrate that the proposed 3D priors achieve superior face\nsuper-resolution results over the state-of-the-arts.', 'Methods for generating synthetic data have become of increasing importance to\nbuild large datasets required for Convolution Neural Networks (CNN) based deep\nlearning techniques for a wide range of computer vision applications. In this\nwork, we extend existing methodologies to show how 2D thermal facial data can\nbe mapped to provide 3D facial models. For the proposed research work we have\nused tufts datasets for generating 3D varying face poses by using a single\nfrontal face pose. The system works by refining the existing image quality by\nperforming fusion based image preprocessing operations. The refined outputs\nhave better contrast adjustments, decreased noise level and higher exposedness\nof the dark regions. It makes the facial landmarks and temperature patterns on\nthe human face more discernible and visible when compared to original raw data.\nDifferent image quality metrics are used to compare the refined version of\nimages with original images. In the next phase of the proposed study, the\nrefined version of images is used to create 3D facial geometry structures by\nusing Convolution Neural Networks (CNN). The generated outputs are then\nimported in blender software to finally extract the 3D thermal facial outputs\nof both males and females. The same technique is also used on our thermal face\ndata acquired using prototype thermal camera (developed under Heliaus EU\nproject) in an indoor lab environment which is then used for generating\nsynthetic 3D face data along with varying yaw face angles and lastly facial\ndepth map is generated.', 'State-of-the-art methods for 3D reconstruction of faces from a single image\nrequire 2D-3D pairs of ground-truth data for supervision. Such data is costly\nto acquire, and most datasets available in the literature are restricted to\npairs for which the input 2D images depict faces in a near fronto-parallel\npose. Therefore, many data-driven methods for single-image 3D facial\nreconstruction perform poorly on profile and near-profile faces. We propose a\nmethod to improve the performance of single-image 3D facial reconstruction\nnetworks by utilizing the network to synthesize its own training data for\nfine-tuning, comprising: (i) single-image 3D reconstruction of faces in\nnear-frontal images without ground-truth 3D shape; (ii) application of a\nrigid-body transformation to the reconstructed face model; (iii) rendering of\nthe face model from new viewpoints; and (iv) use of the rendered image and\ncorresponding 3D reconstruction as additional data for supervised fine-tuning.\nThe new 2D-3D pairs thus produced have the same high-quality observed for near\nfronto-parallel reconstructions, thereby nudging the network towards more\nuniform performance as a function of the viewing angle of input faces.\nApplication of the proposed technique to the fine-tuning of a state-of-the-art\nsingle-image 3D-reconstruction network for faces demonstrates the usefulness of\nthe method, with particularly significant gains for profile or near-profile\nviews.', 'The reconstruction of dense 3D models of face geometry and appearance from a\nsingle image is highly challenging and ill-posed. To constrain the problem,\nmany approaches rely on strong priors, such as parametric face models learned\nfrom limited 3D scan data. However, prior models restrict generalization of the\ntrue diversity in facial geometry, skin reflectance and illumination. To\nalleviate this problem, we present the first approach that jointly learns 1) a\nregressor for face shape, expression, reflectance and illumination on the basis\nof 2) a concurrently learned parametric face model. Our multi-level face model\ncombines the advantage of 3D Morphable Models for regularization with the\nout-of-space generalization of a learned corrective space. We train end-to-end\non in-the-wild images without dense annotations by fusing a convolutional\nencoder with a differentiable expert-designed renderer and a self-supervised\ntraining loss, both defined at multiple detail levels. Our approach compares\nfavorably to the state-of-the-art in terms of reconstruction quality, better\ngeneralizes to real world faces, and runs at over 250 Hz.', 'Existing single view, 3D face reconstruction methods can produce beautifully\ndetailed 3D results, but typically only for near frontal, unobstructed\nviewpoints. We describe a system designed to provide detailed 3D\nreconstructions of faces viewed under extreme conditions, out of plane\nrotations, and occlusions. Motivated by the concept of bump mapping, we propose\na layered approach which decouples estimation of a global shape from its\nmid-level details (e.g., wrinkles). We estimate a coarse 3D face shape which\nacts as a foundation and then separately layer this foundation with details\nrepresented by a bump map. We show how a deep convolutional encoder-decoder can\nbe used to estimate such bump maps. We further show how this approach naturally\nextends to generate plausible details for occluded facial regions. We test our\napproach and its components extensively, quantitatively demonstrating the\ninvariance of our estimated facial details. We further provide numerous\nqualitative examples showing that our method produces detailed 3D face shapes\nin viewing conditions where existing state of the art often break down.', '3D face reconstruction is an important task in the field of computer vision.\nAlthough 3D face reconstruction has being developing rapidly in recent years,\nit is still a challenge for face reconstruction under large pose. That is\nbecause much of the information about a face in a large pose will be\nunknowable. In order to address this issue, this paper proposes a novel 3D face\nreconstruction algorithm (PIFR) based on 3D Morphable Model (3DMM). After input\na single face image, it generates a frontal image by normalizing the image.\nThen we set weighted sum of the 3D parameters of the two images. Our method\nsolves the problem of face reconstruction of a single image of a traditional\nmethod in a large pose, works on arbitrary Pose and Expressions, greatly\nimproves the accuracy of reconstruction. Experiments on the challenging AFW,\nLFPW and AFLW database show that our algorithm significantly improves the\naccuracy of 3D face reconstruction even under extreme poses .', 'With the powerfulness of convolution neural networks (CNN), CNN based face\nreconstruction has recently shown promising performance in reconstructing\ndetailed face shape from 2D face images. The success of CNN-based methods\nrelies on a large number of labeled data. The state-of-the-art synthesizes such\ndata using a coarse morphable face model, which however has difficulty to\ngenerate detailed photo-realistic images of faces (with wrinkles). This paper\npresents a novel face data generation method. Specifically, we render a large\nnumber of photo-realistic face images with different attributes based on\ninverse rendering. Furthermore, we construct a fine-detailed face image dataset\nby transferring different scales of details from one image to another. We also\nconstruct a large number of video-type adjacent frame pairs by simulating the\ndistribution of real video data. With these nicely constructed datasets, we\npropose a coarse-to-fine learning framework consisting of three convolutional\nnetworks. The networks are trained for real-time detailed 3D face\nreconstruction from monocular video as well as from a single image. Extensive\nexperimental results demonstrate that our framework can produce high-quality\nreconstruction but with much less computation time compared to the\nstate-of-the-art. Moreover, our method is robust to pose, expression and\nlighting due to the diversity of data.', 'It has been recently shown that neural networks can recover the geometric\nstructure of a face from a single given image. A common denominator of most\nexisting face geometry reconstruction methods is the restriction of the\nsolution space to some low-dimensional subspace. While such a model\nsignificantly simplifies the reconstruction problem, it is inherently limited\nin its expressiveness. As an alternative, we propose an Image-to-Image\ntranslation network that jointly maps the input image to a depth image and a\nfacial correspondence map. This explicit pixel-based mapping can then be\nutilized to provide high quality reconstructions of diverse faces under extreme\nexpressions, using a purely geometric refinement process. In the spirit of\nrecent approaches, the network is trained only with synthetic data, and is then\nevaluated on in-the-wild facial images. Both qualitative and quantitative\nanalyses demonstrate the accuracy and the robustness of our approach.', 'Recently, 3D face reconstruction and face alignment tasks are gradually\ncombined into one task: 3D dense face alignment. Its goal is to reconstruct the\n3D geometric structure of face with pose information. In this paper, we propose\na graph convolution network to regress 3D face coordinates. Our method directly\nperforms feature learning on the 3D face mesh, where the geometric structure\nand details are well preserved. Extensive experiments show that our approach\ngains superior performance over state-of-the-art methods on several challenging\ndatasets.', 'Existing face completion solutions are primarily driven by end-to-end models\nthat directly generate 2D completions of 2D masked faces. By having to\nimplicitly account for geometric and photometric variations in facial shape and\nappearance, such approaches result in unrealistic completions, especially under\nlarge variations in pose, shape, illumination and mask sizes. To alleviate\nthese limitations, we introduce 3DFaceFill, an analysis-by-synthesis approach\nfor face completion that explicitly considers the image formation process. It\ncomprises three components, (1) an encoder that disentangles the face into its\nconstituent 3D mesh, 3D pose, illumination and albedo factors, (2) an\nautoencoder that inpaints the UV representation of facial albedo, and (3) a\nrenderer that resynthesizes the completed face. By operating on the UV\nrepresentation, 3DFaceFill affords the power of correspondence and allows us to\nnaturally enforce geometrical priors (e.g. facial symmetry) more effectively.\nQuantitatively, 3DFaceFill improves the state-of-the-art by up to 4dB higher\nPSNR and 25% better LPIPS for large masks. And, qualitatively, it leads to\ndemonstrably more photorealistic face completions over a range of masks and\nocclusions while preserving consistency in global and component-wise shape,\npose, illumination and eye-gaze.', 'In the past few years, a lot of work has been done towards reconstructing the\n3D facial structure from single images by capitalizing on the power of Deep\nConvolutional Neural Networks (DCNNs). In the most recent works, differentiable\nrenderers were employed in order to learn the relationship between the facial\nidentity features and the parameters of a 3D morphable model for shape and\ntexture. The texture features either correspond to components of a linear\ntexture space or are learned by auto-encoders directly from in-the-wild images.\nIn all cases, the quality of the facial texture reconstruction of the\nstate-of-the-art methods is still not capable of modeling textures in high\nfidelity. In this paper, we take a radically different approach and harness the\npower of Generative Adversarial Networks (GANs) and DCNNs in order to\nreconstruct the facial texture and shape from single images. That is, we\nutilize GANs to train a very powerful generator of facial texture in UV space.\nThen, we revisit the original 3D Morphable Models (3DMMs) fitting approaches\nmaking use of non-linear optimization to find the optimal latent parameters\nthat best reconstruct the test image but under a new perspective. We optimize\nthe parameters with the supervision of pretrained deep identity features\nthrough our end-to-end differentiable framework. We demonstrate excellent\nresults in photorealistic and identity preserving 3D face reconstructions and\nachieve for the first time, to the best of our knowledge, facial texture\nreconstruction with high-frequency details.', ""Monocular image-based 3D reconstruction of faces is a long-standing problem\nin computer vision. Since image data is a 2D projection of a 3D face, the\nresulting depth ambiguity makes the problem ill-posed. Most existing methods\nrely on data-driven priors that are built from limited 3D face scans. In\ncontrast, we propose multi-frame video-based self-supervised training of a deep\nnetwork that (i) learns a face identity model both in shape and appearance\nwhile (ii) jointly learning to reconstruct 3D faces. Our face model is learned\nusing only corpora of in-the-wild video clips collected from the Internet. This\nvirtually endless source of training data enables learning of a highly general\n3D face model. In order to achieve this, we propose a novel multi-frame\nconsistency loss that ensures consistent shape and appearance across multiple\nframes of a subject's face, thus minimizing depth ambiguity. At test time we\ncan use an arbitrary number of frames, so that we can perform both monocular as\nwell as multi-frame reconstruction."", 'Over the last years, many face analysis tasks have accomplished astounding\nperformance, with applications including face generation and 3D face\nreconstruction from a single ""in-the-wild"" image. Nevertheless, to the best of\nour knowledge, there is no method which can produce render-ready\nhigh-resolution 3D faces from ""in-the-wild"" images and this can be attributed\nto the: (a) scarcity of available data for training, and (b) lack of robust\nmethodologies that can successfully be applied on very high-resolution data. In\nthis work, we introduce the first method that is able to reconstruct\nphotorealistic render-ready 3D facial geometry and BRDF from a single\n""in-the-wild"" image. We capture a large dataset of facial shape and\nreflectance, which we have made public. We define a fast facial photorealistic\ndifferentiable rendering methodology with accurate facial skin diffuse and\nspecular reflection, self-occlusion and subsurface scattering approximation.\nWith this, we train a network that disentangles the facial diffuse and specular\nBRDF components from a shape and texture with baked illumination, reconstructed\nwith a state-of-the-art 3DMM fitting method. Our method outperforms the\nexisting arts by a significant margin and reconstructs high-resolution 3D faces\nfrom a single low-resolution image, that can be rendered in various\napplications, and bridge the uncanny valley.', 'We present an approach to generate high fidelity 3D face avatar with a\nhigh-resolution UV texture map from a single image. To estimate the face\ngeometry, we use a deep neural network to directly predict vertex coordinates\nof the 3D face model from the given image. The 3D face geometry is further\nrefined by a non-rigid deformation process to more accurately capture facial\nlandmarks before texture projection. A key novelty of our approach is to train\nthe shape regression network on facial images synthetically generated using a\nhigh-quality rendering engine. Moreover, our shape estimator fully leverages\nthe discriminative power of deep facial identity features learned from millions\nof facial images. We have conducted extensive experiments to demonstrate the\nsuperiority of our optimized 2D-to-3D rendering approach, especially its\nexcellent generalization property on real-world selfie images. Our proposed\nsystem of rendering 3D avatars from 2D images has a wide range of applications\nfrom virtual/augmented reality (VR/AR) and telepsychiatry to human-computer\ninteraction and social networks.', 'With the increasing demands of applications in virtual reality such as 3D\nfilms, virtual Human-Machine Interactions and virtual agents, the analysis of\n3D human face analysis is considered to be more and more important as a\nfundamental step for those virtual reality tasks. Due to information provided\nby an additional dimension, 3D facial reconstruction enables aforementioned\ntasks to be achieved with higher accuracy than those based on 2D facial\nanalysis. The denser the 3D facial model is, the more information it could\nprovide. However, most existing dense 3D facial reconstruction methods require\ncomplicated processing and high system cost. To this end, this paper presents a\nnovel method that simplifies the process of dense 3D facial reconstruction by\nemploying only one frame of depth data obtained with an off-the-shelf RGB-D\nsensor. The experiments showed competitive results with real world data.', 'We present a novel method to jointly learn a 3D face parametric model and 3D\nface reconstruction from diverse sources. Previous methods usually learn 3D\nface modeling from one kind of source, such as scanned data or in-the-wild\nimages. Although 3D scanned data contain accurate geometric information of face\nshapes, the capture system is expensive and such datasets usually contain a\nsmall number of subjects. On the other hand, in-the-wild face images are easily\nobtained and there are a large number of facial images. However, facial images\ndo not contain explicit geometric information. In this paper, we propose a\nmethod to learn a unified face model from diverse sources. Besides scanned face\ndata and face images, we also utilize a large number of RGB-D images captured\nwith an iPhone X to bridge the gap between the two sources. Experimental\nresults demonstrate that with training data from more sources, we can learn a\nmore powerful face model.', 'Fast and robust three-dimensional reconstruction of facial geometric\nstructure from a single image is a challenging task with numerous applications.\nHere, we introduce a learning-based approach for reconstructing a\nthree-dimensional face from a single image. Recent face recovery methods rely\non accurate localization of key characteristic points. In contrast, the\nproposed approach is based on a Convolutional-Neural-Network (CNN) which\nextracts the face geometry directly from its image. Although such deep\narchitectures outperform other models in complex computer vision problems,\ntraining them properly requires a large dataset of annotated examples. In the\ncase of three-dimensional faces, currently, there are no large volume data\nsets, while acquiring such big-data is a tedious task. As an alternative, we\npropose to generate random, yet nearly photo-realistic, facial images for which\nthe geometric form is known. The suggested model successfully recovers facial\nshapes from real images, even for faces with extreme expressions and under\nvarious lighting conditions.', 'Recently, it was shown that excellent results can be achieved in both face\nlandmark localization and pose-invariant face recognition. These breakthroughs\nare attributed to the efforts of the community to manually annotate facial\nimages in many different poses and to collect 3D faces data. In this paper, we\npropose a novel method for joint face landmark localization and frontal face\nreconstruction (pose correction) using a small set of frontal images only. By\nobserving that the frontal facial image is the one with the minimum rank from\nall different poses we formulate an appropriate model which is able to jointly\nrecover the facial landmarks as well as the frontalized version of the face. To\nthis end, a suitable optimization problem, involving the minimization of the\nnuclear norm and the matrix $\\ell_1$ norm, is solved. The proposed method is\nassessed in frontal face reconstruction (pose correction), face landmark\nlocalization, and pose-invariant face recognition and verification by\nconducting experiments on $6$ facial images databases. The experimental results\ndemonstrate the effectiveness of the proposed method.', 'Face recognition has been of great importance in many applications as a\nbiometric for its throughput, convenience, and non-invasiveness. Recent\nadvancements in deep Convolutional Neural Network (CNN) architectures have\nboosted significantly the performance of face recognition based on\ntwo-dimensional (2D) facial texture images and outperformed the previous state\nof the art using conventional methods. However, the accuracy of 2D face\nrecognition is still challenged by the change of pose, illumination, make-up,\nand expression. On the other hand, the geometric information contained in\nthree-dimensional (3D) face data has the potential to overcome the fundamental\nlimitations of 2D face data.\n  We propose a multi-Channel deep 3D face network for face recognition based on\n3D face data. We compute the geometric information of a 3D face based on its\npiecewise-linear triangular mesh structure and then conformally flatten\ngeometric information along with the color from 3D to 2D plane to leverage the\nstate-of-the-art deep CNN architectures. We modify the input layer of the\nnetwork to take images with nine channels instead of three only such that more\ngeometric information can be explicitly fed to it. We pre-train the network\nusing images from the VGG-Face \\cite{Parkhi2015} and then fine-tune it with the\ngenerated multi-channel face images. The face recognition accuracy of the\nmulti-Channel deep 3D face network has achieved 98.6. The experimental results\nalso clearly show that the network performs much better when a 9-channel image\nis flattened to plane based on the conformal map compared with the orthographic\nprojection.', 'Recently, deep learning based 3D face reconstruction methods have shown\npromising results in both quality and efficiency.However, training deep neural\nnetworks typically requires a large volume of data, whereas face images with\nground-truth 3D face shapes are scarce. In this paper, we propose a novel deep\n3D face reconstruction approach that 1) leverages a robust, hybrid loss\nfunction for weakly-supervised learning which takes into account both low-level\nand perception-level information for supervision, and 2) performs multi-image\nface reconstruction by exploiting complementary information from different\nimages for shape aggregation. Our method is fast, accurate, and robust to\nocclusion and large pose. We provide comprehensive experiments on three\ndatasets, systematically comparing our method with fifteen recent methods and\ndemonstrating its state-of-the-art performance.', 'Facial motion retargeting is an important problem in both computer graphics\nand vision, which involves capturing the performance of a human face and\ntransferring it to another 3D character. Learning 3D morphable model (3DMM)\nparameters from 2D face images using convolutional neural networks is common in\n2D face alignment, 3D face reconstruction etc. However, existing methods either\nrequire an additional face detection step before retargeting or use a cascade\nof separate networks to perform detection followed by retargeting in a\nsequence. In this paper, we present a single end-to-end network to jointly\npredict the bounding box locations and 3DMM parameters for multiple faces.\nFirst, we design a novel multitask learning framework that learns a\ndisentangled representation of 3DMM parameters for a single face. Then, we\nleverage the trained single face model to generate ground truth 3DMM parameters\nfor multiple faces to train another network that performs joint face detection\nand motion retargeting for images with multiple faces. Experimental results\nshow that our joint detection and retargeting network has high face detection\naccuracy and is robust to extreme expressions and poses while being faster than\nstate-of-the-art methods.', 'The ""interpretation through synthesis"" approach to analyze face images,\nparticularly Active Appearance Models (AAMs) method, has become one of the most\nsuccessful face modeling approaches over the last two decades. AAM models have\nability to represent face images through synthesis using a controllable\nparameterized Principal Component Analysis (PCA) model. However, the accuracy\nand robustness of the synthesized faces of AAM are highly depended on the\ntraining sets and inherently on the generalizability of PCA subspaces. This\npaper presents a novel Deep Appearance Models (DAMs) approach, an efficient\nreplacement for AAMs, to accurately capture both shape and texture of face\nimages under large variations. In this approach, three crucial components\nrepresented in hierarchical layers are modeled using the Deep Boltzmann\nMachines (DBM) to robustly capture the variations of facial shapes and\nappearances. DAMs are therefore superior to AAMs in inferencing a\nrepresentation for new face images under various challenging conditions. The\nproposed approach is evaluated in various applications to demonstrate its\nrobustness and capabilities, i.e. facial super-resolution reconstruction,\nfacial off-angle reconstruction or face frontalization, facial occlusion\nremoval and age estimation using challenging face databases, i.e. Labeled Face\nParts in the Wild (LFPW), Helen and FG-NET. Comparing to AAMs and other deep\nlearning based approaches, the proposed DAMs achieve competitive results in\nthose applications, thus this showed their advantages in handling occlusions,\nfacial representation, and reconstruction.', 'A lot of work has been done towards reconstructing the 3D facial structure\nfrom single images by capitalizing on the power of Deep Convolutional Neural\nNetworks (DCNNs). In the recent works, the texture features either correspond\nto components of a linear texture space or are learned by auto-encoders\ndirectly from in-the-wild images. In all cases, the quality of the facial\ntexture reconstruction is still not capable of modeling facial texture with\nhigh-frequency details. In this paper, we take a radically different approach\nand harness the power of Generative Adversarial Networks (GANs) and DCNNs in\norder to reconstruct the facial texture and shape from single images. That is,\nwe utilize GANs to train a very powerful facial texture prior \\edit{from a\nlarge-scale 3D texture dataset}. Then, we revisit the original 3D Morphable\nModels (3DMMs) fitting making use of non-linear optimization to find the\noptimal latent parameters that best reconstruct the test image but under a new\nperspective. In order to be robust towards initialisation and expedite the\nfitting process, we propose a novel self-supervised regression based approach.\nWe demonstrate excellent results in photorealistic and identity preserving 3D\nface reconstructions and achieve for the first time, to the best of our\nknowledge, facial texture reconstruction with high-frequency details.', 'This paper proposes an encoder-decoder network to disentangle shape features\nduring 3D face reconstruction from single 2D images, such that the tasks of\nreconstructing accurate 3D face shapes and learning discriminative shape\nfeatures for face recognition can be accomplished simultaneously. Unlike\nexisting 3D face reconstruction methods, our proposed method directly regresses\ndense 3D face shapes from single 2D images, and tackles identity and residual\n(i.e., non-identity) components in 3D face shapes explicitly and separately\nbased on a composite 3D face shape model with latent representations. We devise\na training process for the proposed network with a joint loss measuring both\nface identification error and 3D face shape reconstruction error. To construct\ntraining data we develop a method for fitting 3D morphable model (3DMM) to\nmultiple 2D images of a subject. Comprehensive experiments have been done on\nMICC, BU3DFE, LFW and YTF databases. The results show that our method expands\nthe capacity of 3DMM for capturing discriminative shape features and facial\ndetail, and thus outperforms existing methods both in 3D face reconstruction\naccuracy and in face recognition accuracy.', 'As a classic statistical model of 3D facial shape and albedo, 3D Morphable\nModel (3DMM) is widely used in facial analysis, e.g., model fitting, image\nsynthesis. Conventional 3DMM is learned from a set of 3D face scans with\nassociated well-controlled 2D face images, and represented by two sets of PCA\nbasis functions. Due to the type and amount of training data, as well as, the\nlinear bases, the representation power of 3DMM can be limited. To address these\nproblems, this paper proposes an innovative framework to learn a nonlinear 3DMM\nmodel from a large set of in-the-wild face images, without collecting 3D face\nscans. Specifically, given a face image as input, a network encoder estimates\nthe projection, lighting, shape and albedo parameters. Two decoders serve as\nthe nonlinear 3DMM to map from the shape and albedo parameters to the 3D shape\nand albedo, respectively. With the projection parameter, lighting, 3D shape,\nand albedo, a novel analytically-differentiable rendering layer is designed to\nreconstruct the original input face. The entire network is end-to-end trainable\nwith only weak supervision. We demonstrate the superior representation power of\nour nonlinear 3DMM over its linear counterpart, and its contribution to face\nalignment, 3D reconstruction, and face editing.', 'This paper investigates the evaluation of dense 3D face reconstruction from a\nsingle 2D image in the wild. To this end, we organise a competition that\nprovides a new benchmark dataset that contains 2000 2D facial images of 135\nsubjects as well as their 3D ground truth face scans. In contrast to previous\ncompetitions or challenges, the aim of this new benchmark dataset is to\nevaluate the accuracy of a 3D dense face reconstruction algorithm using real,\naccurate and high-resolution 3D ground truth face scans. In addition to the\ndataset, we provide a standard protocol as well as a Python script for the\nevaluation. Last, we report the results obtained by three state-of-the-art 3D\nface reconstruction systems on the new benchmark dataset. The competition is\norganised along with the 2018 13th IEEE Conference on Automatic Face & Gesture\nRecognition.', 'In this paper, we address the problem of makeup transfer, which aims at\ntransplanting the makeup from the reference face to the source face while\npreserving the identity of the source. Existing makeup transfer methods have\nmade notable progress in generating realistic makeup faces, but do not perform\nwell in terms of color fidelity and spatial transformation. To tackle these\nissues, we propose a novel Facial Attribute Transformer (FAT) and its variant\nSpatial FAT for high-quality makeup transfer. Drawing inspirations from the\nTransformer in NLP, FAT is able to model the semantic correspondences and\ninteractions between the source face and reference face, and then precisely\nestimate and transfer the facial attributes. To further facilitate shape\ndeformation and transformation of facial parts, we also integrate thin plate\nsplines (TPS) into FAT, thus creating Spatial FAT, which is the first method\nthat can transfer geometric attributes in addition to color and texture.\nExtensive qualitative and quantitative experiments demonstrate the\neffectiveness and superiority of our proposed FATs in the following aspects:\n(1) ensuring high-fidelity color transfer; (2) allowing for geometric\ntransformation of facial parts; (3) handling facial variations (such as poses\nand shadows) and (4) supporting high-resolution face generation.', 'Meaningful facial parts can convey key cues for both facial action unit\ndetection and expression prediction. Textured 3D face scan can provide both\ndetailed 3D geometric shape and 2D texture appearance cues of the face which\nare beneficial for Facial Expression Recognition (FER). However, accurate\nfacial parts extraction as well as their fusion are challenging tasks. In this\npaper, a novel system for 3D FER is designed based on accurate facial parts\nextraction and deep feature fusion of facial parts. In particular, each\ntextured 3D face scan is firstly represented as a 2D texture map and a depth\nmap with one-to-one dense correspondence. Then, the facial parts of both\ntexture map and depth map are extracted using a novel 4-stage process consists\nof facial landmark localization, facial rotation correction, facial resizing,\nfacial parts bounding box extraction and post-processing procedures. Finally,\ndeep fusion Convolutional Neural Networks (CNNs) features of all facial parts\nare learned from both texture maps and depth maps, respectively and nonlinear\nSVMs are used for expression prediction. Experiments are conducted on the\nBU-3DFE database, demonstrating the effectiveness of combing different facial\nparts, texture and depth cues and reporting the state-of-the-art results in\ncomparison with all existing methods under the same setting.', 'Recently, a lot of attention has been focused on the incorporation of 3D data\ninto face analysis and its applications. Despite providing a more accurate\nrepresentation of the face, 3D facial images are more complex to acquire than\n2D pictures. As a consequence, great effort has been invested in developing\nsystems that reconstruct 3D faces from an uncalibrated 2D image. However, the\n3D-from-2D face reconstruction problem is ill-posed, thus prior knowledge is\nneeded to restrict the solutions space. In this work, we review 3D face\nreconstruction methods proposed in the last decade, focusing on those that only\nuse 2D pictures captured under uncontrolled conditions. We present a\nclassification of the proposed methods based on the technique used to add prior\nknowledge, considering three main strategies, namely, statistical model\nfitting, photometry, and deep learning, and reviewing each of them separately.\nIn addition, given the relevance of statistical 3D facial models as prior\nknowledge, we explain the construction procedure and provide a list of the most\npopular publicly available 3D facial models. After the exhaustive study of\n3D-from-2D face reconstruction approaches, we observe that the deep learning\nstrategy is rapidly growing since the last few years, becoming the standard\nchoice in replacement of the widespread statistical model fitting. Unlike the\nother two strategies, photometry-based methods have decreased in number due to\nthe need for strong underlying assumptions that limit the quality of their\nreconstructions compared to statistical model fitting and deep learning\nmethods. The review also identifies current challenges and suggests avenues for\nfuture research.', '3D face reconstruction from a single image is a challenging problem,\nespecially under partial occlusions and extreme poses. This is because the\nuncertainty of the estimated 2D landmarks will affect the quality of face\nreconstruction. In this paper, we propose a novel joint 2D and 3D optimization\nmethod to adaptively reconstruct 3D face shapes from a single image, which\ncombines the depths of 3D landmarks to solve the uncertain detections of\ninvisible landmarks. The strategy of our method involves two aspects: a\ncoarse-to-fine pose estimation using both 2D and 3D landmarks, and an adaptive\n2D and 3D re-weighting based on the refined pose parameter to recover accurate\n3D faces. Experimental results on multiple datasets demonstrate that our method\ncan generate high-quality reconstruction from a single color image and is\nrobust for self-occlusion and large poses.', '3D face reconstruction from a single image is a classical and challenging\nproblem, with wide applications in many areas. Inspired by recent works in face\nanimation from RGB-D or monocular video inputs, we develop a novel method for\nreconstructing 3D faces from unconstrained 2D images, using a coarse-to-fine\noptimization strategy. First, a smooth coarse 3D face is generated from an\nexample-based bilinear face model, by aligning the projection of 3D face\nlandmarks with 2D landmarks detected from the input image. Afterwards, using\nlocal corrective deformation fields, the coarse 3D face is refined using\nphotometric consistency constraints, resulting in a medium face shape. Finally,\na shape-from-shading method is applied on the medium face to recover fine\ngeometric details. Our method outperforms state-of-the-art approaches in terms\nof accuracy and detail recovery, which is demonstrated in extensive experiments\nusing real world models and publicly available datasets.', 'This paper presents a part-based face detection approach where the spatial\nrelationship between the face parts is represented by a hidden 3D model with\nsix parameters. The computational complexity of the search in the six\ndimensional pose space is addressed by proposing meaningful 3D pose candidates\nby image-based regression from detected face keypoint locations. The 3D pose\ncandidates are evaluated using a parameter sensitive classifier based on\ndifference features relative to the 3D pose. A compatible subset of candidates\nis then obtained by non-maximal suppression. Experiments on two standard face\ndetection datasets show that the proposed 3D model based approach obtains\nresults comparable to or better than state of the art.', 'We present a single-image 3D face synthesis technique that can handle\nchallenging facial expressions while recovering fine geometric details. Our\ntechnique employs expression analysis for proxy face geometry generation and\ncombines supervised and unsupervised learning for facial detail synthesis. On\nproxy generation, we conduct emotion prediction to determine a new\nexpression-informed proxy. On detail synthesis, we present a Deep Facial Detail\nNet (DFDN) based on Conditional Generative Adversarial Net (CGAN) that employs\nboth geometry and appearance loss functions. For geometry, we capture 366\nhigh-quality 3D scans from 122 different subjects under 3 facial expressions.\nFor appearance, we use additional 20K in-the-wild face images and apply\nimage-based rendering to accommodate lighting variations. Comprehensive\nexperiments demonstrate that our framework can produce high-quality 3D faces\nwith realistic details under challenging facial expressions.', 'We present a versatile model, FaceAnime, for various video generation tasks\nfrom still images. Video generation from a single face image is an interesting\nproblem and usually tackled by utilizing Generative Adversarial Networks (GANs)\nto integrate information from the input face image and a sequence of sparse\nfacial landmarks. However, the generated face images usually suffer from\nquality loss, image distortion, identity change, and expression mismatching due\nto the weak representation capacity of the facial landmarks. In this paper, we\npropose to ""imagine"" a face video from a single face image according to the\nreconstructed 3D face dynamics, aiming to generate a realistic and\nidentity-preserving face video, with precisely predicted pose and facial\nexpression. The 3D dynamics reveal changes of the facial expression and motion,\nand can serve as a strong prior knowledge for guiding highly realistic face\nvideo generation. In particular, we explore face video prediction and exploit a\nwell-designed 3D dynamic prediction network to predict a 3D dynamic sequence\nfor a single face image. The 3D dynamics are then further rendered by the\nsparse texture mapping algorithm to recover structural details and sparse\ntextures for generating face frames. Our model is versatile for various AR/VR\nand entertainment applications, such as face video retargeting and face video\nprediction. Superior experimental results have well demonstrated its\neffectiveness in generating high-fidelity, identity-preserving, and visually\npleasant face video clips from a single source face image.', 'Facial feature detection from facial images has attracted great attention in\nthe field of computer vision. It is a nontrivial task since the appearance and\nshape of the face tend to change under different conditions. In this paper, we\npropose a hierarchical probabilistic model that could infer the true locations\nof facial features given the image measurements even if the face is with\nsignificant facial expression and pose. The hierarchical model implicitly\ncaptures the lower level shape variations of facial components using the\nmixture model. Furthermore, in the higher level, it also learns the joint\nrelationship among facial components, the facial expression, and the pose\ninformation through automatic structure learning and parameter estimation of\nthe probabilistic model. Experimental results on benchmark databases\ndemonstrate the effectiveness of the proposed hierarchical probabilistic model.', 'With the rapid progress of deepfake techniques in recent years, facial video\nforgery can generate highly deceptive video contents and bring severe security\nthreats. And detection of such forgery videos is much more urgent and\nchallenging. Most existing detection methods treat the problem as a vanilla\nbinary classification problem. In this paper, the problem is treated as a\nspecial fine-grained classification problem since the differences between fake\nand real faces are very subtle. It is observed that most existing face forgery\nmethods left some common artifacts in the spatial domain and time domain,\nincluding generative defects in the spatial domain and inter-frame\ninconsistencies in the time domain. And a spatial-temporal model is proposed\nwhich has two components for capturing spatial and temporal forgery traces in\nglobal perspective respectively. The two components are designed using a novel\nlong distance attention mechanism. The one component of the spatial domain is\nused to capture artifacts in a single frame, and the other component of the\ntime domain is used to capture artifacts in consecutive frames. They generate\nattention maps in the form of patches. The attention method has a broader\nvision which contributes to better assembling global information and extracting\nlocal statistic information. Finally, the attention maps are used to guide the\nnetwork to focus on pivotal parts of the face, just like other fine-grained\nclassification methods. The experimental results on different public datasets\ndemonstrate that the proposed method achieves the state-of-the-art performance,\nand the proposed long distance attention method can effectively capture pivotal\nparts for face forgery.', 'Facial image super-resolution (SR) is an important preprocessing for facial\nimage analysis, face recognition, and image-based 3D face reconstruction.\nRecent convolutional neural network (CNN) based method has shown excellent\nperformance by learning mapping relation using pairs of low-resolution (LR) and\nhigh-resolution (HR) facial images. However, since the HR facial image\nreconstruction using CNN is conventionally aimed to increase the PSNR and SSIM\nmetrics, the reconstructed HR image might not be realistic even with high\nscores. An adversarial framework is proposed in this study to reconstruct the\nHR facial image by simultaneously generating an HR image with and without blur.\nFirst, the spatial resolution of the LR facial image is increased by eight\ntimes using a five-layer CNN. Then, the encoder extracts the features of the\nup-scaled image. These features are finally sent to two branches (decoders) to\ngenerate an HR facial image with and without blur. In addition, local and\nglobal discriminators are combined to focus on the reconstruction of HR facial\nstructures. Experiment results show that the proposed algorithm generates a\nrealistic HR facial image. Furthermore, the proposed method can generate a\nvariety of different facial images.', '3D face reconstruction from a single image is a task that has garnered\nincreased interest in the Computer Vision community, especially due to its\nbroad use in a number of applications such as realistic 3D avatar creation,\npose invariant face recognition and face hallucination. Since the introduction\nof the 3D Morphable Model in the late 90\'s, we witnessed an explosion of\nresearch aiming at particularly tackling this task. Nevertheless, despite the\nincreasing level of detail in the 3D face reconstructions from single images\nmainly attributed to deep learning advances, finer and highly deformable\ncomponents of the face such as the tongue are still absent from all 3D face\nmodels in the literature, although being very important for the realness of the\n3D avatar representations. In this work we present the first, to the best of\nour knowledge, end-to-end trainable pipeline that accurately reconstructs the\n3D face together with the tongue. Moreover, we make this pipeline robust in\n""in-the-wild"" images by introducing a novel GAN method tailored for 3D tongue\nsurface generation. Finally, we make publicly available to the community the\nfirst diverse tongue dataset, consisting of 1,800 raw scans of 700 individuals\nvarying in gender, age, and ethnicity backgrounds. As we demonstrate in an\nextensive series of quantitative as well as qualitative experiments, our model\nproves to be robust and realistically captures the 3D tongue structure, even in\nadverse ""in-the-wild"" conditions.', 'We present a fully automatic approach to real-time 3D face reconstruction\nfrom monocular in-the-wild videos. With the use of a cascaded-regressor based\nface tracking and a 3D Morphable Face Model shape fitting, we obtain a\nsemi-dense 3D face shape. We further use the texture information from multiple\nframes to build a holistic 3D face representation from the video frames. Our\nsystem is able to capture facial expressions and does not require any\nperson-specific training. We demonstrate the robustness of our approach on the\nchallenging 300 Videos in the Wild (300-VW) dataset. Our real-time fitting\nframework is available as an open source library at http://4dface.org.', 'Recently, data-driven single-view reconstruction methods have shown great\nprogress in modeling 3D dressed humans. However, such methods suffer heavily\nfrom depth ambiguities and occlusions inherent to single view inputs. In this\npaper, we tackle this problem by considering a small set of input views and\ninvestigate the best strategy to suitably exploit information from these views.\nWe propose a data-driven end-to-end approach that reconstructs an implicit 3D\nrepresentation of dressed humans from sparse camera views. Specifically, we\nintroduce three key components: first a spatially consistent reconstruction\nthat allows for arbitrary placement of the person in the input views using a\nperspective camera model; second an attention-based fusion layer that learns to\naggregate visual information from several viewpoints; and third a mechanism\nthat encodes local 3D patterns under the multi-view context. In the\nexperiments, we show the proposed approach outperforms the state of the art on\nstandard data both quantitatively and qualitatively. To demonstrate the\nspatially consistent reconstruction, we apply our approach to dynamic scenes.\nAdditionally, we apply our method on real data acquired with a multi-camera\nplatform and demonstrate our approach can obtain results comparable to\nmulti-view stereo with dramatically less views.', 'Face occlusions, covering either the majority or discriminative parts of the\nface, can break facial perception and produce a drastic loss of information.\nBiometric systems such as recent deep face recognition models are not immune to\nobstructions or other objects covering parts of the face. While most of the\ncurrent face recognition methods are not optimized to handle occlusions, there\nhave been a few attempts to improve robustness directly in the training stage.\nUnlike those, we propose to study the effect of generative face completion on\nthe recognition. We offer a face completion encoder-decoder, based on a\nconvolutional operator with a gating mechanism, trained with an ample set of\nface occlusions. To systematically evaluate the impact of realistic occlusions\non recognition, we propose to play the occlusion game: we render 3D objects\nonto different face parts, providing precious knowledge of what the impact is\nof effectively removing those occlusions. Extensive experiments on the Labeled\nFaces in the Wild (LFW), and its more difficult variant LFW-BLUFR, testify that\nface completion is able to partially restore face perception in machine vision\nsystems for improved recognition.', 'Recently, face super-resolution (FSR) methods either feed whole face image\ninto convolutional neural networks (CNNs) or utilize extra facial priors (e.g.,\nfacial parsing maps, facial landmarks) to focus on facial structure, thereby\nmaintaining the consistency of the facial structure while restoring facial\ndetails. However, the limited receptive fields of CNNs and inaccurate facial\npriors will reduce the naturalness and fidelity of the reconstructed face. In\nthis paper, we propose a novel paradigm based on the self-attention mechanism\n(i.e., the core of Transformer) to fully explore the representation capacity of\nthe facial structure feature. Specifically, we design a Transformer-CNN\naggregation network (TANet) consisting of two paths, in which one path uses\nCNNs responsible for restoring fine-grained facial details while the other\nutilizes a resource-friendly Transformer to capture global information by\nexploiting the long-distance visual relation modeling. By aggregating the\nfeatures from the above two paths, the consistency of global facial structure\nand fidelity of local facial detail restoration are strengthened\nsimultaneously. Experimental results of face reconstruction and recognition\nverify that the proposed method can significantly outperform the\nstate-of-the-art methods.', 'We introduce an approach that accurately reconstructs 3D human poses and\ndetailed 3D full-body geometric models from single images in realtime. The key\nidea of our approach is a novel end-to-end multi-task deep learning framework\nthat uses single images to predict five outputs simultaneously: foreground\nsegmentation mask, 2D joints positions, semantic body partitions, 3D part\norientations and uv coordinates (uv map). The multi-task network architecture\nnot only generates more visual cues for reconstruction, but also makes each\nindividual prediction more accurate. The CNN regressor is further combined with\nan optimization based algorithm for accurate kinematic pose reconstruction and\nfull-body shape modeling. We show that the realtime reconstruction reaches\naccurate fitting that has not been seen before, especially for wild images. We\ndemonstrate the results of our realtime 3D pose and human body reconstruction\nsystem on various challenging in-the-wild videos. We show the system advances\nthe frontier of 3D human body and pose reconstruction from single images by\nquantitative evaluations and comparisons with state-of-the-art methods.', 'In this work, we propose a high fidelity face swapping method, called\nHifiFace, which can well preserve the face shape of the source face and\ngenerate photo-realistic results. Unlike other existing face swapping works\nthat only use face recognition model to keep the identity similarity, we\npropose 3D shape-aware identity to control the face shape with the geometric\nsupervision from 3DMM and 3D face reconstruction method. Meanwhile, we\nintroduce the Semantic Facial Fusion module to optimize the combination of\nencoder and decoder features and make adaptive blending, which makes the\nresults more photo-realistic. Extensive experiments on faces in the wild\ndemonstrate that our method can preserve better identity, especially on the\nface shape, and can generate more photo-realistic results than previous\nstate-of-the-art methods.', '3D face reconstruction is a challenging problem but also an important task in\nthe field of computer vision and graphics. Recently, many researchers put\nattention to the problem and a large number of articles have been published.\nSingle image reconstruction is one of the branches of 3D face reconstruction,\nwhich has a lot of applications in our life. This paper is a review of the\nrecent literature on 3D face reconstruction from a single image.', 'Most 3D face reconstruction methods rely on 3D morphable models, which\ndisentangle the space of facial deformations into identity geometry,\nexpressions and skin reflectance. These models are typically learned from a\nlimited number of 3D scans and thus do not generalize well across different\nidentities and expressions. We present the first approach to learn complete 3D\nmodels of face identity geometry, albedo and expression just from images and\nvideos. The virtually endless collection of such data, in combination with our\nself-supervised learning-based approach allows for learning face models that\ngeneralize beyond the span of existing approaches. Our network design and loss\nfunctions ensure a disentangled parameterization of not only identity and\nalbedo, but also, for the first time, an expression basis. Our method also\nallows for in-the-wild monocular reconstruction at test time. We show that our\nlearned models better generalize and lead to higher quality image-based\nreconstructions than existing approaches.', '3D Morphable Model (3DMM) based methods have achieved great success in\nrecovering 3D face shapes from single-view images. However, the facial textures\nrecovered by such methods lack the fidelity as exhibited in the input images.\nRecent work demonstrates high-quality facial texture recovering with generative\nnetworks trained from a large-scale database of high-resolution UV maps of face\ntextures, which is hard to prepare and not publicly available. In this paper,\nwe introduce a method to reconstruct 3D facial shapes with high-fidelity\ntextures from single-view images in-the-wild, without the need to capture a\nlarge-scale face texture database. The main idea is to refine the initial\ntexture generated by a 3DMM based method with facial details from the input\nimage. To this end, we propose to use graph convolutional networks to\nreconstruct the detailed colors for the mesh vertices instead of reconstructing\nthe UV map. Experiments show that our method can generate high-quality results\nand outperforms state-of-the-art methods in both qualitative and quantitative\ncomparisons.', 'Face completion is a challenging generation task because it requires\ngenerating visually pleasing new pixels that are semantically consistent with\nthe unmasked face region. This paper proposes a geometry-aware Face Completion\nand Editing NETwork (FCENet) by systematically studying facial geometry from\nthe unmasked region. Firstly, a facial geometry estimator is learned to\nestimate facial landmark heatmaps and parsing maps from the unmasked face\nimage. Then, an encoder-decoder structure generator serves to complete a face\nimage and disentangle its mask areas conditioned on both the masked face image\nand the estimated facial geometry images. Besides, since low-rank property\nexists in manually labeled masks, a low-rank regularization term is imposed on\nthe disentangled masks, enforcing our completion network to manage occlusion\narea with various shape and size. Furthermore, our network can generate diverse\nresults from the same masked input by modifying estimated facial geometry,\nwhich provides a flexible mean to edit the completed face appearance. Extensive\nexperimental results qualitatively and quantitatively demonstrate that our\nnetwork is able to generate visually pleasing face completion results and edit\nface attributes as well.', 'We propose a straightforward method that simultaneously reconstructs the 3D\nfacial structure and provides dense alignment. To achieve this, we design a 2D\nrepresentation called UV position map which records the 3D shape of a complete\nface in UV space, then train a simple Convolutional Neural Network to regress\nit from a single 2D image. We also integrate a weight mask into the loss\nfunction during training to improve the performance of the network. Our method\ndoes not rely on any prior face model, and can reconstruct full facial geometry\nalong with semantic meaning. Meanwhile, our network is very light-weighted and\nspends only 9.8ms to process an image, which is extremely faster than previous\nworks. Experiments on multiple challenging datasets show that our method\nsurpasses other state-of-the-art methods on both reconstruction and alignment\ntasks by a large margin.', 'In this paper, we propose an effective face completion algorithm using a deep\ngenerative model. Different from well-studied background completion, the face\ncompletion task is more challenging as it often requires to generate\nsemantically new pixels for the missing key components (e.g., eyes and mouths)\nthat contain large appearance variations. Unlike existing nonparametric\nalgorithms that search for patches to synthesize, our algorithm directly\ngenerates contents for missing regions based on a neural network. The model is\ntrained with a combination of a reconstruction loss, two adversarial losses and\na semantic parsing loss, which ensures pixel faithfulness and local-global\ncontents consistency. With extensive experimental results, we demonstrate\nqualitatively and quantitatively that our model is able to deal with a large\narea of missing pixels in arbitrary shapes and generate realistic face\ncompletion results.', 'Video facial expression recognition is useful for many applications and\nreceived much interest lately. Although some solutions give really good results\nin a controlled environment (no occlusion), recognition in the presence of\npartial facial occlusion remains a challenging task. To handle occlusions,\nsolutions based on the reconstruction of the occluded part of the face have\nbeen proposed. These solutions are mainly based on the texture or the geometry\nof the face. However, the similarity of the face movement between different\npersons doing the same expression seems to be a real asset for the\nreconstruction. In this paper we exploit this asset and propose a new solution\nbased on an auto-encoder with skip connections to reconstruct the occluded part\nof the face in the optical flow domain. To the best of our knowledge, this is\nthe first proposition to directly reconstruct the movement for facial\nexpression recognition. We validated our approach in the controlled dataset CK+\non which different occlusions were generated. Our experiments show that the\nproposed method reduce significantly the gap, in terms of recognition accuracy,\nbetween occluded and non-occluded situations. We also compare our approach with\nexisting state-of-the-art solutions. In order to lay the basis of a\nreproducible and fair comparison in the future, we also propose a new\nexperimental protocol that includes occlusion generation and reconstruction\nevaluation.', 'Occlusions are very common in face images in the wild, leading to the\ndegraded performance of face-related tasks. Although much effort has been\ndevoted to removing occlusions from face images, the varying shapes and\ntextures of occlusions still challenge the robustness of current methods. As a\nresult, current methods either rely on manual occlusion masks or only apply to\nspecific occlusions. This paper proposes a novel face de-occlusion model based\non face segmentation and 3D face reconstruction, which automatically removes\nall kinds of face occlusions with even blurred boundaries,e.g., hairs. The\nproposed model consists of a 3D face reconstruction module, a face segmentation\nmodule, and an image generation module. With the face prior and the occlusion\nmask predicted by the first two, respectively, the image generation module can\nfaithfully recover the missing facial textures. To supervise the training, we\nfurther build a large occlusion dataset, with both manually labeled and\nsynthetic occlusions. Qualitative and quantitative results demonstrate the\neffectiveness and robustness of the proposed method.', 'The last few years have witnessed the great success of non-linear generative\nmodels in synthesizing high-quality photorealistic face images. Many recent 3D\nfacial texture reconstruction and pose manipulation from a single image\napproaches still rely on large and clean face datasets to train image-to-image\nGenerative Adversarial Networks (GANs). Yet the collection of such a large\nscale high-resolution 3D texture dataset is still very costly and difficult to\nmaintain age/ethnicity balance. Moreover, regression-based approaches suffer\nfrom generalization to the in-the-wild conditions and are unable to fine-tune\nto a target-image. In this work, we propose an unsupervised approach for\none-shot 3D facial texture completion that does not require large-scale texture\ndatasets, but rather harnesses the knowledge stored in 2D face generators. The\nproposed approach rotates an input image in 3D and fill-in the unseen regions\nby reconstructing the rotated image in a 2D face generator, based on the\nvisible parts. Finally, we stitch the most visible textures at different angles\nin the UV image-plane. Further, we frontalize the target image by projecting\nthe completed texture into the generator. The qualitative and quantitative\nexperiments demonstrate that the completed UV textures and frontalized images\nare of high quality, resembles the original identity, can be used to train a\ntexture GAN model for 3DMM fitting and improve pose-invariant face recognition.', 'Learned 3D representations of human faces are useful for computer vision\nproblems such as 3D face tracking and reconstruction from images, as well as\ngraphics applications such as character generation and animation. Traditional\nmodels learn a latent representation of a face using linear subspaces or\nhigher-order tensor generalizations. Due to this linearity, they can not\ncapture extreme deformations and non-linear expressions. To address this, we\nintroduce a versatile model that learns a non-linear representation of a face\nusing spectral convolutions on a mesh surface. We introduce mesh sampling\noperations that enable a hierarchical mesh representation that captures\nnon-linear variations in shape and expression at multiple scales within the\nmodel. In a variational setting, our model samples diverse realistic 3D faces\nfrom a multivariate Gaussian distribution. Our training data consists of 20,466\nmeshes of extreme expressions captured over 12 different subjects. Despite\nlimited training data, our trained model outperforms state-of-the-art face\nmodels with 50% lower reconstruction error, while using 75% fewer parameters.\nWe also show that, replacing the expression space of an existing\nstate-of-the-art face model with our autoencoder, achieves a lower\nreconstruction error. Our data, model and code are available at\nhttp://github.com/anuragranj/coma', 'Impressive progress has been made in audio-driven 3D facial animation\nrecently, but synthesizing 3D talking-head with rich emotion is still unsolved.\nThis is due to the lack of 3D generative models and available 3D emotional\ndataset with synchronized audios. To address this, we introduce 3D-TalkEmo, a\ndeep neural network that generates 3D talking head animation with various\nemotions. We also create a large 3D dataset with synchronized audios and\nvideos, rich corpus, as well as various emotion states of different persons\nwith the sophisticated 3D face reconstruction methods. In the emotion\ngeneration network, we propose a novel 3D face representation structure -\ngeometry map by classical multi-dimensional scaling analysis. It maps the\ncoordinates of vertices on a 3D face to a canonical image plane, while\npreserving the vertex-to-vertex geodesic distance metric in a least-square\nsense. This maintains the adjacency relationship of each vertex and holds the\neffective convolutional structure for the 3D facial surface. Taking a neutral\n3D mesh and a speech signal as inputs, the 3D-TalkEmo is able to generate vivid\nfacial animations. Moreover, it provides access to change the emotion state of\nthe animated speaker.\n  We present extensive quantitative and qualitative evaluation of our method,\nin addition to user studies, demonstrating the generated talking-heads of\nsignificantly higher quality compared to previous state-of-the-art methods.', 'Satellite multi-view stereo (MVS) imagery is particularly suited for\nlarge-scale Earth surface reconstruction. Differing from the perspective camera\nmodel (pin-hole model) that is commonly used for close-range and aerial\ncameras, the cubic rational polynomial camera (RPC) model is the mainstream\nmodel for push-broom linear-array satellite cameras. However, the homography\nwarping used in the prevailing learning based MVS methods is only applicable to\npin-hole cameras. In order to apply the SOTA learning based MVS technology to\nthe satellite MVS task for large-scale Earth surface reconstruction, RPC\nwarping should be considered. In this work, we propose, for the first time, a\nrigorous RPC warping module. The rational polynomial coefficients are recorded\nas a tensor, and the RPC warping is formulated as a series of tensor\ntransformations. Based on the RPC warping, we propose the deep learning based\nsatellite MVS (SatMVS) framework for large-scale and wide depth range Earth\nsurface reconstruction. We also introduce a large-scale satellite image dataset\nconsisting of 519 5120${\\times}$5120 images, which we call the TLC SatMVS\ndataset. The satellite images were acquired from a three-line camera (TLC) that\ncatches triple-view images simultaneously, forming a valuable supplement to the\nexisting open-source WorldView-3 datasets with single-scanline images.\nExperiments show that the proposed RPC warping module and the SatMVS framework\ncan achieve a superior reconstruction accuracy compared to the pin-hole fitting\nmethod and conventional MVS methods. Code and data are available at\nhttps://github.com/WHU-GPCV/SatMVS.', 'Deep convolutional networks (CNNs) have achieved great success in face\ncompletion to generate plausible facial structures. These methods, however, are\nlimited in maintaining global consistency among face components and recovering\nfine facial details. On the other hand, reflectional symmetry is a prominent\nproperty of face image and benefits face recognition and consistency modeling,\nyet remaining uninvestigated in deep face completion. In this work, we leverage\ntwo kinds of symmetry-enforcing subnets to form a symmetry-consistent CNN model\n(i.e., SymmFCNet) for effective face completion. For missing pixels on only one\nof the half-faces, an illumination-reweighted warping subnet is developed to\nguide the warping and illumination reweighting of the other half-face. As for\nmissing pixels on both of half-faces, we present a generative reconstruction\nsubnet together with a perceptual symmetry loss to enforce symmetry consistency\nof recovered structures. The SymmFCNet is constructed by stacking generative\nreconstruction subnet upon illumination-reweighted warping subnet, and can be\nend-to-end learned from training set of unaligned face images. Experiments show\nthat SymmFCNet can generate high quality results on images with synthetic and\nreal occlusion, and performs favorably against state-of-the-arts.', 'Hyperspectral imaging, providing abundant spatial and spectral information\nsimultaneously, has attracted a lot of interest in recent years. Unfortunately,\ndue to the hardware limitations, the hyperspectral image (HSI) is vulnerable to\nvarious degradations, such noises (random noise, HSI denoising), blurs\n(Gaussian and uniform blur, HSI deblurring), and down-sampled (both spectral\nand spatial downsample, HSI super-resolution). Previous HSI restoration methods\nare designed for one specific task only. Besides, most of them start from the\n1-D vector or 2-D matrix models and cannot fully exploit the structurally\nspectral-spatial correlation in 3-D HSI. To overcome these limitations, in this\nwork, we propose a unified low-rank tensor recovery model for comprehensive HSI\nrestoration tasks, in which non-local similarity between spectral-spatial cubic\nand spectral correlation are simultaneously captured by 3-order tensors.\nFurther, to improve the capability and flexibility, we formulate it as a\nweighted low-rank tensor recovery (WLRTR) model by treating the singular values\ndifferently, and study its analytical solution. We also consider the exclusive\nstripe noise in HSI as the gross error by extending WLRTR to robust principal\ncomponent analysis (WLRTR-RPCA). Extensive experiments demonstrate the proposed\nWLRTR models consistently outperform state-of-the-arts in typical low level\nvision HSI tasks, including denoising, destriping, deblurring and\nsuper-resolution.', 'Face is one of the most important things for communication with the world\naround us. It also forms our identity and expressions. Estimating the face\nstructure is a fundamental task in computer vision with applications in\ndifferent areas such as face recognition and medical surgeries. Recently, deep\nlearning techniques achieved significant results for 3D face reconstruction\nfrom flat images. The main challenge of such techniques is a vital need for\nlarge 3D face datasets. Usually, this challenge is handled by synthetic face\ngeneration. However, synthetic datasets suffer from the existence of\nnon-possible faces. Here, we propose a face manifold learning method for\nsynthetic diverse face dataset generation. First, the face structure is divided\ninto the shape and expression groups. Then, a fully convolutional autoencoder\nnetwork is exploited to deal with the non-possible faces, and, simultaneously,\npreserving the dataset diversity. Simulation results show that the proposed\nmethod is capable of denoising highly corrupted faces. The diversity of the\ngenerated dataset is evaluated qualitatively and quantitatively and compared to\nthe existing methods. Experiments show that our manifold learning method\noutperforms the state of the art methods significantly.', 'In this paper, we aim to address the problem of heterogeneous or\ncross-spectral face recognition using machine learning to synthesize visual\nspectrum face from infrared images. The synthesis of visual-band face images\nallows for more optimal extraction of facial features to be used for face\nidentification and/or verification. We explore the ability to use Generative\nAdversarial Networks (GANs) for face image synthesis, and examine the\nperformance of these images using pre-trained Convolutional Neural Networks\n(CNNs). The features extracted using CNNs are applied in face identification\nand verification. We explore the performance in terms of acceptance rate when\nusing various similarity measures for face verification.', 'Attribute guided face image synthesis aims to manipulate attributes on a face\nimage. Most existing methods for image-to-image translation can either perform\na fixed translation between any two image domains using a single attribute or\nrequire training data with the attributes of interest for each subject.\nTherefore, these methods could only train one specific model for each pair of\nimage domains, which limits their ability in dealing with more than two\ndomains. Another disadvantage of these methods is that they often suffer from\nthe common problem of mode collapse that degrades the quality of the generated\nimages. To overcome these shortcomings, we propose attribute guided face image\ngeneration method using a single model, which is capable to synthesize multiple\nphoto-realistic face images conditioned on the attributes of interest. In\naddition, we adopt the proposed model to increase the realism of the simulated\nface images while preserving the face characteristics. Compared to existing\nmodels, synthetic face images generated by our method present a good\nphotorealistic quality on several face datasets. Finally, we demonstrate that\ngenerated facial images can be used for synthetic data augmentation, and\nimprove the performance of the classifier used for facial expression\nrecognition.', 'An important goal in human-robot-interaction (HRI) is for machines to achieve\na close to human level of face perception. One of the important differences\nbetween machine learning and human intelligence is the lack of\ncompositionality. This paper introduces a new scheme to enable generative\nadversarial networks to learn the distribution of face images composed of\nsmaller parts. This results in a more flexible machine face perception and\neasier generalization to outside training examples. We demonstrate that this\nmodel is able to produce realistic high-quality face images by generating and\npiecing together the parts. Additionally, we demonstrate that this model learns\nthe relations between the facial parts and their distributions. Therefore, the\nspecific facial parts are interchangeable between generated face images.', '3D face reconstruction from a single 2D image is a challenging problem with\nbroad applications. Recent methods typically aim to learn a CNN-based 3D face\nmodel that regresses coefficients of 3D Morphable Model (3DMM) from 2D images\nto render 3D face reconstruction or dense face alignment. However, the shortage\nof training data with 3D annotations considerably limits performance of those\nmethods. To alleviate this issue, we propose a novel 2D-assisted\nself-supervised learning (2DASL) method that can effectively use ""in-the-wild""\n2D face images with noisy landmark information to substantially improve 3D face\nmodel learning. Specifically, taking the sparse 2D facial landmarks as\nadditional information, 2DSAL introduces four novel self-supervision schemes\nthat view the 2D landmark and 3D landmark prediction as a self-mapping process,\nincluding the 2D and 3D landmark self-prediction consistency, cycle-consistency\nover the 2D landmark prediction and self-critic over the predicted 3DMM\ncoefficients based on landmark predictions. Using these four self-supervision\nschemes, the 2DASL method significantly relieves demands on the the\nconventional paired 2D-to-3D annotations and gives much higher-quality 3D face\nmodels without requiring any additional 3D annotations. Experiments on multiple\nchallenging datasets show that our method outperforms state-of-the-arts for\nboth 3D face reconstruction and dense face alignment by a large margin.', 'In this paper, we propose a novel tensor learning and coding model for\nthird-order data completion. Our model is to learn a data-adaptive dictionary\nfrom the given observations, and determine the coding coefficients of\nthird-order tensor tubes. In the completion process, we minimize the\nlow-rankness of each tensor slice containing the coding coefficients. By\ncomparison with the traditional pre-defined transform basis, the advantages of\nthe proposed model are that (i) the dictionary can be learned based on the\ngiven data observations so that the basis can be more adaptively and accurately\nconstructed, and (ii) the low-rankness of the coding coefficients can allow the\nlinear combination of dictionary features more effectively. Also we develop a\nmulti-block proximal alternating minimization algorithm for solving such tensor\nlearning and coding model, and show that the sequence generated by the\nalgorithm can globally converge to a critical point. Extensive experimental\nresults for real data sets such as videos, hyperspectral images, and traffic\ndata are reported to demonstrate these advantages and show the performance of\nthe proposed tensor learning and coding method is significantly better than the\nother tensor completion methods in terms of several evaluation metrics.', 'Faces manifest large variations in many aspects, such as identity,\nexpression, pose, and face styling. Therefore, it is a great challenge to\ndisentangle and extract these characteristics from facial images, especially in\nan unsupervised manner. In this work, we introduce cycle-consistency in facial\ncharacteristics as free supervisory signal to learn facial representations from\nunlabeled facial images. The learning is realized by superimposing the facial\nmotion cycle-consistency and identity cycle-consistency constraints. The main\nidea of the facial motion cycle-consistency is that, given a face with\nexpression, we can perform de-expression to a neutral face via the removal of\nfacial motion and further perform re-expression to reconstruct back to the\noriginal face. The main idea of the identity cycle-consistency is to exploit\nboth de-identity into mean face by depriving the given neutral face of its\nidentity via feature re-normalization and re-identity into neutral face by\nadding the personal attributes to the mean face. At training time, our model\nlearns to disentangle two distinct facial representations to be useful for\nperforming cycle-consistent face reconstruction. At test time, we use the\nlinear protocol scheme for evaluating facial representations on various tasks,\nincluding facial expression recognition and head pose regression. We also can\ndirectly apply the learnt facial representations to person recognition,\nfrontalization and image-to-image translation. Our experiments show that the\nresults of our approach is competitive with those of existing methods,\ndemonstrating the rich and unique information embedded in the disentangled\nrepresentations. Code is available at https://github.com/JiaRenChang/FaceCycle .', '3D face reconstruction and face alignment are two fundamental and highly\nrelated topics in computer vision. Recently, some works start to use deep\nlearning models to estimate the 3DMM coefficients to reconstruct 3D face\ngeometry. However, the performance is restricted due to the limitation of the\npre-defined face templates. To address this problem, some end-to-end methods,\nwhich can completely bypass the calculation of 3DMM coefficients, are proposed\nand attract much attention. In this report, we introduce and analyse three\nstate-of-the-art methods in 3D face reconstruction and face alignment. Some\npotential improvement on PRN are proposed to further enhance its accuracy and\nspeed.', 'In this paper, we address the problem of 3D object mesh reconstruction from\nRGB videos. Our approach combines the best of multi-view geometric and\ndata-driven methods for 3D reconstruction by optimizing object meshes for\nmulti-view photometric consistency while constraining mesh deformations with a\nshape prior. We pose this as a piecewise image alignment problem for each mesh\nface projection. Our approach allows us to update shape parameters from the\nphotometric error without any depth or mask information. Moreover, we show how\nto avoid a degeneracy of zero photometric gradients via rasterizing from a\nvirtual viewpoint. We demonstrate 3D object mesh reconstruction results from\nboth synthetic and real-world videos with our photometric mesh optimization,\nwhich is unachievable with either na\\""ive mesh generation networks or\ntraditional pipelines of surface reconstruction without heavy manual\npost-processing.', 'Face images in the wild undergo large intra-personal variations, such as\nposes, illuminations, occlusions, and low resolutions, which cause great\nchallenges to face-related applications. This paper addresses this challenge by\nproposing a new deep learning framework that can recover the canonical view of\nface images. It dramatically reduces the intra-person variances, while\nmaintaining the inter-person discriminativeness. Unlike the existing face\nreconstruction methods that were either evaluated in controlled 2D environment\nor employed 3D information, our approach directly learns the transformation\nfrom the face images with a complex set of variations to their canonical views.\nAt the training stage, to avoid the costly process of labeling canonical-view\nimages from the training set by hand, we have devised a new measurement to\nautomatically select or synthesize a canonical-view image for each identity. As\nan application, this face recovery approach is used for face verification.\nFacial features are learned from the recovered canonical-view face images by\nusing a facial component-based convolutional neural network. Our approach\nachieves the state-of-the-art performance on the LFW dataset.', 'Face recognition performance has improved remarkably in the last decade. Much\nof this success can be attributed to the development of deep learning\ntechniques such as convolutional neural networks (CNNs). While CNNs have pushed\nthe state-of-the-art forward, their training process requires a large amount of\nclean and correctly labelled training data. If a CNN is intended to tolerate\nfacial pose, then we face an important question: should this training data be\ndiverse in its pose distribution, or should face images be normalized to a\nsingle pose in a pre-processing step? To address this question, we evaluate a\nnumber of popular facial landmarking and pose correction algorithms to\nunderstand their effect on facial recognition performance. Additionally, we\nintroduce a new, automatic, single-image frontalization scheme that exceeds the\nperformance of current algorithms. CNNs trained using sets of different\npre-processing methods are used to extract features from the Point and Shoot\nChallenge (PaSC) and CMU Multi-PIE datasets. We assert that the subsequent\nverification and recognition performance serves to quantify the effectiveness\nof each pose correction scheme.', 'Facial analysis has attracted much attention in the technology for\nhuman-machine interface. Different methods of classification based on sparse\nrepresentation and Gabor kernels have been widely applied in the fields of\nfacial analysis. However, most of these methods treat face from a whole view\nstandpoint. In terms of the importance of different facial views, in this\npaper, we present multi-view face analysis based on sparse representation and\nGabor wavelet coefficients. To evaluate the performance, we conduct face\nanalysis experiments including face recognition (FR) and face expression\nrecognition (FER) on JAFFE database. Experiments are conducted from two parts:\n(1) Face images are divided into three facial parts which are forehead, eye and\nmouth. (2) Face images are divided into 8 parts by the orientation of Gabor\nkernels. Experimental results demonstrate that the proposed methods can\nsignificantly boost the performance and perform better than the other methods.', ""Recovering color images and videos from highly undersampled data is a\nfundamental and challenging task in face recognition and computer vision. By\nthe multi-dimensional nature of color images and videos, in this paper, we\npropose a novel tensor completion approach, which is able to efficiently\nexplore the sparsity of tensor data under the discrete cosine transform (DCT).\nSpecifically, we introduce two ``sparse + low-rank'' tensor completion models\nas well as two implementable algorithms for finding their solutions. The first\none is a DCT-based sparse plus weighted nuclear norm induced low-rank\nminimization model. The second one is a DCT-based sparse plus $p$-shrinking\nmapping induced low-rank optimization model. Moreover, we accordingly propose\ntwo implementable augmented Lagrangian-based algorithms for solving the\nunderlying optimization models. A series of numerical experiments including\ncolor image inpainting and video data recovery demonstrate that our proposed\napproach performs better than many existing state-of-the-art tensor completion\nmethods, especially for the case when the ratio of missing data is high."", 'Face hallucination, which is the task of generating a high-resolution face\nimage from a low-resolution input image, is a well-studied problem that is\nuseful in widespread application areas. Face hallucination is particularly\nchallenging when the input face resolution is very low (e.g., 10 x 12 pixels)\nand/or the image is captured in an uncontrolled setting with large pose and\nillumination variations. In this paper, we revisit the algorithm introduced in\n[1] and present a deep interpretation of this framework that achieves\nstate-of-the-art under such challenging scenarios. In our deep network\narchitecture the global and local constraints that define a face can be\nefficiently modeled and learned end-to-end using training data. Conceptually\nour network design can be partitioned into two sub-networks: the first one\nimplements the holistic face reconstruction according to global constraints,\nand the second one enhances face-specific details and enforces local patch\nstatistics. We optimize the deep network using a new loss function for\nsuper-resolution that combines reconstruction error with a learned face quality\nmeasure in adversarial setting, producing improved visual results. We conduct\nextensive experiments in both controlled and uncontrolled setups and show that\nour algorithm improves the state of the art both numerically and visually.', 'Face parsing is a basic task in face image analysis. It amounts to labeling\neach pixel with appropriate facial parts such as eyes and nose. In the paper,\nwe present a interlinked convolutional neural network (iCNN) for solving this\nproblem in an end-to-end fashion. It consists of multiple convolutional neural\nnetworks (CNNs) taking input in different scales. A special interlinking layer\nis designed to allow the CNNs to exchange information, enabling them to\nintegrate local and contextual information efficiently. The hallmark of iCNN is\nthe extensive use of downsampling and upsampling in the interlinking layers,\nwhile traditional CNNs usually uses downsampling only. A two-stage pipeline is\nproposed for face parsing and both stages use iCNN. The first stage localizes\nfacial parts in the size-reduced image and the second stage labels the pixels\nin the identified facial parts in the original image. On a benchmark dataset we\nhave obtained better results than the state-of-the-art methods.', 'In this paper, we present a novel single shot face-related task analysis\nmethod, called Face-SSD, for detecting faces and for performing various\nface-related (classification/regression) tasks including smile recognition,\nface attribute prediction and valence-arousal estimation in the wild. Face-SSD\nuses a Fully Convolutional Neural Network (FCNN) to detect multiple faces of\ndifferent sizes and recognise/regress one or more face-related classes.\nFace-SSD has two parallel branches that share the same low-level filters, one\nbranch dealing with face detection and the other one with face analysis tasks.\nThe outputs of both branches are spatially aligned heatmaps that are produced\nin parallel - therefore Face-SSD does not require that face detection, facial\nregion extraction, size normalisation, and facial region processing are\nperformed in subsequent steps. Our contributions are threefold: 1) Face-SSD is\nthe first network to perform face analysis without relying on pre-processing\nsuch as face detection and registration in advance - Face-SSD is a simple and a\nsingle FCNN architecture simultaneously performing face detection and\nface-related task analysis - those are conventionally treated as separate\nconsecutive tasks; 2) Face-SSD is a generalised architecture that is applicable\nfor various face analysis tasks without modifying the network structure - this\nis in contrast to designing task-specific architectures; and 3) Face-SSD\nachieves real-time performance (21 FPS) even when detecting multiple faces and\nrecognising multiple classes in a given image. Experimental results show that\nFace-SSD achieves state-of-the-art performance in various face analysis tasks\nby reaching a recognition accuracy of 95.76% for smile detection, 90.29% for\nattribute prediction, and Root Mean Square (RMS) error of 0.44 and 0.39 for\nvalence and arousal estimation.', 'The COVID-19 pandemic has drastically changed accepted norms globally. Within\nthe past year, masks have been used as a public health response to limit the\nspread of the virus. This sudden change has rendered many face recognition\nbased access control, authentication and surveillance systems ineffective.\nOfficial documents such as passports, driving license and national identity\ncards are enrolled with fully uncovered face images. However, in the current\nglobal situation, face matching systems should be able to match these reference\nimages with masked face images. As an example, in an airport or security\ncheckpoint it is safer to match the unmasked image of the identifying document\nto the masked person rather than asking them to remove the mask. We find that\ncurrent facial recognition techniques are not robust to this form of occlusion.\n  To address this unique requirement presented due to the current circumstance,\nwe propose a set of re-purposed datasets and a benchmark for researchers to\nuse. We also propose a contrastive visual representation learning based\npre-training workflow which is specialized to masked vs unmasked face matching.\nWe ensure that our method learns robust features to differentiate people across\nvarying data collection scenarios. We achieve this by training over many\ndifferent datasets and validating our result by testing on various holdout\ndatasets. The specialized weights trained by our method outperform standard\nface recognition features for masked to unmasked face matching. We believe the\nprovided synthetic mask generating code, our novel training approach and the\ntrained weights from the masked face models will help in adopting existing face\nrecognition systems to operate in the current global environment. We\nopen-source all contributions for broader use by the research community.', 'Human face synthesis involves transferring knowledge about the identity and\nidentity-dependent face shape (IDFS) of a human face to target face images\nwhere the context (e.g., facial expressions, head poses, and other background\nfactors) may change dramatically. Human faces are non-rigid, so facial\nexpression leads to deformation of face shape, and head pose also affects the\nface observed in 2D images. A key challenge in face transfer is to match the\nface with unobserved new contexts, adapting the face appearance to different\nposes and expressions accordingly. In this work, we find a way to provide prior\nknowledge for generative models to reason about the appropriate appearance of a\nhuman face in response to various expressions and poses. We propose a novel\ncontext-aware face transfer method, called CarTrans, that incorporates causal\neffects of contextual factors into face representation, and thus is able to be\naware of the uncertainty of new contexts. We estimate the effect of facial\nexpression and head pose in terms of counterfactual inference by designing a\ncontrolled intervention trial, thus avoiding the requirement of a large number\nof observations to cover the pose-expression space well. Moreover, we propose a\nkernel regression-based encoder that eliminates the identity specificity of\ntarget faces when encoding contextual information from target images. The\nresulting method shows impressive performance, allowing fine-grained control\nover face shape and appearance under various contextual conditions.', 'CANDECOMP/PARAFAC (CP) tensor factorization of incomplete data is a powerful\ntechnique for tensor completion through explicitly capturing the multilinear\nlatent factors. The existing CP algorithms require the tensor rank to be\nmanually specified, however, the determination of tensor rank remains a\nchallenging problem especially for CP rank. In addition, existing approaches do\nnot take into account uncertainty information of latent factors, as well as\nmissing entries. To address these issues, we formulate CP factorization using a\nhierarchical probabilistic model and employ a fully Bayesian treatment by\nincorporating a sparsity-inducing prior over multiple latent factors and the\nappropriate hyperpriors over all hyperparameters, resulting in automatic rank\ndetermination. To learn the model, we develop an efficient deterministic\nBayesian inference algorithm, which scales linearly with data size. Our method\nis characterized as a tuning parameter-free approach, which can effectively\ninfer underlying multilinear factors with a low-rank constraint, while also\nproviding predictive distributions over missing entries. Extensive simulations\non synthetic data illustrate the intrinsic capability of our method to recover\nthe ground-truth of CP rank and prevent the overfitting problem, even when a\nlarge amount of entries are missing. Moreover, the results from real-world\napplications, including image inpainting and facial image synthesis,\ndemonstrate that our method outperforms state-of-the-art approaches for both\ntensor factorization and tensor completion in terms of predictive performance.', 'Predicting attributes from face images in the wild is a challenging computer\nvision problem. To automatically describe face attributes from face containing\nimages, traditionally one needs to cascade three technical blocks --- face\nlocalization, facial descriptor construction, and attribute classification ---\nin a pipeline. As a typical classification problem, face attribute prediction\nhas been addressed using deep learning. Current state-of-the-art performance\nwas achieved by using two cascaded Convolutional Neural Networks (CNNs), which\nwere specifically trained to learn face localization and attribute description.\nIn this paper, we experiment with an alternative way of employing the power of\ndeep representations from CNNs. Combining with conventional face localization\ntechniques, we use off-the-shelf architectures trained for face recognition to\nbuild facial descriptors. Recognizing that the describable face attributes are\ndiverse, our face descriptors are constructed from different levels of the CNNs\nfor different attributes to best facilitate face attribute prediction.\nExperiments on two large datasets, LFWA and CelebA, show that our approach is\nentirely comparable to the state-of-the-art. Our findings not only demonstrate\nan efficient face attribute prediction approach, but also raise an important\nquestion: how to leverage the power of off-the-shelf CNN representations for\nnovel tasks.', 'We introduce a deep convolutional neural networks (CNN) architecture to\nclassify facial attributes and recognize face images simultaneously via a\nshared learning paradigm to improve the accuracy for facial attribute\nprediction and face recognition performance. In this method, we use facial\nattributes as an auxiliary source of information to assist CNN features\nextracted from the face images to improve the face recognition performance.\nSpecifically, we use a shared CNN architecture that jointly predicts facial\nattributes and recognize face images simultaneously via a shared learning\nparameters, and then we use facial attribute features an an auxiliary source of\ninformation concatenated by face features to increase the discrimination of the\nCNN for face recognition. This process assists the CNN classifier to better\nrecognize face images. The experimental results show that our model increases\nboth the face recognition and facial attribute prediction performance,\nespecially for the identity attributes such as gender and race. We evaluated\nour method on several standard datasets labeled by identities and face\nattributes and the results show that the proposed method outperforms\nstate-of-the-art face recognition models.', 'Face image quality is an important factor to enable high performance face\nrecognition systems. Face quality assessment aims at estimating the suitability\nof a face image for recognition. Previous work proposed supervised solutions\nthat require artificially or human labelled quality values. However, both\nlabelling mechanisms are error-prone as they do not rely on a clear definition\nof quality and may not know the best characteristics for the utilized face\nrecognition system. Avoiding the use of inaccurate quality labels, we proposed\na novel concept to measure face quality based on an arbitrary face recognition\nmodel. By determining the embedding variations generated from random\nsubnetworks of a face model, the robustness of a sample representation and\nthus, its quality is estimated. The experiments are conducted in a\ncross-database evaluation setting on three publicly available databases. We\ncompare our proposed solution on two face embeddings against six\nstate-of-the-art approaches from academia and industry. The results show that\nour unsupervised solution outperforms all other approaches in the majority of\nthe investigated scenarios. In contrast to previous works, the proposed\nsolution shows a stable performance over all scenarios. Utilizing the deployed\nface recognition model for our face quality assessment methodology avoids the\ntraining phase completely and further outperforms all baseline approaches by a\nlarge margin. Our solution can be easily integrated into current face\nrecognition systems and can be modified to other tasks beyond face recognition.', 'State-of-the-art face recognition systems are based on deep (convolutional)\nneural networks. Therefore, it is imperative to determine to what extent face\ntemplates derived from deep networks can be inverted to obtain the original\nface image. In this paper, we study the vulnerabilities of a state-of-the-art\nface recognition system based on template reconstruction attack. We propose a\nneighborly de-convolutional neural network (\\textit{NbNet}) to reconstruct face\nimages from their deep templates. In our experiments, we assumed that no\nknowledge about the target subject and the deep network are available. To train\nthe \\textit{NbNet} reconstruction models, we augmented two benchmark face\ndatasets (VGG-Face and Multi-PIE) with a large collection of images synthesized\nusing a face generator. The proposed reconstruction was evaluated using type-I\n(comparing the reconstructed images against the original face images used to\ngenerate the deep template) and type-II (comparing the reconstructed images\nagainst a different face image of the same subject) attacks. Given the images\nreconstructed from \\textit{NbNets}, we show that for verification, we achieve\nTAR of 95.20\\% (58.05\\%) on LFW under type-I (type-II) attacks @ FAR of 0.1\\%.\nBesides, 96.58\\% (92.84\\%) of the images reconstruction from templates of\npartition \\textit{fa} (\\textit{fb}) can be identified from partition\n\\textit{fa} in color FERET. Our study demonstrates the need to secure deep\ntemplates in face recognition systems.', 'This work presents a flexible system to reconstruct 3D models of objects\ncaptured with an RGB-D sensor. A major advantage of the method is that our\nreconstruction pipeline allows the user to acquire a full 3D model of the\nobject. This is achieved by acquiring several partial 3D models in different\nsessions that are automatically merged together to reconstruct a full model. In\naddition, the 3D models acquired by our system can be directly used by\nstate-of-the-art object instance recognition and object tracking modules,\nproviding object-perception capabilities for different applications, such as\nhuman-object interaction analysis or robot grasping. The system does not impose\nconstraints in the appearance of objects (textured, untextured) nor in the\nmodelling setup (moving camera with static object or a turn-table setup). The\nproposed reconstruction system has been used to model a large number of objects\nresulting in metrically accurate and visually appealing 3D models.', 'A head-mounted display (HMD) could be an important component of augmented\nreality system. However, as the upper face region is seriously occluded by the\ndevice, the user experience could be affected in applications such as\ntelecommunication and multi-player video games. In this paper, we first present\na novel experimental setup that consists of two near-infrared (NIR) cameras to\npoint to the eye regions and one visible-light RGB camera to capture the\nvisible face region. The main purpose of this paper is to synthesize realistic\nface images without occlusions based on the images captured by these cameras.\nTo this end, we propose a novel synthesis framework that contains four modules:\n3D head reconstruction, face alignment and tracking, face synthesis, and eye\nsynthesis. In face synthesis, we propose a novel algorithm that can robustly\nalign and track a personalized 3D head model given a face that is severely\noccluded by the HMD. In eye synthesis, in order to generate accurate eye\nmovements and dynamic wrinkle variations around eye regions, we propose another\nnovel algorithm to colorize the NIR eye images and further remove the ""red eye""\neffects caused by the colorization. Results show that both hardware setup and\nsystem framework are robust to synthesize realistic face images in video\nsequences.', '3D reconstruction from 2D images is a central problem in computer vision.\nRecent works have been focusing on reconstruction directly from a single image.\nIt is well known however that only one image cannot provide enough information\nfor such a reconstruction. A prior knowledge that has been entertained are 3D\nCAD models due to its online ubiquity. A fundamental question is how to\ncompactly represent millions of CAD models while allowing generalization to new\nunseen objects with fine-scaled geometry. We introduce an approach to compactly\nrepresent a 3D mesh. Our method first selects a 3D model from a graph structure\nby using a novel free-form deformation FFD 3D-2D registration, and then the\nselected 3D model is refined to best fit the image silhouette. We perform a\ncomprehensive quantitative and qualitative analysis that demonstrates\nimpressive dense and realistic 3D reconstruction from single images.', 'Deep learning-based hyperspectral image super-resolution (SR) methods have\nachieved great success recently. However, most existing models can not\neffectively explore spatial information and spectral information between bands\nsimultaneously, obtaining relatively low performance. To address this issue, in\nthis paper, we propose a novel spectral-spatial residual network for\nhyperspectral image super-resolution (SSRNet). Our method can effectively\nexplore spatial-spectral information by using 3D convolution instead of 2D\nconvolution, which enables the network to better extract potential information.\nFurthermore, we design a spectral-spatial residual module (SSRM) to adaptively\nlearn more effective features from all the hierarchical features in units\nthrough local feature fusion, significantly improving the performance of the\nalgorithm. In each unit, we employ spatial and temporal separable 3D\nconvolution to extract spatial and spectral information, which not only reduces\nunaffordable memory usage and high computational cost, but also makes the\nnetwork easier to train. Extensive evaluations and comparisons on three\nbenchmark datasets demonstrate that the proposed approach achieves superior\nperformance in comparison to existing state-of-the-art methods.', 'With the recent success of deep neural networks, remarkable progress has been\nachieved on face recognition. However, collecting large-scale real-world\ntraining data for face recognition has turned out to be challenging, especially\ndue to the label noise and privacy issues. Meanwhile, existing face recognition\ndatasets are usually collected from web images, lacking detailed annotations on\nattributes (e.g., pose and expression), so the influences of different\nattributes on face recognition have been poorly investigated. In this paper, we\naddress the above-mentioned issues in face recognition using synthetic face\nimages, i.e., SynFace. Specifically, we first explore the performance gap\nbetween recent state-of-the-art face recognition models trained with synthetic\nand real face images. We then analyze the underlying causes behind the\nperformance gap, e.g., the poor intra-class variations and the domain gap\nbetween synthetic and real face images. Inspired by this, we devise the SynFace\nwith identity mixup (IM) and domain mixup (DM) to mitigate the above\nperformance gap, demonstrating the great potentials of synthetic data for face\nrecognition. Furthermore, with the controllable face synthesis model, we can\neasily manage different factors of synthetic face generation, including pose,\nexpression, illumination, the number of identities, and samples per identity.\nTherefore, we also perform a systematically empirical analysis on synthetic\nface images to provide some insights on how to effectively utilize synthetic\ndata for face recognition.', 'Face recognition remains a hot topic in computer vision, and it is\nchallenging to tackle the problem that both the training and testing images are\ncorrupted. In this paper, we propose a novel semi-supervised method based on\nthe theory of the low-rank matrix recovery for face recognition, which can\nsimultaneously learn discriminative low-rank and sparse representations for\nboth training and testing images. To this end, a correlation penalty term is\nintroduced into the formulation of our proposed method to learn an incoherent\ndictionary. Experimental results on several face image databases demonstrate\nthe effectiveness of our method, i.e., the proposed method is robust to the\nillumination, expression and pose variations, as well as images with noises\nsuch as block occlusion or uniform noises.', 'Prior knowledge of face shape and structure plays an important role in face\ninpainting. However, traditional face inpainting methods mainly focus on the\ngenerated image resolution of the missing portion without consideration of the\nspecial particularities of the human face explicitly and generally produce\ndiscordant facial parts. To solve this problem, we present a domain embedded\nmulti-model generative adversarial model for inpainting of face images with\nlarge cropped regions. We firstly represent only face regions using the latent\nvariable as the domain knowledge and combine it with the non-face parts\ntextures to generate high-quality face images with plausible contents. Two\nadversarial discriminators are finally used to judge whether the generated\ndistribution is close to the real distribution or not. It can not only\nsynthesize novel image structures but also explicitly utilize the embedded face\ndomain knowledge to generate better predictions with consistency on structures\nand appearance. Experiments on both CelebA and CelebA-HQ face datasets\ndemonstrate that our proposed approach achieved state-of-the-art performance\nand generates higher quality inpainting results than existing ones.', 'Face detection and tracking in low resolution image is not a trivial task due\nto the limitation in the appearance features for face characterization.\nMoreover, facial expression gives additional distortion on this small and noisy\nface. In this paper, we propose deformable symmetric Gabor wavelet network face\nmodel for face detection in low resolution image. Our model optimizes the\nrotation, translation, dilation, perspective and partial deformation amount of\nthe face model with symmetry constraints. Symmetry constraints help our model\nto be more robust to noise and distortion. Experimental results on our low\nresolution face image dataset and videos show promising face detection and\ntracking results under various challenging conditions.', 'Face liveness detection is an essential prerequisite for face recognition\napplications. Previous face liveness detection methods usually train a binary\nclassifier to differentiate between a fake face and a real face before face\nrecognition. The client identity information is not utilized in previous face\nliveness detection methods. However, in practical face recognition\napplications, face spoofing attacks are always aimed at a specific client, and\nthe client identity information can provide useful clues for face liveness\ndetection. In this paper, we propose a face liveness detection method based on\nthe client identity using Siamese network. We detect face liveness after face\nrecognition instead of before face recognition, that is, we detect face\nliveness with the client identity information. We train a Siamese network with\nimage pairs. Each image pair consists of two real face images or one real and\none fake face images. The face images in each pair come from a same client.\nGiven a test face image, the face image is firstly recognized by face\nrecognition system, then the real face image of the identified client is\nretrieved to help the face liveness detection. Experiment results demonstrate\nthe effectiveness of our method.', 'Robust face detection is one of the most important pre-processing steps to\nsupport facial expression analysis, facial landmarking, face recognition, pose\nestimation, building of 3D facial models, etc. Although this topic has been\nintensely studied for decades, it is still challenging due to numerous variants\nof face images in real-world scenarios. In this paper, we present a novel\napproach named Multiple Scale Faster Region-based Convolutional Neural Network\n(MS-FRCNN) to robustly detect human facial regions from images collected under\nvarious challenging conditions, e.g. large occlusions, extremely low\nresolutions, facial expressions, strong illumination variations, etc. The\nproposed approach is benchmarked on two challenging face detection databases,\ni.e. the Wider Face database and the Face Detection Dataset and Benchmark\n(FDDB), and compared against recent other face detection methods, e.g.\nTwo-stage CNN, Multi-scale Cascade CNN, Faceness, Aggregate Chanel Features,\nHeadHunter, Multi-view Face Detection, Cascade CNN, etc. The experimental\nresults show that our proposed approach consistently achieves highly\ncompetitive results with the state-of-the-art performance against other recent\nface detection methods.', 'Due to the COVID-19 situation, face masks have become a main part of our\ndaily life. Wearing mouth-and-nose protection has been made a mandate in many\npublic places, to prevent the spread of the COVID-19 virus. However, face masks\naffect the performance of face recognition, since a large area of the face is\ncovered. The effect of wearing a face mask on the different components of the\nface recognition system in a collaborative environment is a problem that is\nstill to be fully studied. This work studies, for the first time, the effect of\nwearing a face mask on face image quality by utilising state-of-the-art face\nimage quality assessment methods of different natures. This aims at providing\nbetter understanding on the effect of face masks on the operation of face\nrecognition as a whole system. In addition, we further studied the effect of\nsimulated masks on face image utility in comparison to real face masks. We\ndiscuss the correlation between the mask effect on face image quality and that\non the face verification performance by automatic systems and human experts,\nindicating a consistent trend between both factors. The evaluation is conducted\non the database containing (1) no-masked faces, (2) real face masks, and (3)\nsimulated face masks, by synthetically generating digital facial masks on\nno-masked faces. Finally, a visual interpretation of the face areas\ncontributing to the quality score of a selected set of quality assessment\nmethods is provided to give a deeper insight into the difference of network\ndecisions in masked and non-masked faces, among other variations.', 'With the availability of more powerful computers, iterative reconstruction\nalgorithms are the subject of an ongoing work in the design of more efficient\nreconstruction algorithms for X-ray computed tomography. In this work, we show\nhow two analytical reconstruction algorithms can be improved by correcting the\ncorresponding reconstructions using a randomized iterative reconstruction\nalgorithm. The combined analytical reconstruction followed by randomized\niterative reconstruction can also be viewed as a reconstruction algorithm\nwhich, in the experiments we have conducted, uses up to $35\\%$ less projection\nangles as compared to the analytical reconstruction algorithms and produces the\nsame results in terms of quality of reconstruction, without increasing the\nexecution time significantly.', 'Face recognition is a long standing challenge in the field of Artificial\nIntelligence (AI). The goal is to create systems that accurately detect,\nrecognize, verify, and understand human faces. There are significant technical\nhurdles in making these systems accurate, particularly in unconstrained\nsettings due to confounding factors related to pose, resolution, illumination,\nocclusion, and viewpoint. However, with recent advances in neural networks,\nface recognition has achieved unprecedented accuracy, largely built on\ndata-driven deep learning methods. While this is encouraging, a critical aspect\nthat is limiting facial recognition accuracy and fairness is inherent facial\ndiversity. Every face is different. Every face reflects something unique about\nus. Aspects of our heritage - including race, ethnicity, culture, geography -\nand our individual identify - age, gender, and other visible manifestations of\nself-expression, are reflected in our faces. We expect face recognition to work\nequally accurately for every face. Face recognition needs to be fair. As we\nrely on data-driven methods to create face recognition technology, we need to\nensure necessary balance and coverage in training data. However, there are\nstill scientific questions about how to represent and extract pertinent facial\nfeatures and quantitatively measure facial diversity. Towards this goal,\nDiversity in Faces (DiF) provides a data set of one million annotated human\nface images for advancing the study of facial diversity. The annotations are\ngenerated using ten well-established facial coding schemes from the scientific\nliterature. The facial coding schemes provide human-interpretable quantitative\nmeasures of facial features. We believe that by making the extracted coding\nschemes available on a large set of faces, we can accelerate research and\ndevelopment towards creating more fair and accurate facial recognition systems.', 'Face image quality can be defined as a measure of the utility of a face image\nto automatic face recognition. In this work, we propose (and compare) two\nmethods for automatic face image quality based on target face quality values\nfrom (i) human assessments of face image quality (matcher-independent), and\n(ii) quality values computed from similarity scores (matcher-dependent). A\nsupport vector regression model trained on face features extracted using a deep\nconvolutional neural network (ConvNet) is used to predict the quality of a face\nimage. The proposed methods are evaluated on two unconstrained face image\ndatabases, LFW and IJB-A, which both contain facial variations with multiple\nquality factors. Evaluation of the proposed automatic face image quality\nmeasures shows we are able to reduce the FNMR at 1% FMR by at least 13% for two\nface matchers (a COTS matcher and a ConvNet matcher) by using the proposed face\nquality to select subsets of face images and video frames for matching\ntemplates (i.e., multiple faces per subject) in the IJB-A protocol. To our\nknowledge, this is the first work to utilize human assessments of face image\nquality in designing a predictor of unconstrained face quality that is shown to\nbe effective in cross-database evaluation.', 'The widespread deployment of surveillance cameras for facial recognition\ngives rise to many privacy concerns. This study proposes a privacy-friendly\nalternative to large scale facial recognition. While there are multiple\ntechniques to preserve privacy, our work is based on the minimization principle\nwhich implies minimizing the amount of collected personal data. Instead of\nrunning facial recognition software on all video data, we propose to\nautomatically extract a high quality snapshot of each detected person without\nrevealing his or her identity. This snapshot is then encrypted and access is\nonly granted after legal authorization. We introduce a novel unsupervised face\nimage quality assessment method which is used to select the high quality\nsnapshots. For this, we train a variational autoencoder on high quality face\nimages from a publicly available dataset and use the reconstruction probability\nas a metric to estimate the quality of each face crop. We experimentally\nconfirm that the reconstruction probability can be used as biometric quality\npredictor. Unlike most previous studies, we do not rely on a manually defined\nface quality metric as everything is learned from data. Our face quality\nassessment method outperforms supervised, unsupervised and general image\nquality assessment methods on the task of improving face verification\nperformance by rejecting low quality images. The effectiveness of the whole\nsystem is validated qualitatively on still images and videos.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am seeking alternatives to Generative Adversarial Networks (GANs) that can be applied to image datasets, such as CIFAR-10. The alternative should be capable of generating new data points based on the original data distribution and should perform comparably to GANs across various metrics. Could you provide information on the standard metrics typically used to evaluate the performance of GANs? I anticipate that this alternative method would initially estimate and model the original data distribution, possibly using a neural network, and then generate diverse data points that adhere to the same distribution through an intelligent sampling technique. However, I am open to learning about other promising approaches as well.","['This paper presents a methodology and workflow that overcome the limitations\nof the conventional Generative Adversarial Networks (GANs) for geological\nfacies modeling. It attempts to improve the training stability and guarantee\nthe diversity of the generated geology through interpretable latent vectors.\nThe resulting samples are ensured to have the equal probability (or an unbiased\ndistribution) as from the training dataset. This is critical when applying GANs\nto generate unbiased and representative geological models that can be further\nused to facilitate objective uncertainty evaluation and optimal decision-making\nin oil field exploration and development.\n  We proposed and implemented a new variant of GANs called Info-WGAN for the\ngeological facies modeling that combines Information Maximizing Generative\nAdversarial Network (InfoGAN) with Wasserstein distance and Gradient Penalty\n(GP) for learning interpretable latent codes as well as generating stable and\nunbiased distribution from the training data. Different from the original GAN\ndesign, InfoGAN can use the training images with full, partial, or no labels to\nperform disentanglement of the complex sedimentary types exhibited in the\ntraining dataset to achieve the variety and diversity of the generated samples.\nThis is accomplished by adding additional categorical variables that provide\ndisentangled semantic representations besides the mere randomized latent vector\nused in the original GANs. By such means, a regularization term is used to\nmaximize the mutual information between such latent categorical codes and the\ngenerated geological facies in the loss function.\n  Furthermore, the resulting unbiased sampling by Info-WGAN makes the data\nconditioning much easier than the conventional GANs in geological modeling\nbecause of the variety and diversity as well as the equal probability of the\nunconditional sampling by the generator.', 'We introduce a new generative model where samples are produced via Langevin\ndynamics using gradients of the data distribution estimated with score\nmatching. Because gradients can be ill-defined and hard to estimate when the\ndata resides on low-dimensional manifolds, we perturb the data with different\nlevels of Gaussian noise, and jointly estimate the corresponding scores, i.e.,\nthe vector fields of gradients of the perturbed data distribution for all noise\nlevels. For sampling, we propose an annealed Langevin dynamics where we use\ngradients corresponding to gradually decreasing noise levels as the sampling\nprocess gets closer to the data manifold. Our framework allows flexible model\narchitectures, requires no sampling during training or the use of adversarial\nmethods, and provides a learning objective that can be used for principled\nmodel comparisons. Our models produce samples comparable to GANs on MNIST,\nCelebA and CIFAR-10 datasets, achieving a new state-of-the-art inception score\nof 8.87 on CIFAR-10. Additionally, we demonstrate that our models learn\neffective representations via image inpainting experiments.', 'Generative models are undoubtedly a hot topic in Artificial Intelligence,\namong which the most common type is Generative Adversarial Networks (GANs).\nThese architectures let one synthesise artificial datasets by implicitly\nmodelling the underlying probability distribution of a real-world training\ndataset. With the introduction of Conditional GANs and their variants, these\nmethods were extended to generating samples conditioned on ancillary\ninformation available for each sample within the dataset. From a practical\nstandpoint, however, one might desire to generate data conditioned on partial\ninformation. That is, only a subset of the ancillary conditioning variables\nmight be of interest when synthesising data. In this work, we argue that\nstandard Conditional GANs are not suitable for such a task and propose a new\nAdversarial Network architecture and training strategy to deal with the ensuing\nproblems. Experiments illustrating the value of the proposed approach in digit\nand face image synthesis under partial conditioning information are presented,\nshowing that the proposed method can effectively outperform the standard\napproach under these circumstances.', 'We consider inpainting in an unsupervised setting where there is neither\naccess to paired nor unpaired training data. The only available information is\nprovided by the uncomplete observations and the inpainting process statistics.\nIn this context, an observation should give rise to several plausible\nreconstructions which amounts at learning a distribution over the space of\nreconstructed images. We model the reconstruction process by using a\nconditional GAN with constraints on the stochastic component that introduce an\nexplicit dependency between this component and the generated output. This\nallows us sampling from the latent component in order to generate a\ndistribution of images associated to an observation. We demonstrate the\ncapacity of our model on several image datasets: faces (CelebA), food images\n(Recipe-1M) and bedrooms (LSUN Bedrooms) with different types of imputation\nmasks. The approach yields comparable performance to model variants trained\nwith additional supervision.', 'In recent years, research on image generation methods has been developing\nfast. The auto-encoding variational Bayes method (VAEs) was proposed in 2013,\nwhich uses variational inference to learn a latent space from the image\ndatabase and then generates images using the decoder. The generative\nadversarial networks (GANs) came out as a promising framework, which uses\nadversarial training to improve the generative ability of the generator.\nHowever, the generated pictures by GANs are generally blurry. The deep\nconvolutional generative adversarial networks (DCGANs) were then proposed to\nleverage the quality of generated images. Since the input noise vectors are\nrandomly sampled from a Gaussian distribution, the generator has to map from a\nwhole normal distribution to the images. This makes DCGANs unable to reflect\nthe inherent structure of the training data. In this paper, we propose a novel\ndeep model, called generative adversarial networks with decoder-encoder output\nnoise (DE-GANs), which takes advantage of both the adversarial training and the\nvariational Bayesain inference to improve the performance of image generation.\nDE-GANs use a pre-trained decoder-encoder architecture to map the random\nGaussian noise vectors to informative ones and pass them to the generator of\nthe adversarial networks. Since the decoder-encoder architecture is trained by\nthe same images as the generators, the output vectors could carry the intrinsic\ndistribution information of the original images. Moreover, the loss function of\nDE-GANs is different from GANs and DCGANs. A hidden-space loss function is\nadded to the adversarial loss function to enhance the robustness of the model.\nExtensive empirical results show that DE-GANs can accelerate the convergence of\nthe adversarial training process and improve the quality of the generated\nimages.', 'Deep generative models (DGMs) of images are now sufficiently mature that they\nproduce nearly photorealistic samples and obtain scores similar to the data\ndistribution on heuristics such as Frechet Inception Distance (FID). These\nresults, especially on large-scale datasets such as ImageNet, suggest that DGMs\nare learning the data distribution in a perceptually meaningful space and can\nbe used in downstream tasks. To test this latter hypothesis, we use\nclass-conditional generative models from a number of model\nclasses---variational autoencoders, autoregressive models, and generative\nadversarial networks (GANs)---to infer the class labels of real data. We\nperform this inference by training an image classifier using only synthetic\ndata and using the classifier to predict labels on real data. The performance\non this task, which we call Classification Accuracy Score (CAS), reveals some\nsurprising results not identified by traditional metrics and constitute our\ncontributions. First, when using a state-of-the-art GAN (BigGAN-deep), Top-1\nand Top-5 accuracy decrease by 27.9\\% and 41.6\\%, respectively, compared to the\noriginal data; and conditional generative models from other model classes, such\nas Vector-Quantized Variational Autoencoder-2 (VQ-VAE-2) and Hierarchical\nAutoregressive Models (HAMs), substantially outperform GANs on this benchmark.\nSecond, CAS automatically surfaces particular classes for which generative\nmodels failed to capture the data distribution, and were previously unknown in\nthe literature. Third, we find traditional GAN metrics such as Inception Score\n(IS) and FID neither predictive of CAS nor useful when evaluating non-GAN\nmodels. Furthermore, in order to facilitate better diagnoses of generative\nmodels, we open-source the proposed metric.', 'Recent advances in deep learning led to novel generative modeling techniques\nthat achieve unprecedented quality in generated samples and performance in\nlearning complex distributions in imaging data. These new models in medical\nimage computing have important applications that form clinically relevant and\nvery challenging unsupervised learning problems. In this paper, we explore the\nfeasibility of using state-of-the-art auto-encoder-based deep generative\nmodels, such as variational and adversarial auto-encoders, for one such task:\nabnormality detection in medical imaging. We utilize typical, publicly\navailable datasets with brain scans from healthy subjects and patients with\nstroke lesions and brain tumors. We use the data from healthy subjects to train\ndifferent auto-encoder based models to learn the distribution of healthy images\nand detect pathologies as outliers. Models that can better learn the data\ndistribution should be able to detect outliers more accurately. We evaluate the\ndetection performance of deep generative models and compare them with non-deep\nlearning based approaches to provide a benchmark of the current state of\nresearch. We conclude that abnormality detection is a challenging task for deep\ngenerative models and large room exists for improvement. In order to facilitate\nfurther research, we aim to provide carefully pre-processed imaging data\navailable to the research community.', 'A learned generative model often produces biased statistics relative to the\nunderlying data distribution. A standard technique to correct this bias is\nimportance sampling, where samples from the model are weighted by the\nlikelihood ratio under model and true distributions. When the likelihood ratio\nis unknown, it can be estimated by training a probabilistic classifier to\ndistinguish samples from the two distributions. We employ this likelihood-free\nimportance weighting method to correct for the bias in generative models. We\nfind that this technique consistently improves standard goodness-of-fit metrics\nfor evaluating the sample quality of state-of-the-art deep generative models,\nsuggesting reduced bias. Finally, we demonstrate its utility on representative\napplications in a) data augmentation for classification using generative\nadversarial networks, and b) model-based policy evaluation using off-policy\ndata.', ""Generative adversarial networks (GANs) have attracted intense interest in the\nfield of generative models. However, few investigations focusing either on the\ntheoretical analysis or on algorithm design for the approximation ability of\nthe generator of GANs have been reported. This paper will first theoretically\nanalyze GANs' approximation property. Similar to the universal approximation\nproperty of the fully connected neural networks with one hidden layer, we prove\nthat the generator with the input latent variable in GANs can universally\napproximate the potential data distribution given the increasing hidden\nneurons. Furthermore, we propose an approach named stochastic data generation\n(SDG) to enhance GANs'approximation ability. Our approach is based on the\nsimple idea of imposing randomness through data generation in GANs by a prior\ndistribution on the conditional probability between the layers. SDG approach\ncan be easily implemented by using the reparameterization trick. The\nexperimental results on synthetic dataset verify the improved approximation\nability obtained by this SDG approach. In the practical dataset, four GANs\nusing SDG can also outperform the corresponding traditional GANs when the model\narchitectures are smaller."", 'Generative models for deep learning are promising both to improve\nunderstanding of the model, and yield training methods requiring fewer labeled\nsamples.\n  Recent works use generative model approaches to produce the deep net\'s input\ngiven the value of a hidden layer several levels above. However, there is no\naccompanying ""proof of correctness"" for the generative model, showing that the\nfeedforward deep net is the correct inference method for recovering the hidden\nlayer given the input. Furthermore, these models are complicated.\n  The current paper takes a more theoretical tack. It presents a very simple\ngenerative model for RELU deep nets, with the following characteristics: (i)\nThe generative model is just the reverse of the feedforward net: if the forward\ntransformation at a layer is $A$ then the reverse transformation is $A^T$.\n(This can be seen as an explanation of the old weight tying idea for denoising\nautoencoders.) (ii) Its correctness can be proven under a clean theoretical\nassumption: the edge weights in real-life deep nets behave like random numbers.\nUnder this assumption ---which is experimentally tested on real-life nets like\nAlexNet--- it is formally proved that feed forward net is a correct inference\nmethod for recovering the hidden layer.\n  The generative model suggests a simple modification for training: use the\ngenerative model to produce synthetic data with labels and include it in the\ntraining set. Experiments are shown to support this theory of random-like deep\nnets; and that it helps the training.', 'Multimodal generative models should be able to learn a meaningful latent\nrepresentation that enables a coherent joint generation of all modalities\n(e.g., images and text). Many applications also require the ability to\naccurately sample modalities conditioned on observations of a subset of the\nmodalities. Often not all modalities may be observed for all training data\npoints, so semi-supervised learning should be possible. In this study, we\npropose a novel product-of-experts (PoE) based variational autoencoder that\nhave these desired properties. We benchmark it against a mixture-of-experts\n(MoE) approach and an approach of combining the modalities with an additional\nencoder network. An empirical evaluation shows that the PoE based models can\noutperform the contrasted models. Our experiments support the intuition that\nPoE models are more suited for a conjunctive combination of modalities.', 'Insufficient training data is a major bottleneck for most deep learning\npractices, not least in medical imaging where data is difficult to collect and\npublicly available datasets are scarce due to ethics and privacy. This work\ninvestigates the use of synthetic images, created by generative adversarial\nnetworks (GANs), as the only source of training data. We demonstrate that for\nthis application, it is of great importance to make use of multiple GANs to\nimprove the diversity of the generated data, i.e. to sufficiently cover the\ndata distribution. While a single GAN can generate seemingly diverse image\ncontent, training on this data in most cases lead to severe over-fitting. We\ntest the impact of ensembled GANs on synthetic 2D data as well as common image\ndatasets (SVHN and CIFAR-10), and using both DCGANs and progressively growing\nGANs. As a specific use case, we focus on synthesizing digital pathology\npatches to provide anonymized training data.', 'Deep neural network approaches to inverse imaging problems have produced\nimpressive results in the last few years. In this paper, we consider the use of\ngenerative models in a variational regularisation approach to inverse problems.\nThe considered regularisers penalise images that are far from the range of a\ngenerative model that has learned to produce images similar to a training\ndataset. We name this family \\textit{generative regularisers}. The success of\ngenerative regularisers depends on the quality of the generative model and so\nwe propose a set of desired criteria to assess generative models and guide\nfuture research. In our numerical experiments, we evaluate three common\ngenerative models, autoencoders, variational autoencoders and generative\nadversarial networks, against our desired criteria. We also test three\ndifferent generative regularisers on the inverse problems of deblurring,\ndeconvolution, and tomography. We show that restricting solutions of the\ninverse problem to lie exactly in the range of a generative model can give good\nresults but that allowing small deviations from the range of the generator\nproduces more consistent results.', 'Combining Generative Adversarial Networks (GANs) with encoders that learn to\nencode data points has shown promising results in learning data representations\nin an unsupervised way. We propose a framework that combines an encoder and a\ngenerator to learn disentangled representations which encode meaningful\ninformation about the data distribution without the need for any labels. While\ncurrent approaches focus mostly on the generative aspects of GANs, our\nframework can be used to perform inference on both real and generated data\npoints. Experiments on several data sets show that the encoder learns\ninterpretable, disentangled representations which encode descriptive properties\nand can be used to sample images that exhibit specific characteristics.', 'As a new approach to train generative models, \\emph{generative adversarial\nnetworks} (GANs) have achieved considerable success in image generation. This\nframework has also recently been applied to data with graph structures. We\npropose labeled-graph generative adversarial networks (LGGAN) to train deep\ngenerative models for graph-structured data with node labels. We test the\napproach on various types of graph datasets, such as collections of citation\nnetworks and protein graphs. Experiment results show that our model can\ngenerate diverse labeled graphs that match the structural characteristics of\nthe training data and outperforms all alternative approaches in quality and\ngenerality. To further evaluate the quality of the generated graphs, we use\nthem on a downstream task of graph classification, and the results show that\nLGGAN can faithfully capture the important aspects of the graph structure.', 'Generative models are known to be difficult to assess. Recent works,\nespecially on generative adversarial networks (GANs), produce good visual\nsamples of varied categories of images. However, the validation of their\nquality is still difficult to define and there is no existing agreement on the\nbest evaluation process. This paper aims at making a step toward an objective\nevaluation process for generative models. It presents a new method to assess a\ntrained generative model by evaluating the test accuracy of a classifier\ntrained with generated data. The test set is composed of real images.\nTherefore, The classifier accuracy is used as a proxy to evaluate if the\ngenerative model fit the true data distribution. By comparing results with\ndifferent generated datasets we are able to classify and compare generative\nmodels. The motivation of this approach is also to evaluate if generative\nmodels can help discriminative neural networks to learn, i.e., measure if\ntraining on generated data is able to make a model successful at testing on\nreal settings. Our experiments compare different generators from the\nVariational Auto-Encoders (VAE) and Generative Adversarial Network (GAN)\nframeworks on MNIST and fashion MNIST datasets. Our results show that none of\nthe generative models is able to replace completely true data to train a\ndiscriminative model. But they also show that the initial GAN and WGAN are the\nbest choices to generate on MNIST database (Modified National Institute of\nStandards and Technology database) and fashion MNIST database.', 'Neural random fields (NRFs), referring to a class of generative models that\nuse neural networks to implement potential functions in random fields (a.k.a.\nenergy-based models), are not new but receive less attention with slow\nprogress. Different from various directed graphical models such as generative\nadversarial networks (GANs), NRFs provide an interesting family of undirected\ngraphical models for generative modeling. In this paper we propose a new\napproach, the inclusive-NRF approach, to learning NRFs for continuous data\n(e.g. images), by introducing inclusive-divergence minimized auxiliary\ngenerators and developing stochastic gradient sampling in an augmented space.\nBased on the new approach, specific inclusive-NRF models are developed and\nthoroughly evaluated in two important generative modeling applications - image\ngeneration and anomaly detection. The proposed models consistently improve over\nstate-of-the-art results in both applications. Remarkably, in addition to\nsuperior sample generation, one additional benefit of our inclusive-NRF\napproach is that, unlike GANs, it can directly provide (unnormalized) density\nestimate for sample evaluation. With these contributions and results, this\npaper significantly advances the learning and applications of NRFs to a new\nlevel, both theoretically and empirically, which have never been obtained\nbefore.', 'We consider the problem of learning deep generative models from data. We\nformulate a method that generates an independent sample via a single\nfeedforward pass through a multilayer perceptron, as in the recently proposed\ngenerative adversarial networks (Goodfellow et al., 2014). Training a\ngenerative adversarial network, however, requires careful optimization of a\ndifficult minimax program. Instead, we utilize a technique from statistical\nhypothesis testing known as maximum mean discrepancy (MMD), which leads to a\nsimple objective that can be interpreted as matching all orders of statistics\nbetween a dataset and samples from the model, and can be trained by\nbackpropagation. We further boost the performance of this approach by combining\nour generative network with an auto-encoder network, using MMD to learn to\ngenerate codes that can then be decoded to produce samples. We show that the\ncombination of these techniques yields excellent generative models compared to\nbaseline approaches as measured on MNIST and the Toronto Face Database.']","['With the increasing interest in the content creation field in multiple\nsectors such as media, education, and entertainment, there is an increasing\ntrend in the papers that uses AI algorithms to generate content such as images,\nvideos, audio, and text. Generative Adversarial Networks (GANs) in one of the\npromising models that synthesizes data samples that are similar to real data\nsamples. While the variations of GANs models, in general, have been covered to\nsome extent in several survey papers, to the best of our knowledge, this is\namong the first survey papers that reviews the state-of-the-art video GANs\nmodels. This paper first categorized GANs review papers into general GANs\nreview papers, image GANs review papers, and special field GANs review papers\nsuch as anomaly detection, medical imaging, or cybersecurity. The paper then\nsummarizes the main improvements in GANs frameworks that are not initially\ndeveloped for the video domain but have been adopted in multiple video GANs\nvariations. Then, a comprehensive review of video GANs models is provided under\ntwo main divisions according to the presence or non-presence of a condition.\nThe conditional models then further grouped according to the type of condition\ninto audio, text, video, and image. The paper is concluded by highlighting the\nmain challenges and limitations of the current video GANs models. A\ncomprehensive list of datasets, applied loss functions, and evaluation metrics\nis provided in the supplementary material.', ""Measuring the generalization capacity of Deep Generative Models (DGMs) is\ndifficult because of the curse of dimensionality. Evaluation metrics for DGMs\nsuch as Inception Score, Fr\\'echet Inception Distance, Precision-Recall, and\nNeural Net Divergence try to estimate the distance between the generated\ndistribution and the target distribution using a polynomial number of samples.\nThese metrics are the target of researchers when designing new models. Despite\nthe claims, it is still unclear how well can they measure the generalization\ncapacity of a generative model. In this paper, we investigate the capacity of\nthese metrics in measuring the generalization capacity. We introduce a\nframework for comparing the robustness of evaluation metrics. We show that\nbetter scores in these metrics do not imply better generalization. They can be\nfooled easily by a generator that memorizes a small subset of the training set.\nWe propose a fix to the NND metric to make it more robust to noise in the\ngenerated data. Toward building a robust metric for generalization, we propose\nto apply the Minimum Description Length principle to the problem of evaluating\nDGMs. We develop an efficient method for estimating the complexity of\nGenerative Latent Variable Models (GLVMs). Experimental results show that our\nmetric can effectively detect training set memorization and distinguish GLVMs\nof different generalization capacities. Source code is available at\nhttps://github.com/htt210/GeneralizationMetricGAN."", 'Deep generative models, such as Generative Adversarial Networks (GANs) or\nVariational Autoencoders (VAs) have been demonstrated to produce images of high\nvisual quality. However, the existing hardware severely limits the size of the\nimages that can be generated. The rapid growth of high dimensional data in many\nfields of science therefore poses a significant challenge for generative\nmodels. In cosmology, the large-scale, three-dimensional matter distribution,\nmodeled with N-body simulations, plays a crucial role in understanding the\nevolution of the universe. As these simulations are computationally very\nexpensive, GANs have recently generated interest as a possible method to\nemulate these datasets, but they have been, so far, mostly limited to two\ndimensional data. In this work, we introduce a new benchmark for the generation\nof three dimensional N-body simulations, in order to stimulate new ideas in the\nmachine learning community and move closer to the practical use of generative\nmodels in cosmology. As a first benchmark result, we propose a scalable GAN\napproach for training a generator of N-body three-dimensional cubes. Our\ntechnique relies on two key building blocks, (i) splitting the generation of\nthe high-dimensional data into smaller parts, and (ii) using a multi-scale\napproach that efficiently captures global image features that might otherwise\nbe lost in the splitting process. We evaluate the performance of our model for\nthe generation of N-body samples using various statistical measures commonly\nused in cosmology. Our results show that the proposed model produces samples of\nhigh visual quality, although the statistical analysis reveals that capturing\nrare features in the data poses significant problems for the generative models.\nWe make the data, quality evaluation routines, and the proposed GAN\narchitecture publicly available at https://github.com/nperraud/3DcosmoGAN', 'Graphs are important data representations for describing objects and their\nrelationships, which appear in a wide diversity of real-world scenarios. As one\nof a critical problem in this area, graph generation considers learning the\ndistributions of given graphs and generating more novel graphs. Owing to their\nwide range of applications, generative models for graphs, which have a rich\nhistory, however, are traditionally hand-crafted and only capable of modeling a\nfew statistical properties of graphs. Recent advances in deep generative models\nfor graph generation is an important step towards improving the fidelity of\ngenerated graphs and paves the way for new kinds of applications. This article\nprovides an extensive overview of the literature in the field of deep\ngenerative models for graph generation. Firstly, the formal definition of deep\ngenerative models for the graph generation and the preliminary knowledge are\nprovided. Secondly, taxonomies of deep generative models for both unconditional\nand conditional graph generation are proposed respectively; the existing works\nof each are compared and analyzed. After that, an overview of the evaluation\nmetrics in this specific domain is provided. Finally, the applications that\ndeep graph generation enables are summarized and five promising future research\ndirections are highlighted.', 'Generative Adversarial Networks (GANs), as a framework for estimating\ngenerative models via an adversarial process, have attracted huge attention and\nhave proven to be powerful in a variety of tasks. However, training GANs is\nwell known for being delicate and unstable, partially caused by its sig- moid\ncross entropy loss function for the discriminator. To overcome such a problem,\nmany researchers directed their attention on various ways to measure how close\nthe model distribution and real distribution are and have applied dif- ferent\nmetrics as their objective functions. In this paper, we propose a novel\nframework to train GANs based on distance metric learning and we call it Metric\nLearning-based Gener- ative Adversarial Network (MLGAN). The discriminator of\nMLGANs can dynamically learn an appropriate metric, rather than a static one,\nto measure the distance between generated samples and real samples. Afterwards,\nMLGANs update the generator under the newly learned metric. We evaluate our ap-\nproach on several representative datasets and the experimen- tal results\ndemonstrate that MLGANs can achieve superior performance compared with several\nexisting state-of-the-art approaches. We also empirically show that MLGANs\ncould increase the stability of training GANs.', 'Generative Adversarial Networks (GANs) have been extremely successful in\nvarious application domains such as computer vision, medicine, and natural\nlanguage processing. Moreover, transforming an object or person to a desired\nshape become a well-studied research in the GANs. GANs are powerful models for\nlearning complex distributions to synthesize semantically meaningful samples.\nHowever, there is a lack of comprehensive review in this field, especially lack\nof a collection of GANs loss-variant, evaluation metrics, remedies for diverse\nimage generation, and stable training. Given the current fast GANs development,\nin this survey, we provide a comprehensive review of adversarial models for\nimage synthesis. We summarize the synthetic image generation methods, and\ndiscuss the categories including image-to-image translation, fusion image\ngeneration, label-to-image mapping, and text-to-image translation. We organize\nthe literature based on their base models, developed ideas related to\narchitectures, constraints, loss functions, evaluation metrics, and training\ndatasets. We present milestones of adversarial models, review an extensive\nselection of previous works in various categories, and present insights on the\ndevelopment route from the model-based to data-driven methods. Further, we\nhighlight a range of potential future research directions. One of the unique\nfeatures of this review is that all software implementations of these GAN\nmethods and datasets have been collected and made available in one place at\nhttps://github.com/pshams55/GAN-Case-Study.', 'An open secret in contemporary machine learning is that many models work\nbeautifully on standard benchmarks but fail to generalize outside the lab. This\nhas been attributed to biased training data, which provide poor coverage over\nreal world events. Generative models are no exception, but recent advances in\ngenerative adversarial networks (GANs) suggest otherwise - these models can now\nsynthesize strikingly realistic and diverse images. Is generative modeling of\nphotos a solved problem? We show that although current GANs can fit standard\ndatasets very well, they still fall short of being comprehensive models of the\nvisual manifold. In particular, we study their ability to fit simple\ntransformations such as camera movements and color changes. We find that the\nmodels reflect the biases of the datasets on which they are trained (e.g.,\ncentered objects), but that they also exhibit some capacity for generalization:\nby ""steering"" in latent space, we can shift the distribution while still\ncreating realistic images. We hypothesize that the degree of distributional\nshift is related to the breadth of the training data distribution. Thus, we\nconduct experiments to quantify the limits of GAN transformations and introduce\ntechniques to mitigate the problem. Code is released on our project page:\nhttps://ali-design.github.io/gan_steerability/', 'Evaluating generative adversarial networks (GANs) is inherently challenging.\nIn this paper, we revisit several representative sample-based evaluation\nmetrics for GANs, and address the problem of how to evaluate the evaluation\nmetrics. We start with a few necessary conditions for metrics to produce\nmeaningful scores, such as distinguishing real from generated samples,\nidentifying mode dropping and mode collapsing, and detecting overfitting. With\na series of carefully designed experiments, we comprehensively investigate\nexisting sample-based metrics and identify their strengths and limitations in\npractical settings. Based on these results, we observe that kernel Maximum Mean\nDiscrepancy (MMD) and the 1-Nearest-Neighbor (1-NN) two-sample test seem to\nsatisfy most of the desirable properties, provided that the distances between\nsamples are computed in a suitable feature space. Our experiments also unveil\ninteresting properties about the behavior of several popular GAN models, such\nas whether they are memorizing training samples, and how far they are from\nlearning the target distribution.', 'Generative Adversarial Network (GAN) and its variants exhibit\nstate-of-the-art performance in the class of generative models. To capture\nhigher-dimensional distributions, the common learning procedure requires high\ncomputational complexity and a large number of parameters. The problem of\nemploying such massive framework arises when deploying it on a platform with\nlimited computational power such as mobile phones. In this paper, we present a\nnew generative adversarial framework by representing each layer as a tensor\nstructure connected by multilinear operations, aiming to reduce the number of\nmodel parameters by a large factor while preserving the generative performance\nand sample quality. To learn the model, we employ an efficient algorithm which\nalternatively optimizes both discriminator and generator. Experimental outcomes\ndemonstrate that our model can achieve high compression rate for model\nparameters up to $35$ times when compared to the original GAN for MNIST\ndataset.', 'Generative adversarial networks (GAN) have been effective for learning\ngenerative models for real-world data. However, existing GANs (GAN and its\nvariants) tend to suffer from training problems such as instability and mode\ncollapse. In this paper, we propose a novel GAN framework called evolutionary\ngenerative adversarial networks (E-GAN) for stable GAN training and improved\ngenerative performance. Unlike existing GANs, which employ a pre-defined\nadversarial objective function alternately training a generator and a\ndiscriminator, we utilize different adversarial training objectives as mutation\noperations and evolve a population of generators to adapt to the environment\n(i.e., the discriminator). We also utilize an evaluation mechanism to measure\nthe quality and diversity of generated samples, such that only well-performing\ngenerator(s) are preserved and used for further training. In this way, E-GAN\novercomes the limitations of an individual adversarial training objective and\nalways preserves the best offspring, contributing to progress in and the\nsuccess of GANs. Experiments on several datasets demonstrate that E-GAN\nachieves convincing generative performance and reduces the training problems\ninherent in existing GANs.', 'One of the most significant challenges in statistical signal processing and\nmachine learning is how to obtain a generative model that can produce samples\nof large-scale data distribution, such as images and speeches. Generative\nAdversarial Network (GAN) is an effective method to address this problem. The\nGANs provide an appropriate way to learn deep representations without\nwidespread use of labeled training data. This approach has attracted the\nattention of many researchers in computer vision since it can generate a large\namount of data without precise modeling of the probability density function\n(PDF). In GANs, the generative model is estimated via a competitive process\nwhere the generator and discriminator networks are trained simultaneously. The\ngenerator learns to generate plausible data, and the discriminator learns to\ndistinguish fake data created by the generator from real data samples. Given\nthe rapid growth of GANs over the last few years and their application in\nvarious fields, it is necessary to investigate these networks accurately. In\nthis paper, after introducing the main concepts and the theory of GAN, two new\ndeep generative models are compared, the evaluation metrics utilized in the\nliterature and challenges of GANs are also explained. Moreover, the most\nremarkable GAN architectures are categorized and discussed. Finally, the\nessential applications in computer vision are examined.', 'Generative adversarial networks (GANs) aim to generate realistic data from\nsome prior distribution (e.g., Gaussian noises). However, such prior\ndistribution is often independent of real data and thus may lose semantic\ninformation (e.g., geometric structure or content in images) of data. In\npractice, the semantic information might be represented by some latent\ndistribution learned from data, which, however, is hard to be used for sampling\nin GANs. In this paper, rather than sampling from the pre-defined prior\ndistribution, we propose a Local Coordinate Coding (LCC) based sampling method\nto improve GANs. We derive a generalization bound for LCC based GANs and prove\nthat a small dimensional input is sufficient to achieve good generalization.\nExtensive experiments on various real-world datasets demonstrate the\neffectiveness of the proposed method.', 'We propose a novel method to directly learn a stochastic transition operator\nwhose repeated application provides generated samples. Traditional undirected\ngraphical models approach this problem indirectly by learning a Markov chain\nmodel whose stationary distribution obeys detailed balance with respect to a\nparameterized energy function. The energy function is then modified so the\nmodel and data distributions match, with no guarantee on the number of steps\nrequired for the Markov chain to converge. Moreover, the detailed balance\ncondition is highly restrictive: energy based models corresponding to neural\nnetworks must have symmetric weights, unlike biological neural circuits. In\ncontrast, we develop a method for directly learning arbitrarily parameterized\ntransition operators capable of expressing non-equilibrium stationary\ndistributions that violate detailed balance, thereby enabling us to learn more\nbiologically plausible asymmetric neural networks and more general non-energy\nbased dynamical systems. The proposed training objective, which we derive via\nprincipled variational methods, encourages the transition operator to ""walk\nback"" in multi-step trajectories that start at data-points, as quickly as\npossible back to the original data points. We present a series of experimental\nresults illustrating the soundness of the proposed approach, Variational\nWalkback (VW), on the MNIST, CIFAR-10, SVHN and CelebA datasets, demonstrating\nsuperior samples compared to earlier attempts to learn a transition operator.\nWe also show that although each rapid training trajectory is limited to a\nfinite but variable number of steps, our transition operator continues to\ngenerate good samples well past the length of such trajectories, thereby\ndemonstrating the match of its non-equilibrium stationary distribution to the\ndata distribution. Source Code: http://github.com/anirudh9119/walkback_nips17', 'In this paper, we propose a diffusion probabilistic model for handwriting\ngeneration. Diffusion models are a class of generative models where samples\nstart from Gaussian noise and are gradually denoised to produce output. Our\nmethod of handwriting generation does not require using any text-recognition\nbased, writer-style based, or adversarial loss functions, nor does it require\ntraining of auxiliary networks. Our model is able to incorporate writer\nstylistic features directly from image data, eliminating the need for user\ninteraction during sampling. Experiments reveal that our model is able to\ngenerate realistic , high quality images of handwritten text in a similar style\nto a given writer. Our implementation can be found at\nhttps://github.com/tcl9876/Diffusion-Handwriting-Generation', 'Generative Adversarial Networks (GANs) have recently introduced effective\nmethods of performing Image-to-Image translations. These models can be applied\nand generalized to a variety of domains in Image-to-Image translation without\nchanging any parameters. In this paper, we survey and analyze eight\nImage-to-Image Generative Adversarial Networks: Pix2Pix, CycleGAN, CoGAN,\nStarGAN, MUNIT, StarGAN2, DA-GAN, and Self Attention GAN. Each of these models\npresented state-of-the-art results and introduced new techniques to build\nImage-to-Image GANs. In addition to a survey of the models, we also survey the\n18 datasets they were trained on and the 9 metrics they were evaluated on.\nFinally, we present results of a controlled experiment for 6 of these models on\na common set of metrics and datasets. The results were mixed and showed that,\non certain datasets, tasks, and metrics, some models outperformed others. The\nlast section of this paper discusses those results and establishes areas of\nfuture research. As researchers continue to innovate new Image-to-Image GANs,\nit is important to gain a good understanding of the existing methods, datasets,\nand metrics. This paper provides a comprehensive overview and discussion to\nhelp build this foundation.', 'We present a variety of new architectural features and training procedures\nthat we apply to the generative adversarial networks (GANs) framework. We focus\non two applications of GANs: semi-supervised learning, and the generation of\nimages that humans find visually realistic. Unlike most work on generative\nmodels, our primary goal is not to train a model that assigns high likelihood\nto test data, nor do we require the model to be able to learn well without\nusing any labels. Using our new techniques, we achieve state-of-the-art results\nin semi-supervised classification on MNIST, CIFAR-10 and SVHN. The generated\nimages are of high quality as confirmed by a visual Turing test: our model\ngenerates MNIST samples that humans cannot distinguish from real data, and\nCIFAR-10 samples that yield a human error rate of 21.3%. We also present\nImageNet samples with unprecedented resolution and show that our methods enable\nthe model to learn recognizable features of ImageNet classes.', 'Sketch-based image retrieval (SBIR) is the task of retrieving images from a\nnatural image database that correspond to a given hand-drawn sketch. Ideally,\nan SBIR model should learn to associate components in the sketch (say, feet,\ntail, etc.) with the corresponding components in the image having similar shape\ncharacteristics. However, current evaluation methods simply focus only on\ncoarse-grained evaluation where the focus is on retrieving images which belong\nto the same class as the sketch but not necessarily having the same shape\ncharacteristics as in the sketch. As a result, existing methods simply learn to\nassociate sketches with classes seen during training and hence fail to\ngeneralize to unseen classes. In this paper, we propose a new benchmark for\nzero-shot SBIR where the model is evaluated in novel classes that are not seen\nduring training. We show through extensive experiments that existing models for\nSBIR that are trained in a discriminative setting learn only class specific\nmappings and fail to generalize to the proposed zero-shot setting. To\ncircumvent this, we propose a generative approach for the SBIR task by\nproposing deep conditional generative models that take the sketch as an input\nand fill the missing information stochastically. Experiments on this new\nbenchmark created from the ""Sketchy"" dataset, which is a large-scale database\nof sketch-photo pairs demonstrate that the performance of these generative\nmodels is significantly better than several state-of-the-art approaches in the\nproposed zero-shot framework of the coarse-grained SBIR task.', ""Generative neural network models, including Generative Adversarial Network\n(GAN) and Auto-Encoders (AE), are among the most popular neural network models\nto generate adversarial data. The GAN model is composed of a generator that\nproduces synthetic data and of a discriminator that discriminates between the\ngenerator's output and the true data. AE consist of an encoder which maps the\nmodel distribution to a latent manifold and of a decoder which maps the latent\nmanifold to a reconstructed distribution. However, generative models are known\nto provoke chaotically scattered reconstructed distribution during their\ntraining, and consequently, incomplete generated adversarial distributions.\nCurrent distance measures fail to address this problem because they are not\nable to acknowledge the shape of the data manifold, i.e. its topological\nfeatures, and the scale at which the manifold should be analyzed. We propose\nPersistent Homology for Generative Models, PHom-GeM, a new methodology to\nassess and measure the distribution of a generative model. PHom-GeM minimizes\nan objective function between the true and the reconstructed distributions and\nuses persistent homology, the study of the topological features of a space at\ndifferent spatial resolutions, to compare the nature of the true and the\ngenerated distributions. Our experiments underline the potential of persistent\nhomology for Wasserstein GAN in comparison to Wasserstein AE and Variational\nAE. The experiments are conducted on a real-world data set particularly\nchallenging for traditional distance measures and generative neural network\nmodels. PHom-GeM is the first methodology to propose a topological distance\nmeasure, the bottleneck distance, for generative models used to compare\nadversarial samples in the context of credit card transactions."", 'Data sampling acts as a pivotal role in training deep learning models.\nHowever, an effective sampling schedule is difficult to learn due to the\ninherently high dimension of parameters in learning the sampling schedule. In\nthis paper, we propose an AutoSampling method to automatically learn sampling\nschedules for model training, which consists of the multi-exploitation step\naiming for optimal local sampling schedules and the exploration step for the\nideal sampling distribution. More specifically, we achieve sampling schedule\nsearch with shortened exploitation cycle to provide enough supervision. In\naddition, we periodically estimate the sampling distribution from the learned\nsampling schedules and perturb it to search in the distribution space. The\ncombination of two searches allows us to learn a robust sampling schedule. We\napply our AutoSampling method to a variety of image classification tasks\nillustrating the effectiveness of the proposed method.', 'Generative adversarial networks (GANs) have shown tremendous promise in\nlearning to generate data and effective at aiding semi-supervised\nclassification. However, to this point, semi-supervised GAN methods make the\nassumption that the unlabeled data set contains only samples of the joint\ndistribution of the classes of interest, referred to as inliers. Consequently,\nwhen presented with a sample from other distributions, referred to as outliers,\nGANs perform poorly at determining that it is not qualified to make a decision\non the sample. The problem of discriminating outliers from inliers while\nmaintaining classification accuracy is referred to here as the DOIC problem. In\nthis work, we describe an architecture that combines self-organizing maps\n(SOMs) with SS-GANS with the goal of mitigating the DOIC problem and\nexperimental results indicating that the architecture achieves the goal.\nMultiple experiments were conducted on hyperspectral image data sets. The\nSS-GANS performed slightly better than supervised GANS on classification\nproblems with and without the SOM. Incorporating the SOMs into the SS-GANs and\nthe supervised GANS led to substantially mitigation of the DOIC problem when\ncompared to SS-GANS and GANs without the SOMs. Furthermore, the SS-GANS\nperformed much better than GANS on the DOIC problem, even without the SOMs.', 'We extend and improve the work of Model Agnostic Anchors for explanations on\nimage classification through the use of generative adversarial networks (GANs).\nUsing GANs, we generate samples from a more realistic perturbation\ndistribution, by optimizing under a lower dimensional latent space. This\nincreases the trust in an explanation, as results now come from images that are\nmore likely to be found in the original training set of a classifier, rather\nthan an overlay of random images. A large drawback to our method is the\ncomputational complexity of sampling through optimization; to address this, we\nimplement more efficient algorithms, including a diverse encoder. Lastly, we\nshare results from the MNIST and CelebA datasets, and note that our\nexplanations can lead to smaller and higher precision anchors.', 'Prior domain knowledge can greatly help to learn generative models. However,\nit is often too costly to hard-code prior knowledge as a specific model\narchitecture, so we often have to use general-purpose models. In this paper, we\npropose a method to incorporate prior knowledge of feature relations into the\nlearning of general-purpose generative models. To this end, we formulate a\nregularizer that makes the marginals of a generative model to follow prescribed\nrelative dependence of features. It can be incorporated into off-the-shelf\nlearning methods of many generative models, including variational autoencoders\nand generative adversarial networks, as its gradients can be computed using\nstandard backpropagation techniques. We show the effectiveness of the proposed\nmethod with experiments on multiple types of datasets and generative models.', ""The Generative Adversarial Network (GAN) is a state-of-the-art technique in\nthe field of deep learning. A number of recent papers address the theory and\napplications of GANs in various fields of image processing. Fewer studies,\nhowever, have directly evaluated GAN outputs. Those that have been conducted\nfocused on using classification performance, e.g., Inception Score (IS) and\nstatistical metrics, e.g., Fr\\'echet Inception Distance (FID). Here, we\nconsider a fundamental way to evaluate GANs by directly analyzing the images\nthey generate, instead of using them as inputs to other classifiers. We\ncharacterize the performance of a GAN as an image generator according to three\naspects: 1) Creativity: non-duplication of the real images. 2) Inheritance:\ngenerated images should have the same style, which retains key features of the\nreal images. 3) Diversity: generated images are different from each other. A\nGAN should not generate a few different images repeatedly. Based on the three\naspects of ideal GANs, we have designed the Likeness Score (LS) to evaluate GAN\nperformance, and have applied it to evaluate several typical GANs. We compared\nour proposed measure with two commonly used GAN evaluation methods: IS and FID,\nand four additional measures. Furthermore, we discuss how these evaluations\ncould help us deepen our understanding of GANs and improve their performance."", ""We present two new metrics for evaluating generative models in the\nclass-conditional image generation setting. These metrics are obtained by\ngeneralizing the two most popular unconditional metrics: the Inception Score\n(IS) and the Fre'chet Inception Distance (FID). A theoretical analysis shows\nthe motivation behind each proposed metric and links the novel metrics to their\nunconditional counterparts. The link takes the form of a product in the case of\nIS or an upper bound in the FID case. We provide an extensive empirical\nevaluation, comparing the metrics to their unconditional variants and to other\nmetrics, and utilize them to analyze existing generative models, thus providing\nadditional insights about their performance, from unlearned classes to mode\ncollapse."", ""Generative Adversarial Networks (GANs) is a powerful family of models that\nlearn an underlying distribution to generate synthetic data. Many existing\nstudies of GANs focus on improving the realness of the generated image data for\nvisual applications, and few of them concern about improving the quality of the\ngenerated data for training other classifiers -- a task known as the model\ncompatibility problem. As a consequence, existing GANs often prefer generating\n`easier' synthetic data that are far from the boundaries of the classifiers,\nand refrain from generating near-boundary data, which are known to play an\nimportant roles in training the classifiers. To improve GAN in terms of model\ncompatibility, we propose Boundary-Calibration GANs (BCGANs), which leverage\nthe boundary information from a set of pre-trained classifiers using the\noriginal data. In particular, we introduce an auxiliary Boundary-Calibration\nloss (BC-loss) into the generator of GAN to match the statistics between the\nposterior distributions of original data and generated data with respect to the\nboundaries of the pre-trained classifiers. The BC-loss is provably unbiased and\ncan be easily coupled with different GAN variants to improve their model\ncompatibility. Experimental results demonstrate that BCGANs not only generate\nrealistic images like original GANs but also achieves superior model\ncompatibility than the original GANs."", 'Image generation remains a fundamental problem in artificial intelligence in\ngeneral and deep learning in specific. The generative adversarial network (GAN)\nwas successful in generating high quality samples of natural images. We propose\na model called composite generative adversarial network, that reveals the\ncomplex structure of images with multiple generators in which each generator\ngenerates some part of the image. Those parts are combined by alpha blending\nprocess to create a new single image. It can generate, for example, background\nand face sequentially with two generators, after training on face dataset.\nTraining was done in an unsupervised way without any labels about what each\ngenerator should generate. We found possibilities of learning the structure by\nusing this generative model empirically.', 'Generative Adversarial Networks (GANs) became very popular for generation of\nrealistically looking images. In this paper, we propose to use GANs to\nsynthesize artificial financial data for research and benchmarking purposes. We\ntest this approach on three American Express datasets, and show that properly\ntrained GANs can replicate these datasets with high fidelity. For our\nexperiments, we define a novel type of GAN, and suggest methods for data\npreprocessing that allow good training and testing performance of GANs. We also\ndiscuss methods for evaluating the quality of generated data, and their\ncomparison with the original real data.', 'Generative Adversarial Networks (GANs) can accurately model complex\nmulti-dimensional data and generate realistic samples. However, due to their\nimplicit estimation of data distributions, their evaluation is a challenging\ntask. The majority of research efforts associated with tackling this issue were\nvalidated by qualitative visual evaluation. Such approaches do not generalize\nwell beyond the image domain. Since many of those evaluation metrics are\nproposed and bound to the vision domain, they are difficult to apply to other\ndomains. Quantitative measures are necessary to better guide the training and\ncomparison of different GANs models. In this work, we leverage Siamese neural\nnetworks to propose a domain-agnostic evaluation metric: (1) with a qualitative\nevaluation that is consistent with human evaluation, (2) that is robust\nrelative to common GAN issues such as mode dropping and invention, and (3) does\nnot require any pretrained classifier. The empirical results in this paper\ndemonstrate the superiority of this method compared to the popular Inception\nScore and are competitive with the FID score.', 'Generative Adversarial Networks (GANs) have become a very popular tool for\nimplicitly learning high-dimensional probability distributions. Several\nimprovements have been made to the original GAN formulation to address some of\nits shortcomings like mode collapse, convergence issues, entanglement, poor\nvisual quality etc. While a significant effort has been directed towards\nimproving the visual quality of images generated by GANs, it is rather\nsurprising that objective image quality metrics have neither been employed as\ncost functions nor as regularizers in GAN objective functions. In this work, we\nshow how a distance metric that is a variant of the Structural SIMilarity\n(SSIM) index (a popular full-reference image quality assessment algorithm), and\na novel quality aware discriminator gradient penalty function that is inspired\nby the Natural Image Quality Evaluator (NIQE, a popular no-reference image\nquality assessment algorithm) can each be used as excellent regularizers for\nGAN objective functions. Specifically, we demonstrate state-of-the-art\nperformance using the Wasserstein GAN gradient penalty (WGAN-GP) framework over\nCIFAR-10, STL10 and CelebA datasets.', 'A crucial factor to trust Machine Learning (ML) algorithm decisions is a good\nrepresentation of its application field by the training dataset. This is\nparticularly true when parts of the training data have been artificially\ngenerated to overcome common training problems such as lack of data or\nimbalanced dataset. Over the last few years, Generative Adversarial Networks\n(GANs) have shown remarkable results in generating realistic data. However,\nthis ML approach lacks an objective function to evaluate the quality of the\ngenerated data. Numerous GAN applications focus on generating image data mostly\nbecause they can be easily evaluated by a human eye. Less efforts have been\nmade to generate time series data. Assessing their quality is more complicated,\nparticularly for technical data. In this paper, we propose a human-centered\napproach supporting a ML or domain expert to accomplish this task using Visual\nAnalytics (VA) techniques. The presented approach consists of two views, namely\na GAN Iteration View showing similarity metrics between real and generated data\nover the iterations of the generation process and a Detailed Comparative View\nequipped with different time series visualizations such as TimeHistograms, to\ncompare the generated data at different iteration steps. Starting from the GAN\nIteration View, the user can choose suitable iteration steps for detailed\ninspection. We evaluate our approach with a usage scenario that enabled an\nefficient comparison of two different GAN models.', 'We present a novel method and analysis to train generative adversarial\nnetworks (GAN) in a stable manner. As shown in recent analysis, training is\noften undermined by the probability distribution of the data being zero on\nneighborhoods of the data space. We notice that the distributions of real and\ngenerated data should match even when they undergo the same filtering.\nTherefore, to address the limited support problem we propose to train GANs by\nusing different filtered versions of the real and generated data distributions.\nIn this way, filtering does not prevent the exact matching of the data\ndistribution, while helping training by extending the support of both\ndistributions. As filtering we consider adding samples from an arbitrary\ndistribution to the data, which corresponds to a convolution of the data\ndistribution with the arbitrary one. We also propose to learn the generation of\nthese samples so as to challenge the discriminator in the adversarial training.\nWe show that our approach results in a stable and well-behaved training of even\nthe original minimax GAN formulation. Moreover, our technique can be\nincorporated in most modern GAN formulations and leads to a consistent\nimprovement on several common datasets.', ""Generative models estimate the underlying distribution of a dataset to\ngenerate realistic samples according to that distribution. In this paper, we\npresent the first membership inference attacks against generative models: given\na data point, the adversary determines whether or not it was used to train the\nmodel. Our attacks leverage Generative Adversarial Networks (GANs), which\ncombine a discriminative and a generative model, to detect overfitting and\nrecognize inputs that were part of training datasets, using the discriminator's\ncapacity to learn statistical differences in distributions.\n  We present attacks based on both white-box and black-box access to the target\nmodel, against several state-of-the-art generative models, over datasets of\ncomplex representations of faces (LFW), objects (CIFAR-10), and medical images\n(Diabetic Retinopathy). We also discuss the sensitivity of the attacks to\ndifferent training parameters, and their robustness against mitigation\nstrategies, finding that defenses are either ineffective or lead to\nsignificantly worse performances of the generative models in terms of training\nstability and/or sample quality."", ""Generative adversarial networks (GANs) typically require ample data for\ntraining in order to synthesize high-fidelity images. Recent studies have shown\nthat training GANs with limited data remains formidable due to discriminator\noverfitting, the underlying cause that impedes the generator's convergence.\nThis paper introduces a novel strategy called Adaptive Pseudo Augmentation\n(APA) to encourage healthy competition between the generator and the\ndiscriminator. As an alternative method to existing approaches that rely on\nstandard data augmentations or model regularization, APA alleviates overfitting\nby employing the generator itself to augment the real data distribution with\ngenerated images, which deceives the discriminator adaptively. Extensive\nexperiments demonstrate the effectiveness of APA in improving synthesis quality\nin the low-data regime. We provide a theoretical analysis to examine the\nconvergence and rationality of our new training strategy. APA is simple and\neffective. It can be added seamlessly to powerful contemporary GANs, such as\nStyleGAN2, with negligible computational cost."", 'Generative Adversarial Networks (GANs) have achieved a great success in\nunsupervised learning. Despite its remarkable empirical performance, there are\nlimited theoretical studies on the statistical properties of GANs. This paper\nprovides approximation and statistical guarantees of GANs for the estimation of\ndata distributions that have densities in a H\\""{o}lder space. Our main result\nshows that, if the generator and discriminator network architectures are\nproperly chosen, GANs are consistent estimators of data distributions under\nstrong discrepancy metrics, such as the Wasserstein-1 distance. Furthermore,\nwhen the data distribution exhibits low-dimensional structures, we show that\nGANs are capable of capturing the unknown low-dimensional structures in data\nand enjoy a fast statistical convergence, which is free of curse of the ambient\ndimensionality. Our analysis for low-dimensional data builds upon a universal\napproximation theory of neural networks with Lipschitz continuity guarantees,\nwhich may be of independent interest.', 'Generative modeling has recently seen many exciting developments with the\nadvent of deep generative architectures such as Variational Auto-Encoders (VAE)\nor Generative Adversarial Networks (GAN). The ability to draw synthetic i.i.d.\nobservations with the same joint probability distribution as a given dataset\nhas a wide range of applications including representation learning, compression\nor imputation. It appears that it also has many applications in privacy\npreserving data analysis, especially when used in conjunction with differential\nprivacy techniques. This paper focuses on synthetic data generation models with\nprivacy preserving applications in mind. It introduces a novel architecture,\nthe Composable Generative Model (CGM) that is state-of-the-art in tabular data\ngeneration. Any conditional generative model can be used as a sub-component of\nthe CGM, including CGMs themselves, allowing the generation of numerical,\ncategorical data as well as images, text, or time series. The CGM has been\nevaluated on 13 datasets (6 standard datasets and 7 simulated) and compared to\n14 recent generative models. It beats the state of the art in tabular data\ngeneration by a significant margin.', 'Generative Adversarial Networks (GANs) have been used to model the underlying\nprobability distribution of sample based datasets. GANs are notoriuos for\ntraining difficulties and their dependence on arbitrary hyperparameters. One\nrecent improvement in GAN literature is to use the Wasserstein distance as loss\nfunction leading to Wasserstein Generative Adversarial Networks (WGANs). Using\nthis as a basis, we show various ways in which the Wasserstein distance is\nestimated for the task of generative modelling. Additionally, the secrets in\ntraining such models are shown and summarized at the end of this work. Where\napplicable, we extend current works to different algorithms, different cost\nfunctions, and different regularization schemes to improve generative models.', ""A primary difficulty with unsupervised discovery of structure in large data\nsets is a lack of quantitative evaluation criteria. In this work, we propose\nand investigate several metrics for evaluating and comparing generative models\nof behavior learned using imitation learning. Compared to the commonly-used\nmodel log-likelihood, these criteria look at longer temporal relationships in\nbehavior, are relevant if behavior has some properties that are inherently\nunpredictable, and highlight biases in the overall distribution of behaviors\nproduced by the model. Pointwise metrics compare real to model-predicted\ntrajectories given true past information. Distribution metrics compare\nstatistics of the model simulating behavior in open loop, and are inspired by\nhow experimental biologists evaluate the effects of manipulations on animal\nbehavior. We show that the proposed metrics correspond with biologists'\nintuitions about behavior, and allow us to evaluate models, understand their\nbiases, and enable us to propose new research directions."", 'Auto-encoding generative adversarial networks (GANs) combine the standard GAN\nalgorithm, which discriminates between real and model-generated data, with a\nreconstruction loss given by an auto-encoder. Such models aim to prevent mode\ncollapse in the learned generative model by ensuring that it is grounded in all\nthe available training data. In this paper, we develop a principle upon which\nauto-encoders can be combined with generative adversarial networks by\nexploiting the hierarchical structure of the generative model. The underlying\nprinciple shows that variational inference can be used a basic tool for\nlearning, but with the in- tractable likelihood replaced by a synthetic\nlikelihood, and the unknown posterior distribution replaced by an implicit\ndistribution; both synthetic likelihoods and implicit posterior distributions\ncan be learned using discriminators. This allows us to develop a natural fusion\nof variational auto-encoders and generative adversarial networks, combining the\nbest of both these methods. We describe a unified objective for optimization,\ndiscuss the constraints needed to guide learning, connect to the wide range of\nexisting work, and use a battery of tests to systematically and quantitatively\nassess the performance of our method.', 'We study how to generate molecule conformations (i.e., 3D structures) from a\nmolecular graph. Traditional methods, such as molecular dynamics, sample\nconformations via computationally expensive simulations. Recently, machine\nlearning methods have shown great potential by training on a large collection\nof conformation data. Challenges arise from the limited model capacity for\ncapturing complex distributions of conformations and the difficulty in modeling\nlong-range dependencies between atoms. Inspired by the recent progress in deep\ngenerative models, in this paper, we propose a novel probabilistic framework to\ngenerate valid and diverse conformations given a molecular graph. We propose a\nmethod combining the advantages of both flow-based and energy-based models,\nenjoying: (1) a high model capacity to estimate the multimodal conformation\ndistribution; (2) explicitly capturing the complex long-range dependencies\nbetween atoms in the observation space. Extensive experiments demonstrate the\nsuperior performance of the proposed method on several benchmarks, including\nconformation generation and distance modeling tasks, with a significant\nimprovement over existing generative models for molecular conformation\nsampling.', 'Generative adversarial networks (GAN) have shown remarkable results in image\ngeneration tasks. High fidelity class-conditional GAN methods often rely on\nstabilization techniques by constraining the global Lipschitz continuity. Such\nregularization leads to less expressive models and slower convergence speed;\nother techniques, such as the large batch training, require unconventional\ncomputing power and are not widely accessible. In this paper, we develop an\nefficient algorithm, namely FastGAN (Free AdverSarial Training), to improve the\nspeed and quality of GAN training based on the adversarial training technique.\nWe benchmark our method on CIFAR10, a subset of ImageNet, and the full ImageNet\ndatasets. We choose strong baselines such as SNGAN and SAGAN; the results\ndemonstrate that our training algorithm can achieve better generation quality\n(in terms of the Inception score and Frechet Inception distance) with less\noverall training time. Most notably, our training algorithm brings ImageNet\ntraining to the broader public by requiring 2-4 GPUs.', 'An interpretable generative model for handwritten digits synthesis is\nproposed in this work. Modern image generative models, such as Generative\nAdversarial Networks (GANs) and Variational Autoencoders (VAEs), are trained by\nbackpropagation (BP). The training process is complex and the underlying\nmechanism is difficult to explain. We propose an interpretable multi-stage PCA\nmethod to achieve the same goal and use handwritten digit images synthesis as\nan illustrative example. First, we derive principal-component-analysis-based\n(PCA-based) transform kernels at each stage based on the covariance of its\ninputs. This results in a sequence of transforms that convert input images of\ncorrelated pixels to spectral vectors of uncorrelated components. In other\nwords, it is a whitening process. Then, we can synthesize an image based on\nrandom vectors and multi-stage transform kernels through a coloring process.\nThe generative model is a feedforward (FF) design since no BP is used in model\nparameter determination. Its design complexity is significantly lower, and the\nwhole design process is explainable. Finally, we design an FF generative model\nusing the MNIST dataset, compare synthesis results with those obtained by\nstate-of-the-art GAN and VAE methods, and show that the proposed generative\nmodel achieves comparable performance.', 'In this work, we develop an importance sampling estimator by coupling the\nreduced-order model and the generative model in a problem setting of\nuncertainty quantification. The target is to estimate the probability that the\nquantity of interest (QoI) in a complex system is beyond a given threshold. To\navoid the prohibitive cost of sampling a large scale system, the reduced-order\nmodel is usually considered for a trade-off between efficiency and accuracy.\nHowever, the Monte Carlo estimator given by the reduced-order model is biased\ndue to the error from dimension reduction. To correct the bias, we still need\nto sample the fine model. An effective technique to reduce the variance\nreduction is importance sampling, where we employ the generative model to\nestimate the distribution of the data from the reduced-order model and use it\nfor the change of measure in the importance sampling estimator. To compensate\nthe approximation errors of the reduced-order model, more data that induce a\nslightly smaller QoI than the threshold need to be included into the training\nset. Although the amount of these data can be controlled by a posterior error\nestimate, redundant data, which may outnumber the effective data, will be kept\ndue to the epistemic uncertainty. To deal with this issue, we introduce a\nweighted empirical distribution to process the data from the reduced-order\nmodel. The generative model is then trained by minimizing the cross entropy\nbetween it and the weighted empirical distribution. We also introduce a penalty\nterm into the objective function to deal with the overfitting for more\nrobustness. Numerical results are presented to demonstrate the effectiveness of\nthe proposed methodology.', ""The new wave of successful generative models in machine learning has\nincreased the interest in deep learning driven de novo drug design. However,\nassessing the performance of such generative models is notoriously difficult.\nMetrics that are typically used to assess the performance of such generative\nmodels are the percentage of chemically valid molecules or the similarity to\nreal molecules in terms of particular descriptors, such as the partition\ncoefficient (logP) or druglikeness. However, method comparison is difficult\nbecause of the inconsistent use of evaluation metrics, the necessity for\nmultiple metrics, and the fact that some of these measures can easily be\ntricked by simple rule-based systems. We propose a novel distance measure\nbetween two sets of molecules, called Fr\\'echet ChemNet distance (FCD), that\ncan be used as an evaluation metric for generative models. The FCD is similar\nto a recently established performance metric for comparing image generation\nmethods, the Fr\\'echet Inception Distance (FID). Whereas the FID uses one of\nthe hidden layers of InceptionNet, the FCD utilizes the penultimate layer of a\ndeep neural network called ChemNet, which was trained to predict drug\nactivities. Thus, the FCD metric takes into account chemically and biologically\nrelevant information about molecules, and also measures the diversity of the\nset via the distribution of generated molecules. The FCD's advantage over\nprevious metrics is that it can detect if generated molecules are a) diverse\nand have similar b) chemical and c) biological properties as real molecules. We\nfurther provide an easy-to-use implementation that only requires the SMILES\nrepresentation of the generated molecules as input to calculate the FCD.\nImplementations are available at: https://www.github.com/bioinf-jku/FCD"", 'The discriminative approach to classification using deep neural networks has\nbecome the de-facto standard in various fields. Complementing recent\nreservations about safety against adversarial examples, we show that\nconventional discriminative methods can easily be fooled to provide incorrect\nlabels with very high confidence to out of distribution examples. We posit that\na generative approach is the natural remedy for this problem, and propose a\nmethod for classification using generative models. At training time, we learn a\ngenerative model for each class, while at test time, given an example to\nclassify, we query each generator for its most similar generation, and select\nthe class corresponding to the most similar one. Our approach is general and\ncan be used with expressive models such as GANs and VAEs. At test time, our\nmethod accurately ""knows when it does not know,"" and provides resilience to out\nof distribution examples while maintaining competitive performance for standard\nexamples.', 'Real networks exhibit nontrivial topological features such as heavy-tailed\ndegree distribution, high clustering, and small-worldness. Researchers have\ndeveloped several generative models for synthesizing artificial networks that\nare structurally similar to real networks. An important research problem is to\nidentify the generative model that best fits to a target network. In this\npaper, we investigate this problem and our goal is to select the model that is\nable to generate graphs similar to a given network instance. By the means of\ngenerating synthetic networks with seven outstanding generative models, we have\nutilized machine learning methods to develop a decision tree for model\nselection. Our proposed method, which is named ""Generative Model Selection for\nComplex Networks"" (GMSCN), outperforms existing methods with respect to\naccuracy, scalability and size-independence.', 'Generative Adversarial Networks (GANs) is a novel class of deep generative\nmodels which has recently gained significant attention. GANs learns complex and\nhigh-dimensional distributions implicitly over images, audio, and data.\nHowever, there exists major challenges in training of GANs, i.e., mode\ncollapse, non-convergence and instability, due to inappropriate design of\nnetwork architecture, use of objective function and selection of optimization\nalgorithm. Recently, to address these challenges, several solutions for better\ndesign and optimization of GANs have been investigated based on techniques of\nre-engineered network architectures, new objective functions and alternative\noptimization algorithms. To the best of our knowledge, there is no existing\nsurvey that has particularly focused on broad and systematic developments of\nthese solutions. In this study, we perform a comprehensive survey of the\nadvancements in GANs design and optimization solutions proposed to handle GANs\nchallenges. We first identify key research issues within each design and\noptimization technique and then propose a new taxonomy to structure solutions\nby key research issues. In accordance with the taxonomy, we provide a detailed\ndiscussion on different GANs variants proposed within each solution and their\nrelationships. Finally, based on the insights gained, we present the promising\nresearch directions in this rapidly growing field.', 'Recently, generative adversarial networks have gained a lot of popularity for\nimage generation tasks. However, such models are associated with complex\nlearning mechanisms and demand very large relevant datasets. This work borrows\nconcepts from image and video captioning models to form an image generative\nframework. The model is trained in a similar fashion as recurrent captioning\nmodel and uses the learned weights for image generation. This is done in an\ninverse direction, where the input is a caption and the output is an image. The\nvector representation of the sentence and frames are extracted from an\nencoder-decoder model which is initially trained on similar sentence and image\npairs. Our model conditions image generation on a natural language caption. We\nleverage a sequence-to-sequence model to generate synthetic captions that have\nthe same meaning for having a robust image generation. One key advantage of our\nmethod is that the traditional image captioning datasets can be used for\nsynthetic sentence paraphrases. Results indicate that images generated through\nmultiple captions are better at capturing the semantic meaning of the family of\ncaptions.', 'Image captioning is an ambiguous problem, with many suitable captions for an\nimage. To address ambiguity, beam search is the de facto method for sampling\nmultiple captions. However, beam search is computationally expensive and known\nto produce generic captions. To address this concern, some variational\nauto-encoder (VAE) and generative adversarial net (GAN) based methods have been\nproposed. Though diverse, GAN and VAE are less accurate. In this paper, we\nfirst predict a meaningful summary of the image, then generate the caption\nbased on that summary. We use part-of-speech as summaries, since our summary\nshould drive caption generation. We achieve the trifecta: (1) High accuracy for\nthe diverse captions as evaluated by standard captioning metrics and user\nstudies; (2) Faster computation of diverse captions compared to beam search and\ndiverse beam search; and (3) High diversity as evaluated by counting novel\nsentences, distinct n-grams and mutual overlap (i.e., mBleu-4) scores.', 'Generative neural samplers are probabilistic models that implement sampling\nusing feedforward neural networks: they take a random input vector and produce\na sample from a probability distribution defined by the network weights. These\nmodels are expressive and allow efficient computation of samples and\nderivatives, but cannot be used for computing likelihoods or for\nmarginalization. The generative-adversarial training method allows to train\nsuch models through the use of an auxiliary discriminative neural network. We\nshow that the generative-adversarial approach is a special case of an existing\nmore general variational divergence estimation approach. We show that any\nf-divergence can be used for training generative neural samplers. We discuss\nthe benefits of various choices of divergence functions on training complexity\nand the quality of the obtained generative models.', 'Generative Adversarial Network (GAN) and its variants have recently attracted\nintensive research interests due to their elegant theoretical foundation and\nexcellent empirical performance as generative models. These tools provide a\npromising direction in the studies where data availability is limited. One\ncommon issue in GANs is that the density of the learned generative distribution\ncould concentrate on the training data points, meaning that they can easily\nremember training samples due to the high model complexity of deep networks.\nThis becomes a major concern when GANs are applied to private or sensitive data\nsuch as patient medical records, and the concentration of distribution may\ndivulge critical patient information. To address this issue, in this paper we\npropose a differentially private GAN (DPGAN) model, in which we achieve\ndifferential privacy in GANs by adding carefully designed noise to gradients\nduring the learning procedure. We provide rigorous proof for the privacy\nguarantee, as well as comprehensive empirical evidence to support our analysis,\nwhere we demonstrate that our method can generate high quality data points at a\nreasonable privacy level.', ""We introduce a method to stabilize Generative Adversarial Networks (GANs) by\ndefining the generator objective with respect to an unrolled optimization of\nthe discriminator. This allows training to be adjusted between using the\noptimal discriminator in the generator's objective, which is ideal but\ninfeasible in practice, and using the current value of the discriminator, which\nis often unstable and leads to poor solutions. We show how this technique\nsolves the common problem of mode collapse, stabilizes training of GANs with\ncomplex recurrent generators, and increases diversity and coverage of the data\ndistribution by the generator."", 'Generative Adversarial Networks (GANs) have shown remarkable results in\nmodeling complex distributions, but their evaluation remains an unsettled\nissue. Evaluations are essential for: (i) relative assessment of different\nmodels and (ii) monitoring the progress of a single model throughout training.\nThe latter cannot be determined by simply inspecting the generator and\ndiscriminator loss curves as they behave non-intuitively. We leverage the\nnotion of duality gap from game theory to propose a measure that addresses both\n(i) and (ii) at a low computational cost. Extensive experiments show the\neffectiveness of this measure to rank different GAN models and capture the\ntypical GAN failure scenarios, including mode collapse and non-convergent\nbehaviours. This evaluation metric also provides meaningful monitoring on the\nprogression of the loss during training. It highly correlates with FID on\nnatural image datasets, and with domain specific scores for text, sound and\ncosmology data where FID is not directly suitable. In particular, our proposed\nmetric requires no labels or a pretrained classifier, making it domain\nagnostic.', 'Generative models, in particular generative adversarial networks (GANs), have\nreceived significant attention recently. A number of GAN variants have been\nproposed and have been utilized in many applications. Despite large strides in\nterms of theoretical progress, evaluating and comparing GANs remains a daunting\ntask. While several measures have been introduced, as of yet, there is no\nconsensus as to which measure best captures strengths and limitations of models\nand should be used for fair model comparison. As in other areas of computer\nvision and machine learning, it is critical to settle on one or few good\nmeasures to steer the progress in this field. In this paper, I review and\ncritically discuss more than 24 quantitative and 5 qualitative measures for\nevaluating generative models with a particular emphasis on GAN-derived models.\nI also provide a set of 7 desiderata followed by an evaluation of whether a\ngiven measure or a family of measures is compatible with them.', 'Limited data access is a longstanding barrier to data-driven research and\ndevelopment in the networked systems community. In this work, we explore if and\nhow generative adversarial networks (GANs) can be used to incentivize data\nsharing by enabling a generic framework for sharing synthetic datasets with\nminimal expert knowledge. As a specific target, our focus in this paper is on\ntime series datasets with metadata (e.g., packet loss rate measurements with\ncorresponding ISPs). We identify key challenges of existing GAN approaches for\nsuch workloads with respect to fidelity (e.g., long-term dependencies, complex\nmultidimensional relationships, mode collapse) and privacy (i.e., existing\nguarantees are poorly understood and can sacrifice fidelity). To improve\nfidelity, we design a custom workflow called DoppelGANger (DG) and demonstrate\nthat across diverse real-world datasets (e.g., bandwidth measurements, cluster\nrequests, web sessions) and use cases (e.g., structural characterization,\npredictive modeling, algorithm comparison), DG achieves up to 43% better\nfidelity than baseline models. Although we do not resolve the privacy problem\nin this work, we identify fundamental challenges with both classical notions of\nprivacy and recent advances to improve the privacy properties of GANs, and\nsuggest a potential roadmap for addressing these challenges. By shedding light\non the promise and challenges, we hope our work can rekindle the conversation\non workflows for data sharing.', 'A basic, and still largely unanswered, question in the context of Generative\nAdversarial Networks (GANs) is whether they are truly able to capture all the\nfundamental characteristics of the distributions they are trained on. In\nparticular, evaluating the diversity of GAN distributions is challenging and\nexisting methods provide only a partial understanding of this issue. In this\npaper, we develop quantitative and scalable tools for assessing the diversity\nof GAN distributions. Specifically, we take a classification-based perspective\nand view loss of diversity as a form of covariate shift introduced by GANs. We\nexamine two specific forms of such shift: mode collapse and boundary\ndistortion. In contrast to prior work, our methods need only minimal human\nsupervision and can be readily applied to state-of-the-art GANs on large,\ncanonical datasets. Examining popular GANs using our tools indicates that these\nGANs have significant problems in reproducing the more distributional\nproperties of their training dataset.', 'Training and using modern neural-network based latent-variable generative\nmodels (like Variational Autoencoders) often require simultaneously training a\ngenerative direction along with an inferential(encoding) direction, which\napproximates the posterior distribution over the latent variables. Thus, the\nquestion arises: how complex does the inferential model need to be, in order to\nbe able to accurately model the posterior distribution of a given generative\nmodel?\n  In this paper, we identify an important property of the generative map\nimpacting the required size of the encoder. We show that if the generative map\nis ""strongly invertible"" (in a sense we suitably formalize), the inferential\nmodel need not be much more complex. Conversely, we prove that there exist\nnon-invertible generative maps, for which the encoding direction needs to be\nexponentially larger (under standard assumptions in computational complexity).\nImportantly, we do not require the generative model to be layerwise invertible,\nwhich a lot of the related literature assumes and isn\'t satisfied by many\narchitectures used in practice (e.g. convolution and pooling based networks).\nThus, we provide theoretical support for the empirical wisdom that learning\ndeep generative models is harder when data lies on a low-dimensional manifold.', 'Deep generative models are powerful tools that have produced impressive\nresults in recent years. These advances have been for the most part empirically\ndriven, making it essential that we use high quality evaluation metrics. In\nthis paper, we provide new insights into the Inception Score, a recently\nproposed and widely used evaluation metric for generative models, and\ndemonstrate that it fails to provide useful guidance when comparing models. We\ndiscuss both suboptimalities of the metric itself and issues with its\napplication. Finally, we call for researchers to be more systematic and careful\nwhen evaluating and comparing generative models, as the advancement of the\nfield depends upon it.', 'Recent studies have shown that text-to-speech synthesis quality can be\nimproved by using glottal vocoding. This refers to vocoders that parameterize\nspeech into two parts, the glottal excitation and vocal tract, that occur in\nthe human speech production apparatus. Current glottal vocoders generate the\nglottal excitation waveform by using deep neural networks (DNNs). However, the\nsquared error-based training of the present glottal excitation models is\nlimited to generating conditional average waveforms, which fails to capture the\nstochastic variation of the waveforms. As a result, shaped noise is added as\npost-processing. In this study, we propose a new method for predicting glottal\nwaveforms by generative adversarial networks (GANs). GANs are generative models\nthat aim to embed the data distribution in a latent space, enabling generation\nof new instances very similar to the original by randomly sampling the latent\ndistribution. The glottal pulses generated by GANs show a stochastic component\nsimilar to natural glottal pulses. In our experiments, we compare synthetic\nspeech generated using glottal waveforms produced by both DNNs and GANs. The\nresults show that the newly proposed GANs achieve synthesis quality comparable\nto that of widely-used DNNs, without using an additive noise component.', 'We propose a novel architecture which is able to automatically anonymize\nfaces in images while retaining the original data distribution. We ensure total\nanonymization of all faces in an image by generating images exclusively on\nprivacy-safe information. Our model is based on a conditional generative\nadversarial network, generating images considering the original pose and image\nbackground. The conditional information enables us to generate highly realistic\nfaces with a seamless transition between the generated face and the existing\nbackground. Furthermore, we introduce a diverse dataset of human faces,\nincluding unconventional poses, occluded faces, and a vast variability in\nbackgrounds. Finally, we present experimental results reflecting the capability\nof our model to anonymize images while preserving the data distribution, making\nthe data suitable for further training of deep learning models. As far as we\nknow, no other solution has been proposed that guarantees the anonymization of\nfaces while generating realistic images.', 'Generative Adversarial Networks (GANs) are now capable of producing synthetic\nface images of exceptionally high visual quality. In parallel to the\ndevelopment of GANs themselves, efforts have been made to develop metrics to\nobjectively assess the characteristics of the synthetic images, mainly focusing\non visual quality and the variety of images. Little work has been done,\nhowever, to assess overfitting of GANs and their ability to generate new\nidentities. In this paper we apply a state of the art biometric network to\nvarious datasets of synthetic images and perform a thorough assessment of their\nidentity-related characteristics. We conclude that GANs can indeed be used to\ngenerate new, imagined identities meaning that applications such as\nanonymisation of image sets and augmentation of training datasets with\ndistractor images are viable applications. We also assess the ability of GANs\nto disentangle identity from other image characteristics and propose a novel\nGAN triplet loss that we show to improve this disentanglement.', 'This paper studies a training method to jointly estimate an energy-based\nmodel and a flow-based model, in which the two models are iteratively updated\nbased on a shared adversarial value function. This joint training method has\nthe following traits. (1) The update of the energy-based model is based on\nnoise contrastive estimation, with the flow model serving as a strong noise\ndistribution. (2) The update of the flow model approximately minimizes the\nJensen-Shannon divergence between the flow model and the data distribution. (3)\nUnlike generative adversarial networks (GAN) which estimates an implicit\nprobability distribution defined by a generator model, our method estimates two\nexplicit probabilistic distributions on the data. Using the proposed method we\ndemonstrate a significant improvement on the synthesis quality of the flow\nmodel, and show the effectiveness of unsupervised feature learning by the\nlearned energy-based model. Furthermore, the proposed training method can be\neasily adapted to semi-supervised learning. We achieve competitive results to\nthe state-of-the-art semi-supervised learning methods.', 'Class imbalance is a substantial challenge in classifying many real-world\ncases. Synthetic over-sampling methods have been effective to improve the\nperformance of classifiers for imbalance problems. However, most synthetic\nover-sampling methods generate non-diverse synthetic instances within the\nconvex hull formed by the existing minority instances as they only concentrate\non the minority class and ignore the vast information provided by the majority\nclass. They also often do not perform well for extremely imbalanced data as the\nfewer the minority instances, the less information to generate synthetic\ninstances. Moreover, existing methods that generate synthetic instances using\nthe majority class distributional information cannot perform effectively when\nthe majority class has a multi-modal distribution. We propose a new method to\ngenerate diverse and adaptable synthetic instances using Synthetic\nOver-sampling with Minority and Majority classes (SOMM). SOMM generates\nsynthetic instances diversely within the minority data space. It updates the\ngenerated instances adaptively to the neighbourhood including both classes.\nThus, SOMM performs well for both binary and multiclass imbalance problems. We\nexamine the performance of SOMM for binary and multiclass problems using\nbenchmark data sets for different imbalance levels. The empirical results show\nthe superiority of SOMM compared to other existing methods.', 'Generative models are increasingly able to produce remarkably high quality\nimages and text. The community has developed numerous evaluation metrics for\ncomparing generative models. However, these metrics do not effectively quantify\ndata diversity. We develop a new diversity metric that can readily be applied\nto data, both synthetic and natural, of any type. Our method employs random\nnetwork distillation, a technique introduced in reinforcement learning. We\nvalidate and deploy this metric on both images and text. We further explore\ndiversity in few-shot image generation, a setting which was previously\ndifficult to evaluate.', 'Recently, transformation-based self-supervised learning has been applied to\ngenerative adversarial networks (GANs) to mitigate catastrophic forgetting in\nthe discriminator by introducing a stationary learning environment. However,\nthe separate self-supervised tasks in existing self-supervised GANs cause a\ngoal inconsistent with generative modeling due to the fact that their\nself-supervised classifiers are agnostic to the generator distribution. To\naddress this problem, we propose a novel self-supervised GAN that unifies the\nGAN task with the self-supervised task by augmenting the GAN labels (real or\nfake) via self-supervision of data transformation. Specifically, the original\ndiscriminator and self-supervised classifier are unified into a label-augmented\ndiscriminator that predicts the augmented labels to be aware of both the\ngenerator distribution and the data distribution under every transformation,\nand then provide the discrepancy between them to optimize the generator.\nTheoretically, we prove that the optimal generator could converge to replicate\nthe real data distribution. Empirically, we show that the proposed method\nsignificantly outperforms previous self-supervised and data augmentation GANs\non both generative modeling and representation learning across benchmark\ndatasets.', ""Generative Adversarial Networks (GANs) have made great progress in\nsynthesizing realistic images in recent years. However, they are often trained\non image datasets with either too few samples or too many classes belonging to\ndifferent data distributions. Consequently, GANs are prone to underfitting or\noverfitting, making the analysis of them difficult and constrained. Therefore,\nin order to conduct a thorough study on GANs while obviating unnecessary\ninterferences introduced by the datasets, we train them on artificial datasets\nwhere there are infinitely many samples and the real data distributions are\nsimple, high-dimensional and have structured manifolds. Moreover, the\ngenerators are designed such that optimal sets of parameters exist.\nEmpirically, we find that under various distance measures, the generator fails\nto learn such parameters with the GAN training procedure. We also find that\ntraining mixtures of GANs leads to more performance gain compared to increasing\nthe network depth or width when the model complexity is high enough. Our\nexperimental results demonstrate that a mixture of generators can discover\ndifferent modes or different classes automatically in an unsupervised setting,\nwhich we attribute to the distribution of the generation and discrimination\ntasks across multiple generators and discriminators. As an example of the\ngeneralizability of our conclusions to realistic datasets, we train a mixture\nof GANs on the CIFAR-10 dataset and our method significantly outperforms the\nstate-of-the-art in terms of popular metrics, i.e., Inception Score (IS) and\nFr\\'echet Inception Distance (FID)."", 'Probabilistic generative models can be used for compression, denoising,\ninpainting, texture synthesis, semi-supervised learning, unsupervised feature\nlearning, and other tasks. Given this wide range of applications, it is not\nsurprising that a lot of heterogeneity exists in the way these models are\nformulated, trained, and evaluated. As a consequence, direct comparison between\nmodels is often difficult. This article reviews mostly known but often\nunderappreciated properties relating to the evaluation and interpretation of\ngenerative models with a focus on image models. In particular, we show that\nthree of the currently most commonly used criteria---average log-likelihood,\nParzen window estimates, and visual fidelity of samples---are largely\nindependent of each other when the data is high-dimensional. Good performance\nwith respect to one criterion therefore need not imply good performance with\nrespect to the other criteria. Our results show that extrapolation from one\ncriterion to another is not warranted and generative models need to be\nevaluated directly with respect to the application(s) they were intended for.\nIn addition, we provide examples demonstrating that Parzen window estimates\nshould generally be avoided.', 'In terms of Generative Adversarial Networks (GANs), the information metric to\ndiscriminate the generative data from the real data, lies in the key point of\ngeneration efficiency, which plays an important role in GAN-based applications,\nespecially in anomaly detection. As for the original GAN, there exist drawbacks\nfor its hidden information measure based on KL divergence on rare events\ngeneration and training performance for adversarial networks. Therefore, it is\nsignificant to investigate the metrics used in GANs to improve the generation\nability as well as bring gains in the training process. In this paper, we adopt\nthe exponential form, referred from the information measure, i.e. MIM, to\nreplace the logarithm form of the original GAN. This approach is called\nMIM-based GAN, has better performance on networks training and rare events\ngeneration. Specifically, we first discuss the characteristics of training\nprocess in this approach. Moreover, we also analyze its advantages on\ngenerating rare events in theory. In addition, we do simulations on the\ndatasets of MNIST and ODDS to see that the MIM-based GAN achieves\nstate-of-the-art performance on anomaly detection compared with some classical\nGANs.', 'Generative adversarial networks (GANs) are an exciting alternative to\nalgorithms for solving density estimation problems---using data to assess how\nlikely samples are to be drawn from the same distribution. Instead of\nexplicitly computing these probabilities, GANs learn a generator that can match\nthe given probabilistic source. This paper looks particularly at this matching\ncapability in the context of problems with one-dimensional outputs. We identify\na class of function decompositions with properties that make them well suited\nto the critic role in a leading approach to GANs known as Wasserstein GANs. We\nshow that Taylor and Fourier series decompositions belong to our class, provide\nexamples of these critics outperforming standard GAN approaches, and suggest\nhow they can be scaled to higher dimensional problems in the future.', 'Generative Adversarial Networks (GANs) have been used in several machine\nlearning tasks such as domain transfer, super resolution, and synthetic data\ngeneration. State-of-the-art GANs often use tens of millions of parameters,\nmaking them expensive to deploy for applications in low SWAP (size, weight, and\npower) hardware, such as mobile devices, and for applications with real time\ncapabilities. There has been no work found to reduce the number of parameters\nused in GANs. Therefore, we propose a method to compress GANs using knowledge\ndistillation techniques, in which a smaller ""student"" GAN learns to mimic a\nlarger ""teacher"" GAN. We show that the distillation methods used on MNIST,\nCIFAR-10, and Celeb-A datasets can compress teacher GANs at ratios of 1669:1,\n58:1, and 87:1, respectively, while retaining the quality of the generated\nimage. From our experiments, we observe a qualitative limit for GAN\'s\ncompression. Moreover, we observe that, with a fixed parameter budget,\ncompressed GANs outperform GANs trained using standard training methods. We\nconjecture that this is partially owing to the optimization landscape of\nover-parameterized GANs which allows efficient training using alternating\ngradient descent. Thus, training an over-parameterized GAN followed by our\nproposed compression scheme provides a high quality generative model with a\nsmall number of parameters.', 'Recent work has shown that Neural Ordinary Differential Equations (ODEs) can\nserve as generative models of images using the perspective of Continuous\nNormalizing Flows (CNFs). Such models offer exact likelihood calculation, and\ninvertible generation/density estimation. In this work we introduce a\nMulti-Resolution variant of such models (MRCNF), by characterizing the\nconditional distribution over the additional information required to generate a\nfine image that is consistent with the coarse image. We introduce a\ntransformation between resolutions that allows for no change in the log\nlikelihood. We show that this approach yields comparable likelihood values for\nvarious image datasets, with improved performance at higher resolutions, with\nfewer parameters, using only 1 GPU. Further, we examine the out-of-distribution\nproperties of (Multi-Resolution) Continuous Normalizing Flows, and find that\nthey are similar to those of other likelihood-based generative models.', 'Creating noise from data is easy; creating data from noise is generative\nmodeling. We present a stochastic differential equation (SDE) that smoothly\ntransforms a complex data distribution to a known prior distribution by slowly\ninjecting noise, and a corresponding reverse-time SDE that transforms the prior\ndistribution back into the data distribution by slowly removing the noise.\nCrucially, the reverse-time SDE depends only on the time-dependent gradient\nfield (\\aka, score) of the perturbed data distribution. By leveraging advances\nin score-based generative modeling, we can accurately estimate these scores\nwith neural networks, and use numerical SDE solvers to generate samples. We\nshow that this framework encapsulates previous approaches in score-based\ngenerative modeling and diffusion probabilistic modeling, allowing for new\nsampling procedures and new modeling capabilities. In particular, we introduce\na predictor-corrector framework to correct errors in the evolution of the\ndiscretized reverse-time SDE. We also derive an equivalent neural ODE that\nsamples from the same distribution as the SDE, but additionally enables exact\nlikelihood computation, and improved sampling efficiency. In addition, we\nprovide a new way to solve inverse problems with score-based models, as\ndemonstrated with experiments on class-conditional generation, image\ninpainting, and colorization. Combined with multiple architectural\nimprovements, we achieve record-breaking performance for unconditional image\ngeneration on CIFAR-10 with an Inception score of 9.89 and FID of 2.20, a\ncompetitive likelihood of 2.99 bits/dim, and demonstrate high fidelity\ngeneration of 1024 x 1024 images for the first time from a score-based\ngenerative model.', 'Generative adversarial networks (GANs) are increasingly attracting attention\nin the computer vision, natural language processing, speech synthesis and\nsimilar domains. Arguably the most striking results have been in the area of\nimage synthesis. However, evaluating the performance of GANs is still an open\nand challenging problem. Existing evaluation metrics primarily measure the\ndissimilarity between real and generated images using automated statistical\nmethods. They often require large sample sizes for evaluation and do not\ndirectly reflect human perception of image quality. In this work, we describe\nan evaluation metric we call Neuroscore, for evaluating the performance of\nGANs, that more directly reflects psychoperceptual image quality through the\nutilization of brain signals. Our results show that Neuroscore has superior\nperformance to the current evaluation metrics in that: (1) It is more\nconsistent with human judgment; (2) The evaluation process needs much smaller\nnumbers of samples; and (3) It is able to rank the quality of images on a per\nGAN basis. A convolutional neural network (CNN) based neuro-AI interface is\nproposed to predict Neuroscore from GAN-generated images directly without the\nneed for neural responses. Importantly, we show that including neural responses\nduring the training phase of the network can significantly improve the\nprediction capability of the proposed model. Materials related to this work are\nprovided at https://github.com/villawang/Neuro-AI-Interface.', 'Generative adversarial networks (GANs) provide a way to learn deep\nrepresentations without extensively annotated training data. They achieve this\nthrough deriving backpropagation signals through a competitive process\ninvolving a pair of networks. The representations that can be learned by GANs\nmay be used in a variety of applications, including image synthesis, semantic\nimage editing, style transfer, image super-resolution and classification. The\naim of this review paper is to provide an overview of GANs for the signal\nprocessing community, drawing on familiar analogies and concepts where\npossible. In addition to identifying different methods for training and\nconstructing GANs, we also point to remaining challenges in their theory and\napplication.', ""In this paper, we investigate the effect of machine learning based\nanonymization on anomalous subgroup preservation. In particular, we train a\nbinary classifier to discover the most anomalous subgroup in a dataset by\nmaximizing the bias between the group's predicted odds ratio from the model and\nobserved odds ratio from the data. We then perform anonymization using a\nvariational autoencoder (VAE) to synthesize an entirely new dataset that would\nideally be drawn from the distribution of the original data. We repeat the\nanomalous subgroup discovery task on the new data and compare it to what was\nidentified pre-anonymization. We evaluated our approach using publicly\navailable datasets from the financial industry. Our evaluation confirmed that\nthe approach was able to produce synthetic datasets that preserved a high level\nof subgroup differentiation as identified initially in the original dataset.\nSuch a distinction was maintained while having distinctly different records\nbetween the synthetic and original dataset. Finally, we packed the above end to\nend process into what we call Utility Guaranteed Deep Privacy (UGDP) system.\nUGDP can be easily extended to onboard alternative generative approaches such\nas GANs to synthesize tabular data."", 'Generative Adversarial Networks (Goodfellow et al., 2014), a major\nbreakthrough in the field of generative modeling, learn a discriminator to\nestimate some distance between the target and the candidate distributions.\n  This paper examines mathematical issues regarding the way the gradients for\nthe generative model are computed in this context, and notably how to take into\naccount how the discriminator itself depends on the generator parameters.\n  A unifying methodology is presented to define mathematically sound training\nobjectives for generative models taking this dependency into account in a\nrobust way, covering both GAN, VAE and some GAN variants as particular cases.', 'In modern data analysis, random sampling is an efficient and widely-used\nstrategy to overcome the computational difficulties brought by large sample\nsize. In previous studies, researchers conducted random sampling which is\naccording to the input data but independent on the response variable, however\nthe response variable may also be informative for sampling. In this paper we\npropose an adaptive sampling called the gradient-based sampling which is\ndependent on both the input data and the output for fast solving of\nleast-square (LS) problems. We draw the data points by random sampling from the\nfull data according to their gradient values. This sampling is computationally\nsaving, since the running time of computing the sampling probabilities is\nreduced to O(nd) where n is the full sample size and d is the dimension of the\ninput. Theoretically, we establish an error bound analysis of the general\nimportance sampling with respect to LS solution from full data. The result\nestablishes an improved performance of the use of our gradient- based sampling.\nSynthetic and real data sets are used to empirically argue that the\ngradient-based sampling has an obvious advantage over existing sampling methods\nfrom two aspects of statistical efficiency and computational saving.', 'Generative Adversarial Networks (GANs) have extended deep learning to complex\ngeneration and translation tasks across different data modalities. However,\nGANs are notoriously difficult to train: Mode collapse and other instabilities\nin the training process often degrade the quality of the generated results,\nsuch as images. This paper presents a new technique called TaylorGAN for\nimproving GANs by discovering customized loss functions for each of its two\nnetworks. The loss functions are parameterized as Taylor expansions and\noptimized through multiobjective evolution. On an image-to-image translation\nbenchmark task, this approach qualitatively improves generated image quality\nand quantitatively improves two independent GAN performance metrics. It\ntherefore forms a promising approach for applying GANs to more challenging\ntasks in the future.', 'In many applications of computer graphics, art and design, it is desirable\nfor a user to provide intuitive non-image input, such as text, sketch, stroke,\ngraph or layout, and have a computer system automatically generate\nphoto-realistic images that adhere to the input content. While classic works\nthat allow such automatic image content generation have followed a framework of\nimage retrieval and composition, recent advances in deep generative models such\nas generative adversarial networks (GANs), variational autoencoders (VAEs), and\nflow-based methods have enabled more powerful and versatile image generation\ntasks. This paper reviews recent works for image synthesis given intuitive user\ninput, covering advances in input versatility, image generation methodology,\nbenchmark datasets, and evaluation metrics. This motivates new perspectives on\ninput representation and interactivity, cross pollination between major image\ngeneration paradigms, and evaluation and comparison of generation methods.', 'Generative concept representations have three major advantages over\ndiscriminative ones: they can represent uncertainty, they support integration\nof learning and reasoning, and they are good for unsupervised and\nsemi-supervised learning. We discuss probabilistic and generative deep\nlearning, which generative concept representations are based on, and the use of\nvariational autoencoders and generative adversarial networks for learning\ngenerative concept representations, particularly for concepts whose data are\nsequences, structured data or graphs.', 'Generative adversarial networks or GANs are a type of generative modeling\nframework. GANs involve a pair of neural networks engaged in a competition in\niteratively creating fake data, indistinguishable from the real data. One\nnotable application of GANs is developing fake human faces, also known as ""deep\nfakes,"" due to the deep learning algorithms at the core of the GAN framework.\nMeasuring the quality of the generated images is inherently subjective but\nattempts to objectify quality using standardized metrics have been made. One\nexample of objective metrics is the Frechet Inception Distance (FID), which\nmeasures the difference between distributions of feature vectors for two\nseparate datasets of images. There are situations that images with low\nperceptual qualities are not assigned appropriate FID scores. We propose to\nimprove the robustness of the evaluation process by integrating lower-level\nfeatures to cover a wider array of visual defects. Our proposed method\nintegrates three levels of feature abstractions to evaluate the quality of\ngenerated images. Experimental evaluations show better performance of the\nproposed method for distorted images.', 'Generative Adversarial Networks (GAN) have received wide attention in the\nmachine learning field for their potential to learn high-dimensional, complex\nreal data distribution. Specifically, they do not rely on any assumptions about\nthe distribution and can generate real-like samples from latent space in a\nsimple manner. This powerful property leads GAN to be applied to various\napplications such as image synthesis, image attribute editing, image\ntranslation, domain adaptation and other academic fields. In this paper, we aim\nto discuss the details of GAN for those readers who are familiar with, but do\nnot comprehend GAN deeply or who wish to view GAN from various perspectives. In\naddition, we explain how GAN operates and the fundamental meaning of various\nobjective functions that have been suggested recently. We then focus on how the\nGAN can be combined with an autoencoder framework. Finally, we enumerate the\nGAN variants that are applied to various tasks and other fields for those who\nare interested in exploiting GAN for their research.', 'Assessing generative models is not an easy task. Generative models should\nsynthesize graphs which are not replicates of real networks but show\ntopological features similar to real graphs. We introduce an approach for\nassessing graph generative models using graph classifiers. The inability of an\nestablished graph classifier for distinguishing real and synthesized graphs\ncould be considered as a performance measurement for graph generators.', 'This is a tutorial and survey paper on Generative Adversarial Network (GAN),\nadversarial autoencoders, and their variants. We start with explaining\nadversarial learning and the vanilla GAN. Then, we explain the conditional GAN\nand DCGAN. The mode collapse problem is introduced and various methods,\nincluding minibatch GAN, unrolled GAN, BourGAN, mixture GAN, D2GAN, and\nWasserstein GAN, are introduced for resolving this problem. Then, maximum\nlikelihood estimation in GAN are explained along with f-GAN, adversarial\nvariational Bayes, and Bayesian GAN. Then, we cover feature matching in GAN,\nInfoGAN, GRAN, LSGAN, energy-based GAN, CatGAN, MMD GAN, LapGAN, progressive\nGAN, triple GAN, LAG, GMAN, AdaGAN, CoGAN, inverse GAN, BiGAN, ALI, SAGAN,\nFew-shot GAN, SinGAN, and interpolation and evaluation of GAN. Then, we\nintroduce some applications of GAN such as image-to-image translation\n(including PatchGAN, CycleGAN, DeepFaceDrawing, simulated GAN, interactive\nGAN), text-to-image translation (including StackGAN), and mixing image\ncharacteristics (including FineGAN and MixNMatch). Finally, we explain the\nautoencoders based on adversarial learning including adversarial autoencoder,\nPixelGAN, and implicit autoencoder.', 'This paper examines to what degree current deep learning architectures for\nimage caption generation capture spatial language. On the basis of the\nevaluation of examples of generated captions from the literature we argue that\nsystems capture what objects are in the image data but not where these objects\nare located: the captions generated by these systems are the output of a\nlanguage model conditioned on the output of an object detector that cannot\ncapture fine-grained location information. Although language models provide\nuseful knowledge for image captions, we argue that deep learning image\ncaptioning architectures should also model geometric relations between objects.', ""Generative Adversarial Networks (GANs) have been used extensively and quite\nsuccessfully for unsupervised learning. As GANs don't approximate an explicit\nprobability distribution, it's an interesting study to inspect the latent space\nrepresentations learned by GANs. The current work seeks to push the boundaries\nof such inspection methods to further understand in more detail the manifold\nbeing learned by GANs. Various interpolation and extrapolation techniques along\nwith vector arithmetic is used to understand the learned manifold. We show\nthrough experiments that GANs indeed learn a data probability distribution\nrather than memorize images/data. Further, we prove that GANs encode\nsemantically relevant information in the learned probability distribution. The\nexperiments have been performed on two publicly available datasets - Large\nScale Scene Understanding (LSUN) and CelebA."", 'In this paper, a novel strategy of Secure Steganograpy based on Generative\nAdversarial Networks is proposed to generate suitable and secure covers for\nsteganography. The proposed architecture has one generative network, and two\ndiscriminative networks. The generative network mainly evaluates the visual\nquality of the generated images for steganography, and the discriminative\nnetworks are utilized to assess their suitableness for information hiding.\nDifferent from the existing work which adopts Deep Convolutional Generative\nAdversarial Networks, we utilize another form of generative adversarial\nnetworks. By using this new form of generative adversarial networks,\nsignificant improvements are made on the convergence speed, the training\nstability and the image quality. Furthermore, a sophisticated steganalysis\nnetwork is reconstructed for the discriminative network, and the network can\nbetter evaluate the performance of the generated images. Numerous experiments\nare conducted on the publicly available datasets to demonstrate the\neffectiveness and robustness of the proposed method.', ""Devising indicative evaluation metrics for the image generation task remains\nan open problem. The most widely used metric for measuring the similarity\nbetween real and generated images has been the Fr\\'echet Inception Distance\n(FID) score. Because it does not differentiate the fidelity and diversity\naspects of the generated images, recent papers have introduced variants of\nprecision and recall metrics to diagnose those properties separately. In this\npaper, we show that even the latest version of the precision and recall metrics\nare not reliable yet. For example, they fail to detect the match between two\nidentical distributions, they are not robust against outliers, and the\nevaluation hyperparameters are selected arbitrarily. We propose density and\ncoverage metrics that solve the above issues. We analytically and\nexperimentally show that density and coverage provide more interpretable and\nreliable signals for practitioners than the existing metrics. Code:\nhttps://github.com/clovaai/generative-evaluation-prdc."", 'Generative adversarial networks (GANs) are increasingly attracting attention\nin the computer vision, natural language processing, speech synthesis and\nsimilar domains. However, evaluating the performance of GANs is still an open\nand challenging problem. Existing evaluation metrics primarily measure the\ndissimilarity between real and generated images using automated statistical\nmethods. They often require large sample sizes for evaluation and do not\ndirectly reflect human perception of image quality. In this work, we introduce\nan evaluation metric called Neuroscore, for evaluating the performance of GANs,\nthat more directly reflects psychoperceptual image quality through the\nutilization of brain signals. Our results show that Neuroscore has superior\nperformance to the current evaluation metrics in that: (1) It is more\nconsistent with human judgment; (2) The evaluation process needs much smaller\nnumbers of samples; and (3) It is able to rank the quality of images on a per\nGAN basis. A convolutional neural network (CNN) based neuro-AI interface is\nproposed to predict Neuroscore from GAN-generated images directly without the\nneed for neural responses. Importantly, we show that including neural responses\nduring the training phase of the network can significantly improve the\nprediction capability of the proposed model. Codes and data can be referred at\nthis link: https://github.com/villawang/Neuro-AI-Interface.', ""Generative Adversarial Networks (GANs) have become a dominant class of\ngenerative models. In recent years, GAN variants have yielded especially\nimpressive results in the synthesis of a variety of forms of data. Examples\ninclude compelling natural and artistic images, textures, musical sequences,\nand 3D object files. However, one obvious synthesis candidate is missing. In\nthis work, we answer one of deep learning's most pressing questions: GAN you do\nthe GAN GAN? That is, is it possible to train a GAN to model a distribution of\nGANs? We release the full source code for this project under the MIT license."", ""While Generative Adversarial Networks (GANs) show increasing performance and\nthe level of realism is becoming indistinguishable from natural images, this\nalso comes with high demands on data and computation. We show that\nstate-of-the-art GAN models -- such as they are being publicly released by\nresearchers and industry -- can be used for a range of applications beyond\nunconditional image generation. We achieve this by an iterative scheme that\nalso allows gaining control over the image generation process despite the\nhighly non-linear latent spaces of the latest GAN models. We demonstrate that\nthis opens up the possibility to re-use state-of-the-art, difficult to train,\npre-trained GANs with a high level of control even if only black-box access is\ngranted. Our work also raises concerns and awareness that the use cases of a\npublished GAN model may well reach beyond the creators' intention, which needs\nto be taken into account before a full public release. Code is available at\nhttps://github.com/a514514772/hijackgan."", 'Despite recent advances, the remaining bottlenecks in deep generative models\nare necessity of extensive training and difficulties with generalization from\nsmall number of training examples. We develop a new generative model called\nGenerative Matching Network which is inspired by the recently proposed matching\nnetworks for one-shot learning in discriminative tasks. By conditioning on the\nadditional input dataset, our model can instantly learn new concepts that were\nnot available in the training data but conform to a similar generative process.\nThe proposed framework does not explicitly restrict diversity of the\nconditioning data and also does not require an extensive inference procedure\nfor training or adaptation. Our experiments on the Omniglot dataset demonstrate\nthat Generative Matching Networks significantly improve predictive performance\non the fly as more additional data is available and outperform existing state\nof the art conditional generative models.', 'Automated evaluation metrics as a stand-in for manual evaluation are an\nessential part of the development of text-generation tasks such as text\nsummarization. However, while the field has progressed, our standard metrics\nhave not -- for nearly 20 years ROUGE has been the standard evaluation in most\nsummarization papers. In this paper, we make an attempt to re-evaluate the\nevaluation method for text summarization: assessing the reliability of\nautomatic metrics using top-scoring system outputs, both abstractive and\nextractive, on recently popular datasets for both system-level and\nsummary-level evaluation settings. We find that conclusions about evaluation\nmetrics on older datasets do not necessarily hold on modern datasets and\nsystems.', 'Tremendous advances in image restoration tasks such as denoising and\nsuper-resolution have been achieved using neural networks. Such approaches\ngenerally employ very deep architectures, large number of parameters, large\nreceptive fields and high nonlinear modeling capacity. In order to obtain\nefficient and fast image restoration networks one should improve upon the above\nmentioned requirements.\n  In this paper we propose a novel activation function, the multi-bin trainable\nlinear unit (MTLU), for increasing the nonlinear modeling capacity together\nwith lighter and shallower networks. We validate the proposed fast image\nrestoration networks for image denoising (FDnet) and super-resolution (FSRnet)\non standard benchmarks. We achieve large improvements in both memory and\nruntime over current state-of-the-art for comparable or better PSNR accuracies.', 'As information exists in various modalities in real world, effective\ninteraction and fusion among multimodal information plays a key role for the\ncreation and perception of multimodal data in computer vision and deep learning\nresearch. With superb power in modelling the interaction among multimodal\ninformation, multimodal image synthesis and editing has become a hot research\ntopic in recent years. Instead of providing explicit guidance for network\ntraining, multimodal guidance offers intuitive and flexible means for image\nsynthesis and editing. On the other hand, this field is also facing several\nchallenges in alignment of features with inherent modality gaps, synthesis of\nhigh-resolution images, faithful evaluation metrics, etc. In this survey, we\ncomprehensively contextualize the advance of the recent multimodal image\nsynthesis and editing and formulate taxonomies according to data modality and\nmodel architectures. We start with an introduction to different types of\nguidance modalities in image synthesis and editing. We then describe multimodal\nimage synthesis and editing approaches extensively with detailed frameworks\nincluding Generative Adversarial Networks (GANs), Auto-regressive models,\nDiffusion models, Neural Radiance Fields (NeRF) and other methods. This is\nfollowed by a comprehensive description of benchmark datasets and corresponding\nevaluation metrics as widely adopted in multimodal image synthesis and editing,\nas well as detailed comparisons of various synthesis methods with analysis of\nrespective advantages and limitations. Finally, we provide insights about the\ncurrent research challenges and possible directions for future research. We\nhope this survey could lay a sound and valuable foundation for future\ndevelopment of multimodal image synthesis and editing. A project associated\nwith this survey is available at https://github.com/fnzhan/MISE.', 'Generative adversarial networks (GANs) are a novel approach to generative\nmodelling, a task whose goal it is to learn a distribution of real data points.\nThey have often proved difficult to train: GANs are unlike many techniques in\nmachine learning, in that they are best described as a two-player game between\na discriminator and generator. This has yielded both unreliability in the\ntraining process, and a general lack of understanding as to how GANs converge,\nand if so, to what. The purpose of this dissertation is to provide an account\nof the theory of GANs suitable for the mathematician, highlighting both\npositive and negative results. This involves identifying the problems when\ntraining GANs, and how topological and game-theoretic perspectives of GANs have\ncontributed to our understanding and improved our techniques in recent years.', 'With the recent progress in Generative Adversarial Networks (GANs), it is\nimperative for media and visual forensics to develop detectors which can\nidentify and attribute images to the model generating them. Existing works have\nshown to attribute images to their corresponding GAN sources with high\naccuracy. However, these works are limited to a closed set scenario, failing to\ngeneralize to GANs unseen during train time and are therefore, not scalable\nwith a steady influx of new GANs. We present an iterative algorithm for\ndiscovering images generated from previously unseen GANs by exploiting the fact\nthat all GANs leave distinct fingerprints on their generated images. Our\nalgorithm consists of multiple components including network training,\nout-of-distribution detection, clustering, merge and refine steps. Through\nextensive experiments, we show that our algorithm discovers unseen GANs with\nhigh accuracy and also generalizes to GANs trained on unseen real datasets. We\nadditionally apply our algorithm to attribution and discovery of GANs in an\nonline fashion as well as to the more standard task of real/fake detection. Our\nexperiments demonstrate the effectiveness of our approach to discover new GANs\nand can be used in an open-world setup.', 'Generative adversarial networks (GANs) are a hot research topic recently.\nGANs have been widely studied since 2014, and a large number of algorithms have\nbeen proposed. However, there is few comprehensive study explaining the\nconnections among different GANs variants, and how they have evolved. In this\npaper, we attempt to provide a review on various GANs methods from the\nperspectives of algorithms, theory, and applications. Firstly, the motivations,\nmathematical representations, and structure of most GANs algorithms are\nintroduced in details. Furthermore, GANs have been combined with other machine\nlearning algorithms for specific applications, such as semi-supervised\nlearning, transfer learning, and reinforcement learning. This paper compares\nthe commonalities and differences of these GANs methods. Secondly, theoretical\nissues related to GANs are investigated. Thirdly, typical applications of GANs\nin image processing and computer vision, natural language processing, music,\nspeech and audio, medical field, and data science are illustrated. Finally, the\nfuture open research problems for GANs are pointed out.', 'We demonstrate a conditional autoregressive pipeline for efficient music\nrecomposition, based on methods presented in van den Oord et al.(2017).\nRecomposition (Casal & Casey, 2010) focuses on reworking existing musical\npieces, adhering to structure at a high level while also re-imagining other\naspects of the work. This can involve reuse of pre-existing themes or parts of\nthe original piece, while also requiring the flexibility to generate new\ncontent at different levels of granularity. Applying the aforementioned\nmodeling pipeline to recomposition, we show diverse and structured generation\nconditioned on chord sequence annotations.', 'Deep learning has achieved overwhelming success, spanning from discriminative\nmodels to generative models. In particular, deep generative models have\nfacilitated a new level of performance in a myriad of areas, ranging from media\nmanipulation to sanitized dataset generation. Despite the great success, the\npotential risks of privacy breach caused by generative models have not been\nanalyzed systematically. In this paper, we focus on membership inference attack\nagainst deep generative models that reveals information about the training data\nused for victim models. Specifically, we present the first taxonomy of\nmembership inference attacks, encompassing not only existing attacks but also\nour novel ones. In addition, we propose the first generic attack model that can\nbe instantiated in a large range of settings and is applicable to various kinds\nof deep generative models. Moreover, we provide a theoretically grounded attack\ncalibration technique, which consistently boosts the attack performance in all\ncases, across different attack settings, data modalities, and training\nconfigurations. We complement the systematic analysis of attack performance by\na comprehensive experimental study, that investigates the effectiveness of\nvarious attacks w.r.t. model type and training configurations, over three\ndiverse application scenarios (i.e., images, medical data, and location data).', ""Generative models often use human evaluations to measure the perceived\nquality of their outputs. Automated metrics are noisy indirect proxies, because\nthey rely on heuristics or pretrained embeddings. However, up until now, direct\nhuman evaluation strategies have been ad-hoc, neither standardized nor\nvalidated. Our work establishes a gold standard human benchmark for generative\nrealism. We construct Human eYe Perceptual Evaluation (HYPE) a human benchmark\nthat is (1) grounded in psychophysics research in perception, (2) reliable\nacross different sets of randomly sampled outputs from a model, (3) able to\nproduce separable model performances, and (4) efficient in cost and time. We\nintroduce two variants: one that measures visual perception under adaptive time\nconstraints to determine the threshold at which a model's outputs appear real\n(e.g. 250ms), and the other a less expensive variant that measures human error\nrate on fake and real images sans time constraints. We test HYPE across six\nstate-of-the-art generative adversarial networks and two sampling techniques on\nconditional and unconditional image generation using four datasets: CelebA,\nFFHQ, CIFAR-10, and ImageNet. We find that HYPE can track model improvements\nacross training epochs, and we confirm via bootstrap sampling that HYPE\nrankings are consistent and replicable."", 'Identifying the named entities mentioned in text would enrich many semantic\napplications at the downstream level. However, due to the predominant usage of\ncolloquial language in microblogs, the named entity recognition (NER) in\nChinese microblogs experience significant performance deterioration, compared\nwith performing NER in formal Chinese corpus. In this paper, we propose a\nsimple yet effective neural framework to derive the character-level embeddings\nfor NER in Chinese text, named ME-CNER. A character embedding is derived with\nrich semantic information harnessed at multiple granularities, ranging from\nradical, character to word levels. The experimental results demonstrate that\nthe proposed approach achieves a large performance improvement on Weibo dataset\nand comparable performance on MSRA news dataset with lower computational cost\nagainst the existing state-of-the-art alternatives.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently working on a machine learning project in which I aim to train a model to solve a zero-sum game, such as two-player tic-tac-toe. The trained model will be designed to compete against either the system or a human player, and it could also potentially play against another version of itself. I want to utilize gradient descent to train my model. Therefore, the initial task is to formulate this zero-sum game in a manner that makes it differentiable and thus suitable for the application of gradient descent. Moreover, I am seeking to circumvent the potential pitfalls of gradient descent, specifically situations where the algorithm becomes trapped in a suboptimal local minimum and fails to locate the global optimum. I welcome any techniques that could enhance the efficiency of my gradient descent algorithm within this context. Ideally, my gradient learning algorithm should possess robust theoretical properties, ensuring convergence under mild conditions.","['Recent applications that arise in machine learning have surged significant\ninterest in solving min-max saddle point games. This problem has been\nextensively studied in the convex-concave regime for which a global equilibrium\nsolution can be computed efficiently. In this paper, we study the problem in\nthe non-convex regime and show that an \\varepsilon--first order stationary\npoint of the game can be computed when one of the player\'s objective can be\noptimized to global optimality efficiently. In particular, we first consider\nthe case where the objective of one of the players satisfies the\nPolyak-{\\L}ojasiewicz (PL) condition. For such a game, we show that a simple\nmulti-step gradient descent-ascent algorithm finds an \\varepsilon--first order\nstationary point of the problem in \\widetilde{\\mathcal{O}}(\\varepsilon^{-2})\niterations. Then we show that our framework can also be applied to the case\nwhere the objective of the ""max-player"" is concave. In this case, we propose a\nmulti-step gradient descent-ascent algorithm that finds an \\varepsilon--first\norder stationary point of the game in \\widetilde{\\cal O}(\\varepsilon^{-3.5})\niterations, which is the best known rate in the literature. We applied our\nalgorithm to a fair classification problem of Fashion-MNIST dataset and\nobserved that the proposed algorithm results in smoother training and better\ngeneralization.', 'We formulate a general framework for competitive gradient-based learning that\nencompasses a wide breadth of multi-agent learning algorithms, and analyze the\nlimiting behavior of competitive gradient-based learning algorithms using\ndynamical systems theory. For both general-sum and potential games, we\ncharacterize a non-negligible subset of the local Nash equilibria that will be\navoided if each agent employs a gradient-based learning algorithm. We also shed\nlight on the issue of convergence to non-Nash strategies in general- and\nzero-sum games, which may have no relevance to the underlying game, and arise\nsolely due to the choice of algorithm. The existence and frequency of such\nstrategies may explain some of the difficulties encountered when using gradient\ndescent in zero-sum games as, e.g., in the training of generative adversarial\nnetworks. To reinforce the theoretical contributions, we provide empirical\nresults that highlight the frequency of linear quadratic dynamic games (a\nbenchmark for multi-agent reinforcement learning) that admit global Nash\nequilibria that are almost surely avoided by policy gradient.', 'Stochastic gradient descent (\\textsc{Sgd}) methods are the most powerful\noptimization tools in training machine learning and deep learning models.\nMoreover, acceleration (a.k.a. momentum) methods and diagonal scaling (a.k.a.\nadaptive gradient) methods are the two main techniques to improve the slow\nconvergence of \\textsc{Sgd}. While empirical studies have demonstrated\npotential advantages of combining these two techniques, it remains unknown\nwhether these methods can achieve the optimal rate of convergence for\nstochastic optimization. In this paper, we present a new class of adaptive and\naccelerated stochastic gradient descent methods and show that they exhibit the\noptimal sampling and iteration complexity for stochastic optimization. More\nspecifically, we show that diagonal scaling, initially designed to improve\nvanilla stochastic gradient, can be incorporated into accelerated stochastic\ngradient descent to achieve the optimal rate of convergence for smooth\nstochastic optimization. We also show that momentum, apart from being known to\nspeed up the convergence rate of deterministic optimization, also provides us\nnew ways of designing non-uniform and aggressive moving average schemes in\nstochastic optimization. Finally, we present some heuristics that help to\nimplement adaptive accelerated stochastic gradient descent methods and to\nfurther improve their practical performance for machine learning and deep\nlearning.', 'Game-theoretic models of learning are a powerful set of models that optimize\nmulti-objective architectures. Among these models are zero-sum architectures\nthat have inspired adversarial learning frameworks. An important shortcoming of\nthese zeros-sum architectures is that gradient-based training leads to weak\nconvergence and cyclic dynamics.\n  We propose a novel follow the leader training algorithm for zeros-sum\narchitectures that guarantees convergence to mixed Nash equilibrium without\ncyclic behaviors. It is a special type of follow the perturbed leader algorithm\nwhere perturbations are the result of a neural mediating agent.\n  We validate our theoretical results by applying this training algorithm to\ngames with convex and non-convex loss as well as generative adversarial\narchitectures. Moreover, we customize the implementation of this algorithm for\nadversarial imitation learning applications. At every step of the training, the\nmediator agent perturbs the observations with generated codes. As a result of\nthese mediating codes, the proposed algorithm is also efficient for learning in\nenvironments with various factors of variations. We validate our assertion by\nusing a procedurally generated game environment as well as synthetic data.\nGithub implementation is available.', 'We consider the maximum mean discrepancy ($\\mathrm{MMD}$) GAN problem and\npropose a parametric kernelized gradient flow that mimics the min-max game in\ngradient regularized $\\mathrm{MMD}$ GAN. We show that this flow provides a\ndescent direction minimizing the $\\mathrm{MMD}$ on a statistical manifold of\nprobability distributions. We then derive an explicit condition which ensures\nthat gradient descent on the parameter space of the generator in gradient\nregularized $\\mathrm{MMD}$ GAN is globally convergent to the target\ndistribution. Under this condition, we give non asymptotic convergence results\nof gradient descent in MMD GAN. Another contribution of this paper is the\nintroduction of a dynamic formulation of a regularization of $\\mathrm{MMD}$ and\ndemonstrating that the parametric kernelized descent for $\\mathrm{MMD}$ is the\ngradient flow of this functional with respect to the new Riemannian structure.\nOur obtained theoretical result allows ones to treat gradient flows for quite\ngeneral functionals and thus has potential applications to other types of\nvariational inferences on a statistical manifold beyond GANs. Finally,\nnumerical experiments suggest that our parametric kernelized gradient flow\nstabilizes GAN training and guarantees convergence.', 'We introduce a new algorithm for the numerical computation of Nash equilibria\nof competitive two-player games. Our method is a natural generalization of\ngradient descent to the two-player setting where the update is given by the\nNash equilibrium of a regularized bilinear local approximation of the\nunderlying game. It avoids oscillatory and divergent behaviors seen in\nalternating gradient descent. Using numerical experiments and rigorous\nanalysis, we provide a detailed comparison to methods based on \\emph{optimism}\nand \\emph{consensus} and show that our method avoids making any unnecessary\nchanges to the gradient dynamics while achieving exponential (local)\nconvergence for (locally) convex-concave zero sum games. Convergence and\nstability properties of our method are robust to strong interactions between\nthe players, without adapting the stepsize, which is not the case with previous\nmethods. In our numerical experiments on non-convex-concave problems, existing\nmethods are prone to divergence and instability due to their sensitivity to\ninteractions among the players, whereas we never observe divergence of our\nalgorithm. The ability to choose larger stepsizes furthermore allows our\nalgorithm to achieve faster convergence, as measured by the number of model\nevaluations.', 'Existing analyses of optimization in deep learning are either continuous,\nfocusing on (variants of) gradient flow, or discrete, directly treating\n(variants of) gradient descent. Gradient flow is amenable to theoretical\nanalysis, but is stylized and disregards computational efficiency. The extent\nto which it represents gradient descent is an open question in the theory of\ndeep learning. The current paper studies this question. Viewing gradient\ndescent as an approximate numerical solution to the initial value problem of\ngradient flow, we find that the degree of approximation depends on the\ncurvature around the gradient flow trajectory. We then show that over deep\nneural networks with homogeneous activations, gradient flow trajectories enjoy\nfavorable curvature, suggesting they are well approximated by gradient descent.\nThis finding allows us to translate an analysis of gradient flow over deep\nlinear neural networks into a guarantee that gradient descent efficiently\nconverges to global minimum almost surely under random initialization.\nExperiments suggest that over simple deep neural networks, gradient descent\nwith conventional step size is indeed close to gradient flow. We hypothesize\nthat the theory of gradient flows will unravel mysteries behind deep learning.', 'The cornerstone underpinning deep learning is the guarantee that gradient\ndescent on an objective converges to local minima. Unfortunately, this\nguarantee fails in settings, such as generative adversarial nets, where there\nare multiple interacting losses. The behavior of gradient-based methods in\ngames is not well understood -- and is becoming increasingly important as\nadversarial and multi-objective architectures proliferate. In this paper, we\ndevelop new techniques to understand and control the dynamics in general games.\nThe key result is to decompose the second-order dynamics into two components.\nThe first is related to potential games, which reduce to gradient descent on an\nimplicit function; the second relates to Hamiltonian games, a new class of\ngames that obey a conservation law, akin to conservation laws in classical\nmechanical systems. The decomposition motivates Symplectic Gradient Adjustment\n(SGA), a new algorithm for finding stable fixed points in general games. Basic\nexperiments show SGA is competitive with recently proposed algorithms for\nfinding stable fixed points in GANs -- whilst at the same time being applicable\nto -- and having guarantees in -- much more general games.', 'We study the role that a finite timescale separation parameter $\\tau$ has on\ngradient descent-ascent in two-player non-convex, non-concave zero-sum games\nwhere the learning rate of player 1 is denoted by $\\gamma_1$ and the learning\nrate of player 2 is defined to be $\\gamma_2=\\tau\\gamma_1$. Existing work\nanalyzing the role of timescale separation in gradient descent-ascent has\nprimarily focused on the edge cases of players sharing a learning rate ($\\tau\n=1$) and the maximizing player approximately converging between each update of\nthe minimizing player ($\\tau \\rightarrow \\infty$). For the parameter choice of\n$\\tau=1$, it is known that the learning dynamics are not guaranteed to converge\nto a game-theoretically meaningful equilibria in general. In contrast, Jin et\nal. (2020) showed that the stable critical points of gradient descent-ascent\ncoincide with the set of strict local minmax equilibria as\n$\\tau\\rightarrow\\infty$. In this work, we bridge the gap between past work by\nshowing there exists a finite timescale separation parameter $\\tau^{\\ast}$ such\nthat $x^{\\ast}$ is a stable critical point of gradient descent-ascent for all\n$\\tau \\in (\\tau^{\\ast}, \\infty)$ if and only if it is a strict local minmax\nequilibrium. Moreover, we provide an explicit construction for computing\n$\\tau^{\\ast}$ along with corresponding convergence rates and results under\ndeterministic and stochastic gradient feedback. The convergence results we\npresent are complemented by a non-convergence result: given a critical point\n$x^{\\ast}$ that is not a strict local minmax equilibrium, then there exists a\nfinite timescale separation $\\tau_0$ such that $x^{\\ast}$ is unstable for all\n$\\tau\\in (\\tau_0, \\infty)$. Finally, we empirically demonstrate on the CIFAR-10\nand CelebA datasets the significant impact timescale separation has on training\nperformance.', 'Gradient descent ascent (GDA), the simplest single-loop algorithm for\nnonconvex minimax optimization, is widely used in practical applications such\nas generative adversarial networks (GANs) and adversarial training. Albeit its\ndesirable simplicity, recent work shows inferior convergence rates of GDA in\ntheory even assuming strong concavity of the objective on one side. This paper\nestablishes new convergence results for two alternative single-loop algorithms\n-- alternating GDA and smoothed GDA -- under the mild assumption that the\nobjective satisfies the Polyak-Lojasiewicz (PL) condition about one variable.\nWe prove that, to find an $\\epsilon$-stationary point, (i) alternating GDA and\nits stochastic variant (without mini batch) respectively require $O(\\kappa^{2}\n\\epsilon^{-2})$ and $O(\\kappa^{4} \\epsilon^{-4})$ iterations, while (ii)\nsmoothed GDA and its stochastic variant (without mini batch) respectively\nrequire $O(\\kappa \\epsilon^{-2})$ and $O(\\kappa^{2} \\epsilon^{-4})$ iterations.\nThe latter greatly improves over the vanilla GDA and gives the hitherto best\nknown complexity results among single-loop algorithms under similar settings.\nWe further showcase the empirical efficiency of these algorithms in training\nGANs and robust nonlinear regression.', 'In recent years, artificial neural networks have developed into a powerful\ntool for dealing with a multitude of problems for which classical solution\napproaches reach their limits. However, it is still unclear why randomly\ninitialized gradient descent optimization algorithms, such as the well-known\nbatch gradient descent, are able to achieve zero training loss in many\nsituations even though the objective function is non-convex and non-smooth. One\nof the most promising approaches to solving this problem in the field of\nsupervised learning is the analysis of gradient descent optimization in the\nso-called overparameterized regime. In this article we provide a further\ncontribution to this area of research by considering overparameterized\nfully-connected rectified artificial neural networks with biases. Specifically,\nwe show that for a fixed number of training data the mean squared error using\nbatch gradient descent optimization applied to such a randomly initialized\nartificial neural network converges to zero at a linear convergence rate as\nlong as the width of the artificial neural network is large enough, the\nlearning rate is small enough, and the training input data are pairwise\nlinearly independent.', 'Fictitious play with reinforcement learning is a general and effective\nframework for zero-sum games. However, using the current deep neural network\nmodels, the implementation of fictitious play faces crucial challenges. Neural\nnetwork model training employs gradient descent approaches to update all\nconnection weights, and thus is easy to forget the old opponents after training\nto beat the new opponents. Existing approaches often maintain a pool of\nhistorical policy models to avoid the forgetting. However, learning to beat a\npool in stochastic games, i.e., a wide distribution over policy models, is\neither sample-consuming or insufficient to exploit all models with limited\namount of samples. In this paper, we propose a learning process with neural\nfictitious play to alleviate the above issues. We train a single model as our\npolicy model, which consists of sub-models and a selector. Everytime facing a\nnew opponent, the model is expanded by adding a new sub-model, where only the\nnew sub-model is updated instead of the whole model. At the same time, the\nselector is also updated to mix up the new sub-model with the previous ones at\nthe state-level, so that the model is maintained as a behavior strategy instead\nof a wide distribution over policy models. Experiments on Kuhn poker, a\ngrid-world Treasure Hunting game, and Mini-RTS environments show that the\nproposed approach alleviates the forgetting problem, and consequently improves\nthe learning efficiency and the robustness of neural fictitious play.', 'Stochastic convex optimization is a basic and well studied primitive in\nmachine learning. It is well known that convex and Lipschitz functions can be\nminimized efficiently using Stochastic Gradient Descent (SGD). The Normalized\nGradient Descent (NGD) algorithm, is an adaptation of Gradient Descent, which\nupdates according to the direction of the gradients, rather than the gradients\nthemselves. In this paper we analyze a stochastic version of NGD and prove its\nconvergence to a global minimum for a wider class of functions: we require the\nfunctions to be quasi-convex and locally-Lipschitz. Quasi-convexity broadens\nthe con- cept of unimodality to multidimensions and allows for certain types of\nsaddle points, which are a known hurdle for first-order optimization methods\nsuch as gradient descent. Locally-Lipschitz functions are only required to be\nLipschitz in a small region around the optimum. This assumption circumvents\ngradient explosion, which is another known hurdle for gradient descent\nvariants. Interestingly, unlike the vanilla SGD algorithm, the stochastic\nnormalized gradient descent algorithm provably requires a minimal minibatch\nsize.', ""Many economic games and machine learning approaches can be cast as\ncompetitive optimization problems where multiple agents are minimizing their\nrespective objective function, which depends on all agents' actions. While\ngradient descent is a reliable basic workhorse for single-agent optimization,\nit often leads to oscillation in competitive optimization. In this work we\npropose polymatrix competitive gradient descent (PCGD) as a method for solving\ngeneral sum competitive optimization involving arbitrary numbers of agents. The\nupdates of our method are obtained as the Nash equilibria of a local polymatrix\napproximation with a quadratic regularization, and can be computed efficiently\nby solving a linear system of equations. We prove local convergence of PCGD to\nstable fixed points for $n$-player general-sum games, and show that it does not\nrequire adapting the step size to the strength of the player-interactions. We\nuse PCGD to optimize policies in multi-agent reinforcement learning and\ndemonstrate its advantages in Snake, Markov soccer and an electricity market\ngame. Agents trained by PCGD outperform agents trained with simultaneous\ngradient descent, symplectic gradient adjustment, and extragradient in Snake\nand Markov soccer games and on the electricity market game, PCGD trains faster\nthan both simultaneous gradient descent and the extragradient method."", 'Over the last decade, a single algorithm has changed many facets of our lives\n- Stochastic Gradient Descent (SGD). In the era of ever decreasing loss\nfunctions, SGD and its various offspring have become the go-to optimization\ntool in machine learning and are a key component of the success of deep neural\nnetworks (DNNs). While SGD is guaranteed to converge to a local optimum (under\nloose assumptions), in some cases it may matter which local optimum is found,\nand this is often context-dependent. Examples frequently arise in machine\nlearning, from shape-versus-texture-features to ensemble methods and zero-shot\ncoordination. In these settings, there are desired solutions which SGD on\n\'standard\' loss functions will not find, since it instead converges to the\n\'easy\' solutions. In this paper, we present a different approach. Rather than\nfollowing the gradient, which corresponds to a locally greedy direction, we\ninstead follow the eigenvectors of the Hessian, which we call ""ridges"". By\niteratively following and branching amongst the ridges, we effectively span the\nloss surface to find qualitatively different solutions. We show both\ntheoretically and experimentally that our method, called Ridge Rider (RR),\noffers a promising direction for a variety of challenging problems.', 'This paper investigates the convergence of learning dynamics in Stackelberg\ngames. In the class of games we consider, there is a hierarchical game being\nplayed between a leader and a follower with continuous action spaces. We\nestablish a number of connections between the Nash and Stackelberg equilibrium\nconcepts and characterize conditions under which attracting critical points of\nsimultaneous gradient descent are Stackelberg equilibria in zero-sum games.\nMoreover, we show that the only stable critical points of the Stackelberg\ngradient dynamics are Stackelberg equilibria in zero-sum games. Using this\ninsight, we develop a gradient-based update for the leader while the follower\nemploys a best response strategy for which each stable critical point is\nguaranteed to be a Stackelberg equilibrium in zero-sum games. As a result, the\nlearning rule provably converges to a Stackelberg equilibria given an\ninitialization in the region of attraction of a stable critical point. We then\nconsider a follower employing a gradient-play update rule instead of a best\nresponse strategy and propose a two-timescale algorithm with similar asymptotic\nconvergence guarantees. For this algorithm, we also provide finite-time high\nprobability bounds for local convergence to a neighborhood of a stable\nStackelberg equilibrium in general-sum games. Finally, we present extensive\nnumerical results that validate our theory, provide insights into the\noptimization landscape of generative adversarial networks, and demonstrate that\nthe learning dynamics we propose can effectively train generative adversarial\nnetworks.', 'Reinforcement learning from self-play has recently reported many successes.\nSelf-play, where the agents compete with themselves, is often used to generate\ntraining data for iterative policy improvement. In previous work, heuristic\nrules are designed to choose an opponent for the current learner. Typical rules\ninclude choosing the latest agent, the best agent, or a random historical\nagent. However, these rules may be inefficient in practice and sometimes do not\nguarantee convergence even in the simplest matrix games. In this paper, we\npropose a new algorithmic framework for competitive self-play reinforcement\nlearning in two-player zero-sum games. We recognize the fact that the Nash\nequilibrium coincides with the saddle point of the stochastic payoff function,\nwhich motivates us to borrow ideas from classical saddle point optimization\nliterature. Our method trains several agents simultaneously, and intelligently\ntakes each other as opponent based on simple adversarial rules derived from a\nprincipled perturbation-based saddle optimization method. We prove\ntheoretically that our algorithm converges to an approximate equilibrium with\nhigh probability in convex-concave games under standard assumptions. Beyond the\ntheory, we further show the empirical superiority of our method over baseline\nmethods relying on the aforementioned opponent-selection heuristics in matrix\ngames, grid-world soccer, Gomoku, and simulated robot sumo, with neural net\npolicy function approximators.', 'Many modern learning tasks involve fitting nonlinear models to data which are\ntrained in an overparameterized regime where the parameters of the model exceed\nthe size of the training dataset. Due to this overparameterization, the\ntraining loss may have infinitely many global minima and it is critical to\nunderstand the properties of the solutions found by first-order optimization\nschemes such as (stochastic) gradient descent starting from different\ninitializations. In this paper we demonstrate that when the loss has certain\nproperties over a minimally small neighborhood of the initial point, first\norder methods such as (stochastic) gradient descent have a few intriguing\nproperties: (1) the iterates converge at a geometric rate to a global optima\neven when the loss is nonconvex, (2) among all global optima of the loss the\niterates converge to one with a near minimal distance to the initial point, (3)\nthe iterates take a near direct route from the initial point to this global\noptima. As part of our proof technique, we introduce a new potential function\nwhich captures the precise tradeoff between the loss function and the distance\nto the initial point as the iterations progress. For Stochastic Gradient\nDescent (SGD), we develop novel martingale techniques that guarantee SGD never\nleaves a small neighborhood of the initialization, even with rather large\nlearning rates. We demonstrate the utility of our general theory for a variety\nof problem domains spanning low-rank matrix recovery to neural network\ntraining. Underlying our analysis are novel insights that may have implications\nfor training and generalization of more sophisticated learning problems\nincluding those involving deep neural network architectures.']","['The development of machine learning is promoting the search for fast and\nstable minimization algorithms. To this end, we suggest a change in the current\ngradient descent methods that should speed up the motion in flat regions and\nslow it down in steep directions of the function to minimize. It is based on a\n""power gradient"", in which each component of the gradient is replaced by its\nversus-preserving $H$-th power, with $0<H<1$. We test three modern gradient\ndescent methods fed by such variant and by standard gradients, finding the new\nversion to achieve significantly better performances for the Nesterov\naccelerated gradient and AMSGrad. We also propose an effective new take on the\nADAM algorithm, which includes power gradients with varying $H$.', 'Deep learning models are often successfully trained using gradient descent,\ndespite the worst case hardness of the underlying non-convex optimization\nproblem. The key question is then under what conditions can one prove that\noptimization will succeed. Here we provide a strong result of this kind. We\nconsider a neural net with one hidden layer and a convolutional structure with\nno overlap and a ReLU activation function. For this architecture we show that\nlearning is NP-complete in the general case, but that when the input\ndistribution is Gaussian, gradient descent converges to the global optimum in\npolynomial time. To the best of our knowledge, this is the first global\noptimality guarantee of gradient descent on a convolutional neural network with\nReLU activations.', 'Owing to their connection with generative adversarial networks (GANs),\nsaddle-point problems have recently attracted considerable interest in machine\nlearning and beyond. By necessity, most theoretical guarantees revolve around\nconvex-concave (or even linear) problems; however, making theoretical inroads\ntowards efficient GAN training depends crucially on moving beyond this classic\nframework. To make piecemeal progress along these lines, we analyze the\nbehavior of mirror descent (MD) in a class of non-monotone problems whose\nsolutions coincide with those of a naturally associated variational inequality\n- a property which we call coherence. We first show that ordinary, ""vanilla"" MD\nconverges under a strict version of this condition, but not otherwise; in\nparticular, it may fail to converge even in bilinear models with a unique\nsolution. We then show that this deficiency is mitigated by optimism: by taking\nan ""extra-gradient"" step, optimistic mirror descent (OMD) converges in all\ncoherent problems. Our analysis generalizes and extends the results of\nDaskalakis et al. (2018) for optimistic gradient descent (OGD) in bilinear\nproblems, and makes concrete headway for establishing convergence beyond\nconvex-concave games. We also provide stochastic analogues of these results,\nand we validate our analysis by numerical experiments in a wide array of GAN\nmodels (including Gaussian mixture models, as well as the CelebA and CIFAR-10\ndatasets).', 'Games generalize the single-objective optimization paradigm by introducing\ndifferent objective functions for different players. Differentiable games often\nproceed by simultaneous or alternating gradient updates. In machine learning,\ngames are gaining new importance through formulations like generative\nadversarial networks (GANs) and actor-critic systems. However, compared to\nsingle-objective optimization, game dynamics are more complex and less\nunderstood. In this paper, we analyze gradient-based methods with momentum on\nsimple games. We prove that alternating updates are more stable than\nsimultaneous updates. Next, we show both theoretically and empirically that\nalternating gradient updates with a negative momentum term achieves convergence\nin a difficult toy adversarial problem, but also on the notoriously difficult\nto train saturating GANs.', 'Deep neural networks (DNN) are typically optimized using stochastic gradient\ndescent (SGD). However, the estimation of the gradient using stochastic samples\ntends to be noisy and unreliable, resulting in large gradient variance and bad\nconvergence. In this paper, we propose \\textbf{Filter Gradient Decent}~(FGD),\nan efficient stochastic optimization algorithm that makes the consistent\nestimation of the local gradient by solving an adaptive filtering problem with\ndifferent design of filters. Our method reduces variance in stochastic gradient\ndescent by incorporating the historical states to enhance the current\nestimation. It is able to correct noisy gradient direction as well as to\naccelerate the convergence of learning. We demonstrate the effectiveness of the\nproposed Filter Gradient Descent on numerical optimization and training neural\nnetworks, where it achieves superior and robust performance compared with\ntraditional momentum-based methods. To the best of our knowledge, we are the\nfirst to provide a practical solution that integrates filtering into gradient\nestimation by making the analogy between gradient estimation and filtering\nproblems in signal processing. (The code is provided in\nhttps://github.com/Adamdad/Filter-Gradient-Decent)', ""Loss functions with a large number of saddle points are one of the major\nobstacles for training modern machine learning models efficiently. First-order\nmethods such as gradient descent are usually the methods of choice for training\nmachine learning models. However, these methods converge to saddle points for\ncertain choices of initial guesses. In this paper, we propose a modification of\nthe recently proposed Laplacian smoothing gradient descent [Osher et al.,\narXiv:1806.06317], called modified Laplacian smoothing gradient descent\n(mLSGD), and demonstrate its potential to avoid saddle points without\nsacrificing the convergence rate. Our analysis is based on the attraction\nregion, formed by all starting points for which the considered numerical scheme\nconverges to a saddle point. We investigate the attraction region's dimension\nboth analytically and numerically. For a canonical class of quadratic\nfunctions, we show that the dimension of the attraction region for mLSGD is\nfloor((n-1)/2), and hence it is significantly smaller than that of the gradient\ndescent whose dimension is n-1."", 'Many machine learning and data science tasks require solving non-convex\noptimization problems. When the loss function is a sum of multiple terms, a\npopular method is the stochastic gradient descent. Viewed as a process for\nsampling the loss function landscape, the stochastic gradient descent is known\nto prefer flat minima. Though this is desired for certain optimization problems\nsuch as in deep learning, it causes issues when the goal is to find the global\nminimum, especially if the global minimum resides in a sharp valley.\n  Illustrated with a simple motivating example, we show that the fundamental\nreason is that the difference in the Lipschitz constants of multiple terms in\nthe loss function causes stochastic gradient descent to experience different\nvariances at different minima. In order to mitigate this effect and perform\nfaithful optimization, we propose a combined resampling-reweighting scheme to\nbalance the variance at local minima and extend to general loss functions. We\nexplain from the numerical stability perspective how the proposed scheme is\nmore likely to select the true global minimum, and the local convergence\nanalysis perspective how it converges to a minimum faster when compared with\nthe vanilla stochastic gradient descent. Experiments from robust statistics and\ncomputational chemistry are provided to demonstrate the theoretical findings.', 'We study the properties of alignment, a form of implicit regularization, in\nlinear neural networks under gradient descent. We define alignment for fully\nconnected networks with multidimensional outputs and show that it is a natural\nextension of alignment in networks with 1-dimensional outputs as defined by Ji\nand Telgarsky, 2018. While in fully connected networks, there always exists a\nglobal minimum corresponding to an aligned solution, we analyze alignment as it\nrelates to the training process. Namely, we characterize when alignment is an\ninvariant of training under gradient descent by providing necessary and\nsufficient conditions for this invariant to hold. In such settings, the\ndynamics of gradient descent simplify, thereby allowing us to provide an\nexplicit learning rate under which the network converges linearly to a global\nminimum. We then analyze networks with layer constraints such as convolutional\nnetworks. In this setting, we prove that gradient descent is equivalent to\nprojected gradient descent, and that alignment is impossible with sufficiently\nlarge datasets.', 'One of the most important parts of Artificial Neural Networks is minimizing\nthe loss functions which tells us how good or bad our model is. To minimize\nthese losses we need to tune the weights and biases. Also to calculate the\nminimum value of a function we need gradient. And to update our weights we need\ngradient descent. But there are some problems with regular gradient descent ie.\nit is quite slow and not that accurate. This article aims to give an\nintroduction to optimization strategies to gradient descent. In addition, we\nshall also discuss the architecture of these algorithms and further\noptimization of Neural Networks in general', 'We study the hardness of learning unitary transformations in $U(d)$ via\ngradient descent on time parameters of alternating operator sequences. We\nprovide numerical evidence that, despite the non-convex nature of the loss\nlandscape, gradient descent always converges to the target unitary when the\nsequence contains $d^2$ or more parameters. Rates of convergence indicate a\n""computational phase transition."" With less than $d^2$ parameters, gradient\ndescent converges to a sub-optimal solution, whereas with more than $d^2$\nparameters, gradient descent converges exponentially to an optimal solution.', ""This paper considers the multi-agent linear least-squares problem in a\nserver-agent network. In this problem, the system comprises multiple agents,\neach having a set of local data points, that are connected to a server. The\ngoal for the agents is to compute a linear mathematical model that optimally\nfits the collective data points held by all the agents, without sharing their\nindividual local data points. This goal can be achieved, in principle, using\nthe server-agent variant of the traditional iterative gradient-descent method.\nThe gradient-descent method converges linearly to a solution, and its rate of\nconvergence is lower bounded by the conditioning of the agents' collective data\npoints. If the data points are ill-conditioned, the gradient-descent method may\nrequire a large number of iterations to converge.\n  We propose an iterative pre-conditioning technique that mitigates the\ndeleterious effect of the conditioning of data points on the rate of\nconvergence of the gradient-descent method. We rigorously show that the\nresulting pre-conditioned gradient-descent method, with the proposed iterative\npre-conditioning, achieves superlinear convergence when the least-squares\nproblem has a unique solution. In general, the convergence is linear with\nimproved rate of convergence in comparison to the traditional gradient-descent\nmethod and the state-of-the-art accelerated gradient-descent methods. We\nfurther illustrate the improved rate of convergence of our proposed algorithm\nthrough experiments on different real-world least-squares problems in both\nnoise-free and noisy computation environment."", 'In this paper, we propose a geometric framework to analyze the convergence\nproperties of gradient descent trajectories in the context of linear neural\nnetworks. We translate a well-known empirical observation of linear neural nets\ninto a conjecture that we call the \\emph{overfitting conjecture} which states\nthat, for almost all training data and initial conditions, the trajectory of\nthe corresponding gradient descent system converges to a global minimum. This\nwould imply that the solution achieved by vanilla gradient descent algorithms\nis equivalent to that of the least-squares estimation, for linear neural\nnetworks of an arbitrary number of hidden layers. Built upon a key invariance\nproperty induced by the network structure, we first establish convergence of\ngradient descent trajectories to critical points of the square loss function in\nthe case of linear networks of arbitrary depth. Our second result is the proof\nof the \\emph{overfitting conjecture} in the case of single-hidden-layer linear\nnetworks with an argument based on the notion of normal hyperbolicity and under\na generic property on the training data (i.e., holding for almost all training\ndata).', 'In this paper we consider solving saddle point problems using two variants of\nGradient Descent-Ascent algorithms, Extra-gradient (EG) and Optimistic Gradient\nDescent Ascent (OGDA) methods. We show that both of these algorithms admit a\nunified analysis as approximations of the classical proximal point method for\nsolving saddle point problems. This viewpoint enables us to develop a new\nframework for analyzing EG and OGDA for bilinear and strongly convex-strongly\nconcave settings. Moreover, we use the proximal point approximation\ninterpretation to generalize the results for OGDA for a wide range of\nparameters.', 'We generalize gradient descent with momentum for optimization in\ndifferentiable games to have complex-valued momentum. We give theoretical\nmotivation for our method by proving convergence on bilinear zero-sum games for\nsimultaneous and alternating updates. Our method gives real-valued parameter\nupdates, making it a drop-in replacement for standard optimizers. We\nempirically demonstrate that complex-valued momentum can improve convergence in\nrealistic adversarial games - like generative adversarial networks - by showing\nwe can find better solutions with an almost identical computational cost. We\nalso show a practical generalization to a complex-valued Adam variant, which we\nuse to train BigGAN to better inception scores on CIFAR-10.', 'Several widely-used first-order saddle-point optimization methods yield an\nidentical continuous-time ordinary differential equation (ODE) that is\nidentical to that of Gradient Descent Ascent (GDA) method when derived naively.\nHowever, the convergence properties of these methods are qualitatively\ndifferent even on simple bilinear games. Thus the ODE perspective, which has\nproved powerful in analyzing single-objective optimization methods, has not\nplayed a similar role in saddle-point optimization.\n  We adopt a framework studied in fluid dynamics -- known as High-Resolution\nDifferential Equations (HRDEs) -- to design differential equation models for\nseveral saddle-point optimization methods. Critically, these HRDEs are distinct\nfor various saddle-point optimization methods. Moreover, on bilinear games, the\nconvergence properties of the HRDEs match the qualitative features of the\ncorresponding discrete methods. Additionally, we show that the HRDE of\nOptimistic Gradient Descent Ascent (OGDA) exhibits \\emph{last-iterate\nconvergence} for general monotone variational inequalities. Finally, we provide\nrates of convergence for the \\emph{best-iterate convergence} of the OGDA\nmethod, relying solely on the first-order smoothness of the monotone operator.', 'In optimization, the negative gradient of a function denotes the direction of\nsteepest descent. Furthermore, traveling in any direction orthogonal to the\ngradient maintains the value of the function. In this work, we show that these\northogonal directions that are ignored by gradient descent can be critical in\nequilibrium problems. Equilibrium problems have drawn heightened attention in\nmachine learning due to the emergence of the Generative Adversarial Network\n(GAN). We use the framework of Variational Inequalities to analyze popular\ntraining algorithms for a fundamental GAN variant: the Wasserstein\nLinear-Quadratic GAN. We show that the steepest descent direction causes\ndivergence from the equilibrium, and convergence to the equilibrium is achieved\nthrough following a particular orthogonal direction. We call this successful\ntechnique Crossing-the-Curl, named for its mathematical derivation as well as\nits intuition: identify the game\'s axis of rotation and move ""across"" space in\nthe direction towards smaller ""curling"".', 'We develop provably efficient reinforcement learning algorithms for\ntwo-player zero-sum finite-horizon Markov games with simultaneous moves. To\nincorporate function approximation, we consider a family of Markov games where\nthe reward function and transition kernel possess a linear structure. Both the\noffline and online settings of the problems are considered. In the offline\nsetting, we control both players and aim to find the Nash Equilibrium by\nminimizing the duality gap. In the online setting, we control a single player\nplaying against an arbitrary opponent and aim to minimize the regret. For both\nsettings, we propose an optimistic variant of the least-squares minimax value\niteration algorithm. We show that our algorithm is computationally efficient\nand provably achieves an $\\tilde O(\\sqrt{d^3 H^3 T} )$ upper bound on the\nduality gap and regret, where $d$ is the linear dimension, $H$ the horizon and\n$T$ the total number of timesteps. Our results do not require additional\nassumptions on the sampling model.\n  Our setting requires overcoming several new challenges that are absent in\nMarkov decision processes or turn-based Markov games. In particular, to achieve\noptimism with simultaneous moves, we construct both upper and lower confidence\nbounds of the value function, and then compute the optimistic policy by solving\na general-sum matrix game with these bounds as the payoff matrices. As finding\nthe Nash Equilibrium of a general-sum game is computationally hard, our\nalgorithm instead solves for a Coarse Correlated Equilibrium (CCE), which can\nbe obtained efficiently. To our best knowledge, such a CCE-based scheme for\noptimism has not appeared in the literature and might be of interest in its own\nright.', ""This paper studies a distributed multi-agent convex optimization problem. The\nsystem comprises multiple agents in this problem, each with a set of local data\npoints and an associated local cost function. The agents are connected to a\nserver, and there is no inter-agent communication. The agents' goal is to learn\na parameter vector that optimizes the aggregate of their local costs without\nrevealing their local data points. In principle, the agents can solve this\nproblem by collaborating with the server using the traditional distributed\ngradient-descent method. However, when the aggregate cost is ill-conditioned,\nthe gradient-descent method (i) requires a large number of iterations to\nconverge, and (ii) is highly unstable against process noise. We propose an\niterative pre-conditioning technique to mitigate the deleterious effects of the\ncost function's conditioning on the convergence rate of distributed\ngradient-descent. Unlike the conventional pre-conditioning techniques, the\npre-conditioner matrix in our proposed technique updates iteratively to\nfacilitate implementation on the distributed network. In the distributed\nsetting, we provably show that the proposed algorithm converges linearly with\nan improved rate of convergence than the traditional and adaptive\ngradient-descent methods. Additionally, for the special case when the minimizer\nof the aggregate cost is unique, our algorithm converges superlinearly. We\ndemonstrate our algorithm's superior performance compared to prominent\ndistributed algorithms for solving real logistic regression problems and\nemulating neural network training via a noisy quadratic model, thereby\nsignifying the proposed algorithm's efficiency for distributively solving\nnon-convex optimization. Moreover, we empirically show that the proposed\nalgorithm results in faster training without compromising the generalization\nperformance."", 'We provide the first convergence analysis of local gradient descent for\nminimizing the average of smooth and convex but otherwise arbitrary functions.\nProblems of this form and local gradient descent as a solution method are of\nimportance in federated learning, where each function is based on private data\nstored by a user on a mobile device, and the data of different users can be\narbitrarily heterogeneous. We show that in a low accuracy regime, the method\nhas the same communication complexity as gradient descent.', 'The goal of policy gradient approaches is to find a policy in a given class\nof policies which maximizes the expected return. Given a differentiable model\nof the policy, we want to apply a gradient-ascent technique to reach a local\noptimum. We mainly use gradient ascent, because it is theoretically well\nresearched. The main issue is that the policy gradient with respect to the\nexpected return is not available, thus we need to estimate it. As policy\ngradient algorithms also tend to require on-policy data for the gradient\nestimate, their biggest weakness is sample efficiency. For this reason, most\nresearch is focused on finding algorithms with improved sample efficiency. This\npaper provides a formal introduction to policy gradient that shows the\ndevelopment of policy gradient approaches, and should enable the reader to\nfollow current research on the topic.', ""The multinomial logistic regression (MLR) model is widely used in statistics\nand machine learning. Stochastic gradient descent (SGD) is the most common\napproach for determining the parameters of a MLR model in big data scenarios.\nHowever, SGD has slow sub-linear rates of convergence. A way to improve these\nrates of convergence is to use manifold optimization. Along this line,\nstochastic natural gradient descent (SNGD), proposed by Amari, was proven to be\nFisher efficient when it converged. However, SNGD is not guaranteed to converge\nand it is computationally too expensive for MLR models with a large number of\nparameters.\n  Here, we propose a stochastic optimization method for MLR based on manifold\noptimization concepts which (i) has per-iteration computational complexity is\nlinear in the number of parameters and (ii) can be proven to converge.\n  To achieve (i) we establish that the family of joint distributions for MLR is\na dually flat manifold and we use that to speed up calculations.\nS\\'anchez-L\\'opez and Cerquides have recently introduced convergent stochastic\nnatural gradient descent (CSNGD), a variant of SNGD whose convergence is\nguaranteed. To obtain (ii) our algorithm uses the fundamental idea from CSNGD,\nthus relying on an independent sequence to build a bounded approximation of the\nnatural gradient. We call the resulting algorithm dual stochastic natural\ngradient descent (DNSGD). By generalizing a result from Sunehag et al., we\nprove that DSNGD converges. Furthermore, we prove that the computational\ncomplexity of DSNGD iterations are linear on the number of variables of the\nmodel."", ""In this paper, we give explicit descriptions of versions of (Local-)\nBacktracking Gradient Descent and New Q-Newton's method to the Riemannian\nsetting.Here are some easy to state consequences of results in this paper,\nwhere X is a general Riemannian manifold of finite dimension and\n$f:X\\rightarrow \\mathbb{R}$ a $C^2$ function which is Morse (that is, all its\ncritical points are non-degenerate).\n  {\\bf Theorem.} For random choices of the hyperparameters in the Riemanian\nLocal Backtracking Gradient Descent algorithm and for random choices of the\ninitial point $x_0$, the sequence $\\{x_n\\}$ constructed by the algorithm either\n(i) converges to a local minimum of $f$ or (ii) eventually leaves every compact\nsubsets of $X$ (in other words, diverges to infinity on $X$). If $f$ has\ncompact sublevels, then only the former alternative happens. The convergence\nrate is the same as in the classical paper by Armijo.\n  {\\bf Theorem.} Assume that $f$ is $C^3$. For random choices of the\nhyperparametes in the Riemannian New Q-Newton's method, if the sequence\nconstructed by the algorithm converges, then the limit is a critical point of\n$f$. We have a local Stable-Center manifold theorem, near saddle points of $f$,\nfor the dynamical system associated to the algorithm. If the limit point is a\nnon-degenerate minimum point, then the rate of convergence is quadratic. If\nmoreover $X$ is an open subset of a Lie group and the initial point $x_0$ is\nchosen randomly, then we can globally avoid saddle points.\n  As an application, we propose a general method using Riemannian Backtracking\nGD to find minimum of a function on a bounded ball in a Euclidean space, and do\nexplicit calculations for calculating the smallest eigenvalue of a symmetric\nsquare matrix."", 'Large-scale optimization problems require algorithms both effective and\nefficient. One such popular and proven algorithm is Stochastic Gradient Descent\nwhich uses first-order gradient information to solve these problems. This paper\nstudies almost-sure convergence rates of the Stochastic Gradient Descent method\nwhen instead of deterministic, its learning rate becomes stochastic. In\nparticular, its learning rate is equipped with a multiplicative stochasticity,\nproducing a stochastic learning rate scheme. Theoretical results show\naccelerated almost-sure convergence rates of Stochastic Gradient Descent in a\nnonconvex setting when using an appropriate stochastic learning rate, compared\nto a deterministic-learning-rate scheme. The theoretical results are verified\nempirically.', 'We study the convergence properties of gradient descent for training deep\nlinear neural networks, i.e., deep matrix factorizations, by extending a\nprevious analysis for the related gradient flow. We show that under suitable\nconditions on the step sizes gradient descent converges to a critical point of\nthe loss function, i.e., the square loss in this article. Furthermore, we\ndemonstrate that for almost all initializations gradient descent converges to a\nglobal minimum in the case of two layers. In the case of three or more layers\nwe show that gradient descent converges to a global minimum on the manifold\nmatrices of some fixed rank, where the rank cannot be determined a priori.', 'Background: Recent developments have made it possible to accelerate neural\nnetworks training significantly using large batch sizes and data parallelism.\nTraining in an asynchronous fashion, where delay occurs, can make training even\nmore scalable. However, asynchronous training has its pitfalls, mainly a\ndegradation in generalization, even after convergence of the algorithm. This\ngap remains not well understood, as theoretical analysis so far mainly focused\non the convergence rate of asynchronous methods. Contributions: We examine\nasynchronous training from the perspective of dynamical stability. We find that\nthe degree of delay interacts with the learning rate, to change the set of\nminima accessible by an asynchronous stochastic gradient descent algorithm. We\nderive closed-form rules on how the learning rate could be changed, while\nkeeping the accessible set the same. Specifically, for high delay values, we\nfind that the learning rate should be kept inversely proportional to the delay.\nWe then extend this analysis to include momentum. We find momentum should be\neither turned off, or modified to improve training stability. We provide\nempirical experiments to validate our theoretical findings.', 'With the rise of deep neural networks, the challenge of explaining the\npredictions of these networks has become increasingly recognized. While many\nmethods for explaining the decisions of deep neural networks exist, there is\ncurrently no consensus on how to evaluate them. On the other hand, robustness\nis a popular topic for deep learning research; however, it is hardly talked\nabout in explainability until very recently. In this tutorial paper, we start\nby presenting gradient-based interpretability methods. These techniques use\ngradient signals to assign the burden of the decision on the input features.\nLater, we discuss how gradient-based methods can be evaluated for their\nrobustness and the role that adversarial robustness plays in having meaningful\nexplanations. We also discuss the limitations of gradient-based methods.\nFinally, we present the best practices and attributes that should be examined\nbefore choosing an explainability method. We conclude with the future\ndirections for research in the area at the convergence of robustness and\nexplainability.', 'A main puzzle of deep networks revolves around the absence of overfitting\ndespite large overparametrization and despite the large capacity demonstrated\nby zero training error on randomly labeled data. In this note, we show that the\ndynamics associated to gradient descent minimization of nonlinear networks is\ntopologically equivalent, near the asymptotically stable minima of the\nempirical error, to linear gradient system in a quadratic potential with a\ndegenerate (for square loss) or almost degenerate (for logistic or crossentropy\nloss) Hessian. The proposition depends on the qualitative theory of dynamical\nsystems and is supported by numerical results. Our main propositions extend to\ndeep nonlinear networks two properties of gradient descent for linear networks,\nthat have been recently established (1) to be key to their generalization\nproperties: 1. Gradient descent enforces a form of implicit regularization\ncontrolled by the number of iterations, and asymptotically converges to the\nminimum norm solution for appropriate initial conditions of gradient descent.\nThis implies that there is usually an optimum early stopping that avoids\noverfitting of the loss. This property, valid for the square loss and many\nother loss functions, is relevant especially for regression. 2. For\nclassification, the asymptotic convergence to the minimum norm solution implies\nconvergence to the maximum margin solution which guarantees good classification\nerror for ""low noise"" datasets. This property holds for loss functions such as\nthe logistic and cross-entropy loss independently of the initial conditions.\nThe robustness to overparametrization has suggestive implications for the\nrobustness of the architecture of deep convolutional networks with respect to\nthe curse of dimensionality.', 'Generative adversarial networks have been very successful in generative\nmodeling, however they remain relatively challenging to train compared to\nstandard deep neural networks. In this paper, we propose new visualization\ntechniques for the optimization landscapes of GANs that enable us to study the\ngame vector field resulting from the concatenation of the gradient of both\nplayers. Using these visualization techniques we try to bridge the gap between\ntheory and practice by showing empirically that the training of GANs exhibits\nsignificant rotations around Local Stable Stationary Points (LSSP), similar to\nthe one predicted by theory on toy examples. Moreover, we provide empirical\nevidence that GAN training converge to a stable stationary point which is a\nsaddle point for the generator loss, not a minimum, while still achieving\nexcellent performance.', 'Zero-sum stochastic games are easy to solve as they can be cast as simple\nMarkov decision processes. This is however not the case with general-sum\nstochastic games. A fairly general optimization problem formulation is\navailable for general-sum stochastic games by Filar and Vrieze [2004]. However,\nthe optimization problem there has a non-linear objective and non-linear\nconstraints with special structure. Since gradients of both the objective as\nwell as constraints of this optimization problem are well defined, gradient\nbased schemes seem to be a natural choice. We discuss a gradient scheme tuned\nfor two-player stochastic games. We show in simulations that this scheme indeed\nconverges to a Nash equilibrium, for a simple terrain exploration problem\nmodelled as a general-sum stochastic game. However, it turns out that only\nglobal minima of the optimization problem correspond to Nash equilibria of the\nunderlying general-sum stochastic game, while gradient schemes only guarantee\nconvergence to local minima. We then provide important necessary conditions for\ngradient schemes to converge to Nash equilibria in general-sum stochastic\ngames.', 'Gradient descent can be surprisingly good at optimizing deep neural networks\nwithout overfitting and without explicit regularization. We find that the\ndiscrete steps of gradient descent implicitly regularize models by penalizing\ngradient descent trajectories that have large loss gradients. We call this\nImplicit Gradient Regularization (IGR) and we use backward error analysis to\ncalculate the size of this regularization. We confirm empirically that implicit\ngradient regularization biases gradient descent toward flat minima, where test\nerrors are small and solutions are robust to noisy parameter perturbations.\nFurthermore, we demonstrate that the implicit gradient regularization term can\nbe used as an explicit regularizer, allowing us to control this gradient\nregularization directly. More broadly, our work indicates that backward error\nanalysis is a useful theoretical approach to the perennial question of how\nlearning rate, model size, and parameter regularization interact to determine\nthe properties of overparameterized models optimized with gradient descent.', 'There are much recent interests in solving noncovnex min-max optimization\nproblems due to its broad applications in many areas including machine\nlearning, networked resource allocations, and distributed optimization.\nPerhaps, the most popular first-order method in solving min-max optimization is\nthe so-called simultaneous (or single-loop) gradient descent-ascent algorithm\ndue to its simplicity in implementation. However, theoretical guarantees on the\nconvergence of this algorithm is very sparse since it can diverge even in a\nsimple bilinear problem.\n  In this paper, our focus is to characterize the finite-time performance (or\nconvergence rates) of the continuous-time variant of simultaneous gradient\ndescent-ascent algorithm. In particular, we derive the rates of convergence of\nthis method under a number of different conditions on the underlying objective\nfunction, namely, two-sided Polyak-L ojasiewicz (PL), one-sided PL,\nnonconvex-strongly concave, and strongly convex-nonconcave conditions. Our\nconvergence results improve the ones in prior works under the same conditions\nof objective functions. The key idea in our analysis is to use the classic\nsingular perturbation theory and coupling Lyapunov functions to address the\ntime-scale difference and interactions between the gradient descent and ascent\ndynamics. Our results on the behavior of continuous-time algorithm may be used\nto enhance the convergence properties of its discrete-time counterpart.', 'We consider the behavior of gradient flow and of discrete and noisy gradient\ndescent. It is commonly noted that the addition of noise to the process of\ndiscrete gradient descent can affect the trajectory of gradient descent. In\nprevious work, we observed such effects. There, we considered the case where\nthe minima had codimension 1. In this note, we do some computer experiments and\nobserve the behavior of noisy gradient descent in the more complex setting of\nminima of higher codimension.', 'Deep learning is built on the foundational guarantee that gradient descent on\nan objective function converges to local minima. Unfortunately, this guarantee\nfails in settings, such as generative adversarial nets, that exhibit multiple\ninteracting losses. The behavior of gradient-based methods in games is not well\nunderstood -- and is becoming increasingly important as adversarial and\nmulti-objective architectures proliferate. In this paper, we develop new tools\nto understand and control the dynamics in n-player differentiable games.\n  The key result is to decompose the game Jacobian into two components. The\nfirst, symmetric component, is related to potential games, which reduce to\ngradient descent on an implicit function. The second, antisymmetric component,\nrelates to Hamiltonian games, a new class of games that obey a conservation law\nakin to conservation laws in classical mechanical systems. The decomposition\nmotivates Symplectic Gradient Adjustment (SGA), a new algorithm for finding\nstable fixed points in differentiable games. Basic experiments show SGA is\ncompetitive with recently proposed algorithms for finding stable fixed points\nin GANs -- while at the same time being applicable to, and having guarantees\nin, much more general cases.', 'Stochastic gradient descent is an optimisation method that combines classical\ngradient descent with random subsampling within the target functional. In this\nwork, we introduce the stochastic gradient process as a continuous-time\nrepresentation of stochastic gradient descent. The stochastic gradient process\nis a dynamical system that is coupled with a continuous-time Markov process\nliving on a finite state space. The dynamical system -- a gradient flow --\nrepresents the gradient descent part, the process on the finite state space\nrepresents the random subsampling. Processes of this type are, for instance,\nused to model clonal populations in fluctuating environments. After introducing\nit, we study theoretical properties of the stochastic gradient process: We show\nthat it converges weakly to the gradient flow with respect to the full target\nfunction, as the learning rate approaches zero. We give conditions under which\nthe stochastic gradient process with constant learning rate is exponentially\nergodic in the Wasserstein sense. Then we study the case, where the learning\nrate goes to zero sufficiently slowly and the single target functions are\nstrongly convex. In this case, the process converges weakly to the point mass\nconcentrated in the global minimum of the full target function; indicating\nconsistency of the method. We conclude after a discussion of discretisation\nstrategies for the stochastic gradient process and numerical experiments.', 'Generative adversarial networks (GANs) represent a zero-sum game between two\nmachine players, a generator and a discriminator, designed to learn the\ndistribution of data. While GANs have achieved state-of-the-art performance in\nseveral benchmark learning tasks, GAN minimax optimization still poses great\ntheoretical and empirical challenges. GANs trained using first-order\noptimization methods commonly fail to converge to a stable solution where the\nplayers cannot improve their objective, i.e., the Nash equilibrium of the\nunderlying game. Such issues raise the question of the existence of Nash\nequilibrium solutions in the GAN zero-sum game. In this work, we show through\nseveral theoretical and numerical results that indeed GAN zero-sum games may\nnot have any local Nash equilibria. To characterize an equilibrium notion\napplicable to GANs, we consider the equilibrium of a new zero-sum game with an\nobjective function given by a proximal operator applied to the original\nobjective, a solution we call the proximal equilibrium. Unlike the Nash\nequilibrium, the proximal equilibrium captures the sequential nature of GANs,\nin which the generator moves first followed by the discriminator. We prove that\nthe optimal generative model in Wasserstein GAN problems provides a proximal\nequilibrium. Inspired by these results, we propose a new approach, which we\ncall proximal training, for solving GAN problems. We discuss several numerical\nexperiments demonstrating the existence of proximal equilibrium solutions in\nGAN minimax problems.', 'The theory of integral quadratic constraints (IQCs) allows the certification\nof exponential convergence of interconnected systems containing nonlinear or\nuncertain elements. In this work, we adapt the IQC theory to study first-order\nmethods for smooth and strongly-monotone games and show how to design tailored\nquadratic constraints to get tight upper bounds of convergence rates. Using\nthis framework, we recover the existing bound for the gradient method~(GD),\nderive sharper bounds for the proximal point method~(PPM) and optimistic\ngradient method~(OG), and provide \\emph{for the first time} a global\nconvergence rate for the negative momentum method~(NM) with an iteration\ncomplexity $\\mathcal{O}(\\kappa^{1.5})$, which matches its known lower bound. In\naddition, for time-varying systems, we prove that the gradient method with\noptimal step size achieves the fastest provable worst-case convergence rate\nwith quadratic Lyapunov functions. Finally, we further extend our analysis to\nstochastic games and study the impact of multiplicative noise on different\nalgorithms. We show that it is impossible for an algorithm with one step of\nmemory to achieve acceleration if it only queries the gradient once per batch\n(in contrast with the stochastic strongly-convex optimization setting, where\nsuch acceleration has been demonstrated). However, we exhibit an algorithm\nwhich achieves acceleration with two gradient queries per batch.', 'In this paper, we aim at providing an introduction to the gradient descent\nbased optimization algorithms for learning deep neural network models. Deep\nlearning models involving multiple nonlinear projection layers are very\nchallenging to train. Nowadays, most of the deep learning model training still\nrelies on the back propagation algorithm actually. In back propagation, the\nmodel variables will be updated iteratively until convergence with gradient\ndescent based optimization algorithms. Besides the conventional vanilla\ngradient descent algorithm, many gradient descent variants have also been\nproposed in recent years to improve the learning performance, including\nMomentum, Adagrad, Adam, Gadam, etc., which will all be introduced in this\npaper respectively.', 'The objective of a reinforcement learning agent is to behave so as to\nmaximise the sum of a suitable scalar function of state: the reward. These\nrewards are typically given and immutable. In this paper, we instead consider\nthe proposition that the reward function itself can be a good locus of learned\nknowledge. To investigate this, we propose a scalable meta-gradient framework\nfor learning useful intrinsic reward functions across multiple lifetimes of\nexperience. Through several proof-of-concept experiments, we show that it is\nfeasible to learn and capture knowledge about long-term exploration and\nexploitation into a reward function. Furthermore, we show that unlike policy\ntransfer methods that capture ""how"" the agent should behave, the learned reward\nfunctions can generalise to other kinds of agents and to changes in the\ndynamics of the environment by capturing ""what"" the agent should strive to do.', 'Physics-informed neural networks (PINNs) have lately received great attention\nthanks to their flexibility in tackling a wide range of forward and inverse\nproblems involving partial differential equations. However, despite their\nnoticeable empirical success, little is known about how such constrained neural\nnetworks behave during their training via gradient descent. More importantly,\neven less is known about why such models sometimes fail to train at all. In\nthis work, we aim to investigate these questions through the lens of the Neural\nTangent Kernel (NTK); a kernel that captures the behavior of fully-connected\nneural networks in the infinite width limit during training via gradient\ndescent. Specifically, we derive the NTK of PINNs and prove that, under\nappropriate conditions, it converges to a deterministic kernel that stays\nconstant during training in the infinite-width limit. This allows us to analyze\nthe training dynamics of PINNs through the lens of their limiting NTK and find\na remarkable discrepancy in the convergence rate of the different loss\ncomponents contributing to the total training error. To address this\nfundamental pathology, we propose a novel gradient descent algorithm that\nutilizes the eigenvalues of the NTK to adaptively calibrate the convergence\nrate of the total training error. Finally, we perform a series of numerical\nexperiments to verify the correctness of our theory and the practical\neffectiveness of the proposed algorithms. The data and code accompanying this\nmanuscript are publicly available at\n\\url{https://github.com/PredictiveIntelligenceLab/PINNsNTK}.', 'We study the problem of training deep neural networks with Rectified Linear\nUnit (ReLU) activation function using gradient descent and stochastic gradient\ndescent. In particular, we study the binary classification problem and show\nthat for a broad family of loss functions, with proper random weight\ninitialization, both gradient descent and stochastic gradient descent can find\nthe global minima of the training loss for an over-parameterized deep ReLU\nnetwork, under mild assumption on the training data. The key idea of our proof\nis that Gaussian random initialization followed by (stochastic) gradient\ndescent produces a sequence of iterates that stay inside a small perturbation\nregion centering around the initial weights, in which the empirical loss\nfunction of deep ReLU networks enjoys nice local curvature properties that\nensure the global convergence of (stochastic) gradient descent. Our theoretical\nresults shed light on understanding the optimization for deep learning, and\npave the way for studying the optimization dynamics of training modern deep\nneural networks.', 'In this paper, we analyze the optimization landscape of gradient descent\nmethods for static output feedback (SOF) control of discrete-time linear\ntime-invariant systems with quadratic cost. The SOF setting can be quite\ncommon, for example, when there are unmodeled hidden states in the underlying\nprocess. We first establish several important properties of the SOF cost\nfunction, including coercivity, L-smoothness, and M-Lipschitz continuous\nHessian. We then utilize these properties to show that the gradient descent is\nable to converge to a stationary point at a dimension-free rate. Furthermore,\nwe prove that under some mild conditions, gradient descent converges linearly\nto a local minimum if the starting point is close to one. These results not\nonly characterize the performance of gradient descent in optimizing the SOF\nproblem, but also shed light on the efficiency of general policy gradient\nmethods in reinforcement learning.', 'In this paper we investigate how gradient-based algorithms such as gradient\ndescent, (multi-pass) stochastic gradient descent, its persistent variant, and\nthe Langevin algorithm navigate non-convex loss-landscapes and which of them is\nable to reach the best generalization error at limited sample complexity. We\nconsider the loss landscape of the high-dimensional phase retrieval problem as\na prototypical highly non-convex example. We observe that for phase retrieval\nthe stochastic variants of gradient descent are able to reach perfect\ngeneralization for regions of control parameters where the gradient descent\nalgorithm is not. We apply dynamical mean-field theory from statistical physics\nto characterize analytically the full trajectories of these algorithms in their\ncontinuous-time limit, with a warm start, and for large system sizes. We\nfurther unveil several intriguing properties of the landscape and the\nalgorithms such as that the gradient descent can obtain better generalization\nproperties from less informed initializations.', ""Alternating gradient-descent-ascent (AltGDA) is an optimization algorithm\nthat has been widely used for model training in various machine learning\napplications, which aims to solve a nonconvex minimax optimization problem.\nHowever, the existing studies show that it suffers from a high computation\ncomplexity in nonconvex minimax optimization. In this paper, we develop a\nsingle-loop and fast AltGDA-type algorithm that leverages proximal gradient\nupdates and momentum acceleration to solve regularized nonconvex minimax\noptimization problems. By leveraging the momentum acceleration technique, we\nprove that the algorithm converges to a critical point in nonconvex minimax\noptimization and achieves a computation complexity in the order of\n$\\mathcal{O}(\\kappa^{\\frac{11}{6}}\\epsilon^{-2})$, where $\\epsilon$ is the\ndesired level of accuracy and $\\kappa$ is the problem's condition number. {Such\na computation complexity improves the state-of-the-art complexities of\nsingle-loop GDA and AltGDA algorithms (see the summary of comparison in\n\\Cref{table1})}. We demonstrate the effectiveness of our algorithm via an\nexperiment on adversarial deep learning."", 'The success of adversarial formulations in machine learning has brought\nrenewed motivation for smooth games. In this work, we focus on the class of\nstochastic Hamiltonian methods and provide the first convergence guarantees for\ncertain classes of stochastic smooth games. We propose a novel unbiased\nestimator for the stochastic Hamiltonian gradient descent (SHGD) and highlight\nits benefits. Using tools from the optimization literature we show that SHGD\nconverges linearly to the neighbourhood of a stationary point. To guarantee\nconvergence to the exact solution, we analyze SHGD with a decreasing step-size\nand we also present the first stochastic variance reduced Hamiltonian method.\nOur results provide the first global non-asymptotic last-iterate convergence\nguarantees for the class of stochastic unconstrained bilinear games and for the\nmore general class of stochastic games that satisfy a ""sufficiently bilinear""\ncondition, notably including some non-convex non-concave problems. We\nsupplement our analysis with experiments on stochastic bilinear and\nsufficiently bilinear games, where our theory is shown to be tight, and on\nsimple adversarial machine learning formulations.', 'Although recent work in AI has made great progress in solving large,\nzero-sum, extensive-form games, the underlying assumption in most past work is\nthat the parameters of the game itself are known to the agents. This paper\ndeals with the relatively under-explored but equally important ""inverse""\nsetting, where the parameters of the underlying game are not known to all\nagents, but must be learned through observations. We propose a differentiable,\nend-to-end learning framework for addressing this task. In particular, we\nconsider a regularized version of the game, equivalent to a particular form of\nquantal response equilibrium, and develop 1) a primal-dual Newton method for\nfinding such equilibrium points in both normal and extensive form games; and 2)\na backpropagation method that lets us analytically compute gradients of all\nrelevant game parameters through the solution itself. This ultimately lets us\nlearn the game by training in an end-to-end fashion, effectively by integrating\na ""differentiable game solver"" into the loop of larger deep network\narchitectures. We demonstrate the effectiveness of the learning method in\nseveral settings including poker and security game tasks.', 'With the recent advances in solving large, zero-sum extensive form games,\nthere is a growing interest in the inverse problem of inferring underlying game\nparameters given only access to agent actions. Although a recent work provides\na powerful differentiable end-to-end learning frameworks which embed a game\nsolver within a deep-learning framework, allowing unknown game parameters to be\nlearned via backpropagation, this framework faces significant limitations when\napplied to boundedly rational human agents and large scale problems, leading to\npoor practicality. In this paper, we address these limitations and propose a\nframework that is applicable for more practical settings. First, seeking to\nlearn the rationality of human agents in complex two-player zero-sum games, we\ndraw upon well-known ideas in decision theory to obtain a concise and\ninterpretable agent behavior model, and derive solvers and gradients for\nend-to-end learning. Second, to scale up to large, real-world scenarios, we\npropose an efficient first-order primal-dual method which exploits the\nstructure of extensive-form games, yielding significantly faster computation\nfor both game solving and gradient computation. When tested on randomly\ngenerated games, we report speedups of orders of magnitude over previous\napproaches. We also demonstrate the effectiveness of our model on both\nreal-world one-player settings and synthetic data.', 'The advancement of artificial intelligence has cast a new light on the\ndevelopment of optimization algorithm. This paper proposes to learn a two-phase\n(including a minimization phase and an escaping phase) global optimization\nalgorithm for smooth non-convex functions. For the minimization phase, a\nmodel-driven deep learning method is developed to learn the update rule of\ndescent direction, which is formalized as a nonlinear combination of historical\ninformation, for convex functions. We prove that the resultant algorithm with\nthe proposed adaptive direction guarantees convergence for convex functions.\nEmpirical study shows that the learned algorithm significantly outperforms some\nwell-known classical optimization algorithms, such as gradient descent,\nconjugate descent and BFGS, and performs well on ill-posed functions. The\nescaping phase from local optimum is modeled as a Markov decision process with\na fixed escaping policy. We further propose to learn an optimal escaping policy\nby reinforcement learning. The effectiveness of the escaping policies is\nverified by optimizing synthesized functions and training a deep neural network\nfor CIFAR image classification. The learned two-phase global optimization\nalgorithm demonstrates a promising global search capability on some benchmark\nfunctions and machine learning tasks.', 'In this work, we propose an optimization algorithm which we call norm-adapted\ngradient descent. This algorithm is similar to other gradient-based\noptimization algorithms like Adam or Adagrad in that it adapts the learning\nrate of stochastic gradient descent at each iteration. However, rather than\nusing statistical properties of observed gradients, norm-adapted gradient\ndescent relies on a first-order estimate of the effect of a standard gradient\ndescent update step, much like the Newton-Raphson method in many dimensions.\nOur algorithm can also be compared to quasi-Newton methods, but we seek roots\nrather than stationary points. Seeking roots can be justified by the fact that\nfor models with sufficient capacity measured by nonnegative loss functions,\nroots coincide with global optima. This work presents several experiments where\nwe have used our algorithm; in these results, it appears norm-adapted descent\nis particularly strong in regression settings but is also capable of training\nclassifiers.', 'Understanding the implicit bias of training algorithms is of crucial\nimportance in order to explain the success of overparametrised neural networks.\nIn this paper, we study the dynamics of stochastic gradient descent over\ndiagonal linear networks through its continuous time version, namely stochastic\ngradient flow. We explicitly characterise the solution chosen by the stochastic\nflow and prove that it always enjoys better generalisation properties than that\nof gradient flow. Quite surprisingly, we show that the convergence speed of the\ntraining loss controls the magnitude of the biasing effect: the slower the\nconvergence, the better the bias. To fully complete our analysis, we provide\nconvergence guarantees for the dynamics. We also give experimental results\nwhich support our theoretical claims. Our findings highlight the fact that\nstructured noise can induce better generalisation and they help explain the\ngreater performances observed in practice of stochastic gradient descent over\ngradient descent.', 'Natural gradient descent is an optimization method traditionally motivated\nfrom the perspective of information geometry, and works well for many\napplications as an alternative to stochastic gradient descent. In this paper we\ncritically analyze this method and its properties, and show how it can be\nviewed as a type of 2nd-order optimization method, with the Fisher information\nmatrix acting as a substitute for the Hessian. In many important cases, the\nFisher information matrix is shown to be equivalent to the Generalized\nGauss-Newton matrix, which both approximates the Hessian, but also has certain\nproperties that favor its use over the Hessian. This perspective turns out to\nhave significant implications for the design of a practical and robust natural\ngradient optimizer, as it motivates the use of techniques like trust regions\nand Tikhonov regularization. Additionally, we make a series of contributions to\nthe understanding of natural gradient and 2nd-order methods, including: a\nthorough analysis of the convergence speed of stochastic natural gradient\ndescent (and more general stochastic 2nd-order methods) as applied to convex\nquadratics, a critical examination of the oft-used ""empirical"" approximation of\nthe Fisher matrix, and an analysis of the (approximate) parameterization\ninvariance property possessed by natural gradient methods (which we show also\nholds for certain other curvature, but notably not the Hessian).', 'We consider the problem of two-player zero-sum games. This problem is\nformulated as a min-max Markov game in the literature. The solution of this\ngame, which is the min-max payoff, starting from a given state is called the\nmin-max value of the state. In this work, we compute the solution of the\ntwo-player zero-sum game utilizing the technique of successive relaxation that\nhas been successfully applied in the literature to compute a faster value\niteration algorithm in the context of Markov Decision Processes. We extend the\nconcept of successive relaxation to the setting of two-player zero-sum games.\nWe show that, under a special structure on the game, this technique facilitates\nfaster computation of the min-max value of the states. We then derive a\ngeneralized minimax Q-learning algorithm that computes the optimal policy when\nthe model information is not known. Finally, we prove the convergence of the\nproposed generalized minimax Q-learning algorithm utilizing stochastic\napproximation techniques, under an assumption on the boundedness of iterates.\nThrough experiments, we demonstrate the effectiveness of our proposed\nalgorithm.', 'Minimax optimization has found extensive applications in modern machine\nlearning, in settings such as generative adversarial networks (GANs),\nadversarial training and multi-agent reinforcement learning. As most of these\napplications involve continuous nonconvex-nonconcave formulations, a very basic\nquestion arises---""what is a proper definition of local optima?""\n  Most previous work answers this question using classical notions of\nequilibria from simultaneous games, where the min-player and the max-player act\nsimultaneously. In contrast, most applications in machine learning, including\nGANs and adversarial training, correspond to sequential games, where the order\nof which player acts first is crucial (since minimax is in general not equal to\nmaximin due to the nonconvex-nonconcave nature of the problems). The main\ncontribution of this paper is to propose a proper mathematical definition of\nlocal optimality for this sequential setting---local minimax, as well as to\npresent its properties and existence results. Finally, we establish a strong\nconnection to a basic local search algorithm---gradient descent ascent (GDA):\nunder mild conditions, all stable limit points of GDA are exactly local minimax\npoints up to some degenerate points.', ""Gradient descent algorithms have been used in countless applications since\nthe inception of Newton's method. The explosion in the number of applications\nof neural networks has re-energized efforts in recent years to improve the\nstandard gradient descent method in both efficiency and accuracy. These methods\nmodify the effect of the gradient in updating the values of the parameters.\nThese modifications often incorporate hyperparameters: additional variables\nwhose values must be specified at the outset of the program. We provide, below,\na novel gradient descent algorithm, called Gravilon, that uses the geometry of\nthe hypersurface to modify the length of the step in the direction of the\ngradient. Using neural networks, we provide promising experimental results\ncomparing the accuracy and efficiency of the Gravilon method against commonly\nused gradient descent algorithms on MNIST digit classification."", 'Machine learning has become successful in solving wireless interference\nmanagement problems. Different kinds of deep neural networks (DNNs) have been\ntrained to accomplish key tasks such as power control, beamforming and\nadmission control. There are two popular training paradigms for such DNNs-based\ninterference management models: supervised learning (i.e., fitting labels\ngenerated by an optimization algorithm) and unsupervised learning (i.e.,\ndirectly optimizing some system performance measure). Although both of these\nparadigms have been extensively applied in practice, due to the lack of any\ntheoretical understanding about these methods, it is not clear how to\nsystematically understand and compare their performance.\n  In this work, we conduct theoretical studies to provide some in-depth\nunderstanding about these two training paradigms. First, we show a somewhat\nsurprising result, that for some special power control problem, the\nunsupervised learning can perform much worse than its supervised counterpart,\nbecause it is more likely to stuck at some low-quality local solutions. We then\nprovide a series of theoretical results to further understand the properties of\nthe two approaches. Generally speaking, we show that when high-quality labels\nare available, then the supervised learning is less likely to be stuck at a\nsolution than its unsupervised counterpart. Additionally, we develop a\nsemi-supervised learning approach which properly integrates these two training\nparadigms, and can effectively utilize limited number of labels to find\nhigh-quality solutions. To our knowledge, these are the first set of\ntheoretical results trying to understand different training approaches in\nlearning-based wireless communication system design.', 'Batch Normalization (BN) has become a cornerstone of deep learning across\ndiverse architectures, appearing to help optimization as well as\ngeneralization. While the idea makes intuitive sense, theoretical analysis of\nits effectiveness has been lacking. Here theoretical support is provided for\none of its conjectured properties, namely, the ability to allow gradient\ndescent to succeed with less tuning of learning rates. It is shown that even if\nwe fix the learning rate of scale-invariant parameters (e.g., weights of each\nlayer with BN) to a constant (say, $0.3$), gradient descent still approaches a\nstationary point (i.e., a solution where gradient is zero) in the rate of\n$T^{-1/2}$ in $T$ iterations, asymptotically matching the best bound for\ngradient descent with well-tuned learning rates. A similar result with\nconvergence rate $T^{-1/4}$ is also shown for stochastic gradient descent.', ""We study the iteration complexity of the optimistic gradient descent-ascent\n(OGDA) method and the extra-gradient (EG) method for finding a saddle point of\na convex-concave unconstrained min-max problem. To do so, we first show that\nboth OGDA and EG can be interpreted as approximate variants of the proximal\npoint method. This is similar to the approach taken in [Nemirovski, 2004] which\nanalyzes EG as an approximation of the `conceptual mirror prox'. In this paper,\nwe highlight how gradients used in OGDA and EG try to approximate the gradient\nof the Proximal Point method. We then exploit this interpretation to show that\nboth algorithms produce iterates that remain within a bounded set. We further\nshow that the primal dual gap of the averaged iterates generated by both of\nthese algorithms converge with a rate of $\\mathcal{O}(1/k)$. Our theoretical\nanalysis is of interest as it provides a the first convergence rate estimate\nfor OGDA in the general convex-concave setting. Moreover, it provides a simple\nconvergence analysis for the EG algorithm in terms of function value without\nusing compactness assumption."", 'We design and analyze minimax-optimal algorithms for online linear\noptimization games where the player\'s choice is unconstrained. The player\nstrives to minimize regret, the difference between his loss and the loss of a\npost-hoc benchmark strategy. The standard benchmark is the loss of the best\nstrategy chosen from a bounded comparator set. When the the comparison set and\nthe adversary\'s gradients satisfy L_infinity bounds, we give the value of the\ngame in closed form and prove it approaches sqrt(2T/pi) as T -> infinity.\n  Interesting algorithms result when we consider soft constraints on the\ncomparator, rather than restricting it to a bounded set. As a warmup, we\nanalyze the game with a quadratic penalty. The value of this game is exactly\nT/2, and this value is achieved by perhaps the simplest online algorithm of\nall: unprojected gradient descent with a constant learning rate.\n  We then derive a minimax-optimal algorithm for a much softer penalty\nfunction. This algorithm achieves good bounds under the standard notion of\nregret for any comparator point, without needing to specify the comparator set\nin advance. The value of this game converges to sqrt{e} as T ->infinity; we\ngive a closed-form for the exact value as a function of T. The resulting\nalgorithm is natural in unconstrained investment or betting scenarios, since it\nguarantees at worst constant loss, while allowing for exponential reward\nagainst an ""easy"" adversary.', 'Bayesian optimization has demonstrated impressive success in finding the\noptimum input x* and output f* = f(x*) = max f(x) of a black-box function f. In\nsome applications, however, the optimum output f* is known in advance and the\ngoal is to find the corresponding optimum input x*. In this paper, we consider\na new setting in BO in which the knowledge of the optimum output f* is\navailable. Our goal is to exploit the knowledge about f* to search for the\ninput x* efficiently. To achieve this goal, we first transform the Gaussian\nprocess surrogate using the information about the optimum output. Then, we\npropose two acquisition functions, called confidence bound minimization and\nexpected regret minimization. We show that our approaches work intuitively and\ngive quantitatively better performance against standard BO methods. We\ndemonstrate real applications in tuning a deep reinforcement learning algorithm\non the CartPole problem and XGBoost on Skin Segmentation dataset in which the\noptimum values are publicly available.', 'The plain stochastic gradient descent and momentum stochastic gradient\ndescent have extremely wide applications in deep learning due to their simple\nsettings and low computational complexity. The momentum stochastic gradient\ndescent uses the accumulated gradient as the updated direction of the current\nparameters, which has a faster training speed. Because the direction of the\nplain stochastic gradient descent has not been corrected by the accumulated\ngradient. For the parameters that currently need to be updated, it is the\noptimal direction, and its update is more accurate. We combine the advantages\nof the momentum stochastic gradient descent with fast training speed and the\nplain stochastic gradient descent with high accuracy, and propose a scaling\ntransition from momentum stochastic gradient descent to plain stochastic\ngradient descent(TSGD) method. At the same time, a learning rate that decreases\nlinearly with the iterations is used instead of a constant learning rate. The\nTSGD algorithm has a larger step size in the early stage to speed up the\ntraining, and training with a smaller step size in the later stage can steadily\nconverge. Our experimental results show that the TSGD algorithm has faster\ntraining speed, higher accuracy and better stability. Our implementation is\navailable at: https://github.com/kunzeng/TSGD.', 'In this paper, we study the implicit regularization of the gradient descent\nalgorithm in homogeneous neural networks, including fully-connected and\nconvolutional neural networks with ReLU or LeakyReLU activations. In\nparticular, we study the gradient descent or gradient flow (i.e., gradient\ndescent with infinitesimal step size) optimizing the logistic loss or\ncross-entropy loss of any homogeneous model (possibly non-smooth), and show\nthat if the training loss decreases below a certain threshold, then we can\ndefine a smoothed version of the normalized margin which increases over time.\nWe also formulate a natural constrained optimization problem related to margin\nmaximization, and prove that both the normalized margin and its smoothed\nversion converge to the objective value at a KKT point of the optimization\nproblem. Our results generalize the previous results for logistic regression\nwith one-layer or multi-layer linear networks, and provide more quantitative\nconvergence results with weaker assumptions than previous results for\nhomogeneous smooth neural networks. We conduct several experiments to justify\nour theoretical finding on MNIST and CIFAR-10 datasets. Finally, as margin is\nclosely related to robustness, we discuss potential benefits of training longer\nfor improving the robustness of the model.', ""The goal of multi-task learning is to enable more efficient learning than\nsingle task learning by sharing model structures for a diverse set of tasks. A\nstandard multi-task learning objective is to minimize the average loss across\nall tasks. While straightforward, using this objective often results in much\nworse final performance for each task than learning them independently. A major\nchallenge in optimizing a multi-task model is the conflicting gradients, where\ngradients of different task objectives are not well aligned so that following\nthe average gradient direction can be detrimental to specific tasks'\nperformance. Previous work has proposed several heuristics to manipulate the\ntask gradients for mitigating this problem. But most of them lack convergence\nguarantee and/or could converge to any Pareto-stationary point. In this paper,\nwe introduce Conflict-Averse Gradient descent (CAGrad) which minimizes the\naverage loss function, while leveraging the worst local improvement of\nindividual tasks to regularize the algorithm trajectory. CAGrad balances the\nobjectives automatically and still provably converges to a minimum over the\naverage loss. It includes the regular gradient descent (GD) and the multiple\ngradient descent algorithm (MGDA) in the multi-objective optimization (MOO)\nliterature as special cases. On a series of challenging multi-task supervised\nlearning and reinforcement learning tasks, CAGrad achieves improved performance\nover prior state-of-the-art multi-objective gradient manipulation methods."", 'Given that in practice training data is scarce for all but a small set of\nproblems, a core question is how to incorporate prior knowledge into a model.\nIn this paper, we consider the case of prior procedural knowledge for neural\nnetworks, such as knowing how a program should traverse a sequence, but not\nwhat local actions should be performed at each step. To this end, we present an\nend-to-end differentiable interpreter for the programming language Forth which\nenables programmers to write program sketches with slots that can be filled\nwith behaviour trained from program input-output data. We can optimise this\nbehaviour directly through gradient descent techniques on user-specified\nobjectives, and also integrate the program into any larger neural computation\ngraph. We show empirically that our interpreter is able to effectively leverage\ndifferent levels of prior program structure and learn complex behaviours such\nas sequence sorting and addition. When connected to outputs of an LSTM and\ntrained jointly, our interpreter achieves state-of-the-art accuracy for\nend-to-end reasoning about quantities expressed in natural language stories.', 'We analyze stochastic gradient descent for optimizing non-convex functions.\nIn many cases for non-convex functions the goal is to find a reasonable local\nminimum, and the main concern is that gradient updates are trapped in saddle\npoints. In this paper we identify strict saddle property for non-convex problem\nthat allows for efficient optimization. Using this property we show that\nstochastic gradient descent converges to a local minimum in a polynomial number\nof iterations. To the best of our knowledge this is the first work that gives\nglobal convergence guarantees for stochastic gradient descent on non-convex\nfunctions with exponentially many local minima and saddle points. Our analysis\ncan be applied to orthogonal tensor decomposition, which is widely used in\nlearning a rich class of latent variable models. We propose a new optimization\nformulation for the tensor decomposition problem that has strict saddle\nproperty. As a result we get the first online algorithm for orthogonal tensor\ndecomposition with global convergence guarantee.', 'Applications of machine learning inform human decision makers in a broad\nrange of tasks. The resulting problem is usually formulated in terms of a\nsingle decision maker. We argue that it should rather be described as a\ntwo-player learning problem where one player is the machine and the other the\nhuman. While both players try to optimize the final decision, the setup is\noften characterized by (1) the presence of private information and (2) opacity,\nthat is imperfect understanding between the decision makers. We prove that both\nproperties can complicate decision making considerably. A lower bound\nquantifies the worst-case hardness of optimally advising a decision maker who\nis opaque or has access to private information. An upper bound shows that a\nsimple coordination strategy is nearly minimax optimal. More efficient learning\nis possible under certain assumptions on the problem, for example that both\nplayers learn to take actions independently. Such assumptions are implicit in\nexisting literature, for example in medical applications of machine learning,\nbut have not been described or justified theoretically.', 'With the vigorous development of artificial intelligence technology, various\nengineering technology applications have been implemented one after another.\nThe gradient descent method plays an important role in solving various\noptimization problems, due to its simple structure, good stability and easy\nimplementation. In multi-node machine learning system, the gradients usually\nneed to be shared. Shared gradients are generally unsafe. Attackers can obtain\ntraining data simply by knowing the gradient information. In this paper, to\nprevent gradient leakage while keeping the accuracy of model, we propose the\nsuper stochastic gradient descent approach to update parameters by concealing\nthe modulus length of gradient vectors and converting it or them into a unit\nvector. Furthermore, we analyze the security of super stochastic gradient\ndescent approach. Our algorithm can defend against attacks on the gradient.\nExperiment results show that our approach is obviously superior to prevalent\ngradient descent approaches in terms of accuracy, robustness, and adaptability\nto large-scale batches.', 'Gradient descent optimization algorithms are the standard ingredients that\nare used to train artificial neural networks (ANNs). Even though a huge number\nof numerical simulations indicate that gradient descent optimization methods do\nindeed convergence in the training of ANNs, until today there is no rigorous\ntheoretical analysis which proves (or disproves) this conjecture. In\nparticular, even in the case of the most basic variant of gradient descent\noptimization algorithms, the plain vanilla gradient descent method, it remains\nan open problem to prove or disprove the conjecture that gradient descent\nconverges in the training of ANNs. In this article we solve this problem in the\nspecial situation where the target function under consideration is a constant\nfunction. More specifically, in the case of constant target functions we prove\nin the training of rectified fully-connected feedforward ANNs with one-hidden\nlayer that the risk function of the gradient descent method does indeed\nconverge to zero. Our mathematical analysis strongly exploits the property that\nthe rectifier function is the activation function used in the considered ANNs.\nA key contribution of this work is to explicitly specify a Lyapunov function\nfor the gradient flow system of the ANN parameters. This Lyapunov function is\nthe central tool in our convergence proof of the gradient descent method.', 'The rise in machine learning-assisted decision-making has led to concerns\nabout the fairness of the decisions and techniques to mitigate problems of\ndiscrimination. If a negative decision is made about an individual (denying a\nloan, rejecting an application for housing, and so on) justice dictates that we\nbe able to ask how we might change circumstances to get a favorable decision\nthe next time. Moreover, the ability to change circumstances (a better\neducation, improved credentials) should not be limited to only those with\naccess to expensive resources. In other words, \\emph{recourse} for negative\ndecisions should be considered a desirable value that can be equalized across\n(demographically defined) groups. This paper describes how to build models that\nmake accurate predictions while still ensuring that the penalties for a\nnegative outcome do not disadvantage different groups disproportionately. We\nmeasure recourse as the distance of an individual from the decision boundary of\na classifier. We then introduce a regularized objective to minimize the\ndifference in recourse across groups. We explore linear settings and further\nextend recourse to non-linear settings as well as model-agnostic settings where\nthe exact distance from boundary cannot be calculated. Our results show that we\ncan successfully decrease the unfairness in recourse while maintaining\nclassifier performance.', 'We review the theory of, and develop algorithms for transforming a finite\npoint set in ${\\bf R}^d$ into a set in \\emph{radial isotropic position} by a\nnonsingular linear transformation followed by rescaling each image point to the\nunit sphere. This problem arises in a wide spectrum of applications in computer\nscience and mathematics.\n  Our algorithms use gradient descent methods for a particular convex function\n$f$ whose minimum defines the transformation, and our main focus is on\nanalyzing their performance. Although the minimum can be computed exactly, by\nexpensive symbolic algebra techniques, gradient descent only approximates the\ndesired minimum to any desired level of accuracy. We show that computing the\ngradient of $f$ amounts to computing the Singular Value Decomposition (SVD) of\na certain matrix associated with the input set, making it simple to implement.\nWe believe it to be superior to other approximate techniques (mainly the\nellipsoid algorithm) used previously to find this transformation, and it should\nrun much faster in practice.\n  We prove that $f$ is smooth, which yields convergence rate proportional to\n$1/\\epsilon$, where $\\epsilon$ is the desired approximation accuracy. To\ncomplete the analysis, we provide upper bounds on the norm of the optimal\nsolution which depend on new parameters measuring ""the degeneracy"" in our\ninput. We believe that our parameters capture degeneracy better than other,\nseemingly weaker, parameters used in previous works.\n  We next analyze the strong convexity of $f$, and present two worst-case lower\nbounds on the smallest eigenvalue of its Hessian. This gives another worst-case\nbound on the convergence rate of another variant of gradient decent that\ndepends only logarithmically on $1/\\epsilon$.', 'Many binary classification problems minimize misclassification above (or\nbelow) a threshold. We show that instances of ranking problems, accuracy at the\ntop or hypothesis testing may be written in this form. We propose a general\nframework to handle these classes of problems and show which known methods\n(both known and newly proposed) fall into this framework. We provide a\ntheoretical analysis of this framework and mention selected possible pitfalls\nthe methods may encounter. We suggest several numerical improvements including\nthe implicit derivative and stochastic gradient descent. We provide an\nextensive numerical study. Based both on the theoretical properties and\nnumerical experiments, we conclude the paper by suggesting which method should\nbe used in which situation.', 'Learning to learn is a powerful paradigm for enabling models to learn from\ndata more effectively and efficiently. A popular approach to meta-learning is\nto train a recurrent model to read in a training dataset as input and output\nthe parameters of a learned model, or output predictions for new test inputs.\nAlternatively, a more recent approach to meta-learning aims to acquire deep\nrepresentations that can be effectively fine-tuned, via standard gradient\ndescent, to new tasks. In this paper, we consider the meta-learning problem\nfrom the perspective of universality, formalizing the notion of learning\nalgorithm approximation and comparing the expressive power of the\naforementioned recurrent models to the more recent approaches that embed\ngradient descent into the meta-learner. In particular, we seek to answer the\nfollowing question: does deep representation combined with standard gradient\ndescent have sufficient capacity to approximate any learning algorithm? We find\nthat this is indeed true, and further find, in our experiments, that\ngradient-based meta-learning consistently leads to learning strategies that\ngeneralize more widely compared to those represented by recurrent models.', ""We propose gradient adversarial training, an auxiliary deep learning\nframework applicable to different machine learning problems. In gradient\nadversarial training, we leverage a prior belief that in many contexts,\nsimultaneous gradient updates should be statistically indistinguishable from\neach other. We enforce this consistency using an auxiliary network that\nclassifies the origin of the gradient tensor, and the main network serves as an\nadversary to the auxiliary network in addition to performing standard\ntask-based training. We demonstrate gradient adversarial training for three\ndifferent scenarios: (1) as a defense to adversarial examples we classify\ngradient tensors and tune them to be agnostic to the class of their\ncorresponding example, (2) for knowledge distillation, we do binary\nclassification of gradient tensors derived from the student or teacher network\nand tune the student gradient tensor to mimic the teacher's gradient tensor;\nand (3) for multi-task learning we classify the gradient tensors derived from\ndifferent task loss functions and tune them to be statistically\nindistinguishable. For each of the three scenarios we show the potential of\ngradient adversarial training procedure. Specifically, gradient adversarial\ntraining increases the robustness of a network to adversarial attacks, is able\nto better distill the knowledge from a teacher network to a student network\ncompared to soft targets, and boosts multi-task learning by aligning the\ngradient tensors derived from the task specific loss functions. Overall, our\nexperiments demonstrate that gradient tensors contain latent information about\nwhatever tasks are being trained, and can support diverse machine learning\nproblems when intelligently guided through adversarialization using a auxiliary\nnetwork."", 'The backpropagation algorithm, which had been originally introduced in the\n1970s, is the workhorse of learning in neural networks. This backpropagation\nalgorithm makes use of the famous machine learning algorithm known as Gradient\nDescent, which is a first-order iterative optimization algorithm for finding\nthe minimum of a function. To find a local minimum of a function using gradient\ndescent, one takes steps proportional to the negative of the gradient (or of\nthe approximate gradient) of the function at the current point. In this paper,\nwe develop an alternative to the backpropagation without the use of the\nGradient Descent Algorithm, but instead we are going to devise a new algorithm\nto find the error in the weights and biases of an artificial neuron using\nMoore-Penrose Pseudo Inverse. The numerical studies and the experiments\nperformed on various datasets are used to verify the working of this\nalternative algorithm.', 'Tensors are multidimensional arrays of numerical values and therefore\ngeneralize matrices to multiple dimensions. While tensors first emerged in the\npsychometrics community in the $20^{\\text{th}}$ century, they have since then\nspread to numerous other disciplines, including machine learning. Tensors and\ntheir decompositions are especially beneficial in unsupervised learning\nsettings, but are gaining popularity in other sub-disciplines like temporal and\nmulti-relational data analysis, too.\n  The scope of this paper is to give a broad overview of tensors, their\ndecompositions, and how they are used in machine learning. As part of this, we\nare going to introduce basic tensor concepts, discuss why tensors can be\nconsidered more rigid than matrices with respect to the uniqueness of their\ndecomposition, explain the most important factorization algorithms and their\nproperties, provide concrete examples of tensor decomposition applications in\nmachine learning, conduct a case study on tensor-based estimation of mixture\nmodels, talk about the current state of research, and provide references to\navailable software libraries.', 'Domain knowledge is crucial for effective performance in autonomous control\nsystems. Typically, human effort is required to encode this knowledge into a\ncontrol algorithm. In this paper, we present an approach to language grounding\nwhich automatically interprets text in the context of a complex control\napplication, such as a game, and uses domain knowledge extracted from the text\nto improve control performance. Both text analysis and control strategies are\nlearned jointly using only a feedback signal inherent to the application. To\neffectively leverage textual information, our method automatically extracts the\ntext segment most relevant to the current game state, and labels it with a\ntask-centric predicate structure. This labeled text is then used to bias an\naction selection policy for the game, guiding it towards promising regions of\nthe action space. We encode our model for text analysis and game playing in a\nmulti-layer neural network, representing linguistic decisions via latent\nvariables in the hidden layers, and game action quality via the output layer.\nOperating within the Monte-Carlo Search framework, we estimate model parameters\nusing feedback from simulated games. We apply our approach to the complex\nstrategy game Civilization II using the official game manual as the text guide.\nOur results show that a linguistically-informed game-playing agent\nsignificantly outperforms its language-unaware counterpart, yielding a 34%\nabsolute improvement and winning over 65% of games when playing against the\nbuilt-in AI of Civilization.', 'In this note, we observe the behavior of gradient flow and discrete and noisy\ngradient descent in some simple settings. It is commonly noted that addition of\nnoise to gradient descent can affect the trajectory of gradient descent. Here,\nwe run some computer experiments for gradient descent on some simple functions,\nand observe this principle in some concrete examples.', 'We derive a simple and model-independent formula for the change in the\ngeneralization gap due to a gradient descent update. We then compare the change\nin the test error for stochastic gradient descent to the change in test error\nfrom an equivalent number of gradient descent updates and show explicitly that\nstochastic gradient descent acts to regularize generalization error by\ndecorrelating nearby updates. These calculations depends on the details of the\nmodel only through the mean and covariance of the gradient distribution, which\nmay be readily measured for particular models of interest. We discuss further\nimprovements to these calculations and comment on possible implications for\nstochastic optimization.', 'Playing board games is considered a major challenge for both humans and AI\nresearchers. Because some complicated board games are quite hard to learn,\nhumans usually begin with playing on smaller boards and incrementally advance\nto master larger board strategies. Most neural network frameworks that are\ncurrently tasked with playing board games neither perform such incremental\nlearning nor possess capabilities to automatically scale up. In this work, we\nlook at the board as a graph and combine a graph neural network architecture\ninside the AlphaZero framework, along with some other innovative improvements.\nOur ScalableAlphaZero is capable of learning to play incrementally on small\nboards, and advancing to play on large ones. Our model can be trained quickly\nto play different challenging board games on multiple board sizes, without\nusing any domain knowledge. We demonstrate the effectiveness of\nScalableAlphaZero and show, for example, that by training it for only three\ndays on small Othello boards, it can defeat the AlphaZero model on a large\nboard, which was trained to play the large board for $30$ days.', 'Machine learning relies on the assumption that unseen test instances of a\nclassification problem follow the same distribution as observed training data.\nHowever, this principle can break down when machine learning is used to make\nimportant decisions about the welfare (employment, education, health) of\nstrategic individuals. Knowing information about the classifier, such\nindividuals may manipulate their attributes in order to obtain a better\nclassification outcome. As a result of this behavior---often referred to as\ngaming---the performance of the classifier may deteriorate sharply. Indeed,\ngaming is a well-known obstacle for using machine learning methods in practice;\nin financial policy-making, the problem is widely known as Goodhart\'s law. In\nthis paper, we formalize the problem, and pursue algorithms for learning\nclassifiers that are robust to gaming.\n  We model classification as a sequential game between a player named ""Jury""\nand a player named ""Contestant."" Jury designs a classifier, and Contestant\nreceives an input to the classifier, which he may change at some cost. Jury\'s\ngoal is to achieve high classification accuracy with respect to Contestant\'s\noriginal input and some underlying target classification function. Contestant\'s\ngoal is to achieve a favorable classification outcome while taking into account\nthe cost of achieving it.\n  For a natural class of cost functions, we obtain computationally efficient\nlearning algorithms which are near-optimal. Surprisingly, our algorithms are\nefficient even on concept classes that are computationally hard to learn. For\ngeneral cost functions, designing an approximately optimal strategy-proof\nclassifier, for inverse-polynomial approximation, is NP-hard.', 'Nonlinear embedding manifold learning methods provide invaluable visual\ninsights into the structure of high-dimensional data. However, due to a\ncomplicated nonconvex objective function, these methods can easily get stuck in\nlocal minima and their embedding quality can be poor. We propose a natural\nextension to several manifold learning methods aimed at identifying pressured\npoints, i.e. points stuck in poor local minima and have poor embedding quality.\nWe show that the objective function can be decreased by temporarily allowing\nthese points to make use of an extra dimension in the embedding space. Our\nmethod is able to improve the objective function value of existing methods even\nafter they get stuck in a poor local minimum.', 'As the size and complexity of models and datasets grow, so does the need for\ncommunication-efficient variants of stochastic gradient descent that can be\ndeployed to perform parallel model training. One popular\ncommunication-compression method for data-parallel SGD is QSGD (Alistarh et\nal., 2017), which quantizes and encodes gradients to reduce communication\ncosts. The baseline variant of QSGD provides strong theoretical guarantees,\nhowever, for practical purposes, the authors proposed a heuristic variant which\nwe call QSGDinf, which demonstrated impressive empirical gains for distributed\ntraining of large neural networks. In this paper, we build on this work to\npropose a new gradient quantization scheme, and show that it has both stronger\ntheoretical guarantees than QSGD, and matches and exceeds the empirical\nperformance of the QSGDinf heuristic and of other compression methods.', 'When training the parameters of a linear dynamical model, the gradient\ndescent algorithm is likely to fail to converge if the squared-error loss is\nused as the training loss function. Restricting the parameter space to a\nsmaller subset and running the gradient descent algorithm within this subset\ncan allow learning stable dynamical systems, but this strategy does not work\nfor unstable systems. In this work, we look into the dynamics of the gradient\ndescent algorithm and pinpoint what causes the difficulty of learning unstable\nsystems. We show that observations taken at different times from the system to\nbe learned influence the dynamics of the gradient descent algorithm in\nsubstantially different degrees. We introduce a time-weighted logarithmic loss\nfunction to fix this imbalance and demonstrate its effectiveness in learning\nunstable systems.', 'Machine learning methods have gained a great deal of popularity in recent\nyears among public administration scholars and practitioners. These techniques\nopen the door to the analysis of text, image and other types of data that allow\nus to test foundational theories of public administration and to develop new\ntheories. Despite the excitement surrounding machine learning methods, clarity\nregarding their proper use and potential pitfalls is lacking. This paper\nattempts to fill this gap in the literature through providing a machine\nlearning ""guide to practice"" for public administration scholars and\npractitioners. Here, we take a foundational view of machine learning and\ndescribe how these methods can enrich public administration research and\npractice through their ability develop new measures, tap into new sources of\ndata and conduct statistical inference and causal inference in a principled\nmanner. We then turn our attention to the pitfalls of using these methods such\nas unvalidated measures and lack of interpretability. Finally, we demonstrate\nhow machine learning techniques can help us learn about organizational\nreputation in federal agencies through an illustrated example using tweets from\n13 executive federal agencies.', 'Game theory provides essential analysis in many applications of strategic\ninteractions. However, the question of how to construct a game model and what\nis its fidelity is seldom addressed. In this work, we consider learning in a\nclass of repeated zero-sum games with unknown, time-varying payoff matrix, and\nnoisy feedbacks, by making use of an ensemble of benchmark game models. These\nmodels can be pre-trained and collected dynamically during sequential plays.\nThey serve as prior side information and imperfectly underpin the unknown true\ngame model. We propose OFULinMat, an episodic learning algorithm that\nintegrates the adaptive estimation of game models and the learning of the\nstrategies. The proposed algorithm is shown to achieve a sublinear bound on the\nsaddle-point regret. We show that this algorithm is provably efficient through\nboth theoretical analysis and numerical examples. We use a dynamic honeypot\nallocation game as a case study to illustrate and corroborate our results. We\nalso discuss the relationship and highlight the difference between our\nframework and the classical adversarial multi-armed bandit framework.', 'Recently, researchers have discovered that the state-of-the-art object\nclassifiers can be fooled easily by small perturbations in the input\nunnoticeable to human eyes. It is also known that an attacker can generate\nstrong adversarial examples if she knows the classifier parameters. Conversely,\na defender can robustify the classifier by retraining if she has access to the\nadversarial examples. We explain and formulate this adversarial example problem\nas a two-player continuous zero-sum game, and demonstrate the fallacy of\nevaluating a defense or an attack as a static problem. To find the best\nworst-case defense against whitebox attacks, we propose a continuous minimax\noptimization algorithm. We demonstrate the minimax defense with two types of\nattack classes -- gradient-based and neural network-based attacks. Experiments\nwith the MNIST and the CIFAR-10 datasets demonstrate that the defense found by\nnumerical minimax optimization is indeed more robust than non-minimax defenses.\nWe discuss directions for improving the result toward achieving robustness\nagainst multiple types of attack classes.', 'Spectral learning recently generated lots of excitement in machine learning,\nlargely because it is the first known method to produce consistent estimates\n(under suitable conditions) for several latent variable models. In contrast,\nmaximum likelihood estimates may get trapped in local optima due to the\nnon-convex nature of the likelihood function of latent variable models. In this\npaper, we do an empirical evaluation of spectral learning (SL) and expectation\nmaximization (EM), which reveals an important gap between the theory and the\npractice. First, SL often leads to negative probabilities. Second, EM often\nyields better estimates than spectral learning and it does not seem to get\nstuck in local optima. We discuss how the rank of the model parameters and the\namount of training data can yield negative probabilities. We also question the\ncommon belief that maximum likelihood estimators are necessarily inconsistent.', 'We describe an alternative learning method for neural networks, which we call\nBlind Descent. By design, Blind Descent does not face problems like exploding\nor vanishing gradients. In Blind Descent, gradients are not used to guide the\nlearning process. In this paper, we present Blind Descent as a more fundamental\nlearning process compared to gradient descent. We also show that gradient\ndescent can be seen as a specific case of the Blind Descent algorithm. We also\ntrain two neural network architectures, a multilayer perceptron and a\nconvolutional neural network, using the most general Blind Descent algorithm to\ndemonstrate a proof of concept.', 'Predicting and modeling human behavior and finding trends within human\ndecision-making processes is a major problem of social science. Rock Paper\nScissors (RPS) is the fundamental strategic question in many game theory\nproblems and real-world competitions. Finding the right approach to beat a\nparticular human opponent is challenging. Here we use an AI (artificial\nintelligence) algorithm based on Markov Models of one fixed memory length\n(abbreviated as ""single AI"") to compete against humans in an iterated RPS game.\nWe model and predict human competition behavior by combining many Markov Models\nwith different fixed memory lengths (abbreviated as ""multi-AI""), and develop an\narchitecture of multi-AI with changeable parameters to adapt to different\ncompetition strategies. We introduce a parameter called ""focus length"" (a\npositive number such as 5 or 10) to control the speed and sensitivity for our\nmulti-AI to adapt to the opponent\'s strategy change. The focus length is the\nnumber of previous rounds that the multi-AI should look at when determining\nwhich Single-AI has the best performance and should choose to play for the next\ngame. We experimented with 52 different people, each playing 300 rounds\ncontinuously against one specific multi-AI model, and demonstrated that our\nstrategy could win against more than 95% of human opponents.', 'Machine unlearning is the process through which a deployed machine learning\nmodel is made to forget about some of its training data points. While naively\nretraining the model from scratch is an option, it is almost always associated\nwith large computational overheads for deep learning models. Thus, several\napproaches to approximately unlearn have been proposed along with corresponding\nmetrics that formalize what it means for a model to forget about a data point.\nIn this work, we first taxonomize approaches and metrics of approximate\nunlearning. As a result, we identify verification error, i.e., the L2\ndifference between the weights of an approximately unlearned and a naively\nretrained model, as an approximate unlearning metric that should be optimized\nfor as it subsumes a large class of other metrics. We theoretically analyze the\ncanonical training algorithm, stochastic gradient descent (SGD), to surface the\nvariables which are relevant to reducing the verification error of approximate\nunlearning for SGD. From this analysis, we first derive an easy-to-compute\nproxy for verification error (termed unlearning error). The analysis also\ninforms the design of a new training objective penalty that limits the overall\nchange in weights during SGD and as a result facilitates approximate unlearning\nwith lower verification error. We validate our theoretical work through an\nempirical evaluation on learning with CIFAR-10, CIFAR-100, and IMDB sentiment\nanalysis.', 'Machine learning has now become an integral part of research and innovation.\nThe field of machine learning density functional theory has continuously\nexpanded over the years while making several noticeable advances. We briefly\ndiscuss the status of this field and point out some current and future\nchallenges. We also talk about how state-of-the-art science and technology\ntools can help overcome these challenges. This draft is a part of the ""Roadmap\non Machine Learning in Electronic Structure"" to be published in Electronic\nStructure (EST).', 'In the last two years, more than 200 papers have been written on how machine\nlearning (ML) systems can fail because of adversarial attacks on the algorithms\nand data; this number balloons if we were to incorporate papers covering\nnon-adversarial failure modes. The spate of papers has made it difficult for ML\npractitioners, let alone engineers, lawyers, and policymakers, to keep up with\nthe attacks against and defenses of ML systems. However, as these systems\nbecome more pervasive, the need to understand how they fail, whether by the\nhand of an adversary or due to the inherent design of a system, will only\nbecome more pressing. In order to equip software developers, security incident\nresponders, lawyers, and policy makers with a common vernacular to talk about\nthis problem, we developed a framework to classify failures into ""Intentional\nfailures"" where the failure is caused by an active adversary attempting to\nsubvert the system to attain her goals; and ""Unintentional failures"" where the\nfailure is because an ML system produces an inherently unsafe outcome. After\ndeveloping the initial version of the taxonomy last year, we worked with\nsecurity and ML teams across Microsoft, 23 external partners, standards\norganization, and governments to understand how stakeholders would use our\nframework. Throughout the paper, we attempt to highlight how machine learning\nfailure modes are meaningfully different from traditional software failures\nfrom a technology and policy perspective.', 'Machine learning is often used in competitive scenarios: Participants learn\nand fit static models, and those models compete in a shared platform. The\ncommon assumption is that in order to win a competition one has to have the\nbest predictive model, i.e., the model with the smallest out-sample error. Is\nthat necessarily true? Does the best theoretical predictive model for a target\nalways yield the best reward in a competition? If not, can one take the best\nmodel and purposefully change it into a theoretically inferior model which in\npractice results in a higher competitive edge? How does that modification look\nlike? And finally, if all participants modify their prediction models towards\nthe best practical performance, who benefits the most? players with inferior\nmodels, or those with theoretical superiority? The main theme of this paper is\nto raise these important questions and propose a theoretical model to answer\nthem. We consider a study case where two linear predictive models compete over\na shared target. The model with the closest estimate gets the whole reward,\nwhich is equal to the absolute value of the target. We characterize the reward\nfunction of each model, and using a basic game theoretic approach, demonstrate\nthat the inferior competitor can significantly improve his performance by\nchoosing optimal model coefficients that are different from the best\ntheoretical prediction. This is a preliminary study that emphasizes the fact\nthat in many applications where predictive machine learning is at the service\nof competition, much can be gained from practical (back-testing) optimization\nof the model compared to static prediction improvement.', 'Success in the quest for artificial intelligence has the potential to bring\nunprecedented benefits to humanity, and it is therefore worthwhile to\ninvestigate how to maximize these benefits while avoiding potential pitfalls.\nThis article gives numerous examples (which should by no means be construed as\nan exhaustive list) of such worthwhile research aimed at ensuring that AI\nremains robust and beneficial.', 'In this paper, a new sequential surrogate-based optimization (SSBO) algorithm\nis developed, which aims to improve the global search ability and local search\nefficiency for the global optimization of expensive black-box models. The\nproposed method involves three basic sub-criteria to infill new samples\nasynchronously to balance the global exploration and local exploitation. First,\nto capture the promising possible global optimal region, searching for the\nglobal optimum with genetic algorithm (GA) based on the current surrogate\nmodels of the objective and constraint functions. Second, to infill samples in\nthe region with sparse samples to improve the global accuracy of the surrogate\nmodels, a grid searching with Latin hypercube sampling (LHS) with the current\nsurrogate model is adopted to explore the sample space. Third, to accelerate\nthe local searching efficiency, searching for a local optimum with sequential\nquadratic programming (SQP) based on the local surrogate models in the reduced\ninterval, which involves some samples near the current optimum. When the new\nsample is too close to the existing ones, the new sample should be abandoned,\ndue to the poor additional information. According to the three sub-criteria,\nthe new samples are placed in the regions which have not been fully explored\nand includes the possible global optimum point. When a possible global optimum\npoint is found, the local searching sub-criterion captures the local optimum\naround it rapidly. Numerical and engineering examples are used to verify the\nefficiency of the proposed method. The statistical results show that the\nproposed method has good global searching ability and efficiency.', ""We consider a formulation of a non zero-sum n players game by an n+1 players\nzero-sum game. We suppose the existence of the n+1-th player in addition to n\nplayers in the main game, and virtual subsidies to the n players which is\nprovided by the n+1-th player. Its strategic variable affects only the\nsubsidies, and does not affect choice of strategies by the n players in the\nmain game. His objective function is the opposite of the sum of the payoffs of\nthe n players. We will show 1) The minimax theorem by Sion (Sion(1958)) implies\nthe existence of Nash equilibrium in the n players non zero-sum game. 2) The\nmaximin strategy of each player in {1, 2, \\dots, n} with the minimax strategy\nof the n+1-th player is equivalent to the Nash equilibrium strategy of the n\nplayers non zero-sum game. 3) The existence of Nash equilibrium in the n\nplayers non zero-sum game implies Sion's minimax theorem for pairs of each of\nthe n players and the n+1-th player."", ""Our ability to know when to trust the decisions made by machine learning\nsystems has not kept up with the staggering improvements in their performance,\nlimiting their applicability in high-stakes domains. We introduce\nProver-Verifier Games (PVGs), a game-theoretic framework to encourage learning\nagents to solve decision problems in a verifiable manner. The PVG consists of\ntwo learners with competing objectives: a trusted verifier network tries to\nchoose the correct answer, and a more powerful but untrusted prover network\nattempts to persuade the verifier of a particular answer, regardless of its\ncorrectness. The goal is for a reliable justification protocol to emerge from\nthis game. We analyze variants of the framework, including simultaneous and\nsequential games, and narrow the space down to a subset of games which provably\nhave the desired equilibria. We develop instantiations of the PVG for two\nalgorithmic tasks, and show that in practice, the verifier learns a robust\ndecision rule that is able to receive useful and reliable information from an\nuntrusted prover. Importantly, the protocol still works even when the verifier\nis frozen and the prover's messages are directly optimized to convince the\nverifier."", ""Prior work on training generative Visual Dialog models with reinforcement\nlearning(Das et al.) has explored a Qbot-Abot image-guessing game and shown\nthat this 'self-talk' approach can lead to improved performance at the\ndownstream dialog-conditioned image-guessing task. However, this improvement\nsaturates and starts degrading after a few rounds of interaction, and does not\nlead to a better Visual Dialog model. We find that this is due in part to\nrepeated interactions between Qbot and Abot during self-talk, which are not\ninformative with respect to the image. To improve this, we devise a simple\nauxiliary objective that incentivizes Qbot to ask diverse questions, thus\nreducing repetitions and in turn enabling Abot to explore a larger state space\nduring RL ie. be exposed to more visual concepts to talk about, and varied\nquestions to answer. We evaluate our approach via a host of automatic metrics\nand human studies, and demonstrate that it leads to better dialog, ie. dialog\nthat is more diverse (ie. less repetitive), consistent (ie. has fewer\nconflicting exchanges), fluent (ie. more human-like),and detailed, while still\nbeing comparably image-relevant as prior work and ablations."", 'Recent work in reinforcement learning demonstrated that learning solely\nthrough self-play is not only possible, but could also result in novel\nstrategies that humans never would have thought of. However, optimization\nmethods cast as a game between two players require careful tuning to prevent\nsuboptimal results. Hence, we look at random play as an alternative method. In\nthis paper, we train a DQN agent to play Sungka, a two-player turn-based board\ngame wherein the players compete to obtain more stones than the other. We show\nthat even with purely random play, our training algorithm converges very fast\nand is stable. Moreover, we test our trained agent against several baselines\nand show its ability to consistently win against these.', 'In autonomous vehicle (AV) control, allowing mistakes can be quite dangerous\nand costly in the real world. For this reason we investigate methods of\ntraining an AV without allowing the agent to explore and instead having a human\nexplorer collect the data. Supervised learning has been explored for AV\ncontrol, but it encounters the issue of the covariate shift. That is, training\ndata collected from an optimal demonstration consists only of the states\ninduced by the optimal control policy, but at runtime, the trained agent may\nencounter a vastly different state distribution with little relevant training\ndata. To mitigate this issue, we have our human explorer make sub-optimal\ndecisions. In order to have our agent not replicate these sub-optimal\ndecisions, supervised learning requires that we either erase these actions, or\nreplace these action with the correct action. Erasing is wasteful and replacing\nis difficult, since it is not easy to know the correct action without driving.\nWe propose an alternate framework that includes continuous scalar feedback for\neach action, marking which actions we should replicate, which we should avoid,\nand how sure we are. Our framework learns continuous control from sub-optimal\ndemonstration and evaluative feedback collected before training. We find that a\nhuman demonstrator can explore sub-optimal states in a safe manner, while still\ngetting enough gradation to benefit learning. The collection method for data\nand feedback we call ""Backseat Driver."" We call the more general learning\nframework ReNeg, since it learns a regression from states to actions given\nnegative as well as positive examples. We empirically validate several models\nin the ReNeg framework, testing on lane-following with limited data. We find\nthat the best solution is a generalization of mean-squared error and\noutperforms supervised learning on the positive examples alone.', 'Resampling is a key component of sample-based recursive state estimation in\nparticle filters. Recent work explores differentiable particle filters for\nend-to-end learning. However, resampling remains a challenge in these works, as\nit is inherently non-differentiable. We address this challenge by replacing\ntraditional resampling with a learned neural network resampler. We present a\nnovel network architecture, the particle transformer, and train it for particle\nresampling using a likelihood-based loss function over sets of particles.\nIncorporated into a differentiable particle filter, our model can be end-to-end\noptimized jointly with the other particle filter components via gradient\ndescent. Our results show that our learned resampler outperforms traditional\nresampling techniques on synthetic data and in a simulated robot localization\ntask.', ""This paper presents an acoustic echo canceler based on a U-Net convolutional\nneural network for single-talk and double-talk scenarios. U-Net networks have\npreviously been used in the audio processing area for source separation\nproblems because of their ability to reproduce the finest details of audio\nsignals, but to our knowledge, this is the first time they have been used for\nacoustic echo cancellation (AEC). The U-Net hyperparameters have been optimized\nto obtain the best AEC performance, but using a reduced number of parameters to\nmeet a latency restriction of 40 ms. The training and testing of our model have\nbeen carried out within the framework of the 'ICASSP 2021 AEC Challenge'\norganized by Microsoft. We have trained the optimized U-Net model with a\nsynthetic dataset only (S-U-Net) and with a synthetic dataset and the\nsingle-talk set of a real dataset (SR-U-Net), both datasets were released for\nthe challenge. The S-U-Net model presented better results for double-talk\nscenarios, thus their inferred near-end signals from the blind testset were\nsubmitted to the challenge. Our canceler ranked 12th among 17 teams, and 5th\namong 10 academia teams, obtaining an overall mean opinion score of 3.57."", 'An approach for game bot detection in MMORPGs is proposed based on the\nanalysis of game playing behavior. Since MMORPGs are large scale games, users\ncan play in various ways. This variety in playing behavior makes it hard to\ndetect game bots based on play behaviors. In order to cope with this problem,\nthe proposed approach observes game playing behaviors of users and groups them\nby their behavioral similarities. Then, it develops a local bot detection model\nfor each player group. Since the locally optimized models can more accurately\ndetect game bots within each player group, the combination of those models\nbrings about overall improvement. For a practical purpose of reducing the\nworkloads of the game servers in service, the game data is collected at a low\nresolution in time. Behavioral features are selected and developed to\naccurately detect game bots with the low resolution data, considering common\naspects of MMORPG playing. Through the experiment with the real data from a\ngame currently in service, it is shown that the proposed local model approach\nyields more accurate results.', 'Scientific experiments are usually expensive due to complex experimental\npreparation and processing. Experimental design is therefore involved with the\ntask of finding the optimal experimental input that results in the desirable\noutput by using as few experiments as possible. Experimenters can often acquire\nthe knowledge about the location of the global optimum. However, they do not\nknow how to exploit this knowledge to accelerate experimental design. In this\npaper, we adopt the technique of Bayesian optimization for experimental design\nsince Bayesian optimization has established itself as an efficient tool for\noptimizing expensive black-box functions. Again, it is unknown how to\nincorporate the expert prior knowledge about the global optimum into Bayesian\noptimization process. To address it, we represent the expert knowledge about\nthe global optimum via placing a prior distribution on it and we then derive\nits posterior distribution. An efficient Bayesian optimization approach has\nbeen proposed via posterior sampling on the posterior distribution of the\nglobal optimum. We theoretically analyze the convergence of the proposed\nalgorithm and discuss the robustness of incorporating expert prior. We evaluate\nthe efficiency of our algorithm by optimizing synthetic functions and tuning\nhyperparameters of classifiers along with a real-world experiment on the\nsynthesis of short polymer fiber. The results clearly demonstrate the\nadvantages of our proposed method.', 'Multi-agent reinforcement learning offers a way to study how communication\ncould emerge in communities of agents needing to solve specific problems. In\nthis paper, we study the emergence of communication in the negotiation\nenvironment, a semi-cooperative model of agent interaction. We introduce two\ncommunication protocols -- one grounded in the semantics of the game, and one\nwhich is \\textit{a priori} ungrounded and is a form of cheap talk. We show that\nself-interested agents can use the pre-grounded communication channel to\nnegotiate fairly, but are unable to effectively use the ungrounded channel.\nHowever, prosocial agents do learn to use cheap talk to find an optimal\nnegotiating strategy, suggesting that cooperation is necessary for language to\nemerge. We also study communication behaviour in a setting where one agent\ninteracts with agents in a community with different levels of prosociality and\nshow how agent identifiability can aid negotiation.', 'When a channel model is not available, the end-to-end training of encoder and\ndecoder on a fading noisy channel generally requires the repeated use of the\nchannel and of a feedback link. An important limitation of the approach is that\ntraining should be generally carried out from scratch for each new channel. To\ncope with this problem, prior works considered joint training over multiple\nchannels with the aim of finding a single pair of encoder and decoder that\nworks well on a class of channels. In this paper, we propose to obviate the\nlimitations of joint training via meta-learning. The proposed approach is based\non a meta-training phase in which the online gradient-based meta-learning of\nthe decoder is coupled with the joint training of the encoder via the\ntransmission of pilots and the use of a feedback link. Accounting for channel\nvariations during the meta-training phase, this work demonstrates the\nadvantages of meta-learning in terms of number of pilots as compared to\nconventional methods when the feedback link is only available for meta-training\nand not at run time.', 'The game of baccarat has evolved from a parlor game played by French\naristocrats in the first half of the 19th century to a casino game that\ngenerated over US\\$41 billion in revenue for the casinos of Macau in 2013. The\nparlor game was originally a three-person zero-sum game. Later in the 19th\ncentury it was simplified to a two-person zero-sum game. Early in the 20th\ncentury the parlor game became a casino game, no longer zero-sum. In the mid\n20th century, the strategic casino game became a nonstrategic game, with\nplayers competing against the house instead of against each other. We argue\nthat this evolution was motivated by both economic and game-theoretic\nconsiderations.', ""The primary objective is to teach a machine about human emotions, which has\nbecome an essential requirement in the field of social intelligence, also\nexpedites the progress of human-machine interactions. The ability of a machine\nto understand human emotion and act accordingly has been a choice of great\ninterest in today's world. The future generations of computers thus must be\nable to interact with a human being just like another. For example, people who\nhave Autism often find it difficult to talk to someone about their state of\nmind. This model explicitly targets the userbase who are troubled and fail to\nexpress it. Also, this model's speech processing techniques provide an estimate\nof the emotion in the case of poor video quality and vice-versa."", 'Recently we see a rising number of methods in the field of eXplainable\nArtificial Intelligence. To our surprise, their development is driven by model\ndevelopers rather than a study of needs for human end users. The analysis of\nneeds, if done, takes the form of an A/B test rather than a study of open\nquestions. To answer the question ""What would a human operator like to ask the\nML model?"" we propose a conversational system explaining decisions of the\npredictive model. In this experiment, we developed a chatbot called dr_ant to\ntalk about machine learning model trained to predict survival odds on Titanic.\nPeople can talk with dr_ant about different aspects of the model to understand\nthe rationale behind its predictions. Having collected a corpus of 1000+\ndialogues, we analyse the most common types of questions that users would like\nto ask. To our knowledge, it is the first study which uses a conversational\nsystem to collect the needs of human operators from the interactive and\niterative dialogue explorations of a predictive model.', 'With the growing technology of photon-counting detectors (PCD), spectral CT\nis a widely concerned topic which has the potential of material\ndifferentiation. However, due to some non-ideal factors such as cross talk and\npulse pile-up of the detectors, direct reconstruction from detected spectrum\nwithout any corrections will get a wrong result. Conventional methods try to\nmodel these factors using calibration and make corrections accordingly, but\ndepend on the preciseness of the model. To solve this problem, in this paper,\nwe proposed a novel deep learning-based monochromatic image synthesis method\nworking in sinogram domain. Different from previous deep learning-based methods\naimed at this problem, we designed a novel network architecture according to\nthe physical model of cross talk, and it can solve this problem better in an\ningenious way. Our method was tested on a cone-beam CT (CBCT) system equipped\nwith a PCD. After using FDK algorithm on the corrected projection, we got quite\nmore accurate results with less noise, which showed the feasibility of\nmonochromatic image synthesis by our method.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : As a data analyst working with high-dimensional time series data, I am seeking to develop a model capable of handling large datasets, even those with missing values. The model should be able to process data online efficiently, representing high-dimensional data using low-dimensional embeddings without any loss of information. My goal is to identify a method that deconstructs embeddings of large data into simpler components, thereby facilitating efficient learning while simultaneously adjusting to new information online. Subsequently, I aim to train a model that forecasts variables' values based on their previous embeddings. I am particularly interested in testing this method on large, real-world datasets to evaluate its prediction accuracy. Ultimately, my objective is to utilize this model for predicting trends in sectors such as finance, energy, and transportation, especially when dealing with complex time series data.","['In Internet of things (IoT), data is continuously recorded from different\ndata sources and devices can suffer faults in their embedded electronics, thus\nleading to a high-dimensional data sets and concept drift events. Therefore,\nmethods that are capable of high-dimensional non-stationary time series are of\ngreat value in IoT applications. Fuzzy Time Series (FTS) models stand out as\ndata-driven non-parametric models of easy implementation and high accuracy.\nUnfortunately, FTS encounters difficulties when dealing with data sets of many\nvariables and scenarios with concept drift. We present a new approach to handle\nhigh-dimensional non-stationary time series, by projecting the original\nhigh-dimensional data into a low dimensional embedding space and using FTS\napproach. Combining these techniques enables a better representation of the\ncomplex content of non-stationary multivariate time series and accurate\nforecasts. Our model is able to explain 98% of the variance and reach 11.52% of\nRMSE, 2.68% of MAE and 2.91% of MAPE.', 'Efficient processing of large-scale time series data is an intricate problem\nin machine learning. Conventional sensor signal processing pipelines with hand\nengineered feature extraction often involve huge computational cost with high\ndimensional data. Deep recurrent neural networks have shown promise in\nautomated feature learning for improved time-series processing. However,\ngeneric deep recurrent models grow in scale and depth with increased complexity\nof the data. This is particularly challenging in presence of high dimensional\ndata with temporal and spatial characteristics. Consequently, this work\nproposes a novel deep cellular recurrent neural network (DCRNN) architecture to\nefficiently process complex multi-dimensional time series data with spatial\ninformation. The cellular recurrent architecture in the proposed model allows\nfor location-aware synchronous processing of time series data from spatially\ndistributed sensor signal sources. Extensive trainable parameter sharing due to\ncellularity in the proposed architecture ensures efficiency in the use of\nrecurrent processing units with high-dimensional inputs. This study also\ninvestigates the versatility of the proposed DCRNN model for classification of\nmulti-class time series data from different application domains. Consequently,\nthe proposed DCRNN architecture is evaluated using two time-series datasets: a\nmultichannel scalp EEG dataset for seizure detection, and a machine fault\ndetection dataset obtained in-house. The results suggest that the proposed\narchitecture achieves state-of-the-art performance while utilizing\nsubstantially less trainable parameters when compared to comparable methods in\nthe literature.', 'The prediction of residential power usage is essential in assisting a smart\ngrid to manage and preserve energy to ensure efficient use. An accurate energy\nforecasting at the customer level will reflect directly into efficiency\nimprovements across the power grid system, however forecasting building energy\nuse is a complex task due to many influencing factors, such as meteorological\nand occupancy patterns. In addiction, high-dimensional time series increasingly\narise in the Internet of Energy (IoE), given the emergence of multi-sensor\nenvironments and the two way communication between energy consumers and the\nsmart grid. Therefore, methods that are capable of computing high-dimensional\ntime series are of great value in smart building and IoE applications. Fuzzy\nTime Series (FTS) models stand out as data-driven non-parametric models of easy\nimplementation and high accuracy. Unfortunately, the existing FTS models can be\nunfeasible if all features were used to train the model. We present a new\nmethodology for handling high-dimensional time series, by projecting the\noriginal high-dimensional data into a low dimensional embedding space and using\nmultivariate FTS approach in this low dimensional representation. Combining\nthese techniques enables a better representation of the complex content of\nmultivariate time series and more accurate forecasts.', 'Learning dynamical models from data plays a vital role in engineering design,\noptimization, and predictions. Building models describing dynamics of complex\nprocesses (e.g., weather dynamics, or reactive flows) using empirical knowledge\nor first principles are onerous or infeasible. Moreover, these models are\nhigh-dimensional but spatially correlated. It is, however, observed that the\ndynamics of high-fidelity models often evolve in low-dimensional manifolds.\nFurthermore, it is also known that for sufficiently smooth vector fields\ndefining the nonlinear dynamics, a quadratic model can describe it accurately\nin an appropriate coordinate system, conferring to the McCormick relaxation\nidea in nonconvex optimization. Here, we aim at finding a low-dimensional\nembedding of high-fidelity dynamical data, ensuring a simple quadratic model to\nexplain its dynamics. To that aim, this work leverages deep learning to\nidentify low-dimensional quadratic embeddings for high-fidelity dynamical\nsystems. Precisely, we identify the embedding of data using an autoencoder to\nhave the desired property of the embedding. We also embed a Runge-Kutta method\nto avoid the time-derivative computations, which is often a challenge. We\nillustrate the ability of the approach by a couple of examples, arising in\ndescribing flow dynamics and the oscillatory tubular reactor model.', 'Spatial time series forecasting problems arise in a broad range of\napplications, such as environmental and transportation problems. These problems\nare challenging because of the existence of specific spatial, short-term and\nlong-term patterns, and the curse of dimensionality. In this paper, we propose\na deep neural network framework for large-scale spatial time series forecasting\nproblems. We explicitly designed the neural network architecture for capturing\nvarious types of patterns. In preprocessing, a time series decomposition method\nis applied to separately feed short-term, long-term and spatial patterns into\ndifferent components of a neural network. A fuzzy clustering method finds\ncluster of neighboring time series based on similarity of time series\nresiduals; as they can be meaningful short-term patterns for spatial time\nseries. In neural network architecture, each kernel of a multi-kernel\nconvolution layer is applied to a cluster of time series to extract short-term\nfeatures in neighboring areas. The output of convolution layer is concatenated\nby trends and followed by convolution-LSTM layer to capture long-term patterns\nin larger regional areas. To make a robust prediction when faced with missing\ndata, an unsupervised pretrained denoising autoencoder reconstructs the output\nof the model in a fine-tuning step. The experimental results illustrate the\nmodel outperforms baseline and state of the art models in a traffic flow\nprediction dataset.', 'Machine learning and in particular deep learning algorithms are the emerging\napproaches to data analysis. These techniques have transformed traditional data\nmining-based analysis radically into a learning-based model in which existing\ndata sets along with their cluster labels (i.e., train set) are learned to\nbuild a supervised learning model and predict the cluster labels of unseen data\n(i.e., test set). In particular, deep learning techniques are capable of\ncapturing and learning hidden features in a given data sets and thus building a\nmore accurate prediction model for clustering and labeling problem. However,\nthe major problem is that time series data are often unlabeled and thus\nsupervised learning-based deep learning algorithms cannot be directly adapted\nto solve the clustering problems for these special and complex types of data\nsets. To address this problem, this paper introduces a two-stage method for\nclustering time series data. First, a novel technique is introduced to utilize\nthe characteristics (e.g., volatility) of given time series data in order to\ncreate labels and thus be able to transform the problem from unsupervised\nlearning into supervised learning. Second, an autoencoder-based deep learning\nmodel is built to learn and model both known and hidden features of time series\ndata along with their created labels to predict the labels of unseen time\nseries data. The paper reports a case study in which financial and stock time\nseries data of selected 70 stock indices are clustered into distinct groups\nusing the introduced two-stage procedure. The results show that the proposed\nprocedure is capable of achieving 87.5\\% accuracy in clustering and predicting\nthe labels for unseen time series data.', 'We propose a three-stage framework for forecasting high-dimensional\ntime-series data. Our method first estimates parameters for each univariate\ntime series. Next, we use these parameters to cluster the time series. These\nclusters can be viewed as multivariate time series, for which we then compute\nparameters. The forecasted values of a single time series can depend on the\nhistory of other time series in the same cluster, accounting for intra-cluster\nsimilarity while minimizing potential noise in predictions by ignoring\ninter-cluster effects. Our framework -- which we refer to as\n""cluster-and-conquer"" -- is highly general, allowing for any time-series\nforecasting and clustering method to be used in each step. It is\ncomputationally efficient and embarrassingly parallel. We motivate our\nframework with a theoretical analysis in an idealized mixed linear regression\nsetting, where we provide guarantees on the quality of the estimates. We\naccompany these guarantees with experimental results that demonstrate the\nadvantages of our framework: when instantiated with simple linear\nautoregressive models, we are able to achieve state-of-the-art results on\nseveral benchmark datasets, sometimes outperforming deep-learning-based\napproaches.', 'Informative missingness is unavoidable in the digital processing of\ncontinuous time series, where the value for one or more observations at\ndifferent time points are missing. Such missing observations are one of the\nmajor limitations of time series processing using deep learning. Practical\napplications, e.g., sensor data, healthcare, weather, generates data that is in\ntruth continuous in time, and informative missingness is a common phenomenon in\nthese datasets. These datasets often consist of multiple variables, and often\nthere are missing values for one or many of these variables. This\ncharacteristic makes time series prediction more challenging, and the impact of\nmissing input observations on the accuracy of the final output can be\nsignificant. A recent novel deep learning model called GRU-D is one early\nattempt to address informative missingness in time series data. On the other\nhand, a new family of neural networks called Neural ODEs (Ordinary Differential\nEquations) are natural and efficient for processing time series data which is\ncontinuous in time. In this paper, a deep learning model is proposed that\nleverages the effective imputation of GRU-D, and the temporal continuity of\nNeural ODEs. A time series classification task performed on the PhysioNet\ndataset demonstrates the performance of this architecture.', 'Recent advances in matrix completion enable data imputation in full-rank\nmatrices by exploiting low dimensional (nonlinear) latent structure. In this\npaper, we develop a new model for high rank matrix completion (HRMC), together\nwith batch and online methods to fit the model and out-of-sample extension to\ncomplete new data. The method works by (implicitly) mapping the data into a\nhigh dimensional polynomial feature space using the kernel trick; importantly,\nthe data occupies a low dimensional subspace in this feature space, even when\nthe original data matrix is of full-rank. We introduce an explicit\nparametrization of this low dimensional subspace, and an online fitting\nprocedure, to reduce computational complexity compared to the state of the art.\nThe online method can also handle streaming or sequential data and adapt to\nnon-stationary latent structure. We provide guidance on the sampling rate\nrequired these methods to succeed. Experimental results on synthetic data and\nmotion capture data validate the performance of the proposed methods.', 'Non-linear manifold learning enables high-dimensional data analysis, but\nrequires out-of-sample-extension methods to process new data points. In this\npaper, we propose a manifold learning algorithm based on deep learning to\ncreate an encoder, which maps a high-dimensional dataset and its\nlow-dimensional embedding, and a decoder, which takes the embedded data back to\nthe high-dimensional space. Stacking the encoder and decoder together\nconstructs an autoencoder, which we term a diffusion net, that performs\nout-of-sample-extension as well as outlier detection. We introduce new neural\nnet constraints for the encoder, which preserves the local geometry of the\npoints, and we prove rates of convergence for the encoder. Also, our approach\nis efficient in both computational complexity and memory requirements, as\nopposed to previous methods that require storage of all training points in both\nthe high-dimensional and the low-dimensional spaces to calculate the\nout-of-sample-extension and the pre-image.', 'Large-scale and multidimensional spatiotemporal data sets are becoming\nubiquitous in many real-world applications such as monitoring urban traffic and\nair quality. Making predictions on these time series has become a critical\nchallenge due to not only the large-scale and high-dimensional nature but also\nthe considerable amount of missing data. In this paper, we propose a Bayesian\ntemporal factorization (BTF) framework for modeling multidimensional time\nseries -- in particular spatiotemporal data -- in the presence of missing\nvalues. By integrating low-rank matrix/tensor factorization and vector\nautoregressive (VAR) process into a single probabilistic graphical model, this\nframework can characterize both global and local consistencies in large-scale\ntime series data. The graphical model allows us to effectively perform\nprobabilistic predictions and produce uncertainty estimates without imputing\nthose missing values. We develop efficient Gibbs sampling algorithms for model\ninference and model updating for real-time prediction and test the proposed BTF\nframework on several real-world spatiotemporal data sets for both missing data\nimputation and multi-step rolling prediction tasks. The numerical experiments\ndemonstrate the superiority of the proposed BTF approaches over existing\nstate-of-the-art methods.', 'This paper proposes a probabilistic neural network developed on the basis of\ntime-series discriminant component analysis (TSDCA) that can be used to\nclassify high-dimensional time-series patterns. TSDCA involves the compression\nof high-dimensional time series into a lower-dimensional space using a set of\northogonal transformations and the calculation of posterior probabilities based\non a continuous-density hidden Markov model with a Gaussian mixture model\nexpressed in the reduced-dimensional space. The analysis can be incorporated\ninto a neural network, which is named a time-series discriminant component\nnetwork (TSDCN), so that parameters of dimensionality reduction and\nclassification can be obtained simultaneously as network coefficients according\nto a backpropagation through time-based learning algorithm with the Lagrange\nmultiplier method. The TSDCN is considered to enable high-accuracy\nclassification of high-dimensional time-series patterns and to reduce the\ncomputation time taken for network training. The validity of the TSDCN is\ndemonstrated for high-dimensional artificial data and EEG signals in the\nexperiments conducted during the study.', 'Time series data account for a major part of data supply available today.\nTime series mining handles several tasks such as classification, clustering,\nquery-by-content, prediction, and others. Performing data mining tasks on raw\ntime series is inefficient as these data are high-dimensional by nature.\nInstead, time series are first pre-processed using several techniques before\ndifferent data mining tasks can be performed on them. In general, there are two\nmain approaches to reduce time series dimensionality, the first is what we call\nlandmark methods. These methods are based on finding characteristic features in\nthe target time series. The second is based on data transformations. These\nmethods transform the time series from the original space into a reduced space,\nwhere they can be managed more efficiently. The method we present in this paper\napplies a third approach, as it projects a time series onto a lower-dimensional\nspace by selecting important points in the time series. The novelty of our\nmethod is that these points are not chosen according to a geometric criterion,\nwhich is subjective in most cases, but through an optimization process. The\nother important characteristic of our method is that these important points are\nselected on a dataset-level and not on a single time series-level. The direct\nadvantage of this strategy is that the distance defined on the low-dimensional\nspace lower bounds the original distance applied to raw data. This enables us\nto apply the popular GEMINI algorithm. The promising results of our experiments\non a wide variety of time series datasets, using different optimizers, and\napplied to the two major data mining tasks, validate our new method.', ""Multivariate time series forecasting is of great importance to many\nscientific disciplines and industrial sectors. The evolution of a multivariate\ntime series depends on the dynamics of its variables and the connectivity\nnetwork of causal interrelationships among them. Most of the existing time\nseries models do not account for the causal effects among the system's\nvariables and even if they do they rely just on determining the\nbetween-variables causality network. Knowing the structure of such a complex\nnetwork and even more specifically knowing the exact lagged variables that\ncontribute to the underlying process is crucial for the task of multivariate\ntime series forecasting. The latter is a rather unexplored source of\ninformation to leverage. In this direction, here a novel neural network-based\narchitecture is proposed, termed LAgged VAriable Representation NETwork\n(LAVARNET), which intrinsically estimates the importance of lagged variables\nand combines high dimensional latent representations of them to predict future\nvalues of time series. Our model is compared with other baseline and state of\nthe art neural network architectures on one simulated data set and four real\ndata sets from meteorology, music, solar activity, and finance areas. The\nproposed architecture outperforms the competitive architectures in most of the\nexperiments."", 'We propose a novel probabilistic dimensionality reduction framework that can\nnaturally integrate the generative model and the locality information of data.\nBased on this framework, we present a new model, which is able to learn a\nsmooth skeleton of embedding points in a low-dimensional space from\nhigh-dimensional noisy data. The formulation of the new model can be\nequivalently interpreted as two coupled learning problem, i.e., structure\nlearning and the learning of projection matrix. This interpretation motivates\nthe learning of the embedding points that can directly form an explicit graph\nstructure. We develop a new method to learn the embedding points that form a\nspanning tree, which is further extended to obtain a discriminative and compact\nfeature representation for clustering problems. Unlike traditional clustering\nmethods, we assume that centers of clusters should be close to each other if\nthey are connected in a learned graph, and other cluster centers should be\ndistant. This can greatly facilitate data visualization and scientific\ndiscovery in downstream analysis. Extensive experiments are performed that\ndemonstrate that the proposed framework is able to obtain discriminative\nfeature representations, and correctly recover the intrinsic structures of\nvarious real-world datasets.', 'The imputation of missing values in time series has many applications in\nhealthcare and finance. While autoregressive models are natural candidates for\ntime series imputation, score-based diffusion models have recently outperformed\nexisting counterparts including autoregressive models in many tasks such as\nimage generation and audio synthesis, and would be promising for time series\nimputation. In this paper, we propose Conditional Score-based Diffusion models\nfor Imputation (CSDI), a novel time series imputation method that utilizes\nscore-based diffusion models conditioned on observed data. Unlike existing\nscore-based approaches, the conditional diffusion model is explicitly trained\nfor imputation and can exploit correlations between observed values. On\nhealthcare and environmental data, CSDI improves by 40-65% over existing\nprobabilistic imputation methods on popular performance metrics. In addition,\ndeterministic imputation by CSDI reduces the error by 5-20% compared to the\nstate-of-the-art deterministic imputation methods. Furthermore, CSDI can also\nbe applied to time series interpolation and probabilistic forecasting, and is\ncompetitive with existing baselines. The code is available at\nhttps://github.com/ermongroup/CSDI.']","['The high-dimensional data setting, in which p >> n, is a challenging\nstatistical paradigm that appears in many real-world problems. In this setting,\nlearning a compact, low-dimensional representation of the data can\nsubstantially help distinguish signal from noise. One way to achieve this goal\nis to perform subspace learning to estimate a small set of latent features that\ncapture the majority of the variance in the original data. Most existing\nsubspace learning models, such as PCA, assume that the data can be fully\nrepresented by its embedding in one or more latent subspaces. However, in this\nwork, we argue that this assumption is not suitable for many high-dimensional\ndatasets; often only some variables can easily be projected to a\nlow-dimensional space. We propose a hybrid dimensionality reduction technique\nin which some features are mapped to a low-dimensional subspace while others\nremain in the original space. Our model leads to more accurate estimation of\nthe latent space and lower reconstruction error. We present a simple\noptimization procedure for the resulting biconvex problem and show synthetic\ndata results that demonstrate the advantages of our approach over existing\nmethods. Finally, we demonstrate the effectiveness of this method for\nextracting meaningful features from both gene expression and video background\nsubtraction datasets.', 'Identifying systems with high-dimensional inputs and outputs, such as systems\nmeasured by video streams, is a challenging problem with numerous applications\nin robotics, autonomous vehicles and medical imaging. In this paper, we propose\na novel non-linear state-space identification method starting from\nhigh-dimensional input and output data. Multiple computational and conceptual\nadvances are combined to handle the high-dimensional nature of the data. An\nencoder function, represented by a neural network, is introduced to learn a\nreconstructability map to estimate the model states from past inputs and\noutputs. This encoder function is jointly learned with the dynamics.\nFurthermore, multiple computational improvements, such as an improved\nreformulation of multiple shooting and batch optimization, are proposed to keep\nthe computational time under control when dealing with high-dimensional and\nlarge datasets. We apply the proposed method to a video stream of a simulated\nenvironment of a controllable ball in a unit box. The study shows low\nsimulation error with excellent long term prediction capability of the model\nobtained using the proposed method.', 'In the era of big data, practical applications in various domains continually\ngenerate large-scale time-series data. Among them, some data show significant\nor potential periodicity characteristics, such as meteorological and financial\ndata. It is critical to efficiently identify the potential periodic patterns\nfrom massive time-series data and provide accurate predictions. In this paper,\na Periodicity-based Parallel Time Series Prediction (PPTSP) algorithm for\nlarge-scale time-series data is proposed and implemented in the Apache Spark\ncloud computing environment. To effectively handle the massive historical\ndatasets, a Time Series Data Compression and Abstraction (TSDCA) algorithm is\npresented, which can reduce the data scale as well as accurately extracting the\ncharacteristics. Based on this, we propose a Multi-layer Time Series Periodic\nPattern Recognition (MTSPPR) algorithm using the Fourier Spectrum Analysis\n(FSA) method. In addition, a Periodicity-based Time Series Prediction (PTSP)\nalgorithm is proposed. Data in the subsequent period are predicted based on all\nprevious period models, in which a time attenuation factor is introduced to\ncontrol the impact of different periods on the prediction results. Moreover, to\nimprove the performance of the proposed algorithms, we propose a parallel\nsolution on the Apache Spark platform, using the Streaming real-time computing\nmodule. To efficiently process the large-scale time-series datasets in\ndistributed computing environments, Distributed Streams (DStreams) and\nResilient Distributed Datasets (RDDs) are used to store and calculate these\ndatasets. Extensive experimental results show that our PPTSP algorithm has\nsignificant advantages compared with other algorithms in terms of prediction\naccuracy and performance.', 'Multivariate time-series data are used in many classification and regression\npredictive tasks, and recurrent models have been widely used for such tasks.\nMost common recurrent models assume that time-series data elements are of equal\nlength and the ordered observations are recorded at regular intervals. However,\nreal-world time-series data have neither a similar length nor a same number of\nobservations. They also have missing entries, which hinders the performance of\npredictive tasks. In this paper, we approach these issues by presenting a model\nfor the combined task of imputing and predicting values for the irregularly\nobserved and varying length time-series data with missing entries. Our proposed\nmodel (Bi-GAN) uses a bidirectional recurrent network in a generative\nadversarial setting. The generator is a bidirectional recurrent network that\nreceives actual incomplete data and imputes the missing values. The\ndiscriminator attempts to discriminate between the actual and the imputed\nvalues in the output of the generator. Our model learns how to impute missing\nelements in-between (imputation) or outside of the input time steps\n(prediction), hence working as an effective any-time prediction tool for\ntime-series data. Our method has three advantages to the state-of-the-art\nmethods in the field: (a) single model can be used for both imputation and\nprediction tasks; (b) it can perform prediction task for time-series of varying\nlength with missing data; (c) it does not require to know the observation and\nprediction time window during training which provides a flexible length of\nprediction window for both long-term and short-term predictions. We evaluate\nour model on two public datasets and on another large real-world electronic\nhealth records dataset to impute and predict body mass index (BMI) values in\nchildren and show its superior performance in both settings.', 'Many areas of research are characterised by the deluge of large-scale\nhighly-dimensional time-series data. However, using the data available for\nprediction and decision making is hampered by the current lag in our ability to\nuncover and quantify true interactions that explain the outcomes.We are\ninterested in areas such as intensive care medicine, which are characterised by\ni) continuous monitoring of multivariate variables and non-uniform sampling of\ndata streams, ii) the outcomes are generally governed by interactions between a\nsmall set of rare events, iii) these interactions are not necessarily definable\nby specific values (or value ranges) of a given group of variables, but rather,\nby the deviations of these values from the normal state recorded over time, iv)\nthe need to explain the predictions made by the model. Here, while numerous\ndata mining models have been formulated for outcome prediction, they are unable\nto explain their predictions.\n  We present a model for uncovering interactions with the highest likelihood of\ngenerating the outcomes seen from highly-dimensional time series data.\nInteractions among variables are represented by a relational graph structure,\nwhich relies on qualitative abstractions to overcome non-uniform sampling and\nto capture the semantics of the interactions corresponding to the changes and\ndeviations from normality of variables of interest over time. Using the\nassumption that similar templates of small interactions are responsible for the\noutcomes (as prevalent in the medical domains), we reformulate the discovery\ntask to retrieve the most-likely templates from the data.', 'Extracting previously unknown patterns and information in time series is\ncentral to many real-world applications. In this study, we introduce a novel\napproach to modeling financial time series using a deep learning model. We use\na Long Short-Term Memory (LSTM) network equipped with the trainable initial\nhidden states. By learning to reconstruct time series, the proposed model can\nrepresent high-dimensional time series data with its parameters. An experiment\nwith the Korean stock market data showed that the model was able to capture the\nrelative similarity between a large number of stock prices in its latent space.\nBesides, the model was also able to predict the future stock trends from the\nlatent space. The proposed method can help to identify relationships among many\ntime series, and it could be applied to financial applications, such as\noptimizing the investment portfolios.', 'Time series data generation has drawn increasing attention in recent years.\nSeveral generative adversarial network (GAN) based methods have been proposed\nto tackle the problem usually with the assumption that the targeted time series\ndata are well-formatted and complete. However, real-world time series (RTS)\ndata are far away from this utopia, e.g., long sequences with variable lengths\nand informative missing data raise intractable challenges for designing\npowerful generation algorithms. In this paper, we propose a novel generative\nframework for RTS data - RTSGAN to tackle the aforementioned challenges. RTSGAN\nfirst learns an encoder-decoder module which provides a mapping between a time\nseries instance and a fixed-dimension latent vector and then learns a\ngeneration module to generate vectors in the same latent space. By combining\nthe generator and the decoder, RTSGAN is able to generate RTS which respect the\noriginal feature distributions and the temporal dynamics. To generate time\nseries with missing values, we further equip RTSGAN with an observation\nembedding layer and a decide-and-generate decoder to better utilize the\ninformative missing patterns. Experiments on the four RTS datasets show that\nthe proposed framework outperforms the previous generation methods in terms of\nsynthetic data utility for downstream classification and prediction tasks.', 'In the current context of Big Data, the nature of many forecasting problems\nhas changed from predicting isolated time series to predicting many time series\nfrom similar sources. This has opened up the opportunity to develop competitive\nglobal forecasting models that simultaneously learn from many time series. But,\nit still remains unclear when global forecasting models can outperform the\nunivariate benchmarks, especially along the dimensions of the\nhomogeneity/heterogeneity of series, the complexity of patterns in the series,\nthe complexity of forecasting models, and the lengths/number of series. Our\nstudy attempts to address this problem through investigating the effect from\nthese factors, by simulating a number of datasets that have controllable time\nseries characteristics. Specifically, we simulate time series from simple data\ngenerating processes (DGP), such as Auto Regressive (AR) and Seasonal AR, to\ncomplex DGPs, such as Chaotic Logistic Map, Self-Exciting Threshold\nAuto-Regressive, and Mackey-Glass Equations. The data heterogeneity is\nintroduced by mixing time series generated from several DGPs into a single\ndataset. The lengths and the number of series in the dataset are varied in\ndifferent scenarios. We perform experiments on these datasets using global\nforecasting models including Recurrent Neural Networks (RNN), Feed-Forward\nNeural Networks, Pooled Regression (PR) models and Light Gradient Boosting\nModels (LGBM), and compare their performance against standard statistical\nunivariate forecasting techniques. Our experiments demonstrate that when\ntrained as global forecasting models, techniques such as RNNs and LGBMs, which\nhave complex non-linear modelling capabilities, are competitive methods in\ngeneral under challenging forecasting scenarios such as series having short\nlengths, datasets with heterogeneous series and having minimal prior knowledge\nof the patterns of the series.', ""More than 80% of today's data is unstructured in nature, and these\nunstructured datasets evolve over time. A large part of these datasets are text\ndocuments generated by media outlets, scholarly articles in digital libraries,\nfindings from scientific and professional communities, and social media. Vector\nspace models were developed to analyze text data using data mining and machine\nlearning algorithms. While ample vector space models exist for text data, the\nevolutionary aspect of ever-changing text corpora is still missing in\nvector-based representations. The advent of word embeddings has enabled us to\ncreate a contextual vector space, but the embeddings fail to consider the\ntemporal aspects of the feature space successfully. This paper presents an\napproach to include temporal aspects in feature spaces. The inclusion of the\ntime aspect in the feature space provides vectors for every natural language\nelement, such as words or entities, at every timestamp. Such temporal word\nvectors allow us to track how the meaning of a word changes over time, by\nstudying the changes in its neighborhood. Moreover, a time-reflective text\nrepresentation will pave the way to a new set of text analytic abilities\ninvolving time series for text collections. In this paper, we present a\ntime-reflective vector space model for temporal text data that is able to\ncapture short and long-term changes in the meaning of words. We compare our\napproach with the limited literature on dynamic embeddings. We present\nqualitative and quantitative evaluations using the tracking of semantic\nevolution as the target application."", 'Time series forecasting is often fundamental to scientific and engineering\nproblems and enables decision making. With ever increasing data set sizes, a\ntrivial solution to scale up predictions is to assume independence between\ninteracting time series. However, modeling statistical dependencies can improve\naccuracy and enable analysis of interaction effects. Deep learning methods are\nwell suited for this problem, but multivariate models often assume a simple\nparametric distribution and do not scale to high dimensions. In this work we\nmodel the multivariate temporal dynamics of time series via an autoregressive\ndeep learning model, where the data distribution is represented by a\nconditioned normalizing flow. This combination retains the power of\nautoregressive models, such as good performance in extrapolation into the\nfuture, with the flexibility of flows as a general purpose high-dimensional\ndistribution model, while remaining computationally tractable. We show that it\nimproves over the state-of-the-art for standard metrics on many real-world data\nsets with several thousand interacting time-series.', 'The prediction of high-resolution hourly traffic volumes of a given roadway\nis essential for transportation planning. Traditionally, Automatic Traffic\nRecorders (ATR) are used to collect this hourly volume data. These large\ndatasets are time series data characterized by long-term temporal dependencies\nand missing values. Regarding the temporal dependencies, all roadways are\ncharacterized by seasonal variations that can be weekly, monthly or yearly,\ndepending on the cause of the variation. Regarding the missing data in a\ntime-series sequence, traditional time series forecasting models perform poorly\nunder the influence of seasonal variations. To address this limitation, robust,\nRecurrent Neural Network (RNN) based, multi-step ahead forecasting models are\ndeveloped for time-series in this study. The simple RNN, the Gated Recurrent\nUnit (GRU) and the Long Short-Term Memory (LSTM) units are used to develop the\nmodel and evaluate its performance. Two approaches are used to address the\nmissing value issue: masking and imputation, in conjunction with the RNN\nmodels. Six different imputation algorithms are then used to identify the best\nmodel. The analysis indicates that the LSTM model performs better than simple\nRNN and GRU models, and imputation performs better than masking to predict\nfuture traffic volume. Based on analysis using 92 ATRs, the LSTM-Median model\nis deemed the best model in all scenarios for hourly traffic volume and AADT\nprediction, with an average RMSE of 274 and MAPE of 18.91% for hourly traffic\nvolume prediction and average RMSE of 824 and MAPE of 2.10% for AADT\nprediction.', 'Forecasting with multivariate time series, which aims to predict future\nvalues given previous and current several univariate time series data, has been\nstudied for decades, with one example being ARIMA. Because it is difficult to\nmeasure the extent to which noise is mixed with informative signals within\nrapidly fluctuating financial time series data, designing a good predictive\nmodel is not a simple task. Recently, many researchers have become interested\nin recurrent neural networks and attention-based neural networks, applying them\nin financial forecasting. There have been many attempts to utilize these\nmethods for the capturing of long-term temporal dependencies and to select more\nimportant features in multivariate time series data in order to make accurate\npredictions. In this paper, we propose a new prediction framework based on deep\nneural networks and a trend filtering, which converts noisy time series data\ninto a piecewise linear fashion. We reveal that the predictive performance of\ndeep temporal neural networks improves when the training data is temporally\nprocessed by a trend filtering. To verify the effect of our framework, three\ndeep temporal neural networks, state of the art models for predictions in time\nseries finance data, are used and compared with models that contain trend\nfiltering as an input feature. Extensive experiments on real-world multivariate\ntime series data show that the proposed method is effective and significantly\nbetter than existing baseline methods.', 'We present a novel and scalable label embedding framework for large-scale\nmulti-label learning a.k.a ExMLDS (Extreme Multi-Label Learning using\nDistributional Semantics). Our approach draws inspiration from ideas rooted in\ndistributional semantics, specifically the Skip Gram Negative Sampling (SGNS)\napproach, widely used to learn word embeddings for natural language processing\ntasks. Learning such embeddings can be reduced to a certain matrix\nfactorization. Our approach is novel in that it highlights interesting\nconnections between label embedding methods used for multi-label learning and\nparagraph/document embedding methods commonly used for learning representations\nof text data. The framework can also be easily extended to incorporate\nauxiliary information such as label-label correlations; this is crucial\nespecially when there are a lot of missing labels in the training data. We\ndemonstrate the effectiveness of our approach through an extensive set of\nexperiments on a variety of benchmark datasets, and show that the proposed\nlearning methods perform favorably compared to several baselines and\nstate-of-the-art methods for large-scale multi-label learning. To facilitate\nend-to-end learning, we develop a joint learning algorithm that can learn the\nembeddings as well as a regression model that predicts these embeddings given\ninput features, via efficient gradient-based methods.', 'Feature-based time series representations have attracted substantial\nattention in a wide range of time series analysis methods. Recently, the use of\ntime series features for forecast model averaging has been an emerging research\nfocus in the forecasting community. Nonetheless, most of the existing\napproaches depend on the manual choice of an appropriate set of features.\nExploiting machine learning methods to extract features from time series\nautomatically becomes crucial in state-of-the-art time series analysis. In this\npaper, we introduce an automated approach to extract time series features based\non time series imaging. We first transform time series into recurrence plots,\nfrom which local features can be extracted using computer vision algorithms.\nThe extracted features are used for forecast model averaging. Our experiments\nshow that forecasting based on automatically extracted features, with less\nhuman intervention and a more comprehensive view of the raw time series data,\nyields highly comparable performances with the best methods in the largest\nforecasting competition dataset (M4) and outperforms the top methods in the\nTourism forecasting competition dataset.', 'Time series forecasting is an important predictive methodology which can be\napplied to a wide range of problems. Particularly, forecasting the indoor\ntemperature permits an improved utilization of the HVAC (Heating, Ventilating\nand Air Conditioning) systems in a home and thus a better energy efficiency.\nWith such purpose the paper describes how to implement an Artificial Neural\nNetwork (ANN) algorithm in a low cost system-on-chip to develop an autonomous\nintelligent wireless sensor network. The present paper uses a Wireless Sensor\nNetworks (WSN) to monitor and forecast the indoor temperature in a smart home,\nbased on low resources and cost microcontroller technology as the 8051MCU. An\non-line learning approach, based on Back-Propagation (BP) algorithm for ANNs,\nhas been developed for real-time time series learning. It performs the model\ntraining with every new data that arrive to the system, without saving enormous\nquantities of data to create a historical database as usual, i.e., without\nprevious knowledge. Consequently to validate the approach a simulation study\nthrough a Bayesian baseline model have been tested in order to compare with a\ndatabase of a real application aiming to see the performance and accuracy. The\ncore of the paper is a new algorithm, based on the BP one, which has been\ndescribed in detail, and the challenge was how to implement a computational\ndemanding algorithm in a simple architecture with very few hardware resources.', 'In many applications of finance, biology and sociology, complex systems\ninvolve entities interacting with each other. These processes have the\npeculiarity of evolving over time and of comprising latent factors, which\ninfluence the system without being explicitly measured. In this work we present\nlatent variable time-varying graphical lasso (LTGL), a method for multivariate\ntime-series graphical modelling that considers the influence of hidden or\nunmeasurable factors. The estimation of the contribution of the latent factors\nis embedded in the model which produces both sparse and low-rank components for\neach time point. In particular, the first component represents the connectivity\nstructure of observable variables of the system, while the second represents\nthe influence of hidden factors, assumed to be few with respect to the observed\nvariables. Our model includes temporal consistency on both components,\nproviding an accurate evolutionary pattern of the system. We derive a tractable\noptimisation algorithm based on alternating direction method of multipliers,\nand develop a scalable and efficient implementation which exploits proximity\noperators in closed form. LTGL is extensively validated on synthetic data,\nachieving optimal performance in terms of accuracy, structure learning and\nscalability with respect to ground truth and state-of-the-art methods for\ngraphical inference. We conclude with the application of LTGL to real case\nstudies, from biology and finance, to illustrate how our method can be\nsuccessfully employed to gain insights on multivariate time-series data.', 'Sequential and temporal data arise in many fields of research, such as\nquantitative finance, medicine, or computer vision. A novel approach for\nsequential learning, called the signature method and rooted in rough path\ntheory, is considered. Its basic principle is to represent multidimensional\npaths by a graded feature set of their iterated integrals, called the\nsignature. This approach relies critically on an embedding principle, which\nconsists in representing discretely sampled data as paths, i.e., functions from\n$[0,1]$ to $\\mathbb{R}^d$. After a survey of machine learning methodologies for\nsignatures, the influence of embeddings on prediction accuracy is investigated\nwith an in-depth study of three recent and challenging datasets. It is shown\nthat a specific embedding, called lead-lag, is systematically the strongest\nperformer across all datasets and algorithms considered. Moreover, an empirical\nstudy reveals that computing signatures over the whole path domain does not\nlead to a loss of local information. It is concluded that, with a good\nembedding, combining signatures with other simple algorithms achieves results\ncompetitive with state-of-the-art, domain-specific approaches.', 'Forecasting models that are trained across sets of many time series, known as\nGlobal Forecasting Models (GFM), have shown recently promising results in\nforecasting competitions and real-world applications, outperforming many\nstate-of-the-art univariate forecasting techniques. In most cases, GFMs are\nimplemented using deep neural networks, and in particular Recurrent Neural\nNetworks (RNN), which require a sufficient amount of time series to estimate\ntheir numerous model parameters. However, many time series databases have only\na limited number of time series. In this study, we propose a novel, data\naugmentation based forecasting framework that is capable of improving the\nbaseline accuracy of the GFM models in less data-abundant settings. We use\nthree time series augmentation techniques: GRATIS, moving block bootstrap\n(MBB), and dynamic time warping barycentric averaging (DBA) to synthetically\ngenerate a collection of time series. The knowledge acquired from these\naugmented time series is then transferred to the original dataset using two\ndifferent approaches: the pooled approach and the transfer learning approach.\nWhen building GFMs, in the pooled approach, we train a model on the augmented\ntime series alongside the original time series dataset, whereas in the transfer\nlearning approach, we adapt a pre-trained model to the new dataset. In our\nevaluation on competition and real-world time series datasets, our proposed\nvariants can significantly improve the baseline accuracy of GFM models and\noutperform state-of-the-art univariate forecasting methods.', 'Time series data can be subject to changes in the underlying process that\ngenerates them and, because of these changes, models built on old samples can\nbecome obsolete or perform poorly. In this work, we present a way to\nincorporate information about the current data distribution and its evolution\nacross time into machine learning algorithms. Our solution is based on\nefficiently maintaining statistics, particularly the mean and the variance, of\ndata features at different time resolutions. These data summarisations can be\nperformed over the input attributes, in which case they can then be fed into\nthe model as additional input features, or over latent representations learned\nby models, such as those of Recurrent Neural Networks. In classification tasks,\nthe proposed techniques can significantly outperform the prediction\ncapabilities of equivalent architectures with no feature / latent\nsummarisations. Furthermore, these modifications do not introduce notable\ncomputational and memory overhead when properly adjusted.', ""We propose an approximation algorithm for efficient correlation search in\ntime series data. In our method, we use Fourier transform and neural network to\nembed time series into a low-dimensional Euclidean space. The given space is\nlearned such that time series correlation can be effectively approximated from\nEuclidean distance between corresponding embedded vectors. Therefore, search\nfor correlated time series can be done using an index in the embedding space\nfor efficient nearest neighbor search. Our theoretical analysis illustrates\nthat our method's accuracy can be guaranteed under certain regularity\nconditions. We further conduct experiments on real-world datasets and the\nresults show that our method indeed outperforms the baseline solution. In\nparticular, for approximation of correlation, our method reduces the\napproximation loss by a half in most test cases compared to the baseline\nsolution. For top-$k$ highest correlation search, our method improves the\nprecision from 5\\% to 20\\% while the query time is similar to the baseline\napproach query time."", ""Graph analytics can lead to better quantitative understanding and control of\ncomplex networks, but traditional methods suffer from high computational cost\nand excessive memory requirements associated with the high-dimensionality and\nheterogeneous characteristics of industrial size networks. Graph embedding\ntechniques can be effective in converting high-dimensional sparse graphs into\nlow-dimensional, dense and continuous vector spaces, preserving maximally the\ngraph structure properties. Another type of emerging graph embedding employs\nGaussian distribution-based graph embedding with important uncertainty\nestimation. The main goal of graph embedding methods is to pack every node's\nproperties into a vector with a smaller dimension, hence, node similarity in\nthe original complex irregular spaces can be easily quantified in the embedded\nvector spaces using standard metrics. The generated nonlinear and highly\ninformative graph embeddings in the latent space can be conveniently used to\naddress different downstream graph analytics tasks (e.g., node classification,\nlink prediction, community detection, visualization, etc.). In this Review, we\npresent some fundamental concepts in graph analytics and graph embedding\nmethods, focusing in particular on random walk-based and neural network-based\nmethods. We also discuss the emerging deep learning-based dynamic graph\nembedding methods. We highlight the distinct advantages of graph embedding\nmethods in four diverse applications, and present implementation details and\nreferences to open-source software as well as available databases in the\nAppendix for the interested readers to start their exploration into graph\nanalytics."", 'Tasking machine learning to predict segments of a time series requires\nestimating the parameters of a ML model with input/output pairs from the time\nseries. Using the equivalence between statistical data assimilation and\nsupervised machine learning, we revisit this task. The training method for the\nmachine utilizes a precision annealing approach to identifying the global\nminimum of the action (-log[P]). In this way we are able to identify the number\nof training pairs required to produce good generalizations (predictions) for\nthe time series. We proceed from a scalar time series $s(t_n); t_n = t_0 + n\n\\Delta t$ and using methods of nonlinear time series analysis show how to\nproduce a $D_E > 1$ dimensional time delay embedding space in which the time\nseries has no false neighbors as does the observed $s(t_n)$ time series. In\nthat $D_E$-dimensional space we explore the use of feed forward multi-layer\nperceptrons as network models operating on $D_E$-dimensional input and\nproducing $D_E$-dimensional outputs.', 'High-dimensional time series are common in many domains. Since human\ncognition is not optimized to work well in high-dimensional spaces, these areas\ncould benefit from interpretable low-dimensional representations. However, most\nrepresentation learning algorithms for time series data are difficult to\ninterpret. This is due to non-intuitive mappings from data features to salient\nproperties of the representation and non-smoothness over time. To address this\nproblem, we propose a new representation learning framework building on ideas\nfrom interpretable discrete dimensionality reduction and deep generative\nmodeling. This framework allows us to learn discrete representations of time\nseries, which give rise to smooth and interpretable embeddings with superior\nclustering performance. We introduce a new way to overcome the\nnon-differentiability in discrete representation learning and present a\ngradient-based version of the traditional self-organizing map algorithm that is\nmore performant than the original. Furthermore, to allow for a probabilistic\ninterpretation of our method, we integrate a Markov model in the representation\nspace. This model uncovers the temporal transition structure, improves\nclustering performance even further and provides additional explanatory\ninsights as well as a natural representation of uncertainty. We evaluate our\nmodel in terms of clustering performance and interpretability on static\n(Fashion-)MNIST data, a time series of linearly interpolated (Fashion-)MNIST\nimages, a chaotic Lorenz attractor system with two macro states, as well as on\na challenging real world medical time series application on the eICU data set.\nOur learned representations compare favorably with competitor methods and\nfacilitate downstream tasks on the real world data.', 'Hierarchical forecasting is a key problem in many practical multivariate\nforecasting applications - the goal is to simultaneously predict a large number\nof correlated time series that are arranged in a pre-specified aggregation\nhierarchy. The main challenge is to exploit the hierarchical correlations to\nsimultaneously obtain good prediction accuracy for time series at different\nlevels of the hierarchy. In this paper, we propose a new approach for\nhierarchical forecasting which consists of two components. First, decomposing\nthe time series along a global set of basis time series and modeling\nhierarchical constraints using the coefficients of the basis decomposition. And\nsecond, using a linear autoregressive model with coefficients that vary with\ntime. Unlike past methods, our approach is scalable (inference for a specific\ntime series only needs access to its own history) while also modeling the\nhierarchical structure via (approximate) coherence constraints among the time\nseries forecasts. We experiment on several public datasets and demonstrate\nsignificantly improved overall performance on forecasts at different levels of\nthe hierarchy, compared to existing state-of-the-art hierarchical models.', 'Time-series forecasting has been an important research domain for so many\nyears. Its applications include ECG predictions, sales forecasting, weather\nconditions, even COVID-19 spread predictions. These applications have motivated\nmany researchers to figure out an optimal forecasting approach, but the\nmodeling approach also changes as the application domain changes. This work has\nfocused on reviewing different forecasting approaches for telemetry data\npredictions collected at data centers. Forecasting of telemetry data is a\ncritical feature of network and data center management products. However, there\nare multiple options of forecasting approaches that range from a simple linear\nstatistical model to high capacity deep learning architectures. In this paper,\nwe attempted to summarize and evaluate the performance of well known time\nseries forecasting techniques. We hope that this evaluation provides a\ncomprehensive summary to innovate in forecasting approaches for telemetry data.', 'For many machine learning tasks, the input data lie on a low-dimensional\nmanifold embedded in a high dimensional space and, because of this\nhigh-dimensional structure, most algorithms are inefficient. The typical\nsolution is to reduce the dimension of the input data using standard dimension\nreduction algorithms such as ISOMAP, LAPLACIAN EIGENMAPS or LLES. This\napproach, however, does not always work in practice as these algorithms require\nthat we have somewhat ideal data. Unfortunately, most data sets either have\nmissing entries or unacceptably noisy values. That is, real data are far from\nideal and we cannot use these algorithms directly. In this paper, we focus on\nthe case when we have missing data. Some techniques, such as matrix completion,\ncan be used to fill in missing data but these methods do not capture the\nnon-linear structure of the manifold. Here, we present a new algorithm\nMR-MISSING that extends these previous algorithms and can be used to compute\nlow dimensional representation on data sets with missing entries. We\ndemonstrate the effectiveness of our algorithm by running three different\nexperiments. We visually verify the effectiveness of our algorithm on synthetic\nmanifolds, we numerically compare our projections against those computed by\nfirst filling in data using nlPCA and mDRUR on the MNIST data set, and we also\nshow that we can do classification on MNIST with missing data. We also provide\na theoretical guarantee for MR-MISSING under some simplifying assumptions.', 'Time-series data with missing values are commonly encountered in many fields,\nsuch as healthcare, meteorology, and robotics. The imputation aims to fill the\nmissing values with valid values. Most imputation methods trained the models\nimplicitly because missing values have no ground truth. In this paper, we\npropose Random Drop Imputation with Self-training (RDIS), a novel training\nmethod for time-series data imputation models. In RDIS, we generate extra\nmissing values by applying a random drop on the observed values in incomplete\ndata. We can explicitly train the imputation models by filling in the randomly\ndropped values. In addition, we adopt self-training with pseudo values to\nexploit the original missing values. To improve the quality of pseudo values,\nwe set the threshold and filter them by calculating the entropy. To verify the\neffectiveness of RDIS on the time series imputation, we test RDIS to various\nimputation models and achieve competitive results on two real-world datasets.', 'The growing popularity of wearable sensors has generated large quantities of\ntemporal physiological and activity data. Ability to analyze this data offers\nnew opportunities for real-time health monitoring and forecasting. However,\ntemporal physiological data presents many analytic challenges: the data is\nnoisy, contains many missing values, and each series has a different length.\nMost methods proposed for time series analysis and classification do not handle\ndatasets with these characteristics nor do they offer interpretability and\nexplainability, a critical requirement in the health domain. We propose an\nunsupervised method for learning representations of time series based on common\npatterns identified within them. The patterns are, interpretable, variable in\nlength, and extracted using Byte Pair Encoding compression technique. In this\nway the method can capture both long-term and short-term dependencies present\nin the data. We show that this method applies to both univariate and\nmultivariate time series and beats state-of-the-art approaches on a real world\ndataset collected from wearable sensors.', 'Conditional probabilistic graphical models provide a powerful framework for\nstructured regression in spatio-temporal datasets with complex correlation\npatterns. However, in real-life applications a large fraction of observations\nis often missing, which can severely limit the representational power of these\nmodels. In this paper we propose a Marginalized Gaussian Conditional Random\nFields (m-GCRF) structured regression model for dealing with missing labels in\npartially observed temporal attributed graphs. This method is aimed at learning\nwith both labeled and unlabeled parts and effectively predicting future values\nin a graph. The method is even capable of learning from nodes for which the\nresponse variable is never observed in history, which poses problems for many\nstate-of-the-art models that can handle missing data. The proposed model is\ncharacterized for various missingness mechanisms on 500 synthetic graphs. The\nbenefits of the new method are also demonstrated on a challenging application\nfor predicting precipitation based on partial observations of climate variables\nin a temporal graph that spans the entire continental US. We also show that the\nmethod can be useful for optimizing the costs of data collection in climate\napplications via active reduction of the number of weather stations to\nconsider. In experiments on these real-world and synthetic datasets we show\nthat the proposed model is consistently more accurate than alternative\nsemi-supervised structured models, as well as models that either use imputation\nto deal with missing values or simply ignore them altogether.', 'In this paper we propose a data augmentation method for time series with\nirregular sampling, Time-Conditional Generative Adversarial Network (T-CGAN).\nOur approach is based on Conditional Generative Adversarial Networks (CGAN),\nwhere the generative step is implemented by a deconvolutional NN and the\ndiscriminative step by a convolutional NN. Both the generator and the\ndiscriminator are conditioned on the sampling timestamps, to learn the hidden\nrelationship between data and timestamps, and consequently to generate new time\nseries. We evaluate our model with synthetic and real-world datasets. For the\nsynthetic data, we compare the performance of a classifier trained with\nT-CGAN-generated data, against the performance of the same classifier trained\non the original data. Results show that classifiers trained on T-CGAN-generated\ndata perform the same as classifiers trained on real data, even with very short\ntime series and small training sets. For the real world datasets, we compare\nour method with other techniques of data augmentation for time series, such as\ntime slicing and time warping, over a classification problem with unbalanced\ndatasets. Results show that our method always outperforms the other approaches,\nboth in case of regularly sampled and irregularly sampled time series. We\nachieve particularly good performance in case with a small training set and\nshort, noisy, irregularly-sampled time series.', 'Time series forecasting is a growing domain with diverse applications.\nHowever, changes of the system behavior over time due to internal or external\ninfluences are challenging. Therefore, predictions of a previously learned\nfore-casting model might not be useful anymore. In this paper, we present\nEVent-triggered Augmented Refitting of Gaussian Process Regression for Seasonal\nData (EVARS-GPR), a novel online algorithm that is able to handle sudden shifts\nin the target variable scale of seasonal data. For this purpose, EVARS-GPR\ncom-bines online change point detection with a refitting of the prediction\nmodel using data augmentation for samples prior to a change point. Our\nexperiments on sim-ulated data show that EVARS-GPR is applicable for a wide\nrange of output scale changes. EVARS-GPR has on average a 20.8 % lower RMSE on\ndifferent real-world datasets compared to methods with a similar computational\nresource con-sumption. Furthermore, we show that our algorithm leads to a\nsix-fold reduction of the averaged runtime in relation to all comparison\npartners with a periodical refitting strategy. In summary, we present a\ncomputationally efficient online fore-casting algorithm for seasonal time\nseries with changes of the target variable scale and demonstrate its\nfunctionality on simulated as well as real-world data. All code is publicly\navailable on GitHub: https://github.com/grimmlab/evars-gpr.', 'High-dimensional probability density estimation for inference suffers from\nthe ""curse of dimensionality"". For many physical inference problems, the full\nposterior distribution is unwieldy and seldom used in practice. Instead, we\npropose direct estimation of lower-dimensional marginal distributions,\nbypassing high-dimensional density estimation or high-dimensional Markov chain\nMonte Carlo (MCMC) sampling. By evaluating the two-dimensional marginal\nposteriors we can unveil the full-dimensional parameter covariance structure.\nWe additionally propose constructing a simple hierarchy of fast neural\nregression models, called Moment Networks, that compute increasing moments of\nany desired lower-dimensional marginal posterior density; these reproduce exact\nresults from analytic posteriors and those obtained from Masked Autoregressive\nFlows. We demonstrate marginal posterior density estimation using\nhigh-dimensional LIGO-like gravitational wave time series and describe\napplications for problems of fundamental cosmology.', 'Despite the popularisation of machine learning models, more often than not,\nthey still operate as black boxes with no insight into what is happening inside\nthe model. There exist a few methods that allow to visualise and explain why a\nmodel has made a certain prediction. Those methods, however, allow\nvisualisation of the link between the input and output of the model without\npresenting how the model learns to represent the data used to train the model\nas whole. In this paper, a method that addresses that issue is proposed, with a\nfocus on visualising multi-dimensional time-series data. Experiments on a\nhigh-frequency stock market dataset show that the method provides fast and\ndiscernible visualisations. Large datasets can be visualised quickly and on one\nplot, which makes it easy for a user to compare the learned representations of\nthe data. The developed method successfully combines known techniques to\nprovide an insight into the inner workings of time-series classification\nmodels.', 'Temporal anomaly detection looks for irregularities over space-time.\nUnsupervised temporal models employed thus far typically work on sequences of\nfeature vectors, and much less on temporal multiway data. We focus our\ninvestigation on two-way data, in which a data matrix is observed at each time\nstep. Leveraging recent advances in matrix-native recurrent neural networks, we\ninvestigated strategies for data arrangement and unsupervised training for\ntemporal multiway anomaly detection. These include compressing-decompressing,\nencoding-predicting, and temporal data differencing. We conducted a\ncomprehensive suite of experiments to evaluate model behaviors under various\nsettings on synthetic data, moving digits, and ECG recordings. We found\ninteresting phenomena not previously reported. These include the capacity of\nthe compact matrix LSTM to compress noisy data near perfectly, making the\nstrategy of compressing-decompressing data ill-suited for anomaly detection\nunder the noise. Also, long sequence of vectors can be addressed directly by\nmatrix models that allow very long context and multiple step prediction.\nOverall, the encoding-predicting strategy works very well for the matrix LSTMs\nin the conducted experiments, thanks to its compactness and better fit to the\ndata dynamics.', 'Sparse regression such as the Lasso has achieved great success in handling\nhigh-dimensional data. However, one of the biggest practical problems is that\nhigh-dimensional data often contain large amounts of missing values. Convex\nConditioned Lasso (CoCoLasso) has been proposed for dealing with\nhigh-dimensional data with missing values, but it performs poorly when there\nare many missing values, so that the high missing rate problem has not been\nresolved. In this paper, we propose a novel Lasso-type regression method for\nhigh-dimensional data with high missing rates. We effectively incorporate mean\nimputed covariance, overcoming its inherent estimation bias. The result is an\noptimally weighted modification of CoCoLasso according to missing ratios. We\ntheoretically and experimentally show that our proposed method is highly\neffective even when there are many missing values.', 'A cornerstone of the worldwide transition to smart grids are smart meters.\nSmart meters typically collect and provide energy time series that are vital\nfor various applications, such as grid simulations, fault-detection, load\nforecasting, load analysis, and load management. Unfortunately, these time\nseries are often characterized by missing values that must be handled before\nthe data can be used. A common approach to handle missing values in time series\nis imputation. However, existing imputation methods are designed for power time\nseries and do not take into account the total energy of gaps, resulting in\njumps or constant shifts when imputing energy time series. In order to overcome\nthese issues, the present paper introduces the new Copy-Paste Imputation (CPI)\nmethod for energy time series. The CPI method copies data blocks with similar\nproperties and pastes them into gaps of the time series while preserving the\ntotal energy of each gap. The new method is evaluated on a real-world dataset\nthat contains six shares of artificially inserted missing values between 1 and\n30%. It outperforms by far the three benchmark imputation methods selected for\ncomparison. The comparison furthermore shows that the CPI method uses matching\npatterns and preserves the total energy of each gap while requiring only a\nmoderate run-time.', 'In the context of time series forecasting, it is a common practice to\nevaluate multiple methods and choose one of these methods or an ensemble for\nproducing the best forecasts. However, choosing among different ensembles over\nmultiple methods remains a challenging task that undergoes a combinatorial\nexplosion as the number of methods increases. In the context of demand\nforecasting or revenue forecasting, this challenge is further exacerbated by a\nlarge number of time series as well as limited historical data points available\ndue to changing business context. Although deep learning forecasting methods\naim to simultaneously forecast large collections of time series, they become\nchallenging to apply in such scenarios due to the limited history available and\nmight not yield desirable results. We propose a framework for forecasting short\nhigh-dimensional time series data by combining low-rank temporal matrix\nfactorization and optimal model selection on latent time series using\ncross-validation. We demonstrate that forecasting the latent factors leads to\nsignificant performance gains as compared to directly applying different\nuni-variate models on time series. Performance has been validated on a\ntruncated version of the M4 monthly dataset which contains time series data\nfrom multiple domains showing the general applicability of the method.\nMoreover, it is amenable to incorporating the analyst view of the future owing\nto the low number of latent factors which is usually impractical when applying\nforecasting methods directly to high dimensional datasets.', ""Accurate demand forecasts can help on-line retail organizations better plan\ntheir supply-chain processes. The challenge, however, is the large number of\nassociative factors that result in large, non-stationary shifts in demand,\nwhich traditional time series and regression approaches fail to model. In this\npaper, we propose a Neural Network architecture called AR-MDN, that\nsimultaneously models associative factors, time-series trends and the variance\nin the demand. We first identify several causal features and use a combination\nof feature embeddings, MLP and LSTM to represent them. We then model the output\ndensity as a learned mixture of Gaussian distributions. The AR-MDN can be\ntrained end-to-end without the need for additional supervision. We experiment\non a dataset of an year's worth of data over tens-of-thousands of products from\nFlipkart. The proposed architecture yields a significant improvement in\nforecasting accuracy when compared with existing alternatives."", 'This paper presents a time series forecasting framework which combines\nstandard forecasting methods and a machine learning model. The inputs to the\nmachine learning model are not lagged values or regular time series features,\nbut instead forecasts produced by standard methods. The machine learning model\ncan be either a convolutional neural network model or a recurrent neural\nnetwork model. The intuition behind this approach is that forecasts of a time\nseries are themselves good features characterizing the series, especially when\nthe modelling purpose is forecasting. It can also be viewed as a weighted\nensemble method. Tested on the M4 competition dataset, this approach\noutperforms all submissions for quarterly series, and is more accurate than all\nbut the winning algorithm for monthly series.', 'Conditional modeling x \\to y is a central problem in machine learning. A\nsubstantial research effort is devoted to such modeling when x is high\ndimensional. We consider, instead, the case of a high dimensional y, where x is\neither low dimensional or high dimensional. Our approach is based on selecting\na small subset y_L of the dimensions of y, and proceed by modeling (i) x \\to\ny_L and (ii) y_L \\to y. Composing these two models, we obtain a conditional\nmodel x \\to y that possesses convenient statistical properties. Multi-label\nclassification and multivariate regression experiments on several datasets show\nthat this model outperforms the one vs. all approach as well as several\nsophisticated multiple output prediction methods.', 'A networked time series (NETS) is a family of time series on a given graph,\none for each node. It has found a wide range of applications from intelligent\ntransportation, environment monitoring to mobile network management. An\nimportant task in such applications is to predict the future values of a NETS\nbased on its historical values and the underlying graph. Most existing methods\nrequire complete data for training. However, in real-world scenarios, it is not\nuncommon to have missing data due to sensor malfunction, incomplete sensing\ncoverage, etc. In this paper, we study the problem of NETS prediction with\nincomplete data. We propose NETS-ImpGAN, a novel deep learning framework that\ncan be trained on incomplete data with missing values in both history and\nfuture. Furthermore, we propose novel Graph Temporal Attention Networks by\nincorporating the attention mechanism to capture both inter-time series\ncorrelations and temporal correlations. We conduct extensive experiments on\nthree real-world datasets under different missing patterns and missing rates.\nThe experimental results show that NETS-ImpGAN outperforms existing methods\nexcept when data exhibit very low variance, in which case NETS-ImpGAN still\nachieves competitive performance.', 'Finding rare information hidden in a huge amount of data from the Internet is\na necessary but complex issue. Many researchers have studied this issue and\nhave found effective methods to detect anomaly data in low dimensional space.\nHowever, as the dimension increases, most of these existing methods perform\npoorly in detecting outliers because of ""high dimensional curse"". Even though\nsome approaches aim to solve this problem in high dimensional space, they can\nonly detect some anomaly data appearing in low dimensional space and cannot\ndetect all of anomaly data which appear differently in high dimensional space.\nTo cope with this problem, we propose a new k-nearest section-based method\n(k-NS) in a section-based space. Our proposed approach not only detects\noutliers in low dimensional space with section-density ratio but also detects\noutliers in high dimensional space with the ratio of k-nearest section against\naverage value. After taking a series of experiments with the dimension from 10\nto 10000, the experiment results show that our proposed method achieves 100%\nprecision and 100% recall result in the case of extremely high dimensional\nspace, and better improvement in low dimensional space compared to our\npreviously proposed method.', 'We present the collaborative Kalman filter (CKF), a dynamic model for\ncollaborative filtering and related factorization models. Using the matrix\nfactorization approach to collaborative filtering, the CKF accounts for time\nevolution by modeling each low-dimensional latent embedding as a\nmultidimensional Brownian motion. Each observation is a random variable whose\ndistribution is parameterized by the dot product of the relevant Brownian\nmotions at that moment in time. This is naturally interpreted as a Kalman\nfilter with multiple interacting state space vectors. We also present a method\nfor learning a dynamically evolving drift parameter for each location by\nmodeling it as a geometric Brownian motion. We handle posterior intractability\nvia a mean-field variational approximation, which also preserves tractability\nfor downstream calculations in a manner similar to the Kalman filter. We\nevaluate the model on several large datasets, providing quantitative evaluation\non the 10 million Movielens and 100 million Netflix datasets and qualitative\nevaluation on a set of 39 million stock returns divided across roughly 6,500\ncompanies from the years 1962-2014.', 'We present a representation learning framework for financial time series\nforecasting. One challenge of using deep learning models for finance\nforecasting is the shortage of available training data when using small\ndatasets. Direct trend classification using deep neural networks trained on\nsmall datasets is susceptible to the overfitting problem. In this paper, we\npropose to first learn compact representations from time series data, then use\nthe learned representations to train a simpler model for predicting time series\nmovements. We consider a class-conditioned latent variable model. We train an\nencoder network to maximize the mutual information between the latent variables\nand the trend information conditioned on the encoded observed variables. We\nshow that conditional mutual information maximization can be approximated by a\ncontrastive loss. Then, the problem is transformed into a classification task\nof determining whether two encoded representations are sampled from the same\nclass or not. This is equivalent to performing pairwise comparisons of the\ntraining datapoints, and thus, improves the generalization ability of the\nencoder network. We use deep autoregressive models as our encoder to capture\nlong-term dependencies of the sequence data. Empirical experiments indicate\nthat our proposed method has the potential to advance state-of-the-art\nperformance.', 'In this paper, we propose a machine learning approach for forecasting\nhierarchical time series. When dealing with hierarchical time series, apart\nfrom generating accurate forecasts, one needs to select a suitable method for\nproducing reconciled forecasts. Forecast reconciliation is the process of\nadjusting forecasts to make them coherent across the hierarchy. In literature,\ncoherence is often enforced by using a post-processing technique on the base\nforecasts produced by suitable time series forecasting methods. On the\ncontrary, our idea is to use a deep neural network to directly produce accurate\nand reconciled forecasts. We exploit the ability of a deep neural network to\nextract information capturing the structure of the hierarchy. We impose the\nreconciliation at training time by minimizing a customized loss function. In\nmany practical applications, besides time series data, hierarchical time series\ninclude explanatory variables that are beneficial for increasing the\nforecasting accuracy. Exploiting this further information, our approach links\nthe relationship between time series features extracted at any level of the\nhierarchy and the explanatory variables into an end-to-end neural network\nproviding accurate and reconciled point forecasts. The effectiveness of the\napproach is validated on three real-world datasets, where our method\noutperforms state-of-the-art competitors in hierarchical forecasting.', 'Time series analysis is a field of data science which is interested in\nanalyzing sequences of numerical values ordered in time. Time series are\nparticularly interesting because they allow us to visualize and understand the\nevolution of a process over time. Their analysis can reveal trends,\nrelationships and similarities across the data. There exists numerous fields\ncontaining data in the form of time series: health care (electrocardiogram,\nblood sugar, etc.), activity recognition, remote sensing, finance (stock market\nprice), industry (sensors), etc. Time series classification consists of\nconstructing algorithms dedicated to automatically label time series data. The\nsequential aspect of time series data requires the development of algorithms\nthat are able to harness this temporal property, thus making the existing\noff-the-shelf machine learning models for traditional tabular data suboptimal\nfor solving the underlying task. In this context, deep learning has emerged in\nrecent years as one of the most effective methods for tackling the supervised\nclassification task, particularly in the field of computer vision. The main\nobjective of this thesis was to study and develop deep neural networks\nspecifically constructed for the classification of time series data. We thus\ncarried out the first large scale experimental study allowing us to compare the\nexisting deep methods and to position them compared other non-deep learning\nbased state-of-the-art methods. Subsequently, we made numerous contributions in\nthis area, notably in the context of transfer learning, data augmentation,\nensembling and adversarial attacks. Finally, we have also proposed a novel\narchitecture, based on the famous Inception network (Google), which ranks among\nthe most efficient to date.', 'Missing value imputation is crucial for real-world data science workflows.\nImputation is harder in the online setting, as it requires the imputation\nmethod itself to be able to evolve over time. For practical applications,\nimputation algorithms should produce imputations that match the true data\ndistribution, handle data of mixed types, including ordinal, boolean, and\ncontinuous variables, and scale to large datasets. In this work we develop a\nnew online imputation algorithm for mixed data using the Gaussian copula. The\nonline Gaussian copula model meets all the desiderata: its imputations match\nthe data distribution even for mixed data, improve over its offline counterpart\non the accuracy when the streaming data has a changing distribution, and on the\nspeed (up to an order of magnitude) especially on large scale datasets. By\nfitting the copula model to online data, we also provide a new method to detect\nchange points in the multivariate dependence structure with missing values.\nExperimental results on synthetic and real world data validate the performance\nof the proposed methods.', 'Time series prediction with missing values is an important problem of time\nseries analysis since complete data is usually hard to obtain in many\nreal-world applications. To model the generation of time series, autoregressive\n(AR) model is a basic and widely used one, which assumes that each observation\nin the time series is a noisy linear combination of some previous observations\nalong with a constant shift. To tackle the problem of prediction with missing\nvalues, a number of methods were proposed based on various data models. For\nreal application scenarios, how do these methods perform over different types\nof time series with different levels of data missing remains to be\ninvestigated. In this paper, we focus on online methods for AR-model-based time\nseries prediction with missing values. We adapted five mainstream methods to\nfit in such a scenario. We make detailed discussion on each of them by\nintroducing their core ideas about how to estimate the AR coefficients and\ntheir different strategies to deal with missing values. We also present\nalgorithmic implementations for better understanding. In order to\ncomprehensively evaluate these methods and do the comparison, we conduct\nexperiments with various configurations of relative parameters over both\nsynthetic and real data. From the experimental results, we derived several\nnoteworthy conclusions and shows that imputation is a simple but reliable\nstrategy to handle missing values in online prediction tasks.', ""Visualizing high dimensional data by projecting them into two or three\ndimensional space is one of the most effective ways to intuitively understand\nthe data's underlying characteristics, for example their class neighborhood\nstructure. While data visualization in low dimensional space can be efficient\nfor revealing the data's underlying characteristics, classifying a new sample\nin the reduced-dimensional space is not always beneficial because of the loss\nof information in expressing the data. It is possible to classify the data in\nthe high dimensional space, while visualizing them in the low dimensional\nspace, but in this case, the visualization is often meaningless because it\nfails to illustrate the underlying characteristics that are crucial for the\nclassification process.\n  In this paper, the performance-preserving property of the previously proposed\nRestricted Radial Basis Function Network in reducing the dimension of labeled\ndata is explained. Here, it is argued through empirical experiments that the\ninternal representation of the Restricted Radial Basis Function Network, which\nduring the supervised learning process organizes a visualizable two dimensional\nmap, does not only preserve the topographical structure of high dimensional\ndata but also captures their class neighborhood structures that are important\nfor classifying them. Hence, unlike many of the existing dimension reduction\nmethods, the Restricted Radial Basis Function Network offers two dimensional\nvisualization that is strongly correlated with the classification process."", 'We present an approach to analyze $C^1(\\mathbb{R}^m)$ functions that\naddresses limitations present in the Active Subspaces (AS) method of\nConstantine et al.(2015; 2014). Under appropriate hypotheses, our Active\nManifolds (AM) method identifies a 1-D curve in the domain (the active\nmanifold) on which nearly all values of the unknown function are attained, and\nwhich can be exploited for approximation or analysis, especially when $m$ is\nlarge (high-dimensional input space). We provide theorems justifying our AM\ntechnique and an algorithm permitting functional approximation and sensitivity\nanalysis. Using accessible, low-dimensional functions as initial examples, we\nshow AM reduces approximation error by an order of magnitude compared to AS, at\nthe expense of more computation. Following this, we revisit the sensitivity\nanalysis by Glaws et al. (2017), who apply AS to analyze a magnetohydrodynamic\npower generator model, and compare the performance of AM on the same data. Our\nanalysis provides detailed information not captured by AS, exhibiting the\ninfluence of each parameter individually along an active manifold. Overall, AM\nrepresents a novel technique for analyzing functional models with benefits\nincluding: reducing $m$-dimensional analysis to a 1-D analogue, permitting more\naccurate regression than AS (at more computational expense), enabling more\ninformative sensitivity analysis, and granting accessible visualizations(2-D\nplots) of parameter sensitivity along the AM.', 'Accurate time-series forecasting is vital for numerous areas of application\nsuch as transportation, energy, finance, economics, etc. However, while modern\ntechniques are able to explore large sets of temporal data to build forecasting\nmodels, they typically neglect valuable information that is often available\nunder the form of unstructured text. Although this data is in a radically\ndifferent format, it often contains contextual explanations for many of the\npatterns that are observed in the temporal data. In this paper, we propose two\ndeep learning architectures that leverage word embeddings, convolutional layers\nand attention mechanisms for combining text information with time-series data.\nWe apply these approaches for the problem of taxi demand forecasting in event\nareas. Using publicly available taxi data from New York, we empirically show\nthat by fusing these two complementary cross-modal sources of information, the\nproposed models are able to significantly reduce the error in the forecasts.', 'Time-series classification is one of the most frequently performed tasks in\nindustrial data science, and one of the most widely used data representation in\nthe industrial setting is tabular representation. In this work, we propose a\nnovel scalable architecture for learning representations from tabular\ntime-series data and subsequently performing downstream tasks such as\ntime-series classification. The representation learning framework is\nend-to-end, akin to bidirectional encoder representations from transformers\n(BERT) in language modeling, however, we introduce novel masking technique\nsuitable for pretraining of time-series data. Additionally, we also use\none-dimensional convolutions augmented with transformers and explore their\neffectiveness, since the time-series datasets lend themselves naturally for\none-dimensional convolutions. We also propose a novel timestamp embedding\ntechnique, which helps in handling both periodic cycles at different time\ngranularity levels, and aperiodic trends present in the time-series data. Our\nproposed model is end-to-end and can handle both categorical and continuous\nvalued inputs, and does not require any quantization or encoding of continuous\nfeatures.', 'Interactive response time is important in analytical pipelines for users to\nexplore a sufficient number of possibilities and make informed business\ndecisions. We consider a forecasting pipeline with large volumes of\nhigh-dimensional time series data. Real-time forecasting can be conducted in\ntwo steps. First, we specify the part of data to be focused on and the measure\nto be predicted by slicing, dicing, and aggregating the data. Second, a\nforecasting model is trained on the aggregated results to predict the trend of\nthe specified measure. While there are a number of forecasting models\navailable, the first step is the performance bottleneck. A natural idea is to\nutilize sampling to obtain approximate aggregations in real time as the input\nto train the forecasting model. Our scalable real-time forecasting system\nFlashP (Flash Prediction) is built based on this idea, with two major\nchallenges to be resolved in this paper: first, we need to figure out how\napproximate aggregations affect the fitting of forecasting models, and\nforecasting results; and second, accordingly, what sampling algorithms we\nshould use to obtain these approximate aggregations and how large the samples\nare. We introduce a new sampling scheme, called GSW sampling, and analyze error\nbounds for estimating aggregations using GSW samples. We introduce how to\nconstruct compact GSW samples with the existence of multiple measures to be\nanalyzed. We conduct experiments to evaluate our solution and compare it with\nalternatives on real data.', 'In this paper, we present a new approach to time series forecasting. Time\nseries data are prevalent in many scientific and engineering disciplines. Time\nseries forecasting is a crucial task in modeling time series data, and is an\nimportant area of machine learning. In this work we developed a novel method\nthat employs Transformer-based machine learning models to forecast time series\ndata. This approach works by leveraging self-attention mechanisms to learn\ncomplex patterns and dynamics from time series data. Moreover, it is a generic\nframework and can be applied to univariate and multivariate time series data,\nas well as time series embeddings. Using influenza-like illness (ILI)\nforecasting as a case study, we show that the forecasting results produced by\nour approach are favorably comparable to the state-of-the-art.', 'Kernel density estimation is a simple and effective method that lies at the\nheart of many important machine learning applications. Unfortunately, kernel\nmethods scale poorly for large, high dimensional datasets. Approximate kernel\ndensity estimation has a prohibitively high memory and computation cost,\nespecially in the streaming setting. Recent sampling algorithms for high\ndimensional densities can reduce the computation cost but cannot operate\nonline, while streaming algorithms cannot handle high dimensional datasets due\nto the curse of dimensionality. We propose RACE, an efficient sketching\nalgorithm for kernel density estimation on high-dimensional streaming data.\nRACE compresses a set of N high dimensional vectors into a small array of\ninteger counters. This array is sufficient to estimate the kernel density for a\nlarge class of kernels. Our sketch is practical to implement and comes with\nstrong theoretical guarantees. We evaluate our method on real-world\nhigh-dimensional datasets and show that our sketch achieves 10x better\ncompression compared to competing methods.', 'Time series forecasting is essential for agents to make decisions.\nTraditional approaches rely on statistical methods to forecast given past\nnumeric values. In practice, end-users often rely on visualizations such as\ncharts and plots to reason about their forecasts. Inspired by practitioners, we\nre-imagine the topic by creating a novel framework to produce visual forecasts,\nsimilar to the way humans intuitively do. In this work, we leverage advances in\ndeep learning to extend the field of time series forecasting to a visual\nsetting. We capture input data as an image and train a model to produce the\nsubsequent image. This approach results in predicting distributions as opposed\nto pointwise values. We examine various synthetic and real datasets with\ndiverse degrees of complexity. Our experiments show that visual forecasting is\neffective for cyclic data but somewhat less for irregular data such as stock\nprice. Importantly, when using image-based evaluation metrics, we find the\nproposed visual forecasting method to outperform various numerical baselines,\nincluding ARIMA and a numerical variation of our method. We demonstrate the\nbenefits of incorporating vision-based approaches in forecasting tasks -- both\nfor the quality of the forecasts produced, as well as the metrics that can be\nused to evaluate them.', 'Multivariate time-series data are frequently observed in critical care\nsettings and are typically characterized by sparsity (missing information) and\nirregular time intervals. Existing approaches for learning representations in\nthis domain handle these challenges by either aggregation or imputation of\nvalues, which in-turn suppresses the fine-grained information and adds\nundesirable noise/overhead into the machine learning model. To tackle this\nproblem, we propose a Self-supervised Transformer for Time-Series (STraTS)\nmodel which overcomes these pitfalls by treating time-series as a set of\nobservation triplets instead of using the standard dense matrix representation.\nIt employs a novel Continuous Value Embedding technique to encode continuous\ntime and variable values without the need for discretization. It is composed of\na Transformer component with multi-head attention layers which enable it to\nlearn contextual triplet embeddings while avoiding the problems of recurrence\nand vanishing gradients that occur in recurrent architectures. In addition, to\ntackle the problem of limited availability of labeled data (which is typically\nobserved in many healthcare applications), STraTS utilizes self-supervision by\nleveraging unlabeled data to learn better representations by using time-series\nforecasting as an auxiliary proxy task. Experiments on real-world multivariate\nclinical time-series benchmark datasets demonstrate that STraTS has better\nprediction performance than state-of-the-art methods for mortality prediction,\nespecially when labeled data is limited. Finally, we also present an\ninterpretable version of STraTS which can identify important measurements in\nthe time-series data. Our data preprocessing and model implementation codes are\navailable at https://github.com/sindhura97/STraTS.', 'This paper describes a novel approach to change-point detection when the\nobserved high-dimensional data may have missing elements. The performance of\nclassical methods for change-point detection typically scales poorly with the\ndimensionality of the data, so that a large number of observations are\ncollected after the true change-point before it can be reliably detected.\nFurthermore, missing components in the observed data handicap conventional\napproaches. The proposed method addresses these challenges by modeling the\ndynamic distribution underlying the data as lying close to a time-varying\nlow-dimensional submanifold embedded within the ambient observation space.\nSpecifically, streaming data is used to track a submanifold approximation,\nmeasure deviations from this approximation, and calculate a series of\nstatistics of the deviations for detecting when the underlying manifold has\nchanged in a sharp or unexpected manner. The approach described in this paper\nleverages several recent results in the field of high-dimensional data\nanalysis, including subspace tracking with missing data, multiscale analysis\ntechniques for point clouds, online optimization, and change-point detection\nperformance analysis. Simulations and experiments highlight the robustness and\nefficacy of the proposed approach in detecting an abrupt change in an otherwise\nslowly varying low-dimensional manifold.', ""We extract and use player position time-series data, tagged along with the\naction types, to build a competent model for representing team tactics\nbehavioral patterns and use this representation to predict the outcome of\narbitrary movements. We provide a framework for the useful encoding of short\ntactics and space occupations in a more extended sequence of movements or\ntactical plans. We investigate game segments during a match in which the team\nin possession of the ball regularly attempts to reach a position where they can\ntake a shot at goal for a single game. A carefully designed and efficient\nkernel is employed using a triangular fuzzy membership function to create\nmultiple time series for players' potential of presence at different court\nregions. Unsupervised learning is then used for time series using triplet loss\nand deep neural networks with exponentially dilated causal convolutions for the\nderived multivariate time series. This works key contribution lies in its\napproach to model how short scenes contribute to other longer ones and how\nplayers occupies and creates new spaces in-game court. We discuss the\neffectiveness of the proposed approach for prediction and recognition tasks on\nthe professional basketball SportVU dataset for the 2015-16 half-season. The\nproposed system demonstrates descent functionality even with relatively small\ndata."", 'Many businesses and industries nowadays rely on large quantities of time\nseries data making time series forecasting an important research area. Global\nforecasting models that are trained across sets of time series have shown a\nhuge potential in providing accurate forecasts compared with the traditional\nunivariate forecasting models that work on isolated series. However, there are\ncurrently no comprehensive time series archives for forecasting that contain\ndatasets of time series from similar sources available for the research\ncommunity to evaluate the performance of new global forecasting algorithms over\na wide variety of datasets. In this paper, we present such a comprehensive time\nseries forecasting archive containing 20 publicly available time series\ndatasets from varied domains, with different characteristics in terms of\nfrequency, series lengths, and inclusion of missing values. We also\ncharacterise the datasets, and identify similarities and differences among\nthem, by conducting a feature analysis. Furthermore, we present the performance\nof a set of standard baseline forecasting methods over all datasets across\neight error metrics, for the benefit of researchers using the archive to\nbenchmark their forecasting algorithms.', 'This paper presents a machine learning approach to model the electric\nconsumption of electric vehicles at macroscopic level, i.e., in the absence of\na speed profile, while preserving microscopic level accuracy. For this work, we\nleveraged a high-performance, agent-based transportation tool to model trips\nthat occur in the Greater Chicago region under various scenario changes, along\nwith physics-based modeling and simulation tools to provide high-fidelity\nenergy consumption values. The generated results constitute a very large\ndataset of vehicle-route energy outcomes that capture variability in vehicle\nand routing setting, and in which high-fidelity time series of vehicle speed\ndynamics is masked. We show that although all internal dynamics that affect\nenergy consumption are masked, it is possible to learn aggregate-level energy\nconsumption values quite accurately with a deep learning approach. When\nlarge-scale data is available, and with carefully tailored feature engineering,\na well-designed model can overcome and retrieve latent information. This model\nhas been deployed and integrated within POLARIS Transportation System\nSimulation Tool to support real-time behavioral transportation models for\nindividual charging decision-making, and rerouting of electric vehicles.', 'This work contributes to the development of neural forecasting models with\nnovel randomization-based learning methods. These methods improve the fitting\nabilities of the neural model, in comparison to the standard method, by\ngenerating network parameters in accordance with the data and target function\nfeatures. A pattern-based representation of time series makes the proposed\napproach useful for forecasting time series with multiple seasonality. In the\nsimulation study, we evaluate the performance of the proposed models and find\nthat they can compete in terms of forecasting accuracy with fully-trained\nnetworks. Extremely fast and easy training, simple architecture, ease of\nimplementation, high accuracy as well as dealing with nonstationarity and\nmultiple seasonality in time series make the proposed model very attractive for\na wide range of complex time series forecasting problems.', 'This paper addresses the problem of multi-step time series forecasting for\nnon-stationary signals that can present sudden changes. Current\nstate-of-the-art deep learning forecasting methods, often trained with variants\nof the MSE, lack the ability to provide sharp predictions in deterministic and\nprobabilistic contexts. To handle these challenges, we propose to incorporate\nshape and temporal criteria in the training objective of deep models. We define\nshape and temporal similarities and dissimilarities, based on a smooth\nrelaxation of Dynamic Time Warping (DTW) and Temporal Distortion Index (TDI),\nthat enable to build differentiable loss functions and positive semi-definite\n(PSD) kernels. With these tools, we introduce DILATE (DIstortion Loss including\nshApe and TimE), a new objective for deterministic forecasting, that explicitly\nincorporates two terms supporting precise shape and temporal change detection.\nFor probabilistic forecasting, we introduce STRIPE++ (Shape and Time diverRsIty\nin Probabilistic forEcasting), a framework for providing a set of sharp and\ndiverse forecasts, where the structured shape and time diversity is enforced\nwith a determinantal point process (DPP) diversity loss. Extensive experiments\nand ablations studies on synthetic and real-world datasets confirm the benefits\nof leveraging shape and time features in time series forecasting.', 'Decomposing complex time series into trend, seasonality, and remainder\ncomponents is an important task to facilitate time series anomaly detection and\nforecasting. Although numerous methods have been proposed, there are still many\ntime series characteristics exhibiting in real-world data which are not\naddressed properly, including 1) ability to handle seasonality fluctuation and\nshift, and abrupt change in trend and reminder; 2) robustness on data with\nanomalies; 3) applicability on time series with long seasonality period. In the\npaper, we propose a novel and generic time series decomposition algorithm to\naddress these challenges. Specifically, we extract the trend component robustly\nby solving a regression problem using the least absolute deviations loss with\nsparse regularization. Based on the extracted trend, we apply the the non-local\nseasonal filtering to extract the seasonality component. This process is\nrepeated until accurate decomposition is obtained. Experiments on different\nsynthetic and real-world time series datasets demonstrate that our method\noutperforms existing solutions.', 'Additive Manufacturing presents a great application area for Machine Learning\nbecause of the vast volume of data generated and the potential to mine this\ndata to control outcomes. In this paper we present preliminary work on\nclassifying infrared time-series data representing melt-pool temperature in a\nmetal 3D printing process. Our ultimate objective is to use this data to\npredict process outcomes (e.g. hardness, porosity, surface roughness). In the\nwork presented here we simply show that there is a signal in this data that can\nbe used for the classification of different components and stages of the AM\nprocess. In line with other Machine Learning research on time-series\nclassification we use k-Nearest Neighbour classifiers. The results we present\nsuggests that Dynamic Time Warping is an effective distance measure compared\nwith alternatives for 3D printing data of this type.', 'While artificial intelligence (AI)-based decision-making systems are\nincreasingly popular, significant concerns on the potential discrimination\nduring the AI decision-making process have been observed. For example, the\ndistribution of predictions is usually biased and dependents on the sensitive\nattributes (e.g., gender and ethnicity). Numerous approaches have therefore\nbeen proposed to develop decision-making systems that are\ndiscrimination-conscious by-design, which are typically batch-based and require\nthe simultaneous availability of all the training data for model learning.\nHowever, in the real-world, the data streams usually come on the fly which\nrequires the model to process each input data once ""on arrival"" and without the\nneed for storage and reprocessing. In addition, the data streams might also\nevolve over time, which further requires the model to be able to simultaneously\nadapt to non-stationary data distributions and time-evolving bias patterns,\nwith an effective and robust trade-off between accuracy and fairness. In this\npaper, we propose a novel framework of online decision tree with fairness in\nthe data stream with possible distribution drifting. Specifically, first, we\npropose two novel fairness splitting criteria that encode the data as well as\npossible, while simultaneously removing dependence on the sensitive attributes,\nand further adapts to non-stationary distribution with fine-grained control\nwhen needed. Second, we propose two fairness decision tree online growth\nalgorithms that fulfills different online fair decision-making requirements.\nOur experiments show that our algorithms are able to deal with discrimination\nin massive and non-stationary streaming environments, with a better trade-off\nbetween fairness and predictive performance.', 'Time series are widely used as signals in many classification/regression\ntasks. It is ubiquitous that time series contains many missing values. Given\nmultiple correlated time series data, how to fill in missing values and to\npredict their class labels? Existing imputation methods often impose strong\nassumptions of the underlying data generating process, such as linear dynamics\nin the state space. In this paper, we propose BRITS, a novel method based on\nrecurrent neural networks for missing value imputation in time series data. Our\nproposed method directly learns the missing values in a bidirectional recurrent\ndynamical system, without any specific assumption. The imputed values are\ntreated as variables of RNN graph and can be effectively updated during the\nbackpropagation.BRITS has three advantages: (a) it can handle multiple\ncorrelated missing values in time series; (b) it generalizes to time series\nwith nonlinear dynamics underlying; (c) it provides a data-driven imputation\nprocedure and applies to general settings with missing data.We evaluate our\nmodel on three real-world datasets, including an air quality dataset, a\nhealth-care data, and a localization data for human activity. Experiments show\nthat our model outperforms the state-of-the-art methods in both imputation and\nclassification/regression accuracies.', 'Time-Spatial data plays a crucial role for different fields such as traffic\nmanagement. These data can be collected via devices such as surveillance\nsensors or tracking systems. However, how to efficiently an- alyze and\nvisualize these data to capture essential embedded pattern information is\nbecoming a big challenge today. Classic visualization ap- proaches focus on\nrevealing 2D and 3D spatial information and modeling statistical test. Those\nmethods would easily fail when data become mas- sive. Recent attempts concern\non how to simply cluster data and perform prediction with time-oriented\ninformation. However, those approaches could still be further enhanced as they\nalso have limitations for han- dling massive clusters and labels. In this\npaper, we propose a visualiza- tion methodology for mobility data using\nartificial neural net techniques. This method aggregates three main parts that\nare Back-end Data Model, Neural Net Algorithm including clustering method\nSelf-Organizing Map (SOM) and prediction approach Recurrent Neural Net (RNN)\nfor ex- tracting the features and lastly a solid front-end that displays the\nresults to users with an interactive system. SOM is able to cluster the\nvisiting patterns and detect the abnormal pattern. RNN can perform the predic-\ntion for time series analysis using its dynamic architecture. Furthermore, an\ninteractive system will enable user to interpret the result with graph- ics,\nanimation and 3D model for a close-loop feedback. This method can be\nparticularly applied in two tasks that Commercial-based Promotion and abnormal\ntraffic patterns detection.', 'A key obstacle in automated analytics and meta-learning is the inability to\nrecognize when different datasets contain measurements of the same variable.\nBecause provided attribute labels are often uninformative in practice, this\ntask may be more robustly addressed by leveraging the data values themselves\nrather than just relying on their arbitrarily selected variable names. Here, we\npresent a computationally efficient method to identify high-confidence variable\nmatches between a given set of data values and a large repository of previously\nencountered datasets. Our approach enjoys numerous advantages over\ndistributional similarity based techniques because we leverage learned vector\nembeddings of datasets which adaptively account for natural forms of data\nvariation encountered in practice. Based on the neural architecture of deep\nsets, our embeddings can be computed for both numeric and string data. In\ndataset search and schema matching tasks, our methods outperform standard\nstatistical techniques and we find that the learned embeddings generalize well\nto new data sources.', 'Embeddings are one of the fundamental building blocks for data analysis\ntasks. Embeddings are already essential tools for large language models and\nimage analysis, and their use is being extended to many other research domains.\nThe generation of these distributed representations is often a data- and\ncomputation-expensive process; yet the holistic analysis and adjustment of them\nafter they have been created is still a developing area. In this paper, we\nfirst propose a very general quantitatively measure for the presence of\nfeatures in the embedding data based on if it can be learned. We then devise a\nmethod to remove or alleviate undesired features in the embedding while\nretaining the essential structure of the data. We use a Domain Adversarial\nNetwork (DAN) to generate a non-affine transformation, but we add constraints\nto ensure the essential structure of the embedding is preserved. Our empirical\nresults demonstrate that the proposed algorithm significantly outperforms the\nstate-of-art unsupervised algorithm on several data sets, including novel\napplications from the industry.', 'Time series are series of values ordered by time. This kind of data can be\nfound in many real world settings. Classifying time series is a difficult task\nand an active area of research. This paper investigates the use of transfer\nlearning in Deep Neural Networks and a 2D representation of time series known\nas Recurrence Plots. In order to utilize the research done in the area of image\nclassification, where Deep Neural Networks have achieved very good results, we\nuse a Residual Neural Networks architecture known as ResNet. As preprocessing\nof time series is a major part of every time series classification pipeline,\nthe method proposed simplifies this step and requires only few parameters. For\nthe first time we propose a method for multi time series classification:\nTraining a single network to classify all datasets in the archive with one\nnetwork. We are among the first to evaluate the method on the latest 2018\nrelease of the UCR archive, a well established time series classification\nbenchmarking dataset.', 'Encoder-decoder recurrent neural network models (RNN Seq2Seq) have achieved\ngreat success in ubiquitous areas of computation and applications. It was shown\nto be successful in modeling data with both temporal and spatial dependencies\nfor translation or prediction tasks. In this study, we propose an embedding\napproach to visualize and interpret the representation of data by these models.\nFurthermore, we show that the embedding is an effective method for unsupervised\nlearning and can be utilized to estimate the optimality of model training. In\nparticular, we demonstrate that embedding space projections of the decoder\nstates of RNN Seq2Seq model trained on sequences prediction are organized in\nclusters capturing similarities and differences in the dynamics of these\nsequences. Such performance corresponds to an unsupervised clustering of any\nspatio-temporal features and can be employed for time-dependent problems such\nas temporal segmentation, clustering of dynamic activity, self-supervised\nclassification, action recognition, failure prediction, etc. We test and\ndemonstrate the application of the embedding methodology to time-sequences of\n3D human body poses. We show that the methodology provides a high-quality\nunsupervised categorization of movements.', 'Temporal knowledge graph (TKG) reasoning is a crucial task that has gained\nincreasing research interest in recent years. Most existing methods focus on\nreasoning at past timestamps to complete the missing facts, and there are only\na few works of reasoning on known TKGs to forecast future facts. Compared with\nthe completion task, the forecasting task is more difficult that faces two main\nchallenges: (1) how to effectively model the time information to handle future\ntimestamps? (2) how to make inductive inference to handle previously unseen\nentities that emerge over time? To address these challenges, we propose the\nfirst reinforcement learning method for forecasting. Specifically, the agent\ntravels on historical knowledge graph snapshots to search for the answer. Our\nmethod defines a relative time encoding function to capture the timespan\ninformation, and we design a novel time-shaped reward based on Dirichlet\ndistribution to guide the model learning. Furthermore, we propose a novel\nrepresentation method for unseen entities to improve the inductive inference\nability of the model. We evaluate our method for this link prediction task at\nfuture timestamps. Extensive experiments on four benchmark datasets demonstrate\nsubstantial performance improvement meanwhile with higher explainability, less\ncalculation, and fewer parameters when compared with existing state-of-the-art\nmethods.', 'The ability to decompose scenes in terms of abstract building blocks is\ncrucial for general intelligence. Where those basic building blocks share\nmeaningful properties, interactions and other regularities across scenes, such\ndecompositions can simplify reasoning and facilitate imagination of novel\nscenarios. In particular, representing perceptual observations in terms of\nentities should improve data efficiency and transfer performance on a wide\nrange of tasks. Thus we need models capable of discovering useful\ndecompositions of scenes by identifying units with such regularities and\nrepresenting them in a common format. To address this problem, we have\ndeveloped the Multi-Object Network (MONet). In this model, a VAE is trained\nend-to-end together with a recurrent attention network -- in a purely\nunsupervised manner -- to provide attention masks around, and reconstructions\nof, regions of images. We show that this model is capable of learning to\ndecompose and represent challenging 3D scenes into semantically meaningful\ncomponents, such as objects and background elements.', 'Time-series forecasting is important for many applications. Forecasting\nmodels are usually trained using time-series data in a specific target task.\nHowever, sufficient data in the target task might be unavailable, which leads\nto performance degradation. In this paper, we propose a few-shot learning\nmethod that forecasts a future value of a time-series in a target task given a\nfew time-series in the target task. Our model is trained using time-series data\nin multiple training tasks that are different from target tasks. Our model uses\na few time-series to build a forecasting function based on a recurrent neural\nnetwork with an attention mechanism. With the attention mechanism, we can\nretrieve useful patterns in a small number of time-series for the current\nsituation. Our model is trained by minimizing an expected test error of\nforecasting next timestep values. We demonstrate the effectiveness of the\nproposed method using 90 time-series datasets.', 'The problem of missing values in multivariable time series is a key challenge\nin many applications such as clinical data mining. Although many imputation\nmethods show their effectiveness in many applications, few of them are designed\nto accommodate clinical multivariable time series. In this work, we propose a\nmultiple imputation model that capture both cross-sectional information and\ntemporal correlations. We integrate Gaussian processes with mixture models and\nintroduce individualized mixing weights to handle the variance of predictive\nconfidence of Gaussian process models. The proposed model is compared with\nseveral state-of-the-art imputation algorithms on both real-world and synthetic\ndatasets. Experiments show that our best model can provide more accurate\nimputation than the benchmarks on all of our datasets.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', ""Data analysis has become a necessity in the modern era of cricket. Everything\nfrom effective team management to match win predictions use some form of\nanalytics. Meaningful data representations are necessary for efficient analysis\nof data. In this study we investigate the use of adaptive (learnable)\nembeddings to represent inter-related features (such as players, teams, etc).\nThe data used for this study is collected from a classical T20 tournament IPL\n(Indian Premier League). To naturally facilitate the learning of meaningful\nrepresentations of features for accurate data analysis, we formulate a deep\nrepresentation learning framework which jointly learns a custom set of\nembeddings (which represents our features of interest) through the minimization\nof a contrastive loss. We base our objective on a set of classes obtained as a\nresult of hierarchical clustering on the overall run rate of an innings. It's\nbeen assessed that the framework ensures greater generality in the obtained\nembeddings, on top of which a task based analysis of overall run rate\nprediction was done to show the reliability of the framework."", 'Pointwise matches between two time series are of great importance in time\nseries analysis, and dynamic time warping (DTW) is known to provide generally\nreasonable matches. There are situations where time series alignment should be\ninvariant to scaling and offset in amplitude or where local regions of the\nconsidered time series should be strongly reflected in pointwise matches. Two\ndifferent variants of DTW, affine DTW (ADTW) and regional DTW (RDTW), are\nproposed to handle scaling and offset in amplitude and provide regional\nemphasis respectively. Furthermore, ADTW and RDTW can be combined in two\ndifferent ways to generate alignments that incorporate advantages from both\nmethods, where the affine model can be applied either globally to the entire\ntime series or locally to each region. The proposed alignment methods\noutperform DTW on specific simulated datasets, and one-nearest-neighbor\nclassifiers using their associated difference measures are competitive with the\ndifference measures associated with state-of-the-art alignment methods on real\ndatasets.', 'We study the settings for which deep contextual embeddings (e.g., BERT) give\nlarge improvements in performance relative to classic pretrained embeddings\n(e.g., GloVe), and an even simpler baseline---random word embeddings---focusing\non the impact of the training set size and the linguistic properties of the\ntask. Surprisingly, we find that both of these simpler baselines can match\ncontextual embeddings on industry-scale data, and often perform within 5 to 10%\naccuracy (absolute) on benchmark tasks. Furthermore, we identify properties of\ndata for which contextual embeddings give particularly large gains: language\ncontaining complex structure, ambiguous word usage, and words unseen in\ntraining.', 'This paper discusses the prediction of hierarchical time series, where each\nupper-level time series is calculated by summing appropriate lower-level time\nseries. Forecasts for such hierarchical time series should be coherent, meaning\nthat the forecast for an upper-level time series equals the sum of forecasts\nfor corresponding lower-level time series. Previous methods for making coherent\nforecasts consist of two phases: first computing base (incoherent) forecasts\nand then reconciling those forecasts based on their inherent hierarchical\nstructure. With the aim of improving time series predictions, we propose a\nstructured regularization method for completing both phases simultaneously. The\nproposed method is based on a prediction model for bottom-level time series and\nuses a structured regularization term to incorporate upper-level forecasts into\nthe prediction model. We also develop a backpropagation algorithm specialized\nfor application of our method to artificial neural networks for time series\nprediction. Experimental results using synthetic and real-world datasets\ndemonstrate the superiority of our method in terms of prediction accuracy and\ncomputational efficiency.', 'Leveled Homomorphic Encryption (LHE) offers a potential solution that could\nallow sectors with sensitive data to utilize the cloud and securely deploy\ntheir models for remote inference with Deep Neural Networks (DNN). However,\nthis application faces several obstacles due to the limitations of LHE. One of\nthe main problems is the incompatibility of commonly used nonlinear functions\nin DNN with the operations supported by LHE, i.e. addition and multiplication.\nAs common in LHE approaches, we train a model with a nonlinear function, and\nreplace it with a low-degree polynomial approximation at inference time on\nprivate data. While this typically leads to approximation errors and loss in\nprediction accuracy, we propose a method that reduces this loss to small values\nor eliminates it entirely, depending on simple hyper-parameters. This is\nachieved by the introduction of a novel and elegantly simple Min-Max\nnormalization scheme, which scales inputs to nonlinear functions into ranges\nwith low approximation error. While being intuitive in its concept and trivial\nto implement, we empirically show that it offers a stable and effective\napproximation solution to nonlinear functions in DNN. In return, this can\nenable deeper networks with LHE, and facilitate the development of security-\nand privacy-aware analytics applications.', ""For autonomous vehicles (AVs) to behave appropriately on roads populated by\nhuman-driven vehicles, they must be able to reason about the uncertain\nintentions and decisions of other drivers from rich perceptual information.\nTowards these capabilities, we present a probabilistic forecasting model of\nfuture interactions between a variable number of agents. We perform both\nstandard forecasting and the novel task of conditional forecasting, which\nreasons about how all agents will likely respond to the goal of a controlled\nagent (here, the AV). We train models on real and simulated data to forecast\nvehicle trajectories given past positions and LIDAR. Our evaluation shows that\nour model is substantially more accurate in multi-agent driving scenarios\ncompared to existing state-of-the-art. Beyond its general ability to perform\nconditional forecasting queries, we show that our model's predictions of all\nagents improve when conditioned on knowledge of the AV's goal, further\nillustrating its capability to model agent interactions."", 'Additive Manufacturing (AM) is a manufacturing paradigm that builds\nthree-dimensional objects from a computer-aided design model by successively\nadding material layer by layer. AM has become very popular in the past decade\ndue to its utility for fast prototyping such as 3D printing as well as\nmanufacturing functional parts with complex geometries using processes such as\nlaser metal deposition that would be difficult to create using traditional\nmachining. As the process for creating an intricate part for an expensive metal\nsuch as Titanium is prohibitive with respect to cost, computational models are\nused to simulate the behavior of AM processes before the experimental run.\nHowever, as the simulations are computationally costly and time-consuming for\npredicting multiscale multi-physics phenomena in AM, physics-informed\ndata-driven machine-learning systems for predicting the behavior of AM\nprocesses are immensely beneficial. Such models accelerate not only multiscale\nsimulation tools but also empower real-time control systems using in-situ data.\nIn this paper, we design and develop essential components of a scientific\nframework for developing a data-driven model-based real-time control system.\nFinite element methods are employed for solving time-dependent heat equations\nand developing the database. The proposed framework uses extremely randomized\ntrees - an ensemble of bagged decision trees as the regression algorithm\niteratively using temperatures of prior voxels and laser information as inputs\nto predict temperatures of subsequent voxels. The models achieve mean absolute\npercentage errors below 1% for predicting temperature profiles for AM\nprocesses. The code is made available for the research community at\nhttps://github.com/paularindam/ml-iter-additive.', ""Understanding customer behavior is fundamental for many use-cases in\nindustry, especially in accelerated growth areas such as fin-tech and\ne-commerce. Structured data are often expensive, time-consuming and inadequate\nto analyze and study complex customer behaviors. In this paper, we propose a\nmulti-graph embedding approach for creating a non-linear representation of\ncustomers in order to have a better knowledge of their characteristics without\nhaving any prior information about their financial status or their interests.\nBy applying the current method we are able to predict users' future behavior\nwith a reasonably high accuracy only by having the information of their\nfriendship network. Potential applications include recommendation systems and\ncredit risk forecasting."", 'The key contribution of this paper is to propose a classification into two\ndimensions of the load forecasting studies to decide which forecasting tools to\nuse in which case. This classification aims to provide a synthetic view of the\nrelevant forecasting techniques and methodologies by forecasting problem. In\naddition, the key principles of the main techniques and methodologies used are\nsummarized along with the reviews of these papers.\n  The classification process relies on two couples of parameters that define a\nforecasting problem. Each article is classified with key information about the\ndataset used and the forecasting tools implemented: the forecasting techniques\n(probabilistic or deterministic) and methodologies, the data cleansing\ntechniques, and the error metrics.\n  The process to select the articles reviewed in this paper was conducted into\ntwo steps. First, a set of load forecasting studies was built based on relevant\nload forecasting reviews and forecasting competitions. The second step\nconsisted in selecting the most relevant studies of this set based on the\nfollowing criteria: the quality of the description of the forecasting\ntechniques and methodologies implemented, the description of the results, and\nthe contributions.\n  This paper can be read in two passes. The first one by identifying the\nforecasting problem of interest to select the corresponding class into one of\nthe four classification tables. Each one references all the articles classified\nacross a forecasting horizon. They provide a synthetic view of the forecasting\ntools used by articles addressing similar forecasting problems. Then, a second\nlevel composed of four Tables summarizes key information about the forecasting\ntools and the results of these studies. The second pass consists in reading the\nkey principles of the main techniques and methodologies of interest and the\nreviews of the articles.', 'This paper introduces the concept of travel behavior embeddings, a method for\nre-representing discrete variables that are typically used in travel demand\nmodeling, such as mode, trip purpose, education level, family type or\noccupation. This re-representation process essentially maps those variables\ninto a latent space called the \\emph{embedding space}. The benefit of this is\nthat such spaces allow for richer nuances than the typical transformations used\nin categorical variables (e.g. dummy encoding, contrasted encoding, principal\ncomponents analysis). While the usage of latent variable representations is not\nnew per se in travel demand modeling, the idea presented here brings several\ninnovations: it is an entirely data driven algorithm; it is informative and\nconsistent, since the latent space can be visualized and interpreted based on\ndistances between different categories; it preserves interpretability of\ncoefficients, despite being based on Neural Network principles; and it is\ntransferrable, in that embeddings learned from one dataset can be reused for\nother ones, as long as travel behavior keeps consistent between the datasets.\n  The idea is strongly inspired on natural language processing techniques,\nnamely the word2vec algorithm. Such algorithm is behind recent developments\nsuch as in automatic translation or next word prediction. Our method is\ndemonstrated using a model choice model, and shows improvements of up to 60\\%\nwith respect to initial likelihood, and up to 20% with respect to likelihood of\nthe corresponding traditional model (i.e. using dummy variables) in\nout-of-sample evaluation. We provide a new Python package, called PyTre (PYthon\nTRavel Embeddings), that others can straightforwardly use to replicate our\nresults or improve their own models. Our experiments are themselves based on an\nopen dataset (swissmetro).', 'Multi-dimensional data exploration is a classic research topic in\nvisualization. Most existing approaches are designed for identifying record\npatterns in dimensional space or subspace. In this paper, we propose a visual\nanalytics approach to exploring subset patterns. The core of the approach is a\nsubset embedding network (SEN) that represents a group of subsets as\nuniformly-formatted embeddings. We implement the SEN as multiple subnets with\nseparate loss functions. The design enables to handle arbitrary subsets and\ncapture the similarity of subsets on single features, thus achieving accurate\npattern exploration, which in most cases is searching for subsets having\nsimilar values on few features. Moreover, each subnet is a fully-connected\nneural network with one hidden layer. The simple structure brings high training\nefficiency. We integrate the SEN into a visualization system that achieves a\n3-step workflow. Specifically, analysts (1) partition the given dataset into\nsubsets, (2) select portions in a projected latent space created using the SEN,\nand (3) determine the existence of patterns within selected subsets. Generally,\nthe system combines visualizations, interactions, automatic methods, and\nquantitative measures to balance the exploration flexibility and operation\nefficiency, and improve the interpretability and faithfulness of the identified\npatterns. Case studies and quantitative experiments on multiple open datasets\ndemonstrate the general applicability and effectiveness of our approach.', ""The reconstruction of phase spaces is an essential step to analyze time\nseries according to Dynamical System concepts. A regression performed on such\nspaces unveils the relationships among system states from which we can derive\ntheir generating rules, that is, the most probable set of functions responsible\nfor generating observations along time. In this sense, most approaches rely on\nTakens' embedding theorem to unfold the phase space, which requires the\nembedding dimension and the time delay. Moreover, although several methods have\nbeen proposed to empirically estimate those parameters, they still face\nlimitations due to their lack of consistency and robustness, which has\nmotivated this paper. As an alternative, we here propose an artificial neural\nnetwork with a forgetting mechanism to implicitly learn the phase spaces\nproperties, whatever they are. Such network trains on forecasting errors and,\nafter converging, its architecture is used to estimate the embedding\nparameters. Experimental results confirm that our approach is either as\ncompetitive as or better than most state-of-the-art strategies while revealing\nthe temporal relationship among time-series observations."", 'Low-dimensional embeddings, from classical spectral embeddings to modern\nneural-net-inspired methods, are a cornerstone in the modeling and analysis of\ncomplex networks. Recent work by Seshadhri et al. (PNAS 2020) suggests that\nsuch embeddings cannot capture local structure arising in complex networks. In\nparticular, they show that any network generated from a natural low-dimensional\nmodel cannot be both sparse and have high triangle density (high clustering\ncoefficient), two hallmark properties of many real-world networks.\n  In this work we show that the results of Seshadhri et al. are intimately\nconnected to the model they use rather than the low-dimensional structure of\ncomplex networks. Specifically, we prove that a minor relaxation of their model\ncan generate sparse graphs with high triangle density. Surprisingly, we show\nthat this same model leads to exact low-dimensional factorizations of many\nreal-world networks. We give a simple algorithm based on logistic principal\ncomponent analysis (LPCA) that succeeds in finding such exact embeddings.\nFinally, we perform a large number of experiments that verify the ability of\nvery low-dimensional embeddings to capture local structure in real-world\nnetworks.', 'Growing amounts of online user data motivate the need for automated\nprocessing techniques. In case of user ratings, one interesting option is to\nuse neural networks for learning to predict ratings given an item and a user.\nWhile training for prediction, such an approach at the same time learns to map\neach user to a vector, a so-called user embedding. Such embeddings can for\nexample be valuable for estimating user similarity. However, there are various\nways how item and user information can be combined in neural networks, and it\nis unclear how the way of combining affects the resulting embeddings. In this\npaper, we run an experiment on movie ratings data, where we analyze the effect\non embedding quality caused by several fusion strategies in neural networks.\nFor evaluating embedding quality, we propose a novel measure, Pair-Distance\nCorrelation, which quantifies the condition that similar users should have\nsimilar embedding vectors. We find that the fusion strategy affects results in\nterms of both prediction performance and embedding quality. Surprisingly, we\nfind that prediction performance not necessarily reflects embedding quality.\nThis suggests that if embeddings are of interest, the common tendency to select\nmodels based on their prediction ability should be reconsidered.', ""How will my face look when I get older? Or, for a more challenging question:\nHow will my brain look when I get older? To answer this question one must\ndevise (and learn from data) a multivariate auto-regressive function which\ngiven an image and a desired target age generates an output image. While\ncollecting data for faces may be easier, collecting longitudinal brain data is\nnot trivial. We propose a deep learning-based method that learns to simulate\nsubject-specific brain ageing trajectories without relying on longitudinal\ndata. Our method synthesises images conditioned on two factors: age (a\ncontinuous variable), and status of Alzheimer's Disease (AD, an ordinal\nvariable). With an adversarial formulation we learn the joint distribution of\nbrain appearance, age and AD status, and define reconstruction losses to\naddress the challenging problem of preserving subject identity. We compare with\nseveral benchmarks using two widely used datasets. We evaluate the quality and\nrealism of synthesised images using ground-truth longitudinal data and a\npre-trained age predictor. We show that, despite the use of cross-sectional\ndata, our model learns patterns of gray matter atrophy in the middle temporal\ngyrus in patients with AD. To demonstrate generalisation ability, we train on\none dataset and evaluate predictions on the other. In conclusion, our model\nshows an ability to separate age, disease influence and anatomy using only 2D\ncross-sectional data that should be useful in large studies into\nneurodegenerative disease, that aim to combine several data sources. To\nfacilitate such future studies by the community at large our code is made\navailable at https://github.com/xiat0616/BrainAgeing."", 'Quantum computers are expected to surpass the computational capabilities of\nclassical computers during this decade, and achieve disruptive impact on\nnumerous industry sectors, particularly finance. In fact, finance is estimated\nto be the first industry sector to benefit from Quantum Computing not only in\nthe medium and long terms, but even in the short term. This review paper\npresents the state of the art of quantum algorithms for financial applications,\nwith particular focus to those use cases that can be solved via Machine\nLearning.', 'The lack of a large amount of training data has always been the constraining\nfactor in solving a lot of problems in machine learning, making One Shot\nLearning one of the most intriguing ideas in machine learning. It aims to learn\ninformation about object categories from one, or only a few training examples.\nThis process of learning in deep learning is usually accomplished by proper\nobjective function, i.e; loss function and embeddings extraction i.e;\narchitecture. In this paper, we discussed about metrics based deep learning\narchitectures for one shot learning such as Siamese neural networks and present\na method to improve on their accuracy using Kafnets (kernel-based\nnon-parametric activation functions for neural networks) by learning proper\nembeddings with relatively less number of epochs. Using kernel activation\nfunctions, we are able to achieve strong results which exceed those of ReLU\nbased deep learning models in terms of embeddings structure, loss convergence,\nand accuracy.', 'We present an instance-based nearest neighbor approach to entity linking. In\ncontrast to most prior entity retrieval systems which represent each entity\nwith a single vector, we build a contextualized mention-encoder that learns to\nplace similar mentions of the same entity closer in vector space than mentions\nof different entities. This approach allows all mentions of an entity to serve\nas ""class prototypes"" as inference involves retrieving from the full set of\nlabeled entity mentions in the training set and applying the nearest mention\nneighbor\'s entity label. Our model is trained on a large multilingual corpus of\nmention pairs derived from Wikipedia hyperlinks, and performs nearest neighbor\ninference on an index of 700 million mentions. It is simpler to train, gives\nmore interpretable predictions, and outperforms all other systems on two\nmultilingual entity linking benchmarks.', 'We address the problem of solving mixed random linear equations. We have\nunlabeled observations coming from multiple linear regressions, and each\nobservation corresponds to exactly one of the regression models. The goal is to\nlearn the linear regressors from the observations. Classically, Alternating\nMinimization (AM) (which is a variant of Expectation Maximization (EM)) is used\nto solve this problem. AM iteratively alternates between the estimation of\nlabels and solving the regression problems with the estimated labels.\nEmpirically, it is observed that, for a large variety of non-convex problems\nincluding mixed linear regression, AM converges at a much faster rate compared\nto gradient based algorithms. However, the existing theory suggests similar\nrate of convergence for AM and gradient based methods, failing to capture this\nempirical behavior. In this paper, we close this gap between theory and\npractice for the special case of a mixture of $2$ linear regressions. We show\nthat, provided initialized properly, AM enjoys a \\emph{super-linear} rate of\nconvergence in certain parameter regimes. To the best of our knowledge, this is\nthe first work that theoretically establishes such rate for AM. Hence, if we\nwant to recover the unknown regressors upto an error (in $\\ell_2$ norm) of\n$\\epsilon$, AM only takes $\\mathcal{O}(\\log \\log (1/\\epsilon))$ iterations.\nFurthermore, we compare AM with a gradient based heuristic algorithm\nempirically and show that AM dominates in iteration complexity as well as\nwall-clock time.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : In the field of robotic manipulation, I have consistently grappled with the challenge of deploying robots in unfamiliar and evolving environments. These environments could range from outdoor settings to busy traffic roads or even battlefields. My objective is to enable my robot to operate efficiently in these new surroundings, acquiring and applying new knowledge to better adapt to the environment. My current approach is based on the premise that a diverse environment offers a variety of rewards. Consequently, my robotic agent should be capable of learning to optimize a diverse set of rewards. Moreover, the agent should be able to swiftly adapt when the rewards (or objectives) change, especially when the environment remains constant. Ultimately, this system should empower robots to learn from and adapt to changes in their environment and tasks. This would involve incrementally constructing models based on their own experiences, thereby enhancing their overall performance in real-world scenarios.","['One of the great promises of robot learning systems is that they will be able\nto learn from their mistakes and continuously adapt to ever-changing\nenvironments. Despite this potential, most of the robot learning systems today\nare deployed as a fixed policy and they are not being adapted after their\ndeployment. Can we efficiently adapt previously learned behaviors to new\nenvironments, objects and percepts in the real world? In this paper, we present\na method and empirical evidence towards a robot learning framework that\nfacilitates continuous adaption. In particular, we demonstrate how to adapt\nvision-based robotic manipulation policies to new variations by fine-tuning via\noff-policy reinforcement learning, including changes in background, object\nshape and appearance, lighting conditions, and robot morphology. Further, this\nadaptation uses less than 0.2% of the data necessary to learn the task from\nscratch. We find that our approach of adapting pre-trained policies leads to\nsubstantial performance gains over the course of fine-tuning, and that\npre-training via RL is essential: training from scratch or adapting from\nsupervised ImageNet features are both unsuccessful with such small amounts of\ndata. We also find that these positive results hold in a limited continual\nlearning setting, in which we repeatedly fine-tune a single lineage of policies\nusing data from a succession of new tasks. Our empirical conclusions are\nconsistently supported by experiments on simulated manipulation tasks, and by\n52 unique fine-tuning experiments on a real robotic grasping system pre-trained\non 580,000 grasps.', 'AI systems that learn through reward feedback about the actions they take are\nincreasingly deployed in domains that have significant impact on our daily\nlife. However, in many cases the online rewards should not be the only guiding\ncriteria, as there are additional constraints and/or priorities imposed by\nregulations, values, preferences, or ethical principles. We detail a novel\nonline agent that learns a set of behavioral constraints by observation and\nuses these learned constraints as a guide when making decisions in an online\nsetting while still being reactive to reward feedback. To define this agent, we\npropose to adopt a novel extension to the classical contextual multi-armed\nbandit setting and we provide a new algorithm called Behavior Constrained\nThompson Sampling (BCTS) that allows for online learning while obeying\nexogenous constraints. Our agent learns a constrained policy that implements\nthe observed behavioral constraints demonstrated by a teacher agent, and then\nuses this constrained policy to guide the reward-based online exploration and\nexploitation. We characterize the upper bound on the expected regret of the\ncontextual bandit algorithm that underlies our agent and provide a case study\nwith real world data in two application domains. Our experiments show that the\ndesigned agent is able to act within the set of behavior constraints without\nsignificantly degrading its overall reward performance.', 'Learning algorithms are enabling robots to solve increasingly challenging\nreal-world tasks. These approaches often rely on demonstrations and reproduce\nthe behavior shown. Unexpected changes in the environment may require using\ndifferent behaviors to achieve the same effect, for instance to reach and grasp\nan object in changing clutter. An emerging paradigm addressing this robustness\nissue is to learn a diverse set of successful behaviors for a given task, from\nwhich a robot can select the most suitable policy when faced with a new\nenvironment. In this paper, we explore a novel realization of this vision by\nlearning a generative model over policies. Rather than learning a single\npolicy, or a small fixed repertoire, our generative model for policies\ncompactly encodes an unbounded number of policies and allows novel controller\nvariants to be sampled. Leveraging our generative policy network, a robot can\nsample novel behaviors until it finds one that works for a new environment. We\ndemonstrate this idea with an application of robust ball-throwing in the\npresence of obstacles. We show that this approach achieves a greater diversity\nof behaviors than an existing evolutionary approach, while maintaining good\nefficacy of sampled behaviors, allowing a Baxter robot to hit targets more\noften when ball throwing in the presence of obstacles.', 'The success of reinforcement learning for real world robotics has been, in\nmany cases limited to instrumented laboratory scenarios, often requiring\narduous human effort and oversight to enable continuous learning. In this work,\nwe discuss the elements that are needed for a robotic learning system that can\ncontinually and autonomously improve with data collected in the real world. We\npropose a particular instantiation of such a system, using dexterous\nmanipulation as our case study. Subsequently, we investigate a number of\nchallenges that come up when learning without instrumentation. In such\nsettings, learning must be feasible without manually designed resets, using\nonly on-board perception, and without hand-engineered reward functions. We\npropose simple and scalable solutions to these challenges, and then demonstrate\nthe efficacy of our proposed system on a set of dexterous robotic manipulation\ntasks, providing an in-depth analysis of the challenges associated with this\nlearning paradigm. We demonstrate that our complete system can learn without\nany human intervention, acquiring a variety of vision-based skills with a\nreal-world three-fingered hand. Results and videos can be found at\nhttps://sites.google.com/view/realworld-rl/', ""While reinforcement learning provides an appealing formalism for learning\nindividual skills, a general-purpose robotic system must be able to master an\nextensive repertoire of behaviors. Instead of learning a large collection of\nskills individually, can we instead enable a robot to propose and practice its\nown behaviors automatically, learning about the affordances and behaviors that\nit can perform in its environment, such that it can then repurpose this\nknowledge once a new task is commanded by the user? In this paper, we study\nthis question in the context of self-supervised goal-conditioned reinforcement\nlearning. A central challenge in this learning regime is the problem of goal\nsetting: in order to practice useful skills, the robot must be able to\nautonomously set goals that are feasible but diverse. When the robot's\nenvironment and available objects vary, as they do in most open-world settings,\nthe robot must propose to itself only those goals that it can accomplish in its\npresent setting with the objects that are at hand. Previous work only studies\nself-supervised goal-conditioned RL in a single-environment setting, where goal\nproposals come from the robot's past experience or a generative model are\nsufficient. In more diverse settings, this frequently leads to impossible goals\nand, as we show experimentally, prevents effective learning. We propose a\nconditional goal-setting model that aims to propose goals that are feasible\nfrom the robot's current state. We demonstrate that this enables\nself-supervised goal-conditioned off-policy learning with raw image\nobservations in the real world, enabling a robot to manipulate a variety of\nobjects and generalize to new objects that were not seen during training."", 'In order to be effective general purpose machines in real world environments,\nrobots not only will need to adapt their existing manipulation skills to new\ncircumstances, they will need to acquire entirely new skills on-the-fly. A\ngreat promise of continual learning is to endow robots with this ability, by\nusing their accumulated knowledge and experience from prior skills. We take a\nfresh look at this problem, by considering a setting in which the robot is\nlimited to storing that knowledge and experience only in the form of learned\nskill policies. We show that storing skill policies, careful pre-training, and\nappropriately choosing when to transfer those skill policies is sufficient to\nbuild a continual learner in the context of robotic manipulation. We analyze\nwhich conditions are needed to transfer skills in the challenging Meta-World\nsimulation benchmark. Using this analysis, we introduce a pair-wise metric\nrelating skills that allows us to predict the effectiveness of skill transfer\nbetween tasks, and use it to reduce the problem of continual learning to\ncurriculum selection. Given an appropriate curriculum, we show how to\ncontinually acquire robotic manipulation skills without forgetting, and using\nfar fewer samples than needed to train them from scratch.', 'We propose Scheduled Auxiliary Control (SAC-X), a new learning paradigm in\nthe context of Reinforcement Learning (RL). SAC-X enables learning of complex\nbehaviors - from scratch - in the presence of multiple sparse reward signals.\nTo this end, the agent is equipped with a set of general auxiliary tasks, that\nit attempts to learn simultaneously via off-policy RL. The key idea behind our\nmethod is that active (learned) scheduling and execution of auxiliary policies\nallows the agent to efficiently explore its environment - enabling it to excel\nat sparse reward RL. Our experiments in several challenging robotic\nmanipulation settings demonstrate the power of our approach.', 'A core challenge for an autonomous agent acting in the real world is to adapt\nits repertoire of skills to cope with its noisy perception and dynamics. To\nscale learning of skills to long-horizon tasks, robots should be able to learn\nand later refine their skills in a structured manner through trajectories\nrather than making instantaneous decisions individually at each time step. To\nthis end, we propose the Soft Actor-Critic Gaussian Mixture Model (SAC-GMM), a\nnovel hybrid approach that learns robot skills through a dynamical system and\nadapts the learned skills in their own trajectory distribution space through\ninteractions with the environment. Our approach combines classical robotics\ntechniques of learning from demonstration with the deep reinforcement learning\nframework and exploits their complementary nature. We show that our method\nutilizes sensors solely available during the execution of preliminarily learned\nskills to extract relevant features that lead to faster skill refinement.\nExtensive evaluations in both simulation and real-world environments\ndemonstrate the effectiveness of our method in refining robot skills by\nleveraging physical interactions, high-dimensional sensory data, and sparse\ntask completion rewards. Videos, code, and pre-trained models are available at\nhttp://sac-gmm.cs.uni-freiburg.de.', 'Learning predictive models from interaction with the world allows an agent,\nsuch as a robot, to learn about how the world works, and then use this learned\nmodel to plan coordinated sequences of actions to bring about desired outcomes.\nHowever, learning a model that captures the dynamics of complex skills\nrepresents a major challenge: if the agent needs a good model to perform these\nskills, it might never be able to collect the experience on its own that is\nrequired to learn these delicate and complex behaviors. Instead, we can imagine\naugmenting the training set with observational data of other agents, such as\nhumans. Such data is likely more plentiful, but represents a different\nembodiment. For example, videos of humans might show a robot how to use a tool,\nbut (i) are not annotated with suitable robot actions, and (ii) contain a\nsystematic distributional shift due to the embodiment differences between\nhumans and robots. We address the first challenge by formulating the\ncorresponding graphical model and treating the action as an observed variable\nfor the interaction data and an unobserved variable for the observation data,\nand the second challenge by using a domain-dependent prior. In addition to\ninteraction data, our method is able to leverage videos of passive observations\nin a driving dataset and a dataset of robotic manipulation videos. A robotic\nplanning agent equipped with our method can learn to use tools in a tabletop\nrobotic manipulation setting by observing humans without ever seeing a robotic\nvideo of tool use.', 'For many real-world robotics applications, robots need to continually adapt\nand learn new concepts. Further, robots need to learn through limited data\nbecause of scarcity of labeled data in the real-world environments. To this\nend, my research focuses on developing robots that continually learn in dynamic\nunseen environments/scenarios, learn from limited human supervision, remember\npreviously learned knowledge and use that knowledge to learn new concepts. I\ndevelop machine learning models that not only produce State-of-the-results on\nbenchmark datasets but also allow robots to learn new objects and scenes in\nunconstrained environments which lead to a variety of novel robotics\napplications.', 'We are motivated by the goal of generalist robots that can complete a wide\nrange of tasks across many environments. Critical to this is the robot\'s\nability to acquire some metric of task success or reward, which is necessary\nfor reinforcement learning, planning, or knowing when to ask for help. For a\ngeneral-purpose robot operating in the real world, this reward function must\nalso be able to generalize broadly across environments, tasks, and objects,\nwhile depending only on on-board sensor observations (e.g. RGB images). While\ndeep learning on large and diverse datasets has shown promise as a path towards\nsuch generalization in computer vision and natural language, collecting high\nquality datasets of robotic interaction at scale remains an open challenge. In\ncontrast, ""in-the-wild"" videos of humans (e.g. YouTube) contain an extensive\ncollection of people doing interesting tasks across a diverse range of\nsettings. In this work, we propose a simple approach, Domain-agnostic Video\nDiscriminator (DVD), that learns multitask reward functions by training a\ndiscriminator to classify whether two videos are performing the same task, and\ncan generalize by virtue of learning from a small amount of robot data with a\nbroad dataset of human videos. We find that by leveraging diverse human\ndatasets, this reward function (a) can generalize zero shot to unseen\nenvironments, (b) generalize zero shot to unseen tasks, and (c) can be combined\nwith visual model predictive control to solve robotic manipulation tasks on a\nreal WidowX200 robot in an unseen environment from a single human demo.', ""Exploration of indoor environments has recently experienced a significant\ninterest, also thanks to the introduction of deep neural agents built in a\nhierarchical fashion and trained with Deep Reinforcement Learning (DRL) on\nsimulated environments. Current state-of-the-art methods employ a dense\nextrinsic reward that requires the complete a priori knowledge of the layout of\nthe training environment to learn an effective exploration policy. However,\nsuch information is expensive to gather in terms of time and resources. In this\nwork, we propose to train the model with a purely intrinsic reward signal to\nguide exploration, which is based on the impact of the robot's actions on its\ninternal representation of the environment. So far, impact-based rewards have\nbeen employed for simple tasks and in procedurally generated synthetic\nenvironments with countable states. Since the number of states observable by\nthe agent in realistic indoor environments is non-countable, we include a\nneural-based density model and replace the traditional count-based\nregularization with an estimated pseudo-count of previously visited states. The\nproposed exploration approach outperforms DRL-based competitors relying on\nintrinsic rewards and surpasses the agents trained with a dense extrinsic\nreward computed with the environment layouts. We also show that a robot\nequipped with the proposed approach seamlessly adapts to point-goal navigation\nand real-world deployment."", 'While reinforcement learning algorithms can learn effective policies for\ncomplex tasks, these policies are often brittle to even minor task variations,\nespecially when variations are not explicitly provided during training. One\nnatural approach to this problem is to train agents with manually specified\nvariation in the training task or environment. However, this may be infeasible\nin practical situations, either because making perturbations is not possible,\nor because it is unclear how to choose suitable perturbation strategies without\nsacrificing performance. The key insight of this work is that learning diverse\nbehaviors for accomplishing a task can directly lead to behavior that\ngeneralizes to varying environments, without needing to perform explicit\nperturbations during training. By identifying multiple solutions for the task\nin a single environment during training, our approach can generalize to new\nsituations by abandoning solutions that are no longer effective and adopting\nthose that are. We theoretically characterize a robustness set of environments\nthat arises from our algorithm and empirically find that our diversity-driven\napproach can extrapolate to various changes in the environment and task.', 'Training robots for operation in the real world is a complex, time consuming\nand potentially expensive task. Despite significant success of reinforcement\nlearning in games and simulations, research in real robot applications has not\nbeen able to match similar progress. While sample complexity can be reduced by\ntraining policies in simulation, such policies can perform sub-optimally on the\nreal platform given imperfect calibration of model dynamics. We present an\napproach -- supplemental to fine tuning on the real robot -- to further benefit\nfrom parallel access to a simulator during training and reduce sample\nrequirements on the real robot. The developed approach harnesses auxiliary\nrewards to guide the exploration for the real world agent based on the\nproficiency of the agent in simulation and vice versa. In this context, we\ndemonstrate empirically that the reciprocal alignment for both agents provides\nfurther benefit as the agent in simulation can adjust to optimize its behaviour\nfor states commonly visited by the real-world agent.', 'Reinforcement learning has been applied to a wide variety of robotics\nproblems, but most of such applications involve collecting data from scratch\nfor each new task. Since the amount of robot data we can collect for any single\ntask is limited by time and cost considerations, the learned behavior is\ntypically narrow: the policy can only execute the task in a handful of\nscenarios that it was trained on. What if there was a way to incorporate a\nlarge amount of prior data, either from previously solved tasks or from\nunsupervised or undirected environment interaction, to extend and generalize\nlearned behaviors? While most prior work on extending robotic skills using\npre-collected data focuses on building explicit hierarchies or skill\ndecompositions, we show in this paper that we can reuse prior data to extend\nnew skills simply through dynamic programming. We show that even when the prior\ndata does not actually succeed at solving the new task, it can still be\nutilized for learning a better policy, by providing the agent with a broader\nunderstanding of the mechanics of its environment. We demonstrate the\neffectiveness of our approach by chaining together several behaviors seen in\nprior datasets for solving a new task, with our hardest experimental setting\ninvolving composing four robotic skills in a row: picking, placing, drawer\nopening, and grasping, where a +1/0 sparse reward is provided only on task\ncompletion. We train our policies in an end-to-end fashion, mapping\nhigh-dimensional image observations to low-level robot control commands, and\npresent results in both simulated and real world domains. Additional materials\nand source code can be found on our project website:\nhttps://sites.google.com/view/cog-rl', ""This paper contributes a preliminary report on the advantages and\ndisadvantages of incorporating simultaneous human control and feedback signals\nin the training of a reinforcement learning robotic agent. While robotic\nhuman-machine interfaces have become increasingly complex in both form and\nfunction, control remains challenging for users. This has resulted in an\nincreasing gap between user control approaches and the number of robotic motors\nwhich can be controlled. One way to address this gap is to shift some autonomy\nto the robot. Semi-autonomous actions of the robotic agent can then be shaped\nby human feedback, simplifying user control. Most prior work on agent shaping\nby humans has incorporated training with feedback, or has included indirect\ncontrol signals. By contrast, in this paper we explore how a human can provide\nconcurrent feedback signals and real-time myoelectric control signals to train\na robot's actor-critic reinforcement learning control system. Using both a\nphysical and a simulated robotic system, we compare training performance on a\nsimple movement task when reward is derived from the environment, when reward\nis provided by the human, and combinations of these two approaches. Our results\nindicate that some benefit can be gained with the inclusion of human generated\nfeedback."", 'In most real world scenarios, a policy trained by reinforcement learning in\none environment needs to be deployed in another, potentially quite different\nenvironment. However, generalization across different environments is known to\nbe hard. A natural solution would be to keep training after deployment in the\nnew environment, but this cannot be done if the new environment offers no\nreward signal. Our work explores the use of self-supervision to allow the\npolicy to continue training after deployment without using any rewards. While\nprevious methods explicitly anticipate changes in the new environment, we\nassume no prior knowledge of those changes yet still obtain significant\nimprovements. Empirical evaluations are performed on diverse simulation\nenvironments from DeepMind Control suite and ViZDoom, as well as real robotic\nmanipulation tasks in continuously changing environments, taking observations\nfrom an uncalibrated camera. Our method improves generalization in 31 out of 36\nenvironments across various tasks and outperforms domain randomization on a\nmajority of environments.', 'Humans learn all their life long. They accumulate knowledge from a sequence\nof learning experiences and remember the essential concepts without forgetting\nwhat they have learned previously. Artificial neural networks struggle to learn\nsimilarly. They often rely on data rigorously preprocessed to learn solutions\nto specific problems such as classification or regression. In particular, they\nforget their past learning experiences if trained on new ones. Therefore,\nartificial neural networks are often inept to deal with real-life settings such\nas an autonomous-robot that has to learn on-line to adapt to new situations and\novercome new problems without forgetting its past learning-experiences.\nContinual learning (CL) is a branch of machine learning addressing this type of\nproblem. Continual algorithms are designed to accumulate and improve knowledge\nin a curriculum of learning-experiences without forgetting. In this thesis, we\npropose to explore continual algorithms with replay processes. Replay processes\ngather together rehearsal methods and generative replay methods. Generative\nReplay consists of regenerating past learning experiences with a generative\nmodel to remember them. Rehearsal consists of saving a core-set of samples from\npast learning experiences to rehearse them later. The replay processes make\npossible a compromise between optimizing the current learning objective and the\npast ones enabling learning without forgetting in sequences of tasks settings.\nWe show that they are very promising methods for continual learning. Notably,\nthey enable the re-evaluation of past data with new knowledge and the\nconfrontation of data from different learning-experiences. We demonstrate their\nability to learn continually through unsupervised learning, supervised learning\nand reinforcement learning tasks.', 'Reinforcement learning (RL) algorithms find applications in inventory\ncontrol, recommender systems, vehicular traffic management, cloud computing and\nrobotics. The real-world complications of many tasks arising in these domains\nmakes them difficult to solve with the basic assumptions underlying classical\nRL algorithms. RL agents in these applications often need to react and adapt to\nchanging operating conditions. A significant part of research on single-agent\nRL techniques focuses on developing algorithms when the underlying assumption\nof stationary environment model is relaxed. This paper provides a survey of RL\nmethods developed for handling dynamically varying environment models. The goal\nof methods not limited by the stationarity assumption is to help autonomous\nagents adapt to varying operating conditions. This is possible either by\nminimizing the rewards lost during learning by RL agent or by finding a\nsuitable policy for the RL agent which leads to efficient operation of the\nunderlying system. A representative collection of these algorithms is discussed\nin detail in this work along with their categorization and their relative\nmerits and demerits. Additionally we also review works which are tailored to\napplication domains. Finally, we discuss future enhancements for this field.', 'Live fire creates a dynamic, rapidly changing environment that presents a\nworthy challenge for deep learning and artificial intelligence methodologies to\nassist firefighters with scene comprehension in maintaining their situational\nawareness, tracking and relay of important features necessary for key decisions\nas they tackle these catastrophic events. We propose a deep Q-learning based\nagent who is immune to stress induced disorientation and anxiety and thus able\nto make clear decisions for navigation based on the observed and stored facts\nin live fire environments. As a proof of concept, we imitate structural fire in\na gaming engine called Unreal Engine which enables the interaction of the agent\nwith the environment. The agent is trained with a deep Q-learning algorithm\nbased on a set of rewards and penalties as per its actions on the environment.\nWe exploit experience replay to accelerate the learning process and augment the\nlearning of the agent with human-derived experiences. The agent trained under\nthis deep Q-learning approach outperforms agents trained through alternative\npath planning systems and demonstrates this methodology as a promising\nfoundation on which to build a path planning navigation assistant capable of\nsafely guiding fire fighters through live fire environments.', 'In this paper we consider the problem of robot navigation in simple maze-like\nenvironments where the robot has to rely on its onboard sensors to perform the\nnavigation task. In particular, we are interested in solutions to this problem\nthat do not require localization, mapping or planning. Additionally, we require\nthat our solution can quickly adapt to new situations (e.g., changing\nnavigation goals and environments). To meet these criteria we frame this\nproblem as a sequence of related reinforcement learning tasks. We propose a\nsuccessor feature based deep reinforcement learning algorithm that can learn to\ntransfer knowledge from previously mastered navigation tasks to new problem\ninstances. Our algorithm substantially decreases the required learning time\nafter the first task instance has been solved, which makes it easily adaptable\nto changing environments. We validate our method in both simulated and real\nrobot experiments with a Robotino and compare it to a set of baseline methods\nincluding classical planning-based navigation.', 'Deep reinforcement learning (RL) has emerged as a promising approach for\nautonomously acquiring complex behaviors from low level sensor observations.\nAlthough a large portion of deep RL research has focused on applications in\nvideo games and simulated control, which does not connect with the constraints\nof learning in real environments, deep RL has also demonstrated promise in\nenabling physical robots to learn complex skills in the real world. At the same\ntime,real world robotics provides an appealing domain for evaluating such\nalgorithms, as it connects directly to how humans learn; as an embodied agent\nin the real world. Learning to perceive and move in the real world presents\nnumerous challenges, some of which are easier to address than others, and some\nof which are often not considered in RL research that focuses only on simulated\ndomains. In this review article, we present a number of case studies involving\nrobotic deep RL. Building off of these case studies, we discuss commonly\nperceived challenges in deep RL and how they have been addressed in these\nworks. We also provide an overview of other outstanding challenges, many of\nwhich are unique to the real-world robotics setting and are not often the focus\nof mainstream RL research. Our goal is to provide a resource both for\nroboticists and machine learning researchers who are interested in furthering\nthe progress of deep RL in the real world.', 'For a robot to learn a good policy, it often requires expensive equipment\n(such as sophisticated sensors) and a prepared training environment conducive\nto learning. However, it is seldom possible to perfectly equip robots for\neconomic reasons, nor to guarantee ideal learning conditions, when deployed in\nreal-life environments. A solution would be to prepare the robot in the lab\nenvironment, when all necessary material is available to learn a good policy.\nAfter training in the lab, the robot should be able to get by without the\nexpensive equipment that used to be available to it, and yet still be\nguaranteed to perform well on the field. The transition between the lab\n(source) and the real-world environment (target) is related to transfer\nlearning, where the state-space between the source and target tasks differ. We\ntackle a simulated task with continuous states and discrete actions presenting\nthis challenge, using Bootstrapped Dual Policy Iteration, a model-free\nactor-critic reinforcement learning algorithm, and Policy Shaping.\nSpecifically, we train a BDPI agent, embodied by a virtual robot performing a\ntask in the V-Rep simulator, sensing its environment through several proximity\nsensors. The resulting policy is then used by a second agent learning the same\ntask in the same environment, but with camera images as input. The goal is to\nobtain a policy able to perform the task relying on merely camera images.']","['If a robotic agent wants to exploit symbolic planning techniques to achieve\nsome goal, it must be able to properly ground an abstract planning domain in\nthe environment in which it operates. However, if the environment is initially\nunknown by the agent, the agent needs to explore it and discover the salient\naspects of the environment needed to reach its goals. Namely, the agent has to\ndiscover: (i) the objects present in the environment, (ii) the properties of\nthese objects and their relations, and finally (iii) how abstract actions can\nbe successfully executed. The paper proposes a framework that aims to\naccomplish the aforementioned perspective for an agent that perceives the\nenvironment partially and subjectively, through real value sensors (e.g., GPS,\nand on-board camera) and can operate in the environment through low level\nactuators (e.g., move forward of 20 cm). We evaluate the proposed architecture\nin photo-realistic simulated environments, where the sensors are RGB-D on-board\ncamera, GPS and compass, and low level actions include movements,\ngrasping/releasing objects, and manipulating objects. The agent is placed in an\nunknown environment and asked to find objects of a certain type, place an\nobject on top of another, close or open an object of a certain type. We compare\nour approach with the state of the art methods on object goal navigation based\non reinforcement learning, showing better performances.', 'Robots that are trained to perform a task in a fixed environment often fail\nwhen facing unexpected changes to the environment due to a lack of exploration.\nWe propose a principled way to adapt the policy for better exploration in\nchanging sparse-reward environments. Unlike previous works which explicitly\nmodel environmental changes, we analyze the relationship between the value\nfunction and the optimal exploration for a Gaussian-parameterized policy and\nshow that our theory leads to an effective strategy for adjusting the variance\nof the policy, enabling fast adapt to changes in a variety of sparse-reward\nenvironments.', ""Reinforcement Learning AI commonly uses reward/penalty signals that are\nobjective and explicit in an environment -- e.g. game score, completion time,\netc. -- in order to learn the optimal strategy for task performance. However,\nHuman-AI interaction for such AI agents should include additional reinforcement\nthat is implicit and subjective -- e.g. human preferences for certain AI\nbehavior -- in order to adapt the AI behavior to idiosyncratic human\npreferences. Such adaptations would mirror naturally occurring processes that\nincrease trust and comfort during social interactions. Here, we show how a\nhybrid brain-computer-interface (hBCI), which detects an individual's level of\ninterest in objects/events in a virtual environment, can be used to adapt the\nbehavior of a Deep Reinforcement Learning AI agent that is controlling a\nvirtual autonomous vehicle. Specifically, we show that the AI learns a driving\nstrategy that maintains a safe distance from a lead vehicle, and most novelly,\npreferentially slows the vehicle when the human passengers of the vehicle\nencounter objects of interest. This adaptation affords an additional 20\\%\nviewing time for subjectively interesting objects. This is the first\ndemonstration of how an hBCI can be used to provide implicit reinforcement to\nan AI agent in a way that incorporates user preferences into the control\nsystem."", ""Learning a policy capable of moving an agent between any two states in the\nenvironment is important for many robotics problems involving navigation and\nmanipulation. Due to the sparsity of rewards in such tasks, applying\nreinforcement learning in these scenarios can be challenging. Common approaches\nfor tackling this problem include reward engineering with auxiliary rewards,\nrequiring domain-specific knowledge or changing the objective.\n  In this work, we introduce a method based on region-growing that allows\nlearning in an environment with any pair of initial and goal states. Our\nalgorithm first learns how to move between nearby states and then increases the\ndifficulty of the start-goal transitions as the agent's performance improves.\nThis approach creates an efficient curriculum for learning the objective\nbehavior of reaching any goal from any initial state. In addition, we describe\na method to adaptively adjust expansion of the growing region that allows\nautomatic adjustment of the key exploration hyperparameter to environments with\ndifferent requirements. We evaluate our approach on a set of simulated\nnavigation and manipulation tasks, where we demonstrate that our algorithm can\nefficiently learn a policy in the presence of sparse rewards."", 'Autonomous robots may be able to adapt their behavior in response to changes\nin the environment. This is useful, for example, to efficiently handle limited\nresources or to respond appropriately to unexpected events such as faults. The\narchitecture of a self-adaptive robot is complex because it should include\nautomatic mechanisms to dynamically configure the elements that control robot\nbehaviors. To facilitate the construction of this type of architectures, it is\nuseful to have general solutions in the form of software tools that may be\napplicable to different robotic systems. This paper presents an original\nalgorithm to dynamically configure the control architecture, which is\napplicable to the development of self-adaptive autonomous robots. This\nalgorithm uses a constraint-based configuration approach to decide which basic\nrobot behaviors should be activated in response to both reactive and\ndeliberative events. The algorithm uses specific search heuristics and\ninitialization procedures to achieve the performance required by robotic\nsystems. The solution has been implemented as a software development tool\ncalled Behavior Coordinator CBC (Constraint-Based Configuration), which is\nbased on ROS and open source, available to the general public. This tool has\nbeen successfully used for building multiple applications of autonomous aerial\nrobots.', ""One of the central challenges faced by a reinforcement learning (RL) agent is\nto effectively learn a (near-)optimal policy in environments with large state\nspaces having sparse and noisy feedback signals. In real-world applications, an\nexpert with additional domain knowledge can help in speeding up the learning\nprocess via \\emph{shaping the environment}, i.e., making the environment more\nlearner-friendly. A popular paradigm in literature is \\emph{potential-based\nreward shaping}, where the environment's reward function is augmented with\nadditional local rewards using a potential function. However, the applicability\nof potential-based reward shaping is limited in settings where (i) the state\nspace is very large, and it is challenging to compute an appropriate potential\nfunction, (ii) the feedback signals are noisy, and even with shaped rewards the\nagent could be trapped in local optima, and (iii) changing the rewards alone is\nnot sufficient, and effective shaping requires changing the dynamics. We\naddress these limitations of potential-based shaping methods and propose a\nnovel framework of \\emph{environment shaping using state abstraction}. Our key\nidea is to compress the environment's large state space with noisy signals to\nan abstracted space, and to use this abstraction in creating smoother and more\neffective feedback signals for the agent. We study the theoretical\nunderpinnings of our abstraction-based environment shaping, and show that the\nagent's policy learnt in the shaped environment preserves near-optimal behavior\nin the original environment."", ""It is a long-standing challenge to enable an intelligent agent to learn in\none environment and generalize to an unseen environment without further data\ncollection and finetuning. In this paper, we consider a zero shot\ngeneralization problem setup that complies with biological intelligent agents'\nlearning and generalization processes. The agent is first presented with\nprevious experiences in the training environment, along with task description\nin the form of trajectory-level sparse rewards. Later when it is placed in the\nnew testing environment, it is asked to perform the task without any\ninteraction with the testing environment. We find this setting natural for\nbiological creatures and at the same time, challenging for previous methods.\nBehavior cloning, state-of-art RL along with other zero-shot learning methods\nperform poorly on this benchmark. Given a set of experiences in the training\nenvironment, our method learns a neural function that decomposes the sparse\nreward into particular regions in a contingency-aware observation as a per step\nreward. Based on such decomposed rewards, we further learn a dynamics model and\nuse Model Predictive Control (MPC) to obtain a policy. Since the rewards are\ndecomposed to finer-granularity observations, they are naturally generalizable\nto new environments that are composed of similar basic elements. We demonstrate\nour method on a wide range of environments, including a classic video game --\nSuper Mario Bros, as well as a robotic continuous control task. Please refer to\nthe project page for more visualized results."", 'Deep reinforcement learning agents have achieved state-of-the-art results by\ndirectly maximising cumulative reward. However, environments contain a much\nwider variety of possible training signals. In this paper, we introduce an\nagent that also maximises many other pseudo-reward functions simultaneously by\nreinforcement learning. All of these tasks share a common representation that,\nlike unsupervised learning, continues to develop in the absence of extrinsic\nrewards. We also introduce a novel mechanism for focusing this representation\nupon extrinsic rewards, so that learning can rapidly adapt to the most relevant\naspects of the actual task. Our agent significantly outperforms the previous\nstate-of-the-art on Atari, averaging 880\\% expert human performance, and a\nchallenging suite of first-person, three-dimensional \\emph{Labyrinth} tasks\nleading to a mean speedup in learning of 10$\\times$ and averaging 87\\% expert\nhuman performance on Labyrinth.', ""The main novelty of the proposed approach is that it allows a robot to learn\nan end-to-end policy which can adapt to changes in the environment during\nexecution. While goal conditioning of policies has been studied in the RL\nliterature, such approaches are not easily extended to cases where the robot's\ngoal can change during execution. This is something that humans are naturally\nable to do. However, it is difficult for robots to learn such reflexes (i.e.,\nto naturally respond to dynamic environments), especially when the goal\nlocation is not explicitly provided to the robot, and instead needs to be\nperceived through a vision sensor. In the current work, we present a method\nthat can achieve such behavior by combining traditional kinematic planning,\ndeep learning, and deep reinforcement learning in a synergistic fashion to\ngeneralize to arbitrary environments. We demonstrate the proposed approach for\nseveral reaching and pick-and-place tasks in simulation, as well as on a real\nsystem of a 6-DoF industrial manipulator. A video describing our work could be\nfound \\url{https://youtu.be/hE-Ew59GRPQ}."", 'Successfully navigating a complex environment to obtain a desired outcome is\na difficult task, that up to recently was believed to be capable only by\nhumans. This perception has been broken down over time, especially with the\nintroduction of deep reinforcement learning, which has greatly increased the\ndifficulty of tasks that can be automated. However, for traditional\nreinforcement learning agents this requires an environment to be able to\nprovide frequent extrinsic rewards, which are not known or accessible for many\nreal-world environments. This project aims to explore and contrast existing\nreinforcement learning solutions that circumnavigate the difficulties of an\nenvironment that provide sparse rewards. Different reinforcement solutions will\nbe implemented over a several video game environments with varying difficulty\nand varying frequency of rewards, as to properly investigate the applicability\nof these solutions. This project introduces a novel reinforcement learning\nsolution by combining aspects of two existing state of the art sparse reward\nsolutions, curiosity driven exploration and unsupervised auxiliary tasks.', ""A grand goal in AI is to build a robot that can accurately navigate based on\nnatural language instructions, which requires the agent to perceive the scene,\nunderstand and ground language, and act in the real-world environment. One key\nchallenge here is to learn to navigate in new environments that are unseen\nduring training. Most of the existing approaches perform dramatically worse in\nunseen environments as compared to seen ones. In this paper, we present a\ngeneralizable navigational agent. Our agent is trained in two stages. The first\nstage is training via mixed imitation and reinforcement learning, combining the\nbenefits from both off-policy and on-policy optimization. The second stage is\nfine-tuning via newly-introduced 'unseen' triplets (environment, path,\ninstruction). To generate these unseen triplets, we propose a simple but\neffective 'environmental dropout' method to mimic unseen environments, which\novercomes the problem of limited seen environment variability. Next, we apply\nsemi-supervised learning (via back-translation) on these dropped-out\nenvironments to generate new paths and instructions. Empirically, we show that\nour agent is substantially better at generalizability when fine-tuned with\nthese triplets, outperforming the state-of-art approaches by a large margin on\nthe private unseen test set of the Room-to-Room task, and achieving the top\nrank on the leaderboard."", 'Developing agents that can quickly adapt their behavior to new tasks remains\na challenge. Meta-learning has been applied to this problem, but previous\nmethods require either specifying a reward function which can be tedious or\nproviding demonstrations which can be inefficient. In this paper, we\ninvestigate if, and how, a ""helper"" agent can be trained to interactively adapt\ntheir behavior to maximize the reward of another agent, whom we call the\n""prime"" agent, without observing their reward or receiving explicit\ndemonstrations. To this end, we propose to meta-learn a helper agent along with\na prime agent, who, during training, observes the reward function and serves as\na surrogate for a human prime. We introduce a distribution of multi-agent\ncooperative foraging tasks, in which only the prime agent knows the objects\nthat should be collected. We demonstrate that, from the emerged physical\ncommunication, the trained helper rapidly infers and collects the correct\nobjects.', 'The objective of a reinforcement learning agent is to behave so as to\nmaximise the sum of a suitable scalar function of state: the reward. These\nrewards are typically given and immutable. In this paper, we instead consider\nthe proposition that the reward function itself can be a good locus of learned\nknowledge. To investigate this, we propose a scalable meta-gradient framework\nfor learning useful intrinsic reward functions across multiple lifetimes of\nexperience. Through several proof-of-concept experiments, we show that it is\nfeasible to learn and capture knowledge about long-term exploration and\nexploitation into a reward function. Furthermore, we show that unlike policy\ntransfer methods that capture ""how"" the agent should behave, the learned reward\nfunctions can generalise to other kinds of agents and to changes in the\ndynamics of the environment by capturing ""what"" the agent should strive to do.', 'Designing a good reward function is essential to robot planning and\nreinforcement learning, but it can also be challenging and frustrating. The\nreward needs to work across multiple different environments, and that often\nrequires many iterations of tuning. We introduce a novel divide-and-conquer\napproach that enables the designer to specify a reward separately for each\nenvironment. By treating these separate reward functions as observations about\nthe underlying true reward, we derive an approach to infer a common reward\nacross all environments. We conduct user studies in an abstract grid world\ndomain and in a motion planning domain for a 7-DOF manipulator that measure\nuser effort and solution quality. We show that our method is faster, easier to\nuse, and produces a higher quality solution than the typical method of\ndesigning a reward jointly across all environments. We additionally conduct a\nseries of experiments that measure the sensitivity of these results to\ndifferent properties of the reward design task, such as the number of\nenvironments, the number of feasible solutions per environment, and the\nfraction of the total features that vary within each environment. We find that\nindependent reward design outperforms the standard, joint, reward design\nprocess but works best when the design problem can be divided into simpler\nsubproblems.', 'Deep Reinforcement Learning (DRL) has been successfully used to solve\ndifferent challenges, e.g. complex board and computer games, recently. However,\nsolving real-world robotics tasks with DRL seems to be a more difficult\nchallenge. The desired approach would be to train the agent in a simulator and\ntransfer it to the real world. Still, models trained in a simulator tend to\nperform poorly in real-world environments due to the differences. In this\npaper, we present a DRL-based algorithm that is capable of performing\nautonomous robot control using Deep Q-Networks (DQN). In our approach, the\nagent is trained in a simulated environment and it is able to navigate both in\na simulated and real-world environment. The method is evaluated in the\nDuckietown environment, where the agent has to follow the lane based on a\nmonocular camera input. The trained agent is able to run on limited hardware\nresources and its performance is comparable to state-of-the-art approaches.', 'To acquire a new skill, humans learn better and faster if a tutor, based on\ntheir current knowledge level, informs them of how much attention they should\npay to particular content or practice problems. Similarly, a machine learning\nmodel could potentially be trained better with a scorer that ""adapts"" to its\ncurrent learning state and estimates the importance of each training data\ninstance. Training such an adaptive scorer efficiently is a challenging\nproblem; in order to precisely quantify the effect of a data instance at a\ngiven time during the training, it is typically necessary to first complete the\nentire training process. To efficiently optimize data usage, we propose a\nreinforcement learning approach called Differentiable Data Selection (DDS). In\nDDS, we formulate a scorer network as a learnable function of the training\ndata, which can be efficiently updated along with the main model being trained.\nSpecifically, DDS updates the scorer with an intuitive reward signal: it should\nup-weigh the data that has a similar gradient with a dev set upon which we\nwould finally like to perform well. Without significant computing overhead, DDS\ndelivers strong and consistent improvements over several strong baselines on\ntwo very different tasks of machine translation and image classification.', 'To achieve general artificial intelligence, reinforcement learning (RL)\nagents should learn not only to optimize returns for one specific task but also\nto constantly build more complex skills and scaffold their knowledge about the\nworld, without forgetting what has already been learned. In this paper, we\ndiscuss the desired characteristics of environments that can support the\ntraining and evaluation of lifelong reinforcement learning agents, review\nexisting environments from this perspective, and propose recommendations for\ndevising suitable environments in the future.', 'We propose a novel approach to address one aspect of the non-stationarity\nproblem in multi-agent reinforcement learning (RL), where the other agents may\nalter their policies due to environment changes during execution. This violates\nthe Markov assumption that governs most single-agent RL methods and is one of\nthe key challenges in multi-agent RL. To tackle this, we propose to train\nmultiple policies for each agent and postpone the selection of the best policy\nat execution time. Specifically, we model the environment non-stationarity with\na finite set of scenarios and train policies fitting each scenario. In addition\nto multiple policies, each agent also learns a policy predictor to determine\nwhich policy is the best with its local information. By doing so, each agent is\nable to adapt its policy when the environment changes and consequentially the\nother agents alter their policies during execution. We empirically evaluated\nour method on a variety of common benchmark problems proposed for multi-agent\ndeep RL in the literature. Our experimental results show that the agents\ntrained by our algorithm have better adaptiveness in changing environments and\noutperform the state-of-the-art methods in all the tested environments.', 'In many sequential decision making tasks, it is challenging to design reward\nfunctions that help an RL agent efficiently learn behavior that is considered\ngood by the agent designer. A number of different formulations of the\nreward-design problem, or close variants thereof, have been proposed in the\nliterature. In this paper we build on the Optimal Rewards Framework of Singh\net.al. that defines the optimal intrinsic reward function as one that when used\nby an RL agent achieves behavior that optimizes the task-specifying or\nextrinsic reward function. Previous work in this framework has shown how good\nintrinsic reward functions can be learned for lookahead search based planning\nagents. Whether it is possible to learn intrinsic reward functions for learning\nagents remains an open problem. In this paper we derive a novel algorithm for\nlearning intrinsic rewards for policy-gradient based learning agents. We\ncompare the performance of an augmented agent that uses our algorithm to\nprovide additive intrinsic rewards to an A2C-based policy learner (for Atari\ngames) and a PPO-based policy learner (for Mujoco domains) with a baseline\nagent that uses the same policy learners but with only extrinsic rewards. Our\nresults show improved performance on most but not all of the domains.', 'Recent work has shown that deep reinforcement-learning agents can learn to\nfollow language-like instructions from infrequent environment rewards. However,\nthis places on environment designers the onus of designing language-conditional\nreward functions which may not be easily or tractably implemented as the\ncomplexity of the environment and the language scales. To overcome this\nlimitation, we present a framework within which instruction-conditional RL\nagents are trained using rewards obtained not from the environment, but from\nreward models which are jointly trained from expert examples. As reward models\nimprove, they learn to accurately reward agents for completing tasks for\nenvironment configurations---and for instructions---not present amongst the\nexpert data. This framework effectively separates the representation of what\ninstructions require from how they can be executed. In a simple grid world, it\nenables an agent to learn a range of commands requiring interaction with blocks\nand understanding of spatial relations and underspecified abstract\narrangements. We further show the method allows our agent to adapt to changes\nin the environment without requiring new expert examples.', 'Learning reward functions from data is a promising path towards achieving\nscalable Reinforcement Learning (RL) for robotics. However, a major challenge\nin training agents from learned reward models is that the agent can learn to\nexploit errors in the reward model to achieve high reward behaviors that do not\ncorrespond to the intended task. These reward delusions can lead to unintended\nand even dangerous behaviors. On the other hand, adversarial imitation learning\nframeworks tend to suffer the opposite problem, where the discriminator learns\nto trivially distinguish agent and expert behavior, resulting in reward models\nthat produce low reward signal regardless of the input state. In this paper, we\nconnect these two classes of reward learning methods to positive-unlabeled (PU)\nlearning, and we show that by applying a large-scale PU learning algorithm to\nthe reward learning problem, we can address both the reward under- and\nover-estimation problems simultaneously. Our approach drastically improves both\nGAIL and supervised reward learning, without any additional assumptions.', 'Reinforcement learning is concerned with identifying reward-maximizing\nbehaviour policies in environments that are initially unknown. State-of-the-art\nreinforcement learning approaches, such as deep Q-networks, are model-free and\nlearn to act effectively across a wide range of environments such as Atari\ngames, but require huge amounts of data. Model-based techniques are more\ndata-efficient, but need to acquire explicit knowledge about the environment.\n  In this paper, we take a step towards using model-based techniques in\nenvironments with a high-dimensional visual state space by demonstrating that\nit is possible to learn system dynamics and the reward structure jointly. Our\ncontribution is to extend a recently developed deep neural network for video\nframe prediction in Atari games to enable reward prediction as well. To this\nend, we phrase a joint optimization problem for minimizing both video frame and\nreward reconstruction loss, and adapt network parameters accordingly. Empirical\nevaluations on five Atari games demonstrate accurate cumulative reward\nprediction of up to 200 frames. We consider these results as opening up\nimportant directions for model-based reinforcement learning in complex,\ninitially unknown environments.', ""Agents of general intelligence deployed in real-world scenarios must adapt to\never-changing environmental conditions. While such adaptive agents may leverage\nengineered knowledge, they will require the capacity to construct and evaluate\nknowledge themselves from their own experience in a bottom-up, constructivist\nfashion. This position paper builds on the idea of encoding knowledge as\ntemporally extended predictions through the use of general value functions.\nPrior work has focused on learning predictions about externally derived signals\nabout a task or environment (e.g. battery level, joint position). Here we\nadvocate that the agent should also predict internally generated signals\nregarding its own learning process - for example, an agent's confidence in its\nlearned predictions. Finally, we suggest how such information would be\nbeneficial in creating an introspective agent that is able to learn to make\ngood decisions in a complex, changing world."", ""Many reinforcement-learning researchers treat the reward function as a part\nof the environment, meaning that the agent can only know the reward of a state\nif it encounters that state in a trial run. However, we argue that this is an\nunnecessary limitation and instead, the reward function should be provided to\nthe learning algorithm. The advantage is that the algorithm can then use the\nreward function to check the reward for states that the agent hasn't even\nencountered yet. In addition, the algorithm can simultaneously learn policies\nfor multiple reward functions. For each state, the algorithm would calculate\nthe reward using each of the reward functions and add the rewards to its\nexperience replay dataset. The Hindsight Experience Replay algorithm developed\nby Andrychowicz et al. (2017) does just this, and learns to generalize across a\ndistribution of sparse, goal-based rewards. We extend this algorithm to\nlinearly-weighted, multi-objective rewards and learn a single policy that can\ngeneralize across all linear combinations of the multi-objective reward.\nWhereas other multi-objective algorithms teach the Q-function to generalize\nacross the reward weights, our algorithm enables the policy to generalize, and\ncan thus be used with continuous actions."", 'Reinforcement learning usually uses the feedback rewards of environmental to\ntrain agents. But the rewards in the actual environment are sparse, and even\nsome environments will not rewards. Most of the current methods are difficult\nto get good performance in sparse reward or non-reward environments. Although\nusing shaped rewards is effective when solving sparse reward tasks, it is\nlimited to specific problems and learning is also susceptible to local optima.\nWe propose a model-free method that does not rely on environmental rewards to\nsolve the problem of sparse rewards in the general environment. Our method use\nthe minimum number of transitions between states as the distance to replace the\nrewards of environmental, and proposes a goal-distance gradient to achieve\npolicy improvement. We also introduce a bridge point planning method based on\nthe characteristics of our method to improve exploration efficiency, thereby\nsolving more complex tasks. Experiments show that our method performs better on\nsparse reward and local optimal problems in complex environments than previous\nwork.', 'Reinforcement learning with sparse rewards is still an open challenge.\nClassic methods rely on getting feedback via extrinsic rewards to train the\nagent, and in situations where this occurs very rarely the agent learns slowly\nor cannot learn at all. Similarly, if the agent receives also rewards that\ncreate suboptimal modes of the objective function, it will likely prematurely\nstop exploring. More recent methods add auxiliary intrinsic rewards to\nencourage exploration. However, auxiliary rewards lead to a non-stationary\ntarget for the Q-function. In this paper, we present a novel approach that (1)\nplans exploration actions far into the future by using a long-term visitation\ncount, and (2) decouples exploration and exploitation by learning a separate\nfunction assessing the exploration value of the actions. Contrary to existing\nmethods which use models of reward and dynamics, our approach is off-policy and\nmodel-free. We further propose new tabular environments for benchmarking\nexploration in reinforcement learning. Empirical results on classic and novel\nbenchmarks show that the proposed approach outperforms existing methods in\nenvironments with sparse rewards, especially in the presence of rewards that\ncreate suboptimal modes of the objective function. Results also suggest that\nour approach scales gracefully with the size of the environment. Source code is\navailable at https://github.com/sparisi/visit-value-explore', 'Humans and animals show remarkable flexibility in adjusting their behaviour\nwhen their goals, or rewards in the environment change. While such flexibility\nis a hallmark of intelligent behaviour, these multi-task scenarios remain an\nimportant challenge for machine learning algorithms and neurobiological models\nalike. We investigated two approaches that could enable this flexibility:\nfactorized representations, which abstract away general aspects of a task from\nthose prone to change, and nonparametric, memory-based approaches, which can\nprovide a principled way of using similarity to past experiences to guide\ncurrent behaviour. In particular, we combine the successor representation (SR)\nthat factors the value of actions into expected outcomes and corresponding\nrewards with evaluating task similarity through clustering the space of reward\nfunctions. The proposed algorithm inverts a generative model over tasks, and\ndynamically samples from a flexible number of distinct SR maps while\naccumulating evidence about the current task context through amortized\ninference. It improves SR\'s transfer capabilities and outperforms competing\nalgorithms and baselines in settings with both known and unsignalled rewards\nchanges. Further, as a neurobiological model of spatial coding in the\nhippocampus, it explains important signatures of this representation, such as\nthe ""flickering"" behaviour of hippocampal maps, and trajectory-dependent place\ncells (so-called splitter cells) and their dynamics. We thus provide a novel\nalgorithmic approach for multi-task learning, as well as a common normative\nframework that links together these different characteristics of the brain\'s\nspatial representation.', ""An agent learning through interactions should balance its action selection\nprocess between probing the environment to discover new rewards and using the\ninformation acquired in the past to adopt useful behaviour. This trade-off is\nusually obtained by perturbing either the agent's actions (e.g., e-greedy or\nGibbs sampling) or the agent's parameters (e.g., NoisyNet), or by modifying the\nreward it receives (e.g., exploration bonus, intrinsic motivation, or\nhand-shaped rewards). Here, we adopt a disruptive but simple and generic\nperspective, where we explicitly disentangle exploration and exploitation.\nDifferent losses are optimized in parallel, one of them coming from the true\nobjective (maximizing cumulative rewards from the environment) and others being\nrelated to exploration. Every loss is used in turn to learn a policy that\ngenerates transitions, all shared in a single replay buffer. Off-policy methods\nare then applied to these transitions to optimize each loss. We showcase our\napproach on a hard-exploration environment, show its sample-efficiency and\nrobustness, and discuss further implications."", 'Transfer in reinforcement learning refers to the notion that generalization\nshould occur not only within a task but also across tasks. We propose a\ntransfer framework for the scenario where the reward function changes between\ntasks but the environment\'s dynamics remain the same. Our approach rests on two\nkey ideas: ""successor features"", a value function representation that decouples\nthe dynamics of the environment from the rewards, and ""generalized policy\nimprovement"", a generalization of dynamic programming\'s policy improvement\noperation that considers a set of policies rather than a single one. Put\ntogether, the two ideas lead to an approach that integrates seamlessly within\nthe reinforcement learning framework and allows the free exchange of\ninformation across tasks. The proposed method also provides performance\nguarantees for the transferred policy even before any learning has taken place.\nWe derive two theorems that set our approach in firm theoretical ground and\npresent experiments that show that it successfully promotes transfer in\npractice, significantly outperforming alternative methods in a sequence of\nnavigation tasks and in the control of a simulated robotic arm.', 'Learning robust value functions given raw observations and rewards is now\npossible with model-free and model-based deep reinforcement learning\nalgorithms. There is a third alternative, called Successor Representations\n(SR), which decomposes the value function into two components -- a reward\npredictor and a successor map. The successor map represents the expected future\nstate occupancy from any given state and the reward predictor maps states to\nscalar rewards. The value function of a state can be computed as the inner\nproduct between the successor map and the reward weights. In this paper, we\npresent DSR, which generalizes SR within an end-to-end deep reinforcement\nlearning framework. DSR has several appealing properties including: increased\nsensitivity to distal reward changes due to factorization of reward and world\ndynamics, and the ability to extract bottleneck states (subgoals) given\nsuccessor maps trained under a random policy. We show the efficacy of our\napproach on two diverse environments given raw pixel observations -- simple\ngrid-world domains (MazeBase) and the Doom game engine.', ""Predictions and predictive knowledge have seen recent success in improving\nnot only robot control but also other applications ranging from industrial\nprocess control to rehabilitation. A property that makes these predictive\napproaches well suited for robotics is that they can be learned online and\nincrementally through interaction with the environment. However, a remaining\nchallenge for many prediction-learning approaches is an appropriate choice of\nprediction-learning parameters, especially parameters that control the\nmagnitude of a learning machine's updates to its predictions (the learning rate\nor step size). To begin to address this challenge, we examine the use of online\nstep-size adaptation using a sensor-rich robotic arm. Our method of choice,\nTemporal-Difference Incremental Delta-Bar-Delta (TIDBD), learns and adapts step\nsizes on a feature level; importantly, TIDBD allows step-size tuning and\nrepresentation learning to occur at the same time. We show that TIDBD is a\npractical alternative for classic Temporal-Difference (TD) learning via an\nextensive parameter search. Both approaches perform comparably in terms of\npredicting future aspects of a robotic data stream. Furthermore, the use of a\nstep-size adaptation method like TIDBD appears to allow a system to\nautomatically detect and characterize common sensor failures in a robotic\napplication. Together, these results promise to improve the ability of robotic\ndevices to learn from interactions with their environments in a robust way,\nproviding key capabilities for autonomous agents and robots."", ""The ability to transfer skills across tasks has the potential to scale up\nreinforcement learning (RL) agents to environments currently out of reach.\nRecently, a framework based on two ideas, successor features (SFs) and\ngeneralised policy improvement (GPI), has been introduced as a principled way\nof transferring skills. In this paper we extend the SFs & GPI framework in two\nways. One of the basic assumptions underlying the original formulation of SFs &\nGPI is that rewards for all tasks of interest can be computed as linear\ncombinations of a fixed set of features. We relax this constraint and show that\nthe theoretical guarantees supporting the framework can be extended to any set\nof tasks that only differ in the reward function. Our second contribution is to\nshow that one can use the reward functions themselves as features for future\ntasks, without any loss of expressiveness, thus removing the need to specify a\nset of features beforehand. This makes it possible to combine SFs & GPI with\ndeep learning in a more stable way. We empirically verify this claim on a\ncomplex 3D environment where observations are images from a first-person\nperspective. We show that the transfer promoted by SFs & GPI leads to very good\npolicies on unseen tasks almost instantaneously. We also describe how to learn\npolicies specialised to the new tasks in a way that allows them to be added to\nthe agent's set of skills, and thus be reused in the future."", 'Robots are extending their presence in domestic environments every day, being\nmore common to see them carrying out tasks in home scenarios. In the future,\nrobots are expected to increasingly perform more complex tasks and, therefore,\nbe able to acquire experience from different sources as quickly as possible. A\nplausible approach to address this issue is interactive feedback, where a\ntrainer advises a learner on which actions should be taken from specific states\nto speed up the learning process. Moreover, deep reinforcement learning has\nbeen recently widely utilized in robotics to learn the environment and acquire\nnew skills autonomously. However, an open issue when using deep reinforcement\nlearning is the excessive time needed to learn a task from raw input images. In\nthis work, we propose a deep reinforcement learning approach with interactive\nfeedback to learn a domestic task in a human-robot scenario. We compare three\ndifferent learning methods using a simulated robotic arm for the task of\norganizing different objects; the proposed methods are (i) deep reinforcement\nlearning (DeepRL); (ii) interactive deep reinforcement learning using a\npreviously trained artificial agent as an advisor (agent-IDeepRL); and (iii)\ninteractive deep reinforcement learning using a human advisor (human-IDeepRL).\nWe demonstrate that interactive approaches provide advantages for the learning\nprocess. The obtained results show that a learner agent, using either\nagent-IDeepRL or human-IDeepRL, completes the given task earlier and has fewer\nmistakes compared to the autonomous DeepRL approach.', 'In this paper, we propose an AI-FML robotic agent for student learning\nbehavior ontology construction which can be applied in English speaking and\nlistening domain. The AI-FML robotic agent with the ontology contains the\nperception intelligence, computational intelligence, and cognition intelligence\nfor analyzing student learning behavior. In addition, there are three\nintelligent agents, including a perception agent, a computational agent, and a\ncognition agent in the AI-FML robotic agent. We deploy the perception agent and\nthe cognition agent on the robot Kebbi Air. Moreover, the computational agent\nwith the Deep Neural Network (DNN) model is performed in the cloud and can\ncommunicate with the perception agent and cognition agent via the Internet. The\nproposed AI-FML robotic agent is applied in Taiwan and tested in Japan. The\nexperimental results show that the agents can be utilized in the human and\nmachine co-learning model for the future education.', 'In many real-world scenarios where extrinsic rewards to the agent are\nextremely sparse, curiosity has emerged as a useful concept providing intrinsic\nrewards that enable the agent to explore its environment and acquire\ninformation to achieve its goals. Despite their strong performance on many\nsparse-reward tasks, existing curiosity approaches rely on an overly holistic\nview of state transitions, and do not allow for a structured understanding of\nspecific aspects of the environment. In this paper, we formulate curiosity\nbased on grounded question answering by encouraging the agent to ask questions\nabout the environment and be curious when the answers to these questions\nchange. We show that natural language questions encourage the agent to uncover\nspecific knowledge about their environment such as the physical properties of\nobjects as well as their spatial relationships with other objects, which serve\nas valuable curiosity rewards to solve sparse-reward tasks more efficiently.', 'Long-term temporal credit assignment is an important challenge in deep\nreinforcement learning (RL). It refers to the ability of the agent to attribute\nactions to consequences that may occur after a long time interval. Existing\npolicy-gradient and Q-learning algorithms typically rely on dense environmental\nrewards that provide rich short-term supervision and help with credit\nassignment. However, they struggle to solve tasks with delays between an action\nand the corresponding rewarding feedback. To make credit assignment easier,\nrecent works have proposed algorithms to learn dense ""guidance"" rewards that\ncould be used in place of the sparse or delayed environmental rewards. This\npaper is in the same vein -- starting with a surrogate RL objective that\ninvolves smoothing in the trajectory-space, we arrive at a new algorithm for\nlearning guidance rewards. We show that the guidance rewards have an intuitive\ninterpretation, and can be obtained without training any additional neural\nnetworks. Due to the ease of integration, we use the guidance rewards in a few\npopular algorithms (Q-learning, Actor-Critic, Distributional-RL) and present\nresults in single-agent and multi-agent tasks that elucidate the benefit of our\napproach when the environmental rewards are sparse or delayed.', ""In the real world, robots with embodiment face various issues such as dynamic\ncontinuous changes of the environment and input/output disturbances. The key to\nsolving these issues can be found in daily life; people `do actions associated\nwith sensing' and `dynamically change their plans when necessary'. We propose\nthe use of a new concept, enabling robots to do these two things, for\nautonomously controlling mobile robots. We implemented our concept to make two\nexperiments under static/dynamic environments. The results of these experiments\nshow that our idea provides a way to adapt to dynamic changes of the\nenvironment in the real world."", 'Much of the current work on reinforcement learning studies episodic settings,\nwhere the agent is reset between trials to an initial state distribution, often\nwith well-shaped reward functions. Non-episodic settings, where the agent must\nlearn through continuous interaction with the world without resets, and where\nthe agent receives only delayed and sparse reward signals, is substantially\nmore difficult, but arguably more realistic considering real-world environments\ndo not present the learner with a convenient ""reset mechanism"" and easy reward\nshaping. In this paper, instead of studying algorithmic improvements that can\naddress such non-episodic and sparse reward settings, we instead study the\nkinds of environment properties that can make learning under such conditions\neasier. Understanding how properties of the environment impact the performance\nof reinforcement learning agents can help us to structure our tasks in ways\nthat make learning tractable. We first discuss what we term ""environment\nshaping"" -- modifications to the environment that provide an alternative to\nreward shaping, and may be easier to implement. We then discuss an even simpler\nproperty that we refer to as ""dynamism,"" which describes the degree to which\nthe environment changes independent of the agent\'s actions and can be measured\nby environment transition entropy. Surprisingly, we find that even this\nproperty can substantially alleviate the challenges associated with\nnon-episodic RL in sparse reward settings. We provide an empirical evaluation\non a set of new tasks focused on non-episodic learning with sparse rewards.\nThrough this study, we hope to shift the focus of the community towards\nanalyzing how properties of the environment can affect learning and the\nultimate type of behavior that is learned via RL.', 'Robot learning is a very promising topic for the future of automation and\nmachine intelligence. Future robots should be able to autonomously acquire\nskills, learn to represent their environment, and interact with it. While these\ntopics have been explored in simulation, real-world robot learning research\nseems to be still limited. This is due to the additional challenges encountered\nin the real-world, such as noisy sensors and actuators, safe exploration,\nnon-stationary dynamics, autonomous environment resetting as well as the cost\nof running experiments for long periods of time. Unless we develop scalable\nsolutions to these problems, learning complex tasks involving hand-eye\ncoordination and rich contacts will remain an untouched vision that is only\nfeasible in controlled lab environments. We propose dual-arm settings as\nplatforms for robot learning. Such settings enable safe data collection for\nacquiring manipulation skills as well as training perception modules in a\nrobot-supervised manner. They also ease the processes of resetting the\nenvironment. Furthermore, adversarial learning could potentially boost the\ngeneralization capability of robot learning methods by maximizing the\nexploration based on game-theoretic objectives while ensuring safety based on\ncollaborative task spaces. In this paper, we will discuss the potential\nbenefits of this setup as well as the challenges and research directions that\ncan be pursued.', ""Reinforcement learning (RL), particularly in sparse reward settings, often\nrequires prohibitively large numbers of interactions with the environment,\nthereby limiting its applicability to complex problems. To address this,\nseveral prior approaches have used natural language to guide the agent's\nexploration. However, these approaches typically operate on structured\nrepresentations of the environment, and/or assume some structure in the natural\nlanguage commands. In this work, we propose a model that directly maps pixels\nto rewards, given a free-form natural language description of the task, which\ncan then be used for policy learning. Our experiments on the Meta-World robot\nmanipulation domain show that language-based rewards significantly improves the\nsample efficiency of policy learning, both in sparse and dense reward settings."", ""Recent reinforcement learning (RL) approaches have shown strong performance\nin complex domains such as Atari games, but are often highly sample\ninefficient. A common approach to reduce interaction time with the environment\nis to use reward shaping, which involves carefully designing reward functions\nthat provide the agent intermediate rewards for progress towards the goal.\nHowever, designing appropriate shaping rewards is known to be difficult as well\nas time-consuming. In this work, we address this problem by using natural\nlanguage instructions to perform reward shaping. We propose the LanguagE-Action\nReward Network (LEARN), a framework that maps free-form natural language\ninstructions to intermediate rewards based on actions taken by the agent. These\nintermediate language-based rewards can seamlessly be integrated into any\nstandard reinforcement learning algorithm. We experiment with Montezuma's\nRevenge from the Atari Learning Environment, a popular benchmark in RL. Our\nexperiments on a diverse set of 15 tasks demonstrate that, for the same number\nof interactions with the environment, language-based rewards lead to successful\ncompletion of the task 60% more often on average, compared to learning without\nlanguage."", 'The aim of my Ph.D. thesis concerns Reasoning in Highly Reactive\nEnvironments. As reasoning in highly reactive environments, we identify the\nsetting in which a knowledge-based agent, with given goals, is deployed in an\nenvironment subject to repeated, sudden and possibly unknown changes. This is\nfor instance the typical setting in which, e.g., artificial agents for\nvideo-games (the so called ""bots""), cleaning robots, bomb clearing robots, and\nso on are deployed. In all these settings one can follow the classical approach\nin which the operations of the agent are distinguished in ""sensing"" the\nenvironment with proper interface devices, ""thinking"", and then behaving\naccordingly using proper actuators. In order to operate in an highly reactive\nenvironment, an artificial agent needs to be: 1. Responsive -> The agent must\nbe able to react repeatedly and in a reasonable amount of time; 2. Elastic ->\nThe agent must stay reactive also under varying workload; 3. Resilient -> The\nagent must stay responsive also in case of internal failure or failure of one\nof the programmed actions in the environment. Nowadays, thanks to new\ntechnologies in the field of Artificial Intelligence, it is already technically\npossible to create AI agents that are able to operate in reactive environments.\nNevertheless, several issues stay unsolved, and are subject of ongoing\nresearch.', 'Imitation learning is an effective approach for autonomous systems to acquire\ncontrol policies when an explicit reward function is unavailable, using\nsupervision provided as demonstrations from an expert, typically a human\noperator. However, standard imitation learning methods assume that the agent\nreceives examples of observation-action tuples that could be provided, for\ninstance, to a supervised learning algorithm. This stands in contrast to how\nhumans and animals imitate: we observe another person performing some behavior\nand then figure out which actions will realize that behavior, compensating for\nchanges in viewpoint, surroundings, object positions and types, and other\nfactors. We term this kind of imitation learning ""imitation-from-observation,""\nand propose an imitation learning method based on video prediction with context\ntranslation and deep reinforcement learning. This lifts the assumption in\nimitation learning that the demonstration should consist of observations in the\nsame environment configuration, and enables a variety of interesting\napplications, including learning robotic skills that involve tool use simply by\nobserving videos of human tool use. Our experimental results show the\neffectiveness of our approach in learning a wide range of real-world robotic\ntasks modeled after common household chores from videos of a human\ndemonstrator, including sweeping, ladling almonds, pushing objects as well as a\nnumber of tasks in simulation.', 'Solving tasks with sparse rewards is one of the most important challenges in\nreinforcement learning. In the single-agent setting, this challenge is\naddressed by introducing intrinsic rewards that motivate agents to explore\nunseen regions of their state spaces; however, applying these techniques\nnaively to the multi-agent setting results in agents exploring independently,\nwithout any coordination among themselves. Exploration in cooperative\nmulti-agent settings can be accelerated and improved if agents coordinate their\nexploration. In this paper we introduce a framework for designing intrinsic\nrewards which consider what other agents have explored such that the agents can\ncoordinate. Then, we develop an approach for learning how to dynamically select\nbetween several exploration modalities to maximize extrinsic rewards.\nConcretely, we formulate the approach as a hierarchical policy where a\nhigh-level controller selects among sets of policies trained on diverse\nintrinsic rewards and the low-level controllers learn the action policies of\nall agents under these specific rewards. We demonstrate the effectiveness of\nthe proposed approach in cooperative domains with sparse rewards where\nstate-of-the-art methods fail and challenging multi-stage tasks that\nnecessitate changing modes of coordination.', ""There are situations in which an agent should receive rewards only after\nhaving accomplished a series of previous tasks. In other words, the reward that\nthe agent receives is non-Markovian. One natural and quite general way to\nrepresent history-dependent rewards is via a Mealy machine; a finite state\nautomaton that produces output sequences (rewards in our case) from input\nsequences (state/action observations in our case). In our formal setting, we\nconsider a Markov decision process (MDP) that models the dynamic of the\nenvironment in which the agent evolves and a Mealy machine synchronised with\nthis MDP to formalise the non-Markovian reward function. While the MDP is known\nby the agent, the reward function is unknown from the agent and must be learnt.\n  Learning non-Markov reward functions is a challenge. Our approach to overcome\nthis challenging problem is a careful combination of the Angluin's L* active\nlearning algorithm to learn finite automata, testing techniques for\nestablishing conformance of finite model hypothesis and optimisation techniques\nfor computing optimal strategies in Markovian (immediate) reward MDPs. We also\nshow how our framework can be combined with classical heuristics such as Monte\nCarlo Tree Search. We illustrate our algorithms and a preliminary\nimplementation on two typical examples for AI."", 'To solve its task, a robot needs to have the ability to interpret its\nperceptions. In vision, this interpretation is particularly difficult and\nrelies on the understanding of the structure of the scene, at least to the\nextent of its task and sensorimotor abilities. A robot with the ability to\nbuild and adapt this interpretation process according to its own tasks and\ncapabilities would push away the limits of what robots can achieve in a non\ncontrolled environment. A solution is to provide the robot with processes to\nbuild such representations that are not specific to an environment or a\nsituation. A lot of works focus on objects segmentation, recognition and\nmanipulation. Defining an object solely on the basis of its visual appearance\nis challenging given the wide range of possible objects and environments.\nTherefore, current works make simplifying assumptions about the structure of a\nscene. Such assumptions reduce the adaptivity of the object extraction process\nto the environments in which the assumption holds. To limit such assumptions,\nwe introduce an exploration method aimed at identifying moveable elements in a\nscene without considering the concept of object. By using the interactive\nperception framework, we aim at bootstrapping the acquisition process of a\nrepresentation of the environment with a minimum of context specific\nassumptions. The robotic system builds a perceptual map called relevance map\nwhich indicates the moveable parts of the current scene. A classifier is\ntrained online to predict the category of each region (moveable or\nnon-moveable). It is also used to select a region with which to interact, with\nthe goal of minimizing the uncertainty of the classification. A specific\nclassifier is introduced to fit these needs: the collaborative mixture models\nclassifier. The method is tested on a set of scenarios of increasing\ncomplexity, using both simulations and a PR2 robot.', ""Deep reinforcement learning primarily focuses on learning behavior, usually\noverlooking the fact that an agent's function is largely determined by form.\nSo, how should one go about finding a morphology fit for solving tasks in a\ngiven environment? Current approaches that co-adapt morphology and behavior use\na specific task's reward as a signal for morphology optimization. However, this\noften requires expensive policy optimization and results in task-dependent\nmorphologies that are not built to generalize. In this work, we propose a new\napproach, Task-Agnostic Morphology Evolution (TAME), to alleviate both of these\nissues. Without any task or reward specification, TAME evolves morphologies by\nonly applying randomly sampled action primitives on a population of agents.\nThis is accomplished using an information-theoretic objective that efficiently\nranks agents by their ability to reach diverse states in the environment and\nthe causality of their actions. Finally, we empirically demonstrate that across\n2D, 3D, and manipulation environments TAME can evolve morphologies that match\nthe multi-task performance of those learned with task supervised algorithms.\nOur code and videos can be found at\nhttps://sites.google.com/view/task-agnostic-evolution."", 'Deep Reinforcement Learning has been shown to be very successful in complex\ngames, e.g. Atari or Go. These games have clearly defined rules, and hence\nallow simulation. In many practical applications, however, interactions with\nthe environment are costly and a good simulator of the environment is not\navailable. Further, as environments differ by application, the optimal\ninductive bias (architecture, hyperparameters, etc.) of a reinforcement agent\ndepends on the application. In this work, we propose a multi-arm bandit\nframework that selects from a set of different reinforcement learning agents to\nchoose the one with the best inductive bias. To alleviate the problem of sparse\nrewards, the reinforcement learning agents are augmented with surrogate\nrewards. This helps the bandit framework to select the best agents early, since\nthese rewards are smoother and less sparse than the environment reward. The\nbandit has the double objective of maximizing the reward while the agents are\nlearning and selecting the best agent after a finite number of learning steps.\nOur experimental results on standard environments show that the proposed\nframework is able to consistently select the optimal agent after a finite\nnumber of steps, while collecting more cumulative reward compared to selecting\na sub-optimal architecture or uniformly alternating between different agents.', ""A fundamental challenge in multiagent reinforcement learning is to learn\nbeneficial behaviors in a shared environment with other simultaneously learning\nagents. In particular, each agent perceives the environment as effectively\nnon-stationary due to the changing policies of other agents. Moreover, each\nagent is itself constantly learning, leading to natural non-stationarity in the\ndistribution of experiences encountered. In this paper, we propose a novel\nmeta-multiagent policy gradient theorem that directly accounts for the\nnon-stationary policy dynamics inherent to multiagent learning settings. This\nis achieved by modeling our gradient updates to consider both an agent's own\nnon-stationary policy dynamics and the non-stationary policy dynamics of other\nagents in the environment. We show that our theoretically grounded approach\nprovides a general solution to the multiagent learning problem, which\ninherently comprises all key aspects of previous state of the art approaches on\nthis topic. We test our method on a diverse suite of multiagent benchmarks and\ndemonstrate a more efficient ability to adapt to new agents as they learn than\nbaseline methods across the full spectrum of mixed incentive, competitive, and\ncooperative domains."", 'In typical reinforcement learning (RL), the environment is assumed given and\nthe goal of the learning is to identify an optimal policy for the agent taking\nactions through its interactions with the environment. In this paper, we extend\nthis setting by considering the environment is not given, but controllable and\nlearnable through its interaction with the agent at the same time. This\nextension is motivated by environment design scenarios in the real-world,\nincluding game design, shopping space design and traffic signal design.\nTheoretically, we find a dual Markov decision process (MDP) w.r.t. the\nenvironment to that w.r.t. the agent, and derive a policy gradient solution to\noptimizing the parametrized environment. Furthermore, discontinuous\nenvironments are addressed by a proposed general generative framework. Our\nexperiments on a Maze game design task show the effectiveness of the proposed\nalgorithms in generating diverse and challenging Mazes against various agent\nsettings.', 'The development of intelligent machines is one of the biggest unsolved\nchallenges in computer science. In this paper, we propose some fundamental\nproperties these machines should have, focusing in particular on communication\nand learning. We discuss a simple environment that could be used to\nincrementally teach a machine the basics of natural-language-based\ncommunication, as a prerequisite to more complex interaction with human users.\nWe also present some conjectures on the sort of algorithms the machine should\nsupport in order to profitably learn from the environment.', 'Learning to control an environment without hand-crafted rewards or expert\ndata remains challenging and is at the frontier of reinforcement learning\nresearch. We present an unsupervised learning algorithm to train agents to\nachieve perceptually-specified goals using only a stream of observations and\nactions. Our agent simultaneously learns a goal-conditioned policy and a goal\nachievement reward function that measures how similar a state is to the goal\nstate. This dual optimization leads to a co-operative game, giving rise to a\nlearned reward function that reflects similarity in controllable aspects of the\nenvironment instead of distance in the space of observations. We demonstrate\nthe efficacy of our agent to learn, in an unsupervised manner, to reach a\ndiverse set of goals on three domains -- Atari, the DeepMind Control Suite and\nDeepMind Lab.', 'One question central to Reinforcement Learning is how to learn a feature\nrepresentation that supports algorithm scaling and re-use of learned\ninformation from different tasks. Successor Features approach this problem by\nlearning a feature representation that satisfies a temporal constraint. We\npresent an implementation of an approach that decouples the feature\nrepresentation from the reward function, making it suitable for transferring\nknowledge between domains. We then assess the advantages and limitations of\nusing Successor Features for transfer.', 'Reinforcement learning is a promising approach to developing hard-to-engineer\nadaptive solutions for complex and diverse robotic tasks. However, learning\nwith real-world robots is often unreliable and difficult, which resulted in\ntheir low adoption in reinforcement learning research. This difficulty is\nworsened by the lack of guidelines for setting up learning tasks with robots.\nIn this work, we develop a learning task with a UR5 robotic arm to bring to\nlight some key elements of a task setup and study their contributions to the\nchallenges with robots. We find that learning performance can be highly\nsensitive to the setup, and thus oversights and omissions in setup details can\nmake effective learning, reproducibility, and fair comparison hard. Our study\nsuggests some mitigating steps to help future experimenters avoid difficulties\nand pitfalls. We show that highly reliable and repeatable experiments can be\nperformed in our setup, indicating the possibility of reinforcement learning\nresearch extensively based on real-world robots.', 'In this work we create agents that can perform well beyond a single,\nindividual task, that exhibit much wider generalisation of behaviour to a\nmassive, rich space of challenges. We define a universe of tasks within an\nenvironment domain and demonstrate the ability to train agents that are\ngenerally capable across this vast space and beyond. The environment is\nnatively multi-agent, spanning the continuum of competitive, cooperative, and\nindependent games, which are situated within procedurally generated physical 3D\nworlds. The resulting space is exceptionally diverse in terms of the challenges\nposed to agents, and as such, even measuring the learning progress of an agent\nis an open research problem. We propose an iterative notion of improvement\nbetween successive generations of agents, rather than seeking to maximise a\nsingular objective, allowing us to quantify progress despite tasks being\nincomparable in terms of achievable rewards. We show that through constructing\nan open-ended learning process, which dynamically changes the training task\ndistributions and training objectives such that the agent never stops learning,\nwe achieve consistent learning of new behaviours. The resulting agent is able\nto score reward in every one of our humanly solvable evaluation levels, with\nbehaviour generalising to many held-out points in the universe of tasks.\nExamples of this zero-shot generalisation include good performance on Hide and\nSeek, Capture the Flag, and Tag. Through analysis and hand-authored probe tasks\nwe characterise the behaviour of our agent, and find interesting emergent\nheuristic behaviours such as trial-and-error experimentation, simple tool use,\noption switching, and cooperation. Finally, we demonstrate that the general\ncapabilities of this agent could unlock larger scale transfer of behaviour\nthrough cheap finetuning.', 'An important goal of research in Deep Reinforcement Learning in mobile\nrobotics is to train agents capable of solving complex tasks, which require a\nhigh level of scene understanding and reasoning from an egocentric perspective.\nWhen trained from simulations, optimal environments should satisfy a currently\nunobtainable combination of high-fidelity photographic observations, massive\namounts of different environment configurations and fast simulation speeds. In\nthis paper we argue that research on training agents capable of complex\nreasoning can be simplified by decoupling from the requirement of high fidelity\nphotographic observations. We present a suite of tasks requiring complex\nreasoning and exploration in continuous, partially observable 3D environments.\nThe objective is to provide challenging scenarios and a robust baseline agent\narchitecture that can be trained on mid-range consumer hardware in under 24h.\nOur scenarios combine two key advantages: (i) they are based on a simple but\nhighly efficient 3D environment (ViZDoom) which allows high speed simulation\n(12000fps); (ii) the scenarios provide the user with a range of difficulty\nsettings, in order to identify the limitations of current state of the art\nalgorithms and network architectures. We aim to increase accessibility to the\nfield of Deep-RL by providing baselines for challenging scenarios where new\nideas can be iterated on quickly. We argue that the community should be able to\naddress challenging problems in reasoning of mobile agents without the need for\na large compute infrastructure.', 'Despite the potential of reinforcement learning (RL) for building\ngeneral-purpose robotic systems, training RL agents to solve robotics tasks\nstill remains challenging due to the difficulty of exploration in purely\ncontinuous action spaces. Addressing this problem is an active area of research\nwith the majority of focus on improving RL methods via better optimization or\nmore efficient exploration. An alternate but important component to consider\nimproving is the interface of the RL algorithm with the robot. In this work, we\nmanually specify a library of robot action primitives (RAPS), parameterized\nwith arguments that are learned by an RL policy. These parameterized primitives\nare expressive, simple to implement, enable efficient exploration and can be\ntransferred across robots, tasks and environments. We perform a thorough\nempirical study across challenging tasks in three distinct domains with image\ninput and a sparse terminal reward. We find that our simple change to the\naction interface substantially improves both the learning efficiency and task\nperformance irrespective of the underlying RL algorithm, significantly\noutperforming prior methods which learn skills from offline expert data. Code\nand videos at https://mihdalal.github.io/raps/', 'A practical approach to robot reinforcement learning is to first collect a\nlarge batch of real or simulated robot interaction data, using some data\ncollection policy, and then learn from this data to perform various tasks,\nusing offline learning algorithms. Previous work focused on manually designing\nthe data collection policy, and on tasks where suitable policies can easily be\ndesigned, such as random picking policies for collecting data about object\ngrasping. For more complex tasks, however, it may be difficult to find a data\ncollection policy that explores the environment effectively, and produces data\nthat is diverse enough for the downstream task. In this work, we propose that\ndata collection policies should actively explore the environment to collect\ndiverse data. In particular, we develop a simple-yet-effective goal-conditioned\nreinforcement-learning method that actively focuses data collection on novel\nobservations, thereby collecting a diverse data-set. We evaluate our method on\nsimulated robot manipulation tasks with visual inputs and show that the\nimproved diversity of active data collection leads to significant improvements\nin the downstream learning tasks.', 'While recent progress in deep reinforcement learning has enabled robots to\nlearn complex behaviors, tasks with long horizons and sparse rewards remain an\nongoing challenge. In this work, we propose an effective reward shaping method\nthrough predictive coding to tackle sparse reward problems. By learning\npredictive representations offline and using these representations for reward\nshaping, we gain access to reward signals that understand the structure and\ndynamics of the environment. In particular, our method achieves better learning\nby providing reward signals that 1) understand environment dynamics 2)\nemphasize on features most useful for learning 3) resist noise in learned\nrepresentations through reward accumulation. We demonstrate the usefulness of\nthis approach in different domains ranging from robotic manipulation to\nnavigation, and we show that reward signals produced through predictive coding\nare as effective for learning as hand-crafted rewards.', 'Reinforcement learning is a promising method to accomplish robotic control\ntasks. The task of playing musical instruments is, however, largely unexplored\nbecause it involves the challenge of achieving sequential goals - melodies -\nthat have a temporal dimension. In this paper, we address robotic musicianship\nby introducing a temporal extension to goal-conditioned reinforcement learning:\nTime-dependent goals. We demonstrate that these can be used to train a robotic\nmusician to play the theremin instrument. We train the robotic agent in\nsimulation and transfer the acquired policy to a real-world robotic\nthereminist. Supplemental video: https://youtu.be/jvC9mPzdQN4', 'Mastering robotic manipulation skills through reinforcement learning (RL)\ntypically requires the design of shaped reward functions. Recent developments\nin this area have demonstrated that using sparse rewards, i.e. rewarding the\nagent only when the task has been successfully completed, can lead to better\npolicies. However, state-action space exploration is more difficult in this\ncase. Recent RL approaches to learning with sparse rewards have leveraged\nhigh-quality human demonstrations for the task, but these can be costly, time\nconsuming or even impossible to obtain. In this paper, we propose a novel and\neffective approach that does not require human demonstrations. We observe that\nevery robotic manipulation task could be seen as involving a locomotion task\nfrom the perspective of the object being manipulated, i.e. the object could\nlearn how to reach a target state on its own. In order to exploit this idea, we\nintroduce a framework whereby an object locomotion policy is initially obtained\nusing a realistic physics simulator. This policy is then used to generate\nauxiliary rewards, called simulated locomotion demonstration rewards (SLDRs),\nwhich enable us to learn the robot manipulation policy. The proposed approach\nhas been evaluated on 13 tasks of increasing complexity, and can achieve higher\nsuccess rate and faster learning rates compared to alternative algorithms.\nSLDRs are especially beneficial for tasks like multi-object stacking and\nnon-rigid object manipulation.', 'The application of reinforcement learning (RL) in robotic control is still\nlimited in the environments with sparse and delayed rewards. In this paper, we\npropose a practical self-imitation learning method named Self-Imitation\nLearning with Constant Reward (SILCR). Instead of requiring hand-defined\nimmediate rewards from environments, our method assigns the immediate rewards\nat each timestep with constant values according to their final episodic\nrewards. In this way, even if the dense rewards from environments are\nunavailable, every action taken by the agents would be guided properly. We\ndemonstrate the effectiveness of our method in some challenging continuous\nrobotics control tasks in MuJoCo simulation and the results show that our\nmethod significantly outperforms the alternative methods in tasks with sparse\nand delayed rewards. Even compared with alternatives with dense rewards\navailable, our method achieves competitive performance. The ablation\nexperiments also show the stability and reproducibility of our method.', 'In recent years, a variety of tasks have been accomplished by deep\nreinforcement learning (DRL). However, when applying DRL to tasks in a\nreal-world environment, designing an appropriate reward is difficult. Rewards\nobtained via actual hardware sensors may include noise, misinterpretation, or\nfailed observations. The learning instability caused by these unstable signals\nis a problem that remains to be solved in DRL. In this work, we propose an\napproach that extends existing DRL models by adding a subtask to directly\nestimate the variance contained in the reward signal. The model then takes the\nfeature map learned by the subtask in a critic network and sends it to the\nactor network. This enables stable learning that is robust to the effects of\npotential noise. The results of experiments in the Atari game domain with\nunstable reward signals show that our method stabilizes training convergence.\nWe also discuss the extensibility of the model by visualizing feature maps.\nThis approach has the potential to make DRL more practical for use in noisy,\nreal-world scenarios.', 'Transfer learning is an important new subfield of multiagent reinforcement\nlearning that aims to help an agent learn about a problem by using knowledge\nthat it has gained solving another problem, or by using knowledge that is\ncommunicated to it by an agent who already knows the problem. This is useful\nwhen one wishes to change the architecture or learning algorithm of an agent\n(so that the new knowledge need not be built ""from scratch""), when new agents\nare frequently introduced to the environment with no knowledge, or when an\nagent must adapt to similar but different problems. Great progress has been\nmade in the agent-to-agent case using the Teacher/Student framework proposed by\n(Torrey and Taylor 2013). However, that approach requires that learning from a\nteacher be treated differently from learning in every other reinforcement\nlearning context. In this paper, I propose a method which allows the\nteacher/student framework to be applied in a way that fits directly and\nnaturally into the more general reinforcement learning framework by integrating\nthe teacher feedback into the reward signal received by the learning agent. I\nshow that this approach can significantly improve the rate of learning for an\nagent playing a one-player stochastic game; I give examples of potential\npitfalls of the approach; and I propose further areas of research building on\nthis framework.', 'Humans continually expand their learned knowledge to new domains and learn\nnew concepts without any interference with past learned experiences. In\ncontrast, machine learning models perform poorly in a continual learning\nsetting, where input data distribution changes over time. Inspired by the\nnervous system learning mechanisms, we develop a computational model that\nenables a deep neural network to learn new concepts and expand its learned\nknowledge to new domains incrementally in a continual learning setting. We rely\non the Parallel Distributed Processing theory to encode abstract concepts in an\nembedding space in terms of a multimodal distribution. This embedding space is\nmodeled by internal data representations in a hidden network layer. We also\nleverage the Complementary Learning Systems theory to equip the model with a\nmemory mechanism to overcome catastrophic forgetting through implementing\npseudo-rehearsal. Our model can generate pseudo-data points for experience\nreplay and accumulate new experiences to past learned experiences without\ncausing cross-task interference.', 'A common approach to solving physical reasoning tasks is to train a value\nlearner on example tasks. A limitation of such an approach is that it requires\nlearning about object dynamics solely from reward values assigned to the final\nstate of a rollout of the environment. This study aims to address this\nlimitation by augmenting the reward value with self-supervised signals about\nobject dynamics. Specifically, we train the model to characterize the\nsimilarity of two environment rollouts, jointly with predicting the outcome of\nthe reasoning task. This similarity can be defined as a distance measure\nbetween the trajectory of objects in the two rollouts, or learned directly from\npixels using a contrastive formulation. Empirically, we find that this approach\nleads to substantial performance improvements on the PHYRE benchmark for\nphysical reasoning (Bakhtin et al., 2019), establishing a new state-of-the-art.', ""We study the problem of designing autonomous agents that can learn to\ncooperate effectively with a potentially suboptimal partner while having no\naccess to the joint reward function. This problem is modeled as a cooperative\nepisodic two-agent Markov decision process. We assume control over only the\nfirst of the two agents in a Stackelberg formulation of the game, where the\nsecond agent is acting so as to maximise expected utility given the first\nagent's policy. How should the first agent act in order to learn the joint\nreward function as quickly as possible and so that the joint policy is as close\nto optimal as possible? We analyse how knowledge about the reward function can\nbe gained in this interactive two-agent scenario. We show that when the\nlearning agent's policies have a significant effect on the transition function,\nthe reward function can be learned efficiently."", 'Reinforcement learning algorithms can train agents that solve problems in\ncomplex, interesting environments. Normally, the complexity of the trained\nagent is closely related to the complexity of the environment. This suggests\nthat a highly capable agent requires a complex environment for training. In\nthis paper, we point out that a competitive multi-agent environment trained\nwith self-play can produce behaviors that are far more complex than the\nenvironment itself. We also point out that such environments come with a\nnatural curriculum, because for any skill level, an environment full of agents\nof this level will have the right level of difficulty. This work introduces\nseveral competitive multi-agent environments where agents compete in a 3D world\nwith simulated physics. The trained agents learn a wide variety of complex and\ninteresting skills, even though the environment themselves are relatively\nsimple. The skills include behaviors such as running, blocking, ducking,\ntackling, fooling opponents, kicking, and defending using both arms and legs. A\nhighlight of the learned behaviors can be found here: https://goo.gl/eR7fbX', 'A key problem in structured output prediction is direct optimization of the\ntask reward function that matters for test evaluation. This paper presents a\nsimple and computationally efficient approach to incorporate task reward into a\nmaximum likelihood framework. By establishing a link between the log-likelihood\nand expected reward objectives, we show that an optimal regularized expected\nreward is achieved when the conditional distribution of the outputs given the\ninputs is proportional to their exponentiated scaled rewards. Accordingly, we\npresent a framework to smooth the predictive probability of the outputs using\ntheir corresponding rewards. We optimize the conditional log-probability of\naugmented outputs that are sampled proportionally to their exponentiated scaled\nrewards. Experiments on neural sequence to sequence models for speech\nrecognition and machine translation show notable improvements over a maximum\nlikelihood baseline by using reward augmented maximum likelihood (RAML), where\nthe rewards are defined as the negative edit distance between the outputs and\nthe ground truth labels.', 'Many real-world robotic operations that involve high-dimensional humanoid\nrobots require fast-reaction to plan disturbances and probabilistic guarantees\nover collision risks, whereas most probabilistic motion planning approaches\ndeveloped for car-like robots can not be directly applied to high-dimensional\nrobots. In this paper, we present probabilistic Chekov (p-Chekov), a\nfast-reactive motion planning system that can provide safety guarantees for\nhigh-dimensional robots suffering from process noises and observation noises.\nLeveraging recent advances in machine learning as well as our previous work in\ndeterministic motion planning that integrated trajectory optimization into a\nsparse roadmap framework, p-Chekov demonstrates its superiority in terms of\ncollision avoidance ability and planning speed in high-dimensional robotic\nmotion planning tasks in complex environments without the convexification of\nobstacles. Comprehensive theoretical and empirical analysis provided in this\npaper shows that p-Chekov can effectively satisfy user-specified chance\nconstraints over collision risk in practical robotic manipulation tasks.', 'For a multi-robot system equipped with heterogeneous capabilities, this paper\npresents a mechanism to allocate robots to tasks in a resilient manner when\nanomalous environmental conditions such as weather events or adversarial\nattacks affect the performance of robots within the tasks. Our primary\nobjective is to ensure that each task is assigned the requisite level of\nresources, measured as the aggregated capabilities of the robots allocated to\nthe task. By keeping track of task performance deviations under external\nperturbations, our framework quantifies the extent to which robot capabilities\n(e.g., visual sensing or aerial mobility) are affected by environmental\nconditions. This enables an optimization-based framework to flexibly reallocate\nrobots to tasks based on the most degraded capabilities within each task. In\nthe face of resource limitations and adverse environmental conditions, our\nalgorithm minimally relaxes the resource constraints corresponding to some\ntasks, thus exhibiting a graceful degradation of performance. Simulated\nexperiments in a multi-robot coverage and target tracking scenario demonstrate\nthe efficacy of the proposed approach.', 'Mobile robots, performing long-term manipulation activities in human\nenvironments, have to perceive a wide variety of objects possessing very\ndifferent visual characteristics and need to reliably keep track of these\nthroughout the execution of a task. In order to be efficient, robot perception\ncapabilities need to go beyond what is currently perceivable and should be able\nto answer queries about both current and past scenes. In this paper we\ninvestigate a perception system for long-term robot manipulation that keeps\ntrack of the changing environment and builds a representation of the perceived\nworld. Specifically we introduce an amortized component that spreads perception\ntasks throughout the execution cycle. The resulting query driven perception\nsystem asynchronously integrates results from logged images into a symbolic and\nnumeric (what we call sub-symbolic) representation that forms the perceptual\nbelief state of the robot.', 'We describe a simple scheme that allows an agent to learn about its\nenvironment in an unsupervised manner. Our scheme pits two versions of the same\nagent, Alice and Bob, against one another. Alice proposes a task for Bob to\ncomplete; and then Bob attempts to complete the task. In this work we will\nfocus on two kinds of environments: (nearly) reversible environments and\nenvironments that can be reset. Alice will ""propose"" the task by doing a\nsequence of actions and then Bob must undo or repeat them, respectively. Via an\nappropriate reward structure, Alice and Bob automatically generate a curriculum\nof exploration, enabling unsupervised training of the agent. When Bob is\ndeployed on an RL task within the environment, this unsupervised training\nreduces the number of supervised episodes needed to learn, and in some cases\nconverges to a higher reward.', 'Inverse reinforcement learning attempts to reconstruct the reward function in\na Markov decision problem, using observations of agent actions. As already\nobserved in Russell [1998] the problem is ill-posed, and the reward function is\nnot identifiable, even under the presence of perfect information about optimal\nbehavior. We provide a resolution to this non-identifiability for problems with\nentropy regularization. For a given environment, we fully characterize the\nreward functions leading to a given policy and demonstrate that, given\ndemonstrations of actions for the same reward under two distinct discount\nfactors, or under sufficiently different environments, the unobserved reward\ncan be recovered up to a constant. We also give general necessary and\nsufficient conditions for reconstruction of time-homogeneous rewards on finite\nhorizons, and for action-independent rewards, generalizing recent results of\nKim et al. [2021] and Fu et al. [2018].', 'It has been postulated that a good representation is one that disentangles\nthe underlying explanatory factors of variation. However, it remains an open\nquestion what kind of training framework could potentially achieve that.\nWhereas most previous work focuses on the static setting (e.g., with images),\nwe postulate that some of the causal factors could be discovered if the learner\nis allowed to interact with its environment. The agent can experiment with\ndifferent actions and observe their effects. More specifically, we hypothesize\nthat some of these factors correspond to aspects of the environment which are\nindependently controllable, i.e., that there exists a policy and a learnable\nfeature for each such aspect of the environment, such that this policy can\nyield changes in that feature with minimal changes to other features that\nexplain the statistical variations in the observed data. We propose a specific\nobjective function to find such factors, and verify experimentally that it can\nindeed disentangle independently controllable aspects of the environment\nwithout any extrinsic reward signal.', ""Robotic grasping is a crucial area of research as it can result in the\nacceleration of the automation of several Industries utilizing robots ranging\nfrom manufacturing to healthcare. Reinforcement learning is the field of study\nwhere an agent learns a policy to execute an action by exploring and exploiting\nrewards from an environment. Reinforcement learning can thus be used by the\nagent to learn how to execute a certain task, in our case grasping an object.\nWe have used the Pick and Place environment provided by OpenAI's Gym to\nengineer rewards. Hindsight Experience Replay (HER) has shown promising results\nwith problems having a sparse reward. In the default configuration of the\nOpenAI baseline and environment the reward function is calculated using the\ndistance between the target location and the robot end-effector. By weighting\nthe cost based on the distance of the end-effector from the goal in the x,y and\nz-axes we were able to almost halve the learning time compared to the baselines\nprovided by OpenAI, an intuitive strategy that further reduced learning time.\nIn this project, we were also able to introduce certain user desired\ntrajectories in the learnt policies (city-block / Manhattan trajectories). This\nhelps us understand that by engineering the rewards we can tune the agent to\nlearn policies in a certain way even if it might not be the most optimal but is\nthe desired manner."", 'In cooperative multi-agent reinforcement learning (MARL), how to design a\nsuitable reward signal to accelerate learning and stabilize convergence is a\ncritical problem. The global reward signal assigns the same global reward to\nall agents without distinguishing their contributions, while the local reward\nsignal provides different local rewards to each agent based solely on\nindividual behavior. Both of the two reward assignment approaches have some\nshortcomings: the former might encourage lazy agents, while the latter might\nproduce selfish agents.\n  In this paper, we study reward design problem in cooperative MARL based on\npacket routing environments. Firstly, we show that the above two reward signals\nare prone to produce suboptimal policies. Then, inspired by some observations\nand considerations, we design some mixed reward signals, which are\noff-the-shelf to learn better policies. Finally, we turn the mixed reward\nsignals into the adaptive counterparts, which achieve best results in our\nexperiments. Other reward signals are also discussed in this paper. As reward\ndesign is a very fundamental problem in RL and especially in MARL, we hope that\nMARL researchers can rethink the rewards used in their systems.', 'Legg and Hutter, as well as subsequent authors, considered intelligent agents\nthrough the lens of interaction with reward-giving environments, attempting to\nassign numeric intelligence measures to such agents, with the guiding principle\nthat a more intelligent agent should gain higher rewards from environments in\nsome aggregate sense. In this paper, we consider a related question: rather\nthan measure numeric intelligence of one Legg- Hutter agent, how can we compare\nthe relative intelligence of two Legg-Hutter agents? We propose an elegant\nanswer based on the following insight: we can view Legg-Hutter agents as\ncandidates in an election, whose voters are environments, letting each\nenvironment vote (via its rewards) which agent (if either) is more intelligent.\nThis leads to an abstract family of comparators simple enough that we can prove\nsome structural theorems about them. It is an open question whether these\nstructural theorems apply to more practical intelligence measures.', 'With the incremental development of robotic platforms to automate the manual\nprocesses, path planning has become a critical domain with or without the\nknowledge of the indoor and outdoor environment. The algorithms can be\nintelligent or pre-structured and should optimally reach the destination\nefficiently. The major challenge in this domain is to find a path which is free\nfrom static obstacles as well as dynamic obstacles. In this paper, a\nmethodology is proposed with the implementation details of the robotic platform\nto cover the critical key points and to arrive at the original key point in a\ndynamic environment. The main computation is happening inside a Raspberry Pi B+\nmodule, and compass, wheel encoders and ultrasonic sensors were used in the\nimplementation for the localization of the robot to relevant key points.', 'Our world is open-ended, non-stationary, and constantly evolving; thus what\nwe talk about and how we talk about it change over time. This inherent dynamic\nnature of language contrasts with the current static language modelling\nparadigm, which trains and evaluates models on utterances from overlapping time\nperiods. Despite impressive recent progress, we demonstrate that Transformer-XL\nlanguage models perform worse in the realistic setup of predicting future\nutterances from beyond their training period, and that model performance\nbecomes increasingly worse with time. We find that, while increasing model size\nalone -- a key driver behind recent progress -- does not solve this problem,\nhaving models that continually update their knowledge with new information can\nindeed mitigate this performance degradation over time. Hence, given the\ncompilation of ever-larger language modelling datasets, combined with the\ngrowing list of language-model-based NLP applications that require up-to-date\nfactual knowledge about the world, we argue that now is the right time to\nrethink the static way in which we currently train and evaluate our language\nmodels, and develop adaptive language models that can remain up-to-date with\nrespect to our ever-changing and non-stationary world. We publicly release our\ndynamic, streaming language modelling benchmarks for WMT and arXiv to\nfacilitate language model evaluation that takes temporal dynamics into account.', 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', 'Reinforcement Learning (RL) agents require the specification of a reward\nsignal for learning behaviours. However, introduction of corrupt or stochastic\nrewards can yield high variance in learning. Such corruption may be a direct\nresult of goal misspecification, randomness in the reward signal, or\ncorrelation of the reward with external factors that are not known to the\nagent. Corruption or stochasticity of the reward signal can be especially\nproblematic in robotics, where goal specification can be particularly difficult\nfor complex tasks. While many variance reduction techniques have been studied\nto improve the robustness of the RL process, handling such stochastic or\ncorrupted reward structures remains difficult. As an alternative for handling\nthis scenario in model-free RL methods, we suggest using an estimator for both\nrewards and value functions. We demonstrate that this improves performance\nunder corrupted stochastic rewards in both the tabular and non-linear function\napproximation settings for a variety of noise types and environments. The use\nof reward estimation is a robust and easy-to-implement improvement for handling\ncorrupted reward signals in model-free RL.', ""This paper discusses recent research progress in robotic grasping and\nmanipulation in the light of the latest Robotic Grasping and Manipulation\nCompetitions (RGMCs). We first provide an overview of past benchmarks and\ncompetitions related to the robotics manipulation field. Then, we discuss the\nmethodology behind designing the manipulation tasks in RGMCs. We provide a\ndetailed analysis of key challenges for each task and identify the most\ndifficult aspects based on the competing teams' performance in recent years. We\nbelieve that such an analysis is insightful to determine the future research\ndirections for the robotic manipulation domain."", 'We update the method of describing and assessing the process of the study of\nan abstract environment by a system, proposed earlier. We do not model any\nbiological cognition mechanisms and consider the system as an agent equipped\nwith an information processor (or a group of such agents), which makes a move\nin the environment, consumes information supplied by the environment, and gives\nout the next move (hence, the process is considered as a game). The system\nmoves in an unknown environment and should recognize new objects located in it.\nIn this case, the system should build comprehensive images of visible things\nand memorize them if necessary (and it should also choose the current goal\nset). The main problems here are object recognition, and the informational\nreward rating in the game. Thus, the main novelty of the paper is a new method\nof evaluating the amount of visual information about the object as the reward.\nIn such a system, we suggest using a minimally pre-trained neural network to be\nresponsible for the recognition: at first, we train the network only for\nBiederman geons (geometrical primitives). The geons are generated\nprogrammatically and we demonstrate that such a trained network recognizes\ngeons in real objects quite well. We also offer to generate, procedurally, new\nobjects from geon schemes (geon combinations in images) obtained from the\nenvironment and to store them in a database. In this case, we do not obtain new\ninformation about an object (i.e., our reward is maximal, thus the game and the\nobject cognition process stop) when we stop getting new schemes of this kind.\nThese schemes are generated from geons connected with the object. In the case\nof a possibly known item, the informational reward is maximal when we have no\nmore detection uncertainty for any of the objects.', 'Recently we see a rising number of methods in the field of eXplainable\nArtificial Intelligence. To our surprise, their development is driven by model\ndevelopers rather than a study of needs for human end users. The analysis of\nneeds, if done, takes the form of an A/B test rather than a study of open\nquestions. To answer the question ""What would a human operator like to ask the\nML model?"" we propose a conversational system explaining decisions of the\npredictive model. In this experiment, we developed a chatbot called dr_ant to\ntalk about machine learning model trained to predict survival odds on Titanic.\nPeople can talk with dr_ant about different aspects of the model to understand\nthe rationale behind its predictions. Having collected a corpus of 1000+\ndialogues, we analyse the most common types of questions that users would like\nto ask. To our knowledge, it is the first study which uses a conversational\nsystem to collect the needs of human operators from the interactive and\niterative dialogue explorations of a predictive model.', 'We present the process of transitioning machine learning models to the\nTensorFlow framework at a large scale in an online advertising ecosystem. In\nthis talk we address the key challenges we faced and describe how we\nsuccessfully tackled them; notably, implementing the models in TF and serving\nthem efficiently with low latency using various optimization techniques.', 'For the past few decades, man has been trying to create an intelligent\ncomputer which can talk and respond like he can. The task of creating a system\nthat can talk like a human being is the primary objective of Automatic Speech\nRecognition. Various Speech Recognition techniques have been developed in\ntheory and have been applied in practice. This paper discusses the problems\nthat have been encountered in developing Speech Recognition, the techniques\nthat have been applied to automate the task, and a representation of the core\nproblems of present day Speech Recognition by using Fuzzy Mathematics.', 'This paper presents a plausible reasoning system to illustrate some broad\nissues in knowledge representation: dualities between different reasoning\nforms, the difficulty of unifying complementary reasoning styles, and the\napproximate nature of plausible reasoning. These issues have a common\nunderlying theme: there should be an underlying belief calculus of which the\nmany different reasoning forms are special cases, sometimes approximate. The\nsystem presented allows reasoning about defaults, likelihood, necessity and\npossibility in a manner similar to the earlier work of Adams. The system is\nbased on the belief calculus of subjective Bayesian probability which itself is\nbased on a few simple assumptions about how belief should be manipulated.\nApproximations, semantics, consistency and consequence results are presented\nfor the system. While this puts these often discussed plausible reasoning forms\non a probabilistic footing, useful application to practical problems remains an\nissue.', ""This paper presents an acoustic echo canceler based on a U-Net convolutional\nneural network for single-talk and double-talk scenarios. U-Net networks have\npreviously been used in the audio processing area for source separation\nproblems because of their ability to reproduce the finest details of audio\nsignals, but to our knowledge, this is the first time they have been used for\nacoustic echo cancellation (AEC). The U-Net hyperparameters have been optimized\nto obtain the best AEC performance, but using a reduced number of parameters to\nmeet a latency restriction of 40 ms. The training and testing of our model have\nbeen carried out within the framework of the 'ICASSP 2021 AEC Challenge'\norganized by Microsoft. We have trained the optimized U-Net model with a\nsynthetic dataset only (S-U-Net) and with a synthetic dataset and the\nsingle-talk set of a real dataset (SR-U-Net), both datasets were released for\nthe challenge. The S-U-Net model presented better results for double-talk\nscenarios, thus their inferred near-end signals from the blind testset were\nsubmitted to the challenge. Our canceler ranked 12th among 17 teams, and 5th\namong 10 academia teams, obtaining an overall mean opinion score of 3.57."", 'Two interpretations about syllogistic statements are described in this paper.\nOne is the so-called set-based interpretation, which assumes that quantified\nstatements and syllogisms talk about quantity-relationships between sets. The\nother one, the so-called conditional interpretation, assumes that quantified\npropositions talk about conditional propositions and how strong are the links\nbetween the antecedent and the consequent. Both interpretations are compared\nattending to three different questions (existential import, singular statements\nand non-proportional quantifiers) from the point of view of their impact on the\nfurther development of this type of reasoning.', ""There is knowledge. There is belief. And there is tacit agreement.' 'We may\ntalk about objects. We may talk about attributes of the objects. Or we may talk\nboth about objects and their attributes.' This work inspects tacit agreements\non assumptions about the relation between objects and their attributes, and\nstudies a way of expressing them, presenting as the result what we term gradual\nlogic in which the sense of truth gradually shifts. It extends classical logic\ninstances with a new logical connective capturing the object-attribute\nrelation. A formal semantics is presented. Decidability is proved. Para-\nconsistent/epistemic/conditional/intensional/description/combined logics are\ncompared.""]"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I aim to enhance the expressive power of Graph Neural Networks (GNNs) to differentiate between graphs that are structurally different but possess the same local neighborhood structure, a task they traditionally struggle with. To achieve this, I plan to develop structure-related features that can aid GNNs in representing any set of nodes, thereby increasing their expressive power. Could you suggest some standard graph-distance measures, such as the shortest path distance, that could be used to capture the distance between the node set and each individual node in the graph? Furthermore, I want these features to enhance the ability of GNN nodes to gather information from neighboring nodes for self-updating. Importantly, these features should leverage the sparse structure of the underlying graph, ensuring computational efficiency and scalability. Finally, I intend to demonstrate the advantages of my proposed features by testing them on standard network tasks, such as structural role prediction. Could you also suggest other tasks where these features could be tested?","['Recently, the Weisfeiler-Lehman (WL) graph isomorphism test was used to\nmeasure the expressive power of graph neural networks (GNN). It was shown that\nthe popular message passing GNN cannot distinguish between graphs that are\nindistinguishable by the 1-WL test (Morris et al. 2018; Xu et al. 2019).\nUnfortunately, many simple instances of graphs are indistinguishable by the\n1-WL test.\n  In search for more expressive graph learning models we build upon the recent\nk-order invariant and equivariant graph neural networks (Maron et al. 2019a,b)\nand present two results:\n  First, we show that such k-order networks can distinguish between\nnon-isomorphic graphs as good as the k-WL tests, which are provably stronger\nthan the 1-WL test for k>2. This makes these models strictly stronger than\nmessage passing models. Unfortunately, the higher expressiveness of these\nmodels comes with a computational cost of processing high order tensors.\n  Second, setting our goal at building a provably stronger, simple and scalable\nmodel we show that a reduced 2-order network containing just scaled identity\noperator, augmented with a single quadratic operation (matrix multiplication)\nhas a provable 3-WL expressive power. Differently put, we suggest a simple\nmodel that interleaves applications of standard Multilayer-Perceptron (MLP)\napplied to the feature dimension and matrix multiplication. We validate this\nmodel by presenting state of the art results on popular graph classification\nand regression tasks. To the best of our knowledge, this is the first practical\ninvariant/equivariant model with guaranteed 3-WL expressiveness, strictly\nstronger than message passing models.', ""Learning node embeddings that capture a node's position within the broader\ngraph structure is crucial for many prediction tasks on graphs. However,\nexisting Graph Neural Network (GNN) architectures have limited power in\ncapturing the position/location of a given node with respect to all other nodes\nof the graph. Here we propose Position-aware Graph Neural Networks (P-GNNs), a\nnew class of GNNs for computing position-aware node embeddings. P-GNN first\nsamples sets of anchor nodes, computes the distance of a given target node to\neach anchor-set,and then learns a non-linear distance-weighted aggregation\nscheme over the anchor-sets. This way P-GNNs can capture positions/locations of\nnodes with respect to the anchor nodes. P-GNNs have several advantages: they\nare inductive, scalable,and can incorporate node feature information. We apply\nP-GNNs to multiple prediction tasks including link prediction and community\ndetection. We show that P-GNNs consistently outperform state of the art GNNs,\nwith up to 66% improvement in terms of the ROC AUC score."", 'The expressive power of message passing neural networks (MPNNs) is known to\nmatch the expressive power of the 1-dimensional Weisfeiler-Leman graph (1-WL)\nisomorphism test. To boost the expressive power of MPNNs, a number of graph\nneural network architectures have recently been proposed based on\nhigher-dimensional Weisfeiler-Leman tests. In this paper we consider the\ntwo-dimensional (2-WL) test and introduce a new type of MPNNs, referred to as\n$\\ell$-walk MPNNs, which aggregate features along walks of length $\\ell$\nbetween vertices. We show that $2$-walk MPNNs match 2-WL in expressive power.\nMore generally, $\\ell$-walk MPNNs, for any $\\ell\\geq 2$, are shown to match the\nexpressive power of the recently introduced $\\ell$-walk refinement procedure\n(W[$\\ell$]). Based on a correspondence between 2-WL and W[$\\ell$], we observe\nthat $\\ell$-walk MPNNs and $2$-walk MPNNs have the same expressive power, i.e.,\nthey can distinguish the same pairs of graphs, but $\\ell$-walk MPNNs can\npossibly distinguish pairs of graphs faster than $2$-walk MPNNs. When it comes\nto concrete learnable graph neural network (GNN) formalisms that match 2-WL or\nW[$\\ell$] in expressive power, we consider second-order graph neural networks\nthat allow for non-linear layers. In particular, to match W[$\\ell$] in\nexpressive power, we allow $\\ell-1$ matrix multiplications in each layer. We\npropose different versions of second-order GNNs depending on the type of\nfeatures (i.e., coming from a countable set, or coming from an uncountable set)\nas this affects the number of dimensions needed to represent the features. Our\nresults indicate that increasing non-linearity in layers by means of allowing\nmultiple matrix multiplications does not increase expressive power. At the very\nbest, it results in a faster distinction of input graphs.', ""Graph Neural Networks (GNNs) bring the power of deep representation learning\nto graph and relational data and achieve state-of-the-art performance in many\napplications. GNNs compute node representations by taking into account the\ntopology of the node's ego-network and the features of the ego-network's nodes.\nWhen the nodes do not have high-quality features, GNNs learn an embedding layer\nto compute node embeddings and use them as input features. However, the size of\nthe embedding layer is linear to the product of the number of nodes in the\ngraph and the dimensionality of the embedding and does not scale to big data\nand graphs with hundreds of millions of nodes. To reduce the memory associated\nwith this embedding layer, hashing-based approaches, commonly used in\napplications like NLP and recommender systems, can potentially be used.\nHowever, a direct application of these ideas fails to exploit the fact that in\nmany real-world graphs, nodes that are topologically close will tend to be\nrelated to each other (homophily) and as such their representations will be\nsimilar.\n  In this work, we present approaches that take advantage of the nodes'\nposition in the graph to dramatically reduce the memory required, with minimal\nif any degradation in the quality of the resulting GNN model. Our approaches\ndecompose a node's embedding into two components: a position-specific component\nand a node-specific component. The position-specific component models homophily\nand the node-specific component models the node-to-node variation. Extensive\nexperiments using different datasets and GNN models show that our methods are\nable to reduce the memory requirements by 88% to 97% while achieving, in nearly\nall cases, better classification accuracy than other competing approaches,\nincluding the full embeddings."", ""Many tasks that rely on representations of nodes in graphs would benefit if\nthose representations were faithful to distances between nodes in the graph.\nGeometric techniques to extract such representations have poor scaling over\nlarge graph size, and recent advances in Graph Neural Network (GNN) algorithms\nhave limited ability to reflect graph distance information beyond the first\ndegree neighborhood. To enable this highly desired capability, we propose a\ntechnique to learn Isometric Graph Neural Networks (IGNN), which requires\nchanging the input representation space and loss function to enable any GNN\nalgorithm to generate representations that reflect distances between nodes. We\nexperiment with the isometric technique on several GNN architectures for\nmodeling multiple prediction tasks on multiple datasets. In addition to an\nimprovement in AUC-ROC as high as $43\\%$ in these experiments, we observe a\nconsistent and substantial improvement as high as 400% in Kendall's Tau (KT), a\nmeasure that directly reflects distance information, demonstrating that the\nlearned embeddings do account for graph distances."", ""Message-passing neural networks (MPNNs) are the leading architecture for deep\nlearning on graph-structured data, in large part due to their simplicity and\nscalability. Unfortunately, it was shown that these architectures are limited\nin their expressive power. This paper proposes a novel framework called\nEquivariant Subgraph Aggregation Networks (ESAN) to address this issue. Our\nmain observation is that while two graphs may not be distinguishable by an\nMPNN, they often contain distinguishable subgraphs. Thus, we propose to\nrepresent each graph as a set of subgraphs derived by some predefined policy,\nand to process it using a suitable equivariant architecture. We develop novel\nvariants of the 1-dimensional Weisfeiler-Leman (1-WL) test for graph\nisomorphism, and prove lower bounds on the expressiveness of ESAN in terms of\nthese new WL variants. We further prove that our approach increases the\nexpressive power of both MPNNs and more expressive architectures. Moreover, we\nprovide theoretical results that describe how design choices such as the\nsubgraph selection policy and equivariant neural architecture affect our\narchitecture's expressive power. To deal with the increased computational cost,\nwe propose a subgraph sampling scheme, which can be viewed as a stochastic\nversion of our framework. A comprehensive set of experiments on real and\nsynthetic datasets demonstrates that our framework improves the expressive\npower and overall performance of popular GNN architectures."", 'Graph Neural Network (GNN) aggregates the neighborhood of each node into the\nnode embedding and shows its powerful capability for graph representation\nlearning. However, most existing GNN variants aggregate the neighborhood\ninformation in a fixed non-injective fashion, which may map different graphs or\nnodes to the same embedding, reducing the model expressiveness. We present a\ntheoretical framework to design a continuous injective set function for\nneighborhood aggregation in GNN. Using the framework, we propose expressive GNN\nthat aggregates the neighborhood of each node with a continuous injective set\nfunction, so that a GNN layer maps similar nodes with similar neighborhoods to\nsimilar embeddings, different nodes to different embeddings and the equivalent\nnodes or isomorphic graphs to the same embeddings. Moreover, the proposed\nexpressive GNN can naturally learn expressive representations for graphs with\ncontinuous node attributes. We validate the proposed expressive GNN (ExpGNN)\nfor graph classification on multiple benchmark datasets including simple graphs\nand attributed graphs. The experimental results demonstrate that our model\nachieves state-of-the-art performances on most of the benchmarks.', ""Numerous works have proven that existing neighbor-averaging Graph Neural\nNetworks cannot efficiently catch structure features, and many works show that\ninjecting structure, distance, position or spatial features can significantly\nimprove performance of GNNs, however, injecting overall structure and distance\ninto GNNs is an intuitive but remaining untouched idea. In this work, we shed\nlight on the direction. We first extracting hop-wise structure information and\ncompute distance distributional information, gathering with node's intrinsic\nfeatures, embedding them into same vector space and then adding them up. The\nderived embedding vectors are then fed into GATs(like GAT, AGDN) and then\nCorrect and Smooth, experiments show that the DHSEGATs achieve competitive\nresult. The code is available at https://github.com/hzg0601/DHSEGATs."", ""Graph neural networks (GNNs) have emerged recently as a powerful architecture\nfor learning node and graph representations. Standard GNNs have the same\nexpressive power as the Weisfeiler-Leman test of graph isomorphism in terms of\ndistinguishing non-isomorphic graphs. However, it was recently shown that this\ntest cannot identify fundamental graph properties such as connectivity and\ntriangle freeness. We show that GNNs also suffer from the same limitation. To\naddress this limitation, we propose a more expressive architecture, k-hop GNNs,\nwhich updates a node's representation by aggregating information not only from\nits direct neighbors, but from its k-hop neighborhood. We show that the\nproposed architecture can identify fundamental graph properties. We evaluate\nthe proposed architecture on standard node classification and graph\nclassification datasets. Our experimental evaluation confirms our theoretical\nfindings since the proposed model achieves performance better or comparable to\nstandard GNNs and to state-of-the-art algorithms."", 'We present our ongoing work on understanding the limitations of graph\nconvolutional networks (GCNs) as well as our work on generalizations of graph\nconvolutions for representing more complex node attribute dependencies. Based\non an analysis of GCNs with the help of the corresponding computation graphs,\nwe propose a generalization of existing GCNs where the aggregation operations\nare (a) determined by structural properties of the local neighborhood graphs\nand (b) not restricted to weighted averages. We show that the proposed approach\nis strictly more expressive while requiring only a modest increase in the\nnumber of parameters and computations. We also show that the proposed\ngeneralization is identical to standard convolutional layers when applied to\nregular grid graphs.', 'The Transformer architecture has become a dominant choice in many domains,\nsuch as natural language processing and computer vision. Yet, it has not\nachieved competitive performance on popular leaderboards of graph-level\nprediction compared to mainstream GNN variants. Therefore, it remains a mystery\nhow Transformers could perform well for graph representation learning. In this\npaper, we solve this mystery by presenting Graphormer, which is built upon the\nstandard Transformer architecture, and could attain excellent results on a\nbroad range of graph representation learning tasks, especially on the recent\nOGB Large-Scale Challenge. Our key insight to utilizing Transformer in the\ngraph is the necessity of effectively encoding the structural information of a\ngraph into the model. To this end, we propose several simple yet effective\nstructural encoding methods to help Graphormer better model graph-structured\ndata. Besides, we mathematically characterize the expressive power of\nGraphormer and exhibit that with our ways of encoding the structural\ninformation of graphs, many popular GNN variants could be covered as the\nspecial cases of Graphormer.', 'Graph Neural Networks (GNNs) are versatile, powerful machine learning methods\nthat enable graph structure and feature representation learning, and have\napplications across many domains. For applications critically requiring\ninterpretation, attention-based GNNs have been leveraged. However, these\napproaches either rely on specific model architectures or lack a joint\nconsideration of graph structure and node features in their interpretation.\nHere we present a model-agnostic framework for interpreting important graph\nstructure and node features, Graph neural networks Including SparSe\ninTerpretability (GISST). With any GNN model, GISST combines an attention\nmechanism and sparsity regularization to yield an important subgraph and node\nfeature subset related to any graph-based task. Through a single self-attention\nlayer, a GISST model learns an importance probability for each node feature and\nedge in the input graph. By including these importance probabilities in the\nmodel loss function, the probabilities are optimized end-to-end and tied to the\ntask-specific performance. Furthermore, GISST sparsifies these importance\nprobabilities with entropy and L1 regularization to reduce noise in the input\ngraph topology and node features. Our GISST models achieve superior node\nfeature and edge explanation precision in synthetic datasets, as compared to\nalternative interpretation approaches. Moreover, our GISST models are able to\nidentify important graph structure in real-world datasets. We demonstrate in\ntheory that edge feature importance and multiple edge types can be considered\nby incorporating them into the GISST edge probability computation. By jointly\naccounting for topology, node features, and edge features, GISST inherently\nprovides simple and relevant interpretations for any GNN models and tasks.', 'Graph Neural Networks (GNNs) have been shown to be effective models for\ndifferent predictive tasks on graph-structured data. Recent work on their\nexpressive power has focused on isomorphism tasks and countable feature spaces.\nWe extend this theoretical framework to include continuous features - which\noccur regularly in real-world input domains and within the hidden layers of\nGNNs - and we demonstrate the requirement for multiple aggregation functions in\nthis context. Accordingly, we propose Principal Neighbourhood Aggregation\n(PNA), a novel architecture combining multiple aggregators with degree-scalers\n(which generalize the sum aggregator). Finally, we compare the capacity of\ndifferent models to capture and exploit the graph structure via a novel\nbenchmark containing multiple tasks taken from classical graph theory,\nalongside existing benchmarks from real-world domains, all of which demonstrate\nthe strength of our model. With this work, we hope to steer some of the GNN\nresearch towards new aggregation methods which we believe are essential in the\nsearch for powerful and robust models.', 'The spatial convolution layer which is widely used in the Graph Neural\nNetworks (GNNs) aggregates the feature vector of each node with the feature\nvectors of its neighboring nodes. The GNN is not aware of the locations of the\nnodes in the global structure of the graph and when the local structures\ncorresponding to different nodes are similar to each other, the convolution\nlayer maps all those nodes to similar or same feature vectors in the continuous\nfeature space. Therefore, the GNN cannot distinguish two graphs if their\ndifference is not in their local structures. In addition, when the nodes are\nnot labeled/attributed the convolution layers can fail to distinguish even\ndifferent local structures. In this paper, we propose an effective solution to\naddress this problem of the GNNs. The proposed approach leverages a spatial\nrepresentation of the graph which makes the neural network aware of the\ndifferences between the nodes and also their locations in the graph. The\nspatial representation which is equivalent to a point-cloud representation of\nthe graph is obtained by a graph embedding method. Using the proposed approach,\nthe local feature extractor of the GNN distinguishes similar local structures\nin different locations of the graph and the GNN infers the topological\nstructure of the graph from the spatial distribution of the locally extracted\nfeature vectors. Moreover, the spatial representation is utilized to simplify\nthe graph down-sampling problem. A new graph pooling method is proposed and it\nis shown that the proposed pooling method achieves competitive or better\nresults in comparison with the state-of-the-art methods.', 'Graphs can model complicated interactions between entities, which naturally\nemerge in many important applications. These applications can often be cast\ninto standard graph learning tasks, in which a crucial step is to learn\nlow-dimensional graph representations. Graph neural networks (GNNs) are\ncurrently the most popular model in graph embedding approaches. However,\nstandard GNNs in the neighborhood aggregation paradigm suffer from limited\ndiscriminative power in distinguishing \\emph{high-order} graph structures as\nopposed to \\emph{low-order} structures. To capture high-order structures,\nresearchers have resorted to motifs and developed motif-based GNNs. However,\nexisting motif-based GNNs still often suffer from less discriminative power on\nhigh-order structures. To overcome the above limitations, we propose Motif\nGraph Neural Network (MGNN), a novel framework to better capture high-order\nstructures, hinging on our proposed motif redundancy minimization operator and\ninjective motif combination. First, MGNN produces a set of node representations\nw.r.t. each motif. The next phase is our proposed redundancy minimization among\nmotifs which compares the motifs with each other and distills the features\nunique to each motif. Finally, MGNN performs the updating of node\nrepresentations by combining multiple representations from different motifs. In\nparticular, to enhance the discriminative power, MGNN utilizes an injective\nfunction to combine the representations w.r.t. different motifs. We further\nshow that our proposed architecture increases the expressive power of GNNs with\na theoretical analysis. We demonstrate that MGNN outperforms state-of-the-art\nmethods on seven public benchmarks on both node classification and graph\nclassification tasks.', 'Graph Neural Networks (GNNs) rely on the graph structure to define an\naggregation strategy where each node updates its representation by combining\ninformation from its neighbours. A known limitation of GNNs is that, as the\nnumber of layers increases, information gets smoothed and squashed and node\nembeddings become indistinguishable, negatively affecting performance.\nTherefore, practical GNN models employ few layers and only leverage the graph\nstructure in terms of limited, small neighbourhoods around each node.\nInevitably, practical GNNs do not capture information depending on the global\nstructure of the graph. While there have been several works studying the\nlimitations and expressivity of GNNs, the question of whether practical\napplications on graph structured data require global structural knowledge or\nnot, remains unanswered. In this work, we empirically address this question by\ngiving access to global information to several GNN models, and observing the\nimpact it has on downstream performance. Our results show that global\ninformation can in fact provide significant benefits for common graph-related\ntasks. We further identify a novel regularization strategy that leads to an\naverage accuracy improvement of more than 5% on all considered tasks.', 'Message passing has evolved as an effective tool for designing Graph Neural\nNetworks (GNNs). However, most existing methods for message passing simply sum\nor average all the neighboring features to update node representations. They\nare restricted by two problems, i.e., (i) lack of interpretability to identify\nnode features significant to the prediction of GNNs, and (ii) feature\nover-mixing that leads to the over-smoothing issue in capturing long-range\ndependencies and inability to handle graphs under heterophily or low homophily.\nIn this paper, we propose a Node-level Capsule Graph Neural Network (NCGNN) to\naddress these problems with an improved message passing scheme. Specifically,\nNCGNN represents nodes as groups of node-level capsules, in which each capsule\nextracts distinctive features of its corresponding node. For each node-level\ncapsule, a novel dynamic routing procedure is developed to adaptively select\nappropriate capsules for aggregation from a subgraph identified by the designed\ngraph filter. NCGNN aggregates only the advantageous capsules and restrains\nirrelevant messages to avoid over-mixing features of interacting nodes.\nTherefore, it can relieve the over-smoothing issue and learn effective node\nrepresentations over graphs with homophily or heterophily. Furthermore, our\nproposed message passing scheme is inherently interpretable and exempt from\ncomplex post-hoc explanations, as the graph filter and the dynamic routing\nprocedure identify a subset of node features that are most significant to the\nmodel prediction from the extracted subgraph. Extensive experiments on\nsynthetic as well as real-world graphs demonstrate that NCGNN can well address\nthe over-smoothing issue and produce better node representations for\nsemisupervised node classification. It outperforms the state of the arts under\nboth homophily and heterophily.', ""Distinguishing the automorphic equivalence of nodes in a graph plays an\nessential role in many scientific domains, e.g., computational biologist and\nsocial network analysis. However, existing graph neural networks (GNNs) fail to\ncapture such an important property. To make GNN aware of automorphic\nequivalence, we first introduce a localized variant of this concept --\nego-centered automorphic equivalence (Ego-AE). Then, we design a novel variant\nof GNN, i.e., GRAPE, that uses learnable AE-aware aggregators to explicitly\ndifferentiate the Ego-AE of each node's neighbors with the aids of various\nsubgraph templates. While the design of subgraph templates can be hard, we\nfurther propose a genetic algorithm to automatically search them from graph\ndata. Moreover, we theoretically prove that GRAPE is expressive in terms of\ngenerating distinct representations for nodes with different Ego-AE features,\nwhich fills in a fundamental gap of existing GNN variants. Finally, we\nempirically validate our model on eight real-world graph data, including social\nnetwork, e-commerce co-purchase network, and citation network, and show that it\nconsistently outperforms existing GNNs. The source code is public available at\nhttps://github.com/tsinghua-fib-lab/GRAPE."", ""Message Passing Neural Networks (MPNNs) are a common type of Graph Neural\nNetwork (GNN), in which each node's representation is computed recursively by\naggregating representations (messages) from its immediate neighbors akin to a\nstar-shaped pattern. MPNNs are appealing for being efficient and scalable,\nhow-ever their expressiveness is upper-bounded by the 1st-order\nWeisfeiler-Lehman isomorphism test (1-WL). In response, prior works propose\nhighly expressive models at the cost of scalability and sometimes\ngeneralization performance. Our work stands between these two regimes: we\nintroduce a general framework to uplift any MPNN to be more expressive, with\nlimited scalability overhead and greatly improved practical performance. We\nachieve this by extending local aggregation in MPNNs from star patterns to\ngeneral subgraph patterns (e.g.,k-egonets):in our framework, each node\nrepresentation is computed as the encoding of a surrounding induced subgraph\nrather than encoding of immediate neighbors only (i.e. a star). We choose the\nsubgraph encoder to be a GNN (mainly MPNNs, considering scalability) to design\na general framework that serves as a wrapper to up-lift any GNN. We call our\nproposed method GNN-AK(GNN As Kernel), as the framework resembles a\nconvolutional neural network by replacing the kernel with GNNs. Theoretically,\nwe show that our framework is strictly more powerful than 1&2-WL, and is not\nless powerful than 3-WL. We also design subgraph sampling strategies which\ngreatly reduce memory footprint and improve speed while maintaining\nperformance. Our method sets new state-of-the-art performance by large margins\nfor several well-known graph ML tasks; specifically, 0.08 MAE on ZINC,74.79%\nand 86.887% accuracy on CIFAR10 and PATTERN respectively."", 'We propose a generalization of transformer neural network architecture for\narbitrary graphs. The original transformer was designed for Natural Language\nProcessing (NLP), which operates on fully connected graphs representing all\nconnections between the words in a sequence. Such architecture does not\nleverage the graph connectivity inductive bias, and can perform poorly when the\ngraph topology is important and has not been encoded into the node features. We\nintroduce a graph transformer with four new properties compared to the standard\nmodel. First, the attention mechanism is a function of the neighborhood\nconnectivity for each node in the graph. Second, the positional encoding is\nrepresented by the Laplacian eigenvectors, which naturally generalize the\nsinusoidal positional encodings often used in NLP. Third, the layer\nnormalization is replaced by a batch normalization layer, which provides faster\ntraining and better generalization performance. Finally, the architecture is\nextended to edge feature representation, which can be critical to tasks s.a.\nchemistry (bond type) or link prediction (entity relationship in knowledge\ngraphs). Numerical experiments on a graph benchmark demonstrate the performance\nof the proposed graph transformer architecture. This work closes the gap\nbetween the original transformer, which was designed for the limited case of\nline graphs, and graph neural networks, that can work with arbitrary graphs. As\nour architecture is simple and generic, we believe it can be used as a black\nbox for future applications that wish to consider transformer and graphs.', 'It is not until recently that graph neural networks (GNNs) are adopted to\nperform graph representation learning, among which, those based on the\naggregation of features within the neighborhood of a node achieved great\nsuccess. However, despite such achievements, GNNs illustrate defects in\nidentifying some common structural patterns which, unfortunately, play\nsignificant roles in various network phenomena. In this paper, we propose\nGraLSP, a GNN framework which explicitly incorporates local structural patterns\ninto the neighborhood aggregation through random anonymous walks. Specifically,\nwe capture local graph structures via random anonymous walks, powerful and\nflexible tools that represent structural patterns. The walks are then fed into\nthe feature aggregation, where we design various mechanisms to address the\nimpact of structural features, including adaptive receptive radius, attention\nand amplification. In addition, we design objectives that capture similarities\nbetween structures and are optimized jointly with node proximity objectives.\nWith the adequate leverage of structural patterns, our model is able to\noutperform competitive counterparts in various prediction tasks in multiple\ndatasets.', 'Graph neural networks (GNNs) have achieved great success in many graph-based\ntasks. Much work is dedicated to empowering GNNs with the adaptive locality\nability, which enables measuring the importance of neighboring nodes to the\ntarget node by a node-specific mechanism. However, the current node-specific\nmechanisms are deficient in distinguishing the importance of nodes in the\ntopology structure. We believe that the structural importance of neighboring\nnodes is closely related to their importance in aggregation. In this paper, we\nintroduce discrete graph curvature (the Ricci curvature) to quantify the\nstrength of structural connection of pairwise nodes. And we propose Curvature\nGraph Neural Network (CGNN), which effectively improves the adaptive locality\nability of GNNs by leveraging the structural property of graph curvature. To\nimprove the adaptability of curvature to various datasets, we explicitly\ntransform curvature into the weights of neighboring nodes by the necessary\nNegative Curvature Processing Module and Curvature Normalization Module. Then,\nwe conduct numerous experiments on various synthetic datasets and real-world\ndatasets. The experimental results on synthetic datasets show that CGNN\neffectively exploits the topology structure information, and the performance is\nimproved significantly. CGNN outperforms the baselines on 5 dense node\nclassification benchmark datasets. This study deepens the understanding of how\nto utilize advanced topology information and assign the importance of\nneighboring nodes from the perspective of graph curvature and encourages us to\nbridge the gap between graph theory and neural networks.', ""Machine learning on graph-structured data has attracted high research\ninterest due to the emergence of Graph Neural Networks (GNNs). Most of the\nproposed GNNs are based on the node homophily, i.e neighboring nodes share\nsimilar characteristics. However, in many complex networks, nodes that lie to\ndistant parts of the graph share structurally equivalent characteristics and\nexhibit similar roles (e.g chemical properties of distant atoms in a molecule,\ntype of social network users). A growing literature proposed representations\nthat identify structurally equivalent nodes. However, most of the existing\nmethods require high time and space complexity. In this paper, we propose\nVNEstruct, a simple approach, based on entropy measures of the neighborhood's\ntopology, for generating low-dimensional structural representations, that is\ntime-efficient and robust to graph perturbations. Empirically, we observe that\nVNEstruct exhibits robustness on structural role identification tasks.\nMoreover, VNEstruct can achieve state-of-the-art performance on graph\nclassification, without incorporating the graph structure information in the\noptimization, in contrast to GNN competitors."", 'With the great success of deep learning in various domains, graph neural\nnetworks (GNNs) also become a dominant approach to graph classification. By the\nhelp of a global readout operation that simply aggregates all node (or\nnode-cluster) representations, existing GNN classifiers obtain a graph-level\nrepresentation of an input graph and predict its class label using the\nrepresentation. However, such global aggregation does not consider the\nstructural information of each node, which results in information loss on the\nglobal structure. Particularly, it limits the discrimination power by enforcing\nthe same weight parameters of the classifier for all the node representations;\nin practice, each of them contributes to target classes differently depending\non its structural semantic. In this work, we propose structural semantic\nreadout (SSRead) to summarize the node representations at the position-level,\nwhich allows to model the position-specific weight parameters for\nclassification as well as to effectively capture the graph semantic relevant to\nthe global structure. Given an input graph, SSRead aims to identify\nstructurally-meaningful positions by using the semantic alignment between its\nnodes and structural prototypes, which encode the prototypical features of each\nposition. The structural prototypes are optimized to minimize the alignment\ncost for all training graphs, while the other GNN parameters are trained to\npredict the class labels. Our experimental results demonstrate that SSRead\nsignificantly improves the classification performance and interpretability of\nGNN classifiers while being compatible with a variety of aggregation functions,\nGNN architectures, and learning frameworks.', 'While Graph Neural Networks (GNNs) have achieved remarkable results in a\nvariety of applications, recent studies exposed important shortcomings in their\nability to capture the structure of the underlying graph. It has been shown\nthat the expressive power of standard GNNs is bounded by the Weisfeiler-Leman\n(WL) graph isomorphism test, from which they inherit proven limitations such as\nthe inability to detect and count graph substructures. On the other hand, there\nis significant empirical evidence, e.g. in network science and bioinformatics,\nthat substructures are often intimately related to downstream tasks. To this\nend, we propose ""Graph Substructure Networks"" (GSN), a topologically-aware\nmessage passing scheme based on substructure encoding. We theoretically analyse\nthe expressive power of our architecture, showing that it is strictly more\nexpressive than the WL test, and provide sufficient conditions for\nuniversality. Importantly, we do not attempt to adhere to the WL hierarchy;\nthis allows us to retain multiple attractive properties of standard GNNs such\nas locality and linear network complexity, while being able to disambiguate\neven hard instances of graph isomorphism. We perform an extensive experimental\nevaluation on graph classification and regression tasks and obtain\nstate-of-the-art results in diverse real-world settings including molecular\ngraphs and social networks. The code is publicly available at\nhttps://github.com/gbouritsas/graph-substructure-networks.', 'The ability to detect and count certain substructures in graphs is important\nfor solving many tasks on graph-structured data, especially in the contexts of\ncomputational chemistry and biology as well as social network analysis.\nInspired by this, we propose to study the expressive power of graph neural\nnetworks (GNNs) via their ability to count attributed graph substructures,\nextending recent works that examine their power in graph isomorphism testing\nand function approximation. We distinguish between two types of substructure\ncounting: induced-subgraph-count and subgraph-count, and establish both\npositive and negative answers for popular GNN architectures. Specifically, we\nprove that Message Passing Neural Networks (MPNNs), 2-Weisfeiler-Lehman (2-WL)\nand 2-Invariant Graph Networks (2-IGNs) cannot perform induced-subgraph-count\nof substructures consisting of 3 or more nodes, while they can perform\nsubgraph-count of star-shaped substructures. As an intermediary step, we prove\nthat 2-WL and 2-IGNs are equivalent in distinguishing non-isomorphic graphs,\npartly answering an open problem raised in Maron et al. (2019). We also prove\npositive results for k-WL and k-IGNs as well as negative results for k-WL with\na finite number of iterations. We then conduct experiments that support the\ntheoretical results for MPNNs and 2-IGNs. Moreover, motivated by substructure\ncounting and inspired by Murphy et al. (2019), we propose the Local Relational\nPooling model and demonstrate that it is not only effective for substructure\ncounting but also able to achieve competitive performance on molecular\nprediction tasks.', 'Graph Neural Networks (GNNs) have shown advantages in various graph-based\napplications. Most existing GNNs assume strong homophily of graph structure and\napply permutation-invariant local aggregation of neighbors to learn a\nrepresentation for each node. However, they fail to generalize to heterophilic\ngraphs, where most neighboring nodes have different labels or features, and the\nrelevant nodes are distant. Few recent studies attempt to address this problem\nby combining multiple hops of hidden representations of central nodes (i.e.,\nmulti-hop-based approaches) or sorting the neighboring nodes based on attention\nscores (i.e., ranking-based approaches). As a result, these approaches have\nsome apparent limitations. On the one hand, multi-hop-based approaches do not\nexplicitly distinguish relevant nodes from a large number of multi-hop\nneighborhoods, leading to a severe over-smoothing problem. On the other hand,\nranking-based models do not joint-optimize node ranking with end tasks and\nresult in sub-optimal solutions. In this work, we present Graph Pointer Neural\nNetworks (GPNN) to tackle the challenges mentioned above. We leverage a pointer\nnetwork to select the most relevant nodes from a large amount of multi-hop\nneighborhoods, which constructs an ordered sequence according to the\nrelationship with the central node. 1D convolution is then applied to extract\nhigh-level features from the node sequence. The pointer-network-based ranker in\nGPNN is joint-optimized with other parts in an end-to-end manner. Extensive\nexperiments are conducted on six public node classification datasets with\nheterophilic graphs. The results show that GPNN significantly improves the\nclassification performance of state-of-the-art methods. In addition, analyses\nalso reveal the privilege of the proposed GPNN in filtering out irrelevant\nneighbors and reducing over-smoothing.', 'Graph Neural Networks (GNNs) have achieved tremendous success in various\nreal-world applications due to their strong ability in graph representation\nlearning. GNNs explore the graph structure and node features by aggregating and\ntransforming information within node neighborhoods. However, through\ntheoretical and empirical analysis, we reveal that the aggregation process of\nGNNs tends to destroy node similarity in the original feature space. There are\nmany scenarios where node similarity plays a crucial role. Thus, it has\nmotivated the proposed framework SimP-GCN that can effectively and efficiently\npreserve node similarity while exploiting graph structure. Specifically, to\nbalance information from graph structure and node features, we propose a\nfeature similarity preserving aggregation which adaptively integrates graph\nstructure and node features. Furthermore, we employ self-supervised learning to\nexplicitly capture the complex feature similarity and dissimilarity relations\nbetween nodes. We validate the effectiveness of SimP-GCN on seven benchmark\ndatasets including three assortative and four disassorative graphs. The results\ndemonstrate that SimP-GCN outperforms representative baselines. Further probe\nshows various advantages of the proposed framework. The implementation of\nSimP-GCN is available at \\url{https://github.com/ChandlerBang/SimP-GCN}.', 'Graph neural networks (GNNs) emerge as a powerful family of representation\nlearning models on graphs. To derive node representations, they utilize a\nglobal model that recursively aggregates information from the neighboring\nnodes. However, different nodes reside at different parts of the graph in\ndifferent local contexts, making their distributions vary across the graph.\nIdeally, how a node receives its neighborhood information should be a function\nof its local context, to diverge from the global GNN model shared by all nodes.\nTo utilize node locality without overfitting, we propose a node-wise\nlocalization of GNNs by accounting for both global and local aspects of the\ngraph. Globally, all nodes on the graph depend on an underlying global GNN to\nencode the general patterns across the graph; locally, each node is localized\ninto a unique model as a function of the global model and its local context.\nFinally, we conduct extensive experiments on four benchmark graphs, and\nconsistently obtain promising performance surpassing the state-of-the-art GNNs.', ""Graph data widely exist in many high-impact applications. Inspired by the\nsuccess of deep learning in grid-structured data, graph neural network models\nhave been proposed to learn powerful node-level or graph-level representation.\nHowever, most of the existing graph neural networks suffer from the following\nlimitations: (1) there is limited analysis regarding the graph convolution\nproperties, such as seed-oriented, degree-aware and order-free; (2) the node's\ndegree-specific graph structure is not explicitly expressed in graph\nconvolution for distinguishing structure-aware node neighborhoods; (3) the\ntheoretical explanation regarding the graph-level pooling schemes is unclear.\n  To address these problems, we propose a generic degree-specific graph neural\nnetwork named DEMO-Net motivated by Weisfeiler-Lehman graph isomorphism test\nthat recursively identifies 1-hop neighborhood structures. In order to\nexplicitly capture the graph topology integrated with node attributes, we argue\nthat graph convolution should have three properties: seed-oriented,\ndegree-aware, order-free. To this end, we propose multi-task graph convolution\nwhere each task represents node representation learning for nodes with a\nspecific degree value, thus leading to preserving the degree-specific graph\nstructure. In particular, we design two multi-task learning methods:\ndegree-specific weight and hashing functions for graph convolution. In\naddition, we propose a novel graph-level pooling/readout scheme for learning\ngraph representation provably lying in a degree-specific Hilbert kernel space.\nThe experimental results on several node and graph classification benchmark\ndata sets demonstrate the effectiveness and efficiency of our proposed DEMO-Net\nover state-of-the-art graph neural network models."", 'Graph Neural Networks (GNNs) are emerging machine learning models on graphs.\nAlthough sufficiently deep GNNs are shown theoretically capable of fully\npreserving graph structures, most existing GNN models in practice are shallow\nand essentially feature-centric. We show empirically and analytically that the\nexisting shallow GNNs cannot preserve graph structures well. To overcome this\nfundamental challenge, we propose Eigen-GNN, a simple yet effective and general\nplug-in module to boost GNNs ability in preserving graph structures.\nSpecifically, we integrate the eigenspace of graph structures with GNNs by\ntreating GNNs as a type of dimensionality reduction and expanding the initial\ndimensionality reduction bases. Without needing to increase depths, Eigen-GNN\npossesses more flexibilities in handling both feature-driven and\nstructure-driven tasks since the initial bases contain both node features and\ngraph structures. We present extensive experimental results to demonstrate the\neffectiveness of Eigen-GNN for tasks including node classification, link\nprediction, and graph isomorphism tests.', 'Graph data are pervasive in many real-world applications. Recently,\nincreasing attention has been paid on graph neural networks (GNNs), which aim\nto model the local graph structures and capture the hierarchical patterns by\naggregating the information from neighbors with stackable network modules.\nMotivated by the observation that different nodes often require different\niterations of aggregation to fully capture the structural information, in this\npaper, we propose to explicitly sample diverse iterations of aggregation for\ndifferent nodes to boost the performance of GNNs. It is a challenging task to\ndevelop an effective aggregation strategy for each node, given complex graphs\nand sparse features. Moreover, it is not straightforward to derive an efficient\nalgorithm since we need to feed the sampled nodes into different number of\nnetwork layers. To address the above challenges, we propose Policy-GNN, a\nmeta-policy framework that models the sampling procedure and message passing of\nGNNs into a combined learning process. Specifically, Policy-GNN uses a\nmeta-policy to adaptively determine the number of aggregations for each node.\nThe meta-policy is trained with deep reinforcement learning (RL) by exploiting\nthe feedback from the model. We further introduce parameter sharing and a\nbuffer mechanism to boost the training efficiency. Experimental results on\nthree real-world benchmark datasets suggest that Policy-GNN significantly\noutperforms the state-of-the-art alternatives, showing the promise in\naggregation optimization for GNNs.', 'Various recent proposals increase the distinguishing power of Graph Neural\nNetworks GNNs by propagating features between $k$-tuples of vertices. The\ndistinguishing power of these ""higher-order\'\' GNNs is known to be bounded by\nthe $k$-dimensional Weisfeiler-Leman (WL) test, yet their $\\mathcal O(n^k)$\nmemory requirements limit their applicability. Other proposals infuse GNNs with\nlocal higher-order graph structural information from the start, hereby\ninheriting the desirable $\\mathcal O(n)$ memory requirement from GNNs at the\ncost of a one-time, possibly non-linear, preprocessing step. We propose local\ngraph parameter enabled GNNs as a framework for studying the latter kind of\napproaches and precisely characterize their distinguishing power, in terms of a\nvariant of the WL test, and in terms of the graph structural properties that\nthey can take into account. Local graph parameters can be added to any GNN\narchitecture, and are cheap to compute. In terms of expressive power, our\nproposal lies in the middle of GNNs and their higher-order counterparts.\nFurther, we propose several techniques to aide in choosing the right local\ngraph parameters. Our results connect GNNs with deep results in finite model\ntheory and finite variable logics. Our experimental evaluation shows that\nadding local graph parameters often has a positive effect for a variety of\nGNNs, datasets and graph learning tasks.', ""Graph neural networks (GNNs) have achieved tremendous success on multiple\ngraph-based learning tasks by fusing network structure and node features.\nModern GNN models are built upon iterative aggregation of neighbor's/proximity\nfeatures by message passing. Its prediction performance has been shown to be\nstrongly bounded by assortative mixing in the graph, a key property wherein\nnodes with similar attributes mix/connect with each other. We observe that real\nworld networks exhibit heterogeneous or diverse mixing patterns and the\nconventional global measurement of assortativity, such as global assortativity\ncoefficient, may not be a representative statistic in quantifying this mixing.\nWe adopt a generalized concept, node-level assortativity, one that is based at\nthe node level to better represent the diverse patterns and accurately quantify\nthe learnability of GNNs. We find that the prediction performance of a wide\nrange of GNN models is highly correlated with the node level assortativity. To\nbreak this limit, in this work, we focus on transforming the input graph into a\ncomputation graph which contains both proximity and structural information as\ndistinct type of edges. The resulted multi-relational graph has an enhanced\nlevel of assortativity and, more importantly, preserves rich information from\nthe original graph. We then propose to run GNNs on this computation graph and\nshow that adaptively choosing between structure and proximity leads to improved\nperformance under diverse mixing. Empirically, we show the benefits of adopting\nour transformation framework for semi-supervised node classification task on a\nvariety of real world graph learning benchmarks."", ""Message passing Graph Neural Networks (GNNs) provide a powerful modeling\nframework for relational data. However, the expressive power of existing GNNs\nis upper-bounded by the 1-Weisfeiler-Lehman (1-WL) graph isomorphism test,\nwhich means GNNs that are not able to predict node clustering coefficients and\nshortest path distances, and cannot differentiate between different d-regular\ngraphs. Here we develop a class of message passing GNNs, named Identity-aware\nGraph Neural Networks (ID-GNNs), with greater expressive power than the 1-WL\ntest. ID-GNN offers a minimal but powerful solution to limitations of existing\nGNNs. ID-GNN extends existing GNN architectures by inductively considering\nnodes' identities during message passing. To embed a given node, ID-GNN first\nextracts the ego network centered at the node, then conducts rounds of\nheterogeneous message passing, where different sets of parameters are applied\nto the center node than to other surrounding nodes in the ego network. We\nfurther propose a simplified but faster version of ID-GNN that injects node\nidentity information as augmented node features. Altogether, both versions of\nID-GNN represent general extensions of message passing GNNs, where experiments\nshow that transforming existing GNNs to ID-GNNs yields on average 40% accuracy\nimprovement on challenging node, edge, and graph property prediction tasks; 3%\naccuracy improvement on node and graph classification benchmarks; and 15% ROC\nAUC improvement on real-world link prediction tasks. Additionally, ID-GNNs\ndemonstrate improved or comparable performance over other task-specific graph\nnetworks."", 'In recent years, graph neural networks (GNNs) have emerged as a powerful\nneural architecture to learn vector representations of nodes and graphs in a\nsupervised, end-to-end fashion. Up to now, GNNs have only been evaluated\nempirically -- showing promising results. The following work investigates GNNs\nfrom a theoretical point of view and relates them to the $1$-dimensional\nWeisfeiler-Leman graph isomorphism heuristic ($1$-WL). We show that GNNs have\nthe same expressiveness as the $1$-WL in terms of distinguishing non-isomorphic\n(sub-)graphs. Hence, both algorithms also have the same shortcomings. Based on\nthis, we propose a generalization of GNNs, so-called $k$-dimensional GNNs\n($k$-GNNs), which can take higher-order graph structures at multiple scales\ninto account. These higher-order structures play an essential role in the\ncharacterization of social networks and molecule graphs. Our experimental\nevaluation confirms our theoretical findings as well as confirms that\nhigher-order information is useful in the task of graph classification and\nregression.', 'In this paper, we propose a spatial graph convolution (GC) to classify\nsignals on a graph. Existing GC methods are limited to using the structural\ninformation in the feature space. Additionally, the single step of GCs only\nuses features on the one-hop neighboring nodes from the target node. In this\npaper, we propose two methods to improve the performance of GCs: 1) Utilizing\nstructural information in the feature space, and 2) exploiting the multi-hop\ninformation in one GC step. In the first method, we define three structural\nfeatures in the feature space: feature angle, feature distance, and relational\nembedding. The second method aggregates the node-wise features of multi-hop\nneighbors in a GC. Both methods can be simultaneously used. We also propose\ngraph neural networks (GNNs) integrating the proposed GC for classifying nodes\nin 3D point clouds and citation networks. In experiments, the proposed GNNs\nexhibited a higher classification accuracy than existing methods.', 'Graph Neural Networks (GNNs) are deep learning methods which provide the\ncurrent state of the art performance in node classification tasks. GNNs often\nassume homophily -- neighboring nodes having similar features and labels--, and\ntherefore may not be at their full potential when dealing with non-homophilic\ngraphs. In this work, we focus on addressing this limitation and enable Graph\nAttention Networks (GAT), a commonly used variant of GNNs, to explore the\nstructural information within each graph locality. Inspired by the positional\nencoding in the Transformers, we propose a framework, termed Graph Attentional\nNetworks with Positional Embeddings (GAT-POS), to enhance GATs with positional\nembeddings which capture structural and positional information of the nodes in\nthe graph. In this framework, the positional embeddings are learned by a model\npredictive of the graph context, plugged into an enhanced GAT architecture,\nwhich is able to leverage both the positional and content information of each\nnode. The model is trained jointly to optimize for the task of node\nclassification as well as the task of predicting graph context. Experimental\nresults show that GAT-POS reaches remarkable improvement compared to strong GNN\nbaselines and recent structural embedding enhanced GNNs on non-homophilic\ngraphs.', 'Inspired by the immense success of deep learning, graph neural networks\n(GNNs) are widely used to learn powerful node representations and have\ndemonstrated promising performance on different graph learning tasks. However,\nmost real-world graphs often come with high-dimensional and sparse node\nfeatures, rendering the learned node representations from existing GNN\narchitectures less expressive. In this paper, we propose \\textit{Feature\nInteraction-aware Graph Neural Networks (FI-GNNs)}, a plug-and-play GNN\nframework for learning node representations encoded with informative feature\ninteractions. Specifically, the proposed framework is able to highlight\ninformative feature interactions in a personalized manner and further learn\nhighly expressive node representations on feature-sparse graphs. Extensive\nexperiments on various datasets demonstrate the superior capability of FI-GNNs\nfor graph learning tasks.', 'Graph Neural Networks (GNNs) are an effective framework for representation\nlearning of graphs. GNNs follow a neighborhood aggregation scheme, where the\nrepresentation vector of a node is computed by recursively aggregating and\ntransforming representation vectors of its neighboring nodes. Many GNN variants\nhave been proposed and have achieved state-of-the-art results on both node and\ngraph classification tasks. However, despite GNNs revolutionizing graph\nrepresentation learning, there is limited understanding of their\nrepresentational properties and limitations. Here, we present a theoretical\nframework for analyzing the expressive power of GNNs to capture different graph\nstructures. Our results characterize the discriminative power of popular GNN\nvariants, such as Graph Convolutional Networks and GraphSAGE, and show that\nthey cannot learn to distinguish certain simple graph structures. We then\ndevelop a simple architecture that is provably the most expressive among the\nclass of GNNs and is as powerful as the Weisfeiler-Lehman graph isomorphism\ntest. We empirically validate our theoretical findings on a number of graph\nclassification benchmarks, and demonstrate that our model achieves\nstate-of-the-art performance.', 'Graph neural networks (GNNs) work well when the graph structure is provided.\nHowever, this structure may not always be available in real-world applications.\nOne solution to this problem is to infer a task-specific latent structure and\nthen apply a GNN to the inferred graph. Unfortunately, the space of possible\ngraph structures grows super-exponentially with the number of nodes and so the\ntask-specific supervision may be insufficient for learning both the structure\nand the GNN parameters. In this work, we propose the Simultaneous Learning of\nAdjacency and GNN Parameters with Self-supervision, or SLAPS, a method that\nprovides more supervision for inferring a graph structure through\nself-supervision. A comprehensive experimental study demonstrates that SLAPS\nscales to large graphs with hundreds of thousands of nodes and outperforms\nseveral models that have been proposed to learn a task-specific graph structure\non established benchmarks.', ""Graph Neural Networks (GNNs) have achieved remarkable performance on\ngraph-based tasks. The key idea for GNNs is to obtain informative\nrepresentation through aggregating information from local neighborhoods.\nHowever, it remains an open question whether the neighborhood information is\nadequately aggregated for learning representations of nodes with few neighbors.\nTo address this, we propose a simple and efficient data augmentation strategy,\nlocal augmentation, to learn the distribution of the node features of the\nneighbors conditioned on the central node's feature and enhance GNN's\nexpressive power with generated features. Local augmentation is a general\nframework that can be applied to any GNN model in a plug-and-play manner. It\nsamples feature vectors associated with each node from the learned conditional\ndistribution as additional input for the backbone model at each training\niteration. Extensive experiments and analyses show that local augmentation\nconsistently yields performance improvement when applied to various GNN\narchitectures across a diverse set of benchmarks. For example, experiments show\nthat plugging in local augmentation to GCN and GAT improves by an average of\n3.4\\% and 1.6\\% in terms of test accuracy on Cora, Citeseer, and Pubmed.\nBesides, our experimental results on large graphs (OGB) show that our model\nconsistently improves performance over backbones. Code is available at\nhttps://github.com/SongtaoLiu0823/LAGNN."", 'Graph neural networks (GNNs) have been widely used in various graph-related\nproblems such as node classification and graph classification, where superior\nperformance is mainly established when natural node features are available.\nHowever, it is not well understood how GNNs work without natural node features,\nespecially regarding the various ways to construct artificial ones. In this\npaper, we point out the two types of artificial node features, i.e., positional\nand structural node features, and provide insights on why each of them is more\nappropriate for certain tasks, i.e., positional node classification, structural\nnode classification, and graph classification. Extensive experimental results\non 10 benchmark datasets validate our insights, thus leading to a practical\nguideline on the choices between different artificial node features for GNNs on\nnon-attributed graphs. The code is available at\nhttps://github.com/zjzijielu/gnn-positional-structural-node-features.', 'Learning graph-structured data with graph neural networks (GNNs) has been\nrecently emerging as an important field because of its wide applicability in\nbioinformatics, chemoinformatics, social network analysis and data mining.\nRecent GNN algorithms are based on neural message passing, which enables GNNs\nto integrate local structures and node features recursively. However, past GNN\nalgorithms based on 1-hop neighborhood neural message passing are exposed to a\nrisk of loss of information on local structures and relationships. In this\npaper, we propose Neighborhood Edge AggregatoR (NEAR), a framework that\naggregates relations between the nodes in the neighborhood via edges. NEAR,\nwhich can be orthogonally combined with Graph Isomorphism Network (GIN), gives\nintegrated information that describes which nodes in the neighborhood are\nconnected. Therefore, NEAR can reflect additional information of a local\nstructure of each node beyond the nodes themselves in 1-hop neighborhood.\nExperimental results on multiple graph classification tasks show that our\nalgorithm makes a good improvement over other existing 1-hop based GNN-based\nalgorithms.', 'Graph neural networks (GNNs) achieve remarkable success in graph-based\nsemi-supervised node classification, leveraging the information from\nneighboring nodes to improve the representation learning of target node. The\nsuccess of GNNs at node classification depends on the assumption that connected\nnodes tend to have the same label. However, such an assumption does not always\nwork, limiting the performance of GNNs at node classification. In this paper,\nwe propose label-consistency based graph neural network(LC-GNN), leveraging\nnode pairs unconnected but with the same labels to enlarge the receptive field\nof nodes in GNNs. Experiments on benchmark datasets demonstrate the proposed\nLC-GNN outperforms traditional GNNs in graph-based semi-supervised node\nclassification.We further show the superiority of LC-GNN in sparse scenarios\nwith only a handful of labeled nodes.', ""We consider graph representation learning in a self-supervised manner. Graph\nneural networks (GNNs) use neighborhood aggregation as a core component that\nresults in feature smoothing among nodes in proximity. While successful in\nvarious prediction tasks, such a paradigm falls short of capturing nodes'\nsimilarities over a long distance, which proves to be important for\nhigh-quality learning. To tackle this problem, we strengthen the graph with two\nadditional graph views, in which nodes are directly linked to those with the\nmost similar features or local structures. Not restricted by connectivity in\nthe original graph, the generated views allow the model to enhance its\nexpressive power with new and complementary perspectives from which to look at\nthe relationship between nodes. Following a contrastive learning approach, we\npropose a method that aims to maximize the agreement between representations\nacross generated views and the original graph. We also propose a channel-level\ncontrast approach that greatly reduces computation cost, compared to the\ncommonly used node level contrast, which requires computation cost quadratic in\nthe number of nodes. Extensive experiments on seven assortative graphs and four\ndisassortative graphs demonstrate the effectiveness of our approach."", 'We propose a Graph Neural Network with greater expressive power than commonly\nused GNNs - not constrained to only differentiate between graphs that\nWeisfeiler-Lehman test recognizes to be non-isomorphic. We use a graph\nattention network with expanding attention window that aggregates information\nfrom nodes exponentially far away. We also use partially random initial\nembeddings, allowing differentiation between nodes that would otherwise look\nthe same. This could cause problem with a traditional dropout mechanism,\ntherefore we use a ""head dropout"", randomly ignoring some attention heads\nrather than some dimensions of the embedding.', 'Network data can be conveniently modeled as a graph signal, where data values\nare assigned to nodes of a graph that describes the underlying network\ntopology. Successful learning from network data is built upon methods that\neffectively exploit this graph structure. In this work, we leverage graph\nsignal processing to characterize the representation space of graph neural\nnetworks (GNNs). We discuss the role of graph convolutional filters in GNNs and\nshow that any architecture built with such filters has the fundamental\nproperties of permutation equivariance and stability to changes in the\ntopology. These two properties offer insight about the workings of GNNs and\nhelp explain their scalability and transferability properties which, coupled\nwith their local and distributed nature, make GNNs powerful tools for learning\nin physical networks. We also introduce GNN extensions using edge-varying and\nautoregressive moving average graph filters and discuss their properties.\nFinally, we study the use of GNNs in recommender systems and learning\ndecentralized controllers for robot swarms.', ""Graph neural networks (GNNs) have limited expressive power, failing to\nrepresent many graph classes correctly. While more expressive graph\nrepresentation learning (GRL) alternatives can distinguish some of these\nclasses, they are significantly harder to implement, may not scale well, and\nhave not been shown to outperform well-tuned GNNs in real-world tasks. Thus,\ndevising simple, scalable, and expressive GRL architectures that also achieve\nreal-world improvements remains an open challenge. In this work, we show the\nextent to which graph reconstruction -- reconstructing a graph from its\nsubgraphs -- can mitigate the theoretical and practical problems currently\nfaced by GRL architectures. First, we leverage graph reconstruction to build\ntwo new classes of expressive graph representations. Secondly, we show how\ngraph reconstruction boosts the expressive power of any GNN architecture while\nbeing a (provably) powerful inductive bias for invariances to vertex removals.\nEmpirically, we show how reconstruction can boost GNN's expressive power --\nwhile maintaining its invariance to permutations of the vertices -- by solving\nseven graph property tasks not solvable by the original GNN. Further, we\ndemonstrate how it boosts state-of-the-art GNN's performance across nine\nreal-world benchmark datasets."", 'Graph neural networks (GNNs) are powerful machine learning models for various\ngraph learning tasks. Recently, the limitations of the expressive power of\nvarious GNN models have been revealed. For example, GNNs cannot distinguish\nsome non-isomorphic graphs and they cannot learn efficient graph algorithms. In\nthis paper, we demonstrate that GNNs become powerful just by adding a random\nfeature to each node. We prove that the random features enable GNNs to learn\nalmost optimal polynomial-time approximation algorithms for the minimum\ndominating set problem and maximum matching problem in terms of approximation\nratios. The main advantage of our method is that it can be combined with\noff-the-shelf GNN models with slight modifications. Through experiments, we\nshow that the addition of random features enables GNNs to solve various\nproblems that normal GNNs, including the graph convolutional networks (GCNs)\nand graph isomorphism networks (GINs), cannot solve.', 'Graph Neural Networks (GNNs) have led to state-of-the-art performance on a\nvariety of machine learning tasks such as recommendation, node classification\nand link prediction. Graph neural network models generate node embeddings by\nmerging nodes features with the aggregated neighboring nodes information. Most\nexisting GNN models exploit a single type of aggregator (e.g., mean-pooling) to\naggregate neighboring nodes information, and then add or concatenate the output\nof aggregator to the current representation vector of the center node. However,\nusing only a single type of aggregator is difficult to capture the different\naspects of neighboring information and the simple addition or concatenation\nupdate methods limit the expressive capability of GNNs. Not only that, existing\nsupervised or semi-supervised GNN models are trained based on the loss function\nof the node label, which leads to the neglect of graph structure information.\nIn this paper, we propose a novel graph neural network architecture, Graph\nAttention \\& Interaction Network (GAIN), for inductive learning on graphs.\nUnlike the previous GNN models that only utilize a single type of aggregation\nmethod, we use multiple types of aggregators to gather neighboring information\nin different aspects and integrate the outputs of these aggregators through the\naggregator-level attention mechanism. Furthermore, we design a graph\nregularized loss to better capture the topological relationship of the nodes in\nthe graph. Additionally, we first present the concept of graph feature\ninteraction and propose a vector-wise explicit feature interaction mechanism to\nupdate the node embeddings. We conduct comprehensive experiments on two\nnode-classification benchmarks and a real-world financial news dataset. The\nexperiments demonstrate our GAIN model outperforms current state-of-the-art\nperformances on all the tasks.', 'Graph neural networks (GNNs) have become the standard learning architectures\nfor graphs. GNNs have been applied to numerous domains ranging from quantum\nchemistry, recommender systems to knowledge graphs and natural language\nprocessing. A major issue with arbitrary graphs is the absence of canonical\npositional information of nodes, which decreases the representation power of\nGNNs to distinguish e.g. isomorphic nodes and other graph symmetries. An\napproach to tackle this issue is to introduce Positional Encoding (PE) of\nnodes, and inject it into the input layer, like in Transformers. Possible graph\nPE are Laplacian eigenvectors. In this work, we propose to decouple structural\nand positional representations to make easy for the network to learn these two\nessential properties. We introduce a novel generic architecture which we call\nLSPE (Learnable Structural and Positional Encodings). We investigate several\nsparse and fully-connected (Transformer-like) GNNs, and observe a performance\nincrease for molecular datasets, from 1.79% up to 64.14% when considering\nlearnable PE for both GNN classes.', 'Graph neural networks (GNNs) have shown great power in learning on attributed\ngraphs. However, it is still a challenge for GNNs to utilize information\nfaraway from the source node. Moreover, general GNNs require graph attributes\nas input, so they cannot be appled to plain graphs. In the paper, we propose\nnew models named G-GNNs (Global information for GNNs) to address the above\nlimitations. First, the global structure and attribute features for each node\nare obtained via unsupervised pre-training, which preserve the global\ninformation associated to the node. Then, using the global features and the raw\nnetwork attributes, we propose a parallel framework of GNNs to learn different\naspects from these features. The proposed learning methods can be applied to\nboth plain graphs and attributed graphs. Extensive experiments have shown that\nG-GNNs can outperform other state-of-the-art models on three standard\nevaluation graphs. Specially, our methods establish new benchmark records on\nCora (84.31\\%) and Pubmed (80.95\\%) when learning on attributed graphs.', 'This paper addresses the challenging problem of retrieval and matching of\ngraph structured objects, and makes two key contributions. First, we\ndemonstrate how Graph Neural Networks (GNN), which have emerged as an effective\nmodel for various supervised prediction problems defined on structured data,\ncan be trained to produce embedding of graphs in vector spaces that enables\nefficient similarity reasoning. Second, we propose a novel Graph Matching\nNetwork model that, given a pair of graphs as input, computes a similarity\nscore between them by jointly reasoning on the pair through a new cross-graph\nattention-based matching mechanism. We demonstrate the effectiveness of our\nmodels on different domains including the challenging problem of\ncontrol-flow-graph based function similarity search that plays an important\nrole in the detection of vulnerabilities in software systems. The experimental\nanalysis demonstrates that our models are not only able to exploit structure in\nthe context of similarity learning but they can also outperform domain-specific\nbaseline systems that have been carefully hand-engineered for these problems.', 'Graph neural networks (GNNs) have drawn increasing attention in recent years\nand achieved remarkable performance in many graph-based tasks, especially in\nsemi-supervised learning on graphs. However, most existing GNNs excessively\nrely on topological structures and aggregate multi-hop neighborhood information\nby simply stacking network layers, which may introduce superfluous noise\ninformation, limit the expressive power of GNNs and lead to the over-smoothing\nproblem ultimately. In light of this, we propose a novel Dual-Perception Graph\nNeural Network (DPGNN) to address these issues. In DPGNN, we utilize node\nfeatures to construct a feature graph, and perform node representations\nlearning based on the original topology graph and the constructed feature graph\nsimultaneously, which conduce to capture the structural neighborhood\ninformation and the feature-related information. Furthermore, we design a\nMulti-Hop Graph Generator (MHGG), which applies a node-to-hop attention\nmechanism to aggregate node-specific multi-hop neighborhood information\nadaptively. Finally, we apply self-ensembling to form a consistent prediction\nfor unlabeled node representations. Experimental results on five datasets with\ndifferent topological structures demonstrate that our proposed DPGNN\noutperforms all the latest state-of-the-art models on all datasets, which\nproves the superiority and versatility of our model. The source code of our\nmodel is available at https://github.com.', 'Predicting interactions between structured entities lies at the core of\nnumerous tasks such as drug regimen and new material design. In recent years,\ngraph neural networks have become attractive. They represent structured\nentities as graphs and then extract features from each individual graph using\ngraph convolution operations. However, these methods have some limitations: i)\ntheir networks only extract features from a fix-sized subgraph structure (i.e.,\na fix-sized receptive field) of each node, and ignore features in substructures\nof different sizes, and ii) features are extracted by considering each entity\nindependently, which may not effectively reflect the interaction between two\nentities. To resolve these problems, we present MR-GNN, an end-to-end graph\nneural network with the following features: i) it uses a multi-resolution based\narchitecture to extract node features from different neighborhoods of each\nnode, and, ii) it uses dual graph-state long short-term memory networks\n(L-STMs) to summarize local features of each graph and extracts the interaction\nfeatures between pairwise graphs. Experiments conducted on real-world datasets\nshow that MR-GNN improves the prediction of state-of-the-art methods.', 'Graph neural networks (GNNs) have emerged as a powerful tool for graph\nclassification and representation learning. However, GNNs tend to suffer from\nover-smoothing problems and are vulnerable to graph perturbations. To address\nthese challenges, we propose a novel topological neural framework of\ntopological relational inference (TRI) which allows for integrating\nhigher-order graph information to GNNs and for systematically learning a local\ngraph structure. The key idea is to rewire the original graph by using the\npersistent homology of the small neighborhoods of nodes and then to incorporate\nthe extracted topological summaries as the side information into the local\nalgorithm. As a result, the new framework enables us to harness both the\nconventional information on the graph structure and information on the graph\nhigher order topological properties. We derive theoretical stability guarantees\nfor the new local topological representation and discuss their implications on\nthe graph algebraic connectivity. The experimental results on node\nclassification tasks demonstrate that the new TRI-GNN outperforms all 14\nstate-of-the-art baselines on 6 out 7 graphs and exhibit higher robustness to\nperturbations, yielding up to 10\\% better performance under noisy scenarios.', 'Investigating graph feature learning becomes essentially important with the\nemergence of graph data in many real-world applications. Several graph neural\nnetwork approaches are proposed for node feature learning and they generally\nfollow a neighboring information aggregation scheme to learn node features.\nWhile great performance has been achieved, the weights learning for different\nneighboring nodes is still less explored. In this work, we propose a novel\ngraph network layer, known as Node2Seq, to learn node embeddings with\nexplicitly trainable weights for different neighboring nodes. For a target\nnode, our method sorts its neighboring nodes via attention mechanism and then\nemploys 1D convolutional neural networks (CNNs) to enable explicit weights for\ninformation aggregation. In addition, we propose to incorporate non-local\ninformation for feature learning in an adaptive manner based on the attention\nscores. Experimental results demonstrate the effectiveness of our proposed\nNode2Seq layer and show that the proposed adaptively non-local information\nlearning can improve the performance of feature learning.', 'Graph Neural Networks (GNNs) require a relatively large number of labeled\nnodes and a reliable/uncorrupted graph connectivity structure in order to\nobtain good performance on the semi-supervised node classification task. The\nperformance of GNNs can degrade significantly as the number of labeled nodes\ndecreases or the graph connectivity structure is corrupted by adversarial\nattacks or due to noises in data measurement /collection. Therefore, it is\nimportant to develop GNN models that are able to achieve good performance when\nthere is limited supervision knowledge -- a few labeled nodes and noisy graph\nstructures. In this paper, we propose a novel Dual GNN learning framework to\naddress this challenge task. The proposed framework has two GNN based node\nprediction modules. The primary module uses the input graph structure to induce\nregular node embeddings and predictions with a regular GNN baseline, while the\nauxiliary module constructs a new graph structure through fine-grained spectral\nclusterings and learns new node embeddings and predictions. By integrating the\ntwo modules in a dual GNN learning framework, we perform joint learning in an\nend-to-end fashion. This general framework can be applied on many GNN baseline\nmodels. The experimental results validate that the proposed dual GNN framework\ncan greatly outperform the GNN baseline methods when the labeled nodes are\nscarce and the graph connectivity structure is noisy.', ""In complex networks, nodes that share similar structural characteristics\noften exhibit similar roles (e.g type of users in a social network or the\nhierarchical position of employees in a company). In order to leverage this\nrelationship, a growing literature proposed latent representations that\nidentify structurally equivalent nodes. However, most of the existing methods\nrequire high time and space complexity. In this paper, we propose VNEstruct, a\nsimple approach for generating low-dimensional structural node embeddings, that\nis both time efficient and robust to perturbations of the graph structure. The\nproposed approach focuses on the local neighborhood of each node and employs\nthe Von Neumann entropy, an information-theoretic tool, to extract features\nthat capture the neighborhood's topology. Moreover, on graph classification\ntasks, we suggest the utilization of the generated structural embeddings for\nthe transformation of an attributed graph structure into a set of augmented\nnode attributes. Empirically, we observe that the proposed approach exhibits\nrobustness on structural role identification tasks and state-of-the-art\nperformance on graph classification tasks, while maintaining very high\ncomputational speed."", 'Graph neural networks (GNNs) have been demonstrated to be powerful in\nmodeling graph-structured data. However, training GNNs usually requires\nabundant task-specific labeled data, which is often arduously expensive to\nobtain. One effective way to reduce the labeling effort is to pre-train an\nexpressive GNN model on unlabeled data with self-supervision and then transfer\nthe learned model to downstream tasks with only a few labels. In this paper, we\npresent the GPT-GNN framework to initialize GNNs by generative pre-training.\nGPT-GNN introduces a self-supervised attributed graph generation task to\npre-train a GNN so that it can capture the structural and semantic properties\nof the graph. We factorize the likelihood of the graph generation into two\ncomponents: 1) Attribute Generation and 2) Edge Generation. By modeling both\ncomponents, GPT-GNN captures the inherent dependency between node attributes\nand graph structure during the generative process. Comprehensive experiments on\nthe billion-scale Open Academic Graph and Amazon recommendation data\ndemonstrate that GPT-GNN significantly outperforms state-of-the-art GNN models\nwithout pre-training by up to 9.1% across various downstream tasks.', 'Graph neural network (GNN) has shown convincing performance in learning\npowerful node representations that preserve both node attributes and graph\nstructural information. However, many GNNs encounter problems in effectiveness\nand efficiency when they are designed with a deeper network structure or handle\nlarge-sized graphs. Several sampling algorithms have been proposed for\nimproving and accelerating the training of GNNs, yet they ignore understanding\nthe source of GNN performance gain. The measurement of information within graph\ndata can help the sampling algorithms to keep high-value information while\nremoving redundant information and even noise. In this paper, we propose a\nMetric-Guided (MeGuide) subgraph learning framework for GNNs. MeGuide employs\ntwo novel metrics: Feature Smoothness and Connection Failure Distance to guide\nthe subgraph sampling and mini-batch based training. Feature Smoothness is\ndesigned for analyzing the feature of nodes in order to retain the most\nvaluable information, while Connection Failure Distance can measure the\nstructural information to control the size of subgraphs. We demonstrate the\neffectiveness and efficiency of MeGuide in training various GNNs on multiple\ndatasets.', 'Social and information networks are gaining huge popularity recently due to\ntheir various applications. Knowledge representation through graphs in the form\nof nodes and edges should preserve as many characteristics of the original data\nas possible. Some of the interesting and useful applications on these graphs\nare graph classification, node classification, link prediction, etc. The Graph\nNeural Networks have evolved over the last few years. Graph Neural Networks\n(GNNs) are efficient ways to get insight into large and dynamic graph datasets\ncapturing relationships among billions of entities also known as knowledge\ngraphs.\n  In this paper, we discuss the graph convolutional neural networks graph\nautoencoders and spatio-temporal graph neural networks. The representations of\nthe graph in lower dimensions can be learned using these methods. The\nrepresentations in lower dimensions can be used further for downstream machine\nlearning tasks.', ""In order to perform network analysis tasks, representations that capture the\nmost relevant information in the graph structure are needed. However, existing\nmethods do not learn representations that can be interpreted in a\nstraightforward way and that are robust to perturbations to the graph\nstructure. In this work, we address these two limitations by proposing\nnode2coords, a representation learning algorithm for graphs, which learns\nsimultaneously a low-dimensional space and coordinates for the nodes in that\nspace. The patterns that span the low dimensional space reveal the graph's most\nimportant structural information. The coordinates of the nodes reveal the\nproximity of their local structure to the graph structural patterns. In order\nto measure this proximity by taking into account the underlying graph, we\npropose to use Wasserstein distances. We introduce an autoencoder that employs\na linear layer in the encoder and a novel Wasserstein barycentric layer at the\ndecoder. Node connectivity descriptors, that capture the local structure of the\nnodes, are passed through the encoder to learn the small set of graph\nstructural patterns. In the decoder, the node connectivity descriptors are\nreconstructed as Wasserstein barycenters of the graph structural patterns. The\noptimal weights for the barycenter representation of a node's connectivity\ndescriptor correspond to the coordinates of that node in the low-dimensional\nspace. Experimental results demonstrate that the representations learned with\nnode2coords are interpretable, lead to node embeddings that are stable to\nperturbations of the graph structure and achieve competitive or superior\nresults compared to state-of-the-art methods in node classification."", 'Structure inference is an important task for network data processing and\nanalysis in data science. In recent years, quite a few approaches have been\ndeveloped to learn the graph structure underlying a set of observations\ncaptured in a data space. Although real-world data is often acquired in\nsettings where relationships are influenced by a priori known rules, such\ndomain knowledge is still not well exploited in structure inference problems.\nIn this paper, we identify the structure of signals defined in a data space\nwhose inner relationships are encoded by multi-layer graphs. We aim at properly\nexploiting the information originating from each layer to infer the global\nstructure underlying the signals. We thus present a novel method for combining\nthe multiple graphs into a global graph using mask matrices, which are\nestimated through an optimization problem that accommodates the multi-layer\ngraph information and a signal representation model. The proposed mask\ncombination method also estimates the contribution of each graph layer in the\nstructure of signals. The experiments conducted both on synthetic and\nreal-world data suggest that integrating the multi-layer graph representation\nof the data in the structure inference framework enhances the learning\nprocedure considerably by adapting to the quality and the quantity of the input\ndata.', 'While the celebrated graph neural networks yield effective representations\nfor individual nodes of a graph, there has been relatively less success in\nextending to the task of graph similarity learning. Recent work on graph\nsimilarity learning has considered either global-level graph-graph interactions\nor low-level node-node interactions, however ignoring the rich cross-level\ninteractions (e.g., between each node of one graph and the other whole graph).\nIn this paper, we propose a multi-level graph matching network (MGMN) framework\nfor computing the graph similarity between any pair of graph-structured objects\nin an end-to-end fashion. In particular, the proposed MGMN consists of a\nnode-graph matching network for effectively learning cross-level interactions\nbetween each node of one graph and the other whole graph, and a siamese graph\nneural network to learn global-level interactions between two input graphs.\nFurthermore, to compensate for the lack of standard benchmark datasets, we have\ncreated and collected a set of datasets for both the graph-graph classification\nand graph-graph regression tasks with different sizes in order to evaluate the\neffectiveness and robustness of our models. Comprehensive experiments\ndemonstrate that MGMN consistently outperforms state-of-the-art baseline models\non both the graph-graph classification and graph-graph regression tasks.\nCompared with previous work, MGMN also exhibits stronger robustness as the\nsizes of the two input graphs increase.', 'In this paper, we provide a theory of using graph neural networks (GNNs) for\nmulti-node representation learning (where we are interested in learning a\nrepresentation for a set of more than one node, such as link). We know that GNN\nis designed to learn single-node representations. When we want to learn a node\nset representation involving multiple nodes, a common practice in previous\nworks is to directly aggregate the single-node representations obtained by a\nGNN into a joint node set representation. In this paper, we show a fundamental\nconstraint of such an approach, namely the inability to capture the dependence\nbetween nodes in the node set, and argue that directly aggregating individual\nnode representations does not lead to an effective joint representation for\nmultiple nodes. Then, we notice that a few previous successful works for\nmulti-node representation learning, including SEAL, Distance Encoding, and\nID-GNN, all used node labeling. These methods first label nodes in the graph\naccording to their relationships with the target node set before applying a\nGNN. Then, the node representations obtained in the labeled graph are\naggregated into a node set representation. By investigating their inner\nmechanisms, we unify these node labeling techniques into a single and most\ngeneral form -- labeling trick. We prove that with labeling trick a\nsufficiently expressive GNN learns the most expressive node set\nrepresentations, thus in principle solves any joint learning tasks over node\nsets. Experiments on one important two-node representation learning task, link\nprediction, verified our theory. Our work explains the superior performance of\nprevious node-labeling-based methods, and establishes a theoretical foundation\nof using GNNs for multi-node representation learning.', 'Graph neural networks (GNNs) have achieved superior performance on node\nclassification tasks in the last few years. Commonly, this is framed in a\ntransductive semi-supervised learning setup wherein the entire graph, including\nthe target nodes to be labeled, is available for training. Driven in part by\nscalability, recent works have focused on the inductive case where only the\nlabeled portion of a graph is available for training. In this context, our\ncurrent work considers a challenging inductive setting where a set of labeled\ngraphs are available for training while the unlabeled target graph is\ncompletely separate, i.e., there are no connections between labeled and\nunlabeled nodes. Under the implicit assumption that the testing and training\ngraphs come from similar distributions, our goal is to develop a labeling\nfunction that generalizes to unobserved connectivity structures. To that end,\nwe employ a graph neural tangent kernel (GNTK) that corresponds to infinitely\nwide GNNs to find correspondences between nodes in different graphs based on\nboth the topology and the node features. We augment the capabilities of the\nGNTK with residual connections and empirically illustrate its performance gains\non standard benchmarks.', ""Graph Neural Networks (GNNs), which generalize deep neural networks to\ngraph-structured data, have drawn considerable attention and achieved\nstate-of-the-art performance in numerous graph related tasks. However, existing\nGNN models mainly focus on designing graph convolution operations. The graph\npooling (or downsampling) operations, that play an important role in learning\nhierarchical representations, are usually overlooked. In this paper, we propose\na novel graph pooling operator, called Hierarchical Graph Pooling with\nStructure Learning (HGP-SL), which can be integrated into various graph neural\nnetwork architectures. HGP-SL incorporates graph pooling and structure learning\ninto a unified module to generate hierarchical representations of graphs. More\nspecifically, the graph pooling operation adaptively selects a subset of nodes\nto form an induced subgraph for the subsequent layers. To preserve the\nintegrity of graph's topological information, we further introduce a structure\nlearning mechanism to learn a refined graph structure for the pooled graph at\neach layer. By combining HGP-SL operator with graph neural networks, we perform\ngraph level representation learning with focus on graph classification task.\nExperimental results on six widely used benchmarks demonstrate the\neffectiveness of our proposed model."", 'Driven by the outstanding performance of neural networks in the structured\nEuclidean domain, recent years have seen a surge of interest in developing\nneural networks for graphs and data supported on graphs. The graph is leveraged\nat each layer of the neural network as a parameterization to capture detail at\nthe node level with a reduced number of parameters and computational\ncomplexity. Following this rationale, this paper puts forth a general framework\nthat unifies state-of-the-art graph neural networks (GNNs) through the concept\nof EdgeNet. An EdgeNet is a GNN architecture that allows different nodes to use\ndifferent parameters to weigh the information of different neighbors. By\nextrapolating this strategy to more iterations between neighboring nodes, the\nEdgeNet learns edge- and neighbor-dependent weights to capture local detail.\nThis is a general linear and local operation that a node can perform and\nencompasses under one formulation all existing graph convolutional neural\nnetworks (GCNNs) as well as graph attention networks (GATs). In writing\ndifferent GNN architectures with a common language, EdgeNets highlight specific\narchitecture advantages and limitations, while providing guidelines to improve\ntheir capacity without compromising their local implementation. An interesting\nconclusion is the unification of GCNNs and GATs -- approaches that have been so\nfar perceived as separate. In particular, we show that GATs are GCNNs on a\ngraph that is learned from the features. This particularization opens the doors\nto develop alternative attention mechanisms for improving discriminatory power.', 'Recently, graph neural networks (GNNs) have proved to be suitable in tasks on\nunstructured data. Particularly in tasks as community detection, node\nclassification, and link prediction. However, most GNN models still operate\nwith static relationships. We propose the Graph Learning Network (GLN), a\nsimple yet effective process to learn node embeddings and structure prediction\nfunctions. Our model uses graph convolutions to propose expected node features,\nand predict the best structure based on them. We repeat these steps recursively\nto enhance the prediction and the embeddings.', ""We propose interpretable graph neural networks for sampling and recovery of\ngraph signals, respectively. To take informative measurements, we propose a new\ngraph neural sampling module, which aims to select those vertices that\nmaximally express their corresponding neighborhoods. Such expressiveness can be\nquantified by the mutual information between vertices' features and\nneighborhoods' features, which are estimated via a graph neural network. To\nreconstruct an original graph signal from the sampled measurements, we propose\na graph neural recovery module based on the algorithm-unrolling technique.\nCompared to previous analytical sampling and recovery, the proposed methods are\nable to flexibly learn a variety of graph signal models from data by leveraging\nthe learning ability of neural networks; compared to previous\nneural-network-based sampling and recovery, the proposed methods are designed\nthrough exploiting specific graph properties and provide interpretability. We\nfurther design a new multiscale graph neural network, which is a trainable\nmultiscale graph filter bank and can handle various graph-related learning\ntasks. The multiscale network leverages the proposed graph neural sampling and\nrecovery modules to achieve multiscale representations of a graph. In the\nexperiments, we illustrate the effects of the proposed graph neural sampling\nand recovery modules and find that the modules can flexibly adapt to various\ngraph structures and graph signals. In the task of active-sampling-based\nsemi-supervised learning, the graph neural sampling module improves the\nclassification accuracy over 10% in Cora dataset. We further validate the\nproposed multiscale graph neural network on several standard datasets for both\nvertex and graph classification. The results show that our method consistently\nimproves the classification accuracies."", 'Graph distance metric learning serves as the foundation for many graph\nlearning problems, e.g., graph clustering, graph classification and graph\nmatching. Existing research works on graph distance metric (or graph kernels)\nlearning fail to maintain the basic properties of such metrics, e.g.,\nnon-negative, identity of indiscernibles, symmetry and triangle inequality,\nrespectively. In this paper, we will introduce a new graph neural network based\ndistance metric learning approaches, namely GB-DISTANCE (GRAPH-BERT based\nNeural Distance). Solely based on the attention mechanism, GB-DISTANCE can\nlearn graph instance representations effectively based on a pre-trained\nGRAPH-BERT model. Different from the existing supervised/unsupervised metrics,\nGB-DISTANCE can be learned effectively in a semi-supervised manner. In\naddition, GB-DISTANCE can also maintain the distance metric basic properties\nmentioned above. Extensive experiments have been done on several benchmark\ngraph datasets, and the results demonstrate that GB-DISTANCE can out-perform\nthe existing baseline methods, especially the recent graph neural network model\nbased graph metrics, with a significant gap in computing the graph distance.', ""The majority of popular graph kernels is based on the concept of Haussler's\n$\\mathcal{R}$-convolution kernel and defines graph similarities in terms of\nmutual substructures. In this work, we enrich these similarity measures by\nconsidering graph filtrations: Using meaningful orders on the set of edges,\nwhich allow to construct a sequence of nested graphs, we can consider a graph\nat multiple granularities. For one thing, this provides access to features on\ndifferent levels of resolution. Furthermore, rather than to simply compare\nfrequencies of features in graphs, it allows for their comparison in terms of\nwhen and for how long they exist in the sequences. In this work, we propose a\nfamily of graph kernels that incorporate these existence intervals of features.\nWhile our approach can be applied to arbitrary graph features, we particularly\nhighlight Weisfeiler-Lehman vertex labels, leading to efficient kernels. We\nshow that using Weisfeiler-Lehman labels over certain filtrations strictly\nincreases the expressive power over the ordinary Weisfeiler-Lehman procedure in\nterms of deciding graph isomorphism. In fact, this result directly yields more\npowerful graph kernels based on such features and has implications to graph\nneural networks due to their close relationship to the Weisfeiler-Lehman\nmethod. We empirically validate the expressive power of our graph kernels and\nshow significant improvements over state-of-the-art graph kernels in terms of\npredictive performance on various real-world benchmark datasets."", 'Graph Neural Networks (GNNs) are efficient approaches to process\ngraph-structured data. Modelling long-distance node relations is essential for\nGNN training and applications. However, conventional GNNs suffer from bad\nperformance in modelling long-distance node relations due to limited-layer\ninformation propagation. Existing studies focus on building deep GNN\narchitectures, which face the over-smoothing issue and cannot model node\nrelations in particularly long distance. To address this issue, we propose to\nmodel long-distance node relations by simply relying on shallow GNN\narchitectures with two solutions: (1) Implicitly modelling by learning to\npredict node pair relations (2) Explicitly modelling by adding edges between\nnodes that potentially have the same label. To combine our two solutions, we\npropose a model-agnostic training framework named HighwayGraph, which overcomes\nthe challenge of insufficient labeled nodes by sampling node pairs from the\ntraining set and adopting the self-training method. Extensive experimental\nresults show that our HighwayGraph achieves consistent and significant\nimprovements over four representative GNNs on three benchmark datasets.', 'Graph neural networks (GNNs) have achieved great success on various tasks and\nfields that require relational modeling. GNNs aggregate node features using the\ngraph structure as inductive biases resulting in flexible and powerful models.\nHowever, GNNs remain hard to interpret as the interplay between node features\nand graph structure is only implicitly learned. In this paper, we propose a\nnovel method called Kedge for explicitly sparsifying the underlying graph by\nremoving unnecessary neighbors. Our key idea is based on a tractable method for\nsparsification using the Hard Kumaraswamy distribution that can be used in\nconjugation with any GNN model. Kedge learns edge masks in a modular fashion\ntrained with any GNN allowing for gradient based optimization in an end-to-end\nfashion. We demonstrate through extensive experiments that our model Kedge can\nprune a large proportion of the edges with only a minor effect on the test\naccuracy. Specifically, in the PubMed dataset, Kedge learns to drop more than\n80% of the edges with an accuracy drop of merely 2% showing that graph\nstructure has only a small contribution in comparison to node features.\nFinally, we also show that Kedge effectively counters the over-smoothing\nphenomena in deep GNNs by maintaining good task performance with increasing GNN\nlayers.', 'Graph neural networks (GNNs) provide a powerful and scalable solution for\nmodeling continuous spatial data. However, they often rely on Euclidean\ndistances to construct the input graphs. This assumption can be improbable in\nmany real-world settings, where the spatial structure is more complex and\nexplicitly non-Euclidean (e.g., road networks). Here, we propose PE-GNN, a new\nframework that incorporates spatial context and correlation explicitly into the\nmodels. Building on recent advances in geospatial auxiliary task learning and\nsemantic spatial embeddings, our proposed method (1) learns a context-aware\nvector encoding of the geographic coordinates and (2) predicts spatial\nautocorrelation in the data in parallel with the main task. On spatial\ninterpolation and regression tasks, we show the effectiveness of our approach,\nimproving performance over different state-of-the-art GNN approaches. We\nobserve that our approach not only vastly improves over the GNN baselines, but\ncan match Gaussian processes, the most commonly utilized method for spatial\ninterpolation problems.', ""After observing a snapshot of a social network, a link prediction (LP)\nalgorithm identifies node pairs between which new edges will likely materialize\nin future. Most LP algorithms estimate a score for currently non-neighboring\nnode pairs, and rank them by this score. Recent LP systems compute this score\nby comparing dense, low dimensional vector representations of nodes. Graph\nneural networks (GNNs), in particular graph convolutional networks (GCNs), are\npopular examples. For two nodes to be meaningfully compared, their embeddings\nshould be indifferent to reordering of their neighbors. GNNs typically use\nsimple, symmetric set aggregators to ensure this property, but this design\ndecision has been shown to produce representations with limited expressive\npower. Sequence encoders are more expressive, but are permutation sensitive by\ndesign. Recent efforts to overcome this dilemma turn out to be unsatisfactory\nfor LP tasks. In response, we propose PermGNN, which aggregates neighbor\nfeatures using a recurrent, order-sensitive aggregator and directly minimizes\nan LP loss while it is `attacked' by adversarial generator of neighbor\npermutations. By design, PermGNN{} has more expressive power compared to\nearlier symmetric aggregators. Next, we devise an optimization framework to map\nPermGNN's node embeddings to a suitable locality-sensitive hash, which speeds\nup reporting the top-$K$ most likely edges for the LP task. Our experiments on\ndiverse datasets show that \\our outperforms several state-of-the-art link\npredictors by a significant margin, and can predict the most likely edges fast."", 'From the perspective of expressive power, this work compares multi-layer\nGraph Neural Networks (GNNs) with a simplified alternative that we call\nGraph-Augmented Multi-Layer Perceptrons (GA-MLPs), which first augments node\nfeatures with certain multi-hop operators on the graph and then applies an MLP\nin a node-wise fashion. From the perspective of graph isomorphism testing, we\nshow both theoretically and numerically that GA-MLPs with suitable operators\ncan distinguish almost all non-isomorphic graphs, just like the\nWeifeiler-Lehman (WL) test. However, by viewing them as node-level functions\nand examining the equivalence classes they induce on rooted graphs, we prove a\nseparation in expressive power between GA-MLPs and GNNs that grows\nexponentially in depth. In particular, unlike GNNs, GA-MLPs are unable to count\nthe number of attributed walks. We also demonstrate via community detection\nexperiments that GA-MLPs can be limited by their choice of operator family, as\ncompared to GNNs with higher flexibility in learning.', ""Graph Neural Networks (GNNs) have received increasing attention for\nrepresentation learning in various machine learning tasks. However, most\nexisting GNNs applying neighborhood aggregation usually perform poorly on the\ngraph with heterophily where adjacent nodes belong to different classes. In\nthis paper, we show that in typical heterphilous graphs, the edges may be\ndirected, and whether to treat the edges as is or simply make them undirected\ngreatly affects the performance of the GNN models. Furthermore, due to the\nlimitation of heterophily, it is highly beneficial for the nodes to aggregate\nmessages from similar nodes beyond local neighborhood.These motivate us to\ndevelop a model that adaptively learns the directionality of the graph, and\nexploits the underlying long-distance correlations between nodes. We first\ngeneralize the graph Laplacian to digraph based on the proposed Feature-Aware\nPageRank algorithm, which simultaneously considers the graph directionality and\nlong-distance feature similarity between nodes. Then digraph Laplacian defines\na graph propagation matrix that leads to a model called {\\em DiglacianGCN}.\nBased on this, we further leverage the node proximity measured by commute times\nbetween nodes, in order to preserve the nodes' long-distance correlation on the\ntopology level. Extensive experiments on ten datasets with different levels of\nhomophily demonstrate the effectiveness of our method over existing solutions\nin the task of node classification."", 'Most approaches that tackle the problem of node classification consider nodes\nto be similar, if they have shared neighbors or are close to each other in the\ngraph. Recent methods for attributed graphs additionally take attributes of\nneighboring nodes into account. We argue that the class labels of the neighbors\nbear important information and considering them helps to improve classification\nquality. Two nodes which are similar based on class labels in their\nneighborhood do not need to be close-by in the graph and may even belong to\ndifferent connected components. In this work, we propose a novel approach for\nthe semi-supervised node classification. Precisely, we propose a new node\nembedding which is based on the class labels in the local neighborhood of a\nnode. We show that this is a different setting from attribute-based embeddings\nand thus, we propose a new method to learn label-based node embeddings which\ncan mirror a variety of relations between the class labels of neighboring\nnodes. Our experimental evaluation demonstrates that our new methods can\nsignificantly improve the prediction quality on real world data sets.', 'Graph Neural Networks (GNNs) have achieved much success on graph-structured\ndata. In light of this, there have been increasing interests in studying their\nexpressive power. One line of work studies the capability of GNNs to\napproximate permutation-invariant functions on graphs, and another focuses on\nthe their power as tests for graph isomorphism. Our work connects these two\nperspectives and proves their equivalence. We further develop a framework of\nthe expressive power of GNNs that incorporates both of these viewpoints using\nthe language of sigma-algebra, through which we compare the expressive power of\ndifferent types of GNNs together with other graph isomorphism tests. In\nparticular, we prove that the second-order Invariant Graph Network fails to\ndistinguish non-isomorphic regular graphs with the same degree. Then, we extend\nit to a new architecture, Ring-GNN, which succeeds in distinguishing these\ngraphs and achieves good performances on real-world datasets.', 'Graph neural networks for heterogeneous graph embedding is to project nodes\ninto a low-dimensional space by exploring the heterogeneity and semantics of\nthe heterogeneous graph. However, on the one hand, most of existing\nheterogeneous graph embedding methods either insufficiently model the local\nstructure under specific semantic, or neglect the heterogeneity when\naggregating information from it. On the other hand, representations from\nmultiple semantics are not comprehensively integrated to obtain versatile node\nembeddings. To address the problem, we propose a Heterogeneous Graph Neural\nNetwork with Multi-View Representation Learning (named MV-HetGNN) for\nheterogeneous graph embedding by introducing the idea of multi-view\nrepresentation learning. The proposed model consists of node feature\ntransformation, view-specific ego graph encoding and auto multi-view fusion to\nthoroughly learn complex structural and semantic information for generating\ncomprehensive node representations. Extensive experiments on three real-world\nheterogeneous graph datasets show that the proposed MV-HetGNN model\nconsistently outperforms all the state-of-the-art GNN baselines in various\ndownstream tasks, e.g., node classification, node clustering, and link\nprediction.', 'Graph neural networks (GNNs) are shown to be successful in modeling\napplications with graph structures. However, training an accurate GNN model\nrequires a large collection of labeled data and expressive features, which\nmight be inaccessible for some applications. To tackle this problem, we propose\na pre-training framework that captures generic graph structural information\nthat is transferable across tasks. Our framework can leverage the following\nthree tasks: 1) denoising link reconstruction, 2) centrality score ranking, and\n3) cluster preserving. The pre-training procedure can be conducted purely on\nthe synthetic graphs, and the pre-trained GNN is then adapted for downstream\napplications. With the proposed pre-training procedure, the generic structural\ninformation is learned and preserved, thus the pre-trained GNN requires less\namount of labeled data and fewer domain-specific features to achieve high\nperformance on different downstream tasks. Comprehensive experiments\ndemonstrate that our proposed framework can significantly enhance the\nperformance of various tasks at the level of node, link, and graph.']","['Graph neural networks exhibit remarkable performance in graph data analysis.\nHowever, the robustness of GNN models remains a challenge. As a result, they\nare not reliable enough to be deployed in critical applications. Recent studies\ndemonstrate that GNNs could be easily fooled with adversarial perturbations,\nespecially structural perturbations. Such vulnerability is attributed to the\nexcessive dependence on the structure information to make predictions. To\nachieve better robustness, it is desirable to build the prediction of GNNs with\nmore comprehensive features. Graph data, in most cases, has two views of\ninformation, namely structure information and feature information. In this\npaper, we propose CoG, a simple yet effective co-training framework to combine\nthese two views for the purpose of robustness. CoG trains sub-models from the\nfeature view and the structure view independently and allows them to distill\nknowledge from each other by adding their most confident unlabeled data into\nthe training set. The orthogonality of these two views diversifies the\nsub-models, thus enhancing the robustness of their ensemble. We evaluate our\nframework on three popular datasets, and results show that CoG significantly\nimproves the robustness of graph models against adversarial attacks without\nsacrificing their performance on clean data. We also show that CoG still\nachieves good robustness when both node features and graph structures are\nperturbed.', ""Real data collected from different applications that have additional\ntopological structures and connection information are amenable to be\nrepresented as a weighted graph. Considering the node labeling problem, Graph\nNeural Networks (GNNs) is a powerful tool, which can mimic experts' decision on\nnode labeling. GNNs combine node features, connection patterns, and graph\nstructure by using a neural network to embed node information and pass it\nthrough edges in the graph. We want to identify the patterns in the input data\nused by the GNN model to make a decision and examine if the model works as we\ndesire. However, due to the complex data representation and non-linear\ntransformations, explaining decisions made by GNNs is challenging. In this\nwork, we propose new graph features' explanation methods to identify the\ninformative components and important node features. Besides, we propose a\npipeline to identify the key factors used for node classification. We use four\ndatasets (two synthetic and two real) to validate our methods. Our results\ndemonstrate that our explanation approach can mimic data patterns used for node\nclassification by human interpretation and disentangle different features in\nthe graphs. Furthermore, our explanation methods can be used for understanding\ndata, debugging GNN models, and examine model decisions."", 'Recently, Graph Neural Networks (GNNs) have greatly advanced the task of\ngraph classification. Typically, we first build a unified GNN model with graphs\nin a given training set and then use this unified model to predict labels of\nall the unseen graphs in the test set. However, graphs in the same dataset\noften have dramatically distinct structures, which indicates that a unified\nmodel may be sub-optimal given an individual graph. Therefore, in this paper,\nwe aim to develop customized graph neural networks for graph classification.\nSpecifically, we propose a novel customized graph neural network framework,\ni.e., Customized-GNN. Given a graph sample, Customized-GNN can generate a\nsample-specific model for this graph based on its structure. Meanwhile, the\nproposed framework is very general that can be applied to numerous existing\ngraph neural network models. Comprehensive experiments on various graph\nclassification benchmarks demonstrate the effectiveness of the proposed\nframework.', 'In recent years, graph neural networks (GNNs) have gained increasing\npopularity and have shown very promising results for data that are represented\nby graphs. The majority of GNN architectures are designed based on developing\nnew convolutional and/or pooling layers that better extract the hidden and\ndeeper representations of the graphs to be used for different prediction tasks.\nThe inputs to these layers are mainly the three default descriptors of a graph,\nnode features $(X)$, adjacency matrix $(A)$, and edge features $(W)$ (if\navailable). To provide a more enriched input to the network, we propose a\nrandom walk data processing of the graphs based on three selected lengths.\nNamely, (regular) walks of length 1 and 2, and a fractional walk of length\n$\\gamma \\in (0,1)$, in order to capture the different local and global dynamics\non the graphs. We also calculate the stationary distribution of each random\nwalk, which is then used as a scaling factor for the initial node features\n($X$). This way, for each graph, the network receives multiple adjacency\nmatrices along with their individual weighting for the node features. We test\nour method on various molecular datasets by passing the processed node features\nto the network in order to perform several classification and regression tasks.\nInterestingly, our method, not using edge features which are heavily exploited\nin molecular graph learning, let a shallow network outperform well known deep\nGNNs.', 'Graphs play an important role in many applications. Recently, Graph Neural\nNetworks (GNNs) have achieved promising results in graph analysis tasks. Some\nstate-of-the-art GNN models have been proposed, e.g., Graph Convolutional\nNetworks (GCNs), Graph Attention Networks (GATs), etc. Despite these successes,\nmost of the GNNs only have shallow structure. This causes the low expressive\npower of the GNNs. To fully utilize the power of the deep neural network, some\ndeep GNNs have been proposed recently. However, the design of deep GNNs\nrequires significant architecture engineering. In this work, we propose a\nmethod to automate the deep GNNs design. In our proposed method, we add a new\ntype of skip connection to the GNNs search space to encourage feature reuse and\nalleviate the vanishing gradient problem. We also allow our evolutionary\nalgorithm to increase the layers of GNNs during the evolution to generate\ndeeper networks. We evaluate our method in the graph node classification task.\nThe experiments show that the GNNs generated by our method can obtain\nstate-of-the-art results in Cora, Citeseer, Pubmed and PPI datasets.', 'Controversial content refers to any content that attracts both positive and\nnegative feedback. Its automatic identification, especially on social media, is\na challenging task as it should be done on a large number of continuously\nevolving posts, covering a large variety of topics. Most of the existing\napproaches rely on the graph structure of a topic-discussion and/or the content\nof messages. This paper proposes a controversy detection approach based on both\ngraph structure of a discussion and text features. Our proposed approach relies\non Graph Neural Network (gnn) to encode the graph representation (including its\ntexts) in an embedding vector before performing a graph classification task.\nThe latter will classify the post as controversial or not. Two controversy\ndetection strategies are proposed. The first one is based on a hierarchical\ngraph representation learning. Graph user nodes are embedded hierarchically and\niteratively to compute the whole graph embedding vector. The second one is\nbased on the attention mechanism, which allows each user node to give more or\nless importance to its neighbors when computing node embeddings. We conduct\nexperiments to evaluate our approach using different real-world datasets.\nConducted experiments show the positive impact of combining textual features\nand structural information in terms of performance.', 'Graph neural networks (GNNs) have significantly improved the representation\npower for graph-structured data. Despite of the recent success of GNNs, the\ngraph convolution in most GNNs have two limitations. Since the graph\nconvolution is performed in a small local neighborhood on the input graph, it\nis inherently incapable to capture long-range dependencies between distance\nnodes. In addition, when a node has neighbors that belong to different classes,\ni.e., heterophily, the aggregated messages from them often negatively affect\nrepresentation learning. To address the two common problems of graph\nconvolution, in this paper, we propose Deformable Graph Convolutional Networks\n(Deformable GCNs) that adaptively perform convolution in multiple latent spaces\nand capture short/long-range dependencies between nodes. Separated from node\nrepresentations (features), our framework simultaneously learns the node\npositional embeddings (coordinates) to determine the relations between nodes in\nan end-to-end fashion. Depending on node position, the convolution kernels are\ndeformed by deformation vectors and apply different transformations to its\nneighbor nodes. Our extensive experiments demonstrate that Deformable GCNs\nflexibly handles the heterophily and achieve the best performance in node\nclassification tasks on six heterophilic graph datasets.', 'Graph neural networks (GNNs) rely on graph convolutions to extract local\nfeatures from network data. These graph convolutions combine information from\nadjacent nodes using coefficients that are shared across all nodes. Since these\ncoefficients are shared and do not depend on the graph, one can envision using\nthe same coefficients to define a GNN on another graph. This motivates\nanalyzing the transferability of GNNs across graphs. In this paper we introduce\ngraphon NNs as limit objects of GNNs and prove a bound on the difference\nbetween the output of a GNN and its limit graphon-NN. This bound vanishes with\ngrowing number of nodes if the graph convolutional filters are bandlimited in\nthe graph spectral domain. This result establishes a tradeoff between\ndiscriminability and transferability of GNNs.', 'Existing popular methods for semi-supervised learning with Graph Neural\nNetworks (such as the Graph Convolutional Network) provably cannot learn a\ngeneral class of neighborhood mixing relationships. To address this weakness,\nwe propose a new model, MixHop, that can learn these relationships, including\ndifference operators, by repeatedly mixing feature representations of neighbors\nat various distances. Mixhop requires no additional memory or computational\ncomplexity, and outperforms on challenging baselines. In addition, we propose\nsparsity regularization that allows us to visualize how the network prioritizes\nneighborhood information across different graph datasets. Our analysis of the\nlearned architectures reveals that neighborhood mixing varies per datasets.', 'Graph neural networks (GNNs) are important tools for transductive learning\ntasks, such as node classification in graphs, due to their expressive power in\ncapturing complex interdependency between nodes. To enable graph neural network\nlearning, existing works typically assume that labeled nodes, from two or\nmultiple classes, are provided, so that a discriminative classifier can be\nlearned from the labeled data. In reality, this assumption might be too\nrestrictive for applications, as users may only provide labels of interest in a\nsingle class for a small number of nodes. In addition, most GNN models only\naggregate information from short distances (e.g., 1-hop neighbors) in each\nround, and fail to capture long distance relationship in graphs. In this paper,\nwe propose a novel graph neural network framework, long-short distance\naggregation networks (LSDAN), to overcome these limitations. By generating\nmultiple graphs at different distance levels, based on the adjacency matrix, we\ndevelop a long-short distance attention model to model these graphs. The direct\nneighbors are captured via a short-distance attention mechanism, and neighbors\nwith long distance are captured by a long distance attention mechanism. Two\nnovel risk estimators are further employed to aggregate long-short-distance\nnetworks, for PU learning and the loss is back-propagated for model learning.\nExperimental results on real-world datasets demonstrate the effectiveness of\nour algorithm.', 'Graph Representation Learning (GRL) has experienced significant progress as a\nmeans to extract structural information in a meaningful way for subsequent\nlearning tasks. Current approaches including shallow embeddings and Graph\nNeural Networks have mostly been tested with node classification and link\nprediction tasks. In this work, we provide an application oriented perspective\nto a set of popular embedding approaches and evaluate their representational\npower with respect to real-world graph properties. We implement an extensive\nempirical data-driven framework to challenge existing norms regarding the\nexpressive power of embedding approaches in graphs with varying patterns along\nwith a theoretical analysis of the limitations we discovered in this process.\nOur results suggest that ""one-to-fit-all"" GRL approaches are hard to define in\nreal-world scenarios and as new methods are being introduced they should be\nexplicit about their ability to capture graph properties and their\napplicability in datasets with non-trivial structural differences.', 'The nodes of a graph existing in a cluster are more likely to connect to each\nother than with other nodes in the graph. Then revealing some information about\nsome nodes, the structure of the graph (graph edges) provides this opportunity\nto know more information about other nodes. From this perspective, this paper\nrevisits the node classification task in a semi-supervised scenario by graph\nconvolutional networks (GCNs). The goal is to benefit from the flow of\ninformation that circulates around the revealed node labels. The contribution\nof this paper is twofold. First, this paper provides a method for extracting\nside information from a graph realization. Then a new GCN architecture is\npresented that combines the output of traditional GCN and the extracted side\ninformation. Another contribution of this paper is relevant to non-graph\nobservations (independent side information) that exists beside a graph\nrealization in many applications. Indeed, the extracted side information can be\nreplaced by a sequence of side information that is independent of the graph\nstructure. For both cases, the experiments on synthetic and real-world datasets\ndemonstrate that the proposed model achieves a higher prediction accuracy in\ncomparison to the existing state-of-the-art methods for the node classification\ntask.', ""Graphs are the most ubiquitous form of structured data representation used in\nmachine learning. They model, however, only pairwise relations between nodes\nand are not designed for encoding the higher-order relations found in many\nreal-world datasets. To model such complex relations, hypergraphs have proven\nto be a natural representation. Learning the node representations in a\nhypergraph is more complex than in a graph as it involves information\npropagation at two levels: within every hyperedge and across the hyperedges.\nMost current approaches first transform a hypergraph structure to a graph for\nuse in existing geometric deep learning algorithms. This transformation leads\nto information loss, and sub-optimal exploitation of the hypergraph's\nexpressive power. We present HyperSAGE, a novel hypergraph learning framework\nthat uses a two-level neural message passing strategy to accurately and\nefficiently propagate information through hypergraphs. The flexible design of\nHyperSAGE facilitates different ways of aggregating neighborhood information.\nUnlike the majority of related work which is transductive, our approach,\ninspired by the popular GraphSAGE method, is inductive. Thus, it can also be\nused on previously unseen nodes, facilitating deployment in problems such as\nevolving or partially observed hypergraphs. Through extensive experimentation,\nwe show that HyperSAGE outperforms state-of-the-art hypergraph learning methods\non representative benchmark datasets. We also demonstrate that the higher\nexpressive power of HyperSAGE makes it more stable in learning node\nrepresentations as compared to the alternatives."", 'In this paper, we propose an end-to-end graph learning framework, namely Deep\nIterative and Adaptive Learning for Graph Neural Networks (DIAL-GNN), for\njointly learning the graph structure and graph embeddings simultaneously. We\nfirst cast the graph structure learning problem as a similarity metric learning\nproblem and leverage an adapted graph regularization for controlling\nsmoothness, connectivity and sparsity of the generated graph. We further\npropose a novel iterative method for searching for a hidden graph structure\nthat augments the initial graph structure. Our iterative method dynamically\nstops when the learned graph structure approaches close enough to the optimal\ngraph. Our extensive experiments demonstrate that the proposed DIAL-GNN model\ncan consistently outperform or match state-of-the-art baselines in terms of\nboth downstream task performance and computational time. The proposed approach\ncan cope with both transductive learning and inductive learning.', 'Graph Neural Networks (GNNs) have been a latest hot research topic in data\nscience, due to the fact that they use the ubiquitous data structure graphs as\nthe underlying elements for constructing and training neural networks. In a\nGNN, each node has numerous features associated with it. The entire task (for\nexample, classification, or clustering) utilizes the features of the nodes to\nmake decisions, at node level or graph level. In this paper, (1) we extend the\nfeature selection algorithm presented in via Gumbel Softmax to GNNs. We conduct\na series of experiments on our feature selection algorithms, using various\nbenchmark datasets: Cora, Citeseer and Pubmed. (2) We implement a mechanism to\nrank the extracted features. We demonstrate the effectiveness of our\nalgorithms, for both feature selection and ranking. For the Cora dataset, (1)\nwe use the algorithm to select 225 features out of 1433 features. Our\nexperimental results demonstrate their effectiveness for the same\nclassification problem. (2) We extract features such that they are linear\ncombinations of the original features, where the coefficients for each\nextracted features are non-negative and sum up to one. We propose an algorithm\nto rank the extracted features in the sense that when using them for the same\nclassification problem, the accuracy goes down gradually for the extracted\nfeatures within the rank 1 - 50, 51 - 100, 100 - 150, and 151 - 200.', 'Graph Neural Network (GNN) has been demonstrated its effectiveness in dealing\nwith non-Euclidean structural data. Both spatial-based and spectral-based GNNs\nare relying on adjacency matrix to guide message passing among neighbors during\nfeature aggregation. Recent works have mainly focused on powerful message\npassing modules, however, in this paper, we show that none of the message\npassing modules is necessary. Instead, we propose a pure\nmultilayer-perceptron-based framework, Graph-MLP with the supervision signal\nleveraging graph structure, which is sufficient for learning discriminative\nnode representation. In model-level, Graph-MLP only includes multi-layer\nperceptrons, activation function, and layer normalization. In the loss level,\nwe design a neighboring contrastive (NContrast) loss to bridge the gap between\nGNNs and MLPs by utilizing the adjacency information implicitly. This design\nallows our model to be lighter and more robust when facing large-scale graph\ndata and corrupted adjacency information. Extensive experiments prove that even\nwithout adjacency information in testing phase, our framework can still reach\ncomparable and even superior performance against the state-of-the-art models in\nthe graph node classification task.', ""Graph representation learning methods generate numerical vector\nrepresentations for the nodes in a network, thereby enabling their use in\nstandard machine learning models. These methods aim to preserve relational\ninformation, such that nodes that are similar in the graph are found close to\none another in the representation space. Similarity can be based largely on one\nof two notions: connectivity or structural role. In tasks where node structural\nrole is important, connectivity based methods show poor performance. Recent\nwork has begun to focus on scalability of learning methods to massive graphs of\nmillions to billions of nodes and edges. Many unsupervised node representation\nlearning algorithms are incapable of scaling to large graphs, and are unable to\ngenerate node representations for unseen nodes. In this work, we propose\nInferential SIR-GN, a model which is pre-trained on random graphs, then\ncomputes node representations rapidly, including for very large networks. We\ndemonstrate that the model is able to capture node's structural role\ninformation, and show excellent performance at node and graph classification\ntasks, on unseen networks. Additionally, we observe the scalability of\nInferential SIR-GN is comparable to the fastest current approaches for massive\ngraphs."", 'Recently, graph neural networks (GNNs) have revolutionized the field of graph\nrepresentation learning through effectively learned node embeddings, and\nachieved state-of-the-art results in tasks such as node classification and link\nprediction. However, current GNN methods are inherently flat and do not learn\nhierarchical representations of graphs---a limitation that is especially\nproblematic for the task of graph classification, where the goal is to predict\nthe label associated with an entire graph. Here we propose DiffPool, a\ndifferentiable graph pooling module that can generate hierarchical\nrepresentations of graphs and can be combined with various graph neural network\narchitectures in an end-to-end fashion. DiffPool learns a differentiable soft\ncluster assignment for nodes at each layer of a deep GNN, mapping nodes to a\nset of clusters, which then form the coarsened input for the next GNN layer.\nOur experimental results show that combining existing GNN methods with DiffPool\nyields an average improvement of 5-10% accuracy on graph classification\nbenchmarks, compared to all existing pooling approaches, achieving a new\nstate-of-the-art on four out of five benchmark data sets.', 'Graph neural network (GNN) is a deep model for graph representation learning.\nOne advantage of graph neural network is its ability to incorporate node\nfeatures into the learning process. However, this prevents graph neural network\nfrom being applied into featureless graphs. In this paper, we first analyze the\neffects of node features on the performance of graph neural network. We show\nthat GNNs work well if there is a strong correlation between node features and\nnode labels. Based on these results, we propose new feature initialization\nmethods that allows to apply graph neural network to non-attributed graphs. Our\nexperimental results show that the artificial features are highly competitive\nwith real features.', 'Recently, there has been great success in applying deep neural networks on\ngraph structured data. Most work, however, focuses on either node- or\ngraph-level supervised learning, such as node, link or graph classification or\nnode-level unsupervised learning (e.g. node clustering). Despite its wide range\nof possible applications, graph-level unsupervised learning has not received\nmuch attention yet. This might be mainly attributed to the high representation\ncomplexity of graphs, which can be represented by n! equivalent adjacency\nmatrices, where n is the number of nodes. In this work we address this issue by\nproposing a permutation-invariant variational autoencoder for graph structured\ndata. Our proposed model indirectly learns to match the node ordering of input\nand output graph, without imposing a particular node ordering or performing\nexpensive graph matching. We demonstrate the effectiveness of our proposed\nmodel on various graph reconstruction and generation tasks and evaluate the\nexpressive power of extracted representations for downstream graph-level\nclassification and regression.', 'Graph embedding techniques allow to learn high-quality feature vectors from\ngraph structures and are useful in a variety of tasks, from node classification\nto clustering. Existing approaches have only focused on learning feature\nvectors for the nodes in a (knowledge) graph. To the best of our knowledge,\nnone of them has tackled the problem of embedding of graph edges, that is,\nknowledge graph triples. The approaches that are closer to this task have\nfocused on homogeneous graphs involving only one type of edge and obtain edge\nembeddings by applying some operation (e.g., average) on the embeddings of the\nendpoint nodes. The goal of this paper is to introduce Triple2Vec, a new\ntechnique to directly embed edges in (knowledge) graphs. Trple2Vec builds upon\nthree main ingredients. The first is the notion of line graph. The line graph\nof a graph is another graph representing the adjacency between edges of the\noriginal graph. In particular, the nodes of the line graph are the edges of the\noriginal graph. We show that directly applying existing embedding techniques on\nthe nodes of the line graph to learn edge embeddings is not enough in the\ncontext of knowledge graphs. Thus, we introduce the notion of triple line\ngraph. The second is an edge weighting mechanism both for line graphs derived\nfrom knowledge graphs and homogeneous graphs. The third is a strategy based on\ngraph walks on the weighted triple line graph that can preserve proximity\nbetween nodes. Embeddings are finally generated by adopting the SkipGram model,\nwhere sentences are replaced with graph walks. We evaluate our approach on\ndifferent real world (knowledge) graphs and compared it with related work.', 'The expressive power of graph neural network formalisms is commonly measured\nby their ability to distinguish graphs. For many formalisms, the k-dimensional\nWeisfeiler-Leman (k-WL) graph isomorphism test is used as a yardstick. In this\npaper we consider the expressive power of kth-order invariant (linear) graph\nnetworks (k-IGNs). It is known that k-IGNs are expressive enough to simulate\nk-WL. This means that for any two graphs that can be distinguished by k-WL, one\ncan find a k-IGN which also distinguishes those graphs. The question remains\nwhether k-IGNs can distinguish more graphs than k-WL. This was recently shown\nto be false for k=2. Here, we generalise this result to arbitrary k. In other\nwords, we show that k-IGNs are bounded in expressive power by k-WL. This\nimplies that k-IGNs and k-WL are equally powerful in distinguishing graphs.', 'Graph embedding is a transformation of nodes of a network into a set of\nvectors. A good embedding should capture the underlying graph topology and\nstructure, node-to-node relationship, and other relevant information about the\ngraph, its subgraphs, and nodes themselves. If these objectives are achieved,\nan embedding is a meaningful, understandable, and often compressed\nrepresentation of a network. Unfortunately, selecting the best embedding is a\nchallenging task and very often requires domain experts. In this paper, we\nextend the framework for evaluating graph embeddings that was recently\nintroduced by the authors. Now, the framework assigns two scores, local and\nglobal, to each embedding that measure the quality of an evaluated embedding\nfor tasks that require good representation of local and, respectively, global\nproperties of the network. The best embedding, if needed, can be selected in an\nunsupervised way, or the framework can identify a few embeddings that are worth\nfurther investigation. The framework is flexible, scalable, and can deal with\nundirected/directed, weighted/unweighted graphs.', 'Tensor-network techniques have enjoyed outstanding success in physics, and\nhave recently attracted attention in machine learning, both as a tool for the\nformulation of new learning algorithms and for enhancing the mathematical\nunderstanding of existing methods. Inspired by these developments, and the\nnatural correspondence between tensor networks and probabilistic graphical\nmodels, we provide a rigorous analysis of the expressive power of various\ntensor-network factorizations of discrete multivariate probability\ndistributions. These factorizations include non-negative tensor-trains/MPS,\nwhich are in correspondence with hidden Markov models, and Born machines, which\nare naturally related to local quantum circuits. When used to model probability\ndistributions, they exhibit tractable likelihoods and admit efficient learning\nalgorithms. Interestingly, we prove that there exist probability distributions\nfor which there are unbounded separations between the resource requirements of\nsome of these tensor-network factorizations. Particularly surprising is the\nfact that using complex instead of real tensors can lead to an arbitrarily\nlarge reduction in the number of parameters of the network. Additionally, we\nintroduce locally purified states (LPS), a new factorization inspired by\ntechniques for the simulation of quantum systems, with provably better\nexpressive power than all other representations considered. The ramifications\nof this result are explored through numerical experiments. Our findings imply\nthat LPS should be considered over hidden Markov models, and furthermore\nprovide guidelines for the design of local quantum circuits for probabilistic\nmodeling.', 'Improving the resilience of a network is a fundamental problem in network\nscience, which protects the underlying system from natural disasters and\nmalicious attacks. This is traditionally achieved via successive\ndegree-preserving edge rewiring operations, with the major limitation of being\ntransductive. Inductively solving graph-related tasks with sequential actions\nis accomplished by adopting graph neural networks (GNNs) coupled with\nreinforcement learning under the scenario with rich graph features. However,\nsuch frameworks cannot be directly applied to resilience tasks where only pure\ntopological structure is available. In this case, GNNs can barely learn useful\ninformation, resulting in prohibitive difficulty in making actions for\nsuccessively rewiring edges under a reinforcement learning context. In this\npaper, we study in depth the reasons why typical GNNs cause such failure. Based\non this investigation, we propose ResiNet, the first end-to-end trainable\ninductive framework to discover resilient network topologies while balancing\nnetwork utility. To this end, we reformulate resilience optimization as an MDP\nequipped with edge rewiring action space, and propose a pure topology-oriented\nvariant of GNN called filtration enhanced graph neural network (FireGNN), which\ncan learn from graphs without rich features. Extensive experiments demonstrate\nthat ResiNet achieves a near-optimal resilience gain on various graphs while\nbalancing the utility, and outperforms existing approaches by a large margin.', 'Graph Neural Networks (GNNs), which generalize the deep neural networks to\ngraph-structured data, have achieved great success in modeling graphs. However,\nas an extension of deep learning for graphs, GNNs lack explainability, which\nlargely limits their adoption in scenarios that demand the transparency of\nmodels. Though many efforts are taken to improve the explainability of deep\nlearning, they mainly focus on i.i.d data, which cannot be directly applied to\nexplain the predictions of GNNs because GNNs utilize both node features and\ngraph topology to make predictions. There are only very few work on the\nexplainability of GNNs and they focus on post-hoc explanations. Since post-hoc\nexplanations are not directly obtained from the GNNs, they can be biased and\nmisrepresent the true explanations. Therefore, in this paper, we study a novel\nproblem of self-explainable GNNs which can simultaneously give predictions and\nexplanations. We propose a new framework which can find $K$-nearest labeled\nnodes for each unlabeled node to give explainable node classification, where\nnearest labeled nodes are found by interpretable similarity module in terms of\nboth node similarity and local structure similarity. Extensive experiments on\nreal-world and synthetic datasets demonstrate the effectiveness of the proposed\nframework for explainable node classification.', 'Graph embedding methods are becoming increasingly popular in the machine\nlearning community, where they are widely used for tasks such as node\nclassification and link prediction. Embedding graphs in geometric spaces should\naid the identification of network communities as well, because nodes in the\nsame community should be projected close to each other in the geometric space,\nwhere they can be detected via standard data clustering algorithms. In this\npaper, we test the ability of several graph embedding techniques to detect\ncommunities on benchmark graphs. We compare their performance against that of\ntraditional community detection algorithms. We find that the performance is\ncomparable, if the parameters of the embedding techniques are suitably chosen.\nHowever, the optimal parameter set varies with the specific features of the\nbenchmark graphs, like their size, whereas popular community detection\nalgorithms do not require any parameter. So it is not possible to indicate\nbeforehand good parameter sets for the analysis of real networks. This finding,\nalong with the high computational cost of embedding a network and grouping the\npoints, suggests that, for community detection, current embedding techniques do\nnot represent an improvement over network clustering algorithms.', 'Many real-world problems can be formalized as predicting links in a partially\nobserved network. Examples include Facebook friendship suggestions,\nconsumer-product recommendations, and the identification of hidden interactions\nbetween actors in a crime network. Several link prediction algorithms, notably\nthose recently introduced using network embedding, are capable of doing this by\njust relying on the observed part of the network. Often, the link status of a\nnode pair can be queried, which can be used as additional information by the\nlink prediction algorithm. Unfortunately, such queries can be expensive or\ntime-consuming, mandating the careful consideration of which node pairs to\nquery. In this paper we estimate the improvement in link prediction accuracy\nafter querying any particular node pair, to use in an active learning setup.\nSpecifically, we propose ALPINE (Active Link Prediction usIng Network\nEmbedding), the first method to achieve this for link prediction based on\nnetwork embedding. To this end, we generalized the notion of V-optimality from\nexperimental design to this setting, as well as more basic active learning\nheuristics originally developed in standard classification settings. Empirical\nresults on real data show that ALPINE is scalable, and boosts link prediction\naccuracy with far fewer queries.', 'Networks provide a powerful formalism for modeling complex systems by using a\nmodel of pairwise interactions. But much of the structure within these systems\ninvolves interactions that take place among more than two nodes at once; for\nexample, communication within a group rather than person-to person,\ncollaboration among a team rather than a pair of coauthors, or biological\ninteraction between a set of molecules rather than just two. Such higher-order\ninteractions are ubiquitous, but their empirical study has received limited\nattention, and little is known about possible organizational principles of such\nstructures. Here we study the temporal evolution of 19 datasets with explicit\naccounting for higher-order interactions. We show that there is a rich variety\nof structure in our datasets but datasets from the same system types have\nconsistent patterns of higher-order structure. Furthermore, we find that tie\nstrength and edge density are competing positive indicators of higher-order\norganization, and these trends are consistent across interactions involving\ndiffering numbers of nodes. To systematically further the study of theories for\nsuch higher-order structures, we propose higher-order link prediction as a\nbenchmark problem to assess models and algorithms that predict higher-order\nstructure. We find a fundamental differences from traditional pairwise link\nprediction, with a greater role for local rather than long-range information in\npredicting the appearance of new interactions.', 'Graph neural networks provide a powerful toolkit for embedding real-world\ngraphs into low-dimensional spaces according to specific tasks. Up to now,\nthere have been several surveys on this topic. However, they usually lay\nemphasis on different angles so that the readers can not see a panorama of the\ngraph neural networks. This survey aims to overcome this limitation, and\nprovide a comprehensive review on the graph neural networks. First of all, we\nprovide a novel taxonomy for the graph neural networks, and then refer to up to\n400 relevant literatures to show the panorama of the graph neural networks. All\nof them are classified into the corresponding categories. In order to drive the\ngraph neural networks into a new stage, we summarize four future research\ndirections so as to overcome the facing challenges. It is expected that more\nand more scholars can understand and exploit the graph neural networks, and use\nthem in their research community.', ""The traditional setup of link prediction in networks assumes that a test set\nof node pairs, which is usually balanced, is available over which to predict\nthe presence of links. However, in practice, there is no test set: the\nground-truth is not known, so the number of possible pairs to predict over is\nquadratic in the number of nodes in the graph. Moreover, because graphs are\nsparse, most of these possible pairs will not be links. Thus, link prediction\nmethods, which often rely on proximity-preserving embeddings or heuristic\nnotions of node similarity, face a vast search space, with many pairs that are\nin close proximity, but that should not be linked. To mitigate this issue, we\nintroduce LinkWaldo, a framework for choosing from this quadratic,\nmassively-skewed search space of node pairs, a concise set of candidate pairs\nthat, in addition to being in close proximity, also structurally resemble the\nobserved edges. This allows it to ignore some high-proximity but\nlow-resemblance pairs, and also identify high-resemblance, lower-proximity\npairs. Our framework is built on a model that theoretically combines Stochastic\nBlock Models (SBMs) with node proximity models. The block structure of the SBM\nmaps out where in the search space new links are expected to fall, and the\nproximity identifies the most plausible links within these blocks, using\nlocality sensitive hashing to avoid expensive exhaustive search. LinkWaldo can\nuse any node representation learning or heuristic definition of proximity, and\ncan generate candidate pairs for any link prediction method, allowing the\nrepresentation power of current and future methods to be realized for link\nprediction in practice. We evaluate LinkWaldo on 13 networks across multiple\ndomains, and show that on average it returns candidate sets containing 7-33%\nmore missing and future links than both embedding-based and heuristic\nbaselines' sets."", ""Data collection often involves the partial measurement of a larger system. A\ncommon example arises in collecting network data: we often obtain network\ndatasets by recording all of the interactions among a small set of core nodes,\nso that we end up with a measurement of the network consisting of these core\nnodes along with a potentially much larger set of fringe nodes that have links\nto the core. Given the ubiquity of this process for assembling network data, it\nis crucial to understand the role of such a `core-fringe' structure.\n  Here we study how the inclusion of fringe nodes affects the standard task of\nnetwork link prediction. One might initially think the inclusion of any\nadditional data is useful, and hence that it should be beneficial to include\nall fringe nodes that are available. However, we find that this is not true; in\nfact, there is substantial variability in the value of the fringe nodes for\nprediction. Once an algorithm is selected, in some datasets, including any\nadditional data from the fringe can actually hurt prediction performance; in\nother datasets, including some amount of fringe information is useful before\nprediction performance saturates or even declines; and in further cases,\nincluding the entire fringe leads to the best performance. While such variety\nmight seem surprising, we show that these behaviors are exhibited by simple\nrandom graph models."", 'Graphs are widely used as a popular representation of the network structure\nof connected data. Graph data can be found in a broad spectrum of application\ndomains such as social systems, ecosystems, biological networks, knowledge\ngraphs, and information systems. With the continuous penetration of artificial\nintelligence technologies, graph learning (i.e., machine learning on graphs) is\ngaining attention from both researchers and practitioners. Graph learning\nproves effective for many tasks, such as classification, link prediction, and\nmatching. Generally, graph learning methods extract relevant features of graphs\nby taking advantage of machine learning algorithms. In this survey, we present\na comprehensive overview on the state-of-the-art of graph learning. Special\nattention is paid to four categories of existing graph learning methods,\nincluding graph signal processing, matrix factorization, random walk, and deep\nlearning. Major models and algorithms under these categories are reviewed\nrespectively. We examine graph learning applications in areas such as text,\nimages, science, knowledge graphs, and combinatorial optimization. In addition,\nwe discuss several promising research directions in this field.', 'The success of deep convolutional neural network (CNN) in computer vision\nespecially image classification problems requests a new information theory for\nfunction of image, instead of image itself. In this article, after establishing\na deep mathematical connection between image classification problem and quantum\nspin model, we propose to use entanglement entropy, a generalization of\nclassical Boltzmann-Shannon entropy, as a powerful tool to characterize the\ninformation needed for representation of general function of image. We prove\nthat there is a sub-volume-law bound for entanglement entropy of target\nfunctions of reasonable image classification problems. Therefore target\nfunctions of image classification only occupy a small subspace of the whole\nHilbert space. As a result, a neural network with polynomial number of\nparameters is efficient for representation of such target functions of image.\nThe concept of entanglement entropy can also be useful to characterize the\nexpressive power of different neural networks. For example, we show that to\nmaintain the same expressive power, number of channels $D$ in a convolutional\nneural network should scale with the number of convolution layers $n_c$ as\n$D\\sim D_0^{\\frac{1}{n_c}}$. Therefore, deeper CNN with large $n_c$ is more\nefficient than shallow ones.', 'We consider the data shuffling problem in a distributed learning system, in\nwhich a master node is connected to a set of worker nodes, via a shared link,\nin order to communicate a set of files to the worker nodes. The master node has\naccess to a database of files. In every shuffling iteration, each worker node\nprocesses a new subset of files, and has excess storage to partially cache the\nremaining files, assuming the cached files are uncoded. The caches of the\nworker nodes are updated every iteration, and they should be designed to\nsatisfy any possible unknown permutation of the files in subsequent iterations.\nFor this problem, we characterize the exact load-memory trade-off for\nworst-case shuffling by deriving the minimum communication load for a given\nstorage capacity per worker node. As a byproduct, the exact load-memory\ntrade-off for any shuffling is characterized when the number of files is equal\nto the number of worker nodes. We propose a novel deterministic coded shuffling\nscheme, which improves the state of the art, by exploiting the cache memories\nto create coded functions that can be decoded by several worker nodes. Then, we\nprove the optimality of our proposed scheme by deriving a matching lower bound\nand showing that the placement phase of the proposed coded shuffling scheme is\noptimal over all shuffles.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am seeking to comprehend the performance of the MAML algorithm in meta-learning when confronted with adversarial attacks aimed at contaminating its learned parameters. Initially, I require a method to swiftly generate examples of adversarial attacks. Subsequently, I am interested in investigating defense strategies that enable MAML to either identify adversarial examples or prevent learning from them. My current approach involves implementing some form of regularization to ensure that MAML is not disproportionately influenced by a handful of malicious data examples. Given that MAML can perform parameter updates in either its inner or outer loop, it is crucial for me to determine the optimal location to deploy these defense strategies, such as regularization. Ultimately, I aim to develop an end-to-end framework capable of automatically assessing the stability of MAML and other meta-learning algorithms when subjected to adversarial attacks.","[""As we seek to deploy machine learning models beyond virtual and controlled\ndomains, it is critical to analyze not only the accuracy or the fact that it\nworks most of the time, but if such a model is truly robust and reliable. This\npaper studies strategies to implement adversary robustly trained algorithms\ntowards guaranteeing safety in machine learning algorithms. We provide a\ntaxonomy to classify adversarial attacks and defenses, formulate the Robust\nOptimization problem in a min-max setting and divide it into 3 subcategories,\nnamely: Adversarial (re)Training, Regularization Approach, and Certified\nDefenses. We survey the most recent and important results in adversarial\nexample generation, defense mechanisms with adversarial (re)Training as their\nmain defense against perturbations. We also survey mothods that add\nregularization terms that change the behavior of the gradient, making it harder\nfor attackers to achieve their objective. Alternatively, we've surveyed methods\nwhich formally derive certificates of robustness by exactly solving the\noptimization problem or by approximations using upper or lower bounds. In\naddition, we discuss the challenges faced by most of the recent algorithms\npresenting future research perspectives."", 'Background: Machine learning-based security detection models have become\nprevalent in modern malware and intrusion detection systems. However, previous\nstudies show that such models are susceptible to adversarial evasion attacks.\nIn this type of attack, inputs (i.e., adversarial examples) are specially\ncrafted by intelligent malicious adversaries, with the aim of being\nmisclassified by existing state-of-the-art models (e.g., deep neural networks).\nOnce the attackers can fool a classifier to think that a malicious input is\nactually benign, they can render a machine learning-based malware or intrusion\ndetection system ineffective. Goal: To help security practitioners and\nresearchers build a more robust model against non-adaptive, white-box, and\nnon-targeted adversarial evasion attacks through the idea of an ensemble model.\nMethod: We propose an approach called Omni, the main idea of which is to\nexplore methods that create an ensemble of ""unexpected models""; i.e., models\nwhose control hyperparameters have a large distance to the hyperparameters of\nan adversary\'s target model, with which we then make an optimized weighted\nensemble prediction. Result: In studies with five types of adversarial evasion\nattacks (FGSM, BIM, JSMA, DeepFooland Carlini-Wagner) on five security datasets\n(NSL-KDD, CIC-IDS-2017, CSE-CIC-IDS2018, CICAnd-Mal2017, and the Contagio PDF\ndataset), we show Omni is a promising approach as a defense strategy against\nadversarial attacks when compared with other baseline treatments. Conclusion:\nWhen employing ensemble defense against adversarial evasion attacks, we suggest\ncreating an ensemble with unexpected models that are distant from the\nattacker\'s expected model (i.e., target model) through methods such as\nhyperparameter optimization.', 'Deep neural networks (DNNs) are known to be prone to adversarial attacks, for\nwhich many remedies are proposed. While adversarial training (AT) is regarded\nas the most robust defense, it suffers from poor performance both on clean\nexamples and under other types of attacks, e.g. attacks with larger\nperturbations. Meanwhile, regularizers that encourage uncertain outputs, such\nas entropy maximization (EntM) and label smoothing (LS) can maintain accuracy\non clean examples and improve performance under weak attacks, yet their ability\nto defend against strong attacks is still in doubt. In this paper, we revisit\nuncertainty promotion regularizers, including EntM and LS, in the field of\nadversarial learning. We show that EntM and LS alone provide robustness only\nunder small perturbations. Contrarily, we show that uncertainty promotion\nregularizers complement AT in a principled manner, consistently improving\nperformance on both clean examples and under various attacks, especially\nattacks with large perturbations. We further analyze how uncertainty promotion\nregularizers enhance the performance of AT from the perspective of Jacobian\nmatrices $\\nabla_X f(X;\\theta)$, and find out that EntM effectively shrinks the\nnorm of Jacobian matrices and hence promotes robustness.', ""The security of Deep Reinforcement Learning (Deep RL) algorithms deployed in\nreal life applications are of a primary concern. In particular, the robustness\nof RL agents in cyber-physical systems against adversarial attacks are\nespecially vital since the cost of a malevolent intrusions can be extremely\nhigh. Studies have shown Deep Neural Networks (DNN), which forms the core\ndecision-making unit in most modern RL algorithms, are easily subjected to\nadversarial attacks. Hence, it is imperative that RL agents deployed in\nreal-life applications have the capability to detect and mitigate adversarial\nattacks in an online fashion. An example of such a framework is the\nMeta-Learned Advantage Hierarchy (MLAH) agent that utilizes a meta-learning\nframework to learn policies robustly online. Since the mechanism of this\nframework are still not fully explored, we conducted multiple experiments to\nbetter understand the framework's capabilities and limitations. Our results\nshows that the MLAH agent exhibits interesting coping behaviors when subjected\nto different adversarial attacks to maintain a nominal reward. Additionally,\nthe framework exhibits a hierarchical coping capability, based on the\nadaptability of the Master policy and sub-policies themselves. From empirical\nresults, we also observed that as the interval of adversarial attacks increase,\nthe MLAH agent can maintain a higher distribution of rewards, though at the\ncost of higher instabilities."", 'In this paper, we study the problem of how to defend classifiers against\nadversarial attacks that fool the classifiers using subtly modified input data.\nIn contrast to previous works, here we focus on the white-box adversarial\ndefense where the attackers are granted full access to not only the classifiers\nbut also defenders to produce as strong attacks as possible. In such a context\nwe propose viewing a defender as a functional, a higher-order function that\ntakes functions as its argument to represent a function space, rather than\nfixed functions conventionally. From this perspective, a defender should be\nrealized and optimized individually for each adversarial input. To this end, we\npropose RIDE, an efficient and provably convergent self-supervised learning\nalgorithm for individual data estimation to protect the predictions from\nadversarial attacks. We demonstrate the significant improvement of adversarial\ndefense performance on image recognition, eg, 98%, 76%, 43% test accuracy on\nMNIST, CIFAR-10, and ImageNet datasets respectively under the state-of-the-art\nBPDA attacker.', 'Recent works have shown that the input domain of any machine learning\nclassifier is bound to contain adversarial examples. Thus we can no longer hope\nto immune classifiers against adversarial examples and instead can only aim to\nachieve the following two defense goals: 1) making adversarial examples harder\nto find, or 2) weakening their adversarial nature by pushing them further away\nfrom correctly classified data points. Most if not all the previously suggested\ndefense mechanisms attend to just one of those two goals, and as such, could be\nbypassed by adaptive attacks that take the defense mechanism into\nconsideration. In this work we suggest a novel defense mechanism that\nsimultaneously addresses both defense goals: We flatten the gradients of the\nloss surface, making adversarial examples harder to find, using a novel\nstochastic regularization term that explicitly decreases the sensitivity of\nindividual neurons to small input perturbations. In addition, we push the\ndecision boundary away from correctly classified inputs by leveraging Jacobian\nregularization. We present a solid theoretical basis and an empirical testing\nof our suggested approach, demonstrate its superiority over previously\nsuggested defense mechanisms, and show that it is effective against a wide\nrange of adaptive attacks.', 'While great progress has been made at making neural networks effective across\na wide range of visual tasks, most models are surprisingly vulnerable. This\nfrailness takes the form of small, carefully chosen perturbations of their\ninput, known as adversarial examples, which represent a security threat for\nlearned vision models in the wild -- a threat which should be responsibly\ndefended against in safety-critical applications of computer vision. In this\npaper, we advocate for and experimentally investigate the use of a family of\nlogit regularization techniques as an adversarial defense, which can be used in\nconjunction with other methods for creating adversarial robustness at little to\nno marginal cost. We also demonstrate that much of the effectiveness of one\nrecent adversarial defense mechanism can in fact be attributed to logit\nregularization, and show how to improve its defense against both white-box and\nblack-box attacks, in the process creating a stronger black-box attack against\nPGD-based models. We validate our methods on three datasets and include results\non both gradient-free attacks and strong gradient-based iterative attacks with\nas many as 1,000 steps.', ""The arms race between attacks and defenses for machine learning models has\ncome to a forefront in recent years, in both the security community and the\nprivacy community. However, one big limitation of previous research is that the\nsecurity domain and the privacy domain have typically been considered\nseparately. It is thus unclear whether the defense methods in one domain will\nhave any unexpected impact on the other domain.\n  In this paper, we take a step towards resolving this limitation by combining\nthe two domains. In particular, we measure the success of membership inference\nattacks against six state-of-the-art defense methods that mitigate the risk of\nadversarial examples (i.e., evasion attacks). Membership inference attacks\ndetermine whether or not an individual data record has been part of a model's\ntraining set. The accuracy of such attacks reflects the information leakage of\ntraining algorithms about individual members of the training set. Adversarial\ndefense methods against adversarial examples influence the model's decision\nboundaries such that model predictions remain unchanged for a small area around\neach input. However, this objective is optimized on training data. Thus,\nindividual data records in the training set have a significant influence on\nrobust models. This makes the models more vulnerable to inference attacks.\n  To perform the membership inference attacks, we leverage the existing\ninference methods that exploit model predictions. We also propose two new\ninference methods that exploit structural properties of robust models on\nadversarially perturbed data. Our experimental evaluation demonstrates that\ncompared with the natural training (undefended) approach, adversarial defense\nmethods can indeed increase the target model's risk against membership\ninference attacks."", 'Adversarial attacks on deep learning models have compromised their\nperformance considerably. As remedies, a lot of defense methods were proposed,\nwhich however, have been circumvented by newer attacking strategies. In the\nmidst of this ensuing arms race, the problem of robustness against adversarial\nattacks still remains unsolved. This paper proposes a novel, simple yet\neffective defense strategy where adversarial samples are relaxed onto the\nunderlying manifold of the (unknown) target class distribution. Specifically,\nour algorithm drives off-manifold adversarial samples towards high density\nregions of the data generating distribution of the target class by the\nMetroplis-adjusted Langevin algorithm (MALA) with perceptual boundary taken\ninto account. Although the motivation is similar to projection methods, e.g.,\nDefense-GAN, our algorithm, called MALA for DEfense (MALADE), is equipped with\nsignificant dispersion - projection is distributed broadly, and therefore any\nwhitebox attack cannot accurately align the input so that the MALADE moves it\nto a targeted untrained spot where the model predicts a wrong label. In our\nexperiments, MALADE exhibited state-of-the-art performance against various\nelaborate attacking strategies.', 'The growing prospect of deep reinforcement learning (DRL) being used in\ncyber-physical systems has raised concerns around safety and robustness of\nautonomous agents. Recent work on generating adversarial attacks have shown\nthat it is computationally feasible for a bad actor to fool a DRL policy into\nbehaving sub optimally. Although certain adversarial attacks with specific\nattack models have been addressed, most studies are only interested in off-line\noptimization in the data space (e.g., example fitting, distillation). This\npaper introduces a Meta-Learned Advantage Hierarchy (MLAH) framework that is\nattack model-agnostic and more suited to reinforcement learning, via handling\nthe attacks in the decision space (as opposed to data space) and directly\nmitigating learned bias introduced by the adversary. In MLAH, we learn separate\nsub-policies (nominal and adversarial) in an online manner, as guided by a\nsupervisory master agent that detects the presence of the adversary by\nleveraging the advantage function for the sub-policies. We demonstrate that the\nproposed algorithm enables policy learning with significantly lower bias as\ncompared to the state-of-the-art policy learning approaches even in the\npresence of heavy state information attacks. We present algorithm analysis and\nsimulation results using popular OpenAI Gym environments.', 'Adversarial examples are augmented data points generated by imperceptible\nperturbation of input samples. They have recently drawn much attention with the\nmachine learning and data mining community. Being difficult to distinguish from\nreal examples, such adversarial examples could change the prediction of many of\nthe best learning models including the state-of-the-art deep learning models.\nRecent attempts have been made to build robust models that take into account\nadversarial examples. However, these methods can either lead to performance\ndrops or lack mathematical motivations. In this paper, we propose a unified\nframework to build robust machine learning models against adversarial examples.\nMore specifically, using the unified framework, we develop a family of gradient\nregularization methods that effectively penalize the gradient of loss function\nw.r.t. inputs. Our proposed framework is appealing in that it offers a unified\nview to deal with adversarial examples. It incorporates another\nrecently-proposed perturbation based approach as a special case. In addition,\nwe present some visual effects that reveals semantic meaning in those\nperturbations, and thus support our regularization method and provide another\nexplanation for generalizability of adversarial examples. By applying this\ntechnique to Maxout networks, we conduct a series of experiments and achieve\nencouraging results on two benchmark datasets. In particular,we attain the best\naccuracy on MNIST data (without data augmentation) and competitive performance\non CIFAR-10 data.', 'In this paper, we develop improved techniques for defending against\nadversarial examples at scale. First, we implement the state of the art version\nof adversarial training at unprecedented scale on ImageNet and investigate\nwhether it remains effective in this setting - an important open scientific\nquestion (Athalye et al., 2018). Next, we introduce enhanced defenses using a\ntechnique we call logit pairing, a method that encourages logits for pairs of\nexamples to be similar. When applied to clean examples and their adversarial\ncounterparts, logit pairing improves accuracy on adversarial examples over\nvanilla adversarial training; we also find that logit pairing on clean examples\nonly is competitive with adversarial training in terms of accuracy on two\ndatasets. Finally, we show that adversarial logit pairing achieves the state of\nthe art defense on ImageNet against PGD white box attacks, with an accuracy\nimprovement from 1.5% to 27.9%. Adversarial logit pairing also successfully\ndamages the current state of the art defense against black box attacks on\nImageNet (Tramer et al., 2018), dropping its accuracy from 66.6% to 47.1%. With\nthis new accuracy drop, adversarial logit pairing ties with Tramer et al.(2018)\nfor the state of the art on black box attacks on ImageNet.', 'Though deep neural networks have achieved the state of the art performance in\nvisual classification, recent studies have shown that they are all vulnerable\nto the attack of adversarial examples. In this paper, we develop improved\ntechniques for defending against adversarial examples. First, we propose an\nenhanced defense technique denoted Attention and Adversarial Logit\nPairing(AT+ALP), which encourages both attention map and logit for the pairs of\nexamples to be similar. When being applied to clean examples and their\nadversarial counterparts, AT+ALP improves accuracy on adversarial examples over\nadversarial training. We show that AT+ALP can effectively increase the average\nactivations of adversarial examples in the key area and demonstrate that it\nfocuses on discriminate features to improve the robustness of the model.\nFinally, we conduct extensive experiments using a wide range of datasets and\nthe experiment results show that our AT+ALP achieves the state of the art\ndefense performance. For example, on 17 Flower Category Database, under strong\n200-iteration PGD gray-box and black-box attacks where prior art has 34% and\n39% accuracy, our method achieves 50% and 51%. Compared with previous work, our\nwork is evaluated under highly challenging PGD attack: the maximum perturbation\n$\\epsilon \\in \\{0.25,0.5\\}$ i.e. $L_\\infty \\in \\{0.25,0.5\\}$ with 10 to 200\nattack iterations. To the best of our knowledge, such a strong attack has not\nbeen previously explored on a wide range of datasets.', 'Deep neural networks are vulnerable to adversarial attacks. Recent studies\nabout adversarial robustness focus on the loss landscape in the parameter space\nsince it is related to optimization and generalization performance. These\nstudies conclude that the difficulty of adversarial training is caused by the\nnon-smoothness of the loss function: i.e., its gradient is not Lipschitz\ncontinuous. However, this analysis ignores the dependence of adversarial\nattacks on model parameters. Since adversarial attacks are optimized for\nmodels, they should depend on the parameters. Considering this dependence, we\nanalyze the smoothness of the loss function of adversarial training using the\noptimal attacks for the model parameter in more detail. We reveal that the\nconstraint of adversarial attacks is one cause of the non-smoothness and that\nthe smoothness depends on the types of the constraints. Specifically, the\n$L_\\infty$ constraint can cause non-smoothness more than the $L_2$ constraint.\nMoreover, our analysis implies that if we flatten the loss function with\nrespect to input data, the Lipschitz constant of the gradient of adversarial\nloss tends to increase. To address the non-smoothness, we show that EntropySGD\nsmoothens the non-smooth loss and improves the performance of adversarial\ntraining.', 'Various forefront countermeasure methods for automatic speaker verification\n(ASV) with considerable performance in anti-spoofing are proposed in the\nASVspoof 2019 challenge. However, previous work has shown that countermeasure\nmodels are vulnerable to adversarial examples indistinguishable from natural\ndata. A good countermeasure model should not only be robust against spoofing\naudio, including synthetic, converted, and replayed audios; but counteract\ndeliberately generated examples by malicious adversaries. In this work, we\nintroduce a passive defense method, spatial smoothing, and a proactive defense\nmethod, adversarial training, to mitigate the vulnerability of ASV spoofing\ncountermeasure models against adversarial examples. This paper is among the\nfirst to use defense methods to improve the robustness of ASV spoofing\ncountermeasure models under adversarial attacks. The experimental results show\nthat these two defense methods positively help spoofing countermeasure models\ncounter adversarial examples.', 'Meta-learning model can quickly adapt to new tasks using few-shot labeled\ndata. However, despite achieving good generalization on few-shot classification\ntasks, it is still challenging to improve the adversarial robustness of the\nmeta-learning model in few-shot learning. Although adversarial training (AT)\nmethods such as Adversarial Query (AQ) can improve the adversarially robust\nperformance of meta-learning models, AT is still computationally expensive\ntraining. On the other hand, meta-learning models trained with AT will drop\nsignificant accuracy on the original clean images. This paper proposed a\nmeta-learning method on the adversarially robust neural network called\nLong-term Cross Adversarial Training (LCAT). LCAT will update meta-learning\nmodel parameters cross along the natural and adversarial sample distribution\ndirection with long-term to improve both adversarial and clean few-shot\nclassification accuracy. Due to cross-adversarial training, LCAT only needs\nhalf of the adversarial training epoch than AQ, resulting in a low adversarial\ntraining computation. Experiment results show that LCAT achieves superior\nperformance both on the clean and adversarial few-shot classification accuracy\nthan SOTA adversarial training methods for meta-learning models.', 'Meta-learning enables a model to learn from very limited data to undertake a\nnew task. In this paper, we study the general meta-learning with adversarial\nsamples. We present a meta-learning algorithm, ADML (ADversarial Meta-Learner),\nwhich leverages clean and adversarial samples to optimize the initialization of\na learning model in an adversarial manner. ADML leads to the following\ndesirable properties: 1) it turns out to be very effective even in the cases\nwith only clean samples; 2) it is robust to adversarial samples, i.e., unlike\nother meta-learning algorithms, it only leads to a minor performance\ndegradation when there are adversarial samples; 3) it sheds light on tackling\nthe cases with limited and even contaminated samples. It has been shown by\nextensive experimental results that ADML consistently outperforms three\nrepresentative meta-learning algorithms in the cases involving adversarial\nsamples, on two widely-used image datasets, MiniImageNet and CIFAR100, in terms\nof both accuracy and robustness.', 'This paper proposes a classification framework with a rejection option to\nmitigate the performance deterioration caused by adversarial examples. While\nrecent machine learning algorithms achieve high prediction performance, they\nare empirically vulnerable to adversarial examples, which are slightly\nperturbed data samples that are wrongly classified. In real-world applications,\nadversarial attacks using such adversarial examples could cause serious\nproblems. To this end, various methods are proposed to obtain a classifier that\nis robust against adversarial examples. Adversarial training is one of them,\nwhich trains a classifier to minimize the worst-case loss under adversarial\nattacks. In this paper, in order to acquire a more reliable classifier against\nadversarial attacks, we propose the method of Adversarial Training with a\nRejection Option (ATRO). Applying the adversarial training objective to both a\nclassifier and a rejection function simultaneously, classifiers trained by ATRO\ncan choose to abstain from classification when it has insufficient confidence\nto classify a test data point. We examine the feasibility of the framework\nusing the surrogate maximum hinge loss and establish a generalization bound for\nlinear models. Furthermore, we empirically confirmed the effectiveness of ATRO\nusing various models and real-world datasets.', 'Ideally, what confuses neural network should be confusing to humans. However,\nrecent experiments have shown that small, imperceptible perturbations can\nchange the network prediction. To address this gap in perception, we propose a\nnovel approach for learning robust classifier. Our main idea is: adversarial\nexamples for the robust classifier should be indistinguishable from the regular\ndata of the adversarial target. We formulate a problem of learning robust\nclassifier in the framework of Generative Adversarial Networks (GAN), where the\nadversarial attack on classifier acts as a generator, and the critic network\nlearns to distinguish between regular and adversarial images. The classifier\ncost is augmented with the objective that its adversarial examples should\nconfuse the adversary critic. To improve the stability of the adversarial\nmapping, we introduce adversarial cycle-consistency constraint which ensures\nthat the adversarial mapping of the adversarial examples is close to the\noriginal. In the experiments, we show the effectiveness of our defense. Our\nmethod surpasses in terms of robustness networks trained with adversarial\ntraining. Additionally, we verify in the experiments with human annotators on\nMTurk that adversarial examples are indeed visually confusing. Codes for the\nproject are available at https://github.com/aam-at/adversary_critic.', ""Machine Learning (ML) algorithms are vulnerable to poisoning attacks, where a\nfraction of the training data is manipulated to deliberately degrade the\nalgorithms' performance. Optimal poisoning attacks, which can be formulated as\nbilevel optimisation problems, help to assess the robustness of learning\nalgorithms in worst-case scenarios. However, current attacks against algorithms\nwith hyperparameters typically assume that these hyperparameters remain\nconstant ignoring the effect the attack has on them. We show that this approach\nleads to an overly pessimistic view of the robustness of the algorithms. We\npropose a novel optimal attack formulation that considers the effect of the\nattack on the hyperparameters by modelling the attack as a multiobjective\nbilevel optimisation problem. We apply this novel attack formulation to ML\nclassifiers using $L_2$ regularisation and show that, in contrast to results\npreviously reported, $L_2$ regularisation enhances the stability of the\nlearning algorithms and helps to mitigate the attacks. Our empirical evaluation\non different datasets confirms the limitations of previous strategies,\nevidences the benefits of using $L_2$ regularisation to dampen the effect of\npoisoning attacks and shows how the regularisation hyperparameter increases\nwith the fraction of poisoning points."", ""Convolutional Neural Networks and Deep Learning classification systems in\ngeneral have been shown to be vulnerable to attack by specially crafted data\nsamples that appear to belong to one class but are instead classified as\nanother, commonly known as adversarial examples. A variety of attack strategies\nhave been proposed to craft these samples; however, there is no standard model\nthat is used to compare the success of each type of attack. Furthermore, there\nis no literature currently available that evaluates how common hyperparameters\nand optimization strategies may impact a model's ability to resist these\nsamples. This research bridges that lack of awareness and provides a means for\nthe selection of training and model parameters in future research on evasion\nattacks against convolutional neural networks. The findings of this work\nindicate that the selection of model hyperparameters does impact the ability of\na model to resist attack, although they alone cannot prevent the existence of\nadversarial examples."", 'A range of defense methods have been proposed to improve the robustness of\nneural networks on adversarial examples, among which provable defense methods\nhave been demonstrated to be effective to train neural networks that are\ncertifiably robust to the attacker. However, most of these provable defense\nmethods treat all examples equally during training process, which ignore the\ninconsistent constraint of certified robustness between correctly classified\n(natural) and misclassified examples. In this paper, we explore this\ninconsistency caused by misclassified examples and add a novel consistency\nregularization term to make better use of the misclassified examples.\nSpecifically, we identified that the certified robustness of network can be\nsignificantly improved if the constraint of certified robustness on\nmisclassified examples and correctly classified examples is consistent.\nMotivated by this discovery, we design a new defense regularization term called\nMisclassification Aware Adversarial Regularization (MAAR), which constrains the\noutput probability distributions of all examples in the certified region of the\nmisclassified example. Experimental results show that our proposed MAAR\nachieves the best certified robustness and comparable accuracy on CIFAR-10 and\nMNIST datasets in comparison with several state-of-the-art methods.']","['Machine learning classifiers are known to be vulnerable to inputs maliciously\nconstructed by adversaries to force misclassification. Such adversarial\nexamples have been extensively studied in the context of computer vision\napplications. In this work, we show adversarial attacks are also effective when\ntargeting neural network policies in reinforcement learning. Specifically, we\nshow existing adversarial example crafting techniques can be used to\nsignificantly degrade test-time performance of trained policies. Our threat\nmodel considers adversaries capable of introducing small perturbations to the\nraw input of the policy. We characterize the degree of vulnerability across\ntasks and training algorithms, for a subclass of adversarial-example attacks in\nwhite-box and black-box settings. Regardless of the learned task or training\nalgorithm, we observe a significant drop in performance, even with small\nadversarial perturbations that do not interfere with human perception. Videos\nare available at http://rll.berkeley.edu/adversarial.', 'High-performance anti-spoofing models for automatic speaker verification\n(ASV), have been widely used to protect ASV by identifying and filtering\nspoofing audio that is deliberately generated by text-to-speech, voice\nconversion, audio replay, etc. However, it has been shown that high-performance\nanti-spoofing models are vulnerable to adversarial attacks. Adversarial\nattacks, that are indistinguishable from original data but result in the\nincorrect predictions, are dangerous for anti-spoofing models and not in\ndispute we should detect them at any cost. To explore this issue, we proposed\nto employ Mockingjay, a self-supervised learning based model, to protect\nanti-spoofing models against adversarial attacks in the black-box scenario.\nSelf-supervised learning models are effective in improving downstream task\nperformance like phone classification or ASR. However, their effect in defense\nfor adversarial attacks has not been explored yet. In this work, we explore the\nrobustness of self-supervised learned high-level representations by using them\nin the defense against adversarial attacks. A layerwise noise to signal ratio\n(LNSR) is proposed to quantize and measure the effectiveness of deep models in\ncountering adversarial noise. Experimental results on the ASVspoof 2019 dataset\ndemonstrate that high-level representations extracted by Mockingjay can prevent\nthe transferability of adversarial examples, and successfully counter black-box\nattacks.', 'Recent analysis of deep neural networks has revealed their vulnerability to\ncarefully structured adversarial examples. Many effective algorithms exist to\ncraft these adversarial examples, but performant defenses seem to be far away.\nIn this work, we explore the use of edge-aware bilateral filtering as a\nprojection back to the space of natural images. We show that bilateral\nfiltering is an effective defense in multiple attack settings, where the\nstrength of the adversary gradually increases. In the case of an adversary who\nhas no knowledge of the defense, bilateral filtering can remove more than 90%\nof adversarial examples from a variety of different attacks. To evaluate\nagainst an adversary with complete knowledge of our defense, we adapt the\nbilateral filter as a trainable layer in a neural network and show that adding\nthis layer makes ImageNet images significantly more robust to attacks. When\ntrained under a framework of adversarial training, we show that the resulting\nmodel is hard to fool with even the best attack methods.', 'We explore adversarial robustness in the setting in which it is acceptable\nfor a classifier to abstain---that is, output no class---on adversarial\nexamples. Adversarial examples are small perturbations of normal inputs to a\nclassifier that cause the classifier to give incorrect output; they present\nsecurity and safety challenges for machine learning systems. In many\nsafety-critical applications, it is less costly for a classifier to abstain on\nadversarial examples than to give incorrect output for them. We first introduce\na novel objective function for adversarial robustness with an abstain option\nwhich characterizes an explicit tradeoff between robustness and accuracy. We\nthen present a simple baseline in which an adversarially-trained classifier\nabstains on all inputs within a certain distance of the decision boundary,\nwhich we theoretically and experimentally evaluate. Finally, we propose\nCombined Abstention Robustness Learning (CARL), a method for jointly learning a\nclassifier and the region of the input space on which it should abstain. We\nexplore different variations of the PGD and DeepFool adversarial attacks on\nCARL in the abstain setting. Evaluating against these attacks, we demonstrate\nthat training with CARL results in a more accurate, robust, and efficient\nclassifier than the baseline.', 'Recently, the field of adversarial machine learning has been garnering\nattention by showing that state-of-the-art deep neural networks are vulnerable\nto adversarial examples, stemming from small perturbations being added to the\ninput image. Adversarial examples are generated by a malicious adversary by\nobtaining access to the model parameters, such as gradient information, to\nalter the input or by attacking a substitute model and transferring those\nmalicious examples over to attack the victim model. Specifically, one of these\nattack algorithms, Robust Physical Perturbations ($RP_2$), generates\nadversarial images of stop signs with black and white stickers to achieve high\ntargeted misclassification rates against standard-architecture traffic sign\nclassifiers. In this paper, we propose BlurNet, a defense against the $RP_2$\nattack. First, we motivate the defense with a frequency analysis of the first\nlayer feature maps of the network on the LISA dataset, which shows that high\nfrequency noise is introduced into the input image by the $RP_2$ algorithm. To\nremove the high frequency noise, we introduce a depthwise convolution layer of\nstandard blur kernels after the first layer. We perform a blackbox transfer\nattack to show that low-pass filtering the feature maps is more beneficial than\nfiltering the input. We then present various regularization schemes to\nincorporate this low-pass filtering behavior into the training regime of the\nnetwork and perform white-box attacks. We conclude with an adaptive attack\nevaluation to show that the success rate of the attack drops from 90\\% to 20\\%\nwith total variation regularization, one of the proposed defenses.', 'Deep Neural Networks (DNNs) have recently achieved great success in many\ntasks, which encourages DNNs to be widely used as a machine learning service in\nmodel sharing scenarios. However, attackers can easily generate adversarial\nexamples with a small perturbation to fool the DNN models to predict wrong\nlabels. To improve the robustness of shared DNN models against adversarial\nattacks, we propose a novel method called Latent Adversarial Defence (LAD). The\nproposed LAD method improves the robustness of a DNN model through adversarial\ntraining on generated adversarial examples. Different from popular attack\nmethods which are carried in the input space and only generate adversarial\nexamples of repeating patterns, LAD generates myriad of adversarial examples\nthrough adding perturbations to latent features along the normal of the\ndecision boundary which is constructed by an SVM with an attention mechanism.\nOnce adversarial examples are generated, we adversarially train the model\nthrough augmenting the training data with generated adversarial examples.\nExtensive experiments on the MNIST, SVHN, and CelebA dataset demonstrate the\neffectiveness of our model in defending against different types of adversarial\nattacks.', 'In image classification of deep learning, adversarial examples where inputs\nintended to add small magnitude perturbations may mislead deep neural networks\n(DNNs) to incorrect results, which means DNNs are vulnerable to them. Different\nattack and defense strategies have been proposed to better research the\nmechanism of deep learning. However, those research in these networks are only\nfor one aspect, either an attack or a defense, not considering that attacks and\ndefenses should be interdependent and mutually reinforcing, just like the\nrelationship between spears and shields. In this paper, we propose\nCycle-Consistent Adversarial GAN (CycleAdvGAN) to generate adversarial\nexamples, which can learn and approximate the distribution of original\ninstances and adversarial examples. For CycleAdvGAN, once the Generator and are\ntrained, can generate adversarial perturbations efficiently for any instance,\nso as to make DNNs predict wrong, and recovery adversarial examples to clean\ninstances, so as to make DNNs predict correct. We apply CycleAdvGAN under\nsemi-white box and black-box settings on two public datasets MNIST and CIFAR10.\nUsing the extensive experiments, we show that our method has achieved the\nstate-of-the-art adversarial attack method and also efficiently improve the\ndefense ability, which make the integration of adversarial attack and defense\ncome true. In additional, it has improved attack effect only trained on the\nadversarial dataset generated by any kind of adversarial attack.', 'Learning-based pattern classifiers, including deep networks, have shown\nimpressive performance in several application domains, ranging from computer\nvision to cybersecurity. However, it has also been shown that adversarial input\nperturbations carefully crafted either at training or at test time can easily\nsubvert their predictions. The vulnerability of machine learning to such wild\npatterns (also referred to as adversarial examples), along with the design of\nsuitable countermeasures, have been investigated in the research field of\nadversarial machine learning. In this work, we provide a thorough overview of\nthe evolution of this research area over the last ten years and beyond,\nstarting from pioneering, earlier work on the security of non-deep learning\nalgorithms up to more recent work aimed to understand the security properties\nof deep learning algorithms, in the context of computer vision and\ncybersecurity tasks. We report interesting connections between these\napparently-different lines of work, highlighting common misconceptions related\nto the security evaluation of machine-learning algorithms. We review the main\nthreat models and attacks defined to this end, and discuss the main limitations\nof current work, along with the corresponding future challenges towards the\ndesign of more secure learning algorithms.', 'Meta-learning leverages related source tasks to learn an initialization that\ncan be quickly fine-tuned to a target task with limited labeled examples.\nHowever, many popular meta-learning algorithms, such as model-agnostic\nmeta-learning (MAML), only assume access to the target samples for fine-tuning.\nIn this work, we provide a general framework for meta-learning based on\nweighting the loss of different source tasks, where the weights are allowed to\ndepend on the target samples. In this general setting, we provide upper bounds\non the distance of the weighted empirical risk of the source tasks and expected\ntarget risk in terms of an integral probability metric (IPM) and Rademacher\ncomplexity, which apply to a number of meta-learning settings including MAML\nand a weighted MAML variant. We then develop a learning algorithm based on\nminimizing the error bound with respect to an empirical IPM, including a\nweighted MAML algorithm, $\\alpha$-MAML. Finally, we demonstrate empirically on\nseveral regression problems that our weighted meta-learning algorithm is able\nto find better initializations than uniformly-weighted meta-learning\nalgorithms, such as MAML.', ""Adversarial examples are maliciously tweaked images that can easily fool\nmachine learning techniques, such as neural networks, but they are normally not\nvisually distinguishable for human beings. One of the main approaches to solve\nthis problem is to retrain the networks using those adversarial examples,\nnamely adversarial training. However, standard adversarial training might not\nactually change the decision boundaries but cause the problem of gradient\nmasking, resulting in a weaker ability to generate adversarial examples.\nTherefore, it cannot alleviate the problem of black-box attacks, where\nadversarial examples generated from other networks can transfer to the targeted\none. In order to reduce the problem of black-box attacks, we propose a novel\nmethod that allows two networks to learn from each others' adversarial examples\nand become resilient to black-box attacks. We also combine this method with a\nsimple domain adaptation to further improve the performance."", 'Adversarial training (AT) has been demonstrated as one of the most promising\ndefense methods against various adversarial attacks. To our knowledge, existing\nAT-based methods usually train with the locally most adversarial perturbed\npoints and treat all the perturbed points equally, which may lead to\nconsiderably weaker adversarial robust generalization on test data. In this\nwork, we introduce a new adversarial training framework that considers the\ndiversity as well as characteristics of the perturbed points in the vicinity of\nbenign samples. To realize the framework, we propose a Regional Adversarial\nTraining (RAT) defense method that first utilizes the attack path generated by\nthe typical iterative attack method of projected gradient descent (PGD), and\nconstructs an adversarial region based on the attack path. Then, RAT samples\ndiverse perturbed training points efficiently inside this region, and utilizes\na distance-aware label smoothing mechanism to capture our intuition that\nperturbed points at different locations should have different impact on the\nmodel performance. Extensive experiments on several benchmark datasets show\nthat RAT consistently makes significant improvement on standard adversarial\ntraining (SAT), and exhibits better robust generalization.', 'Machine learning (ML) models, e.g., deep neural networks (DNNs), are\nvulnerable to adversarial examples: malicious inputs modified to yield\nerroneous model outputs, while appearing unmodified to human observers.\nPotential attacks include having malicious content like malware identified as\nlegitimate or controlling vehicle behavior. Yet, all existing adversarial\nexample attacks require knowledge of either the model internals or its training\ndata. We introduce the first practical demonstration of an attacker controlling\na remotely hosted DNN with no such knowledge. Indeed, the only capability of\nour black-box adversary is to observe labels given by the DNN to chosen inputs.\nOur attack strategy consists in training a local model to substitute for the\ntarget DNN, using inputs synthetically generated by an adversary and labeled by\nthe target DNN. We use the local substitute to craft adversarial examples, and\nfind that they are misclassified by the targeted DNN. To perform a real-world\nand properly-blinded evaluation, we attack a DNN hosted by MetaMind, an online\ndeep learning API. We find that their DNN misclassifies 84.24% of the\nadversarial examples crafted with our substitute. We demonstrate the general\napplicability of our strategy to many ML techniques by conducting the same\nattack against models hosted by Amazon and Google, using logistic regression\nsubstitutes. They yield adversarial examples misclassified by Amazon and Google\nat rates of 96.19% and 88.94%. We also find that this black-box attack strategy\nis capable of evading defense strategies previously found to make adversarial\nexample crafting harder.', 'Model-agnostic meta-learning (MAML) is a popular method for few-shot learning\nbut assumes that we have access to the meta-training set. In practice, training\non the meta-training set may not always be an option due to data privacy\nconcerns, intellectual property issues, or merely lack of computing resources.\nIn this paper, we consider the novel problem of repurposing pretrained MAML\ncheckpoints to solve new few-shot classification tasks. Because of the\npotential distribution mismatch, the original MAML steps may no longer be\noptimal. Therefore we propose an alternative meta-testing procedure and combine\nMAML gradient steps with adversarial training and uncertainty-based stepsize\nadaptation. Our method outperforms ""vanilla"" MAML on same-domain and\ncross-domains benchmarks using both SGD and Adam optimizers and shows improved\nrobustness to the choice of base stepsize.', ""Model Agnostic Meta-Learning (MAML) has emerged as a standard framework for\nmeta-learning, where a meta-model is learned with the ability of fast adapting\nto new tasks. However, as a double-looped optimization problem, MAML needs to\ndifferentiate through the whole inner-loop optimization path for every\nouter-loop training step, which may lead to both computational inefficiency and\nsub-optimal solutions. In this paper, we generalize MAML to allow meta-learning\nto be defined in function spaces, and propose the first meta-learning paradigm\nin the Reproducing Kernel Hilbert Space (RKHS) induced by the meta-model's\nNeural Tangent Kernel (NTK). Within this paradigm, we introduce two\nmeta-learning algorithms in the RKHS, which no longer need a sub-optimal\niterative inner-loop adaptation as in the MAML framework. We achieve this goal\nby 1) replacing the adaptation with a fast-adaptive regularizer in the RKHS;\nand 2) solving the adaptation analytically based on the NTK theory. Extensive\nexperimental studies demonstrate advantages of our paradigm in both efficiency\nand quality of solutions compared to related meta-learning algorithms. Another\ninteresting feature of our proposed methods is that they are demonstrated to be\nmore robust to adversarial attacks and out-of-distribution adaptation than\npopular baselines, as demonstrated in our experiments."", 'Adversarial examples are intentionally crafted data with the purpose of\ndeceiving neural networks into misclassification. When we talk about strategies\nto create such examples, we usually refer to perturbation-based methods that\nfabricate adversarial examples by applying invisible perturbations onto normal\ndata. The resulting data reserve their visual appearance to human observers,\nyet can be totally unrecognizable to DNN models, which in turn leads to\ncompletely misleading predictions. In this paper, however, we consider crafting\nadversarial examples from existing data as a limitation to example diversity.\nWe propose a non-perturbation-based framework that generates native adversarial\nexamples from class-conditional generative adversarial networks.As such, the\ngenerated data will not resemble any existing data and thus expand example\ndiversity, raising the difficulty in adversarial defense. We then extend this\nframework to pre-trained conditional GANs, in which we turn an existing\ngenerator into an ""adversarial-example generator"". We conduct experiments on\nour approach for MNIST and CIFAR10 datasets and have satisfactory results,\nshowing that this approach can be a potential alternative to previous attack\nstrategies.', ""In this paper, we study the generalization properties of Model-Agnostic\nMeta-Learning (MAML) algorithms for supervised learning problems. We focus on\nthe setting in which we train the MAML model over $m$ tasks, each with $n$ data\npoints, and characterize its generalization error from two points of view:\nFirst, we assume the new task at test time is one of the training tasks, and we\nshow that, for strongly convex objective functions, the expected excess\npopulation loss is bounded by ${\\mathcal{O}}(1/mn)$. Second, we consider the\nMAML algorithm's generalization to an unseen task and show that the resulting\ngeneralization error depends on the total variation distance between the\nunderlying distributions of the new task and the tasks observed during the\ntraining process. Our proof techniques rely on the connections between\nalgorithmic stability and generalization bounds of algorithms. In particular,\nwe propose a new definition of stability for meta-learning algorithms, which\nallows us to capture the role of both the number of tasks $m$ and number of\nsamples per task $n$ on the generalization error of MAML."", 'The vulnerability of deep neural networks to adversarial examples has\nmotivated an increasing number of defense strategies for promoting model\nrobustness. However, the progress is usually hampered by insufficient\nrobustness evaluations. As the de facto standard to evaluate adversarial\nrobustness, adversarial attacks typically solve an optimization problem of\ncrafting adversarial examples with an iterative process. In this work, we\npropose a Model-Agnostic Meta-Attack (MAMA) approach to discover stronger\nattack algorithms automatically. Our method learns the optimizer in adversarial\nattacks parameterized by a recurrent neural network, which is trained over a\nclass of data samples and defenses to produce effective update directions\nduring adversarial example generation. Furthermore, we develop a model-agnostic\ntraining algorithm to improve the generalization ability of the learned\noptimizer when attacking unseen defenses. Our approach can be flexibly\nincorporated with various attacks and consistently improves the performance\nwith little extra computational cost. Extensive experiments demonstrate the\neffectiveness of the learned attacks by MAMA compared to the state-of-the-art\nattacks on different defenses, leading to a more reliable evaluation of\nadversarial robustness.', ""Deep Learning algorithms have achieved the state-of-the-art performance for\nImage Classification and have been used even in security-critical applications,\nsuch as biometric recognition systems and self-driving cars. However, recent\nworks have shown those algorithms, which can even surpass the human\ncapabilities, are vulnerable to adversarial examples. In Computer Vision,\nadversarial examples are images containing subtle perturbations generated by\nmalicious optimization algorithms in order to fool classifiers. As an attempt\nto mitigate these vulnerabilities, numerous countermeasures have been\nconstantly proposed in literature. Nevertheless, devising an efficient defense\nmechanism has proven to be a difficult task, since many approaches have already\nshown to be ineffective to adaptive attackers. Thus, this self-containing paper\naims to provide all readerships with a review of the latest research progress\non Adversarial Machine Learning in Image Classification, however with a\ndefender's perspective. Here, novel taxonomies for categorizing adversarial\nattacks and defenses are introduced and discussions about the existence of\nadversarial examples are provided. Further, in contrast to exisiting surveys,\nit is also given relevant guidance that should be taken into consideration by\nresearchers when devising and evaluating defenses. Finally, based on the\nreviewed literature, it is discussed some promising paths for future research."", 'The fact that deep neural networks are susceptible to crafted perturbations\nseverely impacts the use of deep learning in certain domains of application.\nAmong many developed defense models against such attacks, adversarial training\nemerges as the most successful method that consistently resists a wide range of\nattacks. In this work, based on an observation from a previous study that the\nrepresentations of a clean data example and its adversarial examples become\nmore divergent in higher layers of a deep neural net, we propose the Adversary\nDivergence Reduction Network which enforces local/global compactness and the\nclustering assumption over an intermediate layer of a deep neural network. We\nconduct comprehensive experiments to understand the isolating behavior of each\ncomponent (i.e., local/global compactness and the clustering assumption) and\ncompare our proposed model with state-of-the-art adversarial training methods.\nThe experimental results demonstrate that augmenting adversarial training with\nour proposed components can further improve the robustness of the network,\nleading to higher unperturbed and adversarial predictive performances.', 'Nowadays, Deep Neural Networks (DNNs) report state-of-the-art results in many\nmachine learning areas, including intrusion detection. Nevertheless, recent\nstudies in computer vision have shown that DNNs can be vulnerable to\nadversarial attacks that are capable of deceiving them into misclassification\nby injecting specially crafted data. In security-critical areas, such attacks\ncan cause serious damage; therefore, in this paper, we examine the effect of\nadversarial attacks on deep learning-based intrusion detection. In addition, we\ninvestigate the effectiveness of adversarial training as a defense against such\nattacks. Experimental results show that with sufficient distortion, adversarial\nexamples are able to mislead the detector and that the use of adversarial\ntraining can improve the robustness of intrusion detection.', 'Adversarial attacks by generating examples which are almost indistinguishable\nfrom natural examples, pose a serious threat to learning models. Defending\nagainst adversarial attacks is a critical element for a reliable learning\nsystem. Support vector machine (SVM) is a classical yet still important\nlearning algorithm even in the current deep learning era. Although a wide range\nof researches have been done in recent years to improve the adversarial\nrobustness of learning models, but most of them are limited to deep neural\nnetworks (DNNs) and the work for kernel SVM is still vacant. In this paper, we\naim at kernel SVM and propose adv-SVM to improve its adversarial robustness via\nadversarial training, which has been demonstrated to be the most promising\ndefense techniques. To the best of our knowledge, this is the first work that\ndevotes to the fast and scalable adversarial training of kernel SVM.\nSpecifically, we first build connection of perturbations of samples between\noriginal and kernel spaces, and then give a reduced and equivalent formulation\nof adversarial training of kernel SVM based on the connection. Next, doubly\nstochastic gradients (DSG) based on two unbiased stochastic approximations\n(i.e., one is on training points and another is on random features) are applied\nto update the solution of our objective function. Finally, we prove that our\nalgorithm optimized by DSG converges to the optimal solution at the rate of\nO(1/t) under the constant and diminishing stepsizes. Comprehensive experimental\nresults show that our adversarial training algorithm enjoys robustness against\nvarious attacks and meanwhile has the similar efficiency and scalability with\nclassical DSG algorithm.', 'Delusive attacks aim to substantially deteriorate the test accuracy of the\nlearning model by slightly perturbing the features of correctly labeled\ntraining examples. By formalizing this malicious attack as finding the\nworst-case training data within a specific $\\infty$-Wasserstein ball, we show\nthat minimizing adversarial risk on the perturbed data is equivalent to\noptimizing an upper bound of natural risk on the original data. This implies\nthat adversarial training can serve as a principled defense against delusive\nattacks. Thus, the test accuracy decreased by delusive attacks can be largely\nrecovered by adversarial training. To further understand the internal mechanism\nof the defense, we disclose that adversarial training can resist the delusive\nperturbations by preventing the learner from overly relying on non-robust\nfeatures in a natural setting. Finally, we complement our theoretical findings\nwith a set of experiments on popular benchmark datasets, which show that the\ndefense withstands six different practical attacks. Both theoretical and\nempirical results vote for adversarial training when confronted with delusive\nadversaries.', 'Robustness to adversarial examples of machine learning models remains an open\ntopic of research. Attacks often succeed by repeatedly probing a fixed target\nmodel with adversarial examples purposely crafted to fool it. In this paper, we\nintroduce Morphence, an approach that shifts the defense landscape by making a\nmodel a moving target against adversarial examples. By regularly moving the\ndecision function of a model, Morphence makes it significantly challenging for\nrepeated or correlated attacks to succeed. Morphence deploys a pool of models\ngenerated from a base model in a manner that introduces sufficient randomness\nwhen it responds to prediction queries. To ensure repeated or correlated\nattacks fail, the deployed pool of models automatically expires after a query\nbudget is reached and the model pool is seamlessly replaced by a new model pool\ngenerated in advance. We evaluate Morphence on two benchmark image\nclassification datasets (MNIST and CIFAR10) against five reference attacks (2\nwhite-box and 3 black-box). In all cases, Morphence consistently outperforms\nthe thus-far effective defense, adversarial training, even in the face of\nstrong white-box attacks, while preserving accuracy on clean data.', 'Data sets for fairness relevant tasks can lack examples or be biased\naccording to a specific label in a sensitive attribute. We demonstrate the\nusefulness of weight based meta-learning approaches in such situations. For\nmodels that can be trained through gradient descent, we demonstrate that there\nare some parameter configurations that allow models to be optimized from a few\nnumber of gradient steps and with minimal data which are both fair and\naccurate. To learn such weight sets, we adapt the popular MAML algorithm to\nFair-MAML by the inclusion of a fairness regularization term. In practice,\nFair-MAML allows practitioners to train fair machine learning models from only\na few examples when data from related tasks is available. We empirically\nexhibit the value of this technique by comparing to relevant baselines.', 'The burgeoning success of deep learning has raised the security and privacy\nconcerns as more and more tasks are accompanied with sensitive data.\nAdversarial attacks in deep learning have emerged as one of the dominating\nsecurity threat to a range of mission-critical deep learning systems and\napplications. This paper takes a holistic and principled approach to perform\nstatistical characterization of adversarial examples in deep learning. We\nprovide a general formulation of adversarial examples and elaborate on the\nbasic principle for adversarial attack algorithm design. We introduce easy and\nhard categorization of adversarial attacks to analyze the effectiveness of\nadversarial examples in terms of attack success rate, degree of change in\nadversarial perturbation, average entropy of prediction qualities, and fraction\nof adversarial examples that lead to successful attacks. We conduct extensive\nexperimental study on adversarial behavior in easy and hard attacks under deep\nlearning models with different hyperparameters and different deep learning\nframeworks. We show that the same adversarial attack behaves differently under\ndifferent hyperparameters and across different frameworks due to the different\nfeatures learned under different deep learning model training process. Our\nstatistical characterization with strong empirical evidence provides a\ntransformative enlightenment on mitigation strategies towards effective\ncountermeasures against present and future adversarial attacks.', 'Deep neural networks (DNN) have achieved unprecedented success in numerous\nmachine learning tasks in various domains. However, the existence of\nadversarial examples has raised concerns about applying deep learning to\nsafety-critical applications. As a result, we have witnessed increasing\ninterests in studying attack and defense mechanisms for DNN models on different\ndata types, such as images, graphs and text. Thus, it is necessary to provide a\nsystematic and comprehensive overview of the main threats of attacks and the\nsuccess of corresponding countermeasures. In this survey, we review the state\nof the art algorithms for generating adversarial examples and the\ncountermeasures against adversarial examples, for the three popular data types,\ni.e., images, graphs and text.', 'Meta learning algorithms have been widely applied in many tasks for efficient\nlearning, such as few-shot image classification and fast reinforcement\nlearning. During meta training, the meta learner develops a common learning\nstrategy, or experience, from a variety of learning tasks. Therefore, during\nmeta test, the meta learner can use the learned strategy to quickly adapt to\nnew tasks even with a few training samples. However, there is still a dark side\nabout meta learning in terms of reliability and robustness. In particular, is\nmeta learning vulnerable to adversarial attacks? In other words, would a\nwell-trained meta learner utilize its learned experience to build wrong or\nlikely useless knowledge, if an adversary unnoticeably manipulates the given\ntraining set? Without the understanding of this problem, it is extremely risky\nto apply meta learning in safety-critical applications. Thus, in this paper, we\nperform the initial study about adversarial attacks on meta learning under the\nfew-shot classification problem. In particular, we formally define key elements\nof adversarial attacks unique to meta learning and propose the first attacking\nalgorithm against meta learning under various settings. We evaluate the\neffectiveness of the proposed attacking strategy as well as the robustness of\nseveral representative meta learning algorithms. Experimental results\ndemonstrate that the proposed attacking strategy can easily break the meta\nlearner and meta learning is vulnerable to adversarial attacks. The\nimplementation of the proposed framework will be released upon the acceptance\nof this paper.', 'Model-Agnostic Meta-Learning (MAML) and its variants have achieved success in\nmeta-learning tasks on many datasets and settings. On the other hand, we have\njust started to understand and analyze how they are able to adapt fast to new\ntasks. For example, one popular hypothesis is that the algorithms learn good\nrepresentations for transfer, as in multi-task learning. In this work, we\ncontribute by providing a series of empirical and theoretical studies, and\ndiscover several interesting yet previously unknown properties of the\nalgorithm. We find MAML adapts better with a deep architecture even if the\ntasks need only a shallow one (and thus, no representation learning is needed).\nWhile echoing previous findings by others that the bottom layers in deep\narchitectures enable representation learning, we also find that upper layers\nenable fast adaptation by being meta-learned to perform adaptive gradient\nupdate when generalizing to new tasks. Motivated by these findings, we study\nseveral meta-optimization approaches and propose a new one for learning to\noptimize adaptively. Those approaches attain stronger performance in\nmeta-learning both shallower and deeper architectures than MAML.', 'In few-shot learning scenarios, the challenge is to generalize and perform\nwell on new unseen examples when only very few labeled examples are available\nfor each task. Model-agnostic meta-learning (MAML) has gained the popularity as\none of the representative few-shot learning methods for its flexibility and\napplicability to diverse problems. However, MAML and its variants often resort\nto a simple loss function without any auxiliary loss function or regularization\nterms that can help achieve better generalization. The problem lies in that\neach application and task may require different auxiliary loss function,\nespecially when tasks are diverse and distinct. Instead of attempting to\nhand-design an auxiliary loss function for each application and task, we\nintroduce a new meta-learning framework with a loss function that adapts to\neach task. Our proposed framework, named Meta-Learning with Task-Adaptive Loss\nFunction (MeTAL), demonstrates the effectiveness and the flexibility across\nvarious domains, such as few-shot classification and few-shot regression.', 'Adversarial examples are perturbed inputs designed to fool machine learning\nmodels. Most recent works on adversarial examples for image classification\nfocus on directly modifying pixels with minor perturbations. A common\nrequirement in all these works is that the malicious perturbations should be\nsmall enough (measured by an L_p norm for some p) so that they are\nimperceptible to humans. However, small perturbations can be unnecessarily\nrestrictive and limit the diversity of adversarial examples generated. Further,\nan L_p norm based distance metric ignores important structure patterns hidden\nin images that are important to human perception. Consequently, even the minor\nperturbation introduced in recent works often makes the adversarial examples\nless natural to humans. More importantly, they often do not transfer well and\nare therefore less effective when attacking black-box models especially for\nthose protected by a defense mechanism. In this paper, we propose a\nstructure-preserving transformation (SPT) for generating natural and diverse\nadversarial examples with extremely high transferability. The key idea of our\napproach is to allow perceptible deviation in adversarial examples while\nkeeping structure patterns that are central to a human classifier. Empirical\nresults on the MNIST and the fashion-MNIST datasets show that adversarial\nexamples generated by our approach can easily bypass strong adversarial\ntraining. Further, they transfer well to other target models with no loss or\nlittle loss of successful attack rate.', 'Recent years have witnessed the deployment of adversarial attacks to evaluate\nthe robustness of Neural Networks. Past work in this field has relied on\ntraditional optimization algorithms that ignore the inherent structure of the\nproblem and data, or generative methods that rely purely on learning and often\nfail to generate adversarial examples where they are hard to find. To alleviate\nthese deficiencies, we propose a novel attack based on a graph neural network\n(GNN) that takes advantage of the strengths of both approaches; we call it\nAdvGNN. Our GNN architecture closely resembles the network we wish to attack.\nDuring inference, we perform forward-backward passes through the GNN layers to\nguide an iterative procedure towards adversarial examples. During training, its\nparameters are estimated via a loss function that encourages the efficient\ncomputation of adversarial examples over a time horizon. We show that our\nmethod beats state-of-the-art adversarial attacks, including PGD-attack,\nMI-FGSM, and Carlini and Wagner attack, reducing the time required to generate\nadversarial examples with small perturbation norms by over 65\\%. Moreover,\nAdvGNN achieves good generalization performance on unseen networks. Finally, we\nprovide a new challenging dataset specifically designed to allow for a more\nillustrative comparison of adversarial attacks.', 'There is a growing interest in the learning-to-learn paradigm, also known as\nmeta-learning, where models infer on new tasks using a few training examples.\nRecently, meta-learning based methods have been widely used in few-shot\nclassification, regression, reinforcement learning, and domain adaptation. The\nmodel-agnostic meta-learning (MAML) algorithm is a well-known algorithm that\nobtains model parameter initialization at meta-training phase. In the meta-test\nphase, this initialization is rapidly adapted to new tasks by using gradient\ndescent. However, meta-learning models are prone to overfitting since there are\ninsufficient training tasks resulting in over-parameterized models with poor\ngeneralization performance for unseen tasks. In this paper, we propose a\nBayesian neural network based MAML algorithm, which we refer to as the B-SMALL\nalgorithm. The proposed framework incorporates a sparse variational loss term\nalongside the loss function of MAML, which uses a sparsifying approximated KL\ndivergence as a regularizer. We demonstrate the performance of B-MAML using\nclassification and regression tasks, and highlight that training a sparsifying\nBNN using MAML indeed improves the parameter footprint of the model while\nperforming at par or even outperforming the MAML approach. We also illustrate\napplicability of our approach in distributed sensor networks, where sparsity\nand meta-learning can be beneficial.', 'Deep neural networks are vulnerable to adversarial examples that are crafted\nby imposing imperceptible changes to the inputs. However, these adversarial\nexamples are most successful in white-box settings where the model and its\nparameters are available. Finding adversarial examples that are transferable to\nother models or developed in a black-box setting is significantly more\ndifficult. In this paper, we propose the Direction-Aggregated adversarial\nattacks that deliver transferable adversarial examples. Our method utilizes\naggregated direction during the attack process for avoiding the generated\nadversarial examples overfitting to the white-box model. Extensive experiments\non ImageNet show that our proposed method improves the transferability of\nadversarial examples significantly and outperforms state-of-the-art attacks,\nespecially against adversarial robust models. The best averaged attack success\nrates of our proposed method reaches 94.6\\% against three adversarial trained\nmodels and 94.8\\% against five defense methods. It also reveals that current\ndefense approaches do not prevent transferable adversarial attacks.', ""We present a novel Balanced Incremental Model Agnostic Meta Learning system\n(BI-MAML) for learning multiple tasks. Our method implements a meta-update rule\nto incrementally adapt its model to new tasks without forgetting old tasks.\nSuch a capability is not possible in current state-of-the-art MAML approaches.\nThese methods effectively adapt to new tasks, however, suffer from\n'catastrophic forgetting' phenomena, in which new tasks that are streamed into\nthe model degrade the performance of the model on previously learned tasks. Our\nsystem performs the meta-updates with only a few-shots and can successfully\naccomplish them. Our key idea for achieving this is the design of balanced\nlearning strategy for the baseline model. The strategy sets the baseline model\nto perform equally well on various tasks and incorporates time efficiency. The\nbalanced learning strategy enables BI-MAML to both outperform other\nstate-of-the-art models in terms of classification accuracy for existing tasks\nand also accomplish efficient adaption to similar new tasks with less required\nshots. We evaluate BI-MAML by conducting comparisons on two common benchmark\ndatasets with multiple number of image classification tasks. BI-MAML\nperformance demonstrates advantages in both accuracy and efficiency."", 'Deep neural networks (DNNs) are vulnerable to adversarial examples with small\nperturbations. Adversarial defense thus has been an important means which\nimproves the robustness of DNNs by defending against adversarial examples.\nExisting defense methods focus on some specific types of adversarial examples\nand may fail to defend well in real-world applications. In practice, we may\nface many types of attacks where the exact type of adversarial examples in\nreal-world applications can be even unknown. In this paper, motivated by that\nadversarial examples are more likely to appear near the classification\nboundary, we study adversarial examples from a new perspective that whether we\ncan defend against adversarial examples by pulling them back to the original\nclean distribution. We theoretically and empirically verify the existence of\ndefense affine transformations that restore adversarial examples. Relying on\nthis, we learn a defense transformer to counterattack the adversarial examples\nby parameterizing the affine transformations and exploiting the boundary\ninformation of DNNs. Extensive experiments on both toy and real-world datasets\ndemonstrate the effectiveness and generalization of our defense transformer.', 'Model-agnostic meta-learning (MAML) formulates meta-learning as a bilevel\noptimization problem, where the inner level solves each subtask based on a\nshared prior, while the outer level searches for the optimal shared prior by\noptimizing its aggregated performance over all the subtasks. Despite its\nempirical success, MAML remains less understood in theory, especially in terms\nof its global optimality, due to the nonconvexity of the meta-objective (the\nouter-level objective). To bridge such a gap between theory and practice, we\ncharacterize the optimality gap of the stationary points attained by MAML for\nboth reinforcement learning and supervised learning, where the inner-level and\nouter-level problems are solved via first-order optimization methods. In\nparticular, our characterization connects the optimality gap of such stationary\npoints with (i) the functional geometry of inner-level objectives and (ii) the\nrepresentation power of function approximators, including linear models and\nneural networks. To the best of our knowledge, our analysis establishes the\nglobal optimality of MAML with nonconvex meta-objectives for the first time.', 'Deep neural networks (DNNs) are vulnerable to adversarial noise. Their\nadversarial robustness can be improved by exploiting adversarial examples.\nHowever, given the continuously evolving attacks, models trained on seen types\nof adversarial examples generally cannot generalize well to unseen types of\nadversarial examples. To solve this problem, in this paper, we propose to\nremove adversarial noise by learning generalizable invariant features across\nattacks which maintain semantic classification information. Specifically, we\nintroduce an adversarial feature learning mechanism to disentangle invariant\nfeatures from adversarial noise. A normalization term has been proposed in the\nencoded space of the attack-invariant features to address the bias issue\nbetween the seen and unseen types of attacks. Empirical evaluations demonstrate\nthat our method could provide better protection in comparison to previous\nstate-of-the-art approaches, especially against unseen types of attacks and\nadaptive attacks.', 'In recent years, deep neural networks have demonstrated outstanding\nperformance in many machine learning tasks. However, researchers have\ndiscovered that these state-of-the-art models are vulnerable to adversarial\nexamples: legitimate examples added by small perturbations which are\nunnoticeable to human eyes. Adversarial training, which augments the training\ndata with adversarial examples during the training process, is a well known\ndefense to improve the robustness of the model against adversarial attacks.\nHowever, this robustness is only effective to the same attack method used for\nadversarial training. Madry et al.(2017) suggest that effectiveness of\niterative multi-step adversarial attacks and particularly that projected\ngradient descent (PGD) may be considered the universal first order adversary\nand applying the adversarial training with PGD implies resistance against many\nother first order attacks. However, the computational cost of the adversarial\ntraining with PGD and other multi-step adversarial examples is much higher than\nthat of the adversarial training with other simpler attack techniques. In this\npaper, we show how strong adversarial examples can be generated only at a cost\nsimilar to that of two runs of the fast gradient sign method (FGSM), allowing\ndefense against adversarial attacks with a robustness level comparable to that\nof the adversarial training with multi-step adversarial examples. We\nempirically demonstrate the effectiveness of the proposed two-step defense\napproach against different attack methods and its improvements over existing\ndefense strategies.', 'The existence of adversarial data examples has drawn significant attention in\nthe deep-learning community; such data are seemingly minimally perturbed\nrelative to the original data, but lead to very different outputs from a\ndeep-learning algorithm. Although a significant body of work on developing\ndefensive models has been considered, most such models are heuristic and are\noften vulnerable to adaptive attacks. Defensive methods that provide\ntheoretical robustness guarantees have been studied intensively, yet most fail\nto obtain non-trivial robustness when a large-scale model and data are present.\nTo address these limitations, we introduce a framework that is scalable and\nprovides certified bounds on the norm of the input manipulation for\nconstructing adversarial examples. We establish a connection between robustness\nagainst adversarial perturbation and additive random noise, and propose a\ntraining strategy that can significantly improve the certified bounds. Our\nevaluation on MNIST, CIFAR-10 and ImageNet suggests that the proposed method is\nscalable to complicated models and large data sets, while providing competitive\nrobustness to state-of-the-art provable defense methods.', 'Transferability of adversarial examples is a key issue to apply this kind of\nattacks against multimedia forensics (MMF) techniques based on Deep Learning\n(DL) in a real-life setting. Adversarial example transferability, in fact,\nwould open the way to the deployment of successful counter forensics attacks\nalso in cases where the attacker does not have a full knowledge of the\nto-be-attacked system. Some preliminary works have shown that adversarial\nexamples against CNN-based image forensics detectors are in general\nnon-transferrable, at least when the basic versions of the attacks implemented\nin the most popular libraries are adopted. In this paper, we introduce a\ngeneral strategy to increase the strength of the attacks and evaluate their\ntransferability when such a strength varies. We experimentally show that, in\nthis way, attack transferability can be largely increased, at the expense of a\nlarger distortion. Our research confirms the security threats posed by the\nexistence of adversarial examples even in multimedia forensics scenarios, thus\ncalling for new defense strategies to improve the security of DL-based MMF\ntechniques.', 'Model-agnostic meta-learning (MAML) is a well-known optimization-based\nmeta-learning algorithm that works well in various computer vision tasks, e.g.,\nfew-shot classification. MAML is to learn an initialization so that a model can\nadapt to a new task in a few steps. However, since the gradient norm of a\nclassifier (head) is much bigger than those of backbone layers, the model\nfocuses on learning the decision boundary of the classifier with similar\nrepresentations. Furthermore, gradient norms of high-level layers are small\nthan those of the other layers. So, the backbone of MAML usually learns\ntask-generic features, which results in deteriorated adaptation performance in\nthe inner-loop. To resolve or mitigate this problem, we propose contextual\ngradient scaling (CxGrad), which scales gradient norms of the backbone to\nfacilitate learning task-specific knowledge in the inner-loop. Since the\nscaling factors are generated from task-conditioned parameters, gradient norms\nof the backbone can be scaled in a task-wise fashion. Experimental results show\nthat CxGrad effectively encourages the backbone to learn task-specific\nknowledge in the inner-loop and improves the performance of MAML up to a\nsignificant margin in both same- and cross-domain few-shot classification.', 'Model-Agnostic Meta-Learning (MAML) has become increasingly popular for\ntraining models that can quickly adapt to new tasks via one or few stochastic\ngradient descent steps. However, the MAML objective is significantly more\ndifficult to optimize compared to standard non-adaptive learning (NAL), and\nlittle is understood about how much MAML improves over NAL in terms of the fast\nadaptability of their solutions in various scenarios. We analytically address\nthis issue in a linear regression setting consisting of a mixture of easy and\nhard tasks, where hardness is related to the rate that gradient descent\nconverges on the task. Specifically, we prove that in order for MAML to achieve\nsubstantial gain over NAL, (i) there must be some discrepancy in hardness among\nthe tasks, and (ii) the optimal solutions of the hard tasks must be closely\npacked with the center far from the center of the easy tasks optimal solutions.\nWe also give numerical and analytical results suggesting that these insights\napply to two-layer neural networks. Finally, we provide few-shot image\nclassification experiments that support our insights for when MAML should be\nused and emphasize the importance of training MAML on hard tasks in practice.', 'Model-agnostic meta-learning (MAML) is a meta-learning technique to train a\nmodel on a multitude of learning tasks in a way that primes the model for\nfew-shot learning of new tasks. The MAML algorithm performs well on few-shot\nlearning problems in classification, regression, and fine-tuning of policy\ngradients in reinforcement learning, but comes with the need for costly\nhyperparameter tuning for training stability. We address this shortcoming by\nintroducing an extension to MAML, called Alpha MAML, to incorporate an online\nhyperparameter adaptation scheme that eliminates the need to tune meta-learning\nand learning rates. Our results with the Omniglot database demonstrate a\nsubstantial reduction in the need to tune MAML training hyperparameters and\nimprovement to training stability with less sensitivity to hyperparameter\nchoice.', ""Deep neural network based medical image systems are vulnerable to adversarial\nexamples. Many defense mechanisms have been proposed in the literature,\nhowever, the existing defenses assume a passive attacker who knows little about\nthe defense system and does not change the attack strategy according to the\ndefense. Recent works have shown that a strong adaptive attack, where an\nattacker is assumed to have full knowledge about the defense system, can easily\nbypass the existing defenses. In this paper, we propose a novel adversarial\nexample defense system called Medical Aegis. To the best of our knowledge,\nMedical Aegis is the first defense in the literature that successfully\naddresses the strong adaptive adversarial example attacks to medical images.\nMedical Aegis boasts two-tier protectors: The first tier of Cushion weakens the\nadversarial manipulation capability of an attack by removing its high-frequency\ncomponents, yet posing a minimal effect on classification performance of the\noriginal image; the second tier of Shield learns a set of per-class DNN models\nto predict the logits of the protected model. Deviation from the Shield's\nprediction indicates adversarial examples. Shield is inspired by the\nobservations in our stress tests that there exist robust trails in the shallow\nlayers of a DNN model, which the adaptive attacks can hardly destruct.\nExperimental results show that the proposed defense accurately detects adaptive\nattacks, with negligible overhead for model inference."", 'We identify a trade-off between robustness and accuracy that serves as a\nguiding principle in the design of defenses against adversarial examples.\nAlthough this problem has been widely studied empirically, much remains unknown\nconcerning the theory underlying this trade-off. In this work, we decompose the\nprediction error for adversarial examples (robust error) as the sum of the\nnatural (classification) error and boundary error, and provide a differentiable\nupper bound using the theory of classification-calibrated loss, which is shown\nto be the tightest possible upper bound uniform over all probability\ndistributions and measurable predictors. Inspired by our theoretical analysis,\nwe also design a new defense method, TRADES, to trade adversarial robustness\noff against accuracy. Our proposed algorithm performs well experimentally in\nreal-world datasets. The methodology is the foundation of our entry to the\nNeurIPS 2018 Adversarial Vision Challenge in which we won the 1st place out of\n~2,000 submissions, surpassing the runner-up approach by $11.41\\%$ in terms of\nmean $\\ell_2$ perturbation distance.', 'Adversarial training, in which a network is trained on adversarial examples,\nis one of the few defenses against adversarial attacks that withstands strong\nattacks. Unfortunately, the high cost of generating strong adversarial examples\nmakes standard adversarial training impractical on large-scale problems like\nImageNet. We present an algorithm that eliminates the overhead cost of\ngenerating adversarial examples by recycling the gradient information computed\nwhen updating model parameters. Our ""free"" adversarial training algorithm\nachieves comparable robustness to PGD adversarial training on the CIFAR-10 and\nCIFAR-100 datasets at negligible additional cost compared to natural training,\nand can be 7 to 30 times faster than other strong adversarial training methods.\nUsing a single workstation with 4 P100 GPUs and 2 days of runtime, we can train\na robust model for the large-scale ImageNet classification task that maintains\n40% accuracy against PGD attacks. The code is available at\nhttps://github.com/ashafahi/free_adv_train.', 'We address the problem of adversarial examples in machine learning where an\nadversary tries to misguide a classifier by making functionality-preserving\nmodifications to original samples. We assume a black-box scenario where the\nadversary has access to only the feature set, and the final hard-decision\noutput of the classifier. We propose a method to generate adversarial examples\nusing the minimum description length (MDL) principle. Our final aim is to\nimprove the robustness of the classifier by considering generated examples in\nrebuilding the classifier. We evaluate our method for the application of static\nmalware detection in portable executable (PE) files. We consider API calls of\nPE files as their distinguishing features where the feature vector is a binary\nvector representing the presence-absence of API calls. In our method, we first\ncreate a dataset of benign samples by querying the target classifier. We next\nconstruct a code table of frequent patterns for the compression of this dataset\nusing the MDL principle. We finally generate an adversarial example\ncorresponding to a malware sample by selecting and adding a pattern from the\nbenign code table to the malware sample. The selected pattern is the one that\nminimizes the length of the compressed adversarial example given the code\ntable. This modification preserves the functionalities of the original malware\nsample as all original API calls are kept, and only some new API calls are\nadded. Considering a neural network, we show that the evasion rate is 78.24\npercent for adversarial examples compared to 8.16 percent for original malware\nsamples. This shows the effectiveness of our method in generating examples that\nneed to be considered in rebuilding the classifier.', 'Despite the remarkable performance and generalization levels of deep learning\nmodels in a wide range of artificial intelligence tasks, it has been\ndemonstrated that these models can be easily fooled by the addition of\nimperceptible yet malicious perturbations to natural inputs. These altered\ninputs are known in the literature as adversarial examples. In this paper, we\npropose a novel probabilistic framework to generalize and extend adversarial\nattacks in order to produce a desired probability distribution for the classes\nwhen we apply the attack method to a large number of inputs. This novel attack\nparadigm provides the adversary with greater control over the target model,\nthereby exposing, in a wide range of scenarios, threats against deep learning\nmodels that cannot be conducted by the conventional paradigms. We introduce\nfour different strategies to efficiently generate such attacks, and illustrate\nour approach by extending multiple adversarial attack algorithms. We also\nexperimentally validate our approach for the spoken command classification task\nand the Tweet emotion classification task, two exemplary machine learning\nproblems in the audio and text domain, respectively. Our results demonstrate\nthat we can closely approximate any probability distribution for the classes\nwhile maintaining a high fooling rate and even prevent the attacks from being\ndetected by label-shift detection methods.', 'Recent studies have shown that state-of-the-art deep learning models are\nvulnerable to the inputs with small perturbations (adversarial examples). We\nobserve two critical obstacles in adversarial examples: (i) Strong adversarial\nattacks (e.g., C&W attack) require manually tuning hyper-parameters and take a\nlong time to construct an adversarial example, making it impractical to attack\nreal-time systems; (ii) Most of the studies focus on non-sequential tasks, such\nas image classification, yet only a few consider sequential tasks. In this\nwork, we speed up adversarial attacks, especially on sequential learning tasks.\nBy leveraging the uncertainty of each task, we directly learn the adaptive\nmulti-task weightings, without manually searching hyper-parameters. A unified\narchitecture is developed and evaluated for both non-sequential tasks and\nsequential ones. To validate the effectiveness, we take the scene text\nrecognition task as a case study. To our best knowledge, our proposed method is\nthe first attempt to adversarial attack for scene text recognition. Adaptive\nAttack achieves over 99.9\\% success rate with 3-6X speedup compared to\nstate-of-the-art adversarial attacks.', 'Model Agnostic Meta-Learning (MAML) is one of the most representative of\ngradient-based meta-learning algorithms. MAML learns new tasks with a few data\nsamples using inner updates from a meta-initialization point and learns the\nmeta-initialization parameters with outer updates. It has recently been\nhypothesized that representation reuse, which makes little change in efficient\nrepresentations, is the dominant factor in the performance of the\nmeta-initialized model through MAML in contrast to representation change, which\ncauses a significant change in representations. In this study, we investigate\nthe necessity of representation change for the ultimate goal of few-shot\nlearning, which is solving domain-agnostic tasks. To this aim, we propose a\nnovel meta-learning algorithm, called BOIL (Body Only update in Inner Loop),\nwhich updates only the body (extractor) of the model and freezes the head\n(classifier) during inner loop updates. BOIL leverages representation change\nrather than representation reuse. This is because feature vectors\n(representations) have to move quickly to their corresponding frozen head\nvectors. We visualize this property using cosine similarity, CKA, and empirical\nresults without the head. BOIL empirically shows significant performance\nimprovement over MAML, particularly on cross-domain tasks. The results imply\nthat representation change in gradient-based meta-learning approaches is a\ncritical component.', 'Adversarial training is wildly considered as one of the most effective way to\ndefend against adversarial examples. However, existing adversarial training\nmethods consume unbearable time, due to the fact that they need to generate\nadversarial examples in the large input space. To speed up adversarial\ntraining, we propose a novel adversarial training method that does not need to\ngenerate real adversarial examples. By adding perturbations to logits to\ngenerate Endogenous Adversarial Examples (EAEs) -- the adversarial examples in\nthe latent space, the time consuming gradient calculation can be avoided.\nExtensive experiments are conducted on CIFAR-10 and ImageNet, and the results\nshow that comparing to state-of-the-art methods, our EAE adversarial training\nnot only shortens the training time, but also enhances the robustness of the\nmodel and has less impact on the accuracy of clean examples than the existing\nmethods.', 'As a popular meta-learning approach, the model-agnostic meta-learning (MAML)\nalgorithm has been widely used due to its simplicity and effectiveness.\nHowever, the convergence of the general multi-step MAML still remains\nunexplored. In this paper, we develop a new theoretical framework to provide\nsuch convergence guarantee for two types of objective functions that are of\ninterest in practice: (a) resampling case (e.g., reinforcement learning), where\nloss functions take the form in expectation and new data are sampled as the\nalgorithm runs; and (b) finite-sum case (e.g., supervised learning), where loss\nfunctions take the finite-sum form with given samples. For both cases, we\ncharacterize the convergence rate and the computational complexity to attain an\n$\\epsilon$-accurate solution for multi-step MAML in the general nonconvex\nsetting. In particular, our results suggest that an inner-stage stepsize needs\nto be chosen inversely proportional to the number $N$ of inner-stage steps in\norder for $N$-step MAML to have guaranteed convergence. From the technical\nperspective, we develop novel techniques to deal with the nested structure of\nthe meta gradient for multi-step MAML, which can be of independent interest.', 'White box adversarial perturbations are generated via iterative optimization\nalgorithms most often by minimizing an adversarial loss on a $\\ell_p$\nneighborhood of the original image, the so-called distortion set. Constraining\nthe adversarial search with different norms results in disparately structured\nadversarial examples. Here we explore several distortion sets with\nstructure-enhancing algorithms. These new structures for adversarial examples\nmight provide challenges for provable and empirical robust mechanisms. Because\nadversarial robustness is still an empirical field, defense mechanisms should\nalso reasonably be evaluated against differently structured attacks. Besides,\nthese structured adversarial perturbations may allow for larger distortions\nsize than their $\\ell_p$ counter-part while remaining imperceptible or\nperceptible as natural distortions of the image. We will demonstrate in this\nwork that the proposed structured adversarial examples can significantly bring\ndown the classification accuracy of adversarialy trained classifiers while\nshowing low $\\ell_2$ distortion rate. For instance, on ImagNet dataset the\nstructured attacks drop the accuracy of adversarial model to near zero with\nonly 50\\% of $\\ell_2$ distortion generated using white-box attacks like PGD. As\na byproduct, our finding on structured adversarial examples can be used for\nadversarial regularization of models to make models more robust or improve\ntheir generalization performance on datasets which are structurally different.', 'The problem of adversarial examples, evasion attacks on machine learning\nclassifiers, has proven extremely difficult to solve. This is true even when,\nas is the case in many practical settings, the classifier is hosted as a remote\nservice and so the adversary does not have direct access to the model\nparameters.\n  This paper argues that in such settings, defenders have a much larger space\nof actions than have been previously explored. Specifically, we deviate from\nthe implicit assumption made by prior work that a defense must be a stateless\nfunction that operates on individual examples, and explore the possibility for\nstateful defenses.\n  To begin, we develop a defense designed to detect the process of adversarial\nexample generation. By keeping a history of the past queries, a defender can\ntry to identify when a sequence of queries appears to be for the purpose of\ngenerating an adversarial example. We then introduce query blinding, a new\nclass of attacks designed to bypass defenses that rely on such a defense\napproach.\n  We believe that expanding the study of adversarial examples from stateless\nclassifiers to stateful systems is not only more realistic for many black-box\nsettings, but also gives the defender a much-needed advantage in responding to\nthe adversary.', 'Deep neural networks have been shown to be vulnerable to adversarial examples\ndeliberately constructed to misclassify victim models. As most adversarial\nexamples have restricted their perturbations to $L_{p}$-norm, existing defense\nmethods have focused on these types of perturbations and less attention has\nbeen paid to unrestricted adversarial examples; which can create more realistic\nattacks, able to deceive models without affecting human predictions. To address\nthis problem, the proposed adversarial attack generates an unrestricted\nadversarial example with a limited number of parameters. The attack selects\nthree points on the input image and based on their locations transforms the\nimage into an adversarial example. By limiting the range of movement and\nlocation of these three points and using a discriminatory network, the proposed\nunrestricted adversarial example preserves the image appearance. Experimental\nresults show that the proposed adversarial examples obtain an average success\nrate of 93.5% in terms of human evaluation on the MNIST and SVHN datasets. It\nalso reduces the model accuracy by an average of 73% on six datasets MNIST,\nFMNIST, SVHN, CIFAR10, CIFAR100, and ImageNet. It should be noted that, in the\ncase of attacks, lower accuracy in the victim model denotes a more successful\nattack. The adversarial train of the attack also improves model robustness\nagainst a randomly transformed image.', 'This paper discusses an Enhanced Model-Agnostic Meta-Learning (E-MAML)\nalgorithm that generates fast convergence of the policy function from a small\nnumber of training examples when applied to new learning tasks. Built on top of\nModel-Agnostic Meta-Learning (MAML), E-MAML maintains a set of policy\nparameters learned in the environment for previous tasks. We apply E-MAML to\ndeveloping reinforcement learning (RL)-based online fault tolerant control\nschemes for dynamic systems. The enhancement is applied when a new fault\noccurs, to re-initialize the parameters of a new RL policy that achieves faster\nadaption with a small number of samples of system behavior with the new fault.\nThis replaces the random task sampling step in MAML. Instead, it exploits the\nextant previously generated experiences of the controller. The enhancement is\nsampled to maximally span the parameter space to facilitate adaption to the new\nfault. We demonstrate the performance of our approach combining E-MAML with\nproximal policy optimization (PPO) on the well-known cart pole example, and\nthen on the fuel transfer system of an aircraft.', ""Meta-learning allows an intelligent agent to leverage prior learning episodes\nas a basis for quickly improving performance on a novel task. Bayesian\nhierarchical modeling provides a theoretical framework for formalizing\nmeta-learning as inference for a set of parameters that are shared across\ntasks. Here, we reformulate the model-agnostic meta-learning algorithm (MAML)\nof Finn et al. (2017) as a method for probabilistic inference in a hierarchical\nBayesian model. In contrast to prior methods for meta-learning via hierarchical\nBayes, MAML is naturally applicable to complex function approximators through\nits use of a scalable gradient descent procedure for posterior inference.\nFurthermore, the identification of MAML as hierarchical Bayes provides a way to\nunderstand the algorithm's operation as a meta-learning procedure, as well as\nan opportunity to make use of computational strategies for efficient inference.\nWe use this opportunity to propose an improvement to the MAML algorithm that\nmakes use of techniques from approximate inference and curvature estimation."", 'Model agnostic meta-learning (MAML) is a popular state-of-the-art\nmeta-learning algorithm that provides good weight initialization of a model\ngiven a variety of learning tasks. The model initialized by provided weight can\nbe fine-tuned to an unseen task despite only using a small amount of samples\nand within a few adaptation steps. MAML is simple and versatile but requires\ncostly learning rate tuning and careful design of the task distribution which\naffects its scalability and generalization. This paper proposes a more robust\nMAML based on an adaptive learning scheme and a prioritization task buffer(PTB)\nreferred to as Robust MAML (RMAML) for improving scalability of training\nprocess and alleviating the problem of distribution mismatch. RMAML uses\ngradient-based hyper-parameter optimization to automatically find the optimal\nlearning rate and uses the PTB to gradually adjust train-ing task distribution\ntoward testing task distribution over the course of training. Experimental\nresults on meta reinforcement learning environments demonstrate a substantial\nperformance gain as well as being less sensitive to hyper-parameter choice and\nrobust to distribution mismatch.', 'Advances in machine learning have led to broad deployment of systems with\nimpressive performance on important problems. Nonetheless, these systems can be\ninduced to make errors on data that are surprisingly similar to examples the\nlearned system handles correctly. The existence of these errors raises a\nvariety of questions about out-of-sample generalization and whether bad actors\nmight use such examples to abuse deployed systems. As a result of these\nsecurity concerns, there has been a flurry of recent papers proposing\nalgorithms to defend against such malicious perturbations of correctly handled\nexamples. It is unclear how such misclassifications represent a different kind\nof security problem than other errors, or even other attacker-produced examples\nthat have no specific relationship to an uncorrupted input. In this paper, we\nargue that adversarial example defense papers have, to date, mostly considered\nabstract, toy games that do not relate to any specific security concern.\nFurthermore, defense papers have not yet precisely described all the abilities\nand limitations of attackers that would be relevant in practical security.\nTowards this end, we establish a taxonomy of motivations, constraints, and\nabilities for more plausible adversaries. Finally, we provide a series of\nrecommendations outlining a path forward for future work to more clearly\narticulate the threat model and perform more meaningful evaluation.', 'We introduce ES-MAML, a new framework for solving the model agnostic meta\nlearning (MAML) problem based on Evolution Strategies (ES). Existing algorithms\nfor MAML are based on policy gradients, and incur significant difficulties when\nattempting to estimate second derivatives using backpropagation on stochastic\npolicies. We show how ES can be applied to MAML to obtain an algorithm which\navoids the problem of estimating second derivatives, and is also conceptually\nsimple and easy to implement. Moreover, ES-MAML can handle new types of\nnonsmooth adaptation operators, and other techniques for improving performance\nand estimation of ES methods become applicable. We show empirically that\nES-MAML is competitive with existing methods and often yields better adaptation\nwith fewer queries.', 'The paper studies the complexity of the optimization problem behind the\nModel-Agnostic Meta-Learning (MAML) algorithm. The goal of the study is to\ndetermine the global convergence of MAML on sequential decision-making tasks\npossessing a common structure. We are curious to know when, if at all, the\nbenign landscape of the underlying tasks results in a benign landscape of the\ncorresponding MAML objective. For illustration, we analyze the landscape of the\nMAML objective on LQR tasks to determine what types of similarities in their\nstructures enable the algorithm to converge to the globally optimal solution.', 'A core capability of intelligent systems is the ability to quickly learn new\ntasks by drawing on prior experience. Gradient (or optimization) based\nmeta-learning has recently emerged as an effective approach for few-shot\nlearning. In this formulation, meta-parameters are learned in the outer loop,\nwhile task-specific models are learned in the inner-loop, by using only a small\namount of data from the current task. A key challenge in scaling these\napproaches is the need to differentiate through the inner loop learning\nprocess, which can impose considerable computational and memory burdens. By\ndrawing upon implicit differentiation, we develop the implicit MAML algorithm,\nwhich depends only on the solution to the inner level optimization and not the\npath taken by the inner loop optimizer. This effectively decouples the\nmeta-gradient computation from the choice of inner loop optimizer. As a result,\nour approach is agnostic to the choice of inner loop optimizer and can\ngracefully handle many gradient steps without vanishing gradients or memory\nconstraints. Theoretically, we prove that implicit MAML can compute accurate\nmeta-gradients with a memory footprint that is, up to small constant factors,\nno more than that which is required to compute a single inner loop gradient and\nat no overall increase in the total computational cost. Experimentally, we show\nthat these benefits of implicit MAML translate into empirical gains on few-shot\nimage recognition benchmarks.', ""Deep learning models for graphs have achieved strong performance for the task\nof node classification. Despite their proliferation, currently there is no\nstudy of their robustness to adversarial attacks. Yet, in domains where they\nare likely to be used, e.g. the web, adversaries are common. Can deep learning\nmodels for graphs be easily fooled? In this work, we introduce the first study\nof adversarial attacks on attributed graphs, specifically focusing on models\nexploiting ideas of graph convolutions. In addition to attacks at test time, we\ntackle the more challenging class of poisoning/causative attacks, which focus\non the training phase of a machine learning model. We generate adversarial\nperturbations targeting the node's features and the graph structure, thus,\ntaking the dependencies between instances in account. Moreover, we ensure that\nthe perturbations remain unnoticeable by preserving important data\ncharacteristics. To cope with the underlying discrete domain we propose an\nefficient algorithm Nettack exploiting incremental computations. Our\nexperimental study shows that accuracy of node classification significantly\ndrops even when performing only few perturbations. Even more, our attacks are\ntransferable: the learned attacks generalize to other state-of-the-art node\nclassification models and unsupervised approaches, and likewise are successful\neven when only limited knowledge about the graph is given."", 'Deep learning models are known to be vulnerable to adversarial examples. A\npractical adversarial attack should require as little as possible knowledge of\nattacked models. Current substitute attacks need pre-trained models to generate\nadversarial examples and their attack success rates heavily rely on the\ntransferability of adversarial examples. Current score-based and decision-based\nattacks require lots of queries for the attacked models. In this study, we\npropose a novel adversarial imitation attack. First, it produces a replica of\nthe attacked model by a two-player game like the generative adversarial\nnetworks (GANs). The objective of the generative model is to generate examples\nthat lead the imitation model returning different outputs with the attacked\nmodel. The objective of the imitation model is to output the same labels with\nthe attacked model under the same inputs. Then, the adversarial examples\ngenerated by the imitation model are utilized to fool the attacked model.\nCompared with the current substitute attacks, imitation attacks can use less\ntraining data to produce a replica of the attacked model and improve the\ntransferability of adversarial examples. Experiments demonstrate that our\nimitation attack requires less training data than the black-box substitute\nattacks, but achieves an attack success rate close to the white-box attack on\nunseen data with no query.', 'The field of few-shot learning has recently seen substantial advancements.\nMost of these advancements came from casting few-shot learning as a\nmeta-learning problem. Model Agnostic Meta Learning or MAML is currently one of\nthe best approaches for few-shot learning via meta-learning. MAML is simple,\nelegant and very powerful, however, it has a variety of issues, such as being\nvery sensitive to neural network architectures, often leading to instability\nduring training, requiring arduous hyperparameter searches to stabilize\ntraining and achieve high generalization and being very computationally\nexpensive at both training and inference times. In this paper, we propose\nvarious modifications to MAML that not only stabilize the system, but also\nsubstantially improve the generalization performance, convergence speed and\ncomputational overhead of MAML, which we call MAML++.', ""In this work we present a formal theoretical framework for assessing and\nanalyzing two classes of malevolent action towards generic Artificial\nIntelligence (AI) systems. Our results apply to general multi-class classifiers\nthat map from an input space into a decision space, including artificial neural\nnetworks used in deep learning applications. Two classes of attacks are\nconsidered. The first class involves adversarial examples and concerns the\nintroduction of small perturbations of the input data that cause\nmisclassification. The second class, introduced here for the first time and\nnamed stealth attacks, involves small perturbations to the AI system itself.\nHere the perturbed system produces whatever output is desired by the attacker\non a specific small data set, perhaps even a single input, but performs as\nnormal on a validation set (which is unknown to the attacker). We show that in\nboth cases, i.e., in the case of an attack based on adversarial examples and in\nthe case of a stealth attack, the dimensionality of the AI's decision-making\nspace is a major contributor to the AI's susceptibility. For attacks based on\nadversarial examples, a second crucial parameter is the absence of local\nconcentrations in the data probability distribution, a property known as\nSmeared Absolute Continuity. According to our findings, robustness to\nadversarial examples requires either (a) the data distributions in the AI's\nfeature space to have concentrated probability density functions or (b) the\ndimensionality of the AI's decision variables to be sufficiently small. We also\nshow how to construct stealth attacks on high-dimensional AI systems that are\nhard to spot unless the validation set is made exponentially large."", 'Model Agnostic Meta Learning (MAML) is widely used to find a good\ninitialization for a family of tasks. Despite its success, a critical challenge\nin MAML is to calculate the gradient w.r.t. the initialization of a long\ntraining trajectory for the sampled tasks, because the computation graph can\nrapidly explode and the computational cost is very expensive. To address this\nproblem, we propose Adjoint MAML (A-MAML). We view gradient descent in the\ninner optimization as the evolution of an Ordinary Differential Equation (ODE).\nTo efficiently compute the gradient of the validation loss w.r.t. the\ninitialization, we use the adjoint method to construct a companion, backward\nODE. To obtain the gradient w.r.t. the initialization, we only need to run the\nstandard ODE solver twice -- one is forward in time that evolves a long\ntrajectory of gradient flow for the sampled task; the other is backward and\nsolves the adjoint ODE. We need not create or expand any intermediate\ncomputational graphs, adopt aggressive approximations, or impose proximal\nregularizers in the training loss. Our approach is cheap, accurate, and\nadaptable to different trajectory lengths. We demonstrate the advantage of our\napproach in both synthetic and real-world meta-learning tasks.', 'Deep neural networks are vulnerable to adversarial examples, which poses\nsecurity concerns on these algorithms due to the potentially severe\nconsequences. Adversarial attacks serve as an important surrogate to evaluate\nthe robustness of deep learning models before they are deployed. However, most\nof existing adversarial attacks can only fool a black-box model with a low\nsuccess rate. To address this issue, we propose a broad class of momentum-based\niterative algorithms to boost adversarial attacks. By integrating the momentum\nterm into the iterative process for attacks, our methods can stabilize update\ndirections and escape from poor local maxima during the iterations, resulting\nin more transferable adversarial examples. To further improve the success rates\nfor black-box attacks, we apply momentum iterative algorithms to an ensemble of\nmodels, and show that the adversarially trained models with a strong defense\nability are also vulnerable to our black-box attacks. We hope that the proposed\nmethods will serve as a benchmark for evaluating the robustness of various deep\nmodels and defense methods. With this method, we won the first places in NIPS\n2017 Non-targeted Adversarial Attack and Targeted Adversarial Attack\ncompetitions.', ""An important research direction in machine learning has centered around\ndeveloping meta-learning algorithms to tackle few-shot learning. An especially\nsuccessful algorithm has been Model Agnostic Meta-Learning (MAML), a method\nthat consists of two optimization loops, with the outer loop finding a\nmeta-initialization, from which the inner loop can efficiently learn new tasks.\nDespite MAML's popularity, a fundamental open question remains -- is the\neffectiveness of MAML due to the meta-initialization being primed for rapid\nlearning (large, efficient changes in the representations) or due to feature\nreuse, with the meta initialization already containing high quality features?\nWe investigate this question, via ablation studies and analysis of the latent\nrepresentations, finding that feature reuse is the dominant factor. This leads\nto the ANIL (Almost No Inner Loop) algorithm, a simplification of MAML where we\nremove the inner loop for all but the (task-specific) head of a MAML-trained\nnetwork. ANIL matches MAML's performance on benchmark few-shot image\nclassification and RL and offers computational improvements over MAML. We\nfurther study the precise contributions of the head and body of the network,\nshowing that performance on the test tasks is entirely determined by the\nquality of the learned features, and we can remove even the head of the network\n(the NIL algorithm). We conclude with a discussion of the rapid learning vs\nfeature reuse question for meta-learning algorithms more broadly."", ""Efficiently adapting to new environments and changes in dynamics is critical\nfor agents to successfully operate in the real world. Reinforcement learning\n(RL) based approaches typically rely on external reward feedback for\nadaptation. However, in many scenarios this reward signal might not be readily\navailable for the target task, or the difference between the environments can\nbe implicit and only observable from the dynamics. To this end, we introduce a\nmethod that allows for self-adaptation of learned policies: No-Reward Meta\nLearning (NoRML). NoRML extends Model Agnostic Meta Learning (MAML) for RL and\nuses observable dynamics of the environment instead of an explicit reward\nfunction in MAML's finetune step. Our method has a more expressive update step\nthan MAML, while maintaining MAML's gradient based foundation. Additionally, in\norder to allow more targeted exploration, we implement an extension to MAML\nthat effectively disconnects the meta-policy parameters from the fine-tuned\npolicies' parameters. We first study our method on a number of synthetic\ncontrol problems and then validate our method on common benchmark environments,\nshowing that NoRML outperforms MAML when the dynamics change between tasks."", 'Multi-label learning deals with the problem that each instance is associated\nwith multiple labels simultaneously. Most of the existing approaches aim to\nimprove the performance of multi-label learning by exploiting label\ncorrelations. Although the data augmentation technique is widely used in many\nmachine learning tasks, it is still unclear whether data augmentation is\nhelpful to multi-label learning. In this article, we propose to leverage the\ndata augmentation technique to improve the performance of multi-label learning.\nSpecifically, we first propose a novel data augmentation approach that performs\nclustering on the real examples and treats the cluster centers as virtual\nexamples, and these virtual examples naturally embody the local label\ncorrelations and label importances. Then, motivated by the cluster assumption\nthat examples in the same cluster should have the same label, we propose a\nnovel regularization term to bridge the gap between the real examples and\nvirtual examples, which can promote the local smoothness of the learning\nfunction. Extensive experimental results on a number of real-world multi-label\ndatasets clearly demonstrate that our proposed approach outperforms the\nstate-of-the-art counterparts.', ""In Machine Learning as a Service, a provider trains a deep neural network and\ngives many users access. The hosted (source) model is susceptible to model\nstealing attacks, where an adversary derives a surrogate model from API access\nto the source model. For post hoc detection of such attacks, the provider needs\na robust method to determine whether a suspect model is a surrogate of their\nmodel. We propose a fingerprinting method for deep neural network classifiers\nthat extracts a set of inputs from the source model so that only surrogates\nagree with the source model on the classification of such inputs. These inputs\nare a subclass of transferable adversarial examples which we call conferrable\nadversarial examples that exclusively transfer with a target label from a\nsource model to its surrogates. We propose a new method to generate these\nconferrable adversarial examples. We present an extensive study on the\nirremovability of our fingerprint against fine-tuning, weight pruning,\nretraining, retraining with different architectures, three model extraction\nattacks from related work, transfer learning, adversarial training, and two new\nadaptive attacks. Our fingerprint is robust against distillation, related model\nextraction attacks, and even transfer learning when the attacker has no access\nto the model provider's dataset. Our fingerprint is the first method that\nreaches a ROC AUC of 1.0 in verifying surrogates, compared to a ROC AUC of 0.63\nby previous fingerprints."", 'Model-agnostic meta-learning (MAML) is arguably one of the most popular\nmeta-learning algorithms nowadays. Nevertheless, its performance on few-shot\nclassification is far behind many recent algorithms dedicated to the problem.\nIn this paper, we point out several key facets of how to train MAML to excel in\nfew-shot classification. First, we find that MAML needs a large number of\ngradient steps in its inner loop update, which contradicts its common usage in\nfew-shot classification. Second, we find that MAML is sensitive to the class\nlabel assignments during meta-testing. Concretely, MAML meta-trains the\ninitialization of an $N$-way classifier. These $N$ ways, during meta-testing,\nthen have ""$N!$"" different permutations to be paired with a few-shot task of\n$N$ novel classes. We find that these permutations lead to a huge variance of\naccuracy, making MAML unstable in few-shot classification. Third, we\ninvestigate several approaches to make MAML permutation-invariant, among which\nmeta-training a single vector to initialize all the $N$ weight vectors in the\nclassification head performs the best. On benchmark datasets like MiniImageNet\nand TieredImageNet, our approach, which we name UNICORN-MAML, performs on a par\nwith or even outperforms many recent few-shot classification algorithms,\nwithout sacrificing MAML\'s simplicity.', 'We propose a new computationally-efficient first-order algorithm for\nModel-Agnostic Meta-Learning (MAML). The key enabling technique is to interpret\nMAML as a bilevel optimization (BLO) problem and leverage the sign-based\nSGD(signSGD) as a lower-level optimizer of BLO. We show that MAML, through the\nlens of signSGD-oriented BLO, naturally yields an alternating optimization\nscheme that just requires first-order gradients of a learned meta-model. We\nterm the resulting MAML algorithm Sign-MAML. Compared to the conventional\nfirst-order MAML (FO-MAML) algorithm, Sign-MAML is theoretically-grounded as it\ndoes not impose any assumption on the absence of second-order derivatives\nduring meta training. In practice, we show that Sign-MAML outperforms FO-MAML\nin various few-shot image classification tasks, and compared to MAML, it\nachieves a much more graceful tradeoff between classification accuracy and\ncomputation efficiency.', 'It has been empirically observed that defense mechanisms designed to protect\nneural networks against $\\ell_\\infty$ adversarial examples offer poor\nperformance against $\\ell_2$ adversarial examples and vice versa. In this paper\nwe conduct a geometrical analysis that validates this observation. Then, we\nprovide a number of empirical insights to illustrate the effect of this\nphenomenon in practice. Then, we review some of the existing defense mechanism\nthat attempts to defend against multiple attacks by mixing defense strategies.\nThanks to our numerical experiments, we discuss the relevance of this method\nand state open questions for the adversarial examples community.', 'As an essential tool in security, the intrusion detection system bears the\nresponsibility of the defense to network attacks performed by malicious\ntraffic. Nowadays, with the help of machine learning algorithms, intrusion\ndetection systems develop rapidly. However, the robustness of this system is\nquestionable when it faces adversarial attacks. For the robustness of detection\nsystems, more potential attack approaches are under research. In this paper, a\nframework of the generative adversarial networks, called IDSGAN, is proposed to\ngenerate the adversarial malicious traffic records aiming to attack intrusion\ndetection systems by deceiving and evading the detection. Given that the\ninternal structure and parameters of the detection system are unknown to\nattackers, the adversarial attack examples perform the black-box attacks\nagainst the detection system. IDSGAN leverages a generator to transform\noriginal malicious traffic records into adversarial malicious ones. A\ndiscriminator classifies traffic examples and dynamically learns the real-time\nblack-box detection system. More significantly, the restricted modification\nmechanism is designed for the adversarial generation to preserve original\nattack functionalities of adversarial traffic records. The effectiveness of the\nmodel is indicated by attacking multiple algorithm-based detection models with\ndifferent attack categories. The robustness is verified by changing the number\nof the modified features. A comparative experiment with adversarial attack\nbaselines demonstrates the superiority of our model.', 'The continual learning problem involves training models with limited capacity\nto perform well on a set of an unknown number of sequentially arriving tasks.\nWhile meta-learning shows great potential for reducing interference between old\nand new tasks, the current training procedures tend to be either slow or\noffline, and sensitive to many hyper-parameters. In this work, we propose\nLook-ahead MAML (La-MAML), a fast optimisation-based meta-learning algorithm\nfor online-continual learning, aided by a small episodic memory. Our proposed\nmodulation of per-parameter learning rates in our meta-learning update allows\nus to draw connections to prior work on hypergradients and meta-descent. This\nprovides a more flexible and efficient way to mitigate catastrophic forgetting\ncompared to conventional prior-based methods. La-MAML achieves performance\nsuperior to other replay-based, prior-based and meta-learning based approaches\nfor continual learning on real-world visual classification benchmarks. Source\ncode can be found here: https://github.com/montrealrobotics/La-MAML', 'Recently, model-agnostic meta-learning (MAML) has garnered tremendous\nattention. However, the stochastic optimization of MAML is still immature.\nExisting algorithms for MAML exploit the ""episode"" idea by sampling some tasks\nand some data points for each sampled task at each iteration for updating the\nmeta-model. However, they either do not necessarily guarantee convergence with\na constant mini-batch size or require processing a large number of tasks at\nevery iteration, which is not viable for continual learning or cross-device\nfederated learning where only a small number of tasks are available\nper-iteration or per-round. This paper addresses these issues by (i) proposing\nefficient memory-based stochastic algorithms for MAML with the vanishing\nconvergence error, which only requires sampling a constant number of tasks and\na constant number of data samples per-iteration; (ii) proposing\ncommunication-efficient distributed memory-based MAML algorithms for\npersonalized federated learning in both the cross-device (with client sampling)\nand the cross-silo (without client sampling) settings. The theoretical results\nsignificantly improve the optimization theory for MAML and the empirical\nresults also corroborate the theory.', 'This paper considers meta-learning problems, where there is a distribution of\ntasks, and we would like to obtain an agent that performs well (i.e., learns\nquickly) when presented with a previously unseen task sampled from this\ndistribution. We analyze a family of algorithms for learning a parameter\ninitialization that can be fine-tuned quickly on a new task, using only\nfirst-order derivatives for the meta-learning updates. This family includes and\ngeneralizes first-order MAML, an approximation to MAML obtained by ignoring\nsecond-order derivatives. It also includes Reptile, a new algorithm that we\nintroduce here, which works by repeatedly sampling a task, training on it, and\nmoving the initialization towards the trained weights on that task. We expand\non the results from Finn et al. showing that first-order meta-learning\nalgorithms perform well on some well-established benchmarks for few-shot\nclassification, and we provide theoretical analysis aimed at understanding why\nthese algorithms work.', 'Deep neural networks are vulnerable to adversarial attacks. White-box\nadversarial attacks can fool neural networks with small adversarial\nperturbations, especially for large size images. However, keeping successful\nadversarial perturbations imperceptible is especially challenging for\ntransfer-based black-box adversarial attacks. Often such adversarial examples\ncan be easily spotted due to their unpleasantly poor visual qualities, which\ncompromises the threat of adversarial attacks in practice. In this study, to\nimprove the image quality of black-box adversarial examples perceptually, we\npropose structure-aware adversarial attacks by generating adversarial images\nbased on psychological perceptual models. Specifically, we allow higher\nperturbations on perceptually insignificant regions, while assigning lower or\nno perturbation on visually sensitive regions. In addition to the proposed\nspatial-constrained adversarial perturbations, we also propose a novel\nstructure-aware frequency adversarial attack method in the discrete cosine\ntransform (DCT) domain. Since the proposed attacks are independent of the\ngradient estimation, they can be directly incorporated with existing\ngradient-based attacks. Experimental results show that, with the comparable\nattack success rate (ASR), the proposed methods can produce adversarial\nexamples with considerably improved visual quality for free. With the\ncomparable perceptual quality, the proposed approaches achieve higher attack\nsuccess rates: particularly for the frequency structure-aware attacks, the\naverage ASR improves more than 10% over the baseline attacks.', 'In this paper, we establish the ordinary differential equation (ODE) that\nunderlies the training dynamics of Model-Agnostic Meta-Learning (MAML). Our\ncontinuous-time limit view of the process eliminates the influence of the\nmanually chosen step size of gradient descent and includes the existing\ngradient descent training algorithm as a special case that results from a\nspecific discretization. We show that the MAML ODE enjoys a linear convergence\nrate to an approximate stationary point of the MAML loss function for strongly\nconvex task losses, even when the corresponding MAML loss is non-convex.\nMoreover, through the analysis of the MAML ODE, we propose a new BI-MAML\ntraining algorithm that significantly reduces the computational burden\nassociated with existing MAML training methods. To complement our theoretical\nfindings, we perform empirical experiments to showcase the superiority of our\nproposed methods with respect to the existing work.', ""In past years model-agnostic meta-learning (MAML) has been one of the most\npromising approaches in meta-learning. It can be applied to different kinds of\nproblems, e.g., reinforcement learning, but also shows good results on few-shot\nlearning tasks. Besides their tremendous success in these tasks, it has still\nnot been fully revealed yet, why it works so well. Recent work proposes that\nMAML rather reuses features than rapidly learns. In this paper, we want to\ninspire a deeper understanding of this question by analyzing MAML's\nrepresentation. We apply representation similarity analysis (RSA), a\nwell-established method in neuroscience, to the few-shot learning instantiation\nof MAML. Although some part of our analysis supports their general results that\nfeature reuse is predominant, we also reveal arguments against their\nconclusion. The similarity-increase of layers closer to the input layers arises\nfrom the learning task itself and not from the model. In addition, the\nrepresentations after inner gradient steps make a broader change to the\nrepresentation than the changes during meta-training."", 'Model-agnostic meta-learning (MAML) is one of the most popular and widely\nadopted meta-learning algorithms, achieving remarkable success in various\nlearning problems. Yet, with the unique design of nested inner-loop and\nouter-loop updates, which govern the task-specific and meta-model-centric\nlearning, respectively, the underlying learning objective of MAML remains\nimplicit and thus impedes a more straightforward understanding of it. In this\npaper, we provide a new perspective of the working mechanism of MAML. We\ndiscover that MAML is analogous to a meta-learner using a supervised\ncontrastive objective. The query features are pulled towards the support\nfeatures of the same class and against those of different classes. Such\ncontrastiveness is experimentally verified via an analysis based on the cosine\nsimilarity. Moreover, we reveal that vanilla MAML has an undesirable\ninterference term originating from the random initialization and the cross-task\ninteraction. We thus propose a simple but effective technique, zeroing trick,\nto alleviate the interference. Extensive experiments are conducted on both\nmini-ImageNet and Omniglot datasets to demonstrate the consistent improvement\nbrought by our proposed method, validating its effectiveness.', 'We consider the tracking problem as a special type of object detection\nproblem, which we call instance detection. With proper initialization, a\ndetector can be quickly converted into a tracker by learning the new instance\nfrom a single image. We find that model-agnostic meta-learning (MAML) offers a\nstrategy to initialize the detector that satisfies our needs. We propose a\nprincipled three-step approach to build a high-performance tracker. First, pick\nany modern object detector trained with gradient descent. Second, conduct\noffline training (or initialization) with MAML. Third, perform domain\nadaptation using the initial frame. We follow this procedure to build two\ntrackers, named Retina-MAML and FCOS-MAML, based on two modern detectors\nRetinaNet and FCOS. Evaluations on four benchmarks show that both trackers are\ncompetitive against state-of-the-art trackers. On OTB-100, Retina-MAML achieves\nthe highest ever AUC of 0.712. On TrackingNet, FCOS-MAML ranks the first on the\nleader board with an AUC of 0.757 and the normalized precision of 0.822. Both\ntrackers run in real-time at 40 FPS.', 'We study the convergence of a class of gradient-based Model-Agnostic\nMeta-Learning (MAML) methods and characterize their overall complexity as well\nas their best achievable accuracy in terms of gradient norm for nonconvex loss\nfunctions. We start with the MAML method and its first-order approximation\n(FO-MAML) and highlight the challenges that emerge in their analysis. By\novercoming these challenges not only we provide the first theoretical\nguarantees for MAML and FO-MAML in nonconvex settings, but also we answer some\nof the unanswered questions for the implementation of these algorithms\nincluding how to choose their learning rate and the batch size for both tasks\nand datasets corresponding to tasks. In particular, we show that MAML can find\nan $\\epsilon$-first-order stationary point ($\\epsilon$-FOSP) for any positive\n$\\epsilon$ after at most $\\mathcal{O}(1/\\epsilon^2)$ iterations at the expense\nof requiring second-order information. We also show that FO-MAML which ignores\nthe second-order information required in the update of MAML cannot achieve any\nsmall desired level of accuracy, i.e., FO-MAML cannot find an $\\epsilon$-FOSP\nfor any $\\epsilon>0$. We further propose a new variant of the MAML algorithm\ncalled Hessian-free MAML which preserves all theoretical guarantees of MAML,\nwithout requiring access to second-order information.', 'Deep learning models require a large amount of data to perform well. When\ndata is scarce for a target task, we can transfer the knowledge gained by\ntraining on similar tasks to quickly learn the target. A successful approach is\nmeta-learning, or ""learning to learn"" a distribution of tasks, where ""learning""\nis represented by an outer loop, and ""to learn"" by an inner loop of gradient\ndescent. However, a number of recent empirical studies argue that the inner\nloop is unnecessary and more simple models work equally well or even better. We\nstudy the performance of MAML as a function of the learning rate of the inner\nloop, where zero learning rate implies that there is no inner loop. Using\nrandom matrix theory and exact solutions of linear models, we calculate an\nalgebraic expression for the test loss of MAML applied to mixed linear\nregression and nonlinear regression with overparameterized models.\nSurprisingly, while the optimal learning rate for adaptation is positive, we\nfind that the optimal learning rate for training is always negative, a setting\nthat has never been considered before. Therefore, not only does the performance\nincrease by decreasing the learning rate to zero, as suggested by recent work,\nbut it can be increased even further by decreasing the learning rate to\nnegative values. These results help clarify under what circumstances\nmeta-learning performs best.', ""We formulate a principle for classification with the knowledge of the\nmarginal distribution over the data points (unlabeled data). The principle is\ncast in terms of Tikhonov style regularization where the regularization penalty\narticulates the way in which the marginal density should constrain otherwise\nunrestricted conditional distributions. Specifically, the regularization\npenalty penalizes any information introduced between the examples and labels\nbeyond what is provided by the available labeled examples. The work extends\nSzummer and Jaakkola's information regularization (NIPS 2002) to multiple\ndimensions, providing a regularizer independent of the covering of the space\nused in the derivation. We show in addition how the information regularizer can\nbe used as a measure of complexity of the classification task with unlabeled\ndata and prove a relevant sample-complexity bound. We illustrate the\nregularization principle in practice by restricting the class of conditional\ndistributions to be logistic regression models and constructing the\nregularization penalty from a finite set of unlabeled examples."", 'Machine learning models in health care are often deployed in settings where\nit is important to protect patient privacy. In such settings, methods for\ndifferentially private (DP) learning provide a general-purpose approach to\nlearn models with privacy guarantees. Modern methods for DP learning ensure\nprivacy through mechanisms that censor information judged as too unique. The\nresulting privacy-preserving models, therefore, neglect information from the\ntails of a data distribution, resulting in a loss of accuracy that can\ndisproportionately affect small groups. In this paper, we study the effects of\nDP learning in health care. We use state-of-the-art methods for DP learning to\ntrain privacy-preserving models in clinical prediction tasks, including x-ray\nclassification of images and mortality prediction in time series data. We use\nthese models to perform a comprehensive empirical investigation of the\ntradeoffs between privacy, utility, robustness to dataset shift, and fairness.\nOur results highlight lesser-known limitations of methods for DP learning in\nhealth care, models that exhibit steep tradeoffs between privacy and utility,\nand models whose predictions are disproportionately influenced by large\ndemographic groups in the training data. We discuss the costs and benefits of\ndifferentially private learning in health care.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : My goal is to develop a learning model that can handle multiple tasks simultaneously. This proposed learning model will function as a sub-model selector, meaning that when presented with a new task, it will determine the most suitable sub-model to learn this task. Additionally, my learning model could also operate as a sub-model constructor. If it determines that no existing sub-model can learn the task, it will generate innovative model architectures to construct a new sub-model suitable for the task. At present, I am considering the application of a combination of reinforcement learning and model architecture search algorithms to achieve this. To effectively evaluate the performance, I require a comprehensive benchmark that includes a variety of datasets that can be used as different sub-tasks. My aim is for my learning model to achieve state-of-the-art performance on this benchmark. In general, I am open to any methodologies that enhance the performance and adaptability of my learning model across different problem domains.","['Multi-task learning, as it is understood nowadays, consists of using one\nsingle model to carry out several similar tasks. From classifying hand-written\ncharacters of different alphabets to figuring out how to play several Atari\ngames using reinforcement learning, multi-task models have been able to widen\ntheir performance range across different tasks, although these tasks are\nusually of a similar nature. In this work, we attempt to widen this range even\nfurther, by including heterogeneous tasks in a single learning procedure. To do\nso, we firstly formally define a multi-network model, identifying the necessary\ncomponents and characteristics to allow different adaptations of said model\ndepending on the tasks it is required to fulfill. Secondly, employing the\nformal definition as a starting point, we develop an illustrative model example\nconsisting of three different tasks (classification, regression and data\nsampling). The performance of this model implementation is then analyzed,\nshowing its capabilities. Motivated by the results of the analysis, we\nenumerate a set of open challenges and future research lines over which the\nfull potential of the proposed model definition can be exploited.', 'The usefulness and value of Multi-step Machine Learning (ML), where a task is\norganized into connected sub-tasks with known intermediate inference goals, as\nopposed to a single large model learned end-to-end without intermediate\nsub-tasks, is presented. Pre-optimized ML models are connected and better\nperformance is obtained by re-optimizing the connected one. The selection of an\nML model from several small ML model candidates for each sub-task has been\nperformed by using the idea based on Neural Architecture Search (NAS). In this\npaper, Differentiable Architecture Search (DARTS) and Single Path One-Shot NAS\n(SPOS-NAS) are tested, where the construction of loss functions is improved to\nkeep all ML models smoothly learning. Using DARTS and SPOS-NAS as an\noptimization and selection as well as the connections for multi-step machine\nlearning systems, we find that (1) such a system can quickly and successfully\nselect highly performant model combinations, and (2) the selected models are\nconsistent with baseline algorithms, such as grid search, and their outputs are\nwell controlled.', 'For most deep learning practitioners, sequence modeling is synonymous with\nrecurrent networks. Yet recent results indicate that convolutional\narchitectures can outperform recurrent networks on tasks such as audio\nsynthesis and machine translation. Given a new sequence modeling task or\ndataset, which architecture should one use? We conduct a systematic evaluation\nof generic convolutional and recurrent architectures for sequence modeling. The\nmodels are evaluated across a broad range of standard tasks that are commonly\nused to benchmark recurrent networks. Our results indicate that a simple\nconvolutional architecture outperforms canonical recurrent networks such as\nLSTMs across a diverse range of tasks and datasets, while demonstrating longer\neffective memory. We conclude that the common association between sequence\nmodeling and recurrent networks should be reconsidered, and convolutional\nnetworks should be regarded as a natural starting point for sequence modeling\ntasks. To assist related work, we have made code available at\nhttp://github.com/locuslab/TCN .', 'Deep learning models require extensive architecture design exploration and\nhyperparameter optimization to perform well on a given task. The exploration of\nthe model design space is often made by a human expert, and optimized using a\ncombination of grid search and search heuristics over a large space of possible\nchoices. Neural Architecture Search (NAS) is a Reinforcement Learning approach\nthat has been proposed to automate architecture design. NAS has been\nsuccessfully applied to generate Neural Networks that rival the best\nhuman-designed architectures. However, NAS requires sampling, constructing, and\ntraining hundreds to thousands of models to achieve well-performing\narchitectures. This procedure needs to be executed from scratch for each new\ntask. The application of NAS to a wide set of tasks currently lacks a way to\ntransfer generalizable knowledge across tasks. In this paper, we present the\nMultitask Neural Model Search (MNMS) controller. Our goal is to learn a\ngeneralizable framework that can condition model construction on successful\nmodel searches for previously seen tasks, thus significantly speeding up the\nsearch for new tasks. We demonstrate that MNMS can conduct an automated\narchitecture search for multiple tasks simultaneously while still learning\nwell-performing, specialized models for each task. We then show that\npre-trained MNMS controllers can transfer learning to new tasks. By leveraging\nknowledge from previous searches, we find that pre-trained MNMS models start\nfrom a better location in the search space and reduce search time on unseen\ntasks, while still discovering models that outperform published human-designed\nmodels.', 'Approaches to continual learning aim to successfully learn a set of related\ntasks that arrive in an online manner. Recently, several frameworks have been\ndeveloped which enable deep learning to be deployed in this learning scenario.\nA key modelling decision is to what extent the architecture should be shared\nacross tasks. On the one hand, separately modelling each task avoids\ncatastrophic forgetting but it does not support transfer learning and leads to\nlarge models. On the other hand, rigidly specifying a shared component and a\ntask-specific part enables task transfer and limits the model size, but it is\nvulnerable to catastrophic forgetting and restricts the form of task-transfer\nthat can occur. Ideally, the network should adaptively identify which parts of\nthe network to share in a data driven way. Here we introduce such an approach\ncalled Continual Learning with Adaptive Weights (CLAW), which is based on\nprobabilistic modelling and variational inference. Experiments show that CLAW\nachieves state-of-the-art performance on six benchmarks in terms of overall\ncontinual learning performance, as measured by classification accuracy, and in\nterms of addressing catastrophic forgetting.', 'As deep learning applications continue to become more diverse, an interesting\nquestion arises: Can general problem solving arise from jointly learning\nseveral such diverse tasks? To approach this question, deep multi-task learning\nis extended in this paper to the setting where there is no obvious overlap\nbetween task architectures. The idea is that any set of (architecture,task)\npairs can be decomposed into a set of potentially related subproblems, whose\nsharing is optimized by an efficient stochastic algorithm. The approach is\nfirst validated in a classic synthetic multi-task learning benchmark, and then\napplied to sharing across disparate architectures for vision, NLP, and genomics\ntasks. It discovers regularities across these domains, encodes them into\nsharable modules, and combines these modules systematically to improve\nperformance in the individual tasks. The results confirm that sharing learned\nfunctionality across diverse domains and architectures is indeed beneficial,\nthus establishing a key ingredient for general problem solving in the future.', 'Lifelong learning remains an open problem. One of its main difficulties is\ncatastrophic forgetting. Many dynamic expansion approaches have been proposed\nto address this problem, but they all use homogeneous models of predefined\nstructure for all tasks. The common original model and expansion structures\nignore the requirement of different model structures on different tasks, which\nleads to a less compact model for multiple tasks and causes the model size to\nincrease rapidly as the number of tasks increases. Moreover, they can not\nperform best on all tasks. To solve those problems, in this paper, we propose a\nnew lifelong learning framework named Searchable Extension Units (SEU) by\nintroducing Neural Architecture Search into lifelong learning, which breaks\ndown the need for a predefined original model and searches for specific\nextension units for different tasks, without compromising the performance of\nthe model on different tasks. Our approach can obtain a much more compact model\nwithout catastrophic forgetting. The experimental results on the PMNIST, the\nsplit CIFAR10 dataset, the split CIFAR100 dataset, and the Mixture dataset\nempirically prove that our method can achieve higher accuracy with much smaller\nmodel, whose size is about 25-33 percentage of that of the state-of-the-art\nmethods.', 'Can deep learning solve multiple tasks simultaneously, even when they are\nunrelated and very different? We investigate how the representations of the\nunderlying tasks affect the ability of a single neural network to learn them\njointly. We present theoretical and empirical findings that a single neural\nnetwork is capable of simultaneously learning multiple tasks from a combined\ndata set, for a variety of methods for representing tasks -- for example, when\nthe distinct tasks are encoded by well-separated clusters or decision trees\nover certain task-code attributes. More concretely, we present a novel analysis\nthat shows that families of simple programming-like constructs for the codes\nencoding the tasks are learnable by two-layer neural networks with standard\ntraining. We study more generally how the complexity of learning such combined\ntasks grows with the complexity of the task codes; we find that combining many\ntasks may incur a sample complexity penalty, even though the individual tasks\nare easy to learn. We provide empirical support for the usefulness of the\nlearning bounds by training networks on clusters, decision trees, and SQL-style\naggregation.', 'Continual learning with neural networks is an important learning framework in\nAI that aims to learn a sequence of tasks well. However, it is often confronted\nwith three challenges: (1) overcome the catastrophic forgetting problem, (2)\nadapt the current network to new tasks, and meanwhile (3) control its model\ncomplexity. To reach these goals, we propose a novel approach named as\nContinual Learning with Efficient Architecture Search, or CLEAS in short. CLEAS\nworks closely with neural architecture search (NAS) which leverages\nreinforcement learning techniques to search for the best neural architecture\nthat fits a new task. In particular, we design a neuron-level NAS controller\nthat decides which old neurons from previous tasks should be reused (knowledge\ntransfer), and which new neurons should be added (to learn new knowledge). Such\na fine-grained controller allows one to find a very concise architecture that\ncan fit each new task well. Meanwhile, since we do not alter the weights of the\nreused neurons, we perfectly memorize the knowledge learned from previous\ntasks. We evaluate CLEAS on numerous sequential classification tasks, and the\nresults demonstrate that CLEAS outperforms other state-of-the-art alternative\nmethods, achieving higher classification accuracy while using simpler neural\narchitectures.', 'Many advanced Learning from Demonstration (LfD) methods consider the\ndecomposition of complex, real-world tasks into simpler sub-tasks. By reusing\nthe corresponding sub-policies within and between tasks, they provide training\ndata for each policy from different high-level tasks and compose them to\nperform novel ones. Existing approaches to modular LfD focus either on learning\na single high-level task or depend on domain knowledge and temporal\nsegmentation. In contrast, we propose a weakly supervised, domain-agnostic\napproach based on task sketches, which include only the sequence of sub-tasks\nperformed in each demonstration. Our approach simultaneously aligns the\nsketches with the observed demonstrations and learns the required sub-policies.\nThis improves generalisation in comparison to separate optimisation procedures.\nWe evaluate the approach on multiple domains, including a simulated 3D robot\narm control task using purely image-based observations. The results show that\nour approach performs commensurately with fully supervised approaches, while\nrequiring significantly less annotation effort.', 'Domain similarity measures can be used to gauge adaptability and select\nsuitable data for transfer learning, but existing approaches define ad hoc\nmeasures that are deemed suitable for respective tasks. Inspired by work on\ncurriculum learning, we propose to \\emph{learn} data selection measures using\nBayesian Optimization and evaluate them across models, domains and tasks. Our\nlearned measures outperform existing domain similarity measures significantly\non three tasks: sentiment analysis, part-of-speech tagging, and parsing. We\nshow the importance of complementing similarity with diversity, and that\nlearned measures are -- to some degree -- transferable across models, domains,\nand even tasks.', 'Model-based reinforcement learning (MBRL) methods have shown strong sample\nefficiency and performance across a variety of tasks, including when faced with\nhigh-dimensional visual observations. These methods learn to predict the\nenvironment dynamics and expected reward from interaction and use this\npredictive model to plan and perform the task. However, MBRL methods vary in\ntheir fundamental design choices, and there is no strong consensus in the\nliterature on how these design decisions affect performance. In this paper, we\nstudy a number of design decisions for the predictive model in visual MBRL\nalgorithms, focusing specifically on methods that use a predictive model for\nplanning. We find that a range of design decisions that are often considered\ncrucial, such as the use of latent spaces, have little effect on task\nperformance. A big exception to this finding is that predicting future\nobservations (i.e., images) leads to significant task performance improvement\ncompared to only predicting rewards. We also empirically find that image\nprediction accuracy, somewhat surprisingly, correlates more strongly with\ndownstream task performance than reward prediction accuracy. We show how this\nphenomenon is related to exploration and how some of the lower-scoring models\non standard benchmarks (that require exploration) will perform the same as the\nbest-performing models when trained on the same training data. Simultaneously,\nin the absence of exploration, models that fit the data better usually perform\nbetter on the downstream task as well, but surprisingly, these are often not\nthe same models that perform the best when learning and exploring from scratch.\nThese findings suggest that performance and exploration place important and\npotentially contradictory requirements on the model.', 'We propose an algorithm for meta-learning that is model-agnostic, in the\nsense that it is compatible with any model trained with gradient descent and\napplicable to a variety of different learning problems, including\nclassification, regression, and reinforcement learning. The goal of\nmeta-learning is to train a model on a variety of learning tasks, such that it\ncan solve new learning tasks using only a small number of training samples. In\nour approach, the parameters of the model are explicitly trained such that a\nsmall number of gradient steps with a small amount of training data from a new\ntask will produce good generalization performance on that task. In effect, our\nmethod trains the model to be easy to fine-tune. We demonstrate that this\napproach leads to state-of-the-art performance on two few-shot image\nclassification benchmarks, produces good results on few-shot regression, and\naccelerates fine-tuning for policy gradient reinforcement learning with neural\nnetwork policies.', 'Multi-task learning in Convolutional Networks has displayed remarkable\nsuccess in the field of recognition. This success can be largely attributed to\nlearning shared representations from multiple supervisory tasks. However,\nexisting multi-task approaches rely on enumerating multiple network\narchitectures specific to the tasks at hand, that do not generalize. In this\npaper, we propose a principled approach to learn shared representations in\nConvNets using multi-task learning. Specifically, we propose a new sharing\nunit: ""cross-stitch"" unit. These units combine the activations from multiple\nnetworks and can be trained end-to-end. A network with cross-stitch units can\nlearn an optimal combination of shared and task-specific representations. Our\nproposed method generalizes across multiple tasks and shows dramatically\nimproved performance over baseline methods for categories with few training\nexamples.', 'A core aspect of human intelligence is the ability to learn new tasks quickly\nand switch between them flexibly. Here, we describe a modular continual\nreinforcement learning paradigm inspired by these abilities. We first introduce\na visual interaction environment that allows many types of tasks to be unified\nin a single framework. We then describe a reward map prediction scheme that\nlearns new tasks robustly in the very large state and action spaces required by\nsuch an environment. We investigate how properties of module architecture\ninfluence efficiency of task learning, showing that a module motif\nincorporating specific design principles (e.g. early bottlenecks, low-order\npolynomial nonlinearities, and symmetry) significantly outperforms more\nstandard neural network motifs, needing fewer training examples and fewer\nneurons to achieve high levels of performance. Finally, we present a\nmeta-controller architecture for task switching based on a dynamic neural\nvoting scheme, which allows new modules to use information learned from\npreviously-seen tasks to substantially improve their own learning efficiency.', 'The growing interest in both the automation of machine learning and deep\nlearning has inevitably led to the development of a wide variety of automated\nmethods for neural architecture search. The choice of the network architecture\nhas proven to be critical, and many advances in deep learning spring from its\nimmediate improvements. However, deep learning techniques are computationally\nintensive and their application requires a high level of domain knowledge.\nTherefore, even partial automation of this process helps to make deep learning\nmore accessible to both researchers and practitioners. With this survey, we\nprovide a formalism which unifies and categorizes the landscape of existing\nmethods along with a detailed analysis that compares and contrasts the\ndifferent approaches. We achieve this via a comprehensive discussion of the\ncommonly adopted architecture search spaces and architecture optimization\nalgorithms based on principles of reinforcement learning and evolutionary\nalgorithms along with approaches that incorporate surrogate and one-shot\nmodels. Additionally, we address the new research directions which include\nconstrained and multi-objective architecture search as well as automated data\naugmentation, optimizer and activation function search.', ""The reinforcement learning community has made great strides in designing\nalgorithms capable of exceeding human performance on specific tasks. These\nalgorithms are mostly trained one task at the time, each new task requiring to\ntrain a brand new agent instance. This means the learning algorithm is general,\nbut each solution is not; each agent can only solve the one task it was trained\non. In this work, we study the problem of learning to master not one but\nmultiple sequential-decision tasks at once. A general issue in multi-task\nlearning is that a balance must be found between the needs of multiple tasks\ncompeting for the limited resources of a single learning system. Many learning\nalgorithms can get distracted by certain tasks in the set of tasks to solve.\nSuch tasks appear more salient to the learning process, for instance because of\nthe density or magnitude of the in-task rewards. This causes the algorithm to\nfocus on those salient tasks at the expense of generality. We propose to\nautomatically adapt the contribution of each task to the agent's updates, so\nthat all tasks have a similar impact on the learning dynamics. This resulted in\nstate of the art performance on learning to play all games in a set of 57\ndiverse Atari games. Excitingly, our method learned a single trained policy -\nwith a single set of weights - that exceeds median human performance. To our\nknowledge, this was the first time a single agent surpassed human-level\nperformance on this multi-task domain. The same approach also demonstrated\nstate of the art performance on a set of 30 tasks in the 3D reinforcement\nlearning platform DeepMind Lab."", 'Deep reinforcement learning (DRL) has recently shown its success in tackling\ncomplex combinatorial optimization problems. When these problems are extended\nto multiobjective ones, it becomes difficult for the existing DRL approaches to\nflexibly and efficiently deal with multiple subproblems determined by weight\ndecomposition of objectives. This paper proposes a concise meta-learning-based\nDRL approach. It first trains a meta-model by meta-learning. The meta-model is\nfine-tuned with a few update steps to derive submodels for the corresponding\nsubproblems. The Pareto front is then built accordingly. Compared with other\nlearning-based methods, our method can greatly shorten the training time of\nmultiple submodels. Due to the rapid and excellent adaptability of the\nmeta-model, more submodels can be derived so as to increase the quality and\ndiversity of the found solutions. The computational experiments on\nmultiobjective traveling salesman problems and multiobjective vehicle routing\nproblem with time windows demonstrate the superiority of our method over most\nof learning-based and iteration-based approaches.', 'The application of reinforcement learning algorithms onto real life problems\nalways bears the challenge of filtering the environmental state out of raw\nsensor readings. While most approaches use heuristics, biology suggests that\nthere must exist an unsupervised method to construct such filters\nautomatically. Besides the extraction of environmental states, the filters have\nto represent them in a fashion that support modern reinforcement algorithms.\nMany popular algorithms use a linear architecture, so one should aim at filters\nthat have good approximation properties in combination with linear functions.\nThis thesis wants to propose the unsupervised method slow feature analysis\n(SFA) for this task. Presented with a random sequence of sensor readings, SFA\nlearns a set of filters. With growing model complexity and training examples,\nthe filters converge against trigonometric polynomial functions. These are\nknown to possess excellent approximation capabilities and should therfore\nsupport the reinforcement algorithms well. We evaluate this claim on a robot.\nThe task is to learn a navigational control in a simple environment using the\nleast square policy iteration (LSPI) algorithm. The only accessible sensor is a\nhead mounted video camera, but without meaningful filtering, video images are\nnot suited as LSPI input. We will show that filters learned by SFA, based on a\nrandom walk video of the robot, allow the learned control to navigate\nsuccessfully in ca. 80% of the test trials.', 'Traditional model-based reinforcement learning approaches learn a model of\nthe environment dynamics without explicitly considering how it will be used by\nthe agent. In the presence of misspecified model classes, this can lead to poor\nestimates, as some relevant available information is ignored. In this paper, we\nintroduce a novel model-based policy search approach that exploits the\nknowledge of the current agent policy to learn an approximate transition model,\nfocusing on the portions of the environment that are most relevant for policy\nimprovement. We leverage a weighting scheme, derived from the minimization of\nthe error on the model-based policy gradient estimator, in order to define a\nsuitable objective function that is optimized for learning the approximate\ntransition model. Then, we integrate this procedure into a batch policy\nimprovement algorithm, named Gradient-Aware Model-based Policy Search (GAMPS),\nwhich iteratively learns a transition model and uses it, together with the\ncollected trajectories, to compute the new policy parameters. Finally, we\nempirically validate GAMPS on benchmark domains analyzing and discussing its\nproperties.', 'Meta-learning approaches have shown great success in vision and language\ndomains. However, few studies discuss the practice of meta-learning for\nlarge-scale industrial applications. Although e-commerce companies have spent\nmany efforts on learning representations to provide a better user experience,\nwe argue that such efforts cannot be stopped at this step. In addition to\nlearning a strong profile, the challenging question about how to effectively\ntransfer the learned representation is raised simultaneously. This paper\nintroduces the contributions that we made to address these challenges from\nthree aspects. 1) Meta-learning model: In the context of representation\nlearning with e-commerce user behavior data, we propose a meta-learning\nframework called the Meta-Profile Network, which extends the ideas of matching\nnetwork and relation network for knowledge transfer and fast adaptation; 2)\nEncoding strategy: To keep high fidelity of large-scale long-term sequential\nbehavior data, we propose a time-heatmap encoding strategy that allows the\nmodel to encode data effectively; 3) Deep network architecture: A multi-modal\nmodel combined with multi-task learning architecture is utilized to address the\ncross-domain knowledge learning and insufficient label problems. Moreover, we\nargue that an industrial model should not only have good performance in terms\nof accuracy, but also have better robustness and uncertainty performance under\nextreme conditions. We evaluate the performance of our model with extensive\ncontrol experiments in various extreme scenarios, i.e. out-of-distribution\ndetection, data insufficiency and class imbalance scenarios. The Meta-Profile\nNetwork shows significant improvement in the model performance when compared to\nbaseline models.', 'We consider problems of making sequences of decisions to accomplish tasks,\ninteracting via the medium of language. These problems are often tackled with\nreinforcement learning approaches. We find that these models do not generalize\nwell when applied to novel task domains. However, the large amount of\ncomputation necessary to adequately train and explore the search space of\nsequential decision making, under a reinforcement learning paradigm, precludes\nthe inclusion of large contextualized language models, which might otherwise\nenable the desired generalization ability. We introduce a teacher-student\nimitation learning methodology and a means of converting a reinforcement\nlearning model into a natural language understanding model. Together, these\nmethodologies enable the introduction of contextualized language models into\nthe sequential decision making problem space. We show that models can learn\nfaster and generalize more, leveraging both the imitation learning and the\nreformulation. Our models exceed teacher performance on various held-out\ndecision problems, by up to 7% on in-domain problems and 24% on out-of-domain\nproblems.', 'In order to efficiently learn with small amount of data on new tasks,\nmeta-learning transfers knowledge learned from previous tasks to the new ones.\nHowever, a critical challenge in meta-learning is the task heterogeneity which\ncannot be well handled by traditional globally shared meta-learning methods. In\naddition, current task-specific meta-learning methods may either suffer from\nhand-crafted structure design or lack the capability to capture complex\nrelations between tasks. In this paper, motivated by the way of knowledge\norganization in knowledge bases, we propose an automated relational\nmeta-learning (ARML) framework that automatically extracts the cross-task\nrelations and constructs the meta-knowledge graph. When a new task arrives, it\ncan quickly find the most relevant structure and tailor the learned structure\nknowledge to the meta-learner. As a result, the proposed framework not only\naddresses the challenge of task heterogeneity by a learned meta-knowledge\ngraph, but also increases the model interpretability. We conduct extensive\nexperiments on 2D toy regression and few-shot image classification and the\nresults demonstrate the superiority of ARML over state-of-the-art baselines.', 'Through many recent successes in simulation, model-free reinforcement\nlearning has emerged as a promising approach to solving continuous control\nrobotic tasks. The research community is now able to reproduce, analyze and\nbuild quickly on these results due to open source implementations of learning\nalgorithms and simulated benchmark tasks. To carry forward these successes to\nreal-world applications, it is crucial to withhold utilizing the unique\nadvantages of simulations that do not transfer to the real world and experiment\ndirectly with physical robots. However, reinforcement learning research with\nphysical robots faces substantial resistance due to the lack of benchmark tasks\nand supporting source code. In this work, we introduce several reinforcement\nlearning tasks with multiple commercially available robots that present varying\nlevels of learning difficulty, setup, and repeatability. On these tasks, we\ntest the learning performance of off-the-shelf implementations of four\nreinforcement learning algorithms and analyze sensitivity to their\nhyper-parameters to determine their readiness for applications in various\nreal-world tasks. Our results show that with a careful setup of the task\ninterface and computations, some of these implementations can be readily\napplicable to physical robots. We find that state-of-the-art learning\nalgorithms are highly sensitive to their hyper-parameters and their relative\nordering does not transfer across tasks, indicating the necessity of re-tuning\nthem for each task for best performance. On the other hand, the best\nhyper-parameter configuration from one task may often result in effective\nlearning on held-out tasks even with different robots, providing a reasonable\ndefault. We make the benchmark tasks publicly available to enhance\nreproducibility in real-world reinforcement learning.', ""With the explosive growth of online products and content, recommendation\ntechniques have been considered as an effective tool to overcome information\noverload, improve user experience, and boost business revenue. In recent years,\nwe have observed a new desideratum of considering long-term rewards of multiple\nrelated recommendation tasks simultaneously. The consideration of long-term\nrewards is strongly tied to business revenue and growth. Learning multiple\ntasks simultaneously could generally improve the performance of individual task\ndue to knowledge sharing in multi-task learning. While a few existing works\nhave studied long-term rewards in recommendations, they mainly focus on a\nsingle recommendation task. In this paper, we propose {\\it PoDiRe}: a\n\\underline{po}licy \\underline{di}stilled \\underline{re}commender that can\naddress long-term rewards of recommendations and simultaneously handle multiple\nrecommendation tasks. This novel recommendation solution is based on a marriage\nof deep reinforcement learning and knowledge distillation techniques, which is\nable to establish knowledge sharing among different tasks and reduce the size\nof a learning model. The resulting model is expected to attain better\nperformance and lower response latency for real-time recommendation services.\nIn collaboration with Samsung Game Launcher, one of the world's largest\ncommercial mobile game platforms, we conduct a comprehensive experimental study\non large-scale real data with hundreds of millions of events and show that our\nsolution outperforms many state-of-the-art methods in terms of several standard\nevaluation metrics."", 'Reinforcement learning can train policies that effectively perform complex\ntasks. However for long-horizon tasks, the performance of these methods\ndegrades with horizon, often necessitating reasoning over and chaining\nlower-level skills. Hierarchical reinforcement learning aims to enable this by\nproviding a bank of low-level skills as action abstractions. Hierarchies can\nfurther improve on this by abstracting the space states as well. We posit that\na suitable state abstraction should depend on the capabilities of the available\nlower-level policies. We propose Value Function Spaces: a simple approach that\nproduces such a representation by using the value functions corresponding to\neach lower-level skill. These value functions capture the affordances of the\nscene, thus forming a representation that compactly abstracts task relevant\ninformation and robustly ignores distractors. Empirical evaluations for\nmaze-solving and robotic manipulation tasks demonstrate that our approach\nimproves long-horizon performance and enables better zero-shot generalization\nthan alternative model-free and model-based methods.', 'This study investigates the use of reinforcement learning to guide a general\npurpose cache manager decisions. Cache managers directly impact the overall\nperformance of computer systems. They govern decisions about which objects\nshould be cached, the duration they should be cached for, and decides on which\nobjects to evict from the cache if it is full. These three decisions impact\nboth the cache hit rate and size of the storage that is needed to achieve that\ncache hit rate. An optimal cache manager will avoid unnecessary operations,\nmaximise the cache hit rate which results in fewer round trips to a slower\nbackend storage system, and minimise the size of storage needed to achieve a\nhigh hit-rate.\n  This project investigates using reinforcement learning in cache management by\ndesigning three separate agents for each of the cache manager tasks.\nFurthermore, the project investigates two advanced reinforcement learning\narchitectures for multi-decision problems: a single multi-task agent and a\nmulti-agent. We also introduce a framework to simplify the modelling of\ncomputer systems problems as a reinforcement learning task. The framework\nabstracts delayed experiences observations and reward assignment in computer\nsystems while providing a flexible way to scale to multiple agents.\n  Simulation results based on an established database benchmark system show\nthat reinforcement learning agents can achieve a higher cache hit rate over\nheuristic driven algorithms while minimising the needed space. They are also\nable to adapt to a changing workload and dynamically adjust their caching\nstrategy accordingly. The proposed cache manager model is generic and\napplicable to other types of caches, such as file system caches. This project\nis the first, to our knowledge, to model cache manager decisions as a\nmulti-task control problem.', 'We present here a model to take advantage of the multi-task nature of complex\ndatasets by learning to separate tasks and subtasks in and end to end manner by\nbiasing competitive interactions in the network. This method does not require\nadditional labelling or reformatting of data in a dataset. We propose an\nalternate view to the monolithic one-task-fits-all learning of multi-task\nproblems, and describe a model based on a theory of neuronal attention from\nneuroscience, proposed by Desimone. We create and exhibit a new toy dataset,\nbased on the MNIST dataset, which we call MNIST-QA, for testing Visual Question\nAnswering architectures in a low-dimensional environment while preserving the\nmore difficult components of the Visual Question Answering task, and\ndemonstrate the proposed network architecture on this new dataset, as well as\non COCO-QA and DAQUAR-FULL. We then demonstrate that this model eliminates\ncatastrophic interference between tasks on a newly created toy dataset and\nprovides competitive results in the Visual Question Answering space. We provide\nfurther evidence that Visual Question Answering can be approached as a\nmulti-task problem, and demonstrate that this new architecture based on the\nBiased Competition model is capable of learning to separate and learn the tasks\nin an end-to-end fashion without the need for task labels.', 'We describe a framework for multitask deep reinforcement learning guided by\npolicy sketches. Sketches annotate tasks with sequences of named subtasks,\nproviding information about high-level structural relationships among tasks but\nnot how to implement them---specifically not providing the detailed guidance\nused by much previous work on learning policy abstractions for RL (e.g.\nintermediate rewards, subtask completion signals, or intrinsic motivations). To\nlearn from sketches, we present a model that associates every subtask with a\nmodular subpolicy, and jointly maximizes reward over full task-specific\npolicies by tying parameters across shared subpolicies. Optimization is\naccomplished via a decoupled actor--critic training objective that facilitates\nlearning common behaviors from multiple dissimilar reward functions. We\nevaluate the effectiveness of our approach in three environments featuring both\ndiscrete and continuous control, and with sparse rewards that can be obtained\nonly after completing a number of high-level subgoals. Experiments show that\nusing our approach to learn policies guided by sketches gives better\nperformance than existing techniques for learning task-specific or shared\npolicies, while naturally inducing a library of interpretable primitive\nbehaviors that can be recombined to rapidly adapt to new tasks.', 'We are interested in learning models of non-stationary environments, which\ncan be framed as a multi-task learning problem. Model-free reinforcement\nlearning algorithms can achieve good asymptotic performance in multi-task\nlearning at a cost of extensive sampling, due to their approach, which requires\nlearning from scratch. While model-based approaches are among the most data\nefficient learning algorithms, they still struggle with complex tasks and model\nuncertainties. Meta-reinforcement learning addresses the efficiency and\ngeneralization challenges on multi task learning by quickly leveraging the\nmeta-prior policy for a new task. In this paper, we propose a\nmeta-reinforcement learning approach to learn the dynamic model of a\nnon-stationary environment to be used for meta-policy optimization later. Due\nto the sample efficiency of model-based learning methods, we are able to\nsimultaneously train both the meta-model of the non-stationary environment and\nthe meta-policy until dynamic model convergence. Then, the meta-learned dynamic\nmodel of the environment will generate simulated data for meta-policy\noptimization. Our experiment demonstrates that our proposed method can\nmeta-learn the policy in a non-stationary environment with the data efficiency\nof model-based learning approaches while achieving the high asymptotic\nperformance of model-free meta-reinforcement learning.', 'Meta-reinforcement learning (meta-RL) aims to learn from multiple training\ntasks the ability to adapt efficiently to unseen test tasks. Despite the\nsuccess, existing meta-RL algorithms are known to be sensitive to the task\ndistribution shift. When the test task distribution is different from the\ntraining task distribution, the performance may degrade significantly. To\naddress this issue, this paper proposes Model-based Adversarial\nMeta-Reinforcement Learning (AdMRL), where we aim to minimize the worst-case\nsub-optimality gap -- the difference between the optimal return and the return\nthat the algorithm achieves after adaptation -- across all tasks in a family of\ntasks, with a model-based approach. We propose a minimax objective and optimize\nit by alternating between learning the dynamics model on a fixed task and\nfinding the adversarial task for the current model -- the task for which the\npolicy induced by the model is maximally suboptimal. Assuming the family of\ntasks is parameterized, we derive a formula for the gradient of the\nsuboptimality with respect to the task parameters via the implicit function\ntheorem, and show how the gradient estimator can be efficiently implemented by\nthe conjugate gradient method and a novel use of the REINFORCE estimator. We\nevaluate our approach on several continuous control benchmarks and demonstrate\nits efficacy in the worst-case performance over all tasks, the generalization\npower to out-of-distribution tasks, and in training and test time sample\nefficiency, over existing state-of-the-art meta-RL algorithms.', 'Federated learning is promising for its capability to collaboratively train\nmodels with multiple clients without accessing their data, but vulnerable when\nclients\' data distributions diverge from each other. This divergence further\nleads to a dilemma: ""Should we prioritize the learned model\'s generic\nperformance (for future use at the server) or its personalized performance (for\neach client)?"" These two, seemingly competing goals have divided the community\nto focus on one or the other, yet in this paper we show that it is possible to\napproach both at the same time. Concretely, we propose a novel federated\nlearning framework that explicitly decouples a model\'s dual duties with two\nprediction tasks. On the one hand, we introduce a family of losses that are\nrobust to non-identical class distributions, enabling clients to train a\ngeneric predictor with a consistent objective across them. On the other hand,\nwe formulate the personalized predictor as a lightweight adaptive module that\nis learned to minimize each client\'s empirical risk on top of the generic\npredictor. With this two-loss, two-predictor framework which we name Federated\nRobust Decoupling (Fed-RoD), the learned model can simultaneously achieve\nstate-of-the-art generic and personalized performance, essentially bridging the\ntwo tasks.', 'The aim of multi-task reinforcement learning is two-fold: (1) efficiently\nlearn by training against multiple tasks and (2) quickly adapt, using limited\nsamples, to a variety of new tasks. In this work, the tasks correspond to\nreward functions for environments with the same (or similar) dynamical models.\nWe propose to learn a dynamical model during the training process and use this\nmodel to perform sample-efficient adaptation to new tasks at test time. We use\nsignificantly fewer samples by performing policy optimization only in a\n""virtual"" environment whose transitions are given by our learned dynamical\nmodel. Our algorithm sequentially trains against several tasks. Upon\nencountering a new task, we first warm-up a policy on our learned dynamical\nmodel, which requires no new samples from the environment. We then adapt the\ndynamical model with samples from this policy in the real environment. We\nevaluate our approach on several continuous control benchmarks and demonstrate\nits efficacy over MAML, a state-of-the-art meta-learning algorithm, on these\ntasks.', ""We propose a framework for verifiable and compositional reinforcement\nlearning (RL) in which a collection of RL subsystems, each of which learns to\naccomplish a separate subtask, are composed to achieve an overall task. The\nframework consists of a high-level model, represented as a parametric Markov\ndecision process (pMDP) which is used to plan and to analyze compositions of\nsubsystems, and of the collection of low-level subsystems themselves. By\ndefining interfaces between the subsystems, the framework enables automatic\ndecompositions of task specifications, e.g., reach a target set of states with\na probability of at least 0.95, into individual subtask specifications, i.e.\nachieve the subsystem's exit conditions with at least some minimum probability,\ngiven that its entry conditions are met. This in turn allows for the\nindependent training and testing of the subsystems; if they each learn a policy\nsatisfying the appropriate subtask specification, then their composition is\nguaranteed to satisfy the overall task specification. Conversely, if the\nsubtask specifications cannot all be satisfied by the learned policies, we\npresent a method, formulated as the problem of finding an optimal set of\nparameters in the pMDP, to automatically update the subtask specifications to\naccount for the observed shortcomings. The result is an iterative procedure for\ndefining subtask specifications, and for training the subsystems to meet them.\nAs an additional benefit, this procedure allows for particularly challenging or\nimportant components of an overall task to be determined automatically, and\nfocused on, during training. Experimental results demonstrate the presented\nframework's novel capabilities."", 'Temporal information is essential to learning effective policies with\nReinforcement Learning (RL). However, current state-of-the-art RL algorithms\neither assume that such information is given as part of the state space or,\nwhen learning from pixels, use the simple heuristic of frame-stacking to\nimplicitly capture temporal information present in the image observations. This\nheuristic is in contrast to the current paradigm in video classification\narchitectures, which utilize explicit encodings of temporal information through\nmethods such as optical flow and two-stream architectures to achieve\nstate-of-the-art performance. Inspired by leading video classification\narchitectures, we introduce the Flow of Latents for Reinforcement Learning\n(Flare), a network architecture for RL that explicitly encodes temporal\ninformation through latent vector differences. We show that Flare (i) recovers\noptimal performance in state-based RL without explicit access to the state\nvelocity, solely with positional state information, (ii) achieves\nstate-of-the-art performance on pixel-based challenging continuous control\ntasks within the DeepMind control benchmark suite, namely quadruped walk,\nhopper hop, finger turn hard, pendulum swing, and walker run, and is the most\nsample efficient model-free pixel-based RL algorithm, outperforming the prior\nmodel-free state-of-the-art by 1.9X and 1.5X on the 500k and 1M step\nbenchmarks, respectively, and (iv), when augmented over rainbow DQN,\noutperforms this state-of-the-art level baseline on 5 of 8 challenging Atari\ngames at 100M time step benchmark.', 'In this work, we consider the problem of instance-wise dynamic network model\nselection for multi-task learning. To this end, we propose an efficient\napproach to exploit a compact but accurate model in a backbone architecture for\neach instance of all tasks. The proposed method consists of an estimator and a\nselector. The estimator is based on a backbone architecture and structured\nhierarchically. It can produce multiple different network models of different\nconfigurations in a hierarchical structure. The selector chooses a model\ndynamically from a pool of candidate models given an input instance. The\nselector is a relatively small-size network consisting of a few layers, which\nestimates a probability distribution over the candidate models when an input\ninstance of a task is given. Both estimator and selector are jointly trained in\na unified learning framework in conjunction with a sampling-based learning\nstrategy, without additional computation steps. We demonstrate the proposed\napproach for several image classification tasks compared to existing approaches\nperforming model selection or learning multiple tasks. Experimental results\nshow that our approach gives not only outstanding performance compared to other\ncompetitors but also the versatility to perform instance-wise model selection\nfor multiple tasks.', 'This paper considers the problem of learning a model in model-based\nreinforcement learning (MBRL). We examine how the planning module of an MBRL\nalgorithm uses the model, and propose that the model learning module should\nincorporate the way the planner is going to use the model. This is in contrast\nto conventional model learning approaches, such as those based on maximum\nlikelihood estimate, that learn a predictive model of the environment without\nexplicitly considering the interaction of the model and the planner. We focus\non policy gradient type of planning algorithms and derive new loss functions\nfor model learning that incorporate how the planner uses the model. We call\nthis approach Policy-Aware Model Learning (PAML). We theoretically analyze a\ngeneric model-based policy gradient algorithm and provide a convergence\nguarantee for the optimized policy. We also empirically evaluate PAML on some\nbenchmark problems, showing promising results.', 'The use of imitation learning to learn a single policy for a complex task\nthat has multiple modes or hierarchical structure can be challenging. In fact,\nprevious work has shown that when the modes are known, learning separate\npolicies for each mode or sub-task can greatly improve the performance of\nimitation learning. In this work, we discover the interaction between sub-tasks\nfrom their resulting state-action trajectory sequences using a directed\ngraphical model. We propose a new algorithm based on the generative adversarial\nimitation learning framework which automatically learns sub-task policies from\nunsegmented demonstrations. Our approach maximizes the directed information\nflow in the graphical model between sub-task latent variables and their\ngenerated trajectories. We also show how our approach connects with the\nexisting Options framework, which is commonly used to learn hierarchical\npolicies.', 'Hierarchical reinforcement learning deals with the problem of breaking down\nlarge tasks into meaningful sub-tasks. Autonomous discovery of these sub-tasks\nhas remained a challenging problem. We propose a novel method of learning\nsub-tasks by combining paradigms of routing in computer networks and graph\nbased skill discovery within the options framework to define meaningful\nsub-goals. We apply the recent advancements of learning embeddings using\nRiemannian optimisation in the hyperbolic space to embed the state set into the\nhyperbolic space and create a model of the environment. In doing so we enforce\na global topology on the states and are able to exploit this topology to learn\nmeaningful sub-tasks. We demonstrate empirically, both in discrete and\ncontinuous domains, how these embeddings can improve the learning of meaningful\nsub-tasks.', 'Graph Neural Networks (GNNs) are a framework for graph representation\nlearning, where a model learns to generate low dimensional node embeddings that\nencapsulate structural and feature-related information. GNNs are usually\ntrained in an end-to-end fashion, leading to highly specialized node\nembeddings. However, generating node embeddings that can be used to perform\nmultiple tasks (with performance comparable to single-task models) is an open\nproblem. We propose a novel meta-learning strategy capable of producing\nmulti-task node embeddings. Our method avoids the difficulties arising when\nlearning to perform multiple tasks concurrently by, instead, learning to\nquickly (i.e. with a few steps of gradient descent) adapt to multiple tasks\nsingularly. We show that the embeddings produced by our method can be used to\nperform multiple tasks with comparable or higher performance than classically\ntrained models. Our method is model-agnostic and task-agnostic, thus applicable\nto a wide variety of multi-task domains.', 'When faced with learning challenging new tasks, humans often follow sequences\nof steps that allow them to incrementally build up the necessary skills for\nperforming these new tasks. However, in machine learning, models are most often\ntrained to solve the target tasks directly.Inspired by human learning, we\npropose a novel curriculum learning approach which decomposes challenging tasks\ninto sequences of easier intermediate goals that are used to pre-train a model\nbefore tackling the target task. We focus on classification tasks, and design\nthe intermediate tasks using an automatically constructed label hierarchy. We\ntrain the model at each level of the hierarchy, from coarse labels to fine\nlabels, transferring acquired knowledge across these levels. For instance, the\nmodel will first learn to distinguish animals from objects, and then use this\nacquired knowledge when learning to classify among more fine-grained classes\nsuch as cat, dog, car, and truck. Most existing curriculum learning algorithms\nfor supervised learning consist of scheduling the order in which the training\nexamples are presented to the model. In contrast, our approach focuses on the\noutput space of the model. We evaluate our method on several established\ndatasets and show significant performance gains especially on classification\nproblems with many labels. We also evaluate on a new synthetic dataset which\nallows us to study multiple aspects of our method.', 'Recent research has proposed the lottery ticket hypothesis, suggesting that\nfor a deep neural network, there exist trainable sub-networks performing\nequally or better than the original model with commensurate training steps.\nWhile this discovery is insightful, finding proper sub-networks requires\niterative training and pruning. The high cost incurred limits the applications\nof the lottery ticket hypothesis. We show there exists a subset of the\naforementioned sub-networks that converge significantly faster during the\ntraining process and thus can mitigate the cost issue. We conduct extensive\nexperiments to show such sub-networks consistently exist across various model\nstructures for a restrictive setting of hyperparameters ($e.g.$, carefully\nselected learning rate, pruning ratio, and model capacity). As a practical\napplication of our findings, we demonstrate that such sub-networks can help in\ncutting down the total time of adversarial training, a standard approach to\nimprove robustness, by up to 49\\% on CIFAR-10 to achieve the state-of-the-art\nrobustness.', 'Imitation learning has been commonly applied to solve different tasks in\nisolation. This usually requires either careful feature engineering, or a\nsignificant number of samples. This is far from what we desire: ideally, robots\nshould be able to learn from very few demonstrations of any given task, and\ninstantly generalize to new situations of the same task, without requiring\ntask-specific engineering. In this paper, we propose a meta-learning framework\nfor achieving such capability, which we call one-shot imitation learning.\n  Specifically, we consider the setting where there is a very large set of\ntasks, and each task has many instantiations. For example, a task could be to\nstack all blocks on a table into a single tower, another task could be to place\nall blocks on a table into two-block towers, etc. In each case, different\ninstances of the task would consist of different sets of blocks with different\ninitial states. At training time, our algorithm is presented with pairs of\ndemonstrations for a subset of all tasks. A neural net is trained that takes as\ninput one demonstration and the current state (which initially is the initial\nstate of the other demonstration of the pair), and outputs an action with the\ngoal that the resulting sequence of states and actions matches as closely as\npossible with the second demonstration. At test time, a demonstration of a\nsingle instance of a new task is presented, and the neural net is expected to\nperform well on new instances of this new task. The use of soft attention\nallows the model to generalize to conditions and tasks unseen in the training\ndata. We anticipate that by training this model on a much greater variety of\ntasks and settings, we will obtain a general system that can turn any\ndemonstrations into robust policies that can accomplish an overwhelming variety\nof tasks.\n  Videos available at https://bit.ly/nips2017-oneshot .', ""Hybrid ensemble, an essential branch of ensembles, has flourished in the\nregression field, with studies confirming diversity's importance. However,\nprevious ensembles consider diversity in the sub-model training stage, with\nlimited improvement compared to single models. In contrast, this study\nautomatically selects and weights sub-models from a heterogeneous model pool.\nIt solves an optimization problem using an interior-point filtering\nlinear-search algorithm. The objective function innovatively incorporates\nnegative correlation learning as a penalty term, with which a diverse model\nsubset can be selected. The best sub-models from each model class are selected\nto build the NCL ensemble, which performance is better than the simple average\nand other state-of-the-art weighting methods. It is also possible to improve\nthe NCL ensemble with a regularization term in the objective function. In\npractice, it is difficult to conclude the optimal sub-model for a dataset prior\ndue to the model uncertainty. Regardless, our method would achieve comparable\naccuracy as the potential optimal sub-models. In conclusion, the value of this\nstudy lies in its ease of use and effectiveness, allowing the hybrid ensemble\nto embrace diversity and accuracy."", 'This paper presents a novel framework for automatic learning of complex\nstrategies in human decision making. The task that we are interested in is to\nbetter facilitate long term planning for complex, multi-step events. We observe\ntemporal relationships at the subtask level of expert demonstrations, and\ndetermine the different strategies employed in order to successfully complete a\ntask. To capture the relationship between the subtasks and the overall goal, we\nutilise two external memory modules, one for capturing dependencies within a\nsingle expert demonstration, such as the sequential relationship among\ndifferent sub tasks, and a global memory module for modelling task level\ncharacteristics such as best practice employed by different humans based on\ntheir domain expertise. Furthermore, we demonstrate how the hidden state\nrepresentation of the memory can be used as a reward signal to smooth the state\ntransitions, eradicating subtle changes. We evaluate the effectiveness of the\nproposed model for an autonomous highway driving application, where we\ndemonstrate its capability to learn different expert policies and outperform\nstate-of-the-art methods. The scope in industrial applications extends to any\nrobotics and automation application which requires learning from complex\ndemonstrations containing series of subtasks.']","['In this work, we study the problem of learning a single model for multiple\ndomains. Unlike the conventional machine learning scenario where each domain\ncan have the corresponding model, multiple domains (i.e., applications/users)\nmay share the same machine learning model due to maintenance loads in cloud\ncomputing services. For example, a digit-recognition model should be applicable\nto hand-written digits, house numbers, car plates, etc. Therefore, an ideal\nmodel for cloud computing has to perform well at each applicable domain. To\naddress this new challenge from cloud computing, we develop a framework of\nrobust optimization over multiple domains. In lieu of minimizing the empirical\nrisk, we aim to learn a model optimized to the adversarial distribution over\nmultiple domains. Hence, we propose to learn the model and the adversarial\ndistribution simultaneously with the stochastic algorithm for efficiency.\nTheoretically, we analyze the convergence rate for convex and non-convex\nmodels. To our best knowledge, we first study the convergence rate of learning\na robust non-convex model with a practical algorithm. Furthermore, we\ndemonstrate that the robustness of the framework and the convergence rate can\nbe further enhanced by appropriate regularizers over the adversarial\ndistribution. The empirical study on real-world fine-grained visual\ncategorization and digits recognition tasks verifies the effectiveness and\nefficiency of the proposed framework.', 'In recent years, model-free methods that use deep learning have achieved\ngreat success in many different reinforcement learning environments. Most\nsuccessful approaches focus on solving a single task, while multi-task\nreinforcement learning remains an open problem. In this paper, we present a\nmodel based approach to deep reinforcement learning which we use to solve\ndifferent tasks simultaneously. We show that our approach not only does not\ndegrade but actually benefits from learning multiple tasks. For our model, we\nalso present a new kind of recurrent neural network inspired by residual\nnetworks that decouples memory from computation allowing to model complex\nenvironments that do not require lots of memory.', 'We introduce network with sub-networks, a neural network which its weight\nlayers could be detached into sub-neural networks during inference. To develop\nweights and biases which could be inserted in both base and sub-neural\nnetworks, firstly, the parameters are copied from sub-model to base-model. Each\nmodel is forward-propagated separately. Gradients from a pair of networks are\naveraged and, used to update both networks. Our base model achieves the\ntest-accuracy which is comparable to the regularly trained models, while the\nmodel maintains the ability to detach weight layers.', 'When a neural language model (LM) is adapted to perform a new task, what\naspects of the task predict the eventual performance of the model? In NLP,\nsystematic features of LM generalization to individual examples are well\ncharacterized, but systematic aspects of LM adaptability to new tasks are not\nnearly as well understood. We present a large-scale empirical study of the\nfeatures and limits of LM adaptability using a new benchmark, TaskBench500,\nbuilt from 500 procedurally generated sequence modeling tasks. These tasks\ncombine core aspects of language processing, including lexical semantics,\nsequence processing, memorization, logical reasoning, and world knowledge.\nUsing TaskBench500, we evaluate three facets of adaptability, finding that: (1)\nadaptation procedures differ dramatically in their ability to memorize small\ndatasets; (2) within a subset of task types, adaptation procedures exhibit\ncompositional adaptability to complex tasks; and (3) failure to match training\nlabel distributions is explained by mismatches in the intrinsic difficulty of\npredicting individual labels. Our experiments show that adaptability to new\ntasks, like generalization to new examples, can be systematically described and\nunderstood, and we conclude with a discussion of additional aspects of\nadaptability that could be studied using the new benchmark.', 'Continual learning is the problem of learning new tasks or knowledge while\nprotecting old knowledge and ideally generalizing from old experience to learn\nnew tasks faster. Neural networks trained by stochastic gradient descent often\ndegrade on old tasks when trained successively on new tasks with different data\ndistributions. This phenomenon, referred to as catastrophic forgetting, is\nconsidered a major hurdle to learning with non-stationary data or sequences of\nnew tasks, and prevents networks from continually accumulating knowledge and\nskills. We examine this issue in the context of reinforcement learning, in a\nsetting where an agent is exposed to tasks in a sequence. Unlike most other\nwork, we do not provide an explicit indication to the model of task boundaries,\nwhich is the most general circumstance for a learning agent exposed to\ncontinuous experience. While various methods to counteract catastrophic\nforgetting have recently been proposed, we explore a straightforward, general,\nand seemingly overlooked solution - that of using experience replay buffers for\nall past events - with a mixture of on- and off-policy learning, leveraging\nbehavioral cloning. We show that this strategy can still learn new tasks\nquickly yet can substantially reduce catastrophic forgetting in both Atari and\nDMLab domains, even matching the performance of methods that require task\nidentities. When buffer storage is constrained, we confirm that a simple\nmechanism for randomly discarding data allows a limited size buffer to perform\nalmost as well as an unbounded one.', 'Fictitious play with reinforcement learning is a general and effective\nframework for zero-sum games. However, using the current deep neural network\nmodels, the implementation of fictitious play faces crucial challenges. Neural\nnetwork model training employs gradient descent approaches to update all\nconnection weights, and thus is easy to forget the old opponents after training\nto beat the new opponents. Existing approaches often maintain a pool of\nhistorical policy models to avoid the forgetting. However, learning to beat a\npool in stochastic games, i.e., a wide distribution over policy models, is\neither sample-consuming or insufficient to exploit all models with limited\namount of samples. In this paper, we propose a learning process with neural\nfictitious play to alleviate the above issues. We train a single model as our\npolicy model, which consists of sub-models and a selector. Everytime facing a\nnew opponent, the model is expanded by adding a new sub-model, where only the\nnew sub-model is updated instead of the whole model. At the same time, the\nselector is also updated to mix up the new sub-model with the previous ones at\nthe state-level, so that the model is maintained as a behavior strategy instead\nof a wide distribution over policy models. Experiments on Kuhn poker, a\ngrid-world Treasure Hunting game, and Mini-RTS environments show that the\nproposed approach alleviates the forgetting problem, and consequently improves\nthe learning efficiency and the robustness of neural fictitious play.', 'An important long-term goal in machine learning systems is to build learning\nagents that, like humans, can learn many tasks over their lifetime, and\nmoreover use information from these tasks to improve their ability to do so\nefficiently. In this work, our goal is to provide new theoretical insights into\nthe potential of this paradigm. In particular, we propose a lifelong learning\nframework that adheres to a novel notion of resource efficiency that is\ncritical in many real-world domains where feature evaluations are costly. That\nis, our learner aims to reuse information from previously learned related tasks\nto learn future tasks in a feature-efficient manner. Furthermore, we consider\nnovel combinatorial ways in which learning tasks can relate. Specifically, we\ndesign lifelong learning algorithms for two structurally different and widely\nused families of target functions: decision trees/lists and\nmonomials/polynomials. We also provide strong feature-efficiency guarantees for\nthese algorithms; in fact, we show that in order to learn future targets, we\nneed only slightly more feature evaluations per training example than what is\nneeded to predict on an arbitrary example using those targets. We also provide\nalgorithms with guarantees in an agnostic model where not all the targets are\nrelated to each other. Finally, we also provide lower bounds on the performance\nof a lifelong learner in these models, which are in fact tight under some\nconditions.', ""The motivation of this study is to leverage recent breakthroughs in\nartificial intelligence research to unlock novel solutions to important\nscientific problems encountered in computational science. To address the human\nintelligence limitations in discovering reduced-order models, we propose to\nsupplement human thinking with artificial intelligence. Our three-pronged\nstrategy consists of learning (i) models expressed in analytical form, (ii)\nwhich are evaluated a posteriori, and iii) using exclusively integral\nquantities from the reference solution as prior knowledge. In point (i), we\npursue interpretable models expressed symbolically as opposed to black-box\nneural networks, the latter only being used during learning to efficiently\nparameterize the large search space of possible models. In point (ii), learned\nmodels are dynamically evaluated a posteriori in the computational solver\ninstead of based on a priori information from preprocessed high-fidelity data,\nthereby accounting for the specificity of the solver at hand such as its\nnumerics. Finally in point (iii), the exploration of new models is solely\nguided by predefined integral quantities, e.g., averaged quantities of\nengineering interest in Reynolds-averaged or large-eddy simulations (LES). We\nuse a coupled deep reinforcement learning framework and computational solver to\nconcurrently achieve these objectives. The combination of reinforcement\nlearning with objectives (i), (ii) and (iii) differentiate our work from\nprevious modeling attempts based on machine learning. In this report, we\nprovide a high-level description of the model discovery framework with\nreinforcement learning. The method is detailed for the application of\ndiscovering missing terms in differential equations. An elementary\ninstantiation of the method is described that discovers missing terms in the\nBurgers' equation."", ""Deep learning models are considered to be state-of-the-art in many offline\nmachine learning tasks. However, many of the techniques developed are not\nsuitable for online learning tasks. The problem of using deep learning models\nwith sequential data becomes even harder when several loss functions need to be\nconsidered simultaneously, as in many real-world applications. In this paper,\nwe, therefore, propose a novel online deep learning training procedure which\ncan be used regardless of the neural network's architecture, aiming to deal\nwith the multiple objectives case. We demonstrate and show the effectiveness of\nour algorithm on the Neyman-Pearson classification problem on several benchmark\ndatasets."", 'Many NLP learning tasks can be decomposed into several distinct sub-tasks,\neach associated with a partial label. In this paper we focus on a popular class\nof learning problems, sequence prediction applied to several sentiment analysis\ntasks, and suggest a modular learning approach in which different sub-tasks are\nlearned using separate functional modules, combined to perform the final task\nwhile sharing information. Our experiments show this approach helps constrain\nthe learning process and can alleviate some of the supervision efforts.', 'Few-Shot Learning is the challenge of training a model with only a small\namount of data. Many solutions to this problem use meta-learning algorithms,\ni.e. algorithms that learn to learn. By sampling few-shot tasks from a larger\ndataset, we can teach these algorithms to solve new, unseen tasks. This\ndocument reports my work on meta-learning algorithms for Few-Shot Computer\nVision. This work was done during my internship at Sicara, a French company\nbuilding image recognition solutions for businesses. It contains: 1. an\nextensive review of the state-of-the-art in few-shot computer vision; 2. a\nbenchmark of meta-learning algorithms for few-shot image classification; 3. the\nintroduction to a novel meta-learning algorithm for few-shot object detection,\nwhich is still in development.', 'In the artificial intelligence field, learning often corresponds to changing\nthe parameters of a parameterized function. A learning rule is an algorithm or\nmathematical expression that specifies precisely how the parameters should be\nchanged. When creating an artificial intelligence system, we must make two\ndecisions: what representation should be used (i.e., what parameterized\nfunction should be used) and what learning rule should be used to search\nthrough the resulting set of representable functions. Using most learning\nrules, these two decisions are coupled in a subtle (and often unintentional)\nway. That is, using the same learning rule with two different representations\nthat can represent the same sets of functions can result in two different\noutcomes. After arguing that this coupling is undesirable, particularly when\nusing artificial neural networks, we present a method for partially decoupling\nthese two decisions for a broad class of learning rules that span unsupervised\nlearning, reinforcement learning, and supervised learning.', ""Labeling data (e.g., labeling the people, objects, actions and scene in\nimages) comprehensively and efficiently is a widely needed but challenging\ntask. Numerous models were proposed to label various data and many approaches\nwere designed to enhance the ability of deep learning models or accelerate\nthem. Unfortunately, a single machine-learning model is not powerful enough to\nextract various semantic information from data. Given certain applications,\nsuch as image retrieval platforms and photo album management apps, it is often\nrequired to execute a collection of models to obtain sufficient labels. With\nlimited computing resources and stringent delay, given a data stream and a\ncollection of applicable resource-hungry deep-learning models, we design a\nnovel approach to adaptively schedule a subset of these models to execute on\neach data item, aiming to maximize the value of the model output (e.g., the\nnumber of high-confidence labels). Achieving this lofty goal is nontrivial\nsince a model's output on any data item is content-dependent and unknown until\nwe execute it. To tackle this, we propose an Adaptive Model Scheduling\nframework, consisting of 1) a deep reinforcement learning-based approach to\npredict the value of unexecuted models by mining semantic relationship among\ndiverse models, and 2) two heuristic algorithms to adaptively schedule the\nmodel execution order under a deadline or deadline-memory constraints\nrespectively. The proposed framework doesn't require any prior knowledge of the\ndata, which works as a powerful complement to existing model optimization\ntechnologies. We conduct extensive evaluations on five diverse image datasets\nand 30 popular image labeling models to demonstrate the effectiveness of our\ndesign: our design could save around 53\\% execution time without loss of any\nvaluable labels."", 'The objective of a reinforcement learning agent is to behave so as to\nmaximise the sum of a suitable scalar function of state: the reward. These\nrewards are typically given and immutable. In this paper, we instead consider\nthe proposition that the reward function itself can be a good locus of learned\nknowledge. To investigate this, we propose a scalable meta-gradient framework\nfor learning useful intrinsic reward functions across multiple lifetimes of\nexperience. Through several proof-of-concept experiments, we show that it is\nfeasible to learn and capture knowledge about long-term exploration and\nexploitation into a reward function. Furthermore, we show that unlike policy\ntransfer methods that capture ""how"" the agent should behave, the learned reward\nfunctions can generalise to other kinds of agents and to changes in the\ndynamics of the environment by capturing ""what"" the agent should strive to do.', 'Interpretability has become a necessary feature for machine learning models\ndeployed in critical scenarios, e.g. legal system, healthcare. In these\nsituations, algorithmic decisions may have (potentially negative) long-lasting\neffects on the end-user affected by the decision. In many cases, the\nrepresentational power of deep learning models is not needed, therefore simple\nand interpretable models (e.g. linear models) should be preferred. However, in\nhigh-dimensional and/or complex domains (e.g. computer vision), the universal\napproximation capabilities of neural networks are required. Inspired by linear\nmodels and the Kolmogorov-Arnold representation theorem, we propose a novel\nclass of structurally-constrained neural networks, which we call FLANs\n(Feature-wise Latent Additive Networks). Crucially, FLANs process each input\nfeature separately, computing for each of them a representation in a common\nlatent space. These feature-wise latent representations are then simply summed,\nand the aggregated representation is used for prediction. These constraints\n(which are at the core of the interpretability of linear models) allow a user\nto estimate the effect of each individual feature independently from the\nothers, enhancing interpretability. In a set of experiments across different\ndomains, we show how without compromising excessively the test performance, the\nstructural constraints proposed in FLANs indeed facilitates the\ninterpretability of deep learning models. We quantitatively compare FLANs\ninterpretability to post-hoc methods using recently introduced metrics,\ndiscussing the advantages of natively interpretable models over a post-hoc\nanalysis.', ""This paper presents a meta-learning framework for few-shots One-Class\nClassification (OCC) at test-time, a setting where labeled examples are only\navailable for the positive class, and no supervision is given for the negative\nexample. We consider that we have a set of `one-class classification'\nobjective-tasks with only a small set of positive examples available for each\ntask, and a set of training tasks with full supervision (i.e. highly imbalanced\nclassification). We propose an approach using order-equivariant networks to\nlearn a 'meta' binary-classifier. The model will take as input an example to\nclassify from a given task, as well as the corresponding supervised set of\npositive examples for this OCC task. Thus, the output of the model will be\n'conditioned' on the available positive example of a given task, allowing to\npredict on new tasks and new examples without labeled negative examples. In\nthis paper, we are motivated by an astronomy application. Our goal is to\nidentify if stars belong to a specific stellar group (the 'one-class' for a\ngiven task), called \\textit{stellar streams}, where each stellar stream is a\ndifferent OCC-task. We show that our method transfers well on unseen (test)\nsynthetic streams, and outperforms the baselines even though it is not\nretrained and accesses a much smaller part of the data per task to predict\n(only positive supervision). We see however that it doesn't transfer as well on\nthe real stream GD-1. This could come from intrinsic differences from the\nsynthetic and real stream, highlighting the need for consistency in the\n'nature' of the task for this method. However, light fine-tuning improve\nperformances and outperform our baselines. Our experiments show encouraging\nresults to further explore meta-learning methods for OCC tasks."", 'Curriculum learning in reinforcement learning is a training methodology that\nseeks to speed up learning of a difficult target task, by first training on a\nseries of simpler tasks and transferring the knowledge acquired to the target\ntask. Automatically choosing a sequence of such tasks (i.e. a curriculum) is an\nopen problem that has been the subject of much recent work in this area. In\nthis paper, we build upon a recent method for curriculum design, which\nformulates the curriculum sequencing problem as a Markov Decision Process. We\nextend this model to handle multiple transfer learning algorithms, and show for\nthe first time that a curriculum policy over this MDP can be learned from\nexperience. We explore various representations that make this possible, and\nevaluate our approach by learning curriculum policies for multiple agents in\ntwo different domains. The results show that our method produces curricula that\ncan train agents to perform on a target task as fast or faster than existing\nmethods.', 'Continual learning is a concept of online learning with multiple sequential\ntasks. One of the critical barriers of continual learning is that a network\nshould learn a new task keeping the knowledge of old tasks without access to\nany data of the old tasks. In this paper, we propose a neuron activation\nimportance-based regularization method for stable continual learning regardless\nof the order of tasks. We conduct comprehensive experiments on existing\nbenchmark data sets to evaluate not just the stability and plasticity of our\nmethod with improved classification accuracy also the robustness of the\nperformance along the changes of task order.', 'The work of McCloskey and Cohen popularized the concept of catastrophic\ninterference. They used a neural network that tried to learn addition using two\ngroups of examples as two different tasks. In their case, learning the second\ntask rapidly deteriorated the acquired knowledge about the previous one. We\nhypothesize that this could be a symptom of a fundamental problem: addition is\nan algorithmic task that should not be learned through pattern recognition.\nTherefore, other model architectures better suited for this task would avoid\ncatastrophic forgetting. We use a neural network with a different architecture\nthat can be trained to recover the correct algorithm for the addition of binary\nnumbers. This neural network includes conditional clauses that are naturally\ntreated within the back-propagation algorithm. We test it in the setting\nproposed by McCloskey and Cohen and training on random additions one by one.\nThe neural network not only does not suffer from catastrophic forgetting but it\nimproves its predictive power on unseen pairs of numbers as training\nprogresses. We also show that this is a robust effect, also present when\naveraging many simulations. This work emphasizes the importance that neural\nnetwork architecture has for the emergence of catastrophic forgetting and\nintroduces a neural network that is able to learn an algorithm.', 'Machine understanding of complex images is a key goal of artificial\nintelligence. One challenge underlying this task is that visual scenes contain\nmultiple inter-related objects, and that global context plays an important role\nin interpreting the scene. A natural modeling framework for capturing such\neffects is structured prediction, which optimizes over complex labels, while\nmodeling within-label interactions. However, it is unclear what principles\nshould guide the design of a structured prediction model that utilizes the\npower of deep learning components. Here we propose a design principle for such\narchitectures that follows from a natural requirement of permutation\ninvariance. We prove a necessary and sufficient characterization for\narchitectures that follow this invariance, and discuss its implication on model\ndesign. Finally, we show that the resulting model achieves new state of the art\nresults on the Visual Genome scene graph labeling benchmark, outperforming all\nrecent approaches.', ""Multi-task learning aims to improve generalization performance of multiple\nprediction tasks by appropriately sharing relevant information across them. In\nthe context of deep neural networks, this idea is often realized by\nhand-designed network architectures with layers that are shared across tasks\nand branches that encode task-specific features. However, the space of possible\nmulti-task deep architectures is combinatorially large and often the final\narchitecture is arrived at by manual exploration of this space subject to\ndesigner's bias, which can be both error-prone and tedious. In this work, we\npropose a principled approach for designing compact multi-task deep learning\narchitectures. Our approach starts with a thin network and dynamically widens\nit in a greedy manner during training using a novel criterion that promotes\ngrouping of similar tasks together. Our Extensive evaluation on person\nattributes classification tasks involving facial and clothing attributes\nsuggests that the models produced by the proposed method are fast, compact and\ncan closely match or exceed the state-of-the-art accuracy from strong baselines\nby much more expensive models."", 'It has been a long-standing goal in machine learning, as well as in AI more\ngenerally, to develop life-long learning systems that learn many different\ntasks over time, and reuse insights from tasks learned, ""learning to learn"" as\nthey do so. In this work we pose and provide efficient algorithms for several\nnatural theoretical formulations of this goal. Specifically, we consider the\nproblem of learning many different target functions over time, that share\ncertain commonalities that are initially unknown to the learning algorithm. Our\naim is to learn new internal representations as the algorithm learns new target\nfunctions, that capture this commonality and allow subsequent learning tasks to\nbe solved more efficiently and from less data. We develop efficient algorithms\nfor two very different kinds of commonalities that target functions might\nshare: one based on learning common low-dimensional and unions of\nlow-dimensional subspaces and one based on learning nonlinear Boolean\ncombinations of features. Our algorithms for learning Boolean feature\ncombinations additionally have a dual interpretation, and can be viewed as\ngiving an efficient procedure for constructing near-optimal sparse Boolean\nautoencoders under a natural ""anchor-set"" assumption.', 'One of the questions that arises when designing models that learn to solve\nmultiple tasks simultaneously is how much of the available training budget\nshould be devoted to each individual task. We refer to any formalized approach\nto addressing this problem (learned or otherwise) as a task selection policy.\nIn this work we provide an empirical evaluation of the performance of some\ncommon task selection policies in a synthetic bandit-style setting, as well as\non the GLUE benchmark for natural language understanding. We connect task\nselection policy learning to existing work on automated curriculum learning and\noff-policy evaluation, and suggest a method based on counterfactual estimation\nthat leads to improved model performance in our experimental settings.', 'In this paper, we compare the three most popular algorithms for\nhyperparameter optimization (Grid Search, Random Search, and Genetic Algorithm)\nand attempt to use them for neural architecture search (NAS). We use these\nalgorithms for building a convolutional neural network (search architecture).\nExperimental results on CIFAR-10 dataset further demonstrate the performance\ndifference between compared algorithms. The comparison results are based on the\nexecution time of the above algorithms and accuracy of the proposed models.', 'In open set learning, a model must be able to generalize to novel classes\nwhen it encounters a sample that does not belong to any of the classes it has\nseen before. Open set learning poses a realistic learning scenario that is\nreceiving growing attention. Existing studies on open set learning mainly\nfocused on detecting novel classes, but few studies tried to model them for\ndifferentiating novel classes. In this paper, we recognize that novel classes\nshould be different from each other, and propose distribution networks for open\nset learning that can model different novel classes based on probability\ndistributions. We hypothesize that, through a certain mapping, samples from\ndifferent classes with the same classification criterion should follow\ndifferent probability distributions from the same distribution family. A deep\nneural network is learned to map the samples in the original feature space to a\nlatent space where the distributions of known classes can be jointly learned\nwith the network. We additionally propose a distribution parameter transfer and\nupdating strategy for novel class modeling when a novel class is detected in\nthe latent space. By novel class modeling, the detected novel classes can serve\nas known classes to the subsequent classification. Our experimental results on\nimage datasets MNIST and CIFAR10 show that the distribution networks can detect\nnovel classes accurately, and model them well for the subsequent classification\ntasks.', 'Sharing information between multiple tasks enables algorithms to achieve good\ngeneralization performance even from small amounts of training data. However,\nin a realistic scenario of multi-task learning not all tasks are equally\nrelated to each other, hence it could be advantageous to transfer information\nonly between the most related tasks. In this work we propose an approach that\nprocesses multiple tasks in a sequence with sharing between subsequent tasks\ninstead of solving all tasks jointly. Subsequently, we address the question of\ncurriculum learning of tasks, i.e. finding the best order of tasks to be\nlearned. Our approach is based on a generalization bound criterion for choosing\nthe task order that optimizes the average expected classification performance\nover all tasks. Our experimental results show that learning multiple related\ntasks sequentially can be more effective than learning them jointly, the order\nin which tasks are being solved affects the overall performance, and that our\nmodel is able to automatically discover the favourable order of tasks.', ""Medical image segmentation has significantly benefitted thanks to deep\nlearning architectures. Furthermore, semi-supervised learning (SSL) has\nrecently been a growing trend for improving a model's overall performance by\nleveraging abundant unlabeled data. Moreover, learning multiple tasks within\nthe same model further improves model generalizability. To generate smoother\nand accurate segmentation masks from 3D cardiac MR images, we present a\nMulti-task Cross-task learning consistency approach to enforce the correlation\nbetween the pixel-level (segmentation) and the geometric-level (distance map)\ntasks. Our extensive experimentation with varied quantities of labeled data in\nthe training sets justifies the effectiveness of our model for the segmentation\nof the left atrial cavity from Gadolinium-enhanced magnetic resonance (GE-MR)\nimages. With the incorporation of uncertainty estimates to detect failures in\nthe segmentation masks generated by CNNs, our study further showcases the\npotential of our model to flag low-quality segmentation from a given model."", ""Prior work on training generative Visual Dialog models with reinforcement\nlearning(Das et al.) has explored a Qbot-Abot image-guessing game and shown\nthat this 'self-talk' approach can lead to improved performance at the\ndownstream dialog-conditioned image-guessing task. However, this improvement\nsaturates and starts degrading after a few rounds of interaction, and does not\nlead to a better Visual Dialog model. We find that this is due in part to\nrepeated interactions between Qbot and Abot during self-talk, which are not\ninformative with respect to the image. To improve this, we devise a simple\nauxiliary objective that incentivizes Qbot to ask diverse questions, thus\nreducing repetitions and in turn enabling Abot to explore a larger state space\nduring RL ie. be exposed to more visual concepts to talk about, and varied\nquestions to answer. We evaluate our approach via a host of automatic metrics\nand human studies, and demonstrate that it leads to better dialog, ie. dialog\nthat is more diverse (ie. less repetitive), consistent (ie. has fewer\nconflicting exchanges), fluent (ie. more human-like),and detailed, while still\nbeing comparably image-relevant as prior work and ablations."", ""Despite its potential to improve sample complexity versus model-free\napproaches, model-based reinforcement learning can fail catastrophically if the\nmodel is inaccurate. An algorithm should ideally be able to trust an imperfect\nmodel over a reasonably long planning horizon, and only rely on model-free\nupdates when the model errors get infeasibly large. In this paper, we\ninvestigate techniques for choosing the planning horizon on a state-dependent\nbasis, where a state's planning horizon is determined by the maximum cumulative\nmodel error around that state. We demonstrate that these state-dependent model\nerrors can be learned with Temporal Difference methods, based on a novel\napproach of temporally decomposing the cumulative model errors. Experimental\nresults show that the proposed method can successfully adapt the planning\nhorizon to account for state-dependent model accuracy, significantly improving\nthe efficiency of policy learning compared to model-based and model-free\nbaselines."", 'Machine learning algorithms have enabled computers to predict things by\nlearning from previous data. The data storage and processing power are\nincreasing rapidly, thus increasing machine learning and Artificial\nintelligence applications. Much of the work is done to improve the accuracy of\nthe models built in the past, with little research done to determine the\ncomputational costs of machine learning acquisitions. In this paper, I will\nproceed with this later research work and will make a performance comparison of\nmulti-threaded machine learning clustering algorithms. I will be working on\nLinear Regression, Random Forest, and K-Nearest Neighbors to determine the\nperformance characteristics of the algorithms as well as the computation costs\nto the obtained results. I will be benchmarking system hardware performance by\nrunning these multi-threaded algorithms to train and test the models on a\ndataset to note the differences in performance matrices of the algorithms. In\nthe end, I will state the best performing algorithms concerning the performance\nefficiency of these algorithms on my system.', 'Model complexity is a fundamental problem in deep learning. In this paper we\nconduct a systematic overview of the latest studies on model complexity in deep\nlearning. Model complexity of deep learning can be categorized into expressive\ncapacity and effective model complexity. We review the existing studies on\nthose two categories along four important factors, including model framework,\nmodel size, optimization process and data complexity. We also discuss the\napplications of deep learning model complexity including understanding model\ngeneralization, model optimization, and model selection and design. We conclude\nby proposing several interesting future directions.', 'During sleep and awake rest, the hippocampus replays sequences of place cells\nthat have been activated during prior experiences. These have been interpreted\nas a memory consolidation process, but recent results suggest a possible\ninterpretation in terms of reinforcement learning. The Dyna reinforcement\nlearning algorithms use off-line replays to improve learning. Under limited\nreplay budget, a prioritized sweeping approach, which requires a model of the\ntransitions to the predecessors, can be used to improve performance. We\ninvestigate whether such algorithms can explain the experimentally observed\nreplays. We propose a neural network version of prioritized sweeping\nQ-learning, for which we developed a growing multiple expert algorithm, able to\ncope with multiple predecessors. The resulting architecture is able to improve\nthe learning of simulated agents confronted to a navigation task. We predict\nthat, in animals, learning the world model should occur during rest periods,\nand that the corresponding replays should be shuffled.', 'To solve the problem of redundant information and overlapping relations of\nthe entity and relation extraction model, we propose a joint extraction model.\nThis model can directly extract multiple pairs of related entities without\ngenerating unrelated redundant information. We also propose a recurrent neural\nnetwork named Encoder-LSTM that enhances the ability of recurrent units to\nmodel sentences. Specifically, the joint model includes three sub-modules: the\nNamed Entity Recognition sub-module consisted of a pre-trained language model\nand an LSTM decoder layer, the Entity Pair Extraction sub-module which uses\nEncoder-LSTM network to model the order relationship between related entity\npairs, and the Relation Classification sub-module including Attention\nmechanism. We conducted experiments on the public datasets ADE and CoNLL04 to\nevaluate the effectiveness of our model. The results show that the proposed\nmodel achieves good performance in the task of entity and relation extraction\nand can greatly reduce the amount of redundant information.', 'All famous machine learning algorithms that comprise both supervised and\nsemi-supervised learning work well only under a common assumption: the training\nand test data follow the same distribution. When the distribution changes, most\nstatistical models must be reconstructed from newly collected data, which for\nsome applications can be costly or impossible to obtain. Therefore, it has\nbecome necessary to develop approaches that reduce the need and the effort to\nobtain new labeled samples by exploiting data that are available in related\nareas, and using these further across similar fields. This has given rise to a\nnew machine learning framework known as transfer learning: a learning setting\ninspired by the capability of a human being to extrapolate knowledge across\ntasks to learn more efficiently. Despite a large amount of different transfer\nlearning scenarios, the main objective of this survey is to provide an overview\nof the state-of-the-art theoretical results in a specific, and arguably the\nmost popular, sub-field of transfer learning, called domain adaptation. In this\nsub-field, the data distribution is assumed to change across the training and\nthe test data, while the learning task remains the same. We provide a first\nup-to-date description of existing results related to domain adaptation problem\nthat cover learning bounds based on different statistical learning frameworks.', 'Entity Matching (EM) is a core data cleaning task, aiming to identify\ndifferent mentions of the same real-world entity. Active learning is one way to\naddress the challenge of scarce labeled data in practice, by dynamically\ncollecting the necessary examples to be labeled by an Oracle and refining the\nlearned model (classifier) upon them. In this paper, we build a unified active\nlearning benchmark framework for EM that allows users to easily combine\ndifferent learning algorithms with applicable example selection algorithms. The\ngoal of the framework is to enable concrete guidelines for practitioners as to\nwhat active learning combinations will work well for EM. Towards this, we\nperform comprehensive experiments on publicly available EM datasets from\nproduct and publication domains to evaluate active learning methods, using a\nvariety of metrics including EM quality, #labels and example selection\nlatencies. Our most surprising result finds that active learning with fewer\nlabels can learn a classifier of comparable quality as supervised learning. In\nfact, for several of the datasets, we show that there is an active learning\ncombination that beats the state-of-the-art supervised learning result. Our\nframework also includes novel optimizations that improve the quality of the\nlearned model by roughly 9% in terms of F1-score and reduce example selection\nlatencies by up to 10x without affecting the quality of the model.', 'The idea of using multi-task learning approaches to address the joint\nextraction of entity and relation is motivated by the relatedness between the\nentity recognition task and the relation classification task. Existing methods\nusing multi-task learning techniques to address the problem learn interactions\namong the two tasks through a shared network, where the shared information is\npassed into the task-specific networks for prediction. However, such an\napproach hinders the model from learning explicit interactions between the two\ntasks to improve the performance on the individual tasks. As a solution, we\ndesign a multi-task learning model which we refer to as recurrent interaction\nnetwork which allows the learning of interactions dynamically, to effectively\nmodel task-specific features for classification. Empirical studies on two\nreal-world datasets confirm the superiority of the proposed model.', 'Given the fact of a case, Legal Judgment Prediction (LJP) involves a series\nof sub-tasks such as predicting violated law articles, charges and term of\npenalty. We propose leveraging a unified text-to-text Transformer for LJP,\nwhere the dependencies among sub-tasks can be naturally established within the\nauto-regressive decoder. Compared with previous works, it has three advantages:\n(1) it fits in the pretraining pattern of masked language models, and thereby\ncan benefit from the semantic prompts of each sub-task rather than treating\nthem as atomic labels, (2) it utilizes a single unified architecture, enabling\nfull parameter sharing across all sub-tasks, and (3) it can incorporate both\nclassification and generative sub-tasks. We show that this unified transformer,\nalbeit pretrained on general-domain text, outperforms pretrained models\ntailored specifically for the legal domain. Through an extensive set of\nexperiments, we find that the best order to capture dependencies is different\nfrom human intuitions, and the most reasonable logical order for humans can be\nsub-optimal for the model. We further include two more auxiliary tasks: court\nview generation and article content prediction, showing they can not only\nimprove the prediction accuracy, but also provide interpretable explanations\nfor model outputs even when an error is made. With the best configuration, our\nmodel outperforms both previous SOTA and a single-tasked version of the unified\ntransformer by a large margin.', ""Currently the only techniques for sharing governance of a deep learning model\nare homomorphic encryption and secure multiparty computation. Unfortunately,\nneither of these techniques is applicable to the training of large neural\nnetworks due to their large computational and communication overheads. As a\nscalable technique for shared model governance, we propose splitting deep\nlearning model between multiple parties. This paper empirically investigates\nthe security guarantee of this technique, which is introduced as the problem of\nmodel completion: Given the entire training data set or an environment\nsimulator, and a subset of the parameters of a trained deep learning model, how\nmuch training is required to recover the model's original performance? We\ndefine a metric for evaluating the hardness of the model completion problem and\nstudy it empirically in both supervised learning on ImageNet and reinforcement\nlearning on Atari and DeepMind~Lab. Our experiments show that (1) the model\ncompletion problem is harder in reinforcement learning than in supervised\nlearning because of the unavailability of the trained agent's trajectories, and\n(2) its hardness depends not primarily on the number of parameters of the\nmissing part, but more so on their type and location. Our results suggest that\nmodel splitting might be a feasible technique for shared model governance in\nsome settings where training is very expensive."", 'Identifying relationships between items is a key task of an online\nrecommender system, in order to help users discover items that are functionally\ncomplementary or visually compatible. In domains like clothing recommendation,\nthis task is particularly challenging since a successful system should be\ncapable of handling a large corpus of items, a huge amount of relationships\namong them, as well as the high-dimensional and semantically complicated\nfeatures involved. Furthermore, the human notion of ""compatibility"" to capture\ngoes beyond mere similarity: For two items to be compatible---whether jeans and\na t-shirt, or a laptop and a charger---they should be similar in some ways, but\nsystematically different in others.\n  In this paper we propose a novel method, Monomer, to learn complicated and\nheterogeneous relationships between items in product recommendation settings.\nRecently, scalable methods have been developed that address this task by\nlearning similarity metrics on top of the content of the products involved.\nHere our method relaxes the metricity assumption inherent in previous work and\nmodels multiple localized notions of \'relatedness,\' so as to uncover ways in\nwhich related items should be systematically similar, and systematically\ndifferent. Quantitatively, we show that our system achieves state-of-the-art\nperformance on large-scale compatibility prediction tasks, especially in cases\nwhere there is substantial heterogeneity between related items. Qualitatively,\nwe demonstrate that richer notions of compatibility can be learned that go\nbeyond similarity, and that our model can make effective recommendations of\nheterogeneous content.', 'To act and plan in complex environments, we posit that agents should have a\nmental simulator of the world with three characteristics: (a) it should build\nan abstract state representing the condition of the world; (b) it should form a\nbelief which represents uncertainty on the world; (c) it should go beyond\nsimple step-by-step simulation, and exhibit temporal abstraction. Motivated by\nthe absence of a model satisfying all these requirements, we propose TD-VAE, a\ngenerative sequence model that learns representations containing explicit\nbeliefs about states several steps into the future, and that can be rolled out\ndirectly without single-step transitions. TD-VAE is trained on pairs of\ntemporally separated time points, using an analogue of temporal difference\nlearning used in reinforcement learning.', 'This paper presents a new deep regression model, which we call DeepDistance,\nfor cell detection in images acquired with inverted microscopy. This model\nconsiders cell detection as a task of finding most probable locations that\nsuggest cell centers in an image. It represents this main task with a\nregression task of learning an inner distance metric. However, different than\nthe previously reported regression based methods, the DeepDistance model\nproposes to approach its learning as a multi-task regression problem where\nmultiple tasks are learned by using shared feature representations. To this\nend, it defines a secondary metric, normalized outer distance, to represent a\ndifferent aspect of the problem and proposes to define its learning as\ncomplementary to the main cell detection task. In order to learn these two\ncomplementary tasks more effectively, the DeepDistance model designs a fully\nconvolutional network (FCN) with a shared encoder path and end-to-end trains\nthis FCN to concurrently learn the tasks in parallel. DeepDistance uses the\ninner distances estimated by this FCN in a detection algorithm to locate\nindividual cells in a given image. For further performance improvement on the\nmain task, this paper also presents an extended version of the DeepDistance\nmodel. This extended model includes an auxiliary classification task and learns\nit in parallel to the two regression tasks by sharing feature representations\nwith them. Our experiments on three different human cell lines reveal that the\nproposed multi-task learning models, the DeepDistance model and its extended\nversion, successfully identify cell locations, even for the cell line that was\nnot used in training, and improve the results of the previous deep learning\nmethods.', 'We propose an adversarial training procedure for learning a causal implicit\ngenerative model for a given causal graph. We show that adversarial training\ncan be used to learn a generative model with true observational and\ninterventional distributions if the generator architecture is consistent with\nthe given causal graph. We consider the application of generating faces based\non given binary labels where the dependency structure between the labels is\npreserved with a causal graph. This problem can be seen as learning a causal\nimplicit generative model for the image and labels. We devise a two-stage\nprocedure for this problem. First we train a causal implicit generative model\nover binary labels using a neural network consistent with a causal graph as the\ngenerator. We empirically show that WassersteinGAN can be used to output\ndiscrete labels. Later, we propose two new conditional GAN architectures, which\nwe call CausalGAN and CausalBEGAN. We show that the optimal generator of the\nCausalGAN, given the labels, samples from the image distributions conditioned\non these labels. The conditional GAN combined with a trained causal implicit\ngenerative model for the labels is then a causal implicit generative model over\nthe labels and the generated image. We show that the proposed architectures can\nbe used to sample from observational and interventional image distributions,\neven for interventions which do not naturally occur in the dataset.', 'This doctoral thesis covers some of my advances in electron microscopy with\ndeep learning. Highlights include a comprehensive review of deep learning in\nelectron microscopy; large new electron microscopy datasets for machine\nlearning, dataset search engines based on variational autoencoders, and\nautomatic data clustering by t-distributed stochastic neighbour embedding;\nadaptive learning rate clipping to stabilize learning; generative adversarial\nnetworks for compressed sensing with spiral, uniformly spaced and other fixed\nsparse scan paths; recurrent neural networks trained to piecewise adapt sparse\nscan paths to specimens by reinforcement learning; improving signal-to-noise;\nand conditional generative adversarial networks for exit wavefunction\nreconstruction from single transmission electron micrographs. This thesis adds\nto my publications by presenting their relationships, reflections, and holistic\nconclusions. This version of my thesis is typeset for online dissemination to\nimprove readability, whereas the thesis submitted to the University of Warwick\nin support of my application for the degree of Doctor of Philosophy in Physics\nis typeset for physical printing and binding.', 'This paper makes a first step towards a logic of learning from experiments.\nFor this, we investigate formal frameworks for modeling the interaction of\ncausal and (qualitative) epistemic reasoning. Crucial for our approach is the\nidea that the notion of an intervention can be used as a formal expression of a\n(real or hypothetical) experiment. In a first step we extend the well-known\ncausal models with a simple Hintikka-style representation of the epistemic\nstate of an agent. In the resulting setting, one can talk not only about the\nknowledge of an agent about the values of variables and how interventions\naffect them, but also about knowledge update. The resulting logic can model\nreasoning about thought experiments. However, it is unable to account for\nlearning from experiments, which is clearly brought out by the fact that it\nvalidates the no learning principle for interventions. Therefore, in a second\nstep, we implement a more complex notion of knowledge that allows an agent to\nobserve (measure) certain variables when an experiment is carried out. This\nextended system does allow for learning from experiments. For all the proposed\nlogical systems, we provide a sound and complete axiomatization.', ""When data is unlabelled and the target task is not known a priori, divergent\nsearch offers a strategy for learning a wide range of skills. Having such a\nrepertoire allows a system to adapt to new, unforeseen tasks. Unlabelled image\ndata is plentiful, but it is not always known which features will be required\nfor downstream tasks. We propose a method for divergent search in the few-shot\nimage classification setting and evaluate with Omniglot and Mini-ImageNet. This\nhigh-dimensional behavior space includes all possible ways of partitioning the\ndata. To manage divergent search in this space, we rely on a meta-learning\nframework to integrate useful features from diverse tasks into a single model.\nThe final layer of this model is used as an index into the `archive' of all\npast behaviors. We search for regions in the behavior space that the current\narchive cannot reach. As expected, divergent search is outperformed by models\nwith a strong bias toward the evaluation tasks. But it is able to match and\nsometimes exceed the performance of models that have a weak bias toward the\ntarget task or none at all. This demonstrates that divergent search is a viable\napproach, even in high-dimensional behavior spaces."", ""Despite the potential of active inference for visual-based control, learning\nthe model and the preferences (priors) while interacting with the environment\nis challenging. Here, we study the performance of a deep active inference\n(dAIF) agent on OpenAI's car racing benchmark, where there is no access to the\ncar's state. The agent learns to encode the world's state from high-dimensional\ninput through unsupervised representation learning. State inference and control\nare learned end-to-end by optimizing the expected free energy. Results show\nthat our model achieves comparable performance to deep Q-learning. However,\nvanilla dAIF does not reach state-of-the-art performance compared to other\nworld model approaches. Hence, we discuss the current model implementation's\nlimitations and potential architectures to overcome them."", ""Out-of-distribution (OOD) detection has received much attention lately due to\nits practical importance in enhancing the safe deployment of neural networks.\nOne of the primary challenges is that models often produce highly confident\npredictions on OOD data, which undermines the driving principle in OOD\ndetection that the model should only be confident about in-distribution\nsamples. In this work, we propose ReAct--a simple and effective technique for\nreducing model overconfidence on OOD data. Our method is motivated by novel\nanalysis on internal activations of neural networks, which displays highly\ndistinctive signature patterns for OOD distributions. Our method can generalize\neffectively to different network architectures and different OOD detection\nscores. We empirically demonstrate that ReAct achieves competitive detection\nperformance on a comprehensive suite of benchmark datasets, and give\ntheoretical explication for our method's efficacy. On the ImageNet benchmark,\nReAct reduces the false positive rate (FPR95) by 25.05% compared to the\nprevious best method."", 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', ""Deep neural networks have achieved impressive results on a wide variety of\ntasks. However, quantifying uncertainty in the network's output is a\nchallenging task. Bayesian models offer a mathematical framework to reason\nabout model uncertainty. Variational methods have been used for approximating\nintractable integrals that arise in Bayesian inference for neural networks. In\nthis report, we review the major variational inference concepts pertinent to\nBayesian neural networks and compare various approximation methods used in\nliterature. We also talk about the applications of variational bayes in\nReinforcement learning and continual learning."", ""In model-based reinforcement learning it is typical to decouple the problems\nof learning the dynamics model and learning the reward function. However, when\nthe dynamics model is flawed, it may generate erroneous states that would never\noccur in the true environment. It is not clear a priori what value the reward\nfunction should assign to such states. This paper presents a novel error bound\nthat accounts for the reward model's behavior in states sampled from the model.\nThis bound is used to extend the existing Hallucinated DAgger-MC algorithm,\nwhich offers theoretical performance guarantees in deterministic MDPs that do\nnot assume a perfect model can be learned. Empirically, this approach to reward\nlearning can yield dramatic improvements in control performance when the\ndynamics model is flawed."", 'Objects in the world usually appear in context, participating in spatial\nrelationships and interactions that are predictable and expected. Knowledge of\nthese contexts can be used in the task of using a mobile camera to search for a\nspecified object in a room. We call this the object search task. This paper is\nconcerned with representing this knowledge in a manner facilitating its\napplication to object search while at the same time lending itself to\nautonomous learning by a robot. The ability for the robot to learn such\nknowledge without supervision is crucial due to the vast number of possible\nrelationships that can exist for any given set of objects. Moreover, since a\nrobot will not have an infinite amount of time to learn, it must be able to\ndetermine an order in which to look for possible relationships so as to\nmaximize the rate at which new knowledge is gained. In effect, there must be a\n""focus of interest"" operator that allows the robot to choose which examples are\nlikely to convey the most new information and should be examined first. This\npaper demonstrates how a representation based on statistical confidence\nintervals allows the construction of a system that achieves the above goals. An\nalgorithm, based on the Highest Impact First heuristic, is presented as a means\nfor providing a ""focus of interest"" with which to control the learning process,\nand examples are given.', 'In this article I describe a research agenda for securing machine learning\nmodels against adversarial inputs at test time. This article does not present\nresults but instead shares some of my thoughts about where I think that the\nfield needs to go. Modern machine learning works very well on I.I.D. data: data\nfor which each example is drawn {\\em independently} and for which the\ndistribution generating each example is {\\em identical}. When these assumptions\nare relaxed, modern machine learning can perform very poorly. When machine\nlearning is used in contexts where security is a concern, it is desirable to\ndesign models that perform well even when the input is designed by a malicious\nadversary. So far most research in this direction has focused on an adversary\nwho violates the {\\em identical} assumption, and imposes some kind of\nrestricted worst-case distribution shift. I argue that machine learning\nsecurity researchers should also address the problem of relaxing the {\\em\nindependence} assumption and that current strategies designed for robustness to\ndistribution shift will not do so. I recommend {\\em dynamic models} that change\neach time they are run as a potential solution path to this problem, and show\nan example of a simple attack using correlated data that can be mitigated by a\nsimple dynamic defense. This is not intended as a real-world security measure,\nbut as a recommendation to explore this research direction and develop more\nrealistic defenses.', ""Recently, there has been great interest in learning how to best represent\nproteins, specifically with fixed-length embeddings. Deep learning has become a\npopular tool for protein representation learning as a model's hidden layers\nproduce potentially useful vector embeddings. TAPE introduced a number of\nbenchmark tasks and showed that semi-supervised learning, via pretraining\nlanguage models on a large protein corpus, improved performance on downstream\ntasks. Two of the tasks (fluorescence prediction and stability prediction)\ninvolve learning fitness landscapes. In this paper, we show that CNN models\ntrained solely using supervised learning both compete with and sometimes\noutperform the best models from TAPE that leverage expensive pretraining on\nlarge protein datasets. These CNN models are sufficiently simple and small that\nthey can be trained using a Google Colab notebook. We also find for the\nfluorescence task that linear regression outperforms our models and the TAPE\nmodels. The benchmarking tasks proposed by TAPE are excellent measures of a\nmodel's ability to predict protein function and should be used going forward.\nHowever, we believe it is important to add baselines from simple models to put\nthe performance of the semi-supervised models that have been reported so far\ninto perspective."", 'We consider a many-to-one wireless architecture for federated learning at the\nnetwork edge, where multiple edge devices collaboratively train a model using\nlocal data. The unreliable nature of wireless connectivity, together with\nconstraints in computing resources at edge devices, dictates that the local\nupdates at edge devices should be carefully crafted and compressed to match the\nwireless communication resources available and should work in concert with the\nreceiver. Thus motivated, we propose SGD-based bandlimited coordinate descent\nalgorithms for such settings. Specifically, for the wireless edge employing\nover-the-air computing, a common subset of k-coordinates of the gradient\nupdates across edge devices are selected by the receiver in each iteration, and\nthen transmitted simultaneously over k sub-carriers, each experiencing\ntime-varying channel conditions. We characterize the impact of communication\nerror and compression, in terms of the resulting gradient bias and mean squared\nerror, on the convergence of the proposed algorithms. We then study\nlearning-driven communication error minimization via joint optimization of\npower allocation and learning rates. Our findings reveal that optimal power\nallocation across different sub-carriers should take into account both the\ngradient values and channel conditions, thus generalizing the widely used\nwater-filling policy. We also develop sub-optimal distributed solutions\namenable to implementation.', 'We emphasize the importance of asking the right question when interpreting\nthe decisions of a learning model. We discuss a natural extension of the\ntheoretical machinery from Janzing et. al. 2020, which answers the question\n""Why did my model predict a person has cancer?"" for answering a more involved\nquestion, ""What caused my model to predict a person has cancer?"" While the\nformer quantifies the direct effects of variables on the model, the latter also\naccounts for indirect effects, thereby providing meaningful insights wherever\nhuman beings can reason in terms of cause and effect. We propose three broad\ncategories for interpretations: observational, model-specific and causal each\nof which are significant in their own right. Furthermore, this paper quantifies\nfeature relevance by weaving different natures of interpretations together with\ndifferent measures as characteristic functions for Shapley symmetrization.\nBesides the widely used expected value of the model, we also discuss measures\nof statistical uncertainty and dispersion as informative candidates, and their\nmerits in generating explanations for each data point, some of which are used\nin this context for the first time. These measures are not only useful for\nstudying the influence of variables on the model output, but also on the\npredictive performance of the model, and for that we propose relevant\ncharacteristic functions that are also used for the first time.', 'Model interpretation, or explanation of a machine learning classifier, aims\nto extract generalizable knowledge from a trained classifier into a\nhuman-understandable format, for various purposes such as model assessment,\ndebugging and trust. From a computaional viewpoint, it is formulated as\napproximating the target classifier using a simpler interpretable model, such\nas rule models like a decision set/list/tree. Often, this approximation is\nhandled as standard supervised learning and the only difference is that the\nlabels are provided by the target classifier instead of ground truth. This\nparadigm is particularly popular because there exists a variety of well-studied\nsupervised algorithms for learning an interpretable classifier. However, we\nargue that this paradigm is suboptimal for it does not utilize the unique\nproperty of the model interpretation problem, that is, the ability to generate\nsynthetic instances and query the target classifier for their labels. We call\nthis the active-query property, suggesting that we should consider model\ninterpretation from an active learning perspective. Following this insight, we\nargue that the active-query property should be employed when designing a model\ninterpretation algorithm, and that the generation of synthetic instances should\nbe integrated seamlessly with the algorithm that learns the model\ninterpretation. In this paper, we demonstrate that by doing so, it is possible\nto achieve more faithful interpretation with simpler model complexity. As a\ntechnical contribution, we present an active algorithm Active Decision Set\nInduction (ADS) to learn a decision set, a set of if-else rules, for model\ninterpretation. ADS performs a local search over the space of all decision\nsets. In every iteration, ADS computes confidence intervals for the value of\nthe objective function of all local actions and utilizes active-query to\ndetermine the best one.', 'Entity resolution (ER) refers to the problem of matching records in one or\nmore relations that refer to the same real-world entity. While supervised\nmachine learning (ML) approaches achieve the state-of-the-art results, they\nrequire a large amount of labeled examples that are expensive to obtain and\noften times infeasible. We investigate an important problem that vexes\npractitioners: is it possible to design an effective algorithm for ER that\nrequires Zero labeled examples, yet can achieve performance comparable to\nsupervised approaches? In this paper, we answer in the affirmative through our\nproposed approach dubbed ZeroER. Our approach is based on a simple observation\n-- the similarity vectors for matches should look different from that of\nunmatches. Operationalizing this insight requires a number of technical\ninnovations. First, we propose a simple yet powerful generative model based on\nGaussian Mixture Models for learning the match and unmatch distributions.\nSecond, we propose an adaptive regularization technique customized for ER that\nameliorates the issue of feature overfitting. Finally, we incorporate the\ntransitivity property into the generative model in a novel way resulting in\nimproved accuracy. On five benchmark ER datasets, we show that ZeroER greatly\noutperforms existing unsupervised approaches and achieves comparable\nperformance to supervised approaches.', 'Neural network-based methods for abstractive summarization produce outputs\nthat are more fluent than other techniques, but which can be poor at content\nselection. This work proposes a simple technique for addressing this issue: use\na data-efficient content selector to over-determine phrases in a source\ndocument that should be part of the summary. We use this selector as a\nbottom-up attention step to constrain the model to likely phrases. We show that\nthis approach improves the ability to compress text, while still generating\nfluent summaries. This two-step process is both simpler and higher performing\nthan other end-to-end content selection models, leading to significant\nimprovements on ROUGE for both the CNN-DM and NYT corpus. Furthermore, the\ncontent selector can be trained with as little as 1,000 sentences, making it\neasy to transfer a trained summarizer to a new domain.', ""In this paper, I will discuss the work I am currently doing as a Ph.D.\nstudent at the University of Potsdam, under the tutoring of T. Schaub. I'm\ncurrently looking into action description in ASP. More precisely, my goal is to\nexplore how to represent actions with durations in ASP, in different contexts.\nRight now, I'm focused on Multi-Agent Path Finding (MAPF), looking at how to\nrepresent speeds for different agents and contexts.\n  Before tackling duration, I wanted to explore and compare different\nrepresentations of action taking in ASP. For this, I started comparing\ndifferent simple encodings tackling the MAPF problem. Even in simple code,\nchoices and assumptions have been made in their creations. The objective of my\nwork is to present the consequences of those design decisions in terms of\nperformance and knowledge representation. As far as I know, there is no current\nresearch on this topic.\n  Besides that, I'm also exploring different ways to represent duration and to\nsolve related problems. I planed to compare them the same way I described\nbefore. I also want this to help me find innovative and effective ways to solve\nproblems with duration."", 'My research lies in the intersection of security and machine learning. This\noverview summarizes one component of my research: combining computer vision\nwith malware exploit detection for enhanced security solutions. I will present\nthe perspectives of efficacy, reliability and resiliency to formulate threat\ndetection as computer vision problems and develop state-of-the-art image-based\nmalware classification. Representing malware binary as images provides a direct\nvisualization of data samples, reduces the efforts for feature extraction, and\nconsumes the whole binary for holistic structural analysis. Employing transfer\nlearning of deep neural networks effective for large scale image classification\nto malware classification demonstrates superior classification efficacy\ncompared with classical machine learning algorithms. To enhance reliability of\nthese vision-based malware detectors, interpretation frameworks can be\nconstructed on the malware visual representations and useful for extracting\nfaithful explanation, so that security practitioners have confidence in the\nmodel before deployment. In cyber-security applications, we should always\nassume that a malware writer constantly modifies code to bypass detection.\nAddressing the resiliency of the malware detectors is equivalently important as\nefficacy and reliability. Via understanding the attack surfaces of machine\nlearning models used for malware detection, we can greatly improve the\nrobustness of the algorithms to combat malware adversaries in the wild. Finally\nI will discuss future research directions worth pursuing in this research\ncommunity.', 'Recently we see a rising number of methods in the field of eXplainable\nArtificial Intelligence. To our surprise, their development is driven by model\ndevelopers rather than a study of needs for human end users. The analysis of\nneeds, if done, takes the form of an A/B test rather than a study of open\nquestions. To answer the question ""What would a human operator like to ask the\nML model?"" we propose a conversational system explaining decisions of the\npredictive model. In this experiment, we developed a chatbot called dr_ant to\ntalk about machine learning model trained to predict survival odds on Titanic.\nPeople can talk with dr_ant about different aspects of the model to understand\nthe rationale behind its predictions. Having collected a corpus of 1000+\ndialogues, we analyse the most common types of questions that users would like\nto ask. To our knowledge, it is the first study which uses a conversational\nsystem to collect the needs of human operators from the interactive and\niterative dialogue explorations of a predictive model.', 'Deciding what and when to observe is critical when making observations is\ncostly. In a medical setting where observations can be made sequentially,\nmaking these observations (or not) should be an active choice. We refer to this\nas the active sensing problem. In this paper, we propose a novel deep learning\nframework, which we call ASAC (Active Sensing using Actor-Critic models) to\naddress this problem. ASAC consists of two networks: a selector network and a\npredictor network. The selector network uses previously selected observations\nto determine what should be observed in the future. The predictor network uses\nthe observations selected by the selector network to predict a label, providing\nfeedback to the selector network (well-selected variables should be predictive\nof the label). The goal of the selector network is then to select variables\nthat balance the cost of observing the selected variables with their predictive\npower; we wish to preserve the conditional label distribution. During training,\nwe use the actor-critic models to allow the loss of the selector to be\n""back-propagated"" through the sampling process. The selector network ""acts"" by\nselecting future observations to make. The predictor network acts as a ""critic""\nby feeding predictive errors for the selected variables back to the selector\nnetwork. In our experiments, we show that ASAC significantly outperforms\nstate-of-the-arts in two real-world medical datasets.', 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.', ""We introduce a new task of rephrasing for a more natural virtual assistant.\nCurrently, virtual assistants work in the paradigm of intent slot tagging and\nthe slot values are directly passed as-is to the execution engine. However,\nthis setup fails in some scenarios such as messaging when the query given by\nthe user needs to be changed before repeating it or sending it to another user.\nFor example, for queries like 'ask my wife if she can pick up the kids' or\n'remind me to take my pills', we need to rephrase the content to 'can you pick\nup the kids' and 'take your pills' In this paper, we study the problem of\nrephrasing with messaging as a use case and release a dataset of 3000 pairs of\noriginal query and rephrased query. We show that BART, a pre-trained\ntransformers-based masked language model with auto-regressive decoding, is a\nstrong baseline for the task, and show improvements by adding a copy-pointer\nand copy loss to it. We analyze different tradeoffs of BART-based and\nLSTM-based seq2seq models, and propose a distilled LSTM-based seq2seq as the\nbest practical model."", 'We introduce the task of changing the narrative point of view, where\ncharacters are assigned a narrative perspective that is different from the one\noriginally used by the writer. The resulting shift in the narrative point of\nview alters the reading experience and can be used as a tool in fiction writing\nor to generate types of text ranging from educational to self-help and\nself-diagnosis. We introduce a benchmark dataset containing a wide range of\ntypes of narratives annotated with changes in point of view from deictic (first\nor second person) to anaphoric (third person) and describe a pipeline for\nprocessing raw text that relies on a neural architecture for mention selection.\nEvaluations on the new benchmark dataset show that the proposed architecture\nsubstantially outperforms the baselines by generating mentions that are less\nambiguous and more natural.', 'Two interpretations about syllogistic statements are described in this paper.\nOne is the so-called set-based interpretation, which assumes that quantified\nstatements and syllogisms talk about quantity-relationships between sets. The\nother one, the so-called conditional interpretation, assumes that quantified\npropositions talk about conditional propositions and how strong are the links\nbetween the antecedent and the consequent. Both interpretations are compared\nattending to three different questions (existential import, singular statements\nand non-proportional quantifiers) from the point of view of their impact on the\nfurther development of this type of reasoning.', ""There is knowledge. There is belief. And there is tacit agreement.' 'We may\ntalk about objects. We may talk about attributes of the objects. Or we may talk\nboth about objects and their attributes.' This work inspects tacit agreements\non assumptions about the relation between objects and their attributes, and\nstudies a way of expressing them, presenting as the result what we term gradual\nlogic in which the sense of truth gradually shifts. It extends classical logic\ninstances with a new logical connective capturing the object-attribute\nrelation. A formal semantics is presented. Decidability is proved. Para-\nconsistent/epistemic/conditional/intensional/description/combined logics are\ncompared."", ""Probability theory, epistemically interpreted, provides an excellent, if not\nthe best available account of inductive reasoning. This is so because there are\ngeneral and definite rules for the change of subjective probabilities through\ninformation or experience; induction and belief change are one and same topic,\nafter all. The most basic of these rules is simply to conditionalize with\nrespect to the information received; and there are similar and more general\nrules. 1 Hence, a fundamental reason for the epistemological success of\nprobability theory is that there at all exists a well-behaved concept of\nconditional probability. Still, people have, and have reasons for, various\nconcerns over probability theory. One of these is my starting point:\nIntuitively, we have the notion of plain belief; we believe propositions2 to be\ntrue (or to be false or neither). Probability theory, however, offers no formal\ncounterpart to this notion. Believing A is not the same as having probability 1\nfor A, because probability 1 is incorrigible3; but plain belief is clearly\ncorrigible. And believing A is not the same as giving A a probability larger\nthan some 1 - c, because believing A and believing B is usually taken to be\nequivalent to believing A & B.4 Thus, it seems that the formal representation\nof plain belief has to take a non-probabilistic route. Indeed, representing\nplain belief seems easy enough: simply represent an epistemic state by the set\nof all propositions believed true in it or, since I make the common assumption\nthat plain belief is deductively closed, by the conjunction of all propositions\nbelieved true in it. But this does not yet provide a theory of induction, i.e.\nan answer to the question how epistemic states so represented are changed\ntbrough information or experience. There is a convincing partial answer: if the\nnew information is compatible with the old epistemic state, then the new\nepistemic state is simply represented by the conjunction of the new information\nand the old beliefs. This answer is partial because it does not cover the quite\ncommon case where the new information is incompatible with the old beliefs. It\nis, however, important to complete the answer and to cover this case, too;\notherwise, we would not represent plain belief as conigible. The crucial\nproblem is that there is no good completion. When epistemic states are\nrepresented simply by the conjunction of all propositions believed true in it,\nthe answer cannot be completed; and though there is a lot of fruitful work, no\nother representation of epistemic states has been proposed, as far as I know,\nwhich provides a complete solution to this problem. In this paper, I want to\nsuggest such a solution. In [4], I have more fully argued that this is the only\nsolution, if certain plausible desiderata are to be satisfied. Here, in section\n2, I will be content with formally defining and intuitively explaining my\nproposal. I will compare my proposal with probability theory in section 3. It\nwill turn out that the theory I am proposing is structurally homomorphic to\nprobability theory in important respects and that it is thus equally easily\nimplementable, but moreover computationally simpler. Section 4 contains a very\nbrief comparison with various kinds of logics, in particular conditional logic,\nwith Shackle's functions of potential surprise and related theories, and with\nthe Dempster - Shafer theory of belief functions."", ""I am Joachim Jansen and this is my research summary, part of my application\nto the Doctoral Consortium at ICLP'14. I am a PhD student in the Knowledge\nRepresentation and Reasoning (KRR) research group, a subgroup of the\nDeclarative Languages and Artificial Intelligence (DTAI) group at the\ndepartment of Computer Science at KU Leuven. I started my PhD in September\n2012. My promotor is prof. dr. ir. Gerda Janssens and my co-promotor is prof.\ndr. Marc Denecker. I can be contacted at joachim.jansen@cs.kuleuven.be or at:\nRoom 01.167 Celestijnenlaan 200A 3001 Heverlee Belgium An extended abstract /\nfull version of a paper accepted to be presented at the Doctoral Consortium of\nthe 30th International Conference on Logic Programming (ICLP 2014), July 19-22,\nVienna, Austria"", '""Information Processing"" is a recently launched buzzword whose meaning is\nvague and obscure even for the majority of its users. The reason for this is\nthe lack of a suitable definition for the term ""information"". In my attempt to\namend this bizarre situation, I have realized that, following the insights of\nKolmogorov\'s Complexity theory, information can be defined as a description of\nstructures observable in a given data set. Two types of structures could be\neasily distinguished in every data set - in this regard, two types of\ninformation (information descriptions) should be designated: physical\ninformation and semantic information. Kolmogorov\'s theory also posits that the\ninformation descriptions should be provided as a linguistic text structure.\nThis inevitably leads us to an assertion that information processing has to be\nseen as a kind of text processing. The idea is not new - inspired by the\nobservation that human information processing is deeply rooted in natural\nlanguage handling customs, Lotfi Zadeh and his followers have introduced the\nso-called ""Computing With Words"" paradigm. Despite of promotional efforts, the\nidea is not taking off yet. The reason - a lack of a coherent understanding of\nwhat should be called ""information"", and, as a result, misleading research\nroadmaps and objectives. I hope my humble attempt to clarify these issues would\nbe helpful in avoiding common traps and pitfalls.', 'We argue that logical semantics might have faltered due to its failure in\ndistinguishing between two fundamentally very different types of concepts:\nontological concepts, that should be types in a strongly-typed ontology, and\nlogical concepts, that are predicates corresponding to properties of and\nrelations between objects of various ontological types. We will then show that\naccounting for these differences amounts to the integration of lexical and\ncompositional semantics in one coherent framework, and to an embedding in our\nlogical semantics of a strongly-typed ontology that reflects our commonsense\nview of the world and the way we talk about it in ordinary language. We will\nshow that in such a framework a number of challenges in natural language\nsemantics can be adequately and systematically treated.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : It is understood that training-based models can potentially discriminate against individuals in underrepresented minority demographics. Consequently, I am keen to ensure that machine learning and deep learning models do not foster unfair representations of minority groups. This could be achieved either by introducing constraints during training or by eliminating discriminatory behaviors during model inference. Specifically, I am interested in identifying datasets where a traditionally trained model could exhibit discriminatory behavior. I am also seeking to understand specific terminologies and concepts related to this field of study. Moreover, I am looking for comprehensive evaluation benchmarks that can help determine the fairness of a model. My current hypothesis is that if we can pinpoint the decision boundary of a neural model, we can identify where discrimination occurs. If we can then adjust or modify this decision boundary, we can retroactively correct such discriminatory behaviors in models. Ultimately, my goal is to develop machine learning algorithms that are not only accurate but also fair and transparent in their treatment of protected groups.","['Machine learning models are widely adopted in scenarios that directly affect\npeople. The development of software systems based on these models raises\nsocietal and legal concerns, as their decisions may lead to the unfair\ntreatment of individuals based on attributes like race or gender. Data\npreparation is key in any machine learning pipeline, but its effect on fairness\nis yet to be studied in detail. In this paper, we evaluate how the fairness and\neffectiveness of the learned models are affected by the removal of the\nsensitive attribute, the encoding of the categorical attributes, and instance\nselection methods (including cross-validators and random undersampling). We\nused the Adult Income and the German Credit Data datasets, which are widely\nstudied and known to have fairness concerns. We applied each data preparation\ntechnique individually to analyse the difference in predictive performance and\nfairness, using statistical parity difference, disparate impact, and the\nnormalised prejudice index. The results show that fairness is affected by\ntransformations made to the training data, particularly in imbalanced datasets.\nRemoving the sensitive attribute is insufficient to eliminate all the\nunfairness in the predictions, as expected, but it is key to achieve fairer\nmodels. Additionally, the standard random undersampling with respect to the\ntrue labels is sometimes more prejudicial than performing no random\nundersampling.', 'Automated decision making based on big data and machine learning (ML)\nalgorithms can result in discriminatory decisions against certain protected\ngroups defined upon personal data like gender, race, sexual orientation etc.\nSuch algorithms designed to discover patterns in big data might not only pick\nup any encoded societal biases in the training data, but even worse, they might\nreinforce such biases resulting in more severe discrimination. The majority of\nthus far proposed fairness-aware machine learning approaches focus solely on\nthe pre-, in- or post-processing steps of the machine learning process, that\nis, input data, learning algorithms or derived models, respectively. However,\nthe fairness problem cannot be isolated to a single step of the ML process.\nRather, discrimination is often a result of complex interactions between big\ndata and algorithms, and therefore, a more holistic approach is required. The\nproposed FAE (Fairness-Aware Ensemble) framework combines fairness-related\ninterventions at both pre- and postprocessing steps of the data analysis\nprocess. In the preprocessing step, we tackle the problems of\nunder-representation of the protected group (group imbalance) and of\nclass-imbalance by generating balanced training samples. In the post-processing\nstep, we tackle the problem of class overlapping by shifting the decision\nboundary in the direction of fairness.', 'The rise in machine learning-assisted decision-making has led to concerns\nabout the fairness of the decisions and techniques to mitigate problems of\ndiscrimination. If a negative decision is made about an individual (denying a\nloan, rejecting an application for housing, and so on) justice dictates that we\nbe able to ask how we might change circumstances to get a favorable decision\nthe next time. Moreover, the ability to change circumstances (a better\neducation, improved credentials) should not be limited to only those with\naccess to expensive resources. In other words, \\emph{recourse} for negative\ndecisions should be considered a desirable value that can be equalized across\n(demographically defined) groups. This paper describes how to build models that\nmake accurate predictions while still ensuring that the penalties for a\nnegative outcome do not disadvantage different groups disproportionately. We\nmeasure recourse as the distance of an individual from the decision boundary of\na classifier. We then introduce a regularized objective to minimize the\ndifference in recourse across groups. We explore linear settings and further\nextend recourse to non-linear settings as well as model-agnostic settings where\nthe exact distance from boundary cannot be calculated. Our results show that we\ncan successfully decrease the unfairness in recourse while maintaining\nclassifier performance.', 'Machine Learning models have been deployed across many different aspects of\nsociety, often in situations that affect social welfare. Although these models\noffer streamlined solutions to large problems, they may contain biases and\ntreat groups or individuals unfairly based on protected attributes such as\ngender. In this paper, we introduce several examples of machine learning gender\nbias in practice followed by formalizations of fairness. We provide a survey of\nfairness research by detailing influential pre-processing, in-processing, and\npost-processing bias mitigation algorithms. We then propose an end-to-end bias\nmitigation framework, which employs a fusion of pre-, in-, and post-processing\nmethods to leverage the strengths of each individual technique. We test this\nmethod, along with the standard techniques we review, on a deep neural network\nto analyze bias mitigation in a deep learning setting. We find that our\nend-to-end bias mitigation framework outperforms the baselines with respect to\nseveral fairness metrics, suggesting its promise as a method for improving\nfairness. As society increasingly relies on artificial intelligence to help in\ndecision-making, addressing gender biases present in deep learning models is\nimperative. To provide readers with the tools to assess the fairness of machine\nlearning models and mitigate the biases present in them, we discuss multiple\nopen source packages for fairness in AI.', 'The subject of ""fairness"" in artificial intelligence (AI) refers to assessing\nAI algorithms for potential bias based on demographic characteristics such as\nrace and gender, and the development of algorithms to address this bias. Most\napplications to date have been in computer vision, although some work in\nhealthcare has started to emerge. The use of deep learning (DL) in cardiac MR\nsegmentation has led to impressive results in recent years, and such techniques\nare starting to be translated into clinical practice. However, no work has yet\ninvestigated the fairness of such models. In this work, we perform such an\nanalysis for racial/gender groups, focusing on the problem of training data\nimbalance, using a nnU-Net model trained and evaluated on cine short axis\ncardiac MR data from the UK Biobank dataset, consisting of 5,903 subjects from\n6 different racial groups. We find statistically significant differences in\nDice performance between different racial groups. To reduce the racial bias, we\ninvestigated three strategies: (1) stratified batch sampling, in which batch\nsampling is stratified to ensure balance between racial groups; (2) fair\nmeta-learning for segmentation, in which a DL classifier is trained to classify\nrace and jointly optimized with the segmentation model; and (3) protected group\nmodels, in which a different segmentation model is trained for each racial\ngroup. We also compared the results to the scenario where we have a perfectly\nbalanced database. To assess fairness we used the standard deviation (SD) and\nskewed error ratio (SER) of the average Dice values. Our results demonstrate\nthat the racial bias results from the use of imbalanced training data, and that\nall proposed bias mitigation strategies improved fairness, with the best SD and\nSER resulting from the use of protected group models.', 'Fairness in machine learning is crucial when individuals are subject to\nautomated decisions made by models in high-stake domains. Organizations that\nemploy these models may also need to satisfy regulations that promote\nresponsible and ethical A.I. While fairness metrics relying on comparing model\nerror rates across subpopulations have been widely investigated for the\ndetection and mitigation of bias, fairness in terms of the equalized ability to\nachieve recourse for different protected attribute groups has been relatively\nunexplored. We present a novel formulation for training neural networks that\nconsiders the distance of data points to the decision boundary such that the\nnew objective: (1) reduces the average distance to the decision boundary\nbetween two groups for individuals subject to a negative outcome in each group,\ni.e. the network is more fair with respect to the ability to obtain recourse,\nand (2) increases the average distance of data points to the boundary to\npromote adversarial robustness. We demonstrate that training with this loss\nyields more fair and robust neural networks with similar accuracies to models\ntrained without it. Moreover, we qualitatively motivate and empirically show\nthat reducing recourse disparity across groups also improves fairness measures\nthat rely on error rates. To the best of our knowledge, this is the first time\nthat recourse capabilities across groups are considered to train fairer neural\nnetworks, and a relation between error rates based fairness and recourse based\nfairness is investigated.', 'Context: Machine learning software can generate models that inappropriately\ndiscriminate against specific protected social groups (e.g., groups based on\ngender, ethnicity, etc). Motivated by those results, software engineering\nresearchers have proposed many methods for mitigating those discriminatory\neffects. While those methods are effective in mitigating bias, few of them can\nprovide explanations on what is the root cause of bias.\n  Objective: We aim at better detection and mitigation of algorithmic\ndiscrimination in machine learning software problems.\n  Method: Here we propose xFAIR, a model-based extrapolation method, that is\ncapable of both mitigating bias and explaining the cause. In our xFAIR\napproach, protected attributes are represented by models learned from the other\nindependent variables (and these models offer extrapolations over the space\nbetween existing examples). We then use the extrapolation models to relabel\nprotected attributes later seen in testing data or deployment time. Our\napproach aims to offset the biased predictions of the classification model via\nrebalancing the distribution of protected attributes.\n  Results: The experiments of this paper show that, without compromising\n(original) model performance, xFAIR can achieve significantly better group and\nindividual fairness (as measured in different metrics) than benchmark methods.\nMoreover, when compared to another instance-based rebalancing method, our\nmodel-based approach shows faster runtime and thus better scalability.\n  Conclusion: Algorithmic decision bias can be removed via extrapolation that\nsmooths away outlier points. As evidence for this, our proposed xFAIR is not\nonly performance-wise better (measured by fairness and performance metrics)\nthan two state-of-the-art fairness algorithms.', 'Fairness is becoming a rising concern w.r.t. machine learning model\nperformance. Especially for sensitive fields such as criminal justice and loan\ndecision, eliminating the prediction discrimination towards a certain group of\npopulation (characterized by sensitive features like race and gender) is\nimportant for enhancing the trustworthiness of model. In this paper, we present\na new general framework to improve machine learning fairness. The goal of our\nmodel is to minimize the influence of sensitive feature from the perspectives\nof both the data input and the predictive model. In order to achieve this goal,\nwe reformulate the data input by removing the sensitive information and\nstrengthen model fairness by minimizing the marginal contribution of the\nsensitive feature. We propose to learn the non-sensitive input via sampling\namong features and design an adversarial network to minimize the dependence\nbetween the reformulated input and the sensitive information. Extensive\nexperiments on three benchmark datasets suggest that our model achieve better\nresults than related state-of-the-art methods with respect to both fairness\nmetrics and prediction performance.', ""Data-driven algorithms play a large role in decision making across a variety\nof industries. Increasingly, these algorithms are being used to make decisions\nthat have significant ramifications for people's social and economic\nwell-being, e.g. in sentencing, loan approval, and policing. Amid the\nproliferation of such systems there is a growing concern about their potential\ndiscriminatory impact. In particular, machine learning systems which are\ntrained on biased data have the potential to learn and perpetuate those biases.\nA central challenge for practitioners is thus to determine whether their models\ndisplay discriminatory bias. Here we present a case study in which we frame the\nissue of bias detection as a causal inference problem with observational data.\nWe enumerate two main causes of bias, sampling bias and label bias, and we\ninvestigate the abilities of six different fairness metrics to detect each bias\ntype. Based on these investigations, we propose a set of best practice\nguidelines to select the fairness metric that is most likely to detect bias if\nit is present. Additionally, we aim to identify the conditions in which certain\nfairness metrics may fail to detect bias and instead give practitioners a false\nbelief that their biased model is making fair decisions."", 'The widespread use of ML-based decision making in domains with high societal\nimpact such as recidivism, job hiring and loan credit has raised a lot of\nconcerns regarding potential discrimination. In particular, in certain cases it\nhas been observed that ML algorithms can provide different decisions based on\nsensitive attributes such as gender or race and therefore can lead to\ndiscrimination. Although, several fairness-aware ML approaches have been\nproposed, their focus has been largely on preserving the overall classification\naccuracy while improving fairness in predictions for both protected and\nnon-protected groups (defined based on the sensitive attribute(s)). The overall\naccuracy however is not a good indicator of performance in case of class\nimbalance, as it is biased towards the majority class. As we will see in our\nexperiments, many of the fairness-related datasets suffer from class imbalance\nand therefore, tackling fairness requires also tackling the imbalance problem.\n  To this end, we propose AdaFair, a fairness-aware classifier based on\nAdaBoost that further updates the weights of the instances in each boosting\nround taking into account a cumulative notion of fairness based upon all\ncurrent ensemble members, while explicitly tackling class-imbalance by\noptimizing the number of ensemble members for balanced classification error.\nOur experiments show that our approach can achieve parity in true positive and\ntrue negative rates for both protected and non-protected groups, while it\nsignificantly outperforms existing fairness-aware methods up to 25% in terms of\nbalanced error.', 'In recent years, automated data-driven decision-making systems have enjoyed a\ntremendous success in a variety of fields (e.g., to make product\nrecommendations, or to guide the production of entertainment). More recently,\nthese algorithms are increasingly being used to assist socially sensitive\ndecision-making (e.g., to decide who to admit into a degree program or to\nprioritize individuals for public housing). Yet, these automated tools may\nresult in discriminative decision-making in the sense that they may treat\nindividuals unfairly or unequally based on membership to a category or a\nminority, resulting in disparate treatment or disparate impact and violating\nboth moral and ethical standards. This may happen when the training dataset is\nitself biased (e.g., if individuals belonging to a particular group have\nhistorically been discriminated upon). However, it may also happen when the\ntraining dataset is unbiased, if the errors made by the system affect\nindividuals belonging to a category or minority differently (e.g., if\nmisclassification rates for Blacks are higher than for Whites). In this paper,\nwe unify the definitions of unfairness across classification and regression. We\npropose a versatile mixed-integer optimization framework for learning optimal\nand fair decision trees and variants thereof to prevent disparate treatment\nand/or disparate impact as appropriate. This translates to a flexible schema\nfor designing fair and interpretable policies suitable for socially sensitive\ndecision-making. We conduct extensive computational studies that show that our\nframework improves the state-of-the-art in the field (which typically relies on\nheuristics) to yield non-discriminative decisions at lower cost to overall\naccuracy.', 'Deep learning is increasingly being used in high-stake decision making\napplications that affect individual lives. However, deep learning models might\nexhibit algorithmic discrimination behaviors with respect to protected groups,\npotentially posing negative impacts on individuals and society. Therefore,\nfairness in deep learning has attracted tremendous attention recently. We\nprovide a review covering recent progresses to tackle algorithmic fairness\nproblems of deep learning from the computational perspective. Specifically, we\nshow that interpretability can serve as a useful ingredient to diagnose the\nreasons that lead to algorithmic discrimination. We also discuss fairness\nmitigation approaches categorized according to three stages of deep learning\nlife-cycle, aiming to push forward the area of fairness in deep learning and\nbuild genuinely fair and reliable deep learning systems.', 'Machine learning algorithms are increasingly used for consequential decision\nmaking regarding individuals based on their relevant features. Features that\nare relevant for accurate decisions may however lead to either explicit or\nimplicit forms of discrimination against unprivileged groups, such as those of\ncertain race or gender. This happens due to existing biases in the training\ndata, which are often replicated or even exacerbated by the learning algorithm.\nIdentifying and measuring these biases at the data level is a challenging\nproblem due to the interdependence among the features, and the decision\noutcome. In this work, we develop a framework for fairness-aware feature\nselection which takes into account the correlation among the features and the\ndecision outcome, and is based on information theoretic measures for the\naccuracy and discriminatory impacts of features. In particular, we first\npropose information theoretic measures which quantify the impact of different\nsubsets of features on the accuracy and discrimination of the decision\noutcomes. We then deduce the marginal impact of each feature using Shapley\nvalue function; a solution concept in cooperative game theory used to estimate\nmarginal contributions of players in a coalitional game. Finally, we design a\nfairness utility score for each feature (for feature selection) which\nquantifies how this feature influences accurate as well as nondiscriminatory\ndecisions. Our framework depends on the joint statistics of the data rather\nthan a particular classifier design. We examine our proposed framework on real\nand synthetic data to evaluate its performance.', 'Data-driven decision making is gaining prominence with the popularity of\nvarious machine learning models. Unfortunately, real-life data used in machine\nlearning training may capture human biases, and as a result the learned models\nmay lead to unfair decision making. In this paper, we provide a solution to\nthis problem for decision trees and random forests. Our approach converts any\ndecision tree or random forest into a fair one with respect to a specific data\nset, fairness criteria, and sensitive attributes. The \\emph{FairRepair} tool,\nbuilt based on our approach, is inspired by automated program repair techniques\nfor traditional programs. It uses an SMT solver to decide which paths in the\ndecision tree could have their outcomes flipped to improve the fairness of the\nmodel. Our experiments on the well-known adult dataset from UC Irvine\ndemonstrate that FairRepair scales to realistic decision trees and random\nforests. Furthermore, FairRepair provides formal guarantees about soundness and\ncompleteness of finding a repair. Since our fairness-guided repair technique\nrepairs decision trees and random forests obtained from a given (unfair)\ndata-set, it can help to identify and rectify biases in decision-making in an\norganisation.', 'Algorithmic decision making based on computer vision and machine learning\ntechnologies continue to permeate our lives. But issues related to biases of\nthese models and the extent to which they treat certain segments of the\npopulation unfairly, have led to concern in the general public. It is now\naccepted that because of biases in the datasets we present to the models, a\nfairness-oblivious training will lead to unfair models. An interesting topic is\nthe study of mechanisms via which the de novo design or training of the model\ncan be informed by fairness measures. Here, we study mechanisms that impose\nfairness concurrently while training the model. While existing fairness based\napproaches in vision have largely relied on training adversarial modules\ntogether with the primary classification/regression task, in an effort to\nremove the influence of the protected attribute or variable, we show how ideas\nbased on well-known optimization concepts can provide a simpler alternative. In\nour proposed scheme, imposing fairness just requires specifying the protected\nattribute and utilizing our optimization routine. We provide a detailed\ntechnical analysis and present experiments demonstrating that various fairness\nmeasures from the literature can be reliably imposed on a number of training\ntasks in vision in a manner that is interpretable.', 'Machine learning models are increasingly being used in important\ndecision-making software such as approving bank loans, recommending criminal\nsentencing, hiring employees, and so on. It is important to ensure the fairness\nof these models so that no discrimination is made based on protected attribute\n(e.g., race, sex, age) while decision making. Algorithms have been developed to\nmeasure unfairness and mitigate them to a certain extent. In this paper, we\nhave focused on the empirical evaluation of fairness and mitigations on\nreal-world machine learning models. We have created a benchmark of 40 top-rated\nmodels from Kaggle used for 5 different tasks, and then using a comprehensive\nset of fairness metrics, evaluated their fairness. Then, we have applied 7\nmitigation techniques on these models and analyzed the fairness, mitigation\nresults, and impacts on performance. We have found that some model optimization\ntechniques result in inducing unfairness in the models. On the other hand,\nalthough there are some fairness control mechanisms in machine learning\nlibraries, they are not documented. The mitigation algorithm also exhibit\ncommon patterns such as mitigation in the post-processing is often costly (in\nterms of performance) and mitigation in the pre-processing stage is preferred\nin most cases. We have also presented different trade-off choices of fairness\nmitigation decisions. Our study suggests future research directions to reduce\nthe gap between theoretical fairness aware algorithms and the software\nengineering methods to leverage them in practice.', 'Machine Learning (ML) models trained on data from multiple demographic groups\ncan inherit representation disparity (Hashimoto et al., 2018) that may exist in\nthe data: the model may be less favorable to groups contributing less to the\ntraining process; this in turn can degrade population retention in these groups\nover time, and exacerbate representation disparity in the long run. In this\nstudy, we seek to understand the interplay between ML decisions and the\nunderlying group representation, how they evolve in a sequential framework, and\nhow the use of fairness criteria plays a role in this process. We show that the\nrepresentation disparity can easily worsen over time under a natural user\ndynamics (arrival and departure) model when decisions are made based on a\ncommonly used objective and fairness criteria, resulting in some groups\ndiminishing entirely from the sample pool in the long run. It highlights the\nfact that fairness criteria have to be defined while taking into consideration\nthe impact of decisions on user dynamics. Toward this end, we explain how a\nproper fairness criterion can be selected based on a general user dynamics\nmodel.', 'As more industries integrate machine learning into socially sensitive\ndecision processes like hiring, loan-approval, and parole-granting, we are at\nrisk of perpetuating historical and contemporary socioeconomic disparities.\nThis is a critical problem because on the one hand, organizations who use but\ndo not understand the discriminatory potential of such systems will facilitate\nthe widening of social disparities under the assumption that algorithms are\ncategorically objective. On the other hand, the responsible use of machine\nlearning can help us measure, understand, and mitigate the implicit historical\nbiases in socially sensitive data by expressing implicit decision-making mental\nmodels in terms of explicit statistical models. In this paper we specify,\nimplement, and evaluate a ""fairness-aware"" machine learning interface called\nthemis-ml, which is intended for use by individual data scientists and\nengineers, academic research teams, or larger product teams who use machine\nlearning in production systems.', 'Discrimination is a focal concern in supervised learning algorithms\naugmenting human decision-making. These systems are trained using historical\ndata, which may have been tainted by discrimination, and may learn biases\nagainst the protected groups. An important question is how to train models\nwithout propagating discrimination. In this study, we i) define and model\ndiscrimination as perturbations of a data-generating process and show how\ndiscrimination can be induced via attributes correlated with the protected\nattributes; ii) introduce a measure of resilience of a supervised learning\nalgorithm to potentially discriminatory data perturbations, iii) propose a\nnovel supervised learning algorithm that inhibits discrimination, and iv) show\nthat it is more resilient to discriminatory perturbations in synthetic and\nreal-world datasets than state-of-the-art learning algorithms. The proposed\nmethod can be used with general supervised learning algorithms and avoids\ninducement of discrimination, while maximizing model accuracy.', 'Artificial neural networks perform state-of-the-art in an ever-growing number\nof tasks, nowadays they are used to solve an incredibly large variety of tasks.\nHowever, typical training strategies do not take into account lawful, ethical\nand discriminatory potential issues the trained ANN models could incur in. In\nthis work we propose NDR, a non-discriminatory regularization strategy to\nprevent the ANN model to solve the target task using some discriminatory\nfeatures like, for example, the ethnicity in an image classification task for\nhuman faces. In particular, a part of the ANN model is trained to hide the\ndiscriminatory information such that the rest of the network focuses in\nlearning the given learning task. Our experiments show that NDR can be\nexploited to achieve non-discriminatory models with both minimal computational\noverhead and performance loss.', 'Automated decision making systems are increasingly being used in real-world\napplications. In these systems for the most part, the decision rules are\nderived by minimizing the training error on the available historical data.\nTherefore, if there is a bias related to a sensitive attribute such as gender,\nrace, religion, etc. in the data, say, due to cultural/historical\ndiscriminatory practices against a certain demographic, the system could\ncontinue discrimination in decisions by including the said bias in its decision\nrule. We present an information theoretic framework for designing fair\npredictors from data, which aim to prevent discrimination against a specified\nsensitive attribute in a supervised learning setting. We use equalized odds as\nthe criterion for discrimination, which demands that the prediction should be\nindependent of the protected attribute conditioned on the actual label. To\nensure fairness and generalization simultaneously, we compress the data to an\nauxiliary variable, which is used for the prediction task. This auxiliary\nvariable is chosen such that it is decontaminated from the discriminatory\nattribute in the sense of equalized odds. The final predictor is obtained by\napplying a Bayesian decision rule to the auxiliary variable.', 'Artificial intelligence nowadays plays an increasingly prominent role in our\nlife since decisions that were once made by humans are now delegated to\nautomated systems. A machine learning algorithm trained based on biased data,\nhowever, tends to make unfair predictions. Developing classification algorithms\nthat are fair with respect to protected attributes of the data thus becomes an\nimportant problem. Motivated by concerns surrounding the fairness effects of\nsharing and few-shot machine learning tools, such as the Model Agnostic\nMeta-Learning framework, we propose a novel fair fast-adapted few-shot\nmeta-learning approach that efficiently mitigates biases during meta-train by\nensuring controlling the decision boundary covariance that between the\nprotected variable and the signed distance from the feature vectors to the\ndecision boundary. Through extensive experiments on two real-world image\nbenchmarks over three state-of-the-art meta-learning algorithms, we empirically\ndemonstrate that our proposed approach efficiently mitigates biases on model\noutput and generalizes both accuracy and fairness to unseen tasks with a\nlimited amount of training samples.', 'Considerable research effort has been guided towards algorithmic fairness but\nreal-world adoption of bias reduction techniques is still scarce. Existing\nmethods are either metric- or model-specific, require access to sensitive\nattributes at inference time, or carry high development or deployment costs.\nThis work explores the unfairness that emerges when optimizing ML models solely\nfor predictive performance, and how to mitigate it with a simple and easily\ndeployed intervention: fairness-aware hyperparameter optimization (HO). We\npropose and evaluate fairness-aware variants of three popular HO algorithms:\nFair Random Search, Fair TPE, and Fairband. We validate our approach on a\nreal-world bank account opening fraud case-study, as well as on three datasets\nfrom the fairness literature. Results show that, without extra training cost,\nit is feasible to find models with 111% mean fairness increase and just 6%\ndecrease in performance when compared with fairness-blind HO.', 'Machine learning models are becoming pervasive in high-stakes applications.\nDespite their clear benefits in terms of performance, the models could show\ndiscrimination against minority groups and result in fairness issues in a\ndecision-making process, leading to severe negative impacts on the individuals\nand the society. In recent years, various techniques have been developed to\nmitigate the unfairness for machine learning models. Among them, in-processing\nmethods have drawn increasing attention from the community, where fairness is\ndirectly taken into consideration during model design to induce intrinsically\nfair models and fundamentally mitigate fairness issues in outputs and\nrepresentations. In this survey, we review the current progress of\nin-processing fairness mitigation techniques. Based on where the fairness is\nachieved in the model, we categorize them into explicit and implicit methods,\nwhere the former directly incorporates fairness metrics in training objectives,\nand the latter focuses on refining latent representation learning. Finally, we\nconclude the survey with a discussion of the research challenges in this\ncommunity to motivate future exploration.', 'In recent years, machine learning techniques have been increasingly applied\nin sensitive decision making processes, raising fairness concerns. Past\nresearch has shown that machine learning may reproduce and even exacerbate\nhuman bias due to biased training data or flawed model assumptions, and thus\nmay lead to discriminatory actions. To counteract such biased models,\nresearchers have proposed multiple mathematical definitions of fairness\naccording to which classifiers can be optimized. However, it has also been\nshown that the outcomes generated by some fairness notions may be\nunsatisfactory.\n  In this contribution, we add to this research by considering decision making\nprocesses in time. We establish a theoretic model in which even perfectly\naccurate classifiers which adhere to almost all common fairness definitions\nlead to stable long-term inequalities due to vicious cycles. Only demographic\nparity, which enforces equal rates of positive decisions across groups, avoids\nthese effects and establishes a virtuous cycle, which leads to perfectly\naccurate and fair classification in the long term.', 'Recent work in fairness in machine learning has proposed adjusting for\nfairness by equalizing accuracy metrics across groups and has also studied how\ndatasets affected by historical prejudices may lead to unfair decision\npolicies. We connect these lines of work and study the residual unfairness that\narises when a fairness-adjusted predictor is not actually fair on the target\npopulation due to systematic censoring of training data by existing biased\npolicies. This scenario is particularly common in the same applications where\nfairness is a concern. We characterize theoretically the impact of such\ncensoring on standard fairness metrics for binary classifiers and provide\ncriteria for when residual unfairness may or may not appear. We prove that,\nunder certain conditions, fairness-adjusted classifiers will in fact induce\nresidual unfairness that perpetuates the same injustices, against the same\ngroups, that biased the data to begin with, thus showing that even\nstate-of-the-art fair machine learning can have a ""bias in, bias out"" property.\nWhen certain benchmark data is available, we show how sample reweighting can\nestimate and adjust fairness metrics while accounting for censoring. We use\nthis to study the case of Stop, Question, and Frisk (SQF) and demonstrate that\nattempting to adjust for fairness perpetuates the same injustices that the\npolicy is infamous for.', 'Explicit and implicit bias clouds human judgement, leading to discriminatory\ntreatment of minority groups. A fundamental goal of algorithmic fairness is to\navoid the pitfalls in human judgement by learning policies that improve the\noverall outcomes while providing fair treatment to protected classes. In this\npaper, we propose a causal framework that learns optimal intervention policies\nfrom data subject to fairness constraints. We define two measures of treatment\nbias and infer best treatment assignment that minimizes the bias while\noptimizing overall outcome. We demonstrate that there is a dilemma of balancing\nfairness and overall benefit; however, allowing preferential treatment to\nprotected classes in certain circumstances (affirmative action) can\ndramatically improve the overall benefit while also preserving fairness. We\napply our framework to data containing student outcomes on standardized tests\nand show how it can be used to design real-world policies that fairly improve\nstudent test scores. Our framework provides a principled way to learn fair\ntreatment policies in real-world settings.', ""Software often produces biased outputs. In particular, machine learning (ML)\nbased software are known to produce erroneous predictions when processing\ndiscriminatory inputs. Such unfair program behavior can be caused by societal\nbias. In the last few years, Amazon, Microsoft and Google have provided\nsoftware services that produce unfair outputs, mostly due to societal bias\n(e.g. gender or race). In such events, developers are saddled with the task of\nconducting fairness testing. Fairness testing is challenging; developers are\ntasked with generating discriminatory inputs that reveal and explain biases.\n  We propose a grammar-based fairness testing approach (called ASTRAEA) which\nleverages context-free grammars to generate discriminatory inputs that reveal\nfairness violations in software systems. Using probabilistic grammars, ASTRAEA\nalso provides fault diagnosis by isolating the cause of observed software bias.\nASTRAEA's diagnoses facilitate the improvement of ML fairness.\n  ASTRAEA was evaluated on 18 software systems that provide three major natural\nlanguage processing (NLP) services. In our evaluation, ASTRAEA generated\nfairness violations with a rate of ~18%. ASTRAEA generated over 573K\ndiscriminatory test cases and found over 102K fairness violations. Furthermore,\nASTRAEA improves software fairness by ~76%, via model-retraining."", 'Training classifiers under fairness constraints such as group fairness,\nregularizes the disparities of predictions between the groups. Nevertheless,\neven though the constraints are satisfied during training, they might not\ngeneralize at evaluation time. To improve the generalizability of fair\nclassifiers, we propose fair mixup, a new data augmentation strategy for\nimposing the fairness constraint. In particular, we show that fairness can be\nachieved by regularizing the models on paths of interpolated samples between\nthe groups. We use mixup, a powerful data augmentation strategy to generate\nthese interpolates. We analyze fair mixup and empirically show that it ensures\na better generalization for both accuracy and fairness measurement in tabular,\nvision, and language benchmarks.', 'The discovery of discriminatory bias in human or automated decision making is\na task of increasing importance and difficulty, exacerbated by the pervasive\nuse of machine learning and data mining. Currently, discrimination discovery\nlargely relies upon correlation analysis of decisions records, disregarding the\nimpact of confounding biases. We present a method for causal discrimination\ndiscovery based on propensity score analysis, a statistical tool for filtering\nout the effect of confounding variables. We introduce causal measures of\ndiscrimination which quantify the effect of group membership on the decisions,\nand highlight causal discrimination/favoritism patterns by learning regression\ntrees over the novel measures. We validate our approach on two real world\ndatasets. Our proposed framework for causal discrimination has the potential to\nenhance the transparency of machine learning with tools for detecting\ndiscriminatory bias both in the training data and in the learning algorithms.', 'Classification, a heavily-studied data-driven machine learning task, drives\nan increasing number of prediction systems involving critical human decisions\nsuch as loan approval and criminal risk assessment. However, classifiers often\ndemonstrate discriminatory behavior, especially when presented with biased\ndata. Consequently, fairness in classification has emerged as a high-priority\nresearch area. Data management research is showing an increasing presence and\ninterest in topics related to data and algorithmic fairness, including the\ntopic of fair classification. The interdisciplinary efforts in fair\nclassification, with machine learning research having the largest presence,\nhave resulted in a large number of fairness notions and a wide range of\napproaches that have not been systematically evaluated and compared. In this\npaper, we contribute a broad analysis of 13 fair classification approaches and\nadditional variants, over their correctness, fairness, efficiency, scalability,\nrobustness to data errors, sensitivity to underlying ML model, data efficiency,\nand stability using a variety of metrics and real-world datasets. Our analysis\nhighlights novel insights on the impact of different metrics and high-level\napproach characteristics on different aspects of performance. We also discuss\ngeneral principles for choosing approaches suitable for different practical\nsettings, and identify areas where data-management-centric solutions are likely\nto have the most impact.', ""Recent research has helped to cultivate growing awareness that machine\nlearning systems fueled by big data can create or exacerbate troubling\ndisparities in society. Much of this research comes from outside of the\npracticing data science community, leaving its members with little concrete\nguidance to proactively address these concerns. This article introduces issues\nof discrimination to the data science community on its own terms. In it, we\ntour the familiar data mining process while providing a taxonomy of common\npractices that have the potential to produce unintended discrimination. We also\nsurvey how discrimination is commonly measured, and suggest how familiar\ndevelopment processes can be augmented to mitigate systems' discriminatory\npotential. We advocate that data scientists should be intentional about\nmodeling and reducing discriminatory outcomes. Without doing so, their efforts\nwill result in perpetuating any systemic discrimination that may exist, but\nunder a misleading veil of data-driven objectivity."", 'We study fairness in supervised few-shot meta-learning models that are\nsensitive to discrimination (or bias) in historical data. A machine learning\nmodel trained based on biased data tends to make unfair predictions for users\nfrom minority groups. Although this problem has been studied before, existing\nmethods mainly aim to detect and control the dependency effect of the protected\nvariables (e.g. race, gender) on target prediction based on a large amount of\ntraining data. These approaches carry two major drawbacks that (1) lacking\nshowing a global cause-effect visualization for all variables; (2) lacking\ngeneralization of both accuracy and fairness to unseen tasks. In this work, we\nfirst discover discrimination from data using a causal Bayesian knowledge graph\nwhich not only demonstrates the dependency of the protected variable on target\nbut also indicates causal effects between all variables. Next, we develop a\nnovel algorithm based on risk difference in order to quantify the\ndiscriminatory influence for each protected variable in the graph. Furthermore,\nto protect prediction from unfairness, a fast-adapted bias-control approach in\nmeta-learning is proposed, which efficiently mitigates statistical disparity\nfor each task and it thus ensures independence of protected attributes on\npredictions based on biased and few-shot data samples. Distinct from existing\nmeta-learning models, group unfairness of tasks are efficiently reduced by\nleveraging the mean difference between (un)protected groups for regression\nproblems. Through extensive experiments on both synthetic and real-world data\nsets, we demonstrate that our proposed unfairness discovery and prevention\napproaches efficiently detect discrimination and mitigate biases on model\noutput as well as generalize both accuracy and fairness to unseen tasks with a\nlimited amount of training samples.', 'Nowadays, many decisions are made using predictive models built on historical\ndata.Predictive models may systematically discriminate groups of people even if\nthe computing process is fair and well-intentioned. Discrimination-aware data\nmining studies how to make predictive models free from discrimination, when\nhistorical data, on which they are built, may be biased, incomplete, or even\ncontain past discriminatory decisions. Discrimination refers to disadvantageous\ntreatment of a person based on belonging to a category rather than on\nindividual merit. In this survey we review and organize various discrimination\nmeasures that have been used for measuring discrimination in data, as well as\nin evaluating performance of discrimination-aware predictive models. We also\ndiscuss related measures from other disciplines, which have not been used for\nmeasuring discrimination, but potentially could be suitable for this purpose.\nWe computationally analyze properties of selected measures. We also review and\ndiscuss measuring procedures, and present recommendations for practitioners.\nThe primary target audience is data mining, machine learning, pattern\nrecognition, statistical modeling researchers developing new methods for\nnon-discriminatory predictive modeling. In addition, practitioners and policy\nmakers would use the survey for diagnosing potential discrimination by\npredictive models.', ""Although deep learning has demonstrated astonishing performance in many\napplications, there are still concerns about its dependability. One desirable\nproperty of deep learning applications with societal impact is fairness (i.e.,\nnon-discrimination). Unfortunately, discrimination might be intrinsically\nembedded into the models due to the discrimination in the training data. As a\ncountermeasure, fairness testing systemically identifies discriminatory\nsamples, which can be used to retrain the model and improve the model's\nfairness. Existing fairness testing approaches however have two major\nlimitations. Firstly, they only work well on traditional machine learning\nmodels and have poor performance (e.g., effectiveness and efficiency) on deep\nlearning models. Secondly, they only work on simple structured (e.g., tabular)\ndata and are not applicable for domains such as text. In this work, we bridge\nthe gap by proposing a scalable and effective approach for systematically\nsearching for discriminatory samples while extending existing fairness testing\napproaches to address a more challenging domain, i.e., text classification.\nCompared with state-of-the-art methods, our approach only employs lightweight\nprocedures like gradient computation and clustering, which is significantly\nmore scalable and effective. Experimental results show that on average, our\napproach explores the search space much more effectively (9.62 and 2.38 times\nmore than the state-of-the-art methods respectively on tabular and text\ndatasets) and generates much more discriminatory samples (24.95 and 2.68 times)\nwithin a same reasonable time. Moreover, the retrained models reduce\ndiscrimination by 57.2% and 60.2% respectively on average."", 'The Fairness, Accountability, and Transparency in Machine Learning (FAT-ML)\nliterature proposes a varied set of group fairness metrics to measure\ndiscrimination against socio-demographic groups that are characterized by a\nprotected feature, such as gender or race.Such a system can be deemed as either\nfair or unfair depending on the choice of the metric. Several metrics have been\nproposed, some of them incompatible with each other.We do so empirically, by\nobserving that several of these metrics cluster together in two or three main\nclusters for the same groups and machine learning methods. In addition, we\npropose a robust way to visualize multidimensional fairness in two dimensions\nthrough a Principal Component Analysis (PCA) of the group fairness metrics.\nExperimental results on multiple datasets show that the PCA decomposition\nexplains the variance between the metrics with one to three components.', 'Machine learning models are extensively being used to make decisions that\nhave a significant impact on human life. These models are trained over\nhistorical data that may contain information about sensitive attributes such as\nrace, sex, religion, etc. The presence of such sensitive attributes can impact\ncertain population subgroups unfairly. It is straightforward to remove\nsensitive features from the data; however, a model could pick up prejudice from\nlatent sensitive attributes that may exist in the training data. This has led\nto the growing apprehension about the fairness of the employed models. In this\npaper, we propose a novel algorithm that can effectively identify and treat\nlatent discriminating features. The approach is agnostic of the learning\nalgorithm and generalizes well for classification as well as regression tasks.\nIt can also be used as a key aid in proving that the model is free of\ndiscrimination towards regulatory compliance if the need arises. The approach\nhelps to collect discrimination-free features that would improve the model\nperformance while ensuring the fairness of the model. The experimental results\nfrom our evaluations on publicly available real-world datasets show a\nnear-ideal fairness measurement in comparison to other methods.', ""Machine learning and data mining algorithms have been increasingly used\nrecently to support decision-making systems in many areas of high societal\nimportance such as healthcare, education, or security. While being very\nefficient in their predictive abilities, the deployed algorithms sometimes tend\nto learn an inductive model with a discriminative bias due to the presence of\nthis latter in the learning sample. This problem gave rise to a new field of\nalgorithmic fairness where the goal is to correct the discriminative bias\nintroduced by a certain attribute in order to decorrelate it from the model's\noutput. In this paper, we study the problem of fairness for the task of edge\nprediction in graphs, a largely underinvestigated scenario compared to a more\npopular setting of fair classification. To this end, we formulate the problem\nof fair edge prediction, analyze it theoretically, and propose an\nembedding-agnostic repairing procedure for the adjacency matrix of an arbitrary\ngraph with a trade-off between the group and individual fairness. We\nexperimentally show the versatility of our approach and its capacity to provide\nexplicit control over different notions of fairness and prediction accuracy."", 'Effective machine learning models can automatically learn useful information\nfrom a large quantity of data and provide decisions in a high accuracy. These\nmodels may, however, lead to unfair predictions in certain sense among the\npopulation groups of interest, where the grouping is based on such sensitive\nattributes as race and gender. Various fairness definitions, such as\ndemographic parity and equalized odds, were proposed in prior art to ensure\nthat decisions guided by the machine learning models are equitable.\nUnfortunately, the ""fair"" model trained with these fairness definitions is\nthreshold sensitive, i.e., the condition of fairness may no longer hold true\nwhen tuning the decision threshold. This paper introduces the notion of\nthreshold invariant fairness, which enforces equitable performances across\ndifferent groups independent of the decision threshold. To achieve this goal,\nthis paper proposes to equalize the risk distributions among the groups via two\napproximation methods. Experimental results demonstrate that the proposed\nmethodology is effective to alleviate the threshold sensitivity in machine\nlearning models designed to achieve fairness.', 'Growing use of machine learning in policy and social impact settings have\nraised concerns for fairness implications, especially for racial minorities.\nThese concerns have generated considerable interest among machine learning and\nartificial intelligence researchers, who have developed new methods and\nestablished theoretical bounds for improving fairness, focusing on the source\ndata, regularization and model training, or post-hoc adjustments to model\nscores. However, little work has studied the practical trade-offs between\nfairness and accuracy in real-world settings to understand how these bounds and\nmethods translate into policy choices and impact on society. Our empirical\nstudy fills this gap by investigating the impact of mitigating disparities on\naccuracy, focusing on the common context of using machine learning to inform\nbenefit allocation in resource-constrained programs across education, mental\nhealth, criminal justice, and housing safety. Here we describe applied work in\nwhich we find fairness-accuracy trade-offs to be negligible in practice. In\neach setting studied, explicitly focusing on achieving equity and using our\nproposed post-hoc disparity mitigation methods, fairness was substantially\nimproved without sacrificing accuracy. This observation was robust across\npolicy contexts studied, scale of resources available for intervention, time,\nand relative size of the protected groups. These empirical results challenge a\ncommonly held assumption that reducing disparities either requires accepting an\nappreciable drop in accuracy or the development of novel, complex methods,\nmaking reducing disparities in these applications more practical.', 'Fair machine learning is receiving an increasing attention in machine\nlearning fields. Researchers in fair learning have developed correlation or\nassociation-based measures such as demographic disparity, mistreatment\ndisparity, calibration, causal-based measures such as total effect, direct and\nindirect discrimination, and counterfactual fairness, and fairness notions such\nas equality of opportunity and equal odds that consider both decisions in the\ntraining data and decisions made by predictive models. In this paper, we\ndevelop a new causal-based fairness notation, called equality of effort.\nDifferent from existing fairness notions which mainly focus on discovering the\ndisparity of decisions between two groups of individuals, the proposed equality\nof effort notation helps answer questions like to what extend a legitimate\nvariable should change to make a particular individual achieve a certain\noutcome level and addresses the concerns whether the efforts made to achieve\nthe same outcome level for individuals from the protected group and that from\nthe unprotected group are different. We develop algorithms for determining\nwhether an individual or a group of individuals is discriminated in terms of\nequality of effort. We also develop an optimization-based method for removing\ndiscriminatory effects from the data if discrimination is detected. We conduct\nempirical evaluations to compare the equality of effort and existing fairness\nnotion and show the effectiveness of our proposed algorithms.', 'Representation Learning in a heterogeneous space with mixed variables of\nnumerical and categorical types has interesting challenges due to its complex\nfeature manifold. Moreover, feature learning in an unsupervised setup, without\nclass labels and a suitable learning loss function, adds to the problem\ncomplexity. Further, the learned representation and subsequent predictions\nshould not reflect discriminatory behavior towards certain sensitive groups or\nattributes. The proposed feature map should preserve maximum variations present\nin the data and needs to be fair with respect to the sensitive variables. We\npropose, in the first phase of our work, an efficient encoder-decoder framework\nto capture the mixed-domain information. The second phase of our work focuses\non de-biasing the mixed space representations by adding relevant fairness\nconstraints. This ensures minimal information loss between the representations\nbefore and after the fairness-preserving projections. Both the information\ncontent and the fairness aspect of the final representation learned has been\nvalidated through several metrics where it shows excellent performance. Our\nwork (FairMixRep) addresses the problem of Mixed Space Fair Representation\nlearning from an unsupervised perspective and learns a Universal representation\nthat is timely, unique, and a novel research contribution.', 'We study fairness in collaborative-filtering recommender systems, which are\nsensitive to discrimination that exists in historical data. Biased data can\nlead collaborative filtering methods to make unfair predictions against\nminority groups of users. We identify the insufficiency of existing fairness\nmetrics and propose four new metrics that address different forms of\nunfairness. These fairness metrics can be optimized by adding fairness terms to\nthe learning objective. Experiments on synthetic and real data show that our\nnew metrics can better measure fairness than the baseline, and that the\nfairness objectives effectively help reduce unfairness.', 'Thanks to the increasing growth of computational power and data availability,\nthe research in machine learning has advanced with tremendous rapidity.\nNowadays, the majority of automatic decision making systems are based on data.\nHowever, it is well known that machine learning systems can present problematic\nresults if they are built on partial or incomplete data. In fact, in recent\nyears several studies have found a convergence of issues related to the ethics\nand transparency of these systems in the process of data collection and how\nthey are recorded. Although the process of rigorous data collection and\nanalysis is fundamental in the model design, this step is still largely\noverlooked by the machine learning community. For this reason, we propose a\nmethod of data annotation based on Bayesian statistical inference that aims to\nwarn about the risk of discriminatory results of a given data set. In\nparticular, our method aims to deepen knowledge and promote awareness about the\nsampling practices employed to create the training set, highlighting that the\nprobability of success or failure conditioned to a minority membership is given\nby the structure of the data available. We empirically test our system on three\ndatasets commonly accessed by the machine learning community and we investigate\nthe risk of racial discrimination.', ""Machine learning is being integrated into a growing number of critical\nsystems with far-reaching impacts on society. Unexpected behaviour and unfair\ndecision processes are coming under increasing scrutiny due to this widespread\nuse and its theoretical considerations. Individuals, as well as organisations,\nnotice, test, and criticize unfair results to hold model designers and\ndeployers accountable. We offer a framework that assists these groups in\nmitigating unfair representations stemming from the training datasets. Our\nframework relies on two inter-operating adversaries to improve fairness. First,\na model is trained with the goal of preventing the guessing of protected\nattributes' values while limiting utility losses. This first step optimizes the\nmodel's parameters for fairness. Second, the framework leverages evasion\nattacks from adversarial machine learning to generate new examples that will be\nmisclassified. These new examples are then used to retrain and improve the\nmodel in the first step. These two steps are iteratively applied until a\nsignificant improvement in fairness is obtained. We evaluated our framework on\nwell-studied datasets in the fairness literature -- including COMPAS -- where\nit can surpass other approaches concerning demographic parity, equality of\nopportunity and also the model's utility. We also illustrate our findings on\nthe subtle difficulties when mitigating unfairness and highlight how our\nframework can assist model designers."", 'Algorithms learned from data are increasingly used for deciding many aspects\nin our life: from movies we see, to prices we pay, or medicine we get. Yet\nthere is growing evidence that decision making by inappropriately trained\nalgorithms may unintentionally discriminate people. For example, in automated\nmatching of candidate CVs with job descriptions, algorithms may capture and\npropagate ethnicity related biases. Several repairs for selected algorithms\nhave already been proposed, but the underlying mechanisms how such\ndiscrimination happens from the computational perspective are not yet\nscientifically understood. We need to develop theoretical understanding how\nalgorithms may become discriminatory, and establish fundamental machine\nlearning principles for prevention. We need to analyze machine learning process\nas a whole to systematically explain the roots of discrimination occurrence,\nwhich will allow to devise global machine learning optimization criteria for\nguaranteed prevention, as opposed to pushing empirical constraints into\nexisting algorithms case-by-case. As a result, the state-of-the-art will\nadvance from heuristic repairing, to proactive and theoretically supported\nprevention. This is needed not only because law requires to protect vulnerable\npeople. Penetration of big data initiatives will only increase, and computer\nscience needs to provide solid explanations and accountability to the public,\nbefore public concerns lead to unnecessarily restrictive regulations against\nmachine learning.', 'Thanks to the great progress of machine learning in the last years, several\nArtificial Intelligence (AI) techniques have been increasingly moving from the\ncontrolled research laboratory settings to our everyday life. AI is clearly\nsupportive in many decision-making scenarios, but when it comes to sensitive\nareas such as health care, hiring policies, education, banking or justice, with\nmajor impact on individuals and society, it becomes crucial to establish\nguidelines on how to design, develop, deploy and monitor this technology.\nIndeed the decision rules elaborated by machine learning models are data-driven\nand there are multiple ways in which discriminatory biases can seep into data.\nAlgorithms trained on those data incur the risk of amplifying prejudices and\nsocietal stereotypes by over associating protected attributes such as gender,\nethnicity or disabilities with the prediction task. Starting from the extensive\nexperience of the National Metrology Institute on measurement standards and\ncertification roadmaps, and of Politecnico di Torino on machine learning as\nwell as methods for domain bias evaluation and mastering, we propose a first\njoint effort to define the operational steps needed for AI fairness\ncertification. Specifically we will overview the criteria that should be met by\nan AI system before coming into official service and the conformity assessment\nprocedures useful to monitor its functioning for fair decisions.', 'The lack of bias management in Recommender Systems leads to minority groups\nreceiving unfair recommendations. Moreover, the trade-off between equity and\nprecision makes it difficult to obtain recommendations that meet both criteria.\nHere we propose a Deep Learning based Collaborative Filtering algorithm that\nprovides recommendations with an optimum balance between fairness and accuracy\nwithout knowing demographic information about the users. Experimental results\nshow that it is possible to make fair recommendations without losing a\nsignificant proportion of accuracy.', 'Decision making in crucial applications such as lending, hiring, and college\nadmissions has witnessed increasing use of algorithmic models and techniques as\na result of a confluence of factors such as ubiquitous connectivity, ability to\ncollect, aggregate, and process large amounts of fine-grained data using cloud\ncomputing, and ease of access to applying sophisticated machine learning\nmodels. Quite often, such applications are powered by search and recommendation\nsystems, which in turn make use of personalized ranking algorithms. At the same\ntime, there is increasing awareness about the ethical and legal challenges\nposed by the use of such data-driven systems. Researchers and practitioners\nfrom different disciplines have recently highlighted the potential for such\nsystems to discriminate against certain population groups, due to biases in the\ndatasets utilized for learning their underlying recommendation models. We\npresent a study of fairness in online personalization settings involving the\nranking of individuals. Starting from a fair warm-start machine-learned model,\nwe first demonstrate that online personalization can cause the model to learn\nto act in an unfair manner if the user is biased in his/her responses. For this\npurpose, we construct a stylized model for generating training data with\npotentially biased features as well as potentially biased labels and quantify\nthe extent of bias that is learned by the model when the user responds in a\nbiased manner as in many real-world scenarios. We then formulate the problem of\nlearning personalized models under fairness constraints and present a\nregularization based approach for mitigating biases in machine learning. We\ndemonstrate the efficacy of our approach through extensive simulations with\ndifferent parameter settings. Code:\nhttps://github.com/groshanlal/Fairness-Aware-Online-Personalization', 'Fairness is a critical trait in decision making. As machine-learning models\nare increasingly being used in sensitive application domains (e.g. education\nand employment) for decision making, it is crucial that the decisions computed\nby such models are free of unintended bias. But how can we automatically\nvalidate the fairness of arbitrary machine-learning models? For a given\nmachine-learning model and a set of sensitive input parameters, our AEQUITAS\napproach automatically discovers discriminatory inputs that highlight fairness\nviolation. At the core of AEQUITAS are three novel strategies to employ\nprobabilistic search over the input space with the objective of uncovering\nfairness violation. Our AEQUITAS approach leverages inherent robustness\nproperty in common machine-learning models to design and implement scalable\ntest generation methodologies. An appealing feature of our generated test\ninputs is that they can be systematically added to the training set of the\nunderlying model and improve its fairness. To this end, we design a fully\nautomated module that guarantees to improve the fairness of the underlying\nmodel.\n  We implemented AEQUITAS and we have evaluated it on six state-of-the-art\nclassifiers, including a classifier that was designed with fairness\nconstraints. We show that AEQUITAS effectively generates inputs to uncover\nfairness violation in all the subject classifiers and systematically improves\nthe fairness of the respective models using the generated test inputs. In our\nevaluation, AEQUITAS generates up to 70% discriminatory inputs (w.r.t. the\ntotal number of inputs generated) and leverages these inputs to improve the\nfairness up to 94%.', 'Machine learning systems are often trained using data collected from\nhistorical decisions. If past decisions were biased, then automated systems\nthat learn from historical data will also be biased. We propose a black-box\napproach to identify and remove biased training data. Machine learning models\ntrained on such debiased data (a subset of the original training data) have low\nindividual discrimination, often 0%. These models also have greater accuracy\nand lower statistical disparity than models trained on the full historical\ndata. We evaluated our methodology in experiments using 6 real-world datasets.\nOur approach outperformed seven previous approaches in terms of individual\ndiscrimination and accuracy.', 'The popularity of pretrained language models in natural language processing\nsystems calls for a careful evaluation of such models in down-stream tasks,\nwhich have a higher potential for societal impact. The evaluation of such\nsystems usually focuses on accuracy measures. Our findings in this paper call\nfor attention to be paid to fairness measures as well. Through the analysis of\nmore than a dozen pretrained language models of varying sizes on two toxic text\nclassification tasks (English), we demonstrate that focusing on accuracy\nmeasures alone can lead to models with wide variation in fairness\ncharacteristics. Specifically, we observe that fairness can vary even more than\naccuracy with increasing training data size and different random\ninitializations. At the same time, we find that little of the fairness\nvariation is explained by model size, despite claims in the literature. To\nimprove model fairness without retraining, we show that two post-processing\nmethods developed for structured, tabular data can be successfully applied to a\nrange of pretrained language models. Warning: This paper contains samples of\noffensive text.', 'We observe standard transfer learning can improve prediction accuracies of\ntarget tasks at the cost of lowering their prediction fairness -- a phenomenon\nwe named discriminatory transfer. We examine prediction fairness of a standard\nhypothesis transfer algorithm and a standard multi-task learning algorithm, and\nshow they both suffer discriminatory transfer on the real-world Communities and\nCrime data set. The presented case study introduces an interaction between\nfairness and transfer learning, as an extension of existing fairness studies\nthat focus on single task learning.', 'This paper investigates the application of machine learning when training a\ncredit decision model over real, publicly available data whilst accounting for\n""bias objectives"". We use the term ""bias objective"" to describe the requirement\nthat a trained model displays discriminatory bias against a given groups of\nindividuals that doesn\'t exceed a prescribed level, where such level can be\nzero. This research presents an empirical study examining the tension between\ncompeting model training objectives which in all cases include one or more bias\nobjectives.\n  This work is motivated by the observation that the parties associated with\ncreditworthiness models have requirements that can not certainly be fully met\nsimultaneously. The research herein seeks to highlight the impracticality of\nsatisfying all parties\' objectives, demonstrating the need for ""trade-offs"" to\nbe made. The results and conclusions presented by this paper are of particular\nimportance for all stakeholders within the credit scoring industry that rely\nupon artificial intelligence (AI) models as part of the decision-making process\nwhen determining the creditworthiness of individuals. This paper provides an\nexposition of the difficulty of training AI models that are able to\nsimultaneously satisfy multiple bias objectives whilst maintaining acceptable\nlevels of accuracy. Stakeholders should be aware of this difficulty and should\nacknowledge that some degree of discriminatory bias, across a number of\nprotected characteristics and formulations of bias, cannot be avoided.', 'We consider the problem of improving fairness when one lacks access to a\ndataset labeled with protected groups, making it difficult to take advantage of\nstrategies that can improve fairness but require protected group labels, either\nat training or runtime. To address this, we investigate improving fairness\nmetrics for proxy groups, and test whether doing so results in improved\nfairness for the true sensitive groups. Results on benchmark and real-world\ndatasets demonstrate that such a proxy fairness strategy can work well in\npractice. However, we caution that the effectiveness likely depends on the\nchoice of fairness metric, as well as how aligned the proxy groups are with the\ntrue protected groups in terms of the constrained model parameters.', 'As algorithms increasingly inform and influence decisions made about\nindividuals, it becomes increasingly important to address concerns that these\nalgorithms might be discriminatory. The output of an algorithm can be\ndiscriminatory for many reasons, most notably: (1) the data used to train the\nalgorithm might be biased (in various ways) to favor certain populations over\nothers; (2) the analysis of this training data might inadvertently or\nmaliciously introduce biases that are not borne out in the data. This work\nfocuses on the latter concern.\n  We develop and study multicalbration -- a new measure of algorithmic fairness\nthat aims to mitigate concerns about discrimination that is introduced in the\nprocess of learning a predictor from data. Multicalibration guarantees accurate\n(calibrated) predictions for every subpopulation that can be identified within\na specified class of computations. We think of the class as being quite rich;\nin particular, it can contain many overlapping subgroups of a protected group.\n  We show that in many settings this strong notion of protection from\ndiscrimination is both attainable and aligned with the goal of obtaining\naccurate predictions. Along the way, we present new algorithms for learning a\nmulticalibrated predictor, study the computational complexity of this task, and\ndraw new connections to computational learning models such as agnostic\nlearning.', 'Machine learning algorithms tend to create more accurate models with the\navailability of large datasets. In some cases, highly accurate models can hide\nthe presence of bias in the data. There are several studies published that\ntackle the development of discriminatory-aware machine learning algorithms. We\ncenter on the further evaluation of machine learning models by doing error\nanalysis, to understand under what conditions the model is not working as\nexpected. We focus on the ethical implications of avoiding error analysis, from\na falsification of results and discrimination perspective. Finally, we show\ndifferent ways to approach error analysis in non-interpretable machine learning\nalgorithms such as deep learning.', 'People are rated and ranked, towards algorithmic decision making in an\nincreasing number of applications, typically based on machine learning.\nResearch on how to incorporate fairness into such tasks has prevalently pursued\nthe paradigm of group fairness: giving adequate success rates to specifically\nprotected groups. In contrast, the alternative paradigm of individual fairness\nhas received relatively little attention, and this paper advances this less\nexplored direction. The paper introduces a method for probabilistically mapping\nuser records into a low-rank representation that reconciles individual fairness\nand the utility of classifiers and rankings in downstream applications. Our\nnotion of individual fairness requires that users who are similar in all\ntask-relevant attributes such as job qualification, and disregarding all\npotentially discriminating attributes such as gender, should have similar\noutcomes. We demonstrate the versatility of our method by applying it to\nclassification and learning-to-rank tasks on a variety of real-world datasets.\nOur experiments show substantial improvements over the best prior work for this\nsetting.', 'While Explainable Artificial Intelligence (XAI) is increasingly expanding\nmore areas of application, little has been applied to make deep Reinforcement\nLearning (RL) more comprehensible. As RL becomes ubiquitous and used in\ncritical and general public applications, it is essential to develop methods\nthat make it better understood and more interpretable. This study proposes a\nnovel approach to explain cooperative strategies in multiagent RL using Shapley\nvalues, a game theory concept used in XAI that successfully explains the\nrationale behind decisions taken by Machine Learning algorithms. Through\ntesting common assumptions of this technique in two cooperation-centered\nsocially challenging multi-agent environments environments, this article argues\nthat Shapley values are a pertinent way to evaluate the contribution of players\nin a cooperative multi-agent RL context. To palliate the high overhead of this\nmethod, Shapley values are approximated using Monte Carlo sampling.\nExperimental results on Multiagent Particle and Sequential Social Dilemmas show\nthat Shapley values succeed at estimating the contribution of each agent. These\nresults could have implications that go beyond games in economics, (e.g., for\nnon-discriminatory decision making, ethical and responsible AI-derived\ndecisions or policy making under fairness constraints). They also expose how\nShapley values only give general explanations about a model and cannot explain\na single run, episode nor justify precise actions taken by agents. Future work\nshould focus on addressing these critical aspects.', 'Machine learning models (e.g., speech recognizers) are usually trained to\nminimize average loss, which results in representation disparity---minority\ngroups (e.g., non-native speakers) contribute less to the training objective\nand thus tend to suffer higher loss. Worse, as model accuracy affects user\nretention, a minority group can shrink over time. In this paper, we first show\nthat the status quo of empirical risk minimization (ERM) amplifies\nrepresentation disparity over time, which can even make initially fair models\nunfair. To mitigate this, we develop an approach based on distributionally\nrobust optimization (DRO), which minimizes the worst case risk over all\ndistributions close to the empirical distribution. We prove that this approach\ncontrols the risk of the minority group at each time step, in the spirit of\nRawlsian distributive justice, while remaining oblivious to the identity of the\ngroups. We demonstrate that DRO prevents disparity amplification on examples\nwhere ERM fails, and show improvements in minority group user satisfaction in a\nreal-world text autocomplete task.', ""Increasingly, discrimination by algorithms is perceived as a societal and\nlegal problem. As a response, a number of criteria for implementing algorithmic\nfairness in machine learning have been developed in the literature. This paper\nproposes the Continuous Fairness Algorithm (CFA$\\theta$) which enables a\ncontinuous interpolation between different fairness definitions. More\nspecifically, we make three main contributions to the existing literature.\nFirst, our approach allows the decision maker to continuously vary between\nspecific concepts of individual and group fairness. As a consequence, the\nalgorithm enables the decision maker to adopt intermediate ``worldviews'' on\nthe degree of discrimination encoded in algorithmic processes, adding nuance to\nthe extreme cases of ``we're all equal'' (WAE) and ``what you see is what you\nget'' (WYSIWYG) proposed so far in the literature. Second, we use optimal\ntransport theory, and specifically the concept of the barycenter, to maximize\ndecision maker utility under the chosen fairness constraints. Third, the\nalgorithm is able to handle cases of intersectionality, i.e., of\nmulti-dimensional discrimination of certain groups on grounds of several\ncriteria. We discuss three main examples (credit applications; college\nadmissions; insurance contracts) and map out the legal and policy implications\nof our approach. The explicit formalization of the trade-off between individual\nand group fairness allows this post-processing approach to be tailored to\ndifferent situational contexts in which one or the other fairness criterion may\ntake precedence. Finally, we evaluate our model experimentally."", 'Data-driven learning algorithms are employed in many online applications, in\nwhich data become available over time, like network monitoring, stock price\nprediction, job applications, etc. The underlying data distribution might\nevolve over time calling for model adaptation as new instances arrive and old\ninstances become obsolete. In such dynamic environments, the so-called data\nstreams, fairness-aware learning cannot be considered as a one-off requirement,\nbut rather it should comprise a continual requirement over the stream. Recent\nfairness-aware stream classifiers ignore the problem of class imbalance, which\nmanifests in many real-life applications, and mitigate discrimination mainly\nbecause they ""reject"" minority instances at large due to their inability to\neffectively learn all classes.\n  In this work, we propose \\ours, an online fairness-aware approach that\nmaintains a valid and fair classifier over the stream. \\ours~is an online\nboosting approach that changes the training distribution in an online fashion\nby monitoring stream\'s class imbalance and tweaks its decision boundary to\nmitigate discriminatory outcomes over the stream. Experiments on 8 real-world\nand 1 synthetic datasets from different domains with varying class imbalance\ndemonstrate the superiority of our method over state-of-the-art fairness-aware\nstream approaches with a range (relative) increase [11.2\\%-14.2\\%] in balanced\naccuracy, [22.6\\%-31.8\\%] in gmean, [42.5\\%-49.6\\%] in recall, [14.3\\%-25.7\\%]\nin kappa and [89.4\\%-96.6\\%] in statistical parity (fairness).', ""In consequential domains such as recidivism prediction, facility inspection,\nand benefit assignment, it's important for individuals to know the\ndecision-relevant information for the model's prediction. In addition,\npredictions should be fair both in terms of the outcome and the justification\nof the outcome. In other words, decision-relevant features should provide\nsufficient information for the predicted outcome and should be independent of\nthe membership of individuals in protected groups such as race and gender. In\nthis work, we focus on the problem of (un)fairness in the justification of the\ntext-based neural models. We tie the explanatory power of the model to fairness\nin the outcome and propose a fairness-aware summarization mechanism to detect\nand counteract the bias in such models. Given a potentially biased natural\nlanguage explanation for a decision, we use a multi-task neural model and an\nattribution mechanism based on integrated gradients to extract high-utility and\nlow-bias justifications in form of a summary. The extracted summary is then\nused for training a model to make decisions for individuals. Results on several\nreal world datasets suggest that our method drastically limits the demographic\nleakage in the input (fairness in justification) while moderately enhancing the\nfairness in the outcome. Our model is also effective in detecting and\ncounteracting several types of data poisoning attacks that synthesize\nrace-coded reasoning or irrelevant justifications.""]","[""A machine learning model may exhibit discrimination when used to make\ndecisions involving people. One potential cause for such outcomes is that the\nmodel uses a statistical proxy for a protected demographic attribute. In this\npaper we formulate a definition of proxy use for the setting of linear\nregression and present algorithms for detecting proxies. Our definition follows\nrecent work on proxies in classification models, and characterizes a model's\nconstituent behavior that: 1) correlates closely with a protected random\nvariable, and 2) is causally influential in the overall behavior of the model.\nWe show that proxies in linear regression models can be efficiently identified\nby solving a second-order cone program, and further extend this result to\naccount for situations where the use of a certain input variable is justified\nas a `business necessity'. Finally, we present empirical results on two law\nenforcement datasets that exhibit varying degrees of racial disparity in\nprediction outcomes, demonstrating that proxies shed useful light on the causes\nof discriminatory behavior in models."", 'When machine-learning algorithms are deployed in high-stakes decisions, we\nwant to ensure that their deployment leads to fair and equitable outcomes. This\nconcern has motivated a fast-growing literature that focuses on diagnosing and\naddressing disparities in machine predictions. However, many machine\npredictions are deployed to assist in decisions where a human decision-maker\nretains the ultimate decision authority. In this article, we therefore consider\nhow properties of machine predictions affect the resulting human decisions. We\nshow in a formal model that the inclusion of a biased human decision-maker can\nrevert common relationships between the structure of the algorithm and the\nqualities of resulting decisions. Specifically, we document that excluding\ninformation about protected groups from the prediction may fail to reduce, and\nmay even increase, ultimate disparities. While our concrete results rely on\nspecific assumptions about the data, algorithm, and decision-maker, they show\nmore broadly that any study of critical properties of complex decision systems,\nsuch as the fairness of machine-assisted human decisions, should go beyond\nfocusing on the underlying algorithmic predictions in isolation.', 'The growing use of automated decision-making in critical applications, such\nas crime prediction and college admission, has raised questions about fairness\nin machine learning. How can we decide whether different treatments are\nreasonable or discriminatory? In this paper, we investigate discrimination in\nmachine learning from a visual analytics perspective and propose an interactive\nvisualization tool, DiscriLens, to support a more comprehensive analysis. To\nreveal detailed information on algorithmic discrimination, DiscriLens\nidentifies a collection of potentially discriminatory itemsets based on causal\nmodeling and classification rules mining. By combining an extended Euler\ndiagram with a matrix-based visualization, we develop a novel set visualization\nto facilitate the exploration and interpretation of discriminatory itemsets. A\nuser study shows that users can interpret the visually encoded information in\nDiscriLens quickly and accurately. Use cases demonstrate that DiscriLens\nprovides informative guidance in understanding and reducing algorithmic\ndiscrimination.', 'Developing classification algorithms that are fair with respect to sensitive\nattributes of the data has become an important problem due to the growing\ndeployment of classification algorithms in various social contexts. Several\nrecent works have focused on fairness with respect to a specific metric,\nmodeled the corresponding fair classification problem as a constrained\noptimization problem, and developed tailored algorithms to solve them. Despite\nthis, there still remain important metrics for which we do not have fair\nclassifiers and many of the aforementioned algorithms do not come with\ntheoretical guarantees; perhaps because the resulting optimization problem is\nnon-convex. The main contribution of this paper is a new meta-algorithm for\nclassification that takes as input a large class of fairness constraints, with\nrespect to multiple non-disjoint sensitive attributes, and which comes with\nprovable guarantees. This is achieved by first developing a meta-algorithm for\na large family of classification problems with convex constraints, and then\nshowing that classification problems with general types of fairness constraints\ncan be reduced to those in this family. We present empirical results that show\nthat our algorithm can achieve near-perfect fairness with respect to various\nfairness metrics, and that the loss in accuracy due to the imposed fairness\nconstraints is often small. Overall, this work unifies several prior works on\nfair classification, presents a practical algorithm with theoretical\nguarantees, and can handle fairness metrics that were previously not possible.', 'Many machine learning systems make extensive use of large amounts of data\nregarding human behaviors. Several researchers have found various\ndiscriminatory practices related to the use of human-related machine learning\nsystems, for example in the field of criminal justice, credit scoring and\nadvertising. Fair machine learning is therefore emerging as a new field of\nstudy to mitigate biases that are inadvertently incorporated into algorithms.\nData scientists and computer engineers are making various efforts to provide\ndefinitions of fairness. In this paper, we provide an overview of the most\nwidespread definitions of fairness in the field of machine learning, arguing\nthat the ideas highlighting each formalization are closely related to different\nideas of justice and to different interpretations of democracy embedded in our\nculture. This work intends to analyze the definitions of fairness that have\nbeen proposed to date to interpret the underlying criteria and to relate them\nto different ideas of democracy.', 'The advent of powerful prediction algorithms led to increased automation of\nhigh-stake decisions regarding the allocation of scarce resources such as\ngovernment spending and welfare support. This automation bears the risk of\nperpetuating unwanted discrimination against vulnerable and historically\ndisadvantaged groups. Research on algorithmic discrimination in computer\nscience and other disciplines developed a plethora of fairness metrics to\ndetect and correct discriminatory algorithms. Drawing on robust sociological\nand philosophical discourse on distributive justice, we identify the\nlimitations and problematic implications of prominent fairness metrics. We show\nthat metrics implementing equality of opportunity only apply when resource\nallocations are based on deservingness, but fail when allocations should\nreflect concerns about egalitarianism, sufficiency, and priority. We argue that\nby cleanly distinguishing between prediction tasks and decision tasks, research\non fair machine learning could take better advantage of the rich literature on\ndistributive justice.', 'Real-world applications of machine learning tools in high-stakes domains are\noften regulated to be fair, in the sense that the predicted target should\nsatisfy some quantitative notion of parity with respect to a protected\nattribute. However, the exact tradeoff between fairness and accuracy is not\nentirely clear, even for the basic paradigm of classification problems. In this\npaper, we characterize an inherent tradeoff between statistical parity and\naccuracy in the classification setting by providing a lower bound on the sum of\ngroup-wise errors of any fair classifiers. Our impossibility theorem could be\ninterpreted as a certain uncertainty principle in fairness: if the base rates\ndiffer among groups, then any fair classifier satisfying statistical parity has\nto incur a large error on at least one of the groups. We further extend this\nresult to give a lower bound on the joint error of any (approximately) fair\nclassifiers, from the perspective of learning fair representations. To show\nthat our lower bound is tight, assuming oracle access to Bayes (potentially\nunfair) classifiers, we also construct an algorithm that returns a randomized\nclassifier that is both optimal (in terms of accuracy) and fair. Interestingly,\nwhen the protected attribute can take more than two values, an extension of\nthis lower bound does not admit an analytic solution. Nevertheless, in this\ncase, we show that the lower bound can be efficiently computed by solving a\nlinear program, which we term as the TV-Barycenter problem, a barycenter\nproblem under the TV-distance. On the upside, we prove that if the group-wise\nBayes optimal classifiers are close, then learning fair representations leads\nto an alternative notion of fairness, known as the accuracy parity, which\nstates that the error rates are close between groups. Finally, we also conduct\nexperiments on real-world datasets to confirm our theoretical findings.', ""In this work, the problem of predicting dropout risk in undergraduate studies\nis addressed from a perspective of algorithmic fairness. We develop a machine\nlearning method to predict the risks of university dropout and\nunderperformance. The objective is to understand if such a system can identify\nstudents at risk while avoiding potential discriminatory biases. When modeling\nboth risks, we obtain prediction models with an Area Under the ROC Curve (AUC)\nof 0.77-0.78 based on the data available at the enrollment time, before the\nfirst year of studies starts. This data includes the students' demographics,\nthe high school they attended, and their admission (average) grade. Our models\nare calibrated: they produce estimated probabilities for each risk, not mere\nscores. We analyze if this method leads to discriminatory outcomes for some\nsensitive groups in terms of prediction accuracy (AUC) and error rates\n(Generalized False Positive Rate, GFPR, or Generalized False Negative Rate,\nGFNR). The models exhibit some equity in terms of AUC and GFNR along groups.\nThe similar GFNR means a similar probability of failing to detect risk for\nstudents who drop out. The disparities in GFPR are addressed through a\nmitigation process that does not affect the calibration of the model."", 'As decision-making increasingly relies on Machine Learning (ML) and (big)\ndata, the issue of fairness in data-driven Artificial Intelligence (AI) systems\nis receiving increasing attention from both research and industry. A large\nvariety of fairness-aware machine learning solutions have been proposed which\ninvolve fairness-related interventions in the data, learning algorithms and/or\nmodel outputs. However, a vital part of proposing new approaches is evaluating\nthem empirically on benchmark datasets that represent realistic and diverse\nsettings. Therefore, in this paper, we overview real-world datasets used for\nfairness-aware machine learning. We focus on tabular data as the most common\ndata representation for fairness-aware machine learning. We start our analysis\nby identifying relationships between the different attributes, particularly\nw.r.t. protected attributes and class attribute, using a Bayesian network. For\na deeper understanding of bias in the datasets, we investigate the interesting\nrelationships using exploratory analysis.', 'As an important problem in modern data analytics, classification has\nwitnessed varieties of applications from different domains. Different from\nconventional classification approaches, fair classification concerns the issues\nof unintentional biases against the sensitive features (e.g., gender, race).\nDue to high nonconvexity of fairness measures, existing methods are often\nunable to model exact fairness, which can cause inferior fair classification\noutcomes. This paper fills the gap by developing a novel unified framework to\njointly optimize accuracy and fairness. The proposed framework is versatile and\ncan incorporate different fairness measures studied in literature precisely as\nwell as can be applicable to many classifiers including deep classification\nmodels. Specifically, in this paper, we first prove Fisher consistency of the\nproposed framework. We then show that many classification models within this\nframework can be recast as mixed-integer convex programs, which can be solved\neffectively by off-the-shelf solvers when the instance sizes are moderate and\ncan be used as benchmarks to compare the efficiency of approximation\nalgorithms. We prove that in the proposed framework, when the classification\noutcomes are known, the resulting problem, termed ""unbiased subdata selection,""\nis strongly polynomial-solvable and can be used to enhance the classification\nfairness by selecting more representative data points. This motivates us to\ndevelop an iterative refining strategy (IRS) to solve the large-scale\ninstances, where we improve the classification accuracy and conduct the\nunbiased subdata selection in an alternating fashion. We study the convergence\nproperty of IRS and derive its approximation bound. More broadly, this\nframework can be leveraged to improve classification models with unbalanced\ndata by taking F1 score into consideration.', 'It is tempting to think that machines are less prone to unfairness and\nprejudice. However, machine learning approaches compute their outputs based on\ndata. While biases can enter at any stage of the development pipeline, models\nare particularly receptive to mirror biases of the datasets they are trained on\nand therefore do not necessarily reflect truths about the world but, primarily,\ntruths about the data. To raise awareness about the relationship between modern\nalgorithms and the data that shape them, we use a conditional invertible neural\nnetwork to disentangle the dataset-specific information from the information\nwhich is shared across different datasets. In this way, we can project the same\nimage onto different datasets, thereby revealing their inherent biases. We use\nthis methodology to (i) investigate the impact of dataset quality on the\nperformance of generative models, (ii) show how societal biases of datasets are\nreplicated by generative models, and (iii) present creative applications\nthrough unpaired transfer between diverse datasets such as photographs, oil\nportraits, and animes. Our code and an interactive demonstration are available\nat https://github.com/CompVis/net2net .', 'Fairness in decision-making has been a long-standing issue in our society.\nDespite the increasing number of research activities on unfairness mitigation\nin machine learning models, there is little research focusing on mitigating\nunfairness in human decisions. Fairness in human decisions is as important as,\nif not more important than, fairness in machine learning models since there are\nprocesses where humans make the final decisions and machine learning models can\ninherit bias from the human decisions they were trained on. As a result, this\nwork aims to detect unfairness in human decisions, the very first step of\nsolving the unfair human decision problem.\n  This paper proposes to utilize the existing machine learning fairness\ndetection mechanisms to detect unfairness in human decisions. The rationale\nbehind this is, while it is difficult to directly test whether a human makes\nunfair decisions, with current research on machine learning fairness, it is now\neasy to test, on a large scale at a low cost, whether a machine learning model\nis unfair. By synthesizing unfair labels on four general machine learning\nfairness datasets and one image processing dataset, this paper shows that the\nproposed approach is able to detect (1) whether or not unfair labels exist in\nthe training data and (2) the degree and direction of the unfairness. We\nbelieve that this work demonstrates the potential of utilizing machine learning\nfairness to detect human decision fairness. Following this work, research can\nbe conducted on (1) preventing future unfair decisions, (2) fixing prior unfair\ndecisions, and (3) training a fairer machine learning model.', 'The causes underlying unfair decision making are complex, being internalised\nin different ways by decision makers, other actors dealing with data and\nmodels, and ultimately by the individuals being affected by these decisions.\nOne frequent manifestation of all these latent causes arises in the form of\nmissing values: protected groups are more reluctant to give information that\ncould be used against them, delicate information for some groups can be erased\nby human operators, or data acquisition may simply be less complete and\nsystematic for minority groups. As a result, missing values and bias in data\nare two phenomena that are tightly coupled. However, most recent techniques,\nlibraries and experimental results dealing with fairness in machine learning\nhave simply ignored missing data. In this paper, we claim that fairness\nresearch should not miss the opportunity to deal properly with missing data. To\nsupport this claim, (1) we analyse the sources of missing data and bias, and we\nmap the common causes, (2) we find that rows containing missing values are\nusually fairer than the rest, which should not be treated as the uncomfortable\nugly data that different techniques and libraries get rid of at the first\noccasion, and (3) we study the trade-off between performance and fairness when\nthe rows with missing values are used (either because the technique deals with\nthem directly or by imputation methods). We end the paper with a series of\nrecommended procedures about what to do with missing data when aiming for fair\ndecision making.', 'We present an optimization framework for learning a fair classifier in the\npresence of noisy perturbations in the protected attributes. Compared to prior\nwork, our framework can be employed with a very general class of linear and\nlinear-fractional fairness constraints, can handle multiple, non-binary\nprotected attributes, and outputs a classifier that comes with provable\nguarantees on both accuracy and fairness. Empirically, we show that our\nframework can be used to attain either statistical rate or false positive rate\nfairness guarantees with a minimal loss in accuracy, even when the noise is\nlarge, in two real-world datasets.', 'Recent attempts to achieve fairness in predictive models focus on the balance\nbetween fairness and accuracy. In sensitive applications such as healthcare or\ncriminal justice, this trade-off is often undesirable as any increase in\nprediction error could have devastating consequences. In this work, we argue\nthat the fairness of predictions should be evaluated in context of the data,\nand that unfairness induced by inadequate samples sizes or unmeasured\npredictive variables should be addressed through data collection, rather than\nby constraining the model. We decompose cost-based metrics of discrimination\ninto bias, variance, and noise, and propose actions aimed at estimating and\nreducing each term. Finally, we perform case-studies on prediction of income,\nmortality, and review ratings, confirming the value of this analysis. We find\nthat data collection is often a means to reduce discrimination without\nsacrificing accuracy.', 'The wide spread usage of automated data-driven decision support systems has\nraised a lot of concerns regarding accountability and fairness of the employed\nmodels in the absence of human supervision. Existing fairness-aware approaches\ntackle fairness as a batch learning problem and aim at learning a fair model\nwhich can then be applied to future instances of the problem. In many\napplications, however, the data comes sequentially and its characteristics\nmight evolve with time. In such a setting, it is counter-intuitive to ""fix"" a\n(fair) model over the data stream as changes in the data might incur changes in\nthe underlying model therefore, affecting its fairness. In this work, we\npropose fairness-enhancing interventions that modify the input data so that the\noutcome of any stream classifier applied to that data will be fair. Experiments\non real and synthetic data show that our approach achieves good predictive\nperformance and low discrimination scores over the course of the stream.', 'Considerable research effort has been guided towards algorithmic fairness but\nthere is still no major breakthrough. In practice, an exhaustive search over\nall possible techniques and hyperparameters is needed to find optimal\nfairness-accuracy trade-offs. Hence, coupled with the lack of tools for ML\npractitioners, real-world adoption of bias reduction methods is still scarce.\nTo address this, we present Fairband, a bandit-based fairness-aware\nhyperparameter optimization (HO) algorithm. Fairband is conceptually simple,\nresource-efficient, easy to implement, and agnostic to both the objective\nmetrics, model types and the hyperparameter space being explored. Moreover, by\nintroducing fairness notions into HO, we enable seamless and efficient\nintegration of fairness objectives into real-world ML pipelines. We compare\nFairband with popular HO methods on four real-world decision-making datasets.\nWe show that Fairband can efficiently navigate the fairness-accuracy trade-off\nthrough hyperparameter optimization. Furthermore, without extra training cost,\nit consistently finds configurations attaining substantially improved fairness\nat a comparatively small decrease in predictive accuracy.', 'The newly emerged machine learning (e.g. deep learning) methods have become a\nstrong driving force to revolutionize a wide range of industries, such as smart\nhealthcare, financial technology, and surveillance systems. Meanwhile, privacy\nhas emerged as a big concern in this machine learning-based artificial\nintelligence era. It is important to note that the problem of privacy\npreservation in the context of machine learning is quite different from that in\ntraditional data privacy protection, as machine learning can act as both friend\nand foe. Currently, the work on the preservation of privacy and machine\nlearning (ML) is still in an infancy stage, as most existing solutions only\nfocus on privacy problems during the machine learning process. Therefore, a\ncomprehensive study on the privacy preservation problems and machine learning\nis required. This paper surveys the state of the art in privacy issues and\nsolutions for machine learning. The survey covers three categories of\ninteractions between privacy and machine learning: (i) private machine\nlearning, (ii) machine learning aided privacy protection, and (iii) machine\nlearning-based privacy attack and corresponding protection schemes. The current\nresearch progress in each category is reviewed and the key challenges are\nidentified. Finally, based on our in-depth analysis of the area of privacy and\nmachine learning, we point out future research directions in this field.', 'Since many critical decisions impacting human lives are increasingly being\nmade by algorithms, it is important to ensure that the treatment of individuals\nunder such algorithms is demonstrably fair under reasonable notions of\nfairness. One compelling notion proposed in the literature is that of\nindividual fairness (IF), which advocates that similar individuals should be\ntreated similarly (Dwork et al. 2012). Originally proposed for offline\ndecisions, this notion does not, however, account for temporal considerations\nrelevant for online decision-making. In this paper, we extend the notion of IF\nto account for the time at which a decision is made, in settings where there\nexists a notion of conduciveness of decisions as perceived by the affected\nindividuals. We introduce two definitions: (i) fairness-across-time (FT) and\n(ii) fairness-in-hindsight (FH). FT is the simplest temporal extension of IF\nwhere treatment of individuals is required to be individually fair relative to\nthe past as well as future, while in FH, we require a one-sided notion of\nindividual fairness that is defined relative to only the past decisions. We\nshow that these two definitions can have drastically different implications in\nthe setting where the principal needs to learn the utility model. Linear regret\nrelative to optimal individually fair decisions is inevitable under FT for\nnon-trivial examples. On the other hand, we design a new algorithm: Cautious\nFair Exploration (CaFE), which satisfies FH and achieves sub-linear regret\nguarantees for a broad range of settings. We characterize lower bounds showing\nthat these guarantees are order-optimal in the worst case. FH can thus be\nembedded as a primary safeguard against unfair discrimination in algorithmic\ndeployments, without hindering the ability to take good decisions in the\nlong-run.', ""For supervised learning models, the analysis of generalization ability\n(generalizability) is vital because the generalizability expresses how well a\nmodel will perform on unseen data. Traditional generalization methods, such as\nthe VC dimension, do not apply to deep neural network (DNN) models. Thus, new\ntheories to explain the generalizability of DNNs are required. In this study,\nwe hypothesize that the DNN with a simpler decision boundary has better\ngeneralizability by the law of parsimony (Occam's Razor). We create the\ndecision boundary complexity (DBC) score to define and measure the complexity\nof decision boundary of DNNs. The idea of the DBC score is to generate data\npoints (called adversarial examples) on or near the decision boundary. Our new\napproach then measures the complexity of the boundary using the entropy of\neigenvalues of these data. The method works equally well for high-dimensional\ndata. We use training data and the trained model to compute the DBC score. And,\nthe ground truth for model's generalizability is its test accuracy. Experiments\nbased on the DBC score have verified our hypothesis. The DBC is shown to\nprovide an effective method to measure the complexity of a decision boundary\nand gives a quantitative measure of the generalizability of DNNs."", 'Consider a binary decision making process where a single machine learning\nclassifier replaces a multitude of humans. We raise questions about the\nresulting loss of diversity in the decision making process. We study the\npotential benefits of using random classifier ensembles instead of a single\nclassifier in the context of fairness-aware learning and demonstrate various\nattractive properties: (i) an ensemble of fair classifiers is guaranteed to be\nfair, for several different measures of fairness, (ii) an ensemble of unfair\nclassifiers can still achieve fair outcomes, and (iii) an ensemble of\nclassifiers can achieve better accuracy-fairness trade-offs than a single\nclassifier. Finally, we introduce notions of distributional fairness to\ncharacterize further potential benefits of random classifier ensembles.', 'Fairness of classification and regression has received much attention\nrecently and various, partially non-compatible, criteria have been proposed.\nThe fairness criteria can be enforced for a given classifier or, alternatively,\nthe data can be adapated to ensure that every classifier trained on the data\nwill adhere to desired fairness criteria. We present a practical data adaption\nmethod based on quantile preservation in causal structural equation models. The\ndata adaptation is based on a presumed counterfactual model for the data. While\nthe counterfactual model itself cannot be verified experimentally, we show that\ncertain population notions of fairness are still guaranteed even if the\ncounterfactual model is misspecified. The precise nature of the fulfilled\nnon-causal fairness notion (such as demographic parity, separation or\nsufficiency) depends on the structure of the underlying causal model and the\nchoice of resolving variables. We describe an implementation of the proposed\ndata adaptation procedure based on Random Forests and demonstrate its practical\nuse on simulated and real-world data.', 'What does it mean for an algorithm to be fair? Different papers use different\nnotions of algorithmic fairness, and although these appear internally\nconsistent, they also seem mutually incompatible. We present a mathematical\nsetting in which the distinctions in previous papers can be made formal. In\naddition to characterizing the spaces of inputs (the ""observed"" space) and\noutputs (the ""decision"" space), we introduce the notion of a construct space: a\nspace that captures unobservable, but meaningful variables for the prediction.\n  We show that in order to prove desirable properties of the entire\ndecision-making process, different mechanisms for fairness require different\nassumptions about the nature of the mapping from construct space to decision\nspace. The results in this paper imply that future treatments of algorithmic\nfairness should more explicitly state assumptions about the relationship\nbetween constructs and observations.', ""What does it mean for a machine learning model to be `fair', in terms which\ncan be operationalised? Should fairness consist of ensuring everyone has an\nequal probability of obtaining some benefit, or should we aim instead to\nminimise the harms to the least advantaged? Can the relevant ideal be\ndetermined by reference to some alternative state of affairs in which a\nparticular social pattern of discrimination does not exist? Various definitions\nproposed in recent literature make different assumptions about what terms like\ndiscrimination and fairness mean and how they can be defined in mathematical\nterms. Questions of discrimination, egalitarianism and justice are of\nsignificant interest to moral and political philosophers, who have expended\nsignificant efforts in formalising and defending these central concepts. It is\ntherefore unsurprising that attempts to formalise `fairness' in machine\nlearning contain echoes of these old philosophical debates. This paper draws on\nexisting work in moral and political philosophy in order to elucidate emerging\ndebates about fair machine learning."", 'In recent years, significant progress has been made in developing more\naccurate and efficient machine learning algorithms for segmentation of medical\nand natural images. In this review article, we highlight the imperative role of\nmachine learning algorithms in enabling efficient and accurate segmentation in\nthe field of medical imaging. We specifically focus on several key studies\npertaining to the application of machine learning methods to biomedical image\nsegmentation. We review classical machine learning algorithms such as Markov\nrandom fields, k-means clustering, random forest, etc. Although such classical\nlearning models are often less accurate compared to the deep learning\ntechniques, they are often more sample efficient and have a less complex\nstructure. We also review different deep learning architectures, such as the\nartificial neural networks (ANNs), the convolutional neural networks (CNNs),\nand the recurrent neural networks (RNNs), and present the segmentation results\nattained by those learning models that were published in the past three years.\nWe highlight the successes and limitations of each machine learning paradigm.\nIn addition, we discuss several challenges related to the training of different\nmachine learning models, and we present some heuristics to address those\nchallenges.', 'Deep neural networks and in particular, deep neural classifiers have become\nan integral part of many modern applications. Despite their practical success,\nwe still have limited knowledge of how they work and the demand for such an\nunderstanding is evergrowing. In this regard, one crucial aspect of deep neural\nnetwork classifiers that can help us deepen our knowledge about their\ndecision-making behavior is to investigate their decision boundaries.\nNevertheless, this is contingent upon having access to samples populating the\nareas near the decision boundary. To achieve this, we propose a novel approach\nwe call Deep Decision boundary Instance Generation (DeepDIG). DeepDIG utilizes\na method based on adversarial example generation as an effective way of\ngenerating samples near the decision boundary of any deep neural network model.\nThen, we introduce a set of important principled characteristics that take\nadvantage of the generated instances near the decision boundary to provide\nmultifaceted understandings of deep neural networks. We have performed\nextensive experiments on multiple representative datasets across various deep\nneural network models and characterized their decision boundaries. The code is\npublicly available at https://github.com/hamidkarimi/DeepDIG/.', ""Machine Learning has been successfully applied in systems applications such\nas memory prefetching and caching, where learned models have been shown to\noutperform heuristics. However, the lack of understanding the inner workings of\nthese models -- interpretability -- remains a major obstacle for adoption in\nreal-world deployments. Understanding a model's behavior can help system\nadministrators and developers gain confidence in the model, understand risks,\nand debug unexpected behavior in production. Interpretability for models used\nin computer systems poses a particular challenge: Unlike ML models trained on\nimages or text, the input domain (e.g., memory access patterns, program\ncounters) is not immediately interpretable. A major challenge is therefore to\nexplain the model in terms of concepts that are approachable to a human\npractitioner. By analyzing a state-of-the-art caching model, we provide\nevidence that the model has learned concepts beyond simple statistics that can\nbe leveraged for explanations. Our work provides a first step towards\nexplanability of system ML models and highlights both promises and challenges\nof this emerging research area."", 'Fatal accidents are a major issue hindering the wide acceptance of\nsafety-critical systems that employ machine learning and deep learning models,\nsuch as automated driving vehicles. In order to use machine learning in a\nsafety-critical system, it is necessary to demonstrate the safety and security\nof the system through engineering processes. However, thus far, no such widely\naccepted engineering concepts or frameworks have been established for these\nsystems. The key to using a machine learning model in a deductively engineered\nsystem is decomposing the data-driven training of machine learning models into\nrequirement, design, and verification, particularly for machine learning models\nused in safety-critical systems. Simultaneously, open problems and relevant\ntechnical fields are not organized in a manner that enables researchers to\nselect a theme and work on it. In this study, we identify, classify, and\nexplore the open problems in engineering (safety-critical) machine learning\nsystems --- that is, in terms of requirement, design, and verification of\nmachine learning models and systems --- as well as discuss related works and\nresearch directions, using automated driving vehicles as an example. Our\nresults show that machine learning models are characterized by a lack of\nrequirements specification, lack of design specification, lack of\ninterpretability, and lack of robustness. We also perform a gap analysis on a\nconventional system quality standard SQuARE with the characteristics of machine\nlearning models to study quality models for machine learning systems. We find\nthat a lack of requirements specification and lack of robustness have the\ngreatest impact on conventional quality models.', 'Deep neural networks have been widely deployed in various machine learning\ntasks. However, recent works have demonstrated that they are vulnerable to\nadversarial examples: carefully crafted small perturbations to cause\nmisclassification by the network. In this work, we propose a novel defense\nmechanism called Boundary Conditional GAN to enhance the robustness of deep\nneural networks against adversarial examples. Boundary Conditional GAN, a\nmodified version of Conditional GAN, can generate boundary samples with true\nlabels near the decision boundary of a pre-trained classifier. These boundary\nsamples are fed to the pre-trained classifier as data augmentation to make the\ndecision boundary more robust. We empirically show that the model improved by\nour approach consistently defenses against various types of adversarial attacks\nsuccessfully. Further quantitative investigations about the improvement of\nrobustness and visualization of decision boundaries are also provided to\njustify the effectiveness of our strategy. This new defense mechanism that uses\nboundary samples to enhance the robustness of networks opens up a new way to\ndefense adversarial attacks consistently.', 'We introduce a flexible family of fairness regularizers for (linear and\nlogistic) regression problems. These regularizers all enjoy convexity,\npermitting fast optimization, and they span the rang from notions of group\nfairness to strong individual fairness. By varying the weight on the fairness\nregularizer, we can compute the efficient frontier of the accuracy-fairness\ntrade-off on any given dataset, and we measure the severity of this trade-off\nvia a numerical quantity we call the Price of Fairness (PoF). The centerpiece\nof our results is an extensive comparative study of the PoF across six\ndifferent datasets in which fairness is a primary consideration.', 'Probabilistic models are a critical part of the modern deep learning toolbox\n- ranging from generative models (VAEs, GANs), sequence to sequence models used\nin machine translation and speech processing to models over functional spaces\n(conditional neural processes, neural processes). Given the size and complexity\nof these models, safely deploying them in applications requires the development\nof tools to analyze their behavior rigorously and provide some guarantees that\nthese models are consistent with a list of desirable properties or\nspecifications. For example, a machine translation model should produce\nsemantically equivalent outputs for innocuous changes in the input to the\nmodel. A functional regression model that is learning a distribution over\nmonotonic functions should predict a larger value at a larger input.\nVerification of these properties requires a new framework that goes beyond\nnotions of verification studied in deterministic feedforward networks, since\nrequiring worst-case guarantees in probabilistic models is likely to produce\nconservative or vacuous results. We propose a novel formulation of verification\nfor deep probabilistic models that take in conditioning inputs and sample\nlatent variables in the course of producing an output: We require that the\noutput of the model satisfies a linear constraint with high probability over\nthe sampling of latent variables and for every choice of conditioning input to\nthe model. We show that rigorous lower bounds on the probability that the\nconstraint is satisfied can be obtained efficiently. Experiments with neural\nprocesses show that several properties of interest while modeling functional\nspaces can be modeled within this framework (monotonicity, convexity) and\nverified efficiently using our algorithms', 'Many recent works on knowledge distillation have provided ways to transfer\nthe knowledge of a trained network for improving the learning process of a new\none, but finding a good technique for knowledge distillation is still an open\nproblem. In this paper, we provide a new perspective based on a decision\nboundary, which is one of the most important component of a classifier. The\ngeneralization performance of a classifier is closely related to the adequacy\nof its decision boundary, so a good classifier bears a good decision boundary.\nTherefore, transferring information closely related to the decision boundary\ncan be a good attempt for knowledge distillation. To realize this goal, we\nutilize an adversarial attack to discover samples supporting a decision\nboundary. Based on this idea, to transfer more accurate information about the\ndecision boundary, the proposed algorithm trains a student classifier based on\nthe adversarial samples supporting the decision boundary. Experiments show that\nthe proposed method indeed improves knowledge distillation and achieves the\nstate-of-the-arts performance.', 'Interpretable machine learning tackles the important problem that humans\ncannot understand the behaviors of complex machine learning models and how\nthese models arrive at a particular decision. Although many approaches have\nbeen proposed, a comprehensive understanding of the achievements and challenges\nis still lacking. We provide a survey covering existing techniques to increase\nthe interpretability of machine learning models. We also discuss crucial issues\nthat the community should consider in future work such as designing\nuser-friendly explanations and developing comprehensive evaluation metrics to\nfurther push forward the area of interpretable machine learning.', 'Due to the increasing use of machine learning in practice it becomes more and\nmore important to be able to explain the prediction and behavior of machine\nlearning models. An instance of explanations are counterfactual explanations\nwhich provide an intuitive and useful explanations of machine learning models.\nIn this survey we review model-specific methods for efficiently computing\ncounterfactual explanations of many different machine learning models and\npropose methods for models that have not been considered in literature so far.', ""While machine learning is traditionally a resource intensive task, embedded\nsystems, autonomous navigation and the vision of the Internet-of-Things fuel\nthe interest in resource efficient approaches. These approaches require a\ncarefully chosen trade-off between performance and resource consumption in\nterms of computation and energy. On top of this, it is crucial to treat\nuncertainty in a consistent manner in all but the simplest applications of\nmachine learning systems. In particular, a desideratum for any real-world\nsystem is to be robust in the presence of outliers and corrupted data, as well\nas being `aware' of its limits, i.e.\\ the system should maintain and provide an\nuncertainty estimate over its own predictions. These complex demands are among\nthe major challenges in current machine learning research and key to ensure a\nsmooth transition of machine learning technology into every day's applications.\nIn this article, we provide an overview of the current state of the art of\nmachine learning techniques facilitating these real-world requirements. First\nwe provide a comprehensive review of resource-efficiency in deep neural\nnetworks with focus on techniques for model size reduction, compression and\nreduced precision. These techniques can be applied during training or as\npost-processing and are widely used to reduce both computational complexity and\nmemory footprint. As most (practical) neural networks are limited in their ways\nto treat uncertainty, we contrast them with probabilistic graphical models,\nwhich readily serve these desiderata by means of probabilistic inference. In\nthat way, we provide an extensive overview of the current state-of-the-art of\nrobust and efficient machine learning for real-world systems."", ""Machine learning is a modern approach to problem-solving and task automation.\nIn particular, machine learning is concerned with the development and\napplications of algorithms that can recognize patterns in data and use them for\npredictive modeling. Artificial neural networks are a particular class of\nmachine learning algorithms and models that evolved into what is now described\nas deep learning. Given the computational advances made in the last decade,\ndeep learning can now be applied to massive data sets and in innumerable\ncontexts. Therefore, deep learning has become its own subfield of machine\nlearning. In the context of biological research, it has been increasingly used\nto derive novel insights from high-dimensional biological data. To make the\nbiological applications of deep learning more accessible to scientists who have\nsome experience with machine learning, we solicited input from a community of\nresearchers with varied biological and deep learning interests. These\nindividuals collaboratively contributed to this manuscript's writing using the\nGitHub version control platform and the Manubot manuscript generation toolset.\nThe goal was to articulate a practical, accessible, and concise set of\nguidelines and suggestions to follow when using deep learning. In the course of\nour discussions, several themes became clear: the importance of understanding\nand applying machine learning fundamentals as a baseline for utilizing deep\nlearning, the necessity for extensive model comparisons with careful\nevaluation, and the need for critical thought in interpreting results generated\nby deep learning, among others."", 'Explainable Artificial Intelligence (or xAI) has become an important research\ntopic in the fields of Machine Learning and Deep Learning. In this paper, we\npropose a Genetic Programming (GP) based approach, named Genetic Programming\nExplainer (GPX), to the problem of explaining decisions computed by AI systems.\nThe method generates a noise set located in the neighborhood of the point of\ninterest, whose prediction should be explained, and fits a local explanation\nmodel for the analyzed sample. The tree structure generated by GPX provides a\ncomprehensible analytical, possibly non-linear, symbolic expression which\nreflects the local behavior of the complex model. We considered three machine\nlearning techniques that can be recognized as complex black-box models: Random\nForest, Deep Neural Network and Support Vector Machine in twenty data sets for\nregression and classifications problems. Our results indicate that the GPX is\nable to produce more accurate understanding of complex models than the state of\nthe art. The results validate the proposed approach as a novel way to deploy GP\nto improve interpretability.', 'Many works demonstrate that deep learning system is vulnerable to adversarial\nattack. A deep learning system consists of two parts: the deep learning task\nand the deep model. Nowadays, most existing works investigate the impact of the\ndeep model on robustness of deep learning systems, ignoring the impact of the\nlearning task. In this paper, we adopt the binary and interval label encoding\nstrategy to redefine the classification task and design corresponding loss to\nimprove robustness of the deep learning system. Our method can be viewed as\nimproving the robustness of deep learning systems from both the learning task\nand deep model. Experimental results demonstrate that our learning task-aware\nmethod is much more robust than traditional classification while retaining the\naccuracy.', 'Research in machine learning is at a turning point. While supervised deep\nlearning has conquered the field at a breathtaking pace and demonstrated the\nability to solve inference problems with unprecedented accuracy, it still does\nnot quite live up to its name if we think of learning as the process of\nacquiring knowledge about a subject or problem. Major weaknesses of present-day\ndeep learning models are, for instance, their lack of adaptability to changes\nof environment or their incapability to perform other kinds of tasks than the\none they were trained for. While it is still unclear how to overcome these\nlimitations, one can observe a paradigm shift within the machine learning\ncommunity, with research interests shifting away from increasing the\nperformance of highly parameterized models to exceedingly specific tasks, and\ntowards employing machine learning algorithms in highly diverse domains. This\nresearch question can be approached from different angles. For instance, the\nfield of Informed AI investigates the problem of infusing domain knowledge into\na machine learning model, by using techniques such as regularization, data\naugmentation or post-processing.\n  On the other hand, a remarkable number of works in the recent years has\nfocused on developing models that by themselves guarantee a certain degree of\nversatility and invariance with respect to the domain or problem at hand. Thus,\nrather than investigating how to provide domain-specific knowledge to machine\nlearning models, these works explore methods that equip the models with the\ncapability of acquiring the knowledge by themselves. This white paper provides\nan introduction and discussion of this emerging field in machine learning\nresearch. To this end, it reviews the role of knowledge in machine learning,\nand discusses its relation to the concept of invariance, before providing a\nliterature review of the field.', ""Machine learning models always make a prediction, even when it is likely to\nbe inaccurate. This behavior should be avoided in many decision support\napplications, where mistakes can have severe consequences. Albeit already\nstudied in 1970, machine learning with a reject option recently gained\ninterest. This machine learning subfield enables machine learning models to\nabstain from making a prediction when likely to make a mistake.\n  This survey aims to provide an overview on machine learning with a reject\noption. We introduce the conditions leading to two types of rejection,\nambiguity and novelty rejection. Moreover, we define the existing architectures\nfor models with a reject option, describe the standard learning strategies to\ntrain such models and relate traditional machine learning techniques to\nrejection. Additionally, we review strategies to evaluate a model's predictive\nand rejective quality. Finally, we provide examples of relevant application\ndomains and show how machine learning with rejection relates to other machine\nlearning research areas."", 'Deep learning has achieved a great success in many areas, from computer\nvision to natural language processing, to game playing, and much more. Yet,\nwhat deep learning is really doing is still an open question. There are a lot\nof works in this direction. For example, [5] tried to explain deep learning by\ngroup renormalization, and [6] tried to explain deep learning from the view of\nfunctional approximation. In order to address this very crucial question, here\nwe see deep learning from perspective of mechanical learning and learning\nmachine (see [1], [2]). From this particular angle, we can see deep learning\nmuch better and answer with confidence: What deep learning is really doing? why\nit works well, how it works, and how much data is necessary for learning. We\nalso will discuss advantages and disadvantages of deep learning at the end of\nthis work.', 'Various health-care applications such as assisted living, fall detection\netc., require modeling of user behavior through Human Activity Recognition\n(HAR). HAR using mobile- and wearable-based deep learning algorithms have been\non the rise owing to the advancements in pervasive computing. However, there\nare two other challenges that need to be addressed: first, the deep learning\nmodel should support on-device incremental training (model updation) from\nreal-time incoming data points to learn user behavior over time, while also\nbeing resource-friendly; second, a suitable ground truthing technique (like\nActive Learning) should help establish labels on-the-fly while also selecting\nonly the most informative data points to query from an oracle. Hence, in this\npaper, we propose ActiveHARNet, a resource-efficient deep ensembled model which\nsupports on-device Incremental Learning and inference, with capabilities to\nrepresent model uncertainties through approximations in Bayesian Neural\nNetworks using dropout. This is combined with suitable acquisition functions\nfor active learning. Empirical results on two publicly available wrist-worn HAR\nand fall detection datasets indicate that ActiveHARNet achieves considerable\nefficiency boost during inference across different users, with a substantially\nlow number of acquired pool points (at least 60% reduction) during incremental\nlearning on both datasets experimented with various acquisition functions, thus\ndemonstrating deployment and Incremental Learning feasibility.', 'Machine learning models are widely used for real-world applications, such as\ndocument analysis and vision. Constrained machine learning problems are\nproblems where learned models have to both be accurate and respect constraints.\nFor continuous convex constraints, many works have been proposed, but learning\nunder combinatorial constraints is still a hard problem. The goal of this paper\nis to broaden the modeling capacity of constrained machine learning problems by\nincorporating existing work from combinatorial optimization. We propose first a\ngeneral framework called BaGeL (Branch, Generate and Learn) which applies\nBranch and Bound to constrained learning problems where a learning problem is\ngenerated and trained at each node until only valid models are obtained.\nBecause machine learning has specific requirements, we also propose an extended\ntable constraint to split the space of hypotheses. We validate the approach on\ntwo examples: a linear regression under configuration constraints and a\nnon-negative matrix factorization with prior knowledge for latent semantics\nanalysis.', 'Recent success of machine learning in many domains has been overwhelming,\nwhich often leads to false expectations regarding the capabilities of behavior\nlearning in robotics. In this survey, we analyze the current state of machine\nlearning for robotic behaviors. We will give a broad overview of behaviors that\nhave been learned and used on real robots. Our focus is on kinematically or\nsensorially complex robots. That includes humanoid robots or parts of humanoid\nrobots, for example, legged robots or robotic arms. We will classify presented\nbehaviors according to various categories and we will draw conclusions about\nwhat can be learned and what should be learned. Furthermore, we will give an\noutlook on problems that are challenging today but might be solved by machine\nlearning in the future and argue that classical robotics and other approaches\nfrom artificial intelligence should be integrated more with machine learning to\nform complete, autonomous systems.', 'Overfitting is a major problem in training machine learning models,\nspecifically deep neural networks. This problem may be caused by imbalanced\ndatasets and initialization of the model parameters, which conforms the model\ntoo closely to the training data and negatively affects the generalization\nperformance of the model for unseen data. The original dropout is a\nregularization technique to drop hidden units randomly during training. In this\npaper, we propose an adaptive technique to wisely drop the visible and hidden\nunits in a deep neural network using Ising energy of the network. The\npreliminary results show that the proposed approach can keep the classification\nperformance competitive to the original network while eliminating optimization\nof unnecessary network parameters in each training cycle. The dropout state of\nunits can also be applied to the trained (inference) model. This technique\ncould compress the network in terms of number of parameters up to 41.18% and\n55.86% for the classification task on the MNIST and Fashion-MNIST datasets,\nrespectively.', 'Many end-user evaluations of data visualization techniques have been run\nduring the last decades. Their results are cornerstones to build efficient\nvisualization systems. However, designing such an evaluation is always complex\nand time-consuming and may end in a lack of statistical evidence and\nreproducibility. We believe that modern and efficient computer vision\ntechniques, such as deep convolutional neural networks (CNNs), may help\nvisualization researchers to build and/or adjust their evaluation hypothesis.\nThe basis of our idea is to train machine learning models on several\nvisualization techniques to solve a specific task. Our assumption is that it is\npossible to compare the efficiency of visualization techniques based on the\nperformance of their corresponding model. As current machine learning models\nare not able to strictly reflect human capabilities, including their\nimperfections, such results should be interpreted with caution. However, we\nthink that using machine learning-based pre-evaluation, as a pre-process of\nstandard user evaluations, should help researchers to perform a more exhaustive\nstudy of their design space. Thus, it should improve their final user\nevaluation by providing it better test cases. In this paper, we present the\nresults of two experiments we have conducted to assess how correlated the\nperformance of users and computer vision techniques can be. That study compares\ntwo mainstream graph visualization techniques: node-link (\\NL) and\nadjacency-matrix (\\MD) diagrams. Using two well-known deep convolutional neural\nnetworks, we partially reproduced user evaluations from Ghoniem \\textit{et al.}\nand from Okoe \\textit{et al.}. These experiments showed that some user\nevaluation results can be reproduced automatically.', 'We survey in this article the connections between Machine Learning and\nControl Theory. Control Theory provide useful concepts and tools for Machine\nLearning. Conversely Machine Learning can be used to solve large control\nproblems. In the first part of the paper, we develop the connections between\nreinforcement learning and Markov Decision Processes, which are discrete time\ncontrol problems. In the second part, we review the concept of supervised\nlearning and the relation with static optimization. Deep learning which extends\nsupervised learning, can be viewed as a control problem. In the third part, we\npresent the links between stochastic gradient descent and mean-field theory.\nConversely, in the fourth and fifth parts, we review machine learning\napproaches to stochastic control problems, and focus on the deterministic case,\nto explain, more easily, the numerical algorithms.', ""The increasing incorporation of Artificial Intelligence in the form of\nautomated systems into decision-making procedures highlights not only the\nimportance of decision theory for automated systems but also the need for these\ndecision procedures to be explainable to the people involved in them.\nTraditional realist accounts of explanation, wherein explanation is a relation\nthat holds (or does not hold) eternally between an explanans and an\nexplanandum, are not adequate to account for the notion of explanation required\nfor artificial decision procedures. We offer an alternative account of\nexplanation as used in the context of automated decision-making that makes\nexplanation an epistemic phenomenon, and one that is dependent on context. This\naccount of explanation better accounts for the way that we talk about, and use,\nexplanations and derived concepts, such as `explanatory power', and also allows\nus to differentiate between reasons or causes on the one hand, which do not\nneed to have an epistemic aspect, and explanations on the other, which do have\nsuch an aspect. Against this theoretical backdrop we then review existing\napproaches to explanation in Artificial Intelligence and Machine Learning, and\nsuggest desiderata which truly explainable decision systems should fulfill."", 'The additive margin softmax (AM-Softmax) loss has delivered remarkable\nperformance in speaker verification. A supposed behavior of AM-Softmax is that\nit can shrink within-class variation by putting emphasis on target logits,\nwhich in turn improves margin between target and non-target classes. In this\npaper, we conduct a careful analysis on the behavior of AM-Softmax loss, and\nshow that this loss does not implement real max-margin training. Based on this\nobservation, we present a Real AM-Softmax loss which involves a true margin\nfunction in the softmax training. Experiments conducted on VoxCeleb1, SITW and\nCNCeleb demonstrated that the corrected AM-Softmax loss consistently\noutperforms the original one. The code has been released at\nhttps://gitlab.com/csltstu/sunine.', 'Breakthroughs in machine learning are rapidly changing science and society,\nyet our fundamental understanding of this technology has lagged far behind.\nIndeed, one of the central tenets of the field, the bias-variance trade-off,\nappears to be at odds with the observed behavior of methods used in the modern\nmachine learning practice. The bias-variance trade-off implies that a model\nshould balance under-fitting and over-fitting: rich enough to express\nunderlying structure in data, simple enough to avoid fitting spurious patterns.\nHowever, in the modern practice, very rich models such as neural networks are\ntrained to exactly fit (i.e., interpolate) the data. Classically, such models\nwould be considered over-fit, and yet they often obtain high accuracy on test\ndata. This apparent contradiction has raised questions about the mathematical\nfoundations of machine learning and their relevance to practitioners.\n  In this paper, we reconcile the classical understanding and the modern\npractice within a unified performance curve. This ""double descent"" curve\nsubsumes the textbook U-shaped bias-variance trade-off curve by showing how\nincreasing model capacity beyond the point of interpolation results in improved\nperformance. We provide evidence for the existence and ubiquity of double\ndescent for a wide spectrum of models and datasets, and we posit a mechanism\nfor its emergence. This connection between the performance and the structure of\nmachine learning models delineates the limits of classical analyses, and has\nimplications for both the theory and practice of machine learning.', ""Class labels have been empirically shown useful in improving the sample\nquality of generative adversarial nets (GANs). In this paper, we mathematically\nstudy the properties of the current variants of GANs that make use of class\nlabel information. With class aware gradient and cross-entropy decomposition,\nwe reveal how class labels and associated losses influence GAN's training.\nBased on that, we propose Activation Maximization Generative Adversarial\nNetworks (AM-GAN) as an advanced solution. Comprehensive experiments have been\nconducted to validate our analysis and evaluate the effectiveness of our\nsolution, where AM-GAN outperforms other strong baselines and achieves\nstate-of-the-art Inception Score (8.91) on CIFAR-10. In addition, we\ndemonstrate that, with the Inception ImageNet classifier, Inception Score\nmainly tracks the diversity of the generator, and there is, however, no\nreliable evidence that it can reflect the true sample quality. We thus propose\na new metric, called AM Score, to provide a more accurate estimation of the\nsample quality. Our proposed model also outperforms the baseline methods in the\nnew metric."", 'We present an approach to analyze $C^1(\\mathbb{R}^m)$ functions that\naddresses limitations present in the Active Subspaces (AS) method of\nConstantine et al.(2015; 2014). Under appropriate hypotheses, our Active\nManifolds (AM) method identifies a 1-D curve in the domain (the active\nmanifold) on which nearly all values of the unknown function are attained, and\nwhich can be exploited for approximation or analysis, especially when $m$ is\nlarge (high-dimensional input space). We provide theorems justifying our AM\ntechnique and an algorithm permitting functional approximation and sensitivity\nanalysis. Using accessible, low-dimensional functions as initial examples, we\nshow AM reduces approximation error by an order of magnitude compared to AS, at\nthe expense of more computation. Following this, we revisit the sensitivity\nanalysis by Glaws et al. (2017), who apply AS to analyze a magnetohydrodynamic\npower generator model, and compare the performance of AM on the same data. Our\nanalysis provides detailed information not captured by AS, exhibiting the\ninfluence of each parameter individually along an active manifold. Overall, AM\nrepresents a novel technique for analyzing functional models with benefits\nincluding: reducing $m$-dimensional analysis to a 1-D analogue, permitting more\naccurate regression than AS (at more computational expense), enabling more\ninformative sensitivity analysis, and granting accessible visualizations(2-D\nplots) of parameter sensitivity along the AM.', 'The key to success in machine learning (ML) is the use of effective data\nrepresentations. Traditionally, data representations were hand-crafted.\nRecently it has been demonstrated that, given sufficient data, deep neural\nnetworks can learn effective implicit representations from simple input\nrepresentations. However, for most scientific problems, the use of deep\nlearning is not appropriate as the amount of available data is limited, and/or\nthe output models must be explainable. Nevertheless, many scientific problems\ndo have significant amounts of data available on related tasks, which makes\nthem amenable to multi-task learning, i.e. learning many related problems\nsimultaneously. Here we propose a novel and general representation learning\napproach for multi-task learning that works successfully with small amounts of\ndata. The fundamental new idea is to transform an input intrinsic data\nrepresentation (i.e., handcrafted features), to an extrinsic representation\nbased on what a pre-trained set of models predict about the examples. This\ntransformation has the dual advantages of producing significantly more accurate\npredictions, and providing explainable models. To demonstrate the utility of\nthis transformative learning approach, we have applied it to three real-world\nscientific problems: drug-design (quantitative structure activity relationship\nlearning), predicting human gene expression (across different tissue types and\ndrug treatments), and meta-learning for machine learning (predicting which\nmachine learning methods work best for a given problem). In all three problems,\ntransformative machine learning significantly outperforms the best intrinsic\nrepresentation.', 'This paper surveys the machine learning literature and presents in an\noptimization framework several commonly used machine learning approaches.\nParticularly, mathematical optimization models are presented for regression,\nclassification, clustering, deep learning, and adversarial learning, as well as\nnew emerging applications in machine teaching, empirical model learning, and\nBayesian network structure learning. Such models can benefit from the\nadvancement of numerical optimization techniques which have already played a\ndistinctive role in several machine learning settings. The strengths and the\nshortcomings of these models are discussed and potential research directions\nand open problems are highlighted.', 'In realistic environments, speech is usually interfered by various noise and\nreverberation, which dramatically degrades the performance of automatic speech\nrecognition (ASR) systems. To alleviate this issue, the commonest way is to use\na well-designed speech enhancement approach as the front-end of ASR. However,\nmore complex pipelines, more computations and even higher hardware costs\n(microphone array) are additionally consumed for this kind of methods. In\naddition, speech enhancement would result in speech distortions and mismatches\nto training. In this paper, we propose an adversarial training method to\ndirectly boost noise robustness of acoustic model. Specifically, a jointly\ncompositional scheme of generative adversarial net (GAN) and neural\nnetwork-based acoustic model (AM) is used in the training phase. GAN is used to\ngenerate clean feature representations from noisy features by the guidance of a\ndiscriminator that tries to distinguish between the true clean signals and\ngenerated signals. The joint optimization of generator, discriminator and AM\nconcentrates the strengths of both GAN and AM for speech recognition.\nSystematic experiments on CHiME-4 show that the proposed method significantly\nimproves the noise robustness of AM and achieves the average relative error\nrate reduction of 23.38% and 11.54% on the development and test set,\nrespectively.', 'Deep Learning is one of the newest trends in Machine Learning and Artificial\nIntelligence research. It is also one of the most popular scientific research\ntrends now-a-days. Deep learning methods have brought revolutionary advances in\ncomputer vision and machine learning. Every now and then, new and new deep\nlearning techniques are being born, outperforming state-of-the-art machine\nlearning and even existing deep learning techniques. In recent years, the world\nhas seen many major breakthroughs in this field. Since deep learning is\nevolving at a huge speed, its kind of hard to keep track of the regular\nadvances especially for new researchers. In this paper, we are going to briefly\ndiscuss about recent advances in Deep Learning for past few years.', 'Deep learning has emerged as a powerful machine learning technique that\nlearns multiple layers of representations or features of the data and produces\nstate-of-the-art prediction results. Along with the success of deep learning in\nmany other application domains, deep learning is also popularly used in\nsentiment analysis in recent years. This paper first gives an overview of deep\nlearning and then provides a comprehensive survey of its current applications\nin sentiment analysis.', 'The recent explosion of machine learning (ML) and artificial intelligence\n(AI) shows great potential in the breakthrough of metal additive manufacturing\n(AM) process modeling. However, the success of conventional machine learning\ntools in data science is primarily attributed to the unprecedented large amount\nof labeled data-sets (big data), which can be either obtained by experiments or\nfirst-principle simulations. Unfortunately, these labeled data-sets are\nexpensive to obtain in AM due to the high expense of the AM experiments and\nprohibitive computational cost of high-fidelity simulations.\n  We propose a physics-informed neural network (PINN) framework that fuses both\ndata and first physical principles, including conservation laws of momentum,\nmass, and energy, into the neural network to inform the learning processes. To\nthe best knowledge of the authors, this is the first application of PINN to\nthree dimensional AM processes modeling. Besides, we propose a hard-type\napproach for Dirichlet boundary conditions (BCs) based on a Heaviside function,\nwhich can not only enforce the BCs but also accelerate the learning process.\nThe PINN framework is applied to two representative metal manufacturing\nproblems, including the 2018 NIST AM-Benchmark test series. We carefully assess\nthe performance of the PINN model by comparing the predictions with available\nexperimental data and high-fidelity simulation results. The investigations show\nthat the PINN, owed to the additional physical knowledge, can accurately\npredict the temperature and melt pool dynamics during metal AM processes with\nonly a moderate amount of labeled data-sets. The foray of PINN to metal AM\nshows the great potential of physics-informed deep learning for broader\napplications to advanced manufacturing.', 'The great success of deep learning shows that its technology contains\nprofound truth, and understanding its internal mechanism not only has important\nimplications for the development of its technology and effective application in\nvarious fields, but also provides meaningful insights into the understanding of\nhuman brain mechanism. At present, most of the theoretical research on deep\nlearning is based on mathematics. This dissertation proposes that the neural\nnetwork of deep learning is a physical system, examines deep learning from\nthree different perspectives: microscopic, macroscopic, and physical world\nviews, answers multiple theoretical puzzles in deep learning by using physics\nprinciples. For example, from the perspective of quantum mechanics and\nstatistical physics, this dissertation presents the calculation methods for\nconvolution calculation, pooling, normalization, and Restricted Boltzmann\nMachine, as well as the selection of cost functions, explains why deep learning\nmust be deep, what characteristics are learned in deep learning, why\nConvolutional Neural Networks do not have to be trained layer by layer, and the\nlimitations of deep learning, etc., and proposes the theoretical direction and\nbasis for the further development of deep learning now and in the future. The\nbrilliance of physics flashes in deep learning, we try to establish the deep\nlearning technology based on the scientific theory of physics.', 'Machine learning has made tremendous progress in recent years and received\nlarge amounts of public attention. Though we are still far from designing a\nfull artificially intelligent agent, machine learning has brought us many\napplications in which computers solve human learning tasks remarkably well.\nMuch of this progress comes from a recent trend within machine learning, called\ndeep learning. Deep learning models are responsible for many state-of-the-art\napplications of machine learning. Despite their success, deep learning models\nare hard to train, very difficult to understand, and often times so complex\nthat training is only possible on very large GPU clusters. Lots of work has\nbeen done on enabling neural networks to learn efficiently. However, the design\nand architecture of such neural networks is often done manually through trial\nand error and expert knowledge. This thesis inspects different approaches,\nexisting and novel, to automate the design of deep feedforward neural networks\nin an attempt to create less complex models with good performance that take\naway the burden of deciding on an architecture and make it more efficient to\ndesign and train such deep networks.', 'Machine learning is poised as a very powerful tool that can drastically\nimprove our ability to carry out scientific research. However, many issues need\nto be addressed before this becomes a reality. This article focuses on one\nparticular issue of broad interest: How can we integrate machine learning with\nphysics-based modeling to develop new interpretable and truly reliable physical\nmodels? After introducing the general guidelines, we discuss the two most\nimportant issues for developing machine learning-based physical models:\nImposing physical constraints and obtaining optimal datasets. We also provide a\nsimple and intuitive explanation for the fundamental reasons behind the success\nof modern machine learning, as well as an introduction to the concurrent\nmachine learning framework needed for integrating machine learning with\nphysics-based modeling. Molecular dynamics and moment closure of kinetic\nequations are used as examples to illustrate the main issues discussed. We end\nwith a general discussion on where this integration will lead us to, and where\nthe new frontier will be after machine learning is successfully integrated into\nscientific modeling.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'Speech-driven visual speech synthesis involves mapping features extracted\nfrom acoustic speech to the corresponding lip animation controls for a face\nmodel. This mapping can take many forms, but a powerful approach is to use deep\nneural networks (DNNs). However, a limitation is the lack of synchronized\naudio, video, and depth data required to reliably train the DNNs, especially\nfor speaker-independent models. In this paper, we investigate adapting an\nautomatic speech recognition (ASR) acoustic model (AM) for the visual speech\nsynthesis problem. We train the AM on ten thousand hours of audio-only data.\nThe AM is then adapted to the visual speech synthesis domain using ninety hours\nof synchronized audio-visual speech. Using a subjective assessment test, we\ncompared the performance of the AM-initialized DNN to one with a random\ninitialization. The results show that viewers significantly prefer animations\ngenerated from the AM-initialized DNN than the ones generated using the\nrandomly initialized model. We conclude that visual speech synthesis can\nsignificantly benefit from the powerful representation of speech in the ASR\nacoustic models.', 'Machine learning is usually defined in behaviourist terms, where external\nvalidation is the primary mechanism of learning. In this paper, I argue for a\nmore holistic interpretation in which finding more probable, efficient and\nabstract representations is as central to learning as performance. In other\nwords, machine learning should be extended with strategies to reason over its\nown learning process, leading to so-called meta-cognitive machine learning. As\nsuch, the de facto definition of machine learning should be reformulated in\nthese intrinsically multi-objective terms, taking into account not only the\ntask performance but also internal learning objectives. To this end, we suggest\na ""model entropy function"" to be defined that quantifies the efficiency of the\ninternal learning processes. It is conjured that the minimization of this model\nentropy leads to concept formation. Besides philosophical aspects, some initial\nillustrations are included to support the claims.', ""The paper examines the potential of deep learning to support decisions in\nfinancial risk management. We develop a deep learning model for predicting\nwhether individual spread traders secure profits from future trades. This task\nembodies typical modeling challenges faced in risk and behavior forecasting.\nConventional machine learning requires data that is representative of the\nfeature-target relationship and relies on the often costly development,\nmaintenance, and revision of handcrafted features. Consequently, modeling\nhighly variable, heterogeneous patterns such as trader behavior is challenging.\nDeep learning promises a remedy. Learning hierarchical distributed\nrepresentations of the data in an automatic manner (e.g. risk taking behavior),\nit uncovers generative features that determine the target (e.g., trader's\nprofitability), avoids manual feature engineering, and is more robust toward\nchange (e.g. dynamic market conditions). The results of employing a deep\nnetwork for operational risk forecasting confirm the feature learning\ncapability of deep learning, provide guidance on designing a suitable network\narchitecture and demonstrate the superiority of deep learning over machine\nlearning and rule-based benchmarks."", 'Many important NLP problems can be posed as dual-sequence or\nsequence-to-sequence modeling tasks. Recent advances in building end-to-end\nneural architectures have been highly successful in solving such tasks. In this\nwork we propose a new architecture for dual-sequence modeling that is based on\nassociative memory. We derive AM-RNNs, a recurrent associative memory (AM)\nwhich augments generic recurrent neural networks (RNN). This architecture is\nextended to the Dual AM-RNN which operates on two AMs at once. Our models\nachieve very competitive results on textual entailment. A qualitative analysis\ndemonstrates that long range dependencies between source and target-sequence\ncan be bridged effectively using Dual AM-RNNs. However, an initial experiment\non auto-encoding reveals that these benefits are not exploited by the system\nwhen learning to solve sequence-to-sequence tasks which indicates that\nadditional supervision or regularization is needed.', 'Affiliate Marketing (AM) has become an important and cost effective tool for\ne-commerce. There are numerous risks and vulnerabilities that are typically\nassociated with AM. Though a well-planned AM model can greatly benefit the\ne-commerce strategies of an enterprise, a haphazardly implemented system can\nexpose a business enterprise to major risks and vulnerabilities, which can lead\nto great financial losses through fraudulent activities. This\nresearch-in-progress has identified some of the risks and the technical\nbackground of those scenarios. The research will now move on to build a\nfunctional prototype of an AM network to design and test solutions to control\nthe identified risks.', 'Medical terminology normalization aims to map the clinical mention to\nterminologies come from a knowledge base, which plays an important role in\nanalyzing Electronic Health Record(EHR) and many downstream tasks. In this\npaper, we focus on Chinese procedure terminology normalization. The expression\nof terminologies are various and one medical mention may be linked to multiple\nterminologies. Previous study explores some methods such as multi-class\nclassification or learning to rank(LTR) to sort the terminologies by literature\nand semantic information. However, these information is inadequate to find the\nright terminologies, particularly in multi-implication cases. In this work, we\npropose a combined recall and rank framework to solve the above problems. This\nframework is composed of a multi-task candidate generator(MTCG), a keywords\nattentive ranker(KAR) and a fusion block(FB). MTCG is utilized to predict the\nmention implication number and recall candidates with semantic similarity. KAR\nis based on Bert with a keywords attentive mechanism which focuses on keywords\nsuch as procedure sites and procedure types. FB merges the similarity come from\nMTCG and KAR to sort the terminologies from different perspectives. Detailed\nexperimental analysis shows our proposed framework has a remarkable improvement\non both performance and efficiency.', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', 'The ability to accurately quantify the performance an additively manufactured\n(AM) product is important for a widespread industry adoption of AM as the\ndesign is required to: (1) satisfy geometrical constraints, (2) satisfy\nstructural constraints dictated by its intended function, and (3) be cost\neffective compared to traditional manufacturing methods. Optimization\ntechniques offer design aids in creating cost-effective structures that meet\nthe prescribed structural objectives. The fundamental problem in existing\napproaches lies in the difficulty to quantify the structural performance as\neach unique design leads to a new set of analyses to determine the structural\nrobustness and such analyses can be very costly due to the complexity of in-use\nforces experienced by the structure. This work develops computationally\ntractable methods tailored to maximize the structural performance of AM\nproducts. A geometry preserving build orientation optimization method as well\nas data-driven shape optimization approaches to structural design are\npresented. Proposed methods greatly enhance the value of AM technology by\ntaking advantage of the design space enabled by it for a broad class of\nproblems involving complex in-use loads.', 'Programming by example is the problem of synthesizing a program from a small\nset of input / output pairs. Recent works applying machine learning methods to\nthis task show promise, but are typically reliant on generating synthetic\nexamples for training. A particular challenge lies in generating meaningful\nsets of inputs and outputs, which well-characterize a given program and\naccurately demonstrate its behavior. Where examples used for testing are\ngenerated by the same method as training data then the performance of a model\nmay be partly reliant on this similarity. In this paper we introduce a novel\napproach using an SMT solver to synthesize inputs which cover a diverse set of\nbehaviors for a given program. We carry out a case study comparing this method\nto existing synthetic data generation procedures in the literature, and find\nthat data generated using our approach improves both the discriminatory power\nof example sets and the ability of trained machine learning models to\ngeneralize to unfamiliar data.', 'The study of unsupervised learning can be generally divided into two\ncategories: imitation learning and reinforcement learning. In imitation\nlearning the machine learns by mimicking the behavior of an expert system\nwhereas in reinforcement learning the machine learns via direct environment\nfeedback. Traditional deep reinforcement learning takes a significant time\nbefore the machine starts to converge to an optimal policy. This paper proposes\nAugmented Q-Imitation-Learning, a method by which deep reinforcement learning\nconvergence can be accelerated by applying Q-imitation-learning as the initial\ntraining process in traditional Deep Q-learning.', 'The promise of Additive Manufacturing (AM) includes reduced transportation\nand warehousing costs, reduction of source material waste, and reduced\nenvironmental impact. AM is extremely useful for making prototypes and has\ndemonstrated the ability to manufacture complex parts not possible (or\nprohibitively expensive) with conventional machining. Scientists and\nmanufactures are finding increased uses for AM in creation of all types of\nfinished products including those built from polymers, biological material, and\nmetals. Although companies such as GE have been using 3D printing for Additive\nManufacturing for over thirty years to make mandrels for light bulb\nmanufacturing, application areas of Additive Manufacturing have increased\nsubstantially in recent years, particularly due to the reduction in cost of 3D\nprinters. Like most emergent technologies, there are bound to be growing pains\nwith AM. This paper looks at the software that supports AM and 3D printing and\ntheir vulnerability to cyber-attacks, intellectual property theft, defect rates\nof AM software (which can cause undesired consequences themselves and also\ncreate vulnerabilities that a hacker may exploit), part reliability and safety\nof devices incorporating 3D printed parts (when making mission critical parts),\nand security/throughput issues of computer networks. Literature searches,\nconsulting with technical experts and a relatively new hazard analysis\ntechnique will be used, one especially developed for software intensive systems\ncalled Systemic Theoretic Process Analysis (STPA). The purpose of this white\npaper is to identify risks (or hazards for mission critical parts) for AM in\nthis emergent stage so that mitigations can be applied before accidents occur.\nA second purpose of this white paper is to evaluate the effectiveness of STPA\nas a hazard analysis technique in a field that is still relatively new.', 'Additive Manufacturing (AM) is a manufacturing paradigm that builds\nthree-dimensional objects from a computer-aided design model by successively\nadding material layer by layer. AM has become very popular in the past decade\ndue to its utility for fast prototyping such as 3D printing as well as\nmanufacturing functional parts with complex geometries using processes such as\nlaser metal deposition that would be difficult to create using traditional\nmachining. As the process for creating an intricate part for an expensive metal\nsuch as Titanium is prohibitive with respect to cost, computational models are\nused to simulate the behavior of AM processes before the experimental run.\nHowever, as the simulations are computationally costly and time-consuming for\npredicting multiscale multi-physics phenomena in AM, physics-informed\ndata-driven machine-learning systems for predicting the behavior of AM\nprocesses are immensely beneficial. Such models accelerate not only multiscale\nsimulation tools but also empower real-time control systems using in-situ data.\nIn this paper, we design and develop essential components of a scientific\nframework for developing a data-driven model-based real-time control system.\nFinite element methods are employed for solving time-dependent heat equations\nand developing the database. The proposed framework uses extremely randomized\ntrees - an ensemble of bagged decision trees as the regression algorithm\niteratively using temperatures of prior voxels and laser information as inputs\nto predict temperatures of subsequent voxels. The models achieve mean absolute\npercentage errors below 1% for predicting temperature profiles for AM\nprocesses. The code is made available for the research community at\nhttps://github.com/paularindam/ml-iter-additive.', 'We address the problem of solving mixed random linear equations. We have\nunlabeled observations coming from multiple linear regressions, and each\nobservation corresponds to exactly one of the regression models. The goal is to\nlearn the linear regressors from the observations. Classically, Alternating\nMinimization (AM) (which is a variant of Expectation Maximization (EM)) is used\nto solve this problem. AM iteratively alternates between the estimation of\nlabels and solving the regression problems with the estimated labels.\nEmpirically, it is observed that, for a large variety of non-convex problems\nincluding mixed linear regression, AM converges at a much faster rate compared\nto gradient based algorithms. However, the existing theory suggests similar\nrate of convergence for AM and gradient based methods, failing to capture this\nempirical behavior. In this paper, we close this gap between theory and\npractice for the special case of a mixture of $2$ linear regressions. We show\nthat, provided initialized properly, AM enjoys a \\emph{super-linear} rate of\nconvergence in certain parameter regimes. To the best of our knowledge, this is\nthe first work that theoretically establishes such rate for AM. Hence, if we\nwant to recover the unknown regressors upto an error (in $\\ell_2$ norm) of\n$\\epsilon$, AM only takes $\\mathcal{O}(\\log \\log (1/\\epsilon))$ iterations.\nFurthermore, we compare AM with a gradient based heuristic algorithm\nempirically and show that AM dominates in iteration complexity as well as\nwall-clock time.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : As a data scientist, I am currently developing an early detection system designed to identify the onset of progressive neurodegenerative diseases through the analysis of brain scans. My current dataset consists of a large volume of unlabeled data, supplemented by a limited number of labeled positive examples. My goal is to create a model that can effectively learn from this data to ensure accurate detection. Rather than relying solely on importance reweighting during the model training phase, I am aiming to create a self-learning model. This would involve the model setting its own learning pace, determining its own loss, and gaining insights from its own performance during the training process. As such, the incorporation of some form of reinforcement learning might be necessary. If I go down the route of using RL, given the nature of my dataset, I am particularly interested in reinforcement learning algorithms for hyperparameter selection that can learn from sparse reward signals. Ultimately, my goal is for my trained model to outperform other methods when applied to specific datasets related to progressive neurodegenerative diseases, particularly those containing brain images.","[""Deep learning is attracting significant interest in the neuroimaging\ncommunity as a means to diagnose psychiatric and neurological disorders from\nstructural magnetic resonance images. However, there is a tendency amongst\nresearchers to adopt architectures optimized for traditional computer vision\ntasks, rather than design networks customized for neuroimaging data. We address\nthis by introducing NEURO-DRAM, a 3D recurrent visual attention model tailored\nfor neuroimaging classification. The model comprises an agent which, trained by\nreinforcement learning, learns to navigate through volumetric images,\nselectively attending to the most informative regions for a given task. When\napplied to Alzheimer's disease prediction, NEURODRAM achieves state-of-the-art\nclassification accuracy on an out-of-sample dataset, significantly\noutperforming a baseline convolutional neural network. When further applied to\nthe task of predicting which patients with mild cognitive impairment will be\ndiagnosed with Alzheimer's disease within two years, the model achieves\nstate-of-the-art accuracy with no additional training. Encouragingly, the agent\nlearns, without explicit instruction, a search policy in agreement with\nstandardized radiological hallmarks of Alzheimer's disease, suggesting a route\nto automated biomarker discovery for more poorly understood disorders."", 'Purpose: AI in radiology is hindered chiefly by: 1) Requiring large annotated\ndata sets. 2) Non-generalizability that limits deployment to new scanners /\ninstitutions. And 3) Inadequate explainability and interpretability. We believe\nthat reinforcement learning can address all three shortcomings, with robust and\nintuitive algorithms trainable on small datasets. To the best of our knowledge,\nreinforcement learning has not been directly applied to computer vision tasks\nfor radiological images. In this proof-of-principle work, we train a deep\nreinforcement learning network to predict brain tumor location.\n  Materials and Methods: Using the BraTS brain tumor imaging database, we\ntrained a deep Q network on 70 post-contrast T1-weighted 2D image slices. We\ndid so in concert with image exploration, with rewards and punishments designed\nto localize lesions. To compare with supervised deep learning, we trained a\nkeypoint detection convolutional neural network on the same 70 images. We\napplied both approaches to a separate 30 image testing set.\n  Results: Reinforcement learning predictions consistently improved during\ntraining, whereas those of supervised deep learning quickly diverged.\nReinforcement learning predicted testing set lesion locations with 85%\naccuracy, compared to roughly 7% accuracy for the supervised deep network.\n  Conclusion: Reinforcement learning predicted lesions with high accuracy,\nwhich is unprecedented for such a small training set. We believe that\nreinforcement learning can propel radiology AI well past the inherent\nlimitations of supervised deep learning, with more clinician-driven research\nand finally toward true clinical applicability.']","['Purpose Supervised deep learning in radiology suffers from notorious inherent\nlimitations: 1) It requires large, hand-annotated data sets, 2) It is\nnon-generalizable, and 3) It lacks explainability and intuition. We have\nrecently proposed Reinforcement Learning to address all threes. However, we\napplied it to images with radiologist eye tracking points, which limits the\nstate-action space. Here we generalize the Deep-Q Learning to a gridworld-based\nenvironment, so that only the images and image masks are required.\n  Materials and Methods We trained a Deep Q network on 30 two-dimensional image\nslices from the BraTS brain tumor database. Each image contained one lesion. We\nthen tested the trained Deep Q network on a separate set of 30 testing set\nimages. For comparison, we also trained and tested a keypoint detection\nsupervised deep learning network for the same set of training / testing images.\n  Results Whereas the supervised approach quickly overfit the training data,\nand predicably performed poorly on the testing set (11\\% accuracy), the Deep-Q\nlearning approach showed progressive improved generalizability to the testing\nset over training time, reaching 70\\% accuracy.\n  Conclusion We have shown a proof-of-principle application of reinforcement\nlearning to radiological images, here using 2D contrast-enhanced MRI brain\nimages with the goal of localizing brain tumors. This represents a\ngeneralization of recent work to a gridworld setting, naturally suitable for\nanalyzing medical images.', 'Deep Reinforcement Learning (DRL) is a promising approach for teaching robots\nnew behaviour. However, one of its main limitations is the need for carefully\nhand-coded reward signals by an expert. We argue that it is crucial to automate\nthe reward learning process so that new skills can be taught to robots by their\nusers. To address such automation, we consider task success classifiers using\nvisual observations to estimate the rewards in terms of task success. In this\nwork, we study the performance of multiple state-of-the-art deep reinforcement\nlearning algorithms under different types of reward: Dense, Sparse, Visual\nDense, and Visual Sparse rewards. Our experiments in various simulation tasks\n(Pendulum, Reacher, Pusher, and Fetch Reach) show that while DRL agents can\nlearn successful behaviours using visual rewards when the goal targets are\ndistinguishable, their performance may decrease if the task goal is not clearly\nvisible. Our results also show that visual dense rewards are more successful\nthan visual sparse rewards and that there is no single best algorithm for all\ntasks.', ""Parkinson's disease (PD) is a degenerative and progressive neurological\ncondition. Early diagnosis can improve treatment for patients and is performed\nthrough dopaminergic imaging techniques like the SPECT DaTscan. In this study,\nwe propose a machine learning model that accurately classifies any given\nDaTscan as having Parkinson's disease or not, in addition to providing a\nplausible reason for the prediction. This is kind of reasoning is done through\nthe use of visual indicators generated using Local Interpretable Model-Agnostic\nExplainer (LIME) methods. DaTscans were drawn from the Parkinson's Progression\nMarkers Initiative database and trained on a CNN (VGG16) using transfer\nlearning, yielding an accuracy of 95.2%, a sensitivity of 97.5%, and a\nspecificity of 90.9%. Keeping model interpretability of paramount importance,\nespecially in the healthcare field, this study utilises LIME explanations to\ndistinguish PD from non-PD, using visual superpixels on the DaTscans. It could\nbe concluded that the proposed system, in union with its measured\ninterpretability and accuracy may effectively aid medical workers in the early\ndiagnosis of Parkinson's Disease."", 'We systematically evaluate the performance of deep learning models in the\npresence of diseases not labeled for or present during training. First, we\nevaluate whether deep learning models trained on a subset of diseases (seen\ndiseases) can detect the presence of any one of a larger set of diseases. We\nfind that models tend to falsely classify diseases outside of the subset\n(unseen diseases) as ""no disease"". Second, we evaluate whether models trained\non seen diseases can detect seen diseases when co-occurring with diseases\noutside the subset (unseen diseases). We find that models are still able to\ndetect seen diseases even when co-occurring with unseen diseases. Third, we\nevaluate whether feature representations learned by models may be used to\ndetect the presence of unseen diseases given a small labeled set of unseen\ndiseases. We find that the penultimate layer of the deep neural network\nprovides useful features for unseen disease detection. Our results can inform\nthe safe clinical deployment of deep learning models trained on a\nnon-exhaustive set of disease classes.', ""Alzheimer's disease (AD) is a neurodegenerative disease that affects nearly\n50 million individuals across the globe and is one of the leading causes of\ndeaths globally. It is projected that by 2050, the number of people affected by\nthe disease would more than double. Consequently, the growing advancements in\ntechnology beg the question, can technology be used to predict Alzheimer's for\na better and early diagnosis? In this paper, we focus on this very problem.\nSpecifically, we have trained both ML models and neural networks to predict and\nclassify participants based on their speech patterns. We computed a number of\nlinguistic variables using DementiaBank's Pitt Corpus, a database consisting of\ntranscripts of interviews with subjects suffering from multiple\nneurodegenerative diseases. We then trained both binary classifiers, as well as\nmulticlass classifiers to distinguish AD from normal aging and other\nneurodegenerative diseases. We also worked on establishing the link between\nspecific speech factors that can help determine the onset of AD. Confusion\nmatrices and feature importance graphs have been plotted model-wise to compare\nthe performances of our models. In both multiclass and binary classification,\nneural networks were found to outperform the other models with a testing\naccuracy of 76.44% and 92.05% respectively. For the feature importance, it was\nconcluded that '%_PRESP' (present participle), '%_3S' (3rd person present tense\nmarkers) were two of the most important speech features for our classifiers in\npredicting AD."", ""Alzheimer's disease (AD) is a progressive and incurable neurodegenerative\ndisease which destroys brain cells and causes loss to patient's memory. An\nearly detection can prevent the patient from further damage of the brain cells\nand hence avoid permanent memory loss. In past few years, various automatic\ntools and techniques have been proposed for diagnosis of AD. Several methods\nfocus on fast, accurate and early detection of the disease to minimize the loss\nto patients mental health. Although machine learning and deep learning\ntechniques have significantly improved medical imaging systems for AD by\nproviding diagnostic performance close to human level. But the main problem\nfaced during multi-class classification is the presence of highly correlated\nfeatures in the brain structure. In this paper, we have proposed a smart and\naccurate way of diagnosing AD based on a two-dimensional deep convolutional\nneural network (2D-DCNN) using imbalanced three-dimensional MRI dataset.\nExperimental results on Alzheimer Disease Neuroimaging Initiative magnetic\nresonance imaging (MRI) dataset confirms that the proposed 2D-DCNN model is\nsuperior in terms of accuracy, efficiency, and robustness. The model classifies\nMRI into three categories: AD, mild cognitive impairment, and normal control:\nand has achieved 99.89% classification accuracy with imbalanced classes. The\nproposed model exhibits noticeable improvement in accuracy as compared to the\nstate-fo-the-art methods."", 'Deep reinforcement learning (RL) can acquire complex behaviors from low-level\ninputs, such as images. However, real-world applications of such methods\nrequire generalizing to the vast variability of the real world. Deep networks\nare known to achieve remarkable generalization when provided with massive\namounts of labeled data, but can we provide this breadth of experience to an RL\nagent, such as a robot? The robot might continuously learn as it explores the\nworld around it, even while deployed. However, this learning requires access to\na reward function, which is often hard to measure in real-world domains, where\nthe reward could depend on, for example, unknown positions of objects or the\nemotional state of the user. Conversely, it is often quite practical to provide\nthe agent with reward functions in a limited set of situations, such as when a\nhuman supervisor is present or in a controlled setting. Can we make use of this\nlimited supervision, and still benefit from the breadth of experience an agent\nmight collect on its own? In this paper, we formalize this problem as\nsemisupervised reinforcement learning, where the reward function can only be\nevaluated in a set of ""labeled"" MDPs, and the agent must generalize its\nbehavior to the wide range of states it might encounter in a set of ""unlabeled""\nMDPs, by using experience from both settings. Our proposed method infers the\ntask objective in the unlabeled MDPs through an algorithm that resembles\ninverse RL, using the agent\'s own prior experience in the labeled MDPs as a\nkind of demonstration of optimal behavior. We evaluate our method on\nchallenging tasks that require control directly from images, and show that our\napproach can improve the generalization of a learned deep neural network policy\nby using experience for which no reward function is available. We also show\nthat our method outperforms direct supervised learning of the reward.', 'With the advents of deep learning, improved image classification with complex\ndiscriminative models has been made possible. However, such deep models with\nincreased complexity require a huge set of labeled samples to generalize the\ntraining. Such classification models can easily overfit when applied for\nmedical images because of limited training data, which is a common problem in\nthe field of medical image analysis. This paper proposes and investigates a\nreinforced classifier for improving the generalization under a few available\ntraining data. Partially following the idea of reinforcement learning, the\nproposed classifier uses a generalization-feedback from a subset of the\ntraining data to update its parameter instead of only using the conventional\ncross-entropy loss about the training data. We evaluate the improvement of the\nproposed classifier by applying it on three different classification problems\nagainst the standard deep classifiers equipped with existing\noverfitting-prevention techniques. Besides an overall improvement in\nclassification performance, the proposed classifier showed remarkable\ncharacteristics of generalized learning, which can have great potential in\nmedical classification tasks.', ""In this paper, we propose a principled deep reinforcement learning (RL)\napproach that is able to accelerate the convergence rate of general deep neural\nnetworks (DNNs). With our approach, a deep RL agent (synonym for optimizer in\nthis work) is used to automatically learn policies about how to schedule\nlearning rates during the optimization of a DNN. The state features of the\nagent are learned from the weight statistics of the optimizee during training.\nThe reward function of this agent is designed to learn policies that minimize\nthe optimizee's training time given a certain performance goal. The actions of\nthe agent correspond to changing the learning rate for the optimizee during\ntraining. As far as we know, this is the first attempt to use deep RL to learn\nhow to optimize a large-sized DNN. We perform extensive experiments on a\nstandard benchmark dataset and demonstrate the effectiveness of the policies\nlearned by our approach."", ""Diffusion tensor imaging (DTI) has been used to study the effects of\nneurodegenerative diseases on neural pathways, which may lead to more reliable\nand early diagnosis of these diseases as well as a better understanding of how\nthey affect the brain. We introduce an intelligent visual analytics system for\nstudying patient groups based on their labeled DTI fiber tract data and\ncorresponding statistics. The system's AI-augmented interface guides the user\nthrough an organized and holistic analysis space, including the statistical\nfeature space, the physical space, and the space of patients over different\ngroups. We use a custom machine learning pipeline to help narrow down this\nlarge analysis space, and then explore it pragmatically through a range of\nlinked visualizations. We conduct several case studies using real data from the\nresearch database of Parkinson's Progression Markers Initiative."", 'We model Alzheimer\'s disease (AD) progression by combining differential\nequations (DEs) and reinforcement learning (RL) with domain knowledge. DEs\nprovide relationships between some, but not all, factors relevant to AD. We\nassume that the missing relationships must satisfy general criteria about the\nworking of the brain, for e.g., maximizing cognition while minimizing the cost\nof supporting cognition. This allows us to extract the missing relationships by\nusing RL to optimize an objective (reward) function that captures the above\ncriteria. We use our model consisting of DEs (as a simulator) and the trained\nRL agent to predict individualized 10-year AD progression using baseline (year\n0) features on synthetic and real data. The model was comparable or better at\npredicting 10-year cognition trajectories than state-of-the-art learning-based\nmodels. Our interpretable model demonstrated, and provided insights into,\n""recovery/compensatory"" processes that mitigate the effect of AD, even though\nthose processes were not explicitly encoded in the model. Our framework\ncombines DEs with RL for modelling AD progression and has broad applicability\nfor understanding other neurological disorders.', 'Machine learning applications in medical imaging are frequently limited by\nthe lack of quality labeled data. In this paper, we explore the self training\nmethod, a form of semi-supervised learning, to address the labeling burden. By\nintegrating reinforcement learning, we were able to expand the application of\nself training to complex segmentation networks without any further human\nannotation. The proposed approach, reinforced self training (ReST), fine tunes\na semantic segmentation networks by introducing a policy network that learns to\ngenerate pseudolabels. We incorporate an expert demonstration network, based on\ninverse reinforcement learning, to enhance clinical validity and convergence of\nthe policy network. The model was tested on a pulmonary nodule segmentation\ntask in chest X-rays and achieved the performance of a standard U-Net while\nusing only 50% of the labeled data, by exploiting unlabeled data. When the same\nnumber of labeled data was used, a moderate to significant cross validation\naccuracy improvement was achieved depending on the absolute number of labels\nused.', 'Dynamic treatment recommendation systems based on large-scale electronic\nhealth records (EHRs) become a key to successfully improve practical clinical\noutcomes. Prior relevant studies recommend treatments either use supervised\nlearning (e.g. matching the indicator signal which denotes doctor\nprescriptions), or reinforcement learning (e.g. maximizing evaluation signal\nwhich indicates cumulative reward from survival rates). However, none of these\nstudies have considered to combine the benefits of supervised learning and\nreinforcement learning. In this paper, we propose Supervised Reinforcement\nLearning with Recurrent Neural Network (SRL-RNN), which fuses them into a\nsynergistic learning framework. Specifically, SRL-RNN applies an off-policy\nactor-critic framework to handle complex relations among multiple medications,\ndiseases and individual characteristics. The ""actor"" in the framework is\nadjusted by both the indicator signal and evaluation signal to ensure effective\nprescription and low mortality. RNN is further utilized to solve the\nPartially-Observed Markov Decision Process (POMDP) problem due to the lack of\nfully observed states in real world applications. Experiments on the publicly\nreal-world dataset, i.e., MIMIC-3, illustrate that our model can reduce the\nestimated mortality, while providing promising accuracy in matching doctors\'\nprescriptions.', 'Background: Alzheimers disease is a progressive neurodegenerative disorder\nand the main cause of dementia in aging. Hippocampus is prone to changes in the\nearly stages of Alzheimers disease. Detection and observation of the\nhippocampus changes using magnetic resonance imaging (MRI) before the onset of\nAlzheimers disease leads to the faster preventive and therapeutic measures.\nObjective: The aim of this study was the segmentation of the hippocampus in\nmagnetic resonance (MR) images of Alzheimers patients using deep machine\nlearning method. Methods: U-Net architecture of convolutional neural network\nwas proposed to segment the hippocampus in the real MRI data. The MR images of\nthe 100 and 35 patients available in Alzheimers disease Neuroimaging Initiative\n(ADNI) dataset, was used for the train and test of the model, respectively. The\nperformance of the proposed method was compared with manual segmentation by\nmeasuring the similarity metrics. Results: The desired segmentation achieved\nafter 10 iterations. A Dice similarity coefficient (DSC) = 92.3%, sensitivity =\n96.5%, positive predicted value (PPV) = 90.4%, and Intersection over Union\n(IoU) value for the train 92.94 and test 92.93 sets were obtained which are\nacceptable. Conclusion: The proposed approach is promising and can be extended\nin the prognosis of Alzheimers disease by the prediction of the hippocampus\nvolume changes in the early stage of the disease.', ""Parkinson's Disease (PD) is a neurodegenerative disease that currently does\nnot have a cure. In order to facilitate disease management and reduce the speed\nof symptom progression, early diagnosis is essential. The current clinical,\ndiagnostic approach is to have radiologists perform human visual analysis of\nthe degeneration of dopaminergic neurons in the substantia nigra region of the\nbrain. Clinically, dopamine levels are monitored through observing dopamine\ntransporter (DaT) activity. One method of DaT activity analysis is performed\nwith the injection of an Iodine-123 fluoropropyl (123I-FP-CIT) tracer combined\nwith single photon emission computerized tomography (SPECT) imaging. The tracer\nillustrates the region of interest in the resulting DaTscan SPECT images. Human\nvisual analysis is slow and vulnerable to subjectivity between radiologists, so\nthe goal was to develop an introductory implementation of a deep convolutional\nneural network that can objectively and accurately classify DaTscan SPECT\nimages as Parkinson's Disease or normal. This study illustrates the approach of\nusing a deep convolutional neural network and evaluates its performance on\nDaTscan SPECT image classification. The data used in this study was obtained\nthrough a database provided by the Parkinson's Progression Markers Initiative\n(PPMI). The deep neural network in this study utilizes the InceptionV3\narchitecture, 1st runner up in the 2015 ImageNet Large Scale Visual Recognition\nCompetition (ILSVRC), as a base model. A custom, binary classifier block was\nadded on top of this base. In order to account for the small dataset size, a\nten fold cross validation was implemented to evaluate the model's performance."", ""Dementia is a group of irreversible, chronic, and progressive\nneurodegenerative disorders resulting in impaired memory, communication, and\nthought processes. In recent years, clinical research advances in brain aging\nhave focused on the earliest clinically detectable stage of incipient dementia,\ncommonly known as mild cognitive impairment (MCI). Currently, these disorders\nare diagnosed using a manual analysis of neuropsychological examinations. We\nmeasure the feasibility of using the linguistic characteristics of verbal\nutterances elicited during neuropsychological exams of elderly subjects to\ndistinguish between elderly control groups, people with MCI, people diagnosed\nwith possible Alzheimer's disease (AD), and probable AD. We investigated the\nperformance of both theory-driven psycholinguistic features and data-driven\ncontextual language embeddings in identifying different clinically diagnosed\ngroups. Our experiments show that a combination of contextual and\npsycholinguistic features extracted by a Support Vector Machine improved\ndistinguishing the verbal utterances of elderly controls, people with MCI,\npossible AD, and probable AD. This is the first work to identify four clinical\ndiagnosis groups of dementia in a highly imbalanced dataset. Our work shows\nthat machine learning algorithms built on contextual and psycholinguistic\nfeatures can learn the linguistic biomarkers from verbal utterances and assist\nclinical diagnosis of different stages and types of dementia, even with limited\ndata."", 'Purpose: Lesion segmentation in medical imaging is key to evaluating\ntreatment response. We have recently shown that reinforcement learning can be\napplied to radiological images for lesion localization. Furthermore, we\ndemonstrated that reinforcement learning addresses important limitations of\nsupervised deep learning; namely, it can eliminate the requirement for large\namounts of annotated training data and can provide valuable intuition lacking\nin supervised approaches. However, we did not address the fundamental task of\nlesion/structure-of-interest segmentation. Here we introduce a method combining\nunsupervised deep learning clustering with reinforcement learning to segment\nbrain lesions on MRI.\n  Materials and Methods: We initially clustered images using unsupervised deep\nlearning clustering to generate candidate lesion masks for each MRI image. The\nuser then selected the best mask for each of 10 training images. We then\ntrained a reinforcement learning algorithm to select the masks. We tested the\ncorresponding trained deep Q network on a separate testing set of 10 images.\nFor comparison, we also trained and tested a U-net supervised deep learning\nnetwork on the same set of training/testing images.\n  Results: Whereas the supervised approach quickly overfit the training data\nand predictably performed poorly on the testing set (16% average Dice score),\nthe unsupervised deep clustering and reinforcement learning achieved an average\nDice score of 83%.\n  Conclusion: We have demonstrated a proof-of-principle application of\nunsupervised deep clustering and reinforcement learning to segment brain\ntumors. The approach represents human-allied AI that requires minimal input\nfrom the radiologist without the need for hand-traced annotation.', 'Rapid advances of hardware-based technologies during the past decades have\nopened up new possibilities for Life scientists to gather multimodal data in\nvarious application domains (e.g., Omics, Bioimaging, Medical Imaging, and\n[Brain/Body]-Machine Interfaces), thus generating novel opportunities for\ndevelopment of dedicated data intensive machine learning techniques. Overall,\nrecent research in Deep learning (DL), Reinforcement learning (RL), and their\ncombination (Deep RL) promise to revolutionize Artificial Intelligence. The\ngrowth in computational power accompanied by faster and increased data storage\nand declining computing costs have already allowed scientists in various fields\nto apply these techniques on datasets that were previously intractable for\ntheir size and complexity. This review article provides a comprehensive survey\non the application of DL, RL, and Deep RL techniques in mining Biological data.\nIn addition, we compare performances of DL techniques when applied to different\ndatasets across various application domains. Finally, we outline open issues in\nthis challenging research area and discuss future development perspectives.', ""Artificial Intelligence allows the improvement of our daily life, for\ninstance, speech and handwritten text recognition, real time translation and\nweather forecasting are common used applications. In the livestock sector,\nmachine learning algorithms have the potential for early detection and warning\nof problems, which represents a significant milestone in the poultry industry.\nProduction problems generate economic loss that could be avoided by acting in a\ntimely manner. In the current study, training and testing of support vector\nmachines are addressed, for an early detection of problems in the production\ncurve of commercial eggs, using farm's egg production data of 478,919 laying\nhens grouped in 24 flocks. Experiments using support vector machines with a 5\nk-fold cross-validation were performed at different previous time intervals, to\nalert with up to 5 days of forecasting interval, whether a flock will\nexperience a problem in production curve. Performance metrics such as accuracy,\nspecificity, sensitivity, and positive predictive value were evaluated,\nreaching 0-day values of 0.9874, 0.9876, 0.9783 and 0.6518 respectively on\nunseen data (test-set). The optimal forecasting interval was from zero to three\ndays, performance metrics decreases as the forecasting interval is increased.\nIt should be emphasized that this technique was able to issue an alert a day in\nadvance, achieving an accuracy of 0.9854, a specificity of 0.9865, a\nsensitivity of 0.9333 and a positive predictive value of 0.6135. This novel\napplication embedded in a computer system of poultry management is able to\nprovide significant improvements in early detection and warning of problems\nrelated to the production curve."", 'The emergence of continuous health monitoring and the availability of an\nenormous amount of time series data has provided a great opportunity for the\nadvancement of personal health tracking. In recent years, unsupervised learning\nmethods have drawn special attention of researchers to tackle the sparse\nannotation of health data and real-time detection of anomalies has been a\ncentral problem of interest. However, one problem that has not been well\naddressed before is the early prediction of forthcoming negative health events.\nEarly signs of an event can introduce subtle and gradual changes in the health\nsignal prior to its onset, detection of which can be invaluable in effective\nprevention. In this study, we first demonstrate our observations on the\nshortcoming of widely adopted anomaly detection methods in uncovering the\nchanges prior to a negative health event. We then propose a framework which\nrelies on online clustering of signal segment representations which are\nautomatically learned by a specially designed LSTM auto-encoder. We show the\neffectiveness of our approach by predicting Bradycardia events in infants using\nMIT-PICS dataset 1.3 minutes ahead of time with 68\\% AUC score on average,\nusing no label supervision. Results of our study can indicate the viability of\nour approach in the early detection of health events in other applications as\nwell.', ""Early detection of lung nodules is of great importance in lung cancer\nscreening. Existing research recognizes the critical role played by CAD systems\nin early detection and diagnosis of lung nodules. However, many CAD systems,\nwhich are used as cancer detection tools, produce a lot of false positives (FP)\nand require a further FP reduction step. Furthermore, guidelines for early\ndiagnosis and treatment of lung cancer are consist of different shape and\nvolume measurements of abnormalities. Segmentation is at the heart of our\nunderstanding of nodules morphology making it a major area of interest within\nthe field of computer aided diagnosis systems. This study set out to test the\nhypothesis that joint learning of false positive (FP) nodule reduction and\nnodule segmentation can improve the computer aided diagnosis (CAD) systems'\nperformance on both tasks. To support this hypothesis we propose a 3D deep\nmulti-task CNN to tackle these two problems jointly. We tested our system on\nLUNA16 dataset and achieved an average dice similarity coefficient (DSC) of 91%\nas segmentation accuracy and a score of nearly 92% for FP reduction. As a proof\nof our hypothesis, we showed improvements of segmentation and FP reduction\ntasks over two baselines. Our results support that joint training of these two\ntasks through a multi-task learning approach improves system performance on\nboth. We also showed that a semi-supervised approach can be used to overcome\nthe limitation of lack of labeled data for the 3D segmentation task."", 'In this paper we present a Recurrent neural networks (RNN) based architecture\nthat achieves an AUCROC of 0.9147 for predicting the onset of Congestive Heart\nFailure (CHF) 15 months in advance using a 12-month observation window on a\nlarge cohort of 216,394 patients. We believe this to be the largest study in\nCHF onset prediction with respect to the number of CHF case patients in the\ncohort and the test set (3,332 CHF patients) on which the AUC metrics are\nreported. We explore the extent to which LSTM (Long Short Term Memory) based\nmodel, a variant of RNNs, can accurately predict the onset of CHF when compared\nto known linear baselines like Logistic Regression, Random Forests and deep\nlearning based models such as Multi-Layer Perceptron and Convolutional Neural\nNetworks. We utilize demographics, medical diagnosis and procedure data from\n21,405 CHF and 194,989 control patients to as our features. We describe our\nfeature embedding strategy for medical diagnosis codes that accommodates the\nsparse, irregular, longitudinal, and high-dimensional characteristics of EHR\ndata. We empirically show that LSTMs can capture the longitudinal aspects of\nEHR data better than the proposed baselines. As an attempt to interpret the\nmodel, we present a temporal data analysis-based technique on false positives\nto attribute feature importance. A model capable of predicting the onset of\ncongestive heart failure months in the future with this level of accuracy and\nprecision can support efforts of practitioners to implement risk factor\nreduction strategies and researchers to begin to systematically evaluate\ninterventions to potentially delay or avert development of the disease with\nhigh mortality, morbidity and significant costs.', 'Most modern neural networks for classification fail to take into account the\nconcept of the unknown. Trained neural networks are usually tested in an\nunrealistic scenario with only examples from a closed set of known classes. In\nan attempt to develop a more realistic model, the concept of working in an open\nset environment has been introduced. This in turn leads to the concept of\nincremental learning where a model with its own architecture and initial\ntrained set of data can identify unknown classes during the testing phase and\nautonomously update itself if evidence of a new class is detected. Some\nproblems that arise in incremental learning are inefficient use of resources to\nretrain the classifier repeatedly and the decrease of classification accuracy\nas multiple classes are added over time. This process of instantiating new\nclasses is repeated as many times as necessary, accruing errors. To address\nthese problems, this paper proposes the Classification Confidence Threshold\napproach to prime neural networks for incremental learning to keep accuracies\nhigh by limiting forgetting. A lean method is also used to reduce resources\nused in the retraining of the neural network. The proposed method is based on\nthe idea that a network is able to incrementally learn a new class even when\nexposed to a limited number samples associated with the new class. This method\ncan be applied to most existing neural networks with minimal changes to network\narchitecture.', 'In recent years deep reinforcement learning (RL) systems have attained\nsuperhuman performance in a number of challenging task domains. However, a\nmajor limitation of such applications is their demand for massive amounts of\ntraining data. A critical present objective is thus to develop deep RL methods\nthat can adapt rapidly to new tasks. In the present work we introduce a novel\napproach to this challenge, which we refer to as deep meta-reinforcement\nlearning. Previous work has shown that recurrent networks can support\nmeta-learning in a fully supervised context. We extend this approach to the RL\nsetting. What emerges is a system that is trained using one RL algorithm, but\nwhose recurrent dynamics implement a second, quite separate RL procedure. This\nsecond, learned RL algorithm can differ from the original one in arbitrary\nways. Importantly, because it is learned, it is configured to exploit structure\nin the training domain. We unpack these points in a series of seven\nproof-of-concept experiments, each of which examines a key aspect of deep\nmeta-RL. We consider prospects for extending and scaling up the approach, and\nalso point out some potentially important implications for neuroscience.', ""Alzheimer's disease (AD) is an irreversible devastative neurodegenerative\ndisorder associated with progressive impairment of memory and cognitive\nfunctions. Its early diagnosis is crucial for the development of possible\nfuture treatment option(s). Structural magnetic resonance images (sMRI) plays\nan important role to help in understanding the anatomical changes related to AD\nespecially in its early stages. Conventional methods require the expertise of\ndomain experts and extract hand-picked features such as gray matter\nsubstructures and train a classifier to distinguish AD subjects from healthy\nsubjects. Different from these methods, this paper proposes to construct\nmultiple deep 2D convolutional neural networks (2D-CNNs) to learn the various\nfeatures from local brain images which are combined to make the final\nclassification for AD diagnosis. The whole brain image was passed through two\ntransfer learning architectures; Inception version 3 and Xception; as well as\ncustom Convolutional Neural Network (CNN) built with the help of separable\nconvolutional layers which can automatically learn the generic features from\nimaging data for classification. Our study is conducted using cross-sectional\nT1-weighted structural MRI brain images from Open Access Series of Imaging\nStudies (OASIS) database to maintain the size and contrast over different MRI\nscans. Experimental results show that the transfer learning approaches exceed\nthe performance of non-transfer learning based approaches demonstrating the\neffectiveness of these approaches for the binary AD classification task."", 'While recent progress in deep reinforcement learning has enabled robots to\nlearn complex behaviors, tasks with long horizons and sparse rewards remain an\nongoing challenge. In this work, we propose an effective reward shaping method\nthrough predictive coding to tackle sparse reward problems. By learning\npredictive representations offline and using these representations for reward\nshaping, we gain access to reward signals that understand the structure and\ndynamics of the environment. In particular, our method achieves better learning\nby providing reward signals that 1) understand environment dynamics 2)\nemphasize on features most useful for learning 3) resist noise in learned\nrepresentations through reward accumulation. We demonstrate the usefulness of\nthis approach in different domains ranging from robotic manipulation to\nnavigation, and we show that reward signals produced through predictive coding\nare as effective for learning as hand-crafted rewards.', 'Over the past decade a wide spectrum of machine learning models have been\ndeveloped to model the neurodegenerative diseases, associating biomarkers,\nespecially non-intrusive neuroimaging markers, with key clinical scores\nmeasuring the cognitive status of patients. Multi-task learning (MTL) has been\ncommonly utilized by these studies to address high dimensionality and small\ncohort size challenges. However, most existing MTL approaches are based on\nlinear models and suffer from two major limitations: 1) they cannot explicitly\nconsider upper/lower bounds in these clinical scores; 2) they lack the\ncapability to capture complicated non-linear interactions among the variables.\nIn this paper, we propose Subspace Network, an efficient deep modeling approach\nfor non-linear multi-task censored regression. Each layer of the subspace\nnetwork performs a multi-task censored regression to improve upon the\npredictions from the last layer via sketching a low-dimensional subspace to\nperform knowledge transfer among learning tasks. Under mild assumptions, for\neach layer the parametric subspace can be recovered using only one pass of\ntraining data. Empirical results demonstrate that the proposed subspace network\nquickly picks up the correct parameter subspaces, and outperforms\nstate-of-the-arts in predicting neurodegenerative clinical scores using\ninformation in brain imaging.', ""To study neurodegenerative diseases, longitudinal studies are carried on\nvolunteer patients. During a time span of several months to several years, they\ngo through regular medical visits to acquire data from different modalities,\nsuch as biological samples, cognitive tests, structural and functional imaging.\nThese variables are heterogeneous but they all depend on the patient's health\ncondition, meaning that there are possibly unknown relationships between all\nmodalities. Some information may be specific to some modalities, others may be\ncomplementary, and others may be redundant. Some data may also be missing. In\nthis work we present a neural network architecture for multimodal learning,\nable to use imaging and clinical data from two time points to predict the\nevolution of a neurodegenerative disease, and robust to missing values. Our\nmultimodal network achieves 92.5\\% accuracy and an AUC score of 0.978 over a\ntest set of 57 subjects. We also show the superiority of the multimodal\narchitecture, for up to 37.5\\% of missing values in test set subjects' clinical\nmeasurements, compared to a model using only the clinical modality."", ""We introduce Disease Knowledge Transfer (DKT), a novel technique for\ntransferring biomarker information between related neurodegenerative diseases.\nDKT infers robust multimodal biomarker trajectories in rare neurodegenerative\ndiseases even when only limited, unimodal data is available, by transferring\ninformation from larger multimodal datasets from common neurodegenerative\ndiseases. DKT is a joint-disease generative model of biomarker progressions,\nwhich exploits biomarker relationships that are shared across diseases. Our\nproposed method allows, for the first time, the estimation of plausible,\nmultimodal biomarker trajectories in Posterior Cortical Atrophy (PCA), a rare\nneurodegenerative disease where only unimodal MRI data is available. For this\nwe train DKT on a combined dataset containing subjects with two distinct\ndiseases and sizes of data available: 1) a larger, multimodal typical AD (tAD)\ndataset from the TADPOLE Challenge, and 2) a smaller unimodal Posterior\nCortical Atrophy (PCA) dataset from the Dementia Research Centre (DRC), for\nwhich only a limited number of Magnetic Resonance Imaging (MRI) scans are\navailable. Although validation is challenging due to lack of data in PCA, we\nvalidate DKT on synthetic data and two patient datasets (TADPOLE and PCA\ncohorts), showing it can estimate the ground truth parameters in the simulation\nand predict unseen biomarkers on the two patient datasets. While we\ndemonstrated DKT on Alzheimer's variants, we note DKT is generalisable to other\nforms of related neurodegenerative diseases. Source code for DKT is available\nonline: https://github.com/mrazvan22/dkt."", ""Alzheimer's disease is the most common dementia leading to an irreversible\nneurodegenerative process. To date, subject revealed advanced brain structural\nalterations when the diagnosis is established. Therefore, an earlier diagnosis\nof this dementia is crucial although it is a challenging task. Recently, many\nstudies have proposed biomarkers to perform early detection of Alzheimer's\ndisease. Some of them have proposed methods based on inter-subject similarity\nwhile other approaches have investigated framework using intra-subject\nvariability. In this work, we propose a novel framework combining both\napproaches within an efficient graph of brain structures grading. Subsequently,\nwe demonstrate the competitive performance of the proposed method compared to\nstate-of-the-art methods."", 'This doctoral thesis covers some of my advances in electron microscopy with\ndeep learning. Highlights include a comprehensive review of deep learning in\nelectron microscopy; large new electron microscopy datasets for machine\nlearning, dataset search engines based on variational autoencoders, and\nautomatic data clustering by t-distributed stochastic neighbour embedding;\nadaptive learning rate clipping to stabilize learning; generative adversarial\nnetworks for compressed sensing with spiral, uniformly spaced and other fixed\nsparse scan paths; recurrent neural networks trained to piecewise adapt sparse\nscan paths to specimens by reinforcement learning; improving signal-to-noise;\nand conditional generative adversarial networks for exit wavefunction\nreconstruction from single transmission electron micrographs. This thesis adds\nto my publications by presenting their relationships, reflections, and holistic\nconclusions. This version of my thesis is typeset for online dissemination to\nimprove readability, whereas the thesis submitted to the University of Warwick\nin support of my application for the degree of Doctor of Philosophy in Physics\nis typeset for physical printing and binding.', 'While current machine learning models have impressive performance over a wide\nrange of applications, their large size and complexity render them unsuitable\nfor tasks such as remote monitoring on edge devices with limited storage and\ncomputational power. A naive approach to resolve this on the model level is to\nuse simpler architectures, but this sacrifices prediction accuracy and is\nunsuitable for monitoring applications requiring accurate detection of the\nonset of adverse events. In this paper, we propose an alternative solution to\nthis problem by decomposing the predictive model as the sum of a simple\nfunction which serves as a local monitoring tool, and a complex correction term\nto be evaluated on the server. A sign requirement is imposed on the latter to\nensure that the local monitoring function is safe, in the sense that it can\neffectively serve as an early warning system. Our analysis quantifies the\ntrade-offs between model complexity and performance, and serves as a guidance\nfor architecture design. We validate our proposed framework on a series of\nmonitoring experiments, where we succeed at learning monitoring models with\nsignificantly reduced complexity that minimally violate the safety requirement.\nMore broadly, our framework is useful for learning classifiers in applications\nwhere false negatives are significantly more costly compared to false\npositives.', 'To enable a deep learning-based system to be used in the medical domain as a\ncomputer-aided diagnosis system, it is essential to not only classify diseases\nbut also present the locations of the diseases. However, collecting\ninstance-level annotations for various thoracic diseases is expensive.\nTherefore, weakly supervised localization methods have been proposed that use\nonly image-level annotation. While the previous methods presented the disease\nlocation as the most discriminative part for classification, this causes a deep\nnetwork to localize wrong areas for indistinguishable X-ray images. To solve\nthis issue, we propose a spatial attention method using disease masks that\ndescribe the areas where diseases mainly occur. We then apply the spatial\nattention to find the precise disease area by highlighting the highest\nprobability of disease occurrence. Meanwhile, the various sizes, rotations and\nnoise in chest X-ray images make generating the disease masks challenging. To\nreduce the variation among images, we employ an alignment module to transform\nan input X-ray image into a generalized image. Through extensive experiments on\nthe NIH-Chest X-ray dataset with eight kinds of diseases, we show that the\nproposed method results in superior localization performances compared to\nstate-of-the-art methods.', 'Cancer is a complex disease, the understanding and treatment of which are\nbeing aided through increases in the volume of collected data and in the scale\nof deployed computing power. Consequently, there is a growing need for the\ndevelopment of data-driven and, in particular, deep learning methods for\nvarious tasks such as cancer diagnosis, detection, prognosis, and prediction.\nDespite recent successes, however, designing high-performing deep learning\nmodels for nonimage and nontext cancer data is a time-consuming,\ntrial-and-error, manual task that requires both cancer domain and deep learning\nexpertise. To that end, we develop a reinforcement-learning-based neural\narchitecture search to automate deep-learning-based predictive model\ndevelopment for a class of representative cancer data. We develop custom\nbuilding blocks that allow domain experts to incorporate the\ncancer-data-specific characteristics. We show that our approach discovers deep\nneural network architectures that have significantly fewer trainable\nparameters, shorter training time, and accuracy similar to or higher than those\nof manually designed architectures. We study and demonstrate the scalability of\nour approach on up to 1,024 Intel Knights Landing nodes of the Theta\nsupercomputer at the Argonne Leadership Computing Facility.', 'To acquire a new skill, humans learn better and faster if a tutor, based on\ntheir current knowledge level, informs them of how much attention they should\npay to particular content or practice problems. Similarly, a machine learning\nmodel could potentially be trained better with a scorer that ""adapts"" to its\ncurrent learning state and estimates the importance of each training data\ninstance. Training such an adaptive scorer efficiently is a challenging\nproblem; in order to precisely quantify the effect of a data instance at a\ngiven time during the training, it is typically necessary to first complete the\nentire training process. To efficiently optimize data usage, we propose a\nreinforcement learning approach called Differentiable Data Selection (DDS). In\nDDS, we formulate a scorer network as a learnable function of the training\ndata, which can be efficiently updated along with the main model being trained.\nSpecifically, DDS updates the scorer with an intuitive reward signal: it should\nup-weigh the data that has a similar gradient with a dev set upon which we\nwould finally like to perform well. Without significant computing overhead, DDS\ndelivers strong and consistent improvements over several strong baselines on\ntwo very different tasks of machine translation and image classification.', 'We consider the problem of anomaly detection with a small set of partially\nlabeled anomaly examples and a large-scale unlabeled dataset. This is a common\nscenario in many important applications. Existing related methods either\nexclusively fit the limited anomaly examples that typically do not span the\nentire set of anomalies, or proceed with unsupervised learning from the\nunlabeled data. We propose here instead a deep reinforcement learning-based\napproach that enables an end-to-end optimization of the detection of both\nlabeled and unlabeled anomalies. This approach learns the known abnormality by\nautomatically interacting with an anomaly-biased simulation environment, while\ncontinuously extending the learned abnormality to novel classes of anomaly\n(i.e., unknown anomalies) by actively exploring possible anomalies in the\nunlabeled data. This is achieved by jointly optimizing the exploitation of the\nsmall labeled anomaly data and the exploration of the rare unlabeled anomalies.\nExtensive experiments on 48 real-world datasets show that our model\nsignificantly outperforms five state-of-the-art competing methods.', ""Detection of Alzheimer's Disease (AD) from neuroimaging data such as MRI\nthrough machine learning have been a subject of intense research in recent\nyears. Recent success of deep learning in computer vision have progressed such\nresearch further. However, common limitations with such algorithms are reliance\non a large number of training images, and requirement of careful optimization\nof the architecture of deep networks. In this paper, we attempt solving these\nissues with transfer learning, where state-of-the-art architectures such as VGG\nand Inception are initialized with pre-trained weights from large benchmark\ndatasets consisting of natural images, and the fully-connected layer is\nre-trained with only a small number of MRI images. We employ image entropy to\nselect the most informative slices for training. Through experimentation on the\nOASIS MRI dataset, we show that with training size almost 10 times smaller than\nthe state-of-the-art, we reach comparable or even better performance than\ncurrent deep-learning based methods."", 'The application of Deep Learning (DL) for medical diagnosis is often hampered\nby two problems. First, the amount of training data may be scarce, as it is\nlimited by the number of patients who have acquired the condition to be\ndiagnosed. Second, the training data may be corrupted by various types of\nnoise. Here, we study the problem of brain tumor detection from magnetic\nresonance spectroscopy (MRS) data, where both types of problems are prominent.\nTo overcome these challenges, we propose a new method for training a deep\nneural network that distills particularly representative training examples and\naugments the training data by mixing these samples from one class with those\nfrom the same and other classes to create additional training samples. We\ndemonstrate that this technique substantially improves performance, allowing\nour method to reach human-expert-level accuracy with just a few thousand\ntraining examples. Interestingly, the network learns to rely on features of the\ndata that are usually ignored by human experts, suggesting new directions for\nfuture research.', 'Reinforcement learning methods have recently been very successful at\nperforming complex sequential tasks like playing Atari games, Go and Poker.\nThese algorithms have outperformed humans in several tasks by learning from\nscratch, using only scalar rewards obtained through interaction with their\nenvironment. While there certainly has been considerable independent innovation\nto produce such results, many core ideas in reinforcement learning are inspired\nby phenomena in animal learning, psychology and neuroscience. In this paper, we\ncomprehensively review a large number of findings in both neuroscience and\npsychology that evidence reinforcement learning as a promising candidate for\nmodeling learning and decision making in the brain. In doing so, we construct a\nmapping between various classes of modern RL algorithms and specific findings\nin both neurophysiological and behavioral literature. We then discuss the\nimplications of this observed relationship between RL, neuroscience and\npsychology and its role in advancing research in both AI and brain science.', 'Neurodegenerative diseases are frequently associated with structural changes\nin the brain. Magnetic Resonance Imaging (MRI) scans can show these variations\nand therefore be used as a supportive feature for a number of neurodegenerative\ndiseases. The hippocampus has been known to be a biomarker for Alzheimer\ndisease and other neurological and psychiatric diseases. However, it requires\naccurate, robust and reproducible delineation of hippocampal structures. Fully\nautomatic methods are usually the voxel based approach, for each voxel a number\nof local features were calculated. In this paper we compared four different\ntechniques for feature selection from a set of 315 features extracted for each\nvoxel: (i) filter method based on the Kolmogorov-Smirnov test; two wrapper\nmethods, respectively, (ii) Sequential Forward Selection and (iii) Sequential\nBackward Elimination; and (iv) embedded method based on the Random Forest\nClassifier on a set of 10 T1-weighted brain MRIs and tested on an independent\nset of 25 subjects. The resulting segmentations were compared with manual\nreference labelling. By using only 23 features for each voxel (sequential\nbackward elimination) we obtained comparable state of-the-art performances with\nrespect to the standard tool FreeSurfer.', 'Reinforcement learning has recently shown promise as a technique for training\nan artificial neural network to parse sentences in some unknown format, through\na body of work known as RL-GRIT. A key aspect of the RL-GRIT approach is that\nrather than explicitly inferring a grammar that describes the format, the\nneural network learns to perform various parsing actions (such as merging two\ntokens) over a corpus of sentences, with the goal of maximizing the estimated\nfrequency of the resulting parse structures. This can allow the learning\nprocess to more easily explore different action choices, since a given choice\nmay change the optimality of the parse (as expressed by the total reward), but\nwill not result in the failure to parse a sentence. However, this also presents\na limitation: because the trained neural network can successfully parse any\nsentence, it cannot be directly used to identify sentences that deviate from\nthe format of the training sentences, i.e., that are anomalous. In this paper,\nwe address this limitation by presenting procedures for extracting production\nrules from the neural network, and for using these rules to determine whether a\ngiven sentence is nominal or anomalous. When a sentence is anomalous, an\nattempt is made to identify the location of the anomaly. We empirically\ndemonstrate that our approach is capable of grammatical inference and anomaly\ndetection for both non-regular formats and those containing regions of high\nrandomness/entropy. While a format with high randomness typically requires\nlarge sets of production rules, we propose a two pass grammatical inference\nmethod to generate parsimonious rule sets for such formats. By further\nimproving parser learning, and leveraging the presented rule extraction and\nanomaly detection algorithms, one might begin to understand common errors,\neither benign or malicious, in practical formats.', 'We propose a new algorithm for hyperparameter selection in machine learning\nalgorithms. The algorithm is a novel modification of Harmonica, a spectral\nhyperparameter selection approach using sparse recovery methods. In particular,\nwe show that a special encoding of hyperparameter space enables a natural\ngroup-sparse recovery formulation, which when coupled with HyperBand (a\nmulti-armed bandit strategy) leads to improvement over existing hyperparameter\noptimization methods such as Successive Halving and Random Search. Experimental\nresults on image datasets such as CIFAR-10 confirm the benefits of our\napproach.', 'With a mortality rate of 5.4 million lives worldwide every year and a\nhealthcare cost of more than 16 billion dollars in the USA alone, sepsis is one\nof the leading causes of hospital mortality and an increasing concern in the\nageing western world. Recently, medical and technological advances have helped\nre-define the illness criteria of this disease, which is otherwise poorly\nunderstood by the medical society. Together with the rise of widely accessible\nElectronic Health Records, the advances in data mining and complex nonlinear\nalgorithms are a promising avenue for the early detection of sepsis. This work\ncontributes to the research effort in the field of automated sepsis detection\nwith an open-access labelling of the medical MIMIC-III data set. Moreover, we\npropose MGP-AttTCN: a joint multitask Gaussian Process and attention-based deep\nlearning model to early predict the occurrence of sepsis in an interpretable\nmanner. We show that our model outperforms the current state-of-the-art and\npresent evidence that different labelling heuristics lead to discrepancies in\ntask difficulty. For instance, when predicting sepsis five hours prior to onset\non our new realistic labels, our proposed model achieves an area under the ROC\ncurve of 0.660 and an area under the PR curve of 0.483, whereas the (less\ninterpretable) previous state-of-the-art model (MGP-TCN) achieves 0.635 AUROC\nand 0.460 AUPR and the popular commercial InSight model achieves 0.490 AUROC\nand 0.359 AUPR.', 'The timeliness of detection of a sepsis event in progress is a crucial factor\nin the outcome for the patient. Machine learning models built from data in\nelectronic health records can be used as an effective tool for improving this\ntimeliness, but so far the potential for clinical implementations has been\nlargely limited to studies in intensive care units. This study will employ a\nricher data set that will expand the applicability of these models beyond\nintensive care units. Furthermore, we will circumvent several important\nlimitations that have been found in the literature: 1) Models are evaluated\nshortly before sepsis onset without considering interventions already\ninitiated. 2) Machine learning models are built on a restricted set of clinical\nparameters, which are not necessarily measured in all departments. 3) Model\nperformance is limited by current knowledge of sepsis, as feature interactions\nand time dependencies are hardcoded into the model. In this study, we present a\nmodel to overcome these shortcomings using a deep learning approach on a\ndiverse multicenter data set. We used retrospective data from multiple Danish\nhospitals over a seven-year period. Our sepsis detection system is constructed\nas a combination of a convolutional neural network and a long short-term memory\nnetwork. We suggest a retrospective assessment of interventions by looking at\nintravenous antibiotics and blood cultures preceding the prediction time.\nResults show performance ranging from AUROC 0.856 (3 hours before sepsis onset)\nto AUROC 0.756 (24 hours before sepsis onset). We present a deep learning\nsystem for early detection of sepsis that is able to learn characteristics of\nthe key factors and interactions from the raw event sequence data itself,\nwithout relying on a labor-intensive feature extraction work.', 'Despite significant advances in modeling of friction-induced vibrations and\nbrake squeal, the majority of industrial research and design is still conducted\nexperimentally, since many aspects of squeal and its mechanisms involved remain\nunknown. We report here for the first time on novel strategies for handling\ndata-intensive vibration testings to gain better insights into friction brake\nsystem vibrations and noise generation mechanisms. Machine learning-based\nmethods to detect and characterize vibrations, to understand sensitivities and\nto predict brake squeal are applied with the aim to illustrate how\ninterdisciplinary approaches can leverage the potential of data science\ntechniques for classical mechanical engineering challenges. In the first part,\na deep learning brake squeal detector is developed to identify several classes\nof typical friction noise recordings. The detection method is rooted in recent\ncomputer vision techniques for object detection based on convolutional neural\nnetworks. It allows to overcome limitations of classical approaches that solely\nrely on instantaneous spectral properties of the recorded noise. Results\nindicate superior detection and characterization quality when compared to a\nstate-of-the-art brake squeal detector. In the second part, a recurrent neural\nnetwork is employed to learn the parametric patterns that determine the dynamic\nstability of an operating brake system. Given a set of multivariate loading\nconditions, the RNN learns to predict the noise generation of the structure.\nThe validated RNN represents a virtual twin model for the squeal behavior of a\nspecific brake system. It is found that this model can predict the occurrence\nand the onset of brake squeal with high accuracy and that it can identify the\ncomplicated patterns and temporal dependencies in the loading conditions that\ndrive the dynamical structure into regimes of instability.', 'We present a framework capable of tackilng the problem of continual object\nrecognition in a setting which resembles that under whichhumans see and learn.\nThis setting has a set of unique characteristics:it assumes an egocentric\npoint-of-view bound to the needs of a singleperson, which implies a relatively\nlow diversity of data and a coldstart with no data; it requires to operate in\nan open world, where newobjects can be encounteredat any time; supervision is\nscarce and hasto be solicited to the user, and completelyunsupervised\nrecognitionof new objects should be possible. Note that this setting differs\nfromthe one addressed in the open world recognition literature, where\nsupervised feedback is always requested to be able to incorporate newobjects.\nWe propose a first solution to this problem in the form ofa memory-based\nincremental framework that is capable of storinginformation of each and any\nobject it encounters, while using the supervision of the user to learn to\ndiscriminate between known and unknown objects. Our approach is based on four\nmain features: the useof time and space persistence (i.e., the appearance of\nobjects changesrelatively slowly), the use of similarity as the main driving\nprinciplefor object recognition and novelty detection, the progressive\nintroduction of new objects in a developmental fashion and the\nselectiveelicitation of user feedback in an online active learning fashion.\nExperimental results show the feasibility of open world, generic\nobjectrecognition, the ability to recognize, memorize and re-identify\nnewobjects even in complete absence of user supervision, and the utilityof\npersistence and incrementality in boosting performance.', 'Few-Shot Learning is the challenge of training a model with only a small\namount of data. Many solutions to this problem use meta-learning algorithms,\ni.e. algorithms that learn to learn. By sampling few-shot tasks from a larger\ndataset, we can teach these algorithms to solve new, unseen tasks. This\ndocument reports my work on meta-learning algorithms for Few-Shot Computer\nVision. This work was done during my internship at Sicara, a French company\nbuilding image recognition solutions for businesses. It contains: 1. an\nextensive review of the state-of-the-art in few-shot computer vision; 2. a\nbenchmark of meta-learning algorithms for few-shot image classification; 3. the\nintroduction to a novel meta-learning algorithm for few-shot object detection,\nwhich is still in development.', 'We present Earliness-Aware Deep Convolutional Networks (EA-ConvNets), an\nend-to-end deep learning framework, for early classification of time series\ndata. Unlike most existing methods for early classification of time series\ndata, that are designed to solve this problem under the assumption of the\navailability of a good set of pre-defined (often hand-crafted) features, our\nframework can jointly perform feature learning (by learning a deep hierarchy of\n\\emph{shapelets} capturing the salient characteristics in each time series),\nalong with a dynamic truncation model to help our deep feature learning\narchitecture focus on the early parts of each time series. Consequently, our\nframework is able to make highly reliable early predictions, outperforming\nvarious state-of-the-art methods for early time series classification, while\nalso being competitive when compared to the state-of-the-art time series\nclassification algorithms that work with \\emph{fully observed} time series\ndata. To the best of our knowledge, the proposed framework is the first to\nperform data-driven (deep) feature learning in the context of early\nclassification of time series data. We perform a comprehensive set of\nexperiments, on several benchmark data sets, which demonstrate that our method\nyields significantly better predictions than various state-of-the-art methods\ndesigned for early time series classification. In addition to obtaining high\naccuracies, our experiments also show that the learned deep shapelets based\nfeatures are also highly interpretable and can help gain better understanding\nof the underlying characteristics of time series data.', 'Many practical applications of reinforcement learning require agents to learn\nfrom sparse and delayed rewards. It challenges the ability of agents to\nattribute their actions to future outcomes. In this paper, we consider the\nproblem formulation of episodic reinforcement learning with trajectory\nfeedback. It refers to an extreme delay of reward signals, in which the agent\ncan only obtain one reward signal at the end of each trajectory. A popular\nparadigm for this problem setting is learning with a designed auxiliary dense\nreward function, namely proxy reward, instead of sparse environmental signals.\nBased on this framework, this paper proposes a novel reward redistribution\nalgorithm, randomized return decomposition (RRD), to learn a proxy reward\nfunction for episodic reinforcement learning. We establish a surrogate problem\nby Monte-Carlo sampling that scales up least-squares-based reward\nredistribution to long-horizon problems. We analyze our surrogate loss function\nby connection with existing methods in the literature, which illustrates the\nalgorithmic properties of our approach. In experiments, we extensively evaluate\nour proposed method on a variety of benchmark tasks with episodic rewards and\ndemonstrate substantial improvement over baseline algorithms.', 'Fake tweets are observed to be ever-increasing, demanding immediate\ncountermeasures to combat their spread. During COVID-19, tweets with\nmisinformation should be flagged and neutralized in their early stages to\nmitigate the damages. Most of the existing methods for early detection of fake\nnews assume to have enough propagation information for large labeled tweets --\nwhich may not be an ideal setting for cases like COVID-19 where both aspects\nare largely absent. In this work, we present ENDEMIC, a novel early detection\nmodel which leverages exogenous and endogenous signals related to tweets, while\nlearning on limited labeled data. We first develop a novel dataset, called CTF\nfor early COVID-19 Twitter fake news, with additional behavioral test sets to\nvalidate early detection. We build a heterogeneous graph with\nfollower-followee, user-tweet, and tweet-retweet connections and train a graph\nembedding model to aggregate propagation information. Graph embeddings and\ncontextual features constitute endogenous, while time-relative web-scraped\ninformation constitutes exogenous signals. ENDEMIC is trained in a\nsemi-supervised fashion, overcoming the challenge of limited labeled data. We\npropose a co-attention mechanism to fuse signal representations optimally.\nExperimental results on ECTF, PolitiFact, and GossipCop show that ENDEMIC is\nhighly reliable in detecting early fake tweets, outperforming nine\nstate-of-the-art methods significantly.', ""Early diagnosis of treatable diseases is essential for improving healthcare,\nand many diseases' onsets are predictable from annual lab tests and their\ntemporal trends. We introduce a multi-resolution convolutional neural network\nfor early detection of multiple diseases from irregularly measured sparse lab\nvalues. Our novel architecture takes as input both an imputed version of the\ndata and a binary observation matrix. For imputing the temporal sparse\nobservations, we develop a flexible, fast to train method for differentiable\nmultivariate kernel regression. Our experiments on data from 298K individuals\nover 8 years, 18 common lab measurements, and 171 diseases show that the\ntemporal signatures learned via convolution are significantly more predictive\nthan baselines commonly used for early disease diagnosis."", 'In statistical modelling the biggest threat is concept drift which makes the\nmodel gradually showing deteriorating performance over time. There are state of\nthe art methodologies to detect the impact of concept drift, however general\nstrategy considered to overcome the issue in performance is to rebuild or\nre-calibrate the model periodically as the variable patterns for the model\nchanges significantly due to market change or consumer behavior change etc.\nQuantitative research is the most widely spread application of data science in\nMarketing or financial domain where applicability of state of the art\nreinforcement learning for auto-learning is less explored paradigm.\nReinforcement learning is heavily dependent on having a simulated environment\nwhich is majorly available for gaming or online systems, to learn from the live\nfeedback. However, there are some research happened on the area of online\nadvertisement, pricing etc where due to the nature of the online learning\nenvironment scope of reinforcement learning is explored. Our proposed solution\nis a reinforcement learning based, true self-learning algorithm which can adapt\nto the data change or concept drift and auto learn and self-calibrate for the\nnew patterns of the data solving the problem of concept drift.\n  Keywords - Reinforcement learning, Genetic Algorithm, Q-learning,\nClassification modelling, CMA-ES, NES, Multi objective optimization, Concept\ndrift, Population stability index, Incremental learning, F1-measure, Predictive\nModelling, Self-learning, MCTS, AlphaGo, AlphaZero', 'The recent explosion of machine learning (ML) and artificial intelligence\n(AI) shows great potential in the breakthrough of metal additive manufacturing\n(AM) process modeling. However, the success of conventional machine learning\ntools in data science is primarily attributed to the unprecedented large amount\nof labeled data-sets (big data), which can be either obtained by experiments or\nfirst-principle simulations. Unfortunately, these labeled data-sets are\nexpensive to obtain in AM due to the high expense of the AM experiments and\nprohibitive computational cost of high-fidelity simulations.\n  We propose a physics-informed neural network (PINN) framework that fuses both\ndata and first physical principles, including conservation laws of momentum,\nmass, and energy, into the neural network to inform the learning processes. To\nthe best knowledge of the authors, this is the first application of PINN to\nthree dimensional AM processes modeling. Besides, we propose a hard-type\napproach for Dirichlet boundary conditions (BCs) based on a Heaviside function,\nwhich can not only enforce the BCs but also accelerate the learning process.\nThe PINN framework is applied to two representative metal manufacturing\nproblems, including the 2018 NIST AM-Benchmark test series. We carefully assess\nthe performance of the PINN model by comparing the predictions with available\nexperimental data and high-fidelity simulation results. The investigations show\nthat the PINN, owed to the additional physical knowledge, can accurately\npredict the temperature and melt pool dynamics during metal AM processes with\nonly a moderate amount of labeled data-sets. The foray of PINN to metal AM\nshows the great potential of physics-informed deep learning for broader\napplications to advanced manufacturing.', 'Early outbreak detection is a key aspect in the containment of infectious\ndiseases, as it enables the identification and isolation of infected\nindividuals before the disease can spread to a larger population. Instead of\ndetecting unexpected increases of infections by monitoring confirmed cases,\nsyndromic surveillance aims at the detection of cases with early symptoms,\nwhich allows a more timely disclosure of outbreaks. However, the definition of\nthese disease patterns is often challenging, as early symptoms are usually\nshared among many diseases and a particular disease can have several clinical\npictures in the early phase of an infection. To support epidemiologists in the\nprocess of defining reliable disease patterns, we present a novel, data-driven\napproach to discover such patterns in historic data. The key idea is to take\ninto account the correlation between indicators in a health-related data source\nand the reported number of infections in the respective geographic region. In\nan experimental evaluation, we use data from several emergency departments to\ndiscover disease patterns for three infectious diseases. Our results suggest\nthat the proposed approach is able to find patterns that correlate with the\nreported infections and often identifies indicators that are related to the\nrespective diseases.', 'Diseases and pests cause huge economic loss to the apple industry every year.\nThe identification of various apple diseases is challenging for the farmers as\nthe symptoms produced by different diseases may be very similar, and may be\npresent simultaneously. This paper is an attempt to provide the timely and\naccurate detection and identification of apple diseases. In this study, we\npropose a deep learning based approach for identification and classification of\napple diseases. The first part of the study is dataset creation which includes\ndata collection and data labelling. Next, we train a Convolutional Neural\nNetwork (CNN) model on the prepared dataset for automatic classification of\napple diseases. CNNs are end-to-end learning algorithms which perform automatic\nfeature extraction and learn complex features directly from raw images, making\nthem suitable for wide variety of tasks like image classification, object\ndetection, segmentation etc. We applied transfer learning to initialize the\nparameters of the proposed deep model. Data augmentation techniques like\nrotation, translation, reflection and scaling were also applied to prevent\noverfitting. The proposed CNN model obtained encouraging results, reaching\naround 97.18% of accuracy on our prepared dataset. The results validate that\nthe proposed method is effective in classifying various types of apple diseases\nand can be used as a practical tool by farmers.', 'In this paper we propose a novel framework for decentralized, online learning\nby many learners. At each moment of time, an instance characterized by a\ncertain context may arrive to each learner; based on the context, the learner\ncan select one of its own actions (which gives a reward and provides\ninformation) or request assistance from another learner. In the latter case,\nthe requester pays a cost and receives the reward but the provider learns the\ninformation. In our framework, learners are modeled as cooperative contextual\nbandits. Each learner seeks to maximize the expected reward from its arrivals,\nwhich involves trading off the reward received from its own actions, the\ninformation learned from its own actions, the reward received from the actions\nrequested of others and the cost paid for these actions - taking into account\nwhat it has learned about the value of assistance from each other learner. We\ndevelop distributed online learning algorithms and provide analytic bounds to\ncompare the efficiency of these with algorithms with the complete knowledge\n(oracle) benchmark (in which the expected reward of every action in every\ncontext is known by every learner). Our estimates show that regret - the loss\nincurred by the algorithm - is sublinear in time. Our theoretical framework can\nbe used in many practical applications including Big Data mining, event\ndetection in surveillance sensor networks and distributed online recommendation\nsystems.', 'What goals should a multi-goal reinforcement learning agent pursue during\ntraining in long-horizon tasks? When the desired (test time) goal distribution\nis too distant to offer a useful learning signal, we argue that the agent\nshould not pursue unobtainable goals. Instead, it should set its own intrinsic\ngoals that maximize the entropy of the historical achieved goal distribution.\nWe propose to optimize this objective by having the agent pursue past achieved\ngoals in sparsely explored areas of the goal space, which focuses exploration\non the frontier of the achievable goal set. We show that our strategy achieves\nan order of magnitude better sample efficiency than the prior state of the art\non long-horizon multi-goal tasks including maze navigation and block stacking.', 'A key challenge in scaling up robot learning to many skills and environments\nis removing the need for human supervision, so that robots can collect their\nown data and improve their own performance without being limited by the cost of\nrequesting human feedback. Model-based reinforcement learning holds the promise\nof enabling an agent to learn to predict the effects of its actions, which\ncould provide flexible predictive models for a wide range of tasks and\nenvironments, without detailed human supervision. We develop a method for\ncombining deep action-conditioned video prediction models with model-predictive\ncontrol that uses entirely unlabeled training data. Our approach does not\nrequire a calibrated camera, an instrumented training set-up, nor precise\nsensing and actuation. Our results show that our method enables a real robot to\nperform nonprehensile manipulation -- pushing objects -- and can handle novel\nobjects not seen during training.', ""Parkinson's disease (PD) is debilitating, progressive, and clinically marked\nby motor symptoms. As the second most common neurodegenerative disease in the\nworld, it affects over 10 million lives globally. Existing diagnoses methods\nhave limitations, such as the expense of visiting doctors and the challenge of\nautomated early detection, considering that behavioral differences in patients\nand healthy individuals are often indistinguishable in the early stages.\nHowever, micrographia, a handwriting disorder that leads to abnormally small\nhandwriting, tremors, dystonia, and slow movement in the hands and fingers, is\ncommonly observed in the early stages of PD. In this work, we apply machine\nlearning techniques to extract signs of micrographia from drawing samples\ngathered from two open-source datasets and achieve a predictive accuracy of\n94%. This work also sets the foundations for a publicly available and\nuser-friendly web portal that anyone with access to a pen, printer, and phone\ncan use for early PD detection."", 'Healthcare is one of the most important aspects of human life. Heart disease\nis known to be one of the deadliest diseases which is hampering the lives of\nmany people around the world. Heart disease must be detected early so the loss\nof lives can be prevented. The availability of large-scale data for medical\ndiagnosis has helped developed complex machine learning and deep learning-based\nmodels for automated early diagnosis of heart diseases. The classical\napproaches have been limited in terms of not generalizing well to new data\nwhich have not been seen in the training set. This is indicated by a large gap\nin training and test accuracies. This paper proposes a novel deep learning\narchitecture using a 1D convolutional neural network for classification between\nhealthy and non-healthy persons to overcome the limitations of classical\napproaches. Various clinical parameters are used for assessing the risk profile\nin the patients which helps in early diagnosis. Various techniques are used to\navoid overfitting in the proposed network. The proposed network achieves over\n97% training accuracy and 96% test accuracy on the dataset. The accuracy of the\nmodel is compared in detail with other classification algorithms using various\nperformance parameters which proves the effectiveness of the proposed\narchitecture.', ""The proliferation of web platforms has created incentives for online abuse.\nMany graph-based anomaly detection techniques are proposed to identify the\nsuspicious accounts and behaviors. However, most of them detect the anomalies\nonce the users have performed many such behaviors. Their performance is\nsubstantially hindered when the users' observed data is limited at an early\nstage, which needs to be improved to minimize financial loss. In this work, we\npropose Eland, a novel framework that uses action sequence augmentation for\nearly anomaly detection. Eland utilizes a sequence predictor to predict next\nactions of every user and exploits the mutual enhancement between action\nsequence augmentation and user-action graph anomaly detection. Experiments on\nthree real-world datasets show that Eland improves the performance of a variety\nof graph-based anomaly detection methods. With Eland, anomaly detection\nperformance at an earlier stage is better than non-augmented methods that need\nsignificantly more observed data by up to 15% on the Area under the ROC curve."", ""Amyotrophic lateral sclerosis (ALS) is a progressive neurodegenerative\ndisease of the brain and the spinal cord, which leads to paralysis of motor\nfunctions. Patients retain their ability to blink, which can be used for\ncommunication. Here, We present an Artificial Intelligence (AI) system that\nuses eye-blinks to communicate with the outside world, running on real-time\nInternet-of-Things (IoT) devices. The system uses a Convolutional Neural\nNetwork (CNN) to find the blinking pattern, which is defined as a series of\nOpen and Closed states. Each pattern is mapped to a collection of words that\nmanifest the patient's intent. To investigate the best trade-off between\naccuracy and latency, we investigated several Convolutional Network\narchitectures, such as ResNet, SqueezeNet, DenseNet, and InceptionV3, and\nevaluated their performance. We found that the InceptionV3 architecture, after\nhyper-parameter fine-tuning on the specific task led to the best performance\nwith an accuracy of 99.20% and 94ms latency. This work demonstrates how the\nlatest advances in deep learning architectures can be adapted for clinical\nsystems that ameliorate the patient's quality of life regardless of the\npoint-of-care."", ""In order to detect unknown intrusions and runtime errors of computer\nprograms, the cyber-security community has developed various detection\ntechniques. Anomaly detection is an approach that is designed to profile the\nnormal runtime behavior of computer programs in order to detect intrusions and\nerrors as anomalous deviations from the observed normal. However, normal but\nunobserved behavior can trigger false positives. This limitation has\nsignificantly decreased the practical viability of anomaly detection\ntechniques. Reported approaches to this limitation span a simple alert\nthreshold definition to distribution models for approximating all normal\nbehavior based on the limited observation. However, each assumption or\napproximation poses the potential for even greater false positive rates. This\npaper presents our study on how to explain the presence of anomalies using a\nneural network, particularly Long Short-Term Memory, independent of actual data\ndistributions. We present and compare three anomaly detection models, and\nreport on our experience running different types of attacks on an Apache\nHypertext Transfer Protocol server. We performed a comparative study, focusing\non each model's ability to detect the onset of each attack while avoiding false\npositives resulting from unknown normal behavior. Our best-performing model\ndetected the true onset of every attack with zero false positives."", 'In medical science, the use of computer science in disease detection and\ndiagnosis is gaining popularity. Previously, the detection of disease used to\ntake a significant amount of time and was less reliable. Machine learning (ML)\ntechniques employed in recent biomedical researches are making revolutionary\nchanges by gaining higher accuracy with more concise timing. At present, it is\neven possible to automatically detect diseases from the scanned images with the\nhelp of ML. In this research, we have taken such an attempt to detect retinal\ndiseases from optical coherence tomography (OCT) X-ray images. Here, we propose\na deep learning (DL) based approach in detecting retinal diseases from OCT\nimages which can identify three conditions of the retina. Four different models\nused in this approach are compared with each other. On the test set, the\ndetection accuracy is 98.00\\% for a vanilla convolutional neural network (CNN)\nmodel, 99.07\\% for Xception model, 97.00\\% for ResNet50 model, and 99.17\\% for\nMobileNetV2 model. The MobileNetV2 model acquires the highest accuracy, and the\nclosest to the highest is the Xception model. The proposed approach has a\npotential impact on creating a tool for automatically detecting retinal\ndiseases.', 'Fever can provide valuable information for diagnosis and prognosis of various\ndiseases such as pneumonia, dengue, sepsis, etc., therefore, predicting fever\nearly can help in the effectiveness of treatment options and expediting the\ntreatment process. This study aims to develop novel algorithms that can\naccurately predict fever onset in critically ill patients by applying machine\nlearning technique on continuous physiological data. We analyzed continuous\nphysiological data collected every 5-minute from a cohort of over 200,000\ncritically ill patients admitted to an Intensive Care Unit (ICU) over a 2-year\nperiod. Each episode of fever from the same patient were considered as an\nindependent event, with separations of at least 24 hours. We extracted\ndescriptive statistical features from six physiological data streams, including\nheart rate, respiration, systolic and diastolic blood pressure, mean arterial\npressure, and oxygen saturation, and use these features to independently\npredict the onset of fever. Using a bootstrap aggregation method, we created a\nbalanced dataset of 7,801 afebrile and febrile patients and analyzed features\nup to 4 hours before the fever onset. We found that supervised machine learning\nmethods can predict fever up to 4 hours before onset in critically ill patients\nwith high recall, precision, and F1-score. This study demonstrates the\nviability of using machine learning to predict fever among hospitalized adults.\nThe discovery of salient physiomarkers through machine learning and deep\nlearning techniques has the potential to further accelerate the development and\nimplementation of innovative care delivery protocols and strategies for\nmedically vulnerable patients.', 'The recent history of machine learning research has taught us that machine\nlearning methods can be most effective when they are provided with very large,\nhigh-capacity models, and trained on very large and diverse datasets. This has\nspurred the community to search for ways to remove any bottlenecks to scale.\nOften the foremost among such bottlenecks is the need for human effort,\nincluding the effort of curating and labeling datasets. As a result,\nconsiderable attention in recent years has been devoted to utilizing unlabeled\ndata, which can be collected in vast quantities. However, some of the most\nwidely used methods for training on such unlabeled data themselves require\nhuman-designed objective functions that must correlate in some meaningful way\nto downstream tasks. I will argue that a general, principled, and powerful\nframework for utilizing unlabeled data can be derived from reinforcement\nlearning, using general purpose unsupervised or self-supervised reinforcement\nlearning objectives in concert with offline reinforcement learning methods that\ncan leverage large datasets. I will discuss how such a procedure is more\nclosely aligned with potential downstream tasks, and how it could build on\nexisting techniques that have been developed in recent years.', ""Disparate areas of machine learning have benefited from models that can take\nraw data with little preprocessing as input and learn rich representations of\nthat raw data in order to perform well on a given prediction task. We evaluate\nthis approach in healthcare by using longitudinal measurements of lab tests,\none of the more raw signals of a patient's health state widely available in\nclinical data, to predict disease onsets. In particular, we train a Long\nShort-Term Memory (LSTM) recurrent neural network and two novel convolutional\nneural networks for multi-task prediction of disease onset for 133 conditions\nbased on 18 common lab tests measured over time in a cohort of 298K patients\nderived from 8 years of administrative claims data. We compare the neural\nnetworks to a logistic regression with several hand-engineered, clinically\nrelevant features. We find that the representation-based learning approaches\nsignificantly outperform this baseline. We believe that our work suggests a new\navenue for patient risk stratification based solely on lab results."", ""Machine Learning requires large amounts of labeled data to fit a model. Many\ndatasets are already publicly available, nevertheless forcing application\npossibilities of machine learning to the domains of those public datasets. The\never-growing penetration of machine learning algorithms in new application\nareas requires solutions for the need for data in those new domains. This\nthesis works on active learning as one possible solution to reduce the amount\nof data that needs to be processed by hand, by processing only those datapoints\nthat specifically benefit the training of a strong model for the task. A newly\nproposed framework for framing the active learning workflow as a reinforcement\nlearning problem is adapted for image classification and a series of three\nexperiments is conducted. Each experiment is evaluated and potential issues\nwith the approach are outlined. Each following experiment then proposes\nimprovements to the framework and evaluates their impact. After the last\nexperiment, a final conclusion is drawn, unfortunately rejecting this work's\nhypothesis and outlining that the proposed framework at the moment is not\ncapable of improving active learning for image classification with a trained\nreinforcement learning agent."", 'Chest X-ray scan is a most often used modality by radiologists to diagnose\nmany chest related diseases in their initial stages. The proposed system aids\nthe radiologists in making decision about the diseases found in the scans more\nefficiently. Our system combines the techniques of image processing for feature\nenhancement and deep learning for classification among diseases. We have used\nthe ChestX-ray14 database in order to train our deep learning model on the 14\ndifferent labeled diseases found in it. The proposed research shows the\nsignificant improvement in the results by using wavelet transforms as\npre-processing technique.', 'Several neurodegenerative diseases involve the accumulation of cellular DNA\ndamage. Comet assays are a popular way of estimating the extent of DNA damage.\nCurrent literature on the use of deep learning to quantify DNA damage presents\nan empirical approach to hyper-parameter optimization and does not include\nuncertainty estimates. Deep ensemble averaging is a standard approach to\nestimating uncertainty but it requires several iterations of network training,\nwhich makes it time-consuming. Here we present an approach to quantify the\nextent of DNA damage that combines deep learning with a rigorous and\ncomprehensive method to optimize the hyper-parameters with the help of\nstatistical tests. We also use an architecture that allows for a faster\ncomputation of deep ensemble averaging and performs statistical tests\napplicable to networks using transfer learning. We applied our approach to a\ncomet assay dataset with more than 1300 images and achieved an $R^2$ of 0.84,\nwhere the output included the confidence interval for each prediction. The\nproposed architecture is an improvement over the current approaches since it\nspeeds up the uncertainty estimation by 30X while being statistically more\nrigorous.', 'Crypto-ransomware is characterized by its irreversible effect even after the\ndetection and removal. As such, the early detection is crucial to protect user\ndata and files of being held to ransom. Several solutions have proposed\nutilizing the data extracted during the initial phases of the attacks before\nthe encryption takes place. However, the lack of enough data at the early\nphases of the attack along with high dimensional features space renders the\nmodel prone to overfitting which decreases its detection accuracy. To this end,\nthis paper proposed a novel redundancy coefficient gradual up-weighting\napproach that was incorporated to the calculation of redundancy term of mutual\ninformation to improve the feature selection process and enhance the accuracy\nof the detection model. Several machine learning classifiers were used to\nevaluate the detection performance of the proposed techniques. The experimental\nresults show that the accuracy of proposed techniques achieved higher detection\naccuracy. Those results demonstrate the efficacy of the proposed techniques for\nthe early detection tasks.', 'Making disguise between real and fake news propagation through online social\nnetworks is an important issue in many applications. The time gap between the\nnews release time and detection of its label is a significant step towards\nbroadcasting the real information and avoiding the fake. Therefore, one of the\nchallenging tasks in this area is to identify fake and real news in early\nstages of propagation. However, there is a trade-off between minimizing the\ntime gap and maximizing accuracy. Despite recent efforts in detection of fake\nnews, there has been no significant work that explicitly incorporates early\ndetection in its model. In this paper, we focus on accurate early labeling of\nnews, and propose a model by considering earliness both in modeling and\nprediction. The proposed method utilizes recurrent neural networks with a novel\nloss function, and a new stopping rule. Given the context of news, we first\nembed it with a class-specific text representation. Then, we utilize the\navailable public profile of users, and speed of news diffusion, for early\nlabeling of the news. Experiments on real datasets demonstrate the\neffectiveness of our model both in terms of early labelling and accuracy,\ncompared to the state of the art baseline and models.', 'Consider a 2-player normal-form game repeated over time. We introduce an\nadaptive learning procedure, where the players only observe their own realized\npayoff at each stage. We assume that agents do not know their own payoff\nfunction, and have no information on the other player. Furthermore, we assume\nthat they have restrictions on their own action set such that, at each stage,\ntheir choice is limited to a subset of their action set. We prove that the\nempirical distributions of play converge to the set of Nash equilibria for\nzero-sum and potential games, and games where one player has two actions.', ""How will my face look when I get older? Or, for a more challenging question:\nHow will my brain look when I get older? To answer this question one must\ndevise (and learn from data) a multivariate auto-regressive function which\ngiven an image and a desired target age generates an output image. While\ncollecting data for faces may be easier, collecting longitudinal brain data is\nnot trivial. We propose a deep learning-based method that learns to simulate\nsubject-specific brain ageing trajectories without relying on longitudinal\ndata. Our method synthesises images conditioned on two factors: age (a\ncontinuous variable), and status of Alzheimer's Disease (AD, an ordinal\nvariable). With an adversarial formulation we learn the joint distribution of\nbrain appearance, age and AD status, and define reconstruction losses to\naddress the challenging problem of preserving subject identity. We compare with\nseveral benchmarks using two widely used datasets. We evaluate the quality and\nrealism of synthesised images using ground-truth longitudinal data and a\npre-trained age predictor. We show that, despite the use of cross-sectional\ndata, our model learns patterns of gray matter atrophy in the middle temporal\ngyrus in patients with AD. To demonstrate generalisation ability, we train on\none dataset and evaluate predictions on the other. In conclusion, our model\nshows an ability to separate age, disease influence and anatomy using only 2D\ncross-sectional data that should be useful in large studies into\nneurodegenerative disease, that aim to combine several data sources. To\nfacilitate such future studies by the community at large our code is made\navailable at https://github.com/xiat0616/BrainAgeing."", ""Early identification and prevention of various plant diseases in commercial\nfarms and orchards is a key feature of precision agriculture technology. This\npaper presents a high-performance real-time fine-grain object detection\nframework that addresses several obstacles in plant disease detection that\nhinder the performance of traditional methods, such as, dense distribution,\nirregular morphology, multi-scale object classes, textural similarity, etc. The\nproposed model is built on an improved version of the You Only Look Once\n(YOLOv4) algorithm. The modified network architecture maximizes both detection\naccuracy and speed by including the DenseNet in the back-bone to optimize\nfeature transfer and reuse, two new residual blocks in the backbone and neck\nenhance feature extraction and reduce computing cost; the Spatial Pyramid\nPooling (SPP) enhances receptive field, and a modified Path Aggregation Network\n(PANet) preserves fine-grain localized information and improve feature fusion.\nAdditionally, the use of the Hard-Swish function as the primary activation\nimproved the model's accuracy due to better nonlinear feature extraction. The\nproposed model is tested in detecting four different diseases in tomato plants\nunder various challenging environments. The model outperforms the existing\nstate-of-the-art detection models in detection accuracy and speed. At a\ndetection rate of 70.19 FPS, the proposed model obtained a precision value of\n$90.33 \\%$, F1-score of $93.64 \\%$, and a mean average precision ($mAP$) value\nof $96.29 \\%$. Current work provides an effective and efficient method for\ndetecting different plant diseases in complex scenarios that can be extended to\ndifferent fruit and crop detection, generic disease detection, and various\nautomated agricultural detection processes."", 'The segmentation of brain tumors in multimodal MRIs is one of the most\nchallenging tasks in medical image analysis. The recent state of the art\nalgorithms solving this task is based on machine learning approaches and deep\nlearning in particular. The amount of data used for training such models and\nits variability is a keystone for building an algorithm with high\nrepresentation power. In this paper, we study the relationship between the\nperformance of the model and the amount of data employed during the training\nprocess. On the example of brain tumor segmentation challenge, we compare the\nmodel trained with labeled data provided by challenge organizers, and the same\nmodel trained in omni-supervised manner using additional unlabeled data\nannotated with the ensemble of heterogeneous models. As a result, a single\nmodel trained with additional data achieves performance close to the ensemble\nof multiple models and outperforms individual methods.', 'This paper presents a methodology for early detection of audio events from\naudio streams. Early detection is the ability to infer an ongoing event during\nits initial stage. The proposed system consists of a novel inference step\ncoupled with dual parallel tailored-loss deep neural networks (DNNs). The DNNs\nshare a similar architecture except for their loss functions, i.e. weighted\nloss and multitask loss, which are designed to efficiently cope with issues\ncommon to audio event detection. The inference step is newly introduced to make\nuse of the network outputs for recognizing ongoing events. The monotonicity of\nthe detection function is required for reliable early detection, and will also\nbe proved. Experiments on the ITC-Irst database show that the proposed system\nachieves state-of-the-art detection performance. Furthermore, even partial\nevents are sufficient to achieve good performance similar to that obtained when\nan entire event is observed, enabling early event detection.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'Transfer learning techniques are important to handle small training sets and\nto allow for quick generalization even from only a few examples. The following\npaper is the introduction as well as the literature overview part of my thesis\nrelated to the topic of transfer learning for visual recognition problems.', 'When learning from positive and unlabelled data, it is a strong assumption\nthat the positive observations are randomly sampled from the distribution of\n$X$ conditional on $Y = 1$, where X stands for the feature and Y the label.\nMost existing algorithms are optimally designed under the assumption. However,\nfor many real-world applications, the observed positive examples are dependent\non the conditional probability $P(Y = 1|X)$ and should be sampled biasedly. In\nthis paper, we assume that a positive example with a higher $P(Y = 1|X)$ is\nmore likely to be labelled and propose a probabilistic-gap based PU learning\nalgorithms. Specifically, by treating the unlabelled data as noisy negative\nexamples, we could automatically label a group positive and negative examples\nwhose labels are identical to the ones assigned by a Bayesian optimal\nclassifier with a consistency guarantee. The relabelled examples have a biased\ndomain, which is remedied by the kernel mean matching technique. The proposed\nalgorithm is model-free and thus do not have any parameters to tune.\nExperimental results demonstrate that our method works well on both generated\nand real-world datasets.', 'Nearly all existing counting methods are designed for a specific object\nclass. Our work, however, aims to create a counting model able to count any\nclass of object. To achieve this goal, we formulate counting as a matching\nproblem, enabling us to exploit the image self-similarity property that\nnaturally exists in object counting problems. We make the following three\ncontributions: first, a Generic Matching Network (GMN) architecture that can\npotentially count any object in a class-agnostic manner; second, by\nreformulating the counting problem as one of matching objects, we can take\nadvantage of the abundance of video data labeled for tracking, which contains\nnatural repetitions suitable for training a counting model. Such data enables\nus to train the GMN. Third, to customize the GMN to different user\nrequirements, an adapter module is used to specialize the model with minimal\neffort, i.e. using a few labeled examples, and adapting only a small fraction\nof the trained parameters. This is a form of few-shot learning, which is\npractical for domains where labels are limited due to requiring expert\nknowledge (e.g. microbiology). We demonstrate the flexibility of our method on\na diverse set of existing counting benchmarks: specifically cells, cars, and\nhuman crowds. The model achieves competitive performance on cell and crowd\ncounting datasets, and surpasses the state-of-the-art on the car dataset using\nonly three training images. When training on the entire dataset, the proposed\nmethod outperforms all previous methods by a large margin.', 'In realistic environments, speech is usually interfered by various noise and\nreverberation, which dramatically degrades the performance of automatic speech\nrecognition (ASR) systems. To alleviate this issue, the commonest way is to use\na well-designed speech enhancement approach as the front-end of ASR. However,\nmore complex pipelines, more computations and even higher hardware costs\n(microphone array) are additionally consumed for this kind of methods. In\naddition, speech enhancement would result in speech distortions and mismatches\nto training. In this paper, we propose an adversarial training method to\ndirectly boost noise robustness of acoustic model. Specifically, a jointly\ncompositional scheme of generative adversarial net (GAN) and neural\nnetwork-based acoustic model (AM) is used in the training phase. GAN is used to\ngenerate clean feature representations from noisy features by the guidance of a\ndiscriminator that tries to distinguish between the true clean signals and\ngenerated signals. The joint optimization of generator, discriminator and AM\nconcentrates the strengths of both GAN and AM for speech recognition.\nSystematic experiments on CHiME-4 show that the proposed method significantly\nimproves the noise robustness of AM and achieves the average relative error\nrate reduction of 23.38% and 11.54% on the development and test set,\nrespectively.', 'Data science requires time-consuming iterative manual activities. In\nparticular, activities such as data selection, preprocessing, transformation,\nand mining, highly depend on iterative trial-and-error processes that could be\nsped-up significantly by providing quick feedback on the impact of changes. The\nidea of progressive data science is to compute the results of changes in a\nprogressive manner, returning a first approximation of results quickly and\nallow iterative refinements until converging to a final result. Enabling the\nuser to interact with the intermediate results allows an early detection of\nerroneous or suboptimal choices, the guided definition of modifications to the\npipeline and their quick assessment. In this paper, we discuss the\nprogressiveness challenges arising in different steps of the data science\npipeline. We describe how changes in each step of the pipeline impact the\nsubsequent steps and outline why progressive data science will help to make the\nprocess more effective. Computing progressive approximations of outcomes\nresulting from changes creates numerous research challenges, especially if the\nchanges are made in the early steps of the pipeline. We discuss these\nchallenges and outline first steps towards progressiveness, which, we argue,\nwill ultimately help to significantly speed-up the overall data science\nprocess.', ""In this technical report I present my method for automatic synthetic dataset\ngeneration for object detection and demonstrate it on the video game League of\nLegends. This report furthermore serves as a handbook on how to automatically\ngenerate datasets and as an introduction on the dataset generation part of the\nLeagueAI framework. The LeagueAI framework is a software framework that\nprovides detailed information about the game League of Legends based on the\nsame input a human player would have, namely vision. The framework allows\nresearchers and enthusiasts to develop their own intelligent agents or to\nextract detailed information about the state of the game. A big problem of\nmachine vision applications usually is the laborious work of gathering large\namounts of hand labeled data. Thus, a crucial part of the vision pipeline of\nthe LeagueAI framework, the dataset generation, is presented in this report.\nThe method involves extracting image raw data from the game's 3D models and\ncombining them with the game background to create game-like synthetic images\nand to generate the corresponding labels automatically. In an experiment I\ncompared a model trained on synthetic data to a model trained on hand labeled\ndata and a model trained on a combined dataset. The model trained on the\nsynthetic data showed higher detection precision on more classes and more\nreliable tracking performance of the player character. The model trained on the\ncombined dataset did not perform better because of the different formats of the\nolder hand labeled dataset and the synthetic data."", ""A system is presented that segments, clusters and predicts musical audio in\nan unsupervised manner, adjusting the number of (timbre) clusters\ninstantaneously to the audio input. A sequence learning algorithm adapts its\nstructure to a dynamically changing clustering tree. The flow of the system is\nas follows: 1) segmentation by onset detection, 2) timbre representation of\neach segment by Mel frequency cepstrum coefficients, 3) discretization by\nincremental clustering, yielding a tree of different sound classes (e.g.\ninstruments) that can grow or shrink on the fly driven by the instantaneous\nsound events, resulting in a discrete symbol sequence, 4) extraction of\nstatistical regularities of the symbol sequence, using hierarchical N-grams and\nthe newly introduced conceptual Boltzmann machine, and 5) prediction of the\nnext sound event in the sequence. The system's robustness is assessed with\nrespect to complexity and noisiness of the signal. Clustering in isolation\nyields an adjusted Rand index (ARI) of 82.7% / 85.7% for data sets of singing\nvoice and drums. Onset detection jointly with clustering achieve an ARI of\n81.3% / 76.3% and the prediction of the entire system yields an ARI of 27.2% /\n39.2%."", 'Detection of protein-protein interactions (PPIs) plays a vital role in\nmolecular biology. Particularly, infections are caused by the interactions of\nhost and pathogen proteins. It is important to identify host-pathogen\ninteractions (HPIs) to discover new drugs to counter infectious diseases.\nConventional wet lab PPI prediction techniques have limitations in terms of\nlarge scale application and budget. Hence, computational approaches are\ndeveloped to predict PPIs. This study aims to develop large margin machine\nlearning models to predict interspecies PPIs with a special interest in\nhost-pathogen protein interactions (HPIs). Especially, we focus on seeking\nanswers to three queries that arise while developing an HPI predictor. 1) How\nshould we select negative samples? 2) What should be the size of negative\nsamples as compared to the positive samples? 3) What type of margin violation\npenalty should be used to train the predictor? We compare two available methods\nfor negative sampling. Moreover, we propose a new method of assigning weights\nto each training example in weighted SVM depending on the distance of the\nnegative examples from the positive examples. We have also developed a web\nserver for our HPI predictor called HoPItor (Host Pathogen Interaction\npredicTOR) that can predict interactions between human and viral proteins. This\nwebserver can be accessed at the URL:\nhttp://faculty.pieas.edu.pk/fayyaz/software.html#HoPItor.', 'We present an approach to analyze $C^1(\\mathbb{R}^m)$ functions that\naddresses limitations present in the Active Subspaces (AS) method of\nConstantine et al.(2015; 2014). Under appropriate hypotheses, our Active\nManifolds (AM) method identifies a 1-D curve in the domain (the active\nmanifold) on which nearly all values of the unknown function are attained, and\nwhich can be exploited for approximation or analysis, especially when $m$ is\nlarge (high-dimensional input space). We provide theorems justifying our AM\ntechnique and an algorithm permitting functional approximation and sensitivity\nanalysis. Using accessible, low-dimensional functions as initial examples, we\nshow AM reduces approximation error by an order of magnitude compared to AS, at\nthe expense of more computation. Following this, we revisit the sensitivity\nanalysis by Glaws et al. (2017), who apply AS to analyze a magnetohydrodynamic\npower generator model, and compare the performance of AM on the same data. Our\nanalysis provides detailed information not captured by AS, exhibiting the\ninfluence of each parameter individually along an active manifold. Overall, AM\nrepresents a novel technique for analyzing functional models with benefits\nincluding: reducing $m$-dimensional analysis to a 1-D analogue, permitting more\naccurate regression than AS (at more computational expense), enabling more\ninformative sensitivity analysis, and granting accessible visualizations(2-D\nplots) of parameter sensitivity along the AM.', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', 'We emphasize the importance of asking the right question when interpreting\nthe decisions of a learning model. We discuss a natural extension of the\ntheoretical machinery from Janzing et. al. 2020, which answers the question\n""Why did my model predict a person has cancer?"" for answering a more involved\nquestion, ""What caused my model to predict a person has cancer?"" While the\nformer quantifies the direct effects of variables on the model, the latter also\naccounts for indirect effects, thereby providing meaningful insights wherever\nhuman beings can reason in terms of cause and effect. We propose three broad\ncategories for interpretations: observational, model-specific and causal each\nof which are significant in their own right. Furthermore, this paper quantifies\nfeature relevance by weaving different natures of interpretations together with\ndifferent measures as characteristic functions for Shapley symmetrization.\nBesides the widely used expected value of the model, we also discuss measures\nof statistical uncertainty and dispersion as informative candidates, and their\nmerits in generating explanations for each data point, some of which are used\nin this context for the first time. These measures are not only useful for\nstudying the influence of variables on the model output, but also on the\npredictive performance of the model, and for that we propose relevant\ncharacteristic functions that are also used for the first time.', ""We introduce a new task of rephrasing for a more natural virtual assistant.\nCurrently, virtual assistants work in the paradigm of intent slot tagging and\nthe slot values are directly passed as-is to the execution engine. However,\nthis setup fails in some scenarios such as messaging when the query given by\nthe user needs to be changed before repeating it or sending it to another user.\nFor example, for queries like 'ask my wife if she can pick up the kids' or\n'remind me to take my pills', we need to rephrase the content to 'can you pick\nup the kids' and 'take your pills' In this paper, we study the problem of\nrephrasing with messaging as a use case and release a dataset of 3000 pairs of\noriginal query and rephrased query. We show that BART, a pre-trained\ntransformers-based masked language model with auto-regressive decoding, is a\nstrong baseline for the task, and show improvements by adding a copy-pointer\nand copy loss to it. We analyze different tradeoffs of BART-based and\nLSTM-based seq2seq models, and propose a distilled LSTM-based seq2seq as the\nbest practical model."", 'Artificial self-perception is the machine ability to perceive its own body,\ni.e., the mastery of modal and intermodal contingencies of performing an action\nwith a specific sensors/actuators body configuration. In other words, the\nspatio-temporal patterns that relate its sensors (e.g. visual, proprioceptive,\ntactile, etc.), its actions and its body latent variables are responsible of\nthe distinction between its own body and the rest of the world. This paper\ndescribes some of the latest approaches for modelling artificial body\nself-perception: from Bayesian estimation to deep learning. Results show the\npotential of these free-model unsupervised or semi-supervised\ncrossmodal/intermodal learning approaches. However, there are still challenges\nthat should be overcome before we achieve artificial multisensory body\nperception.', 'Natural language processing and machine learning algorithms have been shown\nto be effective in a variety of applications. In this work, we contribute to\nthe area of AI adoption in the public sector. We present an automated system\nthat was used to process textual information, generate important keywords, and\nautomatically summarize key elements of the Meadville community statements. We\nalso describe the process of collaboration with My Meadville administrators\nduring the development of our system. My Meadville, a community initiative,\nsupported by the city of Meadville conducted a large number of interviews with\nthe residents of Meadville during the community events and transcribed these\ninterviews into textual data files. Their goal was to uncover the issues of\nimportance to the Meadville residents in an attempt to enhance public services.\nOur AI system cleans and pre-processes the interview data, then using machine\nlearning algorithms it finds important keywords and key excerpts from each\ninterview. It also provides searching functionality to find excerpts from\nrelevant interviews based on specific keywords. Our automated system allowed\nthe city to save over 300 hours of human labor that would have taken to read\nall interviews and highlight important points. Our findings are being used by\nMy Meadville initiative to locate important information from the collected data\nset for ongoing community enhancement projects, to highlight relevant community\nassets, and to assist in identifying the steps to be taken based on the\nconcerns and areas of improvement identified by the community members.', 'In this dissertation, I derive a new method to estimate the\nVapnik-Chervonenkis Dimension (VCD) for the class of linear functions. This\nmethod is inspired by the technique developed by Vapnik et al. Vapnik et al.\n(1994). My contribution rests on the approximation of the expected maximum\ndifference between two empirical Losses (EMDBTEL). In fact, I use a\ncross-validated form of the error to compute the EMDBTEL, and I make the bound\non the EMDBTEL tighter by minimizing a constant in of its right upper bound. I\nalso derive two bounds for the true unknown risk using the additive (ERM1) and\nthe multiplicative (ERM2) Chernoff bounds. These bounds depend on the estimated\nVCD and the empirical risk. These bounds can be used to perform model selection\nand to declare with high probability, the chosen model will perform better\nwithout making strong assumptions about the data generating process (DG).\n  I measure the accuracy of my technique on simulated datasets and also on\nthree real datasets. The model selection provided by VCD was always as good as\nif not better than the other methods under reasonable conditions.', 'Voice Onset Time (VOT), a key measurement of speech for basic research and\napplied medical studies, is the time between the onset of a stop burst and the\nonset of voicing. When the voicing onset precedes burst onset the VOT is\nnegative; if voicing onset follows the burst, it is positive. In this work, we\npresent a deep-learning model for accurate and reliable measurement of VOT in\nnaturalistic speech. The proposed system addresses two critical issues: it can\nmeasure positive and negative VOT equally well, and it is trained to be robust\nto variation across annotations. Our approach is based on the structured\nprediction framework, where the feature functions are defined to be RNNs. These\nlearn to capture segmental variation in the signal. Results suggest that our\nmethod substantially improves over the current state-of-the-art. In contrast to\nprevious work, our Deep and Robust VOT annotator, Dr.VOT, can successfully\nestimate negative VOTs while maintaining state-of-the-art performance on\npositive VOTs. This high level of performance generalizes to new corpora\nwithout further retraining. Index Terms: structured prediction, multi-task\nlearning, adversarial training, recurrent neural networks, sequence\nsegmentation.', 'We study the low rank regression problem $\\my = M\\mx + \\epsilon$, where $\\mx$\nand $\\my$ are $d_1$ and $d_2$ dimensional vectors respectively. We consider the\nextreme high-dimensional setting where the number of observations $n$ is less\nthan $d_1 + d_2$. Existing algorithms are designed for settings where $n$ is\ntypically as large as $\\Rank(M)(d_1+d_2)$. This work provides an efficient\nalgorithm which only involves two SVD, and establishes statistical guarantees\non its performance. The algorithm decouples the problem by first estimating the\nprecision matrix of the features, and then solving the matrix denoising\nproblem. To complement the upper bound, we introduce new techniques for\nestablishing lower bounds on the performance of any algorithm for this problem.\nOur preliminary experiments confirm that our algorithm often out-performs\nexisting baselines, and is always at least competitive.', 'Machine learning algorithms operating on mobile networks can be characterized\ninto three different categories. First is the classical situation in which the\nend-user devices send their data to a central server where this data is used to\ntrain a model. Second is the distributed setting in which each device trains\nits own model and send its model parameters to a central server where these\nmodel parameters are aggregated to create one final model. Third is the\nfederated learning setting in which, at any given time $t$, a certain number of\nactive end users train with their own local data along with feedback provided\nby the central server and then send their newly estimated model parameters to\nthe central server. The server, then, aggregates these new parameters, updates\nits own model, and feeds the updated parameters back to all the end users,\ncontinuing this process until it converges.\n  The main objective of this work is to provide an information-theoretic\nframework for all of the aforementioned learning paradigms. Moreover, using the\nprovided framework, we develop upper and lower bounds on the generalization\nerror together with bounds on the privacy leakage in the classical, distributed\nand federated learning settings.\n  Keywords: Federated Learning, Distributed Learning, Machine Learning, Model\nAggregation.', 'Additive Manufacturing (AM) is a manufacturing paradigm that builds\nthree-dimensional objects from a computer-aided design model by successively\nadding material layer by layer. AM has become very popular in the past decade\ndue to its utility for fast prototyping such as 3D printing as well as\nmanufacturing functional parts with complex geometries using processes such as\nlaser metal deposition that would be difficult to create using traditional\nmachining. As the process for creating an intricate part for an expensive metal\nsuch as Titanium is prohibitive with respect to cost, computational models are\nused to simulate the behavior of AM processes before the experimental run.\nHowever, as the simulations are computationally costly and time-consuming for\npredicting multiscale multi-physics phenomena in AM, physics-informed\ndata-driven machine-learning systems for predicting the behavior of AM\nprocesses are immensely beneficial. Such models accelerate not only multiscale\nsimulation tools but also empower real-time control systems using in-situ data.\nIn this paper, we design and develop essential components of a scientific\nframework for developing a data-driven model-based real-time control system.\nFinite element methods are employed for solving time-dependent heat equations\nand developing the database. The proposed framework uses extremely randomized\ntrees - an ensemble of bagged decision trees as the regression algorithm\niteratively using temperatures of prior voxels and laser information as inputs\nto predict temperatures of subsequent voxels. The models achieve mean absolute\npercentage errors below 1% for predicting temperature profiles for AM\nprocesses. The code is made available for the research community at\nhttps://github.com/paularindam/ml-iter-additive.', 'The ability to accurately quantify the performance an additively manufactured\n(AM) product is important for a widespread industry adoption of AM as the\ndesign is required to: (1) satisfy geometrical constraints, (2) satisfy\nstructural constraints dictated by its intended function, and (3) be cost\neffective compared to traditional manufacturing methods. Optimization\ntechniques offer design aids in creating cost-effective structures that meet\nthe prescribed structural objectives. The fundamental problem in existing\napproaches lies in the difficulty to quantify the structural performance as\neach unique design leads to a new set of analyses to determine the structural\nrobustness and such analyses can be very costly due to the complexity of in-use\nforces experienced by the structure. This work develops computationally\ntractable methods tailored to maximize the structural performance of AM\nproducts. A geometry preserving build orientation optimization method as well\nas data-driven shape optimization approaches to structural design are\npresented. Proposed methods greatly enhance the value of AM technology by\ntaking advantage of the design space enabled by it for a broad class of\nproblems involving complex in-use loads.', 'We address the problem of solving mixed random linear equations. We have\nunlabeled observations coming from multiple linear regressions, and each\nobservation corresponds to exactly one of the regression models. The goal is to\nlearn the linear regressors from the observations. Classically, Alternating\nMinimization (AM) (which is a variant of Expectation Maximization (EM)) is used\nto solve this problem. AM iteratively alternates between the estimation of\nlabels and solving the regression problems with the estimated labels.\nEmpirically, it is observed that, for a large variety of non-convex problems\nincluding mixed linear regression, AM converges at a much faster rate compared\nto gradient based algorithms. However, the existing theory suggests similar\nrate of convergence for AM and gradient based methods, failing to capture this\nempirical behavior. In this paper, we close this gap between theory and\npractice for the special case of a mixture of $2$ linear regressions. We show\nthat, provided initialized properly, AM enjoys a \\emph{super-linear} rate of\nconvergence in certain parameter regimes. To the best of our knowledge, this is\nthe first work that theoretically establishes such rate for AM. Hence, if we\nwant to recover the unknown regressors upto an error (in $\\ell_2$ norm) of\n$\\epsilon$, AM only takes $\\mathcal{O}(\\log \\log (1/\\epsilon))$ iterations.\nFurthermore, we compare AM with a gradient based heuristic algorithm\nempirically and show that AM dominates in iteration complexity as well as\nwall-clock time.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : Deep learning models are notoriously challenging to interpret due to their inherent black-box nature. In my project, I aim to demystify these models by developing alternative, more transparent models that maintain high performance. My current approach involves simplifying the network and reducing its parameters as much as possible without compromising performance. Once the network is sufficiently simplified, we can break it down into more interpretable components. Each component should be simple enough to visualize, and by visualizing the interaction of each component, we can gain a clearer understanding of the entire mechanism and decision process of the black-box model. Consequently, I plan to develop a visualization toolkit to illustrate each component of the simplified neural network model. I will also apply my proposed methodology and toolkit to various real-world scenarios where transparent decision-making is crucial, such as in credit score assessment and loan approval processes.","['Deep neural network (DNN) models have achieved phenomenal success for\napplications in many domains, ranging from academic research in science and\nengineering to industry and business. The modeling power of DNN is believed to\nhave come from the complexity and over-parameterization of the model, which on\nthe other hand has been criticized for the lack of interpretation. Although\ncertainly not true for every application, in some applications, especially in\neconomics, social science, healthcare industry, and administrative decision\nmaking, scientists or practitioners are resistant to use predictions made by a\nblack-box system for multiple reasons. One reason is that a major purpose of a\nstudy can be to make discoveries based upon the prediction function, e.g., to\nreveal the relationships between measurements. Another reason can be that the\ntraining dataset is not large enough to make researchers feel completely sure\nabout a purely data-driven result. Being able to examine and interpret the\nprediction function will enable researchers to connect the result with existing\nknowledge or gain insights about new directions to explore. Although classic\nstatistical models are much more explainable, their accuracy often falls\nconsiderably below DNN. In this paper, we propose an approach to fill the gap\nbetween relatively simple explainable models and DNN such that we can more\nflexibly tune the trade-off between interpretability and accuracy. Our main\nidea is a mixture of discriminative models that is trained with the guidance\nfrom a DNN. Although mixtures of discriminative models have been studied\nbefore, our way of generating the mixture is quite different.', 'Deep learning has been shown to outperform traditional machine learning\nalgorithms across a wide range of problem domains. However, current deep\nlearning algorithms have been criticized as uninterpretable ""black-boxes"" which\ncannot explain their decision making processes. This is a major shortcoming\nthat prevents the widespread application of deep learning to domains with\nregulatory processes such as finance. As such, industries such as finance have\nto rely on traditional models like decision trees that are much more\ninterpretable but less effective than deep learning for complex problems. In\nthis paper, we propose CLEAR-Trade, a novel financial AI visualization\nframework for deep learning-driven stock market prediction that mitigates the\ninterpretability issue of deep learning methods. In particular, CLEAR-Trade\nprovides a effective way to visualize and explain decisions made by deep stock\nmarket prediction models. We show the efficacy of CLEAR-Trade in enhancing the\ninterpretability of stock market prediction by conducting experiments based on\nS&P 500 stock index prediction. The results demonstrate that CLEAR-Trade can\nprovide significant insight into the decision-making process of deep\nlearning-driven financial models, particularly for regulatory processes, thus\nimproving their potential uptake in the financial industry.', ""Deep neural network (DNN) models have recently obtained state-of-the-art\nprediction accuracy for the transcription factor binding (TFBS) site\nclassification task. However, it remains unclear how these approaches identify\nmeaningful DNA sequence signals and give insights as to why TFs bind to certain\nlocations. In this paper, we propose a toolkit called the Deep Motif Dashboard\n(DeMo Dashboard) which provides a suite of visualization strategies to extract\nmotifs, or sequence patterns from deep neural network models for TFBS\nclassification. We demonstrate how to visualize and understand three important\nDNN models: convolutional, recurrent, and convolutional-recurrent networks. Our\nfirst visualization method is finding a test sequence's saliency map which uses\nfirst-order derivatives to describe the importance of each nucleotide in making\nthe final prediction. Second, considering recurrent models make predictions in\na temporal manner (from one end of a TFBS sequence to the other), we introduce\ntemporal output scores, indicating the prediction score of a model over time\nfor a sequential input. Lastly, a class-specific visualization strategy finds\nthe optimal input sequence for a given TFBS positive class via stochastic\ngradient optimization. Our experimental results indicate that a\nconvolutional-recurrent architecture performs the best among the three\narchitectures. The visualization techniques indicate that CNN-RNN makes\npredictions by modeling both motifs as well as dependencies among them."", 'For reliability, it is important that the predictions made by machine\nlearning methods are interpretable by human. In general, deep neural networks\n(DNNs) can provide accurate predictions, although it is difficult to interpret\nwhy such predictions are obtained by DNNs. On the other hand, interpretation of\nlinear models is easy, although their predictive performance would be low since\nreal-world data is often intrinsically non-linear. To combine both the benefits\nof the high predictive performance of DNNs and high interpretability of linear\nmodels into a single model, we propose neural generators of sparse local linear\nmodels (NGSLLs). The sparse local linear models have high flexibility as they\ncan approximate non-linear functions. The NGSLL generates sparse linear weights\nfor each sample using DNNs that take original representations of each sample\n(e.g., word sequence) and their simplified representations (e.g., bag-of-words)\nas input. By extracting features from the original representations, the weights\ncan contain rich information to achieve high predictive performance.\nAdditionally, the prediction is interpretable because it is obtained by the\ninner product between the simplified representations and the sparse weights,\nwhere only a small number of weights are selected by our gate module in the\nNGSLL. In experiments with real-world datasets, we demonstrate the\neffectiveness of the NGSLL quantitatively and qualitatively by evaluating\nprediction performance and visualizing generated weights on image and text\nclassification tasks.', 'The practical impact of deep learning on complex supervised learning problems\nhas been significant, so much so that almost every Artificial Intelligence\nproblem, or at least a portion thereof, has been somehow recast as a deep\nlearning problem. The applications appeal is significant, but this appeal is\nincreasingly challenged by what some call the challenge of explainability, or\nmore generally the more traditional challenge of debuggability: if the outcomes\nof a deep learning process produce unexpected results (e.g., less than expected\nperformance of a classifier), then there is little available in the way of\ntheories or tools to help investigate the potential causes of such unexpected\nbehavior, especially when this behavior could impact people\'s lives. We\ndescribe a preliminary framework to help address this issue, which we call\n""deep visual explanation"" (DVE). ""Deep,"" because it is the development and\nperformance of deep neural network models that we want to understand. ""Visual,""\nbecause we believe that the most rapid insight into a complex multi-dimensional\nmodel is provided by appropriate visualization techniques, and ""Explanation,""\nbecause in the spectrum from instrumentation by inserting print statements to\nthe abductive inference of explanatory hypotheses, we believe that the key to\nunderstanding deep learning relies on the identification and exposure of\nhypotheses about the performance behavior of a learned deep model. In the\nexposition of our preliminary framework, we use relatively straightforward\nimage classification examples and a variety of choices on initial configuration\nof a deep model building scenario. By careful but not complicated\ninstrumentation, we expose classification outcomes of deep models using\nvisualization, and also show initial results for one potential application of\ninterpretability.', 'Automated machine learning (AutoML) usually involves several crucial\ncomponents, such as Data Augmentation (DA) policy, Hyper-Parameter Optimization\n(HPO), and Neural Architecture Search (NAS). Although many strategies have been\ndeveloped for automating these components in separation, joint optimization of\nthese components remains challenging due to the largely increased search\ndimension and the variant input types of each component. In parallel to this,\nthe common practice of searching for the optimal architecture first and then\nretraining it before deployment in NAS often suffers from low performance\ncorrelation between the searching and retraining stages. An end-to-end solution\nthat integrates the AutoML components and returns a ready-to-use model at the\nend of the search is desirable. In view of these, we propose DHA, which\nachieves joint optimization of Data augmentation policy, Hyper-parameter and\nArchitecture. Specifically, end-to-end NAS is achieved in a differentiable\nmanner by optimizing a compressed lower-dimensional feature space, while DA\npolicy and HPO are regarded as dynamic schedulers, which adapt themselves to\nthe update of network parameters and network architecture at the same time.\nExperiments show that DHA achieves state-of-the-art (SOTA) results on various\ndatasets and search spaces. To the best of our knowledge, we are the first to\nefficiently and jointly optimize DA policy, NAS, and HPO in an end-to-end\nmanner without retraining.', 'With the availability of large databases and recent improvements in deep\nlearning methodology, the performance of AI systems is reaching or even\nexceeding the human level on an increasing number of complex tasks. Impressive\nexamples of this development can be found in domains such as image\nclassification, sentiment analysis, speech understanding or strategic game\nplaying. However, because of their nested non-linear structure, these highly\nsuccessful machine learning and artificial intelligence models are usually\napplied in a black box manner, i.e., no information is provided about what\nexactly makes them arrive at their predictions. Since this lack of transparency\ncan be a major drawback, e.g., in medical applications, the development of\nmethods for visualizing, explaining and interpreting deep learning models has\nrecently attracted increasing attention. This paper summarizes recent\ndevelopments in this field and makes a plea for more interpretability in\nartificial intelligence. Furthermore, it presents two approaches to explaining\npredictions of deep learning models, one method which computes the sensitivity\nof the prediction with respect to changes in the input and one approach which\nmeaningfully decomposes the decision in terms of the input variables. These\nmethods are evaluated on three classification tasks.', 'The ability of deep learning (DL) to improve the practice of medicine and its\nclinical outcomes faces a looming obstacle: model interpretation. Without\ndescription of how outputs are generated, a collaborating physician can neither\nresolve when the model\'s conclusions are in conflict with his or her own, nor\nlearn to anticipate model behavior. Current research aims to interpret networks\nthat diagnose ECG recordings, which has great potential impact as recordings\nbecome more personalized and widely deployed. A generalizable impact beyond\nECGs lies in the ability to provide a rich test-bed for the development of\ninterpretive techniques in medicine. Interpretive techniques for Deep Neural\nNetworks (DNNs), however, tend to be heuristic and observational in nature,\nlacking the mathematical rigor one might expect in the analysis of math\nequations. The motivation of this paper is to offer a third option, a\nscientific approach. We treat the model output itself as a phenomenon to be\nexplained through component parts and equations governing their behavior. We\nargue that these component parts should also be ""black boxes"" --additional\ntargets to interpret heuristically with clear functional connection to the\noriginal. We show how to rigorously factor a DNN into a hierarchical equation\nconsisting of black box variables. This is not a subdivision into physical\nparts, like an organism into its cells; it is but one choice of an equation\ninto a collection of abstract functions. Yet, for DNNs trained to identify\nnormal ECG waveforms on PhysioNet 2017 Challenge data, we demonstrate this\nchoice yields interpretable component models identified with visual composite\nsketches of ECG samples in corresponding input regions. Moreover, the recursion\ndistills this interpretation: additional factorization of component black boxes\ncorresponds to ECG partitions that are more morphologically pure.', 'Model interpretability is a requirement in many applications in which crucial\ndecisions are made by users relying on a model\'s outputs. The recent movement\nfor ""algorithmic fairness"" also stipulates explainability, and therefore\ninterpretability of learning models. And yet the most successful contemporary\nMachine Learning approaches, the Deep Neural Networks, produce models that are\nhighly non-interpretable. We attempt to address this challenge by proposing a\ntechnique called CNN-INTE to interpret deep Convolutional Neural Networks (CNN)\nvia meta-learning. In this work, we interpret a specific hidden layer of the\ndeep CNN model on the MNIST image dataset. We use a clustering algorithm in a\ntwo-level structure to find the meta-level training data and Random Forest as\nbase learning algorithms to generate the meta-level test data. The\ninterpretation results are displayed visually via diagrams, which clearly\nindicates how a specific test instance is classified. Our method achieves\nglobal interpretation for all the test instances without sacrificing the\naccuracy obtained by the original deep CNN model. This means our model is\nfaithful to the deep CNN model, which leads to reliable interpretations.', 'With the unprecedented proliferation of machine learning software, there is\nan ever-increasing need to generate efficient code for such applications.\nState-of-the-art deep-learning compilers like TVM and Halide incorporate a\nlearning-based performance model to search the space of valid implementations\nof a given deep learning algorithm. For a given application, the model\ngenerates a performance metric such as the run time without executing the\napplication on hardware. Such models speed up the compilation process by\nobviating the need to benchmark an enormous number of candidate\nimplementations, referred to as schedules, on hardware. Existing performance\nmodels employ feed-forward networks, recurrent networks, or decision tree\nensembles to estimate the performance of different implementations of a neural\nnetwork. Graphs present a natural and intuitive way to model deep-learning\nnetworks where each node represents a computational stage or operation.\nIncorporating the inherent graph structure of these workloads in the\nperformance model can enable a better representation and learning of\ninter-stage interactions. The accuracy of a performance model has direct\nimplications on the efficiency of the search strategy, making it a crucial\ncomponent of this class of deep-learning compilers. In this work, we develop a\nnovel performance model that adopts a graph representation. In our model, each\nstage of computation represents a node characterized by features that capture\nthe operations performed by the stage. The interaction between nodes is\nachieved using graph convolutions. Experimental evaluation shows a 7:75x and\n12x reduction in prediction error compared to the Halide and TVM models,\nrespectively.', 'Interpretable machine learning has become a strong competitor for traditional\nblack-box models. However, the possible loss of the predictive performance for\ngaining interpretability is often inevitable, putting practitioners in a\ndilemma of choosing between high accuracy (black-box models) and\ninterpretability (interpretable models). In this work, we propose a novel\nframework for building a Hybrid Predictive Model (HPM) that integrates an\ninterpretable model with any black-box model to combine their strengths. The\ninterpretable model substitutes the black-box model on a subset of data where\nthe black-box is overkill or nearly overkill, gaining transparency at no or low\ncost of the predictive accuracy. We design a principled objective function that\nconsiders predictive accuracy, model interpretability, and model transparency\n(defined as the percentage of data processed by the interpretable substitute.)\nUnder this framework, we propose two hybrid models, one substituting with\nassociation rules and the other with linear models, and we design customized\ntraining algorithms for both models. We test the hybrid models on structured\ndata and text data where interpretable models collaborate with various\nstate-of-the-art black-box models. Results show that hybrid models obtain an\nefficient trade-off between transparency and predictive performance,\ncharacterized by our proposed efficient frontiers.', 'State of the art machine learning algorithms are highly optimized to provide\nthe optimal prediction possible, naturally resulting in complex models. While\nthese models often outperform simpler more interpretable models by order of\nmagnitudes, in terms of understanding the way the model functions, we are often\nfacing a ""black box"".\n  In this paper we suggest a simple method to interpret the behavior of any\npredictive model, both for regression and classification. Given a particular\nmodel, the information required to interpret it can be obtained by studying the\npartial derivatives of the model with respect to the input. We exemplify this\ninsight by interpreting convolutional and multi-layer neural networks in the\nfield of natural language processing.', 'Recent advances in deep learning have achieved impressive gains in\nclassification accuracy on a variety of types of data, including images and\ntext. Despite these gains, however, concerns have been raised about the\ncalibration, robustness, and interpretability of these models. In this paper we\npropose a simple way to modify any conventional deep architecture to\nautomatically provide more transparent explanations for classification\ndecisions, as well as an intuitive notion of the credibility of each\nprediction. Specifically, we draw on ideas from nonparametric kernel\nregression, and propose to predict labels based on a weighted sum of training\ninstances, where the weights are determined by distance in a learned\ninstance-embedding space. Working within the framework of conformal methods, we\npropose a new measure of nonconformity suggested by our model, and\nexperimentally validate the accompanying theoretical expectations,\ndemonstrating improved transparency, controlled error rates, and robustness to\nout-of-domain data, without compromising on accuracy or calibration.', ""The work concerns the problem of reducing a pre-trained deep neuronal network\nto a smaller network, with just few layers, whilst retaining the network's\nfunctionality on a given task\n  The proposed approach is motivated by the observation that the aim to deliver\nthe highest accuracy possible in the broadest range of operational conditions,\nwhich many deep neural networks models strive to achieve, may not necessarily\nbe always needed, desired, or even achievable due to the lack of data or\ntechnical constraints. In relation to the face recognition problem, we\nformulated an example of such a usecase, the `backyard dog' problem. The\n`backyard dog', implemented by a lean network, should correctly identify\nmembers from a limited group of individuals, a `family', and should distinguish\nbetween them. At the same time, the network must produce an alarm to an image\nof an individual who is not in a member of the family. To produce such a\nnetwork, we propose a shallowing algorithm. The algorithm takes an existing\ndeep learning model on its input and outputs a shallowed version of it. The\nalgorithm is non-iterative and is based on the Advanced Supervised Principal\nComponent Analysis. Performance of the algorithm is assessed in exhaustive\nnumerical experiments. In the above usecase, the `backyard dog' problem, the\nmethod is capable of drastically reducing the depth of deep learning neural\nnetworks, albeit at the cost of mild performance deterioration.\n  We developed a simple non-iterative method for shallowing down pre-trained\ndeep networks. The method is generic in the sense that it applies to a broad\nclass of feed-forward networks, and is based on the Advanced Supervise\nPrincipal Component Analysis. The method enables generation of families of\nsmaller-size shallower specialized networks tuned for specific operational\nconditions and tasks from a single larger and more universal legacy network."", 'Due to the capability of deep learning to perform well in high dimensional\nproblems, deep reinforcement learning agents perform well in challenging tasks\nsuch as Atari 2600 games. However, clearly explaining why a certain action is\ntaken by the agent can be as important as the decision itself. Deep\nreinforcement learning models, as other deep learning models, tend to be opaque\nin their decision-making process. In this work, we propose to make deep\nreinforcement learning more transparent by visualizing the evidence on which\nthe agent bases its decision. In this work, we emphasize the importance of\nproducing a justification for an observed action, which could be applied to a\nblack-box decision agent.', 'Deep learning has become popular because of its potential to achieve high\naccuracy in prediction tasks. However, accuracy is not always the only goal of\nstatistical modelling, especially for models developed as part of scientific\nresearch. Rather, many scientific models are developed to facilitate scientific\ndiscovery, by which we mean to abstract a human-understandable representation\nof the natural world. Unfortunately, the opacity of deep neural networks limit\ntheir role in scientific discovery, creating a new demand for models that are\ntransparently interpretable. This article is a field guide to transparent model\ndesign. It provides a taxonomy of transparent model design concepts, a\npractical workflow for putting design concepts into practice, and a general\ntemplate for reporting design choices. We hope this field guide will help\nresearchers more effectively design transparently interpretable models, and\nthus enable them to use deep learning for scientific discovery.', 'In recent years, machine learning (ML) has become a key enabling technology\nfor the sciences and industry. Especially through improvements in methodology,\nthe availability of large databases and increased computational power, today\'s\nML algorithms are able to achieve excellent performance (at times even\nexceeding the human level) on an increasing number of complex tasks. Deep\nlearning models are at the forefront of this development. However, due to their\nnested non-linear structure, these powerful models have been generally\nconsidered ""black boxes"", not providing any information about what exactly\nmakes them arrive at their predictions. Since in many applications, e.g., in\nthe medical domain, such lack of transparency may be not acceptable, the\ndevelopment of methods for visualizing, explaining and interpreting deep\nlearning models has recently attracted increasing attention. This introductory\npaper presents recent developments and applications in this field and makes a\nplea for a wider use of explainable learning algorithms in practice.', 'Deep neural networks are learning models with a very high capacity and\ntherefore prone to over-fitting. Many regularization techniques such as\nDropout, DropConnect, and weight decay all attempt to solve the problem of\nover-fitting by reducing the capacity of their respective models (Srivastava et\nal., 2014), (Wan et al., 2013), (Krogh & Hertz, 1992). In this paper we\nintroduce a new form of regularization that guides the learning problem in a\nway that reduces over-fitting without sacrificing the capacity of the model.\nThe mistakes that models make in early stages of training carry information\nabout the learning problem. By adjusting the labels of the current epoch of\ntraining through a weighted average of the real labels, and an exponential\naverage of the past soft-targets we achieved a regularization scheme as\npowerful as Dropout without necessarily reducing the capacity of the model, and\nsimplified the complexity of the learning problem. SoftTarget regularization\nproved to be an effective tool in various neural network architectures.', 'We propose a possible solution to a public challenge posed by the Fair Isaac\nCorporation (FICO), which is to provide an explainable model for credit risk\nassessment. Rather than present a black box model and explain it afterwards, we\nprovide a globally interpretable model that is as accurate as other neural\nnetworks. Our ""two-layer additive risk model"" is decomposable into subscales,\nwhere each node in the second layer represents a meaningful subscale, and all\nof the nonlinearities are transparent. We provide three types of explanations\nthat are simpler than, but consistent with, the global model. One of these\nexplanation methods involves solving a minimum set cover problem to find\nhigh-support globally-consistent explanations. We present a new online\nvisualization tool to allow users to explore the global model and its\nexplanations.', 'Physical motion models offer interpretable predictions for the motion of\nvehicles. However, some model parameters, such as those related to aero- and\nhydrodynamics, are expensive to measure and are often only roughly approximated\nreducing prediction accuracy. Recurrent neural networks achieve high prediction\naccuracy at low cost, as they can use cheap measurements collected during\nroutine operation of the vehicle, but their results are hard to interpret. To\nprecisely predict vehicle states without expensive measurements of physical\nparameters, we propose a hybrid approach combining deep learning and physical\nmotion models including a novel two-phase training procedure. We achieve\ninterpretability by restricting the output range of the deep neural network as\npart of the hybrid model, which limits the uncertainty introduced by the neural\nnetwork to a known quantity. We have evaluated our approach for the use case of\nship and quadcopter motion. The results show that our hybrid model can improve\nmodel interpretability with no decrease in accuracy compared to existing deep\nlearning approaches.', 'While deep learning has achieved great success in many fields, one common\ncriticism about deep learning is its lack of interpretability. In most cases,\nthe hidden units in a deep neural network do not have a clear semantic meaning\nor correspond to any physical entities. However, model interpretability and\nexplainability are crucial in many biomedical applications. To address this\nchallenge, we developed the Factor Graph Neural Network model that is\ninterpretable and predictable by combining probabilistic graphical models with\ndeep learning. We directly encode biological knowledge such as Gene Ontology as\na factor graph into the model architecture, making the model transparent and\ninterpretable. Furthermore, we devised an attention mechanism that can capture\nmulti-scale hierarchical interactions among biological entities such as genes\nand Gene Ontology terms. With parameter sharing mechanism, the unrolled Factor\nGraph Neural Network model can be trained with stochastic depth and generalize\nwell. We applied our model to two cancer genomic datasets to predict target\nclinical variables and achieved better results than other traditional machine\nlearning and deep learning models. Our model can also be used for gene set\nenrichment analysis and selecting Gene Ontology terms that are important to\ntarget clinical variables.', 'In many machine learning tasks, learning a good representation of the data\ncan be the key to building a well-performant solution. This is because most\nlearning algorithms operate with the features in order to find models for the\ndata. For instance, classification performance can improve if the data is\nmapped to a space where classes are easily separated, and regression can be\nfacilitated by finding a manifold of data in the feature space. As a general\nrule, features are transformed by means of statistical methods such as\nprincipal component analysis, or manifold learning techniques such as Isomap or\nlocally linear embedding. From a plethora of representation learning methods,\none of the most versatile tools is the autoencoder. In this paper we aim to\ndemonstrate how to influence its learned representations to achieve the desired\nlearning behavior. To this end, we present a series of learning tasks: data\nembedding for visualization, image denoising, semantic hashing, detection of\nabnormal behaviors and instance generation. We model them from the\nrepresentation learning perspective, following the state of the art\nmethodologies in each field. A solution is proposed for each task employing\nautoencoders as the only learning method. The theoretical developments are put\ninto practice using a selection of datasets for the different problems and\nimplementing each solution, followed by a discussion of the results in each\ncase study and a brief explanation of other six learning applications. We also\nexplore the current challenges and approaches to explainability in the context\nof autoencoders. All of this helps conclude that, thanks to alterations in\ntheir structure as well as their objective function, autoencoders may be the\ncore of a possible solution to many problems which can be modeled as a\ntransformation of the feature space.', ""Machine learning has recently been widely adopted to address the managerial\ndecision making problems, in which the decision maker needs to be able to\ninterpret the contributions of individual attributes in an explicit form.\nHowever, there is a trade-off between performance and interpretability. Full\ncomplexity models are non-traceable black-box, whereas classic interpretable\nmodels are usually simplified with lower accuracy. This trade-off limits the\napplication of state-of-the-art machine learning models in management problems,\nwhich requires high prediction performance, as well as the understanding of\nindividual attributes' contributions to the model outcome. Multiple criteria\ndecision aiding (MCDA) is a family of analytic approaches to depicting the\nrationale of human decision. It is also limited by strong assumptions. To meet\nthe decision maker's demand for more interpretable machine learning models, we\npropose a novel hybrid method, namely Neural Network-based Multiple Criteria\nDecision Aiding, which combines an additive value model and a fully-connected\nmultilayer perceptron (MLP) to achieve good performance while capturing the\nexplicit relationships between individual attributes and the prediction.\nNN-MCDA has a linear component to characterize such relationships through\nproviding explicit marginal value functions, and a nonlinear component to\ncapture the implicit high-order interactions between attributes and their\ncomplex nonlinear transformations. We demonstrate the effectiveness of NN-MCDA\nwith extensive simulation studies and three real-world datasets. To the best of\nour knowledge, this research is the first to enhance the interpretability of\nmachine learning models with MCDA techniques. The proposed framework also sheds\nlight on how to use machine learning techniques to free MCDA from strong\nassumptions."", 'Contemporary deep learning based video captioning follows encoder-decoder\nframework. In encoder, visual features are extracted with 2D/3D Convolutional\nNeural Networks (CNNs) and a transformed version of those features is passed to\nthe decoder. The decoder uses word embeddings and a language model to map\nvisual features to natural language captions. Due to its composite nature, the\nencoder-decoder pipeline provides the freedom of multiple choices for each of\nits components, e.g the choices of CNNs models, feature transformations, word\nembeddings, and language models etc. Component selection can have drastic\neffects on the overall video captioning performance. However, current\nliterature is void of any systematic investigation in this regard. This article\nfills this gap by providing the first thorough empirical analysis of the role\nthat each major component plays in a contemporary video captioning pipeline. We\nperform extensive experiments by varying the constituent components of the\nvideo captioning framework, and quantify the performance gains that are\npossible by mere component selection. We use the popular MSVD dataset as the\ntest-bed, and demonstrate that substantial performance gains are possible by\ncareful selection of the constituent components without major changes to the\npipeline itself. These results are expected to provide guiding principles for\nfuture research in the fast growing direction of video captioning.', 'Among interpretable machine learning methods, the class of Generalised\nAdditive Neural Networks (GANNs) is referred to as Self-Explaining Neural\nNetworks (SENN) because of the linear dependence on explicit functions of the\ninputs. In binary classification this shows the precise weight that each input\ncontributes towards the logit. The nomogram is a graphical representation of\nthese weights. We show that functions of individual and pairs of variables can\nbe derived from a functional Analysis of Variance (ANOVA) representation,\nenabling an efficient feature selection to be carried by application of the\nlogistic Lasso. This process infers the structure of GANNs which otherwise\nneeds to be predefined. As this method is particularly suited for tabular data,\nit starts by fitting a generic flexible model, in this case a Multi-layer\nPerceptron (MLP) to which the ANOVA decomposition is applied. This has the\nfurther advantage that the resulting GANN can be replicated as a SENN, enabling\nfurther refinement of the univariate and bivariate component functions to take\nplace. The component functions are partial responses hence the SENN is a\npartial response network. The Partial Response Network (PRN) is equally as\ntransparent as a traditional logistic regression model, but capable of\nnon-linear classification with comparable or superior performance to the\noriginal MLP. In other words, the PRN is a fully interpretable representation\nof the MLP, at the level of univariate and bivariate effects. The performance\nof the PRN is shown to be competitive for benchmark data, against\nstate-of-the-art machine learning methods including GBM, SVM and Random\nForests. It is also compared with spline-based Sparse Additive Models (SAM)\nshowing that a semi-parametric representation of the GAM as a neural network\ncan be as effective as the SAM though less constrained by the need to set\nspline nodes.', 'Deep Neural Networks have achieved huge success at a wide spectrum of\napplications from language modeling, computer vision to speech recognition.\nHowever, nowadays, good performance alone is not sufficient to satisfy the\nneeds of practical deployment where interpretability is demanded for cases\ninvolving ethics and mission critical applications. The complex models of Deep\nNeural Networks make it hard to understand and reason the predictions, which\nhinders its further progress. To tackle this problem, we apply the Knowledge\nDistillation technique to distill Deep Neural Networks into decision trees in\norder to attain good performance and interpretability simultaneously. We\nformulate the problem at hand as a multi-output regression problem and the\nexperiments demonstrate that the student model achieves significantly better\naccuracy performance (about 1\\% to 5\\%) than vanilla decision trees at the same\nlevel of tree depth. The experiments are implemented on the TensorFlow platform\nto make it scalable to big datasets. To the best of our knowledge, we are the\nfirst to distill Deep Neural Networks into vanilla decision trees on\nmulti-class datasets.', ""The black-box nature of deep learning models prevents them from being\ncompletely trusted in domains like biomedicine. Most explainability techniques\ndo not capture the concept-based reasoning that human beings follow. In this\nwork, we attempt to understand the behavior of trained models that perform\nimage processing tasks in the medical domain by building a graphical\nrepresentation of the concepts they learn. Extracting such a graphical\nrepresentation of the model's behavior on an abstract, higher conceptual level\nwould unravel the learnings of these models and would help us to evaluate the\nsteps taken by the model for predictions. We show the application of our\nproposed implementation on two biomedical problems - brain tumor segmentation\nand fundus image classification. We provide an alternative graphical\nrepresentation of the model by formulating a concept level graph as discussed\nabove, which makes the problem of intervention to find active inference trails\nmore tractable. Understanding these trails would provide an understanding of\nthe hierarchy of the decision-making process followed by the model. [As well as\noverall nature of model]. Our framework is available at\nhttps://github.com/koriavinash1/BioExp"", 'Despite outperforming the human in many tasks, deep neural network models are\nalso criticized for the lack of transparency and interpretability in decision\nmaking. The opaqueness results in uncertainty and low confidence when deploying\nsuch a model in model sharing scenarios, when the model is developed by a third\nparty. For a supervised machine learning model, sharing training process\nincluding training data provides an effective way to gain trust and to better\nunderstand model predictions. However, it is not always possible to share all\ntraining data due to privacy and policy constraints. In this paper, we propose\na method to disclose a small set of training data that is just sufficient for\nusers to get the insight of a complicated model. The method constructs a\nboundary tree using selected training data and the tree is able to approximate\nthe complicated model with high fidelity. We show that traversing data points\nin the tree gives users significantly better understanding of the model and\npaves the way for trustworthy model sharing.']","['The widespread use of deep neural networks has achieved substantial success\nin many tasks. However, there still exists a huge gap between the operating\nmechanism of deep learning models and human-understandable decision making, so\nthat humans cannot fully trust the predictions made by these models. To date,\nlittle work has been done on how to align the behaviors of deep learning models\nwith human perception in order to train a human-understandable model. To fill\nthis gap, we propose a new framework to train a deep neural network by\nincorporating the prior of human perception into the model learning process.\nOur proposed model mimics the process of perceiving conceptual parts from\nimages and assessing their relative contributions towards the final\nrecognition. The effectiveness of our proposed model is evaluated on two\nclassical visual recognition tasks. The experimental results and analysis\nconfirm our model is able to provide interpretable explanations for its\npredictions, but also maintain competitive recognition accuracy.', ""Is a deep learning model capable of understanding systems governed by certain\nfirst principle laws by only observing the system's output? Can deep learning\nlearn the underlying physics and honor the physics when making predictions? The\nanswers are both positive. In an effort to simulate two-dimensional subsurface\nfluid dynamics in porous media, we found that an accurate deep-learning-based\nproxy model can be taught efficiently by a computationally expensive\nfinite-volume-based simulator. We pose the problem as an image-to-image\nregression, running the simulator with different input parameters to furnish a\nsynthetic training dataset upon which we fit the deep learning models. Since\nthe data is spatiotemporal, we compare the performance of two alternative\ntreatments of time; a convolutional LSTM versus an autoencoder network that\ntreats time as a direct input. Adversarial methods are adopted to address the\nsharp spatial gradient in the fluid dynamic problems. Compared to traditional\nsimulation, the proposed deep learning approach enables much faster forward\ncomputation, which allows us to explore more scenarios with a much larger\nparameter space given the same time. It is shown that the improved forward\ncomputation efficiency is particularly valuable in solving inversion problems,\nwhere the physics model has unknown parameters to be determined by history\nmatching. By computing the pixel-level attention of the trained model, we\nquantify the sensitivity of the deep learning model to key physical parameters\nand hence demonstrate that the inversion problems can be solved with great\nacceleration. We assess the efficacy of the machine learning surrogate in terms\nof its training speed and accuracy. The network can be trained within minutes\nusing limited training data and achieve accuracy that scales desirably with the\namount of training data supplied."", ""Virtual reality (VR) offers immersive visualization and intuitive\ninteraction. We leverage VR to enable any biomedical professional to deploy a\ndeep learning (DL) model for image classification. While DL models can be\npowerful tools for data analysis, they are also challenging to understand and\ndevelop. To make deep learning more accessible and intuitive, we have built a\nvirtual reality-based DL development environment. Within our environment, the\nuser can move tangible objects to construct a neural network only using their\nhands. Our software automatically translates these configurations into a\ntrainable model and then reports its resulting accuracy on a test dataset in\nreal-time. Furthermore, we have enriched the virtual objects with\nvisualizations of the model's components such that users can achieve insight\nabout the DL models that they are developing. With this approach, we bridge the\ngap between professionals in different fields of expertise while offering a\nnovel perspective for model analysis and data interaction. We further suggest\nthat techniques of development and visualization in deep learning can benefit\nby integrating virtual reality."", 'Histopathology image analysis can be considered as a Multiple instance\nlearning (MIL) problem, where the whole slide histopathology image (WSI) is\nregarded as a bag of instances (i.e, patches) and the task is to predict a\nsingle class label to the WSI. However, in many real-life applications such as\ncomputational pathology, discovering the key instances that trigger the bag\nlabel is of great interest because it provides reasons for the decision made by\nthe system. In this paper, we propose a deep convolutional neural network (CNN)\nmodel that addresses the primary task of a bag classification on a WSI and also\nlearns to identify the response of each instance to provide interpretable\nresults to the final prediction. We incorporate the attention mechanism into\nthe proposed model to operate the transformation of instances and learn\nattention weights to allow us to find key patches. To perform a balanced\ntraining, we introduce adaptive weighing in each training bag to explicitly\nadjust the weight distribution in order to concentrate more on the contribution\nof hard samples. Based on the learned attention weights, we further develop a\nsolution to boost the classification performance by generating the bags with\nhard negative instances. We conduct extensive experiments on colon and breast\ncancer histopathology data and show that our framework achieves\nstate-of-the-art performance.', 'Deep learning is very effective at jointly learning feature representations\nand classification models, especially when dealing with high dimensional input\npatterns. Probabilistic logic reasoning, on the other hand, is capable to take\nconsistent and robust decisions in complex environments. The integration of\ndeep learning and logic reasoning is still an open-research problem and it is\nconsidered to be the key for the development of real intelligent agents. This\npaper presents Deep Logic Models, which are deep graphical models integrating\ndeep learning and logic reasoning both for learning and inference. Deep Logic\nModels create an end-to-end differentiable architecture, where deep learners\nare embedded into a network implementing a continuous relaxation of the logic\nknowledge. The learning process allows to jointly learn the weights of the deep\nlearners and the meta-parameters controlling the high-level reasoning. The\nexperimental results show that the proposed methodology overtakes the\nlimitations of the other approaches that have been proposed to bridge deep\nlearning and reasoning.', 'Many seemingly unrelated computer vision tasks can be viewed as a special\ncase of image decomposition into separate layers. For example, image\nsegmentation (separation into foreground and background layers); transparent\nlayer separation (into reflection and transmission layers); Image dehazing\n(separation into a clear image and a haze map), and more. In this paper we\npropose a unified framework for unsupervised layer decomposition of a single\nimage, based on coupled ""Deep-image-Prior"" (DIP) networks. It was shown\n[Ulyanov et al] that the structure of a single DIP generator network is\nsufficient to capture the low-level statistics of a single image. We show that\ncoupling multiple such DIPs provides a powerful tool for decomposing images\ninto their basic components, for a wide variety of applications. This\ncapability stems from the fact that the internal statistics of a mixture of\nlayers is more complex than the statistics of each of its individual\ncomponents. We show the power of this approach for Image-Dehazing, Fg/Bg\nSegmentation, Watermark-Removal, Transparency Separation in images and video,\nand more. These capabilities are achieved in a totally unsupervised way, with\nno training examples other than the input image/video itself.', 'We present an interpretable companion model for any pre-trained black-box\nclassifiers. The idea is that for any input, a user can decide to either\nreceive a prediction from the black-box model, with high accuracy but no\nexplanations, or employ a companion rule to obtain an interpretable prediction\nwith slightly lower accuracy. The companion model is trained from data and the\npredictions of the black-box model, with the objective combining area under the\ntransparency--accuracy curve and model complexity. Our model provides flexible\nchoices for practitioners who face the dilemma of choosing between always using\ninterpretable models and always using black-box models for a predictive task,\nso users can, for any given input, take a step back to resort to an\ninterpretable prediction if they find the predictive performance satisfying, or\nstick to the black-box model if the rules are unsatisfying. To show the value\nof companion models, we design a human evaluation on more than a hundred people\nto investigate the tolerable accuracy loss to gain interpretability for humans.', 'We propose Black Box Explanations through Transparent Approximations (BETA),\na novel model agnostic framework for explaining the behavior of any black-box\nclassifier by simultaneously optimizing for fidelity to the original model and\ninterpretability of the explanation. To this end, we develop a novel objective\nfunction which allows us to learn (with optimality guarantees), a small number\nof compact decision sets each of which explains the behavior of the black box\nmodel in unambiguous, well-defined regions of feature space. Furthermore, our\nframework also is capable of accepting user input when generating these\napproximations, thus allowing users to interactively explore how the black-box\nmodel behaves in different subspaces that are of interest to the user. To the\nbest of our knowledge, this is the first approach which can produce global\nexplanations of the behavior of any given black box model through joint\noptimization of unambiguity, fidelity, and interpretability, while also\nallowing users to explore model behavior based on their preferences.\nExperimental evaluation with real-world datasets and user studies demonstrates\nthat our approach can generate highly compact, easy-to-understand, yet accurate\napproximations of various kinds of predictive models compared to\nstate-of-the-art baselines.', 'Interpretable machine learning (IML) becomes increasingly important in highly\nregulated industry sectors related to the health and safety or fundamental\nrights of human beings. In general, the inherently IML models should be adopted\nbecause of their transparency and explainability, while black-box models with\nmodel-agnostic explainability can be more difficult to defend under regulatory\nscrutiny. For assessing inherent interpretability of a machine learning model,\nwe propose a qualitative template based on feature effects and model\narchitecture constraints. It provides the design principles for\nhigh-performance IML model development, with examples given by reviewing our\nrecent works on ExNN, GAMI-Net, SIMTree, and the Aletheia toolkit for local\nlinear interpretability of deep ReLU networks. We further demonstrate how to\ndesign an interpretable ReLU DNN model with evaluation of conceptual soundness\nfor a real case study of predicting credit default in home lending. We hope\nthat this work will provide a practical guide of developing inherently IML\nmodels in high risk applications in banking industry, as well as other sectors.', 'As the discipline has evolved, research in machine learning has been focused\nmore and more on creating more powerful neural networks, without regard for the\ninterpretability of these networks. Such ""black-box models"" yield\nstate-of-the-art results, but we cannot understand why they make a particular\ndecision or prediction. Sometimes this is acceptable, but often it is not. We\npropose a novel architecture, Regression Networks, which combines the power of\nneural networks with the understandability of regression analysis. While some\nmethods for combining these exist in the literature, our architecture\ngeneralizes these approaches by taking interactions into account, offering the\npower of a dense neural network without forsaking interpretability. We\ndemonstrate that the models exceed the state-of-the-art performance of\ninterpretable models on several benchmark datasets, matching the power of a\ndense neural network. Finally, we discuss how these techniques can be\ngeneralized to other neural architectures, such as convolutional and recurrent\nneural networks.', 'Nonlinear methods such as Deep Neural Networks (DNNs) are the gold standard\nfor various challenging machine learning problems, e.g., image classification,\nnatural language processing or human action recognition. Although these methods\nperform impressively well, they have a significant disadvantage, the lack of\ntransparency, limiting the interpretability of the solution and thus the scope\nof application in practice. Especially DNNs act as black boxes due to their\nmultilayer nonlinear structure. In this paper we introduce a novel methodology\nfor interpreting generic multilayer neural networks by decomposing the network\nclassification decision into contributions of its input elements. Although our\nfocus is on image classification, the method is applicable to a broad set of\ninput data, learning tasks and network architectures. Our method is based on\ndeep Taylor decomposition and efficiently utilizes the structure of the network\nby backpropagating the explanations from the output to the input layer. We\nevaluate the proposed method empirically on the MNIST and ILSVRC data sets.', 'Deep neural networks have been well-known for their superb handling of\nvarious machine learning and artificial intelligence tasks. However, due to\ntheir over-parameterized black-box nature, it is often difficult to understand\nthe prediction results of deep models. In recent years, many interpretation\ntools have been proposed to explain or reveal how deep models make decisions.\nIn this paper, we review this line of research and try to make a comprehensive\nsurvey. Specifically, we first introduce and clarify two basic concepts --\ninterpretations and interpretability -- that people usually get confused about.\nTo address the research efforts in interpretations, we elaborate the designs of\na number of interpretation algorithms, from different perspectives, by\nproposing a new taxonomy. Then, to understand the interpretation results, we\nalso survey the performance metrics for evaluating interpretation algorithms.\nFurther, we summarize the current works in evaluating models\' interpretability\nusing ""trustworthy"" interpretation algorithms. Finally, we review and discuss\nthe connections between deep models\' interpretations and other factors, such as\nadversarial robustness and learning from interpretations, and we introduce\nseveral open-source libraries for interpretation algorithms and evaluation\napproaches.', 'Interpretability and explainability of deep neural networks are challenging\ndue to their scale, complexity, and the agreeable notions on which the\nexplaining process rests. Previous work, in particular, has focused on\nrepresenting internal components of neural networks through human-friendly\nvisuals and concepts. On the other hand, in real life, when making a decision,\nhuman tends to rely on similar situations and/or associations in the past.\nHence arguably, a promising approach to make the model transparent is to design\nit in a way such that the model explicitly connects the current sample with the\nseen ones, and bases its decision on these samples. Grounded on that principle,\nwe propose in this paper an explainable, evidence-based memory network\narchitecture, which learns to summarize the dataset and extract supporting\nevidences to make its decision. Our model achieves state-of-the-art performance\non two popular question answering datasets (i.e. TrecQA and WikiQA). Via\nfurther analysis, we show that this model can reliably trace the errors it has\nmade in the validation step to the training instances that might have caused\nthese errors. We believe that this error-tracing capability provides\nsignificant benefit in improving dataset quality in many applications.', 'Deep Learning algorithms are often used as black box type learning and they\nare too complex to understand. The widespread usability of Deep Learning\nalgorithms to solve various machine learning problems demands deep and\ntransparent understanding of the internal representation as well as decision\nmaking. Moreover, the learning models, trained on sequential data, such as\naudio and video data, have intricate internal reasoning process due to their\ncomplex distribution of features. Thus, a visual simulator might be helpful to\ntrace the internal decision making mechanisms in response to adversarial input\ndata, and it would help to debug and design appropriate deep learning models.\nHowever, interpreting the internal reasoning of deep learning model is not well\nstudied in the literature. In this work, we have developed a visual interactive\nweb application, namely d-DeVIS, which helps to visualize the internal\nreasoning of the learning model which is trained on the audio data. The\nproposed system allows to perceive the behavior as well as to debug the model\nby interactively generating adversarial audio data point. The web application\nof d-DeVIS is available at ddevis.herokuapp.com.', 'Credit assessments activities are essential for financial institutions and\nallow the global economy to grow. Building robust, solid and accurate models\nthat estimate the probability of a default of a company is mandatory for credit\ninsurance companies, moreover when it comes to bridging the trade finance gap.\nAutomating the risk assessment process will allow credit risk experts to reduce\ntheir workload and focus on the critical and complex cases, as well as to\nimprove the loan approval process by reducing the time to process the\napplication. The recent developments in Artificial Intelligence are offering\nnew powerful opportunities. However, most AI techniques are labelled as\nblackbox models due to their lack of explainability. For both users and\nregulators, in order to deploy such technologies at scale, being able to\nunderstand the model logic is a must to grant accurate and ethical decision\nmaking. In this study, we focus on companies credit scoring and we benchmark\ndifferent machine learning models. The aim is to build a model to predict\nwhether a company will experience financial problems in a given time horizon.\nWe address the black box problem using eXplainable Artificial Techniques in\nparticular, post-hoc explanations using SHapley Additive exPlanations. We bring\nlight by providing an expert-aligned feature relevance score highlighting the\ndisagreement between a credit risk expert and a model feature attribution\nexplanation in order to better quantify the convergence towards a better\nhuman-aligned decision making.', 'Machine learning solutions for pattern classification problems are nowadays\nwidely deployed in society and industry. However, the lack of transparency and\naccountability of most accurate models often hinders their safe use. Thus,\nthere is a clear need for developing explainable artificial intelligence\nmechanisms. There exist model-agnostic methods that summarize feature\ncontributions, but their interpretability is limited to predictions made by\nblack-box models. An open challenge is to develop models that have intrinsic\ninterpretability and produce their own explanations, even for classes of models\nthat are traditionally considered black boxes like (recurrent) neural networks.\nIn this paper, we propose a Long-Term Cognitive Network for interpretable\npattern classification of structured data. Our method brings its own mechanism\nfor providing explanations by quantifying the relevance of each feature in the\ndecision process. For supporting the interpretability without affecting the\nperformance, the model incorporates more flexibility through a quasi-nonlinear\nreasoning rule that allows controlling nonlinearity. Besides, we propose a\nrecurrence-aware decision model that evades the issues posed by the unique\nfixed point while introducing a deterministic learning algorithm to compute the\ntunable parameters. The simulations show that our interpretable model obtains\ncompetitive results when compared to state-of-the-art white and black-box\nmodels.', 'Deep learning has been widely used within learning algorithms for robotics.\nOne disadvantage of deep networks is that these networks are black-box\nrepresentations. Therefore, the learned approximations ignore the existing\nknowledge of physics or robotics. Especially for learning dynamics models,\nthese black-box models are not desirable as the underlying principles are well\nunderstood and the standard deep networks can learn dynamics that violate these\nprinciples. To learn dynamics models with deep networks that guarantee\nphysically plausible dynamics, we introduce physics-inspired deep networks that\ncombine first principles from physics with deep learning. We incorporate\nLagrangian mechanics within the model learning such that all approximated\nmodels adhere to the laws of physics and conserve energy. Deep Lagrangian\nNetworks (DeLaN) parametrize the system energy using two networks. The\nparameters are obtained by minimizing the squared residual of the\nEuler-Lagrange differential equation. Therefore, the resulting model does not\nrequire specific knowledge of the individual system, is interpretable, and can\nbe used as a forward, inverse, and energy model. Previously these properties\nwere only obtained when using system identification techniques that require\nknowledge of the kinematic structure. We apply DeLaN to learning dynamics\nmodels and apply these models to control simulated and physical rigid body\nsystems. The results show that the proposed approach obtains dynamics models\nthat can be applied to physical systems for real-time control. Compared to\nstandard deep networks, the physics-inspired models learn better models and\ncapture the underlying structure of the dynamics.', 'Deep neural networks have introduced novel and useful tools to the machine\nlearning community. Other types of classifiers can potentially make use of\nthese tools as well to improve their performance and generality. This paper\nreviews the current state of the art for deep learning classifier technologies\nthat are being used outside of deep neural networks. Non-network classifiers\ncan employ many components found in deep neural network architectures. In this\npaper, we review the feature learning, optimization, and regularization methods\nthat form a core of deep network technologies. We then survey non-neural\nnetwork learning algorithms that make innovative use of these methods to\nimprove classification. Because many opportunities and challenges still exist,\nwe discuss directions that can be pursued to expand the area of deep learning\nfor a variety of classification algorithms.', 'Embeddings are a fundamental component of many modern machine learning and\nnatural language processing models. Understanding them and visualizing them is\nessential for gathering insights about the information they capture and the\nbehavior of the models. State of the art in analyzing embeddings consists in\nprojecting them in two-dimensional planes without any interpretable semantics\nassociated to the axes of the projection, which makes detailed analyses and\ncomparison among multiple sets of embeddings challenging. In this work, we\npropose to use explicit axes defined as algebraic formulae over embeddings to\nproject them into a lower dimensional, but semantically meaningful subspace, as\na simple yet effective analysis and visualization methodology. This methodology\nassigns an interpretable semantics to the measures of variability and the axes\nof visualizations, allowing for both comparisons among different sets of\nembeddings and fine-grained inspection of the embedding spaces. We demonstrate\nthe power of the proposed methodology through a series of case studies that\nmake use of visualizations constructed around the underlying methodology and\nthrough a user study. The results show how the methodology is effective at\nproviding more profound insights than classical projection methods and how it\nis widely applicable to many other use cases.', 'Despite the remarkable success of deep learning in pattern recognition, deep\nnetwork models face the problem of training a large number of parameters. In\nthis paper, we propose and evaluate a novel multi-path wavelet neural network\narchitecture for image classification with far less number of trainable\nparameters. The model architecture consists of a multi-path layout with several\nlevels of wavelet decompositions performed in parallel followed by fully\nconnected layers. These decomposition operations comprise wavelet neurons with\nlearnable parameters, which are updated during the training phase using the\nback-propagation algorithm. We evaluate the performance of the introduced\nnetwork using common image datasets without data augmentation except for SVHN\nand compare the results with influential deep learning models. Our findings\nsupport the possibility of reducing the number of parameters significantly in\ndeep neural networks without compromising its accuracy.', 'The lack of interpretability in current deep learning models causes serious\nconcerns as they are extensively used for various life-critical applications.\nHence, it is of paramount importance to develop interpretable deep learning\nmodels. In this paper, we consider the problem of blind deconvolution and\npropose a novel model-aware deep architecture that allows for the recovery of\nboth the blur kernel and the sharp image from the blurred image. In particular,\nwe propose the Deep Unfolded Richardson-Lucy (Deep-URL) framework -- an\ninterpretable deep-learning architecture that can be seen as an amalgamation of\nclassical estimation technique and deep neural network, and consequently leads\nto improved performance. Our numerical investigations demonstrate significant\nimprovement compared to state-of-the-art algorithms.', 'Deep Learning is a state-of-the-art technique to make inference on extensive\nor complex data. As a black box model due to their multilayer nonlinear\nstructure, Deep Neural Networks are often criticized to be non-transparent and\ntheir predictions not traceable by humans. Furthermore, the models learn from\nartificial datasets, often with bias or contaminated discriminating content.\nThrough their increased distribution, decision-making algorithms can contribute\npromoting prejudge and unfairness which is not easy to notice due to lack of\ntransparency. Hence, scientists developed several so-called explanators or\nexplainers which try to point out the connection between input and output to\nrepresent in a simplified way the inner structure of machine learning black\nboxes. In this survey we differ the mechanisms and properties of explaining\nsystems for Deep Neural Networks for Computer Vision tasks. We give a\ncomprehensive overview about taxonomy of related studies and compare several\nsurvey papers that deal with explainability in general. We work out the\ndrawbacks and gaps and summarize further research ideas.', 'In recent years, the use of sophisticated statistical models that influence\ndecisions in domains of high societal relevance is on the rise. Although these\nmodels can often bring substantial improvements in the accuracy and efficiency\nof organizations, many governments, institutions, and companies are reluctant\nto their adoption as their output is often difficult to explain in\nhuman-interpretable ways. Hence, these models are often regarded as\nblack-boxes, in the sense that their internal mechanisms can be opaque to human\naudit. In real-world applications, particularly in domains where decisions can\nhave a sensitive impact--e.g., criminal justice, estimating credit scores,\ninsurance risk, health risks, etc.--model interpretability is desired.\nRecently, the academic literature has proposed a substantial amount of methods\nfor providing interpretable explanations to machine learning models. This\nsurvey reviews the most relevant and novel methods that form the\nstate-of-the-art for addressing the particular problem of explaining individual\ninstances in machine learning. It seeks to provide a succinct review that can\nguide data science and machine learning practitioners in the search for\nappropriate methods to their problem domain.', 'Interpretability of deep neural networks (DNNs) is essential since it enables\nusers to understand the overall strengths and weaknesses of the models, conveys\nan understanding of how the models will behave in the future, and how to\ndiagnose and correct potential problems. However, it is challenging to reason\nabout what a DNN actually does due to its opaque or black-box nature. To\naddress this issue, we propose a novel technique to improve the\ninterpretability of DNNs by leveraging the rich semantic information embedded\nin human descriptions. By concentrating on the video captioning task, we first\nextract a set of semantically meaningful topics from the human descriptions\nthat cover a wide range of visual concepts, and integrate them into the model\nwith an interpretive loss. We then propose a prediction difference maximization\nalgorithm to interpret the learned features of each neuron. Experimental\nresults demonstrate its effectiveness in video captioning using the\ninterpretable features, which can also be transferred to video action\nrecognition. By clearly understanding the learned features, users can easily\nrevise false predictions via a human-in-the-loop procedure.', ""Machine learning models that first learn a representation of a domain in\nterms of human-understandable concepts, then use it to make predictions, have\nbeen proposed to facilitate interpretation and interaction with models trained\non high-dimensional data. However these methods have important limitations: the\nway they define concepts are not inherently interpretable, and they assume that\nconcept labels either exist for individual instances or can easily be acquired\nfrom users. These limitations are particularly acute for high-dimensional\ntabular features. We propose an approach for learning a set of transparent\nconcept definitions in high-dimensional tabular data that relies on users\nlabeling concept features instead of individual instances. Our method produces\nconcepts that both align with users' intuitive sense of what a concept means,\nand facilitate prediction of the downstream label by a transparent machine\nlearning model. This ensures that the full model is transparent and intuitive,\nand as predictive as possible given this constraint. We demonstrate with\nsimulated user feedback on real prediction problems, including one in a\nclinical domain, that this kind of direct feedback is much more efficient at\nlearning solutions that align with ground truth concept definitions than\nalternative transparent approaches that rely on labeling instances or other\nexisting interaction mechanisms, while maintaining similar predictive\nperformance."", ""The decision-making process of many state-of-the-art machine learning models\nis inherently inscrutable to the extent that it is impossible for a human to\ninterpret the model directly: they are black box models. This has led to a call\nfor research on explaining black box models, for which there are two main\napproaches. Global explanations that aim to explain a model's decision making\nprocess in general, and local explanations that aim to explain a single\nprediction. Since it remains challenging to establish fidelity to black box\nmodels in globally interpretable approximations, much attention is put on local\nexplanations. However, whether local explanations are able to reliably\nrepresent the black box model and provide useful insights remains an open\nquestion. We present Global Aggregations of Local Explanations (GALE) with the\nobjective to provide insights in a model's global decision making process.\nOverall, our results reveal that the choice of aggregation matters. We find\nthat the global importance introduced by Local Interpretable Model-agnostic\nExplanations (LIME) does not reliably represent the model's global behavior.\nOur proposed aggregations are better able to represent how features affect the\nmodel's predictions, and to provide global insights by identifying\ndistinguishing features."", 'Black-box risk scoring models permeate our lives, yet are typically\nproprietary or opaque. We propose Distill-and-Compare, a model distillation and\ncomparison approach to audit such models. To gain insight into black-box\nmodels, we treat them as teachers, training transparent student models to mimic\nthe risk scores assigned by black-box models. We compare the student model\ntrained with distillation to a second un-distilled transparent model trained on\nground-truth outcomes, and use differences between the two models to gain\ninsight into the black-box model. Our approach can be applied in a realistic\nsetting, without probing the black-box model API. We demonstrate the approach\non four public data sets: COMPAS, Stop-and-Frisk, Chicago Police, and Lending\nClub. We also propose a statistical test to determine if a data set is missing\nkey features used to train the black-box model. Our test finds that the\nProPublica data is likely missing key feature(s) used in COMPAS.', ""Fisher information and natural gradient provided deep insights and powerful\ntools to artificial neural networks. However related analysis becomes more and\nmore difficult as the learner's structure turns large and complex. This paper\nmakes a preliminary step towards a new direction. We extract a local component\nof a large neuron system, and defines its relative Fisher information metric\nthat describes accurately this small component, and is invariant to the other\nparts of the system. This concept is important because the geometry structure\nis much simplified and it can be easily applied to guide the learning of neural\nnetworks. We provide an analysis on a list of commonly used components, and\ndemonstrate how to use this concept to further improve optimization."", 'Integrating physics models within machine learning models holds considerable\npromise toward learning robust models with improved interpretability and\nabilities to extrapolate. In this work, we focus on the integration of\nincomplete physics models into deep generative models. In particular, we\nintroduce an architecture of variational autoencoders (VAEs) in which a part of\nthe latent space is grounded by physics. A key technical challenge is to strike\na balance between the incomplete physics and trainable components such as\nneural networks for ensuring that the physics part is used in a meaningful\nmanner. To this end, we propose a regularized learning method that controls the\neffect of the trainable components and preserves the semantics of the\nphysics-based latent variables as intended. We not only demonstrate generative\nperformance improvements over a set of synthetic and real-world datasets, but\nwe also show that we learn robust models that can consistently extrapolate\nbeyond the training distribution in a meaningful manner. Moreover, we show that\nwe can control the generative process in an interpretable manner.', 'The methods of extracting image features are the key to many image processing\ntasks. At present, the most popular method is the deep neural network which can\nautomatically extract robust features through end-to-end training instead of\nhand-crafted feature extraction. However, the deep neural network currently\nfaces many challenges: 1) its effectiveness is heavily dependent on large\ndatasets, so the computational complexity is very high; 2) it is usually\nregarded as a black box model with poor interpretability. To meet the above\nchallenges, a more interpretable and scalable feature learning method, i.e.,\ndeep image feature learning with fuzzy rules (DIFL-FR), is proposed in the\npaper, which combines the rule-based fuzzy modeling technique and the deep\nstacked learning strategy. The method progressively learns image features\nthrough a layer-by-layer manner based on fuzzy rules, so the feature learning\nprocess can be better explained by the generated rules. More importantly, the\nlearning process of the method is only based on forward propagation without\nback propagation and iterative learning, which results in the high learning\nefficiency. In addition, the method is under the settings of unsupervised\nlearning and can be easily extended to scenes of supervised and semi-supervised\nlearning. Extensive experiments are conducted on image datasets of different\nscales. The results obviously show the effectiveness of the proposed method.', ""Interpretable deep learning is a fundamental building block towards safer AI,\nespecially when the deployment possibilities of deep learning-based\ncomputer-aided medical diagnostic systems are so eminent. However, without a\ncomputational formulation of black-box interpretation, general interpretability\nresearch rely heavily on subjective bias. Clear decision structure of the\nmedical diagnostics lets us approximate the decision process of a radiologist\nas a model - removed from subjective bias. We define the process of\ninterpretation as a finite communication between a known model and a black-box\nmodel to optimally map the black box's decision process in the known model.\nConsequently, we define interpretability as maximal information gain over the\ninitial uncertainty about the black-box's decision within finite communication.\nWe relax this definition based on the observation that diagnostic\ninterpretation is typically achieved by a process of minimal querying. We\nderive an algorithm to calculate diagnostic interpretability. The usual\nquestion of accuracy-interpretability tradeoff, i.e. whether a black-box\nmodel's prediction accuracy is dependent on its ability to be interpreted by a\nknown source model, does not arise in this theory. With multiple example\nsimulation experiments of various complexity levels, we demonstrate the working\nof such a theoretical model in synthetic supervised classification scenarios."", 'While deep learning has resulted in major breakthroughs in many application\ndomains, the frameworks commonly used in deep learning remain fragile to\nartificially-crafted and imperceptible changes in the data. In response to this\nfragility, adversarial training has emerged as a principled approach for\nenhancing the robustness of deep learning with respect to norm-bounded\nperturbations. However, there are other sources of fragility for deep learning\nthat are arguably more common and less thoroughly studied. Indeed, natural\nvariation such as lighting or weather conditions can significantly degrade the\naccuracy of trained neural networks, proving that such natural variation\npresents a significant challenge for deep learning.\n  In this paper, we propose a paradigm shift from perturbation-based\nadversarial robustness toward model-based robust deep learning. Our objective\nis to provide general training algorithms that can be used to train deep neural\nnetworks to be robust against natural variation in data. Critical to our\nparadigm is first obtaining a model of natural variation which can be used to\nvary data over a range of natural conditions. Such models may be either known a\npriori or else learned from data. In the latter case, we show that deep\ngenerative models can be used to learn models of natural variation that are\nconsistent with realistic conditions. We then exploit such models in three\nnovel model-based robust training algorithms in order to enhance the robustness\nof deep learning with respect to the given model. Our extensive experiments\nshow that across a variety of naturally-occurring conditions and across various\ndatasets, deep neural networks trained with our model-based algorithms\nsignificantly outperform both standard deep learning algorithms as well as\nnorm-bounded robust deep learning algorithms.', 'A new ensemble framework for interpretable model called Linear Iterative\nFeature Embedding (LIFE) has been developed to achieve high prediction\naccuracy, easy interpretation and efficient computation simultaneously. The\nLIFE algorithm is able to fit a wide single-hidden-layer neural network (NN)\naccurately with three steps: defining the subsets of a dataset by the linear\nprojections of neural nodes, creating the features from multiple narrow\nsingle-hidden-layer NNs trained on the different subsets of the data, combining\nthe features with a linear model. The theoretical rationale behind LIFE is also\nprovided by the connection to the loss ambiguity decomposition of stack\nensemble methods. Both simulation and empirical experiments confirm that LIFE\nconsistently outperforms directly trained single-hidden-layer NNs and also\noutperforms many other benchmark models, including multi-layers Feed Forward\nNeural Network (FFNN), Xgboost, and Random Forest (RF) in many experiments. As\na wide single-hidden-layer NN, LIFE is intrinsically interpretable. Meanwhile,\nboth variable importance and global main and interaction effects can be easily\ncreated and visualized. In addition, the parallel nature of the base learner\nbuilding makes LIFE computationally efficient by leveraging parallel computing.', ""Machine learning models are becoming increasingly popular in different types\nof settings. This is mainly caused by their ability to achieve a level of\npredictive performance that is hard to match by human experts in this new era\nof big data. With this usage growth comes an increase of the requirements for\naccountability and understanding of the models' predictions. However, the\ndegree of sophistication of the most successful models (e.g. ensembles, deep\nlearning) is becoming a large obstacle to this endeavour as these models are\nessentially black boxes. In this paper we describe two general approaches that\ncan be used to provide interpretable descriptions of the expected performance\nof any black box classification model. These approaches are of high practical\nrelevance as they provide means to uncover and describe in an interpretable way\nsituations where the models are expected to have a performance that deviates\nsignificantly from their average behaviour. This may be of critical relevance\nfor applications where costly decisions are driven by the predictions of the\nmodels, as it can be used to warn end users against the usage of the models in\nsome specific cases."", 'There is a rich and growing literature on producing local\ncontrastive/counterfactual explanations for black-box models (e.g. neural\nnetworks).\n  In these methods, for an input, an explanation is in the form of a contrast\npoint differing in very few features from the original input and lying in a\ndifferent class. Other works try to build globally interpretable models like\ndecision trees and rule lists based on the data using actual labels or based on\nthe black-box models predictions. Although these interpretable global models\ncan be useful, they may not be consistent with local explanations from a\nspecific black-box of choice. In this work, we explore the question: Can we\nproduce a transparent global model that is simultaneously accurate and\nconsistent with the local (contrastive) explanations of the black-box model? We\nintroduce a natural local consistency metric that quantifies if the local\nexplanations and predictions of the black-box model are also consistent with\nthe proxy global transparent model. Based on a key insight we propose a novel\nmethod where we create custom boolean features from sparse local contrastive\nexplanations of the black-box model and then train a globally transparent model\non just these, and showcase empirically that such models have higher local\nconsistency compared with other known strategies, while still being close in\nperformance to models that are trained with access to the original data.', 'Deep neural networks have frequently been used to directly learn\nrepresentations useful for a given task from raw input data. In terms of\noverall performance metrics, machine learning solutions employing deep\nrepresentations frequently have been reported to greatly outperform those using\nhand-crafted feature representations. At the same time, they may pick up on\naspects that are predominant in the data, yet not actually meaningful or\ninterpretable. In this paper, we therefore propose a systematic way to test the\ntrustworthiness of deep music representations, considering musical semantics.\nThe underlying assumption is that in case a deep representation is to be\ntrusted, distance consistency between known related points should be maintained\nboth in the input audio space and corresponding latent deep space. We generate\nknown related points through semantically meaningful transformations, both\nconsidering imperceptible and graver transformations. Then, we examine within-\nand between-space distance consistencies, both considering audio space and\nlatent embedded space, the latter either being a result of a conventional\nfeature extractor or a deep encoder. We illustrate how our method, as a\ncomplement to task-specific performance, provides interpretable insight into\nwhat a network may have captured from training data signals.', 'Training deep learning models on mobile devices recently becomes possible,\nbecause of increasing computation power on mobile hardware and the advantages\nof enabling high user experiences. Most of the existing work on machine\nlearning at mobile devices is focused on the inference of deep learning models\n(particularly convolutional neural network and recurrent neural network), but\nnot training. The performance characterization of training deep learning models\non mobile devices is largely unexplored, although understanding the performance\ncharacterization is critical for designing and implementing deep learning\nmodels on mobile devices.\n  In this paper, we perform a variety of experiments on a representative mobile\ndevice (the NVIDIA TX2) to study the performance of training deep learning\nmodels. We introduce a benchmark suite and tools to study performance of\ntraining deep learning models on mobile devices, from the perspectives of\nmemory consumption, hardware utilization, and power consumption. The tools can\ncorrelate performance results with fine-grained operations in deep learning\nmodels, providing capabilities to capture performance variance and problems at\na fine granularity. We reveal interesting performance problems and\nopportunities, including under-utilization of heterogeneous hardware, large\nenergy consumption of the memory, and high predictability of workload\ncharacterization. Based on the performance analysis, we suggest interesting\nresearch directions.', 'Deep neural networks work well at approximating complicated functions when\nprovided with data and trained by gradient descent methods. At the same time,\nthere is a vast amount of existing functions that programmatically solve\ndifferent tasks in a precise manner eliminating the need for training. In many\ncases, it is possible to decompose a task to a series of functions, of which\nfor some we may prefer to use a neural network to learn the functionality,\nwhile for others the preferred method would be to use existing black-box\nfunctions. We propose a method for end-to-end training of a base neural network\nthat integrates calls to existing black-box functions. We do so by\napproximating the black-box functionality with a differentiable neural network\nin a way that drives the base network to comply with the black-box function\ninterface during the end-to-end optimization process. At inference time, we\nreplace the differentiable estimator with its external black-box\nnon-differentiable counterpart such that the base network output matches the\ninput arguments of the black-box function. Using this ""Estimate and Replace""\nparadigm, we train a neural network, end to end, to compute the input to\nblack-box functionality while eliminating the need for intermediate labels. We\nshow that by leveraging the existing precise black-box function during\ninference, the integrated model generalizes better than a fully differentiable\nmodel, and learns more efficiently compared to RL-based methods.', 'Being able to interpret, or explain, the predictions made by a machine\nlearning model is of fundamental importance. This is especially true when there\nis interest in deploying data-driven models to make high-stakes decisions, e.g.\nin healthcare. While recent years have seen an increasing interest in\ninterpretable machine learning research, this field is currently lacking an\nagreed-upon definition of interpretability, and some researchers have called\nfor a more active conversation towards a rigorous approach to interpretability.\nJoining this conversation, we claim in this paper that the difficulty of\ninterpreting a complex model stems from the existing interactions among\nfeatures. We argue that by enforcing monotonicity between features and outputs,\nwe are able to reason about the effect of a single feature on an output\nindependently from other features, and consequently better understand the\nmodel. We show how to structurally introduce this constraint in deep learning\nmodels by adding new simple layers. We validate our model on benchmark\ndatasets, and compare our results with previously proposed interpretable\nmodels.', 'Black box machine learning models are currently being used for high stakes\ndecision-making throughout society, causing problems throughout healthcare,\ncriminal justice, and in other domains. People have hoped that creating methods\nfor explaining these black box models will alleviate some of these problems,\nbut trying to \\textit{explain} black box models, rather than creating models\nthat are \\textit{interpretable} in the first place, is likely to perpetuate bad\npractices and can potentially cause catastrophic harm to society. There is a\nway forward -- it is to design models that are inherently interpretable. This\nmanuscript clarifies the chasm between explaining black boxes and using\ninherently interpretable models, outlines several key reasons why explainable\nblack boxes should be avoided in high-stakes decisions, identifies challenges\nto interpretable machine learning, and provides several example applications\nwhere interpretable models could potentially replace black box models in\ncriminal justice, healthcare, and computer vision.', 'Classical learning theory suggests that the optimal generalization\nperformance of a machine learning model should occur at an intermediate model\ncomplexity, with simpler models exhibiting high bias and more complex models\nexhibiting high variance of the predictive function. However, such a simple\ntrade-off does not adequately describe deep learning models that simultaneously\nattain low bias and variance in the heavily overparameterized regime. A primary\nobstacle in explaining this behavior is that deep learning algorithms typically\ninvolve multiple sources of randomness whose individual contributions are not\nvisible in the total variance. To enable fine-grained analysis, we describe an\ninterpretable, symmetric decomposition of the variance into terms associated\nwith the randomness from sampling, initialization, and the labels. Moreover, we\ncompute the high-dimensional asymptotic behavior of this decomposition for\nrandom feature kernel regression, and analyze the strikingly rich phenomenology\nthat arises. We find that the bias decreases monotonically with the network\nwidth, but the variance terms exhibit non-monotonic behavior and can diverge at\nthe interpolation boundary, even in the absence of label noise. The divergence\nis caused by the \\emph{interaction} between sampling and initialization and can\ntherefore be eliminated by marginalizing over samples (i.e. bagging) \\emph{or}\nover the initial parameters (i.e. ensemble learning).', 'In the last years many accurate decision support systems have been\nconstructed as black boxes, that is as systems that hide their internal logic\nto the user. This lack of explanation constitutes both a practical and an\nethical issue. The literature reports many approaches aimed at overcoming this\ncrucial weakness sometimes at the cost of scarifying accuracy for\ninterpretability. The applications in which black box decision systems can be\nused are various, and each approach is typically developed to provide a\nsolution for a specific problem and, as a consequence, delineating explicitly\nor implicitly its own definition of interpretability and explanation. The aim\nof this paper is to provide a classification of the main problems addressed in\nthe literature with respect to the notion of explanation and the type of black\nbox system. Given a problem definition, a black box type, and a desired\nexplanation this survey should help the researcher to find the proposals more\nuseful for his own work. The proposed classification of approaches to open\nblack box models should also be useful for putting the many research open\nquestions in perspective.', ""Deep Learning based models are currently dominating most state-of-the-art\nsolutions for disease prediction. Existing works employ RNNs along with\nmultiple levels of attention mechanisms to provide interpretability. These deep\nlearning models, with trainable parameters running into millions, require huge\namounts of compute and data to train and deploy. These requirements are\nsometimes so huge that they render usage of such models as unfeasible. We\naddress these challenges by developing a simpler yet interpretable non-deep\nlearning based model for application to EHR data. We model and showcase our\nwork's results on the task of predicting first occurrence of a diagnosis, often\noverlooked in existing works. We push the capabilities of a tree based model\nand come up with a strong baseline for more sophisticated models. Its\nperformance shows an improvement over deep learning based solutions (both, with\nand without the first-occurrence constraint) all the while maintaining\ninterpretability."", 'Deep ensembles excel in large-scale image classification tasks both in terms\nof prediction accuracy and calibration. Despite being simple to train, the\ncomputation and memory cost of deep ensembles limits their practicability.\nWhile some recent works propose to distill an ensemble model into a single\nmodel to reduce such costs, there is still a performance gap between the\nensemble and distilled models. We propose a simple approach for reducing this\ngap, i.e., making the distilled performance close to the full ensemble. Our key\nassumption is that a distilled model should absorb as much function diversity\ninside the ensemble as possible. We first empirically show that the typical\ndistillation procedure does not effectively transfer such diversity, especially\nfor complex models that achieve near-zero training error. To fix this, we\npropose a perturbation strategy for distillation that reveals diversity by\nseeking inputs for which ensemble member outputs disagree. We empirically show\nthat a model distilled with such perturbed samples indeed exhibits enhanced\ndiversity, leading to improved performance.', 'It is known that unsupervised nonlinear dimensionality reduction and\nclustering is sensitive to the selection of hyperparameters, particularly for\ndeep learning based methods, which hinders its practical use. How to select a\nproper network structure that may be dramatically different in different\napplications is a hard issue for deep models, given little prior knowledge of\ndata. In this paper, we aim to automatically determine the optimal network\nstructure of a deep model, named multilayer bootstrap networks (MBN), via\nsimple ensemble learning and selection techniques. Specifically, we first\npropose an MBN ensemble (MBN-E) algorithm which concatenates the sparse outputs\nof a set of MBN base models with different network structures into a new\nrepresentation. Then, we take the new representation produced by MBN-E as a\nreference for selecting the optimal MBN base models. Moreover, we propose a\nfast version of MBN-E (fMBN-E), which is not only theoretically even faster\nthan a single standard MBN but also does not increase the estimation error of\nMBN-E. Importantly, MBN-E and its ensemble selection techniques maintain the\nsimple formulation of MBN that is based on one-nearest-neighbor learning.\nEmpirically, comparing to a number of advanced deep clustering methods and as\nmany as 20 representative unsupervised ensemble learning and selection methods,\nthe proposed methods reach the state-of-the-art performance without manual\nhyperparameter tuning. fMBN-E is empirically even hundreds of times faster than\nMBN-E without suffering performance degradation. The applications to image\nsegmentation and graph data mining further demonstrate the advantage of the\nproposed methods.', 'We study the task of replicating the functionality of black-box neural\nmodels, for which we only know the output class probabilities provided for a\nset of input images. We assume back-propagation through the black-box model is\nnot possible and its training images are not available, e.g. the model could be\nexposed only through an API. In this context, we present a teacher-student\nframework that can distill the black-box (teacher) model into a student model\nwith minimal accuracy loss. To generate useful data samples for training the\nstudent, our framework (i) learns to generate images on a proxy data set (with\nimages and classes different from those used to train the black-box) and (ii)\napplies an evolutionary strategy to make sure that each generated data sample\nexhibits a high response for a specific class when given as input to the black\nbox. Our framework is compared with several baseline and state-of-the-art\nmethods on three benchmark data sets. The empirical evidence indicates that our\nmodel is superior to the considered baselines. Although our method does not\nback-propagate through the black-box network, it generally surpasses\nstate-of-the-art methods that regard the teacher as a glass-box model. Our code\nis available at: https://github.com/antoniobarbalau/black-box-ripper.', 'Deep learning has achieved a great success in many areas, from computer\nvision to natural language processing, to game playing, and much more. Yet,\nwhat deep learning is really doing is still an open question. There are a lot\nof works in this direction. For example, [5] tried to explain deep learning by\ngroup renormalization, and [6] tried to explain deep learning from the view of\nfunctional approximation. In order to address this very crucial question, here\nwe see deep learning from perspective of mechanical learning and learning\nmachine (see [1], [2]). From this particular angle, we can see deep learning\nmuch better and answer with confidence: What deep learning is really doing? why\nit works well, how it works, and how much data is necessary for learning. We\nalso will discuss advantages and disadvantages of deep learning at the end of\nthis work.', 'Learning, taking into account full distribution of the data, referred to as\ngenerative, is not feasible with deep neural networks (DNNs) because they model\nonly the conditional distribution of the outputs given the inputs. Current\nsolutions are either based on joint probability models facing difficult\nestimation problems or learn two separate networks, mapping inputs to outputs\n(recognition) and vice-versa (generation). We propose an intermediate approach.\nFirst, we show that forward computation in DNNs with logistic sigmoid\nactivations corresponds to a simplified approximate Bayesian inference in a\ndirected probabilistic multi-layer model. This connection allows to interpret\nDNN as a probabilistic model of the output and all hidden units given the\ninput. Second, we propose that in order for the recognition and generation\nnetworks to be more consistent with the joint model of the data, weights of the\nrecognition and generator network should be related by transposition. We\ndemonstrate in a tentative experiment that such a coupled pair can be learned\ngeneratively, modelling the full distribution of the data, and has enough\ncapacity to perform well in both recognition and generation.', 'Deep learning has recently become very popular on account of its incredible\nsuccess in many complex data-driven applications, such as image classification\nand speech recognition. The database community has worked on data-driven\napplications for many years, and therefore should be playing a lead role in\nsupporting this new wave. However, databases and deep learning are different in\nterms of both techniques and applications. In this paper, we discuss research\nproblems at the intersection of the two fields. In particular, we discuss\npossible improvements for deep learning systems from a database perspective,\nand analyze database applications that may benefit from deep learning\ntechniques.', 'Deep neural networks are able to solve tasks across a variety of domains and\nmodalities of data. Despite many empirical successes, we lack the ability to\nclearly understand and interpret the learned internal mechanisms that\ncontribute to such effective behaviors or, more critically, failure modes. In\nthis work, we present a general method for visualizing an arbitrary neural\nnetwork\'s inner mechanisms and their power and limitations. Our dataset-centric\nmethod produces visualizations of how a trained network attends to components\nof its inputs. The computed ""attention masks"" support improved interpretability\nby highlighting which input attributes are critical in determining output. We\ndemonstrate the effectiveness of our framework on a variety of deep neural\nnetwork architectures in domains from computer vision, natural language\nprocessing, and reinforcement learning. The primary contribution of our\napproach is an interpretable visualization of attention that provides unique\ninsights into the network\'s underlying decision-making process irrespective of\nthe data modality.', ""With machine learning models being increasingly used to aid decision making\neven in high-stakes domains, there has been a growing interest in developing\ninterpretable models. Although many supposedly interpretable models have been\nproposed, there have been relatively few experimental studies investigating\nwhether these models achieve their intended effects, such as making people more\nclosely follow a model's predictions when it is beneficial for them to do so or\nenabling them to detect when a model has made a mistake. We present a sequence\nof pre-registered experiments (N=3,800) in which we showed participants\nfunctionally identical models that varied only in two factors commonly thought\nto make machine learning models more or less interpretable: the number of\nfeatures and the transparency of the model (i.e., whether the model internals\nare clear or black box). Predictably, participants who saw a clear model with\nfew features could better simulate the model's predictions. However, we did not\nfind that participants more closely followed its predictions. Furthermore,\nshowing participants a clear model meant that they were less able to detect and\ncorrect for the model's sizable mistakes, seemingly due to information\noverload. These counterintuitive findings emphasize the importance of testing\nover intuition when developing interpretable models."", 'We consider the task of solving generic inverse problems, where one wishes to\ndetermine the hidden parameters of a natural system that will give rise to a\nparticular set of measurements. Recently many new approaches based upon deep\nlearning have arisen generating impressive results. We conceptualize these\nmodels as different schemes for efficiently, but randomly, exploring the space\nof possible inverse solutions. As a result, the accuracy of each approach\nshould be evaluated as a function of time rather than a single estimated\nsolution, as is often done now. Using this metric, we compare several\nstate-of-the-art inverse modeling approaches on four benchmark tasks: two\nexisting tasks, one simple task for visualization and one new task from\nmetamaterial design. Finally, inspired by our conception of the inverse\nproblem, we explore a solution that uses a deep learning model to approximate\nthe forward model, and then uses backpropagation to search for good inverse\nsolutions. This approach, termed the neural-adjoint, achieves the best\nperformance in many scenarios.', 'Kernel methods are powerful machine learning techniques which implement\ngeneric non-linear functions to solve complex tasks in a simple way. They Have\na solid mathematical background and exhibit excellent performance in practice.\nHowever, kernel machines are still considered black-box models as the feature\nmapping is not directly accessible and difficult to interpret.The aim of this\nwork is to show that it is indeed possible to interpret the functions learned\nby various kernel methods is intuitive despite their complexity. Specifically,\nwe show that derivatives of these functions have a simple mathematical\nformulation, are easy to compute, and can be applied to many different\nproblems. We note that model function derivatives in kernel machines is\nproportional to the kernel function derivative. We provide the explicit\nanalytic form of the first and second derivatives of the most common kernel\nfunctions with regard to the inputs as well as generic formulas to compute\nhigher order derivatives. We use them to analyze the most used supervised and\nunsupervised kernel learning methods: Gaussian Processes for regression,\nSupport Vector Machines for classification, Kernel Entropy Component Analysis\nfor density estimation, and the Hilbert-Schmidt Independence Criterion for\nestimating the dependency between random variables. For all cases we expressed\nthe derivative of the learned function as a linear combination of the kernel\nfunction derivative. Moreover we provide intuitive explanations through\nillustrative toy examples and show how to improve the interpretation of real\napplications in the context of spatiotemporal Earth system data cubes. This\nwork reflects on the observation that function derivatives may play a crucial\nrole in kernel methods analysis and understanding.', 'Due to the dominant position of deep learning (mostly deep neural networks)\nin various artificial intelligence applications, recently, ensemble learning\nbased on deep neural networks (ensemble deep learning) has shown significant\nperformances in improving the generalization of learning system. However, since\nmodern deep neural networks usually have millions to billions of parameters,\nthe time and space overheads for training multiple base deep learners and\ntesting with the ensemble deep learner are far greater than that of traditional\nensemble learning. Though several algorithms of fast ensemble deep learning\nhave been proposed to promote the deployment of ensemble deep learning in some\napplications, further advances still need to be made for many applications in\nspecific fields, where the developing time and computing resources are usually\nrestricted or the data to be processed is of large dimensionality. An urgent\nproblem needs to be solved is how to take the significant advantages of\nensemble deep learning while reduce the required expenses so that many more\napplications in specific fields can benefit from it. For the alleviation of\nthis problem, it is essential to know about how ensemble learning has developed\nunder the era of deep learning. Thus, in this article, we present fundamental\ndiscussions focusing on data analyses of published works, methodologies, recent\nadvances and unattainability of traditional ensemble learning and ensemble deep\nlearning. We hope this article will be helpful to realize the intrinsic\nproblems and technical challenges faced by future developments of ensemble\nlearning under the era of deep learning.', 'Should we input known genome sequence features or input sequence itself in\ndeep learning framework? As deep learning more popular in various applications,\nresearchers often come to question whether to generate features or use raw\nsequences for deep learning. To answer this question, we study the prediction\naccuracy of precursor miRNA prediction of feature-based deep belief network and\nsequence-based convolution neural network. Tested on a variant of six-layer\nconvolution neural net and three-layer deep belief network, we find the raw\nsequence input based convolution neural network model performs similar or\nslightly better than feature based deep belief networks with best accuracy\nvalues of 0.995 and 0.990, respectively. Both the models outperform existing\nbenchmarks models. The results shows us that if provided large enough data,\nwell devised raw sequence based deep learning models can replace feature based\ndeep learning models. However, construction of well behaved deep learning model\ncan be very challenging. In cased features can be easily extracted,\nfeature-based deep learning models may be a better alternative.', 'Advancements in genomic research such as high-throughput sequencing\ntechniques have driven modern genomic studies into ""big data"" disciplines. This\ndata explosion is constantly challenging conventional methods used in genomics.\nIn parallel with the urgent demand for robust algorithms, deep learning has\nsucceeded in a variety of fields such as vision, speech, and text processing.\nYet genomics entails unique challenges to deep learning since we are expecting\nfrom deep learning a superhuman intelligence that explores beyond our knowledge\nto interpret the genome. A powerful deep learning model should rely on\ninsightful utilization of task-specific knowledge. In this paper, we briefly\ndiscuss the strengths of different deep learning models from a genomic\nperspective so as to fit each particular task with a proper deep architecture,\nand remark on practical considerations of developing modern deep learning\narchitectures for genomics. We also provide a concise review of deep learning\napplications in various aspects of genomic research, as well as pointing out\npotential opportunities and obstacles for future genomics applications.', 'Many end-user evaluations of data visualization techniques have been run\nduring the last decades. Their results are cornerstones to build efficient\nvisualization systems. However, designing such an evaluation is always complex\nand time-consuming and may end in a lack of statistical evidence and\nreproducibility. We believe that modern and efficient computer vision\ntechniques, such as deep convolutional neural networks (CNNs), may help\nvisualization researchers to build and/or adjust their evaluation hypothesis.\nThe basis of our idea is to train machine learning models on several\nvisualization techniques to solve a specific task. Our assumption is that it is\npossible to compare the efficiency of visualization techniques based on the\nperformance of their corresponding model. As current machine learning models\nare not able to strictly reflect human capabilities, including their\nimperfections, such results should be interpreted with caution. However, we\nthink that using machine learning-based pre-evaluation, as a pre-process of\nstandard user evaluations, should help researchers to perform a more exhaustive\nstudy of their design space. Thus, it should improve their final user\nevaluation by providing it better test cases. In this paper, we present the\nresults of two experiments we have conducted to assess how correlated the\nperformance of users and computer vision techniques can be. That study compares\ntwo mainstream graph visualization techniques: node-link (\\NL) and\nadjacency-matrix (\\MD) diagrams. Using two well-known deep convolutional neural\nnetworks, we partially reproduced user evaluations from Ghoniem \\textit{et al.}\nand from Okoe \\textit{et al.}. These experiments showed that some user\nevaluation results can be reproduced automatically.', 'Understanding how a learned black box works is of crucial interest for the\nfuture of Machine Learning. In this paper, we pioneer the question of the\nglobal interpretability of learned black box models that assign numerical\nvalues to symbolic sequential data. To tackle that task, we propose a spectral\nalgorithm for the extraction of weighted automata (WA) from such black boxes.\nThis algorithm does not require the access to a dataset or to the inner\nrepresentation of the black box: the inferred model can be obtained solely by\nquerying the black box, feeding it with inputs and analyzing its outputs.\nExperiments using Recurrent Neural Networks (RNN) trained on a wide collection\nof 48 synthetic datasets and 2 real datasets show that the obtained\napproximation is of great quality.', 'My research lies in the intersection of security and machine learning. This\noverview summarizes one component of my research: combining computer vision\nwith malware exploit detection for enhanced security solutions. I will present\nthe perspectives of efficacy, reliability and resiliency to formulate threat\ndetection as computer vision problems and develop state-of-the-art image-based\nmalware classification. Representing malware binary as images provides a direct\nvisualization of data samples, reduces the efforts for feature extraction, and\nconsumes the whole binary for holistic structural analysis. Employing transfer\nlearning of deep neural networks effective for large scale image classification\nto malware classification demonstrates superior classification efficacy\ncompared with classical machine learning algorithms. To enhance reliability of\nthese vision-based malware detectors, interpretation frameworks can be\nconstructed on the malware visual representations and useful for extracting\nfaithful explanation, so that security practitioners have confidence in the\nmodel before deployment. In cyber-security applications, we should always\nassume that a malware writer constantly modifies code to bypass detection.\nAddressing the resiliency of the malware detectors is equivalently important as\nefficacy and reliability. Via understanding the attack surfaces of machine\nlearning models used for malware detection, we can greatly improve the\nrobustness of the algorithms to combat malware adversaries in the wild. Finally\nI will discuss future research directions worth pursuing in this research\ncommunity.', ""Deep learning models have achieved huge success in numerous fields, such as\ncomputer vision and natural language processing. However, unlike such fields,\nit is hard to apply traditional deep learning models on the graph data due to\nthe 'node-orderless' property. Normally, adjacency matrices will cast an\nartificial and random node-order on the graphs, which renders the performance\nof deep models on graph classification tasks extremely erratic, and the\nrepresentations learned by such models lack clear interpretability. To\neliminate the unnecessary node-order constraint, we propose a novel model named\nIsomorphic Neural Network (IsoNN), which learns the graph representation by\nextracting its isomorphic features via the graph matching between input graph\nand templates. IsoNN has two main components: graph isomorphic feature\nextraction component and classification component. The graph isomorphic feature\nextraction component utilizes a set of subgraph templates as the kernel\nvariables to learn the possible subgraph patterns existing in the input graph\nand then computes the isomorphic features. A set of permutation matrices is\nused in the component to break the node-order brought by the matrix\nrepresentation. Three fully-connected layers are used as the classification\ncomponent in IsoNN. Extensive experiments are conducted on benchmark datasets,\nthe experimental results can demonstrate the effectiveness of ISONN, especially\ncompared with both classic and state-of-the-art graph classification methods."", 'Questions convey information about the questioner, namely what one does not\nknow. In this paper, we propose a novel approach to allow a learning agent to\nask what it considers as tricky to predict, in the course of producing a final\noutput. By analyzing when and what it asks, we can make our model more\ntransparent and interpretable. We first develop this idea to propose a general\nframework of deep neural networks that can ask questions, which we call asking\nnetworks. A specific architecture and training process for an asking network is\nproposed for the task of colorization, which is an exemplar one-to-many task\nand thus a task where asking questions is helpful in performing the task\naccurately. Our results show that the model learns to generate meaningful\nquestions, asks difficult questions first, and utilizes the provided hint more\nefficiently than baseline models. We conclude that the proposed asking\nframework makes the learning agent reveal its weaknesses, which poses a\npromising new direction in developing interpretable and interactive models.', 'Deep learning models are widely used for various industrial and scientific\napplications. Even though these models have achieved considerable success in\nrecent years, there exists a lack of understanding of the rationale behind\ndecisions made by such systems in the machine learning community. This problem\nof interpretability is further aggravated by the increasing complexity of such\nmodels. This paper utilizes concepts from machine learning, quantum computation\nand quantum field theory to demonstrate how a many valued quantum logic system\nnaturally arises in a specific class of generative deep learning models called\nConvolutional Deep Belief Networks. It provides a robust theoretical framework\nfor constructing deep learning models equipped with the interpretability of\nmany valued quantum logic systems without compromising their computing\nefficiency.', 'To achieve seamless human-robot interactions, robots need to intimately\nreason about complex interaction dynamics and future human behaviors within\ntheir motion planning process. However, there is a disconnect between\nstate-of-the-art neural network-based human behavior models and robot motion\nplanners -- either the behavior models are limited in their consideration of\ndownstream planning or a simplified behavior model is used to ensure\ntractability of the planning problem. In this work, we present a framework that\nfuses together the interpretability and flexibility of trajectory optimization\n(TO) with the predictive power of state-of-the-art human trajectory prediction\nmodels. In particular, we leverage gradient information from data-driven\nprediction models to explicitly reason about human-robot interaction dynamics\nwithin a gradient-based TO problem. We demonstrate the efficacy of our approach\nin a multi-agent scenario whereby a robot is required to safely and efficiently\nnavigate through a crowd of up to ten pedestrians. We compare against a variety\nof planning methods, and show that by explicitly accounting for interaction\ndynamics within the planner, our method offers safer and more efficient\nbehaviors, even yielding proactive and nuanced behaviors such as waiting for a\npedestrian to pass before moving.', 'Analog hardware implemented deep learning models are promising for\ncomputation and energy constrained systems such as edge computing devices.\nHowever, the analog nature of the device and the associated many noise sources\nwill cause changes to the value of the weights in the trained deep learning\nmodels deployed on such devices. In this study, systematic evaluation of the\ninference performance of trained popular deep learning models for image\nclassification deployed on analog devices has been carried out, where additive\nwhite Gaussian noise has been added to the weights of the trained models during\ninference. It is observed that deeper models and models with more redundancy in\ndesign such as VGG are more robust to the noise in general. However, the\nperformance is also affected by the design philosophy of the model, the\ndetailed structure of the model, the exact machine learning task, as well as\nthe datasets.', ""Understanding generalization in deep learning is arguably one of the most\nimportant questions in deep learning. Deep learning has been successfully\nadopted to a large number of problems ranging from pattern recognition to\ncomplex decision making, but many recent researchers have raised many concerns\nabout deep learning, among which the most important is generalization. Despite\nnumerous attempts, conventional statistical learning approaches have yet been\nable to provide a satisfactory explanation on why deep learning works. A recent\nline of works aims to address the problem by trying to predict the\ngeneralization performance through complexity measures. In this competition, we\ninvite the community to propose complexity measures that can accurately predict\ngeneralization of models. A robust and general complexity measure would\npotentially lead to a better understanding of deep learning's underlying\nmechanism and behavior of deep models on unseen data, or shed light on better\ngeneralization bounds. All these outcomes will be important for making deep\nlearning more robust and reliable."", 'In recent years, deep neural networks (DNN) have become a highly active area\nof research, and shown remarkable achievements on a variety of computer vision\ntasks. DNNs, however, are known to often make overconfident yet incorrect\npredictions on out-of-distribution samples, which can be a major obstacle to\nreal-world deployments because the training dataset is always limited compared\nto diverse real-world samples. Thus, it is fundamental to provide guarantees of\nrobustness to the distribution shift between training and test time when we\nconstruct DNN models in practice. Moreover, in many cases, the deep learning\nmodels are deployed as black boxes and the performance has been already\noptimized for a training dataset, thus changing the black box itself can lead\nto performance degradation. We here study the robustness to the geometric\ntransformations in a specific condition where the black-box image classifier is\ngiven. We propose an additional learner, \\emph{REinforcement Spatial Transform\nlearner (REST)}, that transforms the warped input data into samples regarded as\nin-distribution by the black-box models. Our work aims to improve the\nrobustness by adding a REST module in front of any black boxes and training\nonly the REST module without retraining the original black box model in an\nend-to-end manner, i.e. we try to convert the real-world data into training\ndistribution which the performance of the black-box model is best suited for.\nWe use a confidence score that is obtained from the black-box model to\ndetermine whether the transformed input is drawn from in-distribution. We\nempirically show that our method has an advantage in generalization to\ngeometric transformations and sample efficiency.', 'Building surrogate models is one common approach when we attempt to learn\nunknown black-box functions. Bayesian optimization provides a framework which\nallows us to build surrogate models based on sequential samples drawn from the\nfunction and find the optimum. Tuning algorithmic parameters to optimize the\nperformance of large, complicated ""black-box"" application codes is a specific\nimportant application, which aims at finding the optima of black-box functions.\nWithin the Bayesian optimization framework, the Gaussian process model produces\nsmooth or continuous sample paths. However, the black-box function in the\ntuning problem is often non-smooth. This difficult tuning problem is worsened\nby the fact that we usually have limited sequential samples from the black-box\nfunction. Motivated by these issues encountered in tuning, we propose a novel\nadditive Gaussian process model called clustered Gaussian process (cGP), where\nthe additive components are induced by clustering. In the examples we studied,\nthe performance can be improved by as much as 90% among repetitive experiments.\nBy using this surrogate model, we want to capture the non-smoothness of the\nblack-box function. In addition to an algorithm for constructing this model, we\nalso apply the model to several artificial and real applications to evaluate\nit.', 'Nowadays, the use of machine learning models is becoming a utility in many\napplications. Companies deliver pre-trained models encapsulated as application\nprogramming interfaces (APIs) that developers combine with third party\ncomponents and their own models and data to create complex data products to\nsolve specific problems. The complexity of such products and the lack of\ncontrol and knowledge of the internals of each component used cause unavoidable\neffects, such as lack of transparency, difficulty in auditability, and\nemergence of potential uncontrolled risks. They are effectively black-boxes.\nAccountability of such solutions is a challenge for the auditors and the\nmachine learning community. In this work, we propose a wrapper that given a\nblack-box model enriches its output prediction with a measure of uncertainty.\nBy using this wrapper, we make the black-box auditable for the accuracy risk\n(risk derived from low quality or uncertain decisions) and at the same time we\nprovide an actionable mechanism to mitigate that risk in the form of decision\nrejection; we can choose not to issue a prediction when the risk or uncertainty\nin that decision is significant. Based on the resulting uncertainty measure, we\nadvocate for a rejection system that selects the more confident predictions,\ndiscarding those more uncertain, leading to an improvement in the trustability\nof the resulting system. We showcase the proposed technique and methodology in\na practical scenario where a simulated sentiment analysis API based on natural\nlanguage processing is applied to different domains. Results demonstrate the\neffectiveness of the uncertainty computed by the wrapper and its high\ncorrelation to bad quality predictions and misclassifications.', 'Recently, the market on deep learning including not only software but also\nhardware is developing rapidly. Big data is collected through IoT devices and\nthe industry world will analyze them to improve their manufacturing process.\nDeep Learning has the hierarchical network architecture to represent the\ncomplicated features of input patterns. Although deep learning can show the\nhigh capability of classification, prediction, and so on, the implementation on\nGPU devices are required. We may meet the trade-off between the higher\nprecision by deep learning and the higher cost with GPU devices. We can success\nthe knowledge extraction from the trained deep learning with high\nclassification capability. The knowledge that can realize faster inference of\npre-trained deep network is extracted as IF-THEN rules from the network signal\nflow given input data. Some experiment results with benchmark tests for time\nseries data sets showed the effectiveness of our proposed method related to the\ncomputational speed.', 'Machine learning models are increasingly used to automate decisions that\naffect humans - deciding who should receive a loan, a job interview, or a\nsocial service. In such applications, a person should have the ability to\nchange the decision of a model. When a person is denied a loan by a credit\nscore, for example, they should be able to alter its input variables in a way\nthat guarantees approval. Otherwise, they will be denied the loan as long as\nthe model is deployed. More importantly, they will lack the ability to\ninfluence a decision that affects their livelihood.\n  In this paper, we frame these issues in terms of recourse, which we define as\nthe ability of a person to change the decision of a model by altering\nactionable input variables (e.g., income vs. age or marital status). We present\ninteger programming tools to ensure recourse in linear classification problems\nwithout interfering in model development. We demonstrate how our tools can\ninform stakeholders through experiments on credit scoring problems. Our results\nshow that recourse can be significantly affected by standard practices in model\ndevelopment, and motivate the need to evaluate recourse in practice.', 'Many works demonstrate that deep learning system is vulnerable to adversarial\nattack. A deep learning system consists of two parts: the deep learning task\nand the deep model. Nowadays, most existing works investigate the impact of the\ndeep model on robustness of deep learning systems, ignoring the impact of the\nlearning task. In this paper, we adopt the binary and interval label encoding\nstrategy to redefine the classification task and design corresponding loss to\nimprove robustness of the deep learning system. Our method can be viewed as\nimproving the robustness of deep learning systems from both the learning task\nand deep model. Experimental results demonstrate that our learning task-aware\nmethod is much more robust than traditional classification while retaining the\naccuracy.', 'We describe an open-source toolkit for neural machine translation (NMT). The\ntoolkit prioritizes efficiency, modularity, and extensibility with the goal of\nsupporting NMT research into model architectures, feature representations, and\nsource modalities, while maintaining competitive performance and reasonable\ntraining requirements. The toolkit consists of modeling and translation\nsupport, as well as detailed pedagogical documentation about the underlying\ntechniques.', ""Deep neural networks have achieved state of the art accuracy at classifying\nmolecules with respect to whether they bind to specific protein targets. A key\nbreakthrough would occur if these models could reveal the fragment\npharmacophores that are causally involved in binding. Extracting chemical\ndetails of binding from the networks could potentially lead to scientific\ndiscoveries about the mechanisms of drug actions. But doing so requires shining\nlight into the black box that is the trained neural network model, a task that\nhas proved difficult across many domains. Here we show how the binding\nmechanism learned by deep neural network models can be interrogated, using a\nrecently described attribution method. We first work with carefully constructed\nsynthetic datasets, in which the 'fragment logic' of binding is fully known. We\nfind that networks that achieve perfect accuracy on held out test datasets\nstill learn spurious correlations due to biases in the datasets, and we are\nable to exploit this non-robustness to construct adversarial examples that fool\nthe model. The dataset bias makes these models unreliable for accurately\nrevealing information about the mechanisms of protein-ligand binding. In light\nof our findings, we prescribe a test that checks for dataset bias given a\nhypothesis. If the test fails, it indicates that either the model must be\nsimplified or regularized and/or that the training dataset requires\naugmentation."", 'Which components of the singular value decomposition of a signal-plus-noise\ndata matrix are most informative for the inferential task of detecting or\nestimating an embedded low-rank signal matrix? Principal component analysis\nascribes greater importance to the components that capture the greatest\nvariation, i.e., the singular vectors associated with the largest singular\nvalues. This choice is often justified by invoking the Eckart-Young theorem\neven though that work addresses the problem of how to best represent a\nsignal-plus-noise matrix using a low-rank approximation and not how to\nbest_infer_ the underlying low-rank signal component.\n  Here we take a first-principles approach in which we start with a\nsignal-plus-noise data matrix and show how the spectrum of the noise-only\ncomponent governs whether the principal or the middle components of the\nsingular value decomposition of the data matrix will be the informative\ncomponents for inference. Simply put, if the noise spectrum is supported on a\nconnected interval, in a sense we make precise, then the use of the principal\ncomponents is justified. When the noise spectrum is supported on multiple\nintervals, then the middle components might be more informative than the\nprincipal components.\n  The end result is a proper justification of the use of principal components\nin the setting where the noise matrix is i.i.d. Gaussian and the identification\nof scenarios, generically involving heterogeneous noise models such as mixtures\nof Gaussians, where the middle components might be more informative than the\nprincipal components so that they may be exploited to extract additional\nprocessing gain. Our results show how the blind use of principal components can\nlead to suboptimal or even faulty inference because of phase transitions that\nseparate a regime where the principal components are informative from a regime\nwhere they are uninformative.', 'Gene expression profiles have been widely used to characterize patterns of\ncellular responses to diseases. As data becomes available, scalable learning\ntoolkits become essential to processing large datasets using deep learning\nmodels to model complex biological processes. We present an autoencoder to\ncapture nonlinear relationships recovered from gene expression profiles. The\nautoencoder is a nonlinear dimension reduction technique using an artificial\nneural network, which learns hidden representations of unlabeled data. We train\nthe autoencoder on a large collection of tumor samples from the National Cancer\nInstitute Genomic Data Commons, and obtain a generalized and unsupervised\nlatent representation. We leverage a HPC-focused deep learning toolkit,\nLivermore Big Artificial Neural Network (LBANN) to efficiently parallelize the\ntraining algorithm, reducing computation times from several hours to a few\nminutes. With the trained autoencoder, we generate latent representations of a\nsmall dataset, containing pairs of normal and cancer cells of various tumor\ntypes. A novel measure called autoencoder node saliency (ANS) is introduced to\nidentify the hidden nodes that best differentiate various pairs of cells. We\ncompare our findings of the best classifying nodes with principal component\nanalysis and the visualization of t-distributed stochastic neighbor embedding.\nWe demonstrate that the autoencoder effectively extracts distinct gene features\nfor multiple learning tasks in the dataset.', 'Sparse Principal Component Analysis (sPCA) is a popular matrix factorization\napproach based on Principal Component Analysis (PCA) that combines variance\nmaximization and sparsity with the ultimate goal of improving data\ninterpretation. When moving from PCA to sPCA, there are a number of\nimplications that the practitioner needs to be aware of. A relevant one is that\nscores and loadings in sPCA may not be orthogonal. For this reason, the\ntraditional way of computing scores, residuals and variance explained that is\nused in the classical PCA cannot directly be applied to sPCA models. This also\naffects how sPCA components should be visualized. In this paper we illustrate\nthis problem both theoretically and numerically using simulations for several\nstate-of-the-art sPCA algorithms, and provide proper computation of the\ndifferent elements mentioned. We show that sPCA approaches present disparate\nand limited performance when modeling noise-free, sparse data. In a follow-up\npaper, we discuss the theoretical properties that lead to this problem.', 'Probabilistic deep learning is deep learning that accounts for uncertainty,\nboth model uncertainty and data uncertainty. It is based on the use of\nprobabilistic models and deep neural networks. We distinguish two approaches to\nprobabilistic deep learning: probabilistic neural networks and deep\nprobabilistic models. The former employs deep neural networks that utilize\nprobabilistic layers which can represent and process uncertainty; the latter\nuses probabilistic models that incorporate deep neural network components which\ncapture complex non-linear stochastic relationships between the random\nvariables. We discuss some major examples of each approach including Bayesian\nneural networks and mixture density networks (for probabilistic neural\nnetworks), and variational autoencoders, deep Gaussian processes and deep mixed\neffects models (for deep probabilistic models). TensorFlow Probability is a\nlibrary for probabilistic modeling and inference which can be used for both\napproaches of probabilistic deep learning. We include its code examples for\nillustration.', ""The thin-file borrowers are customers for whom a creditworthiness assessment\nis uncertain due to their lack of credit history; many researchers have used\nborrowers' relationships and interactions networks in the form of graphs as an\nalternative data source to address this. Incorporating network data is\ntraditionally made by hand-crafted feature engineering, and lately, the graph\nneural network has emerged as an alternative, but it still does not improve\nover the traditional method's performance. Here we introduce a framework to\nimprove credit scoring models by blending several Graph Representation Learning\nmethods: feature engineering, graph embeddings, and graph neural networks. We\nstacked their outputs to produce a single score in this approach. We validated\nthis framework using a unique multi-source dataset that characterizes the\nrelationships and credit history for the entire population of a Latin American\ncountry, applying it to credit risk models, application, and behavior,\ntargeting both individuals and companies.\n  Our results show that the graph representation learning methods should be\nused as complements, and these should not be seen as self-sufficient methods as\nis currently done. In terms of AUC and KS, we enhance the statistical\nperformance, outperforming traditional methods.\n  In Corporate lending, where the gain is much higher, it confirms that\nevaluating an unbanked company cannot solely consider its features. The\nbusiness ecosystem where these firms interact with their owners, suppliers,\ncustomers, and other companies provides novel knowledge that enables financial\ninstitutions to enhance their creditworthiness assessment.\n  Our results let us know when and which group to use graph data and what\neffects on performance to expect. They also show the enormous value of graph\ndata on the unbanked credit scoring problem, principally to help companies'\nbanking."", 'In this paper, we present a statistical-mechanical analysis of deep learning.\nWe elucidate some of the essential components of deep learning---pre-training\nby unsupervised learning and fine tuning by supervised learning. We formulate\nthe extraction of features from the training data as a margin criterion in a\nhigh-dimensional feature-vector space. The self-organized classifier is then\nsupplied with small amounts of labelled data, as in deep learning. Although we\nemploy a simple single-layer perceptron model, rather than directly analyzing a\nmulti-layer neural network, we find a nontrivial phase transition that is\ndependent on the number of unlabelled data in the generalization error of the\nresultant classifier. In this sense, we evaluate the efficacy of the\nunsupervised learning component of deep learning. The analysis is performed by\nthe replica method, which is a sophisticated tool in statistical mechanics. We\nvalidate our result in the manner of deep learning, using a simple iterative\nalgorithm to learn the weight vector on the basis of belief propagation.', 'The great success of deep learning shows that its technology contains\nprofound truth, and understanding its internal mechanism not only has important\nimplications for the development of its technology and effective application in\nvarious fields, but also provides meaningful insights into the understanding of\nhuman brain mechanism. At present, most of the theoretical research on deep\nlearning is based on mathematics. This dissertation proposes that the neural\nnetwork of deep learning is a physical system, examines deep learning from\nthree different perspectives: microscopic, macroscopic, and physical world\nviews, answers multiple theoretical puzzles in deep learning by using physics\nprinciples. For example, from the perspective of quantum mechanics and\nstatistical physics, this dissertation presents the calculation methods for\nconvolution calculation, pooling, normalization, and Restricted Boltzmann\nMachine, as well as the selection of cost functions, explains why deep learning\nmust be deep, what characteristics are learned in deep learning, why\nConvolutional Neural Networks do not have to be trained layer by layer, and the\nlimitations of deep learning, etc., and proposes the theoretical direction and\nbasis for the further development of deep learning now and in the future. The\nbrilliance of physics flashes in deep learning, we try to establish the deep\nlearning technology based on the scientific theory of physics.', 'With the growing popularity of deep-learning based NLP models, comes a need\nfor interpretable systems. But what is interpretability, and what constitutes a\nhigh-quality interpretation? In this opinion piece we reflect on the current\nstate of interpretability evaluation research. We call for more clearly\ndifferentiating between different desired criteria an interpretation should\nsatisfy, and focus on the faithfulness criteria. We survey the literature with\nrespect to faithfulness evaluation, and arrange the current approaches around\nthree assumptions, providing an explicit form to how faithfulness is ""defined""\nby the community. We provide concrete guidelines on how evaluation of\ninterpretation methods should and should not be conducted. Finally, we claim\nthat the current binary definition for faithfulness sets a potentially\nunrealistic bar for being considered faithful. We call for discarding the\nbinary notion of faithfulness in favor of a more graded one, which we believe\nwill be of greater practical utility.', 'Social network analysis is the process of investigating social structures\nthrough the use of networks and graph theory. It combines a variety of\ntechniques for analyzing the structure of social networks as well as theories\nthat aim at explaining the underlying dynamics and patterns observed in these\nstructures. It is an inherently interdisciplinary field which originally\nemerged from the fields of social psychology, statistics and graph theory. This\ntalk will covers the theory of social network analysis, with a short\nintroduction to graph theory and information spread. Then we will deep dive\ninto Python code with NetworkX to get a better understanding of the network\ncomponents, followed-up by constructing and implying social networks from real\nPandas and textual datasets. Finally we will go over code examples of practical\nuse-cases such as visualization with matplotlib, social-centrality analysis and\ninfluence maximization for information spread.', 'In this study, we define interaction components of different orders between\ntwo input variables based on game theory. We further prove that interaction\ncomponents of different orders satisfy several desirable properties.', ""Deep neural networks have achieved impressive results on a wide variety of\ntasks. However, quantifying uncertainty in the network's output is a\nchallenging task. Bayesian models offer a mathematical framework to reason\nabout model uncertainty. Variational methods have been used for approximating\nintractable integrals that arise in Bayesian inference for neural networks. In\nthis report, we review the major variational inference concepts pertinent to\nBayesian neural networks and compare various approximation methods used in\nliterature. We also talk about the applications of variational bayes in\nReinforcement learning and continual learning."", 'Deep learning models are increasingly used in mobile applications as critical\ncomponents. Unlike the program bytecode whose vulnerabilities and threats have\nbeen widely-discussed, whether and how the deep learning models deployed in the\napplications can be compromised are not well-understood since neural networks\nare usually viewed as a black box. In this paper, we introduce a highly\npractical backdoor attack achieved with a set of reverse-engineering techniques\nover compiled deep learning models. The core of the attack is a neural\nconditional branch constructed with a trigger detector and several operators\nand injected into the victim model as a malicious payload. The attack is\neffective as the conditional logic can be flexibly customized by the attacker,\nand scalable as it does not require any prior knowledge from the original\nmodel. We evaluated the attack effectiveness using 5 state-of-the-art deep\nlearning models and real-world samples collected from 30 users. The results\ndemonstrated that the injected backdoor can be triggered with a success rate of\n93.5%, while only brought less than 2ms latency overhead and no more than 1.4%\naccuracy decrease. We further conducted an empirical study on real-world mobile\ndeep learning apps collected from Google Play. We found 54 apps that were\nvulnerable to our attack, including popular and security-critical ones. The\nresults call for the awareness of deep learning application developers and\nauditors to enhance the protection of deployed models.', 'We offer a general theoretical framework for brain and behavior that is\nevolutionarily and computationally plausible. The brain in our abstract model\nis a network of nodes and edges. Although it has some similarities to standard\nneural network models, as we show, there are some significant differences. Both\nnodes and edges in our network have weights and activation levels. They act as\nprobabilistic transducers that use a set of relatively simple rules to\ndetermine how activation levels and weights are affected by input, generate\noutput, and affect each other. We show that these simple rules enable a\nlearning process that allows the network to represent increasingly complex\nknowledge, and simultaneously to act as a computing device that facilitates\nplanning, decision-making, and the execution of behavior. By specifying the\ninnate (genetic) components of the network, we show how evolution could endow\nthe network with initial adaptive rules and goals that are then enriched\nthrough learning. We demonstrate how the developing structure of the network\n(which determines what the brain can do and how well) is critically affected by\nthe co-evolved coordination between the mechanisms affecting the distribution\nof data input and those determining the learning parameters (used in the\nprograms run by nodes and edges). Finally, we consider how the model accounts\nfor various findings in the field of learning and decision making, how it can\naddress some challenging problems in mind and behavior, such as those related\nto setting goals and self-control, and how it can help understand some\ncognitive disorders.', 'The transfer learning toolkit wraps the codes of 17 transfer learning models\nand provides integrated interfaces, allowing users to use those models by\ncalling a simple function. It is easy for primary researchers to use this\ntoolkit and to choose proper models for real-world applications. The toolkit is\nwritten in Python and distributed under MIT open source license. In this paper,\nthe current state of this toolkit is described and the necessary environment\nsetting and usage are introduced.', 'This paper is a review of the evolutionary history of deep learning models.\nIt covers from the genesis of neural networks when associationism modeling of\nthe brain is studied, to the models that dominate the last decade of research\nin deep learning like convolutional neural networks, deep belief networks, and\nrecurrent neural networks. In addition to a review of these models, this paper\nprimarily focuses on the precedents of the models above, examining how the\ninitial ideas are assembled to construct the early models and how these\npreliminary models are developed into their current forms. Many of these\nevolutionary paths last more than half a century and have a diversity of\ndirections. For example, CNN is built on prior knowledge of biological vision\nsystem; DBN is evolved from a trade-off of modeling power and computation\ncomplexity of graphical models and many nowadays models are neural counterparts\nof ancient linear models. This paper reviews these evolutionary paths and\noffers a concise thought flow of how these models are developed, and aims to\nprovide a thorough background for deep learning. More importantly, along with\nthe path, this paper summarizes the gist behind these milestones and proposes\nmany directions to guide the future research of deep learning.', 'Recent advances in artificial intelligence and the increasing need for\npowerful defensive measures in the domain of network security, have led to the\nadoption of deep learning approaches for use in network intrusion detection\nsystems. These methods have achieved superior performance against conventional\nnetwork attacks, which enable the deployment of practical security systems to\nunique and dynamic sectors. Adversarial machine learning, unfortunately, has\nrecently shown that deep learning models are inherently vulnerable to\nadversarial modifications on their input data. Because of this susceptibility,\nthe deep learning models deployed to power a network defense could in fact be\nthe weakest entry point for compromising a network system. In this paper, we\nshow that by modifying on average as little as 1.38 of the input features, an\nadversary can generate malicious inputs which effectively fool a deep learning\nbased NIDS. Therefore, when designing such systems, it is crucial to consider\nthe performance from not only the conventional network security perspective but\nalso the adversarial machine learning domain.', ""What should I work on first? What can wait until later? Which projects should\nI prioritize and which tasks are not worth my time? These are challenging\nquestions that many people face every day. People's intuitive strategy is to\nprioritize their immediate experience over the long-term consequences. This\nleads to procrastination and the neglect of important long-term projects in\nfavor of seemingly urgent tasks that are less important. Optimal gamification\nstrives to help people overcome these problems by incentivizing each task by a\nnumber of points that communicates how valuable it is in the long-run.\nUnfortunately, computing the optimal number of points with standard dynamic\nprogramming methods quickly becomes intractable as the number of a person's\nprojects and the number of tasks required by each project increase. Here, we\nintroduce and evaluate a scalable method for identifying which tasks are most\nimportant in the long run and incentivizing each task according to its\nlong-term value. Our method makes it possible to create to-do list gamification\napps that can handle the size and complexity of people's to-do lists in the\nreal world."", ""Probability theory, epistemically interpreted, provides an excellent, if not\nthe best available account of inductive reasoning. This is so because there are\ngeneral and definite rules for the change of subjective probabilities through\ninformation or experience; induction and belief change are one and same topic,\nafter all. The most basic of these rules is simply to conditionalize with\nrespect to the information received; and there are similar and more general\nrules. 1 Hence, a fundamental reason for the epistemological success of\nprobability theory is that there at all exists a well-behaved concept of\nconditional probability. Still, people have, and have reasons for, various\nconcerns over probability theory. One of these is my starting point:\nIntuitively, we have the notion of plain belief; we believe propositions2 to be\ntrue (or to be false or neither). Probability theory, however, offers no formal\ncounterpart to this notion. Believing A is not the same as having probability 1\nfor A, because probability 1 is incorrigible3; but plain belief is clearly\ncorrigible. And believing A is not the same as giving A a probability larger\nthan some 1 - c, because believing A and believing B is usually taken to be\nequivalent to believing A & B.4 Thus, it seems that the formal representation\nof plain belief has to take a non-probabilistic route. Indeed, representing\nplain belief seems easy enough: simply represent an epistemic state by the set\nof all propositions believed true in it or, since I make the common assumption\nthat plain belief is deductively closed, by the conjunction of all propositions\nbelieved true in it. But this does not yet provide a theory of induction, i.e.\nan answer to the question how epistemic states so represented are changed\ntbrough information or experience. There is a convincing partial answer: if the\nnew information is compatible with the old epistemic state, then the new\nepistemic state is simply represented by the conjunction of the new information\nand the old beliefs. This answer is partial because it does not cover the quite\ncommon case where the new information is incompatible with the old beliefs. It\nis, however, important to complete the answer and to cover this case, too;\notherwise, we would not represent plain belief as conigible. The crucial\nproblem is that there is no good completion. When epistemic states are\nrepresented simply by the conjunction of all propositions believed true in it,\nthe answer cannot be completed; and though there is a lot of fruitful work, no\nother representation of epistemic states has been proposed, as far as I know,\nwhich provides a complete solution to this problem. In this paper, I want to\nsuggest such a solution. In [4], I have more fully argued that this is the only\nsolution, if certain plausible desiderata are to be satisfied. Here, in section\n2, I will be content with formally defining and intuitively explaining my\nproposal. I will compare my proposal with probability theory in section 3. It\nwill turn out that the theory I am proposing is structurally homomorphic to\nprobability theory in important respects and that it is thus equally easily\nimplementable, but moreover computationally simpler. Section 4 contains a very\nbrief comparison with various kinds of logics, in particular conditional logic,\nwith Shackle's functions of potential surprise and related theories, and with\nthe Dempster - Shafer theory of belief functions."", ""There is knowledge. There is belief. And there is tacit agreement.' 'We may\ntalk about objects. We may talk about attributes of the objects. Or we may talk\nboth about objects and their attributes.' This work inspects tacit agreements\non assumptions about the relation between objects and their attributes, and\nstudies a way of expressing them, presenting as the result what we term gradual\nlogic in which the sense of truth gradually shifts. It extends classical logic\ninstances with a new logical connective capturing the object-attribute\nrelation. A formal semantics is presented. Decidability is proved. Para-\nconsistent/epistemic/conditional/intensional/description/combined logics are\ncompared."", 'We argue that logical semantics might have faltered due to its failure in\ndistinguishing between two fundamentally very different types of concepts:\nontological concepts, that should be types in a strongly-typed ontology, and\nlogical concepts, that are predicates corresponding to properties of and\nrelations between objects of various ontological types. We will then show that\naccounting for these differences amounts to the integration of lexical and\ncompositional semantics in one coherent framework, and to an embedding in our\nlogical semantics of a strongly-typed ontology that reflects our commonsense\nview of the world and the way we talk about it in ordinary language. We will\nshow that in such a framework a number of challenges in natural language\nsemantics can be adequately and systematically treated.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I aim to develop a neural machine translation (NMT) model capable of effectively translating between multiple languages, each with unique morpheme usage. I wish to address several limitations of traditional NMT models. Firstly, my model should not rely on a fixed vocabulary but rather be capable of processing, embedding, and understanding new, unseen words. To achieve this, I plan to implement an intelligent subword tokenization scheme, which should enable my model to surpass the performance of basic character-level NMT. Moreover, the model needs to be parameter-efficient to facilitate deployment on compact wearable devices. It should also be capable of managing long-distance context and grammar dependencies within paragraphs. My current approach involves using a form of hierarchical decoding. This method initially translates larger and longer word or phrase-level information swiftly, then fine-tunes subword or even character-level information with greater precision. My model needs to reach state-of-the-art performance on standard machine translation datasets.","['In neural machine translation (NMT), researchers face the challenge of\nun-seen (or out-of-vocabulary OOV) words translation. To solve this, some\nresearchers propose the splitting of western languages such as English and\nGerman into sub-words or compounds. In this paper, we try to address this OOV\nissue and improve the NMT adequacy with a harder language Chinese whose\ncharacters are even more sophisticated in composition. We integrate the Chinese\nradicals into the NMT model with different settings to address the unseen words\nchallenge in Chinese to English translation. On the other hand, this also can\nbe considered as semantic part of the MT system since the Chinese radicals\nusually carry the essential meaning of the words they are constructed in.\nMeaningful radicals and new characters can be integrated into the NMT systems\nwith our models. We use an attention-based NMT system as a strong baseline\nsystem. The experiments on standard Chinese-to-English NIST translation shared\ntask data 2006 and 2008 show that our designed models outperform the baseline\nmodel in a wide range of state-of-the-art evaluation metrics including LEPOR,\nBEER, and CharacTER, in addition to BLEU and NIST scores, especially on the\nadequacy-level translation. We also have some interesting findings from the\nresults of our various experiment settings about the performance of words and\ncharacters in Chinese NMT, which is different with other languages. For\ninstance, the fully character level NMT may perform well or the state of the\nart in some other languages as researchers demonstrated recently, however, in\nthe Chinese NMT model, word boundary knowledge is important for the model\nlearning.', 'Neural machine translation (NMT), a new approach to machine translation, has\nbeen proved to outperform conventional statistical machine translation (SMT)\nacross a variety of language pairs. Translation is an open-vocabulary problem,\nbut most existing NMT systems operate with a fixed vocabulary, which causes the\nincapability of translating rare words. This problem can be alleviated by using\ndifferent translation granularities, such as character, subword and hybrid\nword-character. Translation involving Chinese is one of the most difficult\ntasks in machine translation, however, to the best of our knowledge, there has\nnot been any other work exploring which translation granularity is most\nsuitable for Chinese in NMT. In this paper, we conduct an extensive comparison\nusing Chinese-English NMT as a case study. Furthermore, we discuss the\nadvantages and disadvantages of various translation granularities in detail.\nOur experiments show that subword model performs best for Chinese-to-English\ntranslation with the vocabulary which is not so big while hybrid word-character\nmodel is most suitable for English-to-Chinese translation. Moreover,\nexperiments of different granularities show that Hybrid_BPE method can achieve\nbest result on Chinese-to-English translation task.', ""The existing machine translation systems, whether phrase-based or neural,\nhave relied almost exclusively on word-level modelling with explicit\nsegmentation. In this paper, we ask a fundamental question: can neural machine\ntranslation generate a character sequence without any explicit segmentation? To\nanswer this question, we evaluate an attention-based encoder-decoder with a\nsubword-level encoder and a character-level decoder on four language\npairs--En-Cs, En-De, En-Ru and En-Fi-- using the parallel corpora from WMT'15.\nOur experiments show that the models with a character-level decoder outperform\nthe ones with a subword-level decoder on all of the four language pairs.\nFurthermore, the ensembles of neural models with a character-level decoder\noutperform the state-of-the-art non-neural machine translation systems on\nEn-Cs, En-De and En-Fi and perform comparably on En-Ru.""]","['Neural machine translation (NMT), a new approach to machine translation, has\nachieved promising results comparable to those of traditional approaches such\nas statistical machine translation (SMT). Despite its recent success, NMT\ncannot handle a larger vocabulary because training complexity and decoding\ncomplexity proportionally increase with the number of target words. This\nproblem becomes even more serious when translating patent documents, which\ncontain many technical terms that are observed infrequently. In NMTs, words\nthat are out of vocabulary are represented by a single unknown token. In this\npaper, we propose a method that enables NMT to translate patent sentences\ncomprising a large vocabulary of technical terms. We train an NMT system on\nbilingual data wherein technical terms are replaced with technical term tokens;\nthis allows it to translate most of the source sentences except technical\nterms. Further, we use it as a decoder to translate source sentences with\ntechnical term tokens and replace the tokens with technical term translations\nusing SMT. We also use it to rerank the 1,000-best SMT translations on the\nbasis of the average of the SMT score and that of the NMT rescoring of the\ntranslated sentences with technical term tokens. Our experiments on\nJapanese-Chinese patent sentences show that the proposed NMT system achieves a\nsubstantial improvement of up to 3.1 BLEU points and 2.3 RIBES points over\ntraditional SMT systems and an improvement of approximately 0.6 BLEU points and\n0.8 RIBES points over an equivalent NMT system without our proposed technique.', 'Neural machine translation (NMT) models typically operate with a fixed\nvocabulary, but translation is an open-vocabulary problem. Previous work\naddresses the translation of out-of-vocabulary words by backing off to a\ndictionary. In this paper, we introduce a simpler and more effective approach,\nmaking the NMT model capable of open-vocabulary translation by encoding rare\nand unknown words as sequences of subword units. This is based on the intuition\nthat various word classes are translatable via smaller units than words, for\ninstance names (via character copying or transliteration), compounds (via\ncompositional translation), and cognates and loanwords (via phonological and\nmorphological transformations). We discuss the suitability of different word\nsegmentation techniques, including simple character n-gram models and a\nsegmentation based on the byte pair encoding compression algorithm, and\nempirically show that subword models improve over a back-off dictionary\nbaseline for the WMT 15 translation tasks English-German and English-Russian by\n1.1 and 1.3 BLEU, respectively.', 'The non-autoregressive models have boosted the efficiency of neural machine\ntranslation through parallelized decoding at the cost of effectiveness when\ncomparing with the autoregressive counterparts. In this paper, we claim that\nthe syntactic and semantic structures among natural language are critical for\nnon-autoregressive machine translation and can further improve the performance.\nHowever, these structures are rarely considered in the existing\nnon-autoregressive models. Inspired by this intuition, we propose to\nincorporate the explicit syntactic and semantic structures of languages into a\nnon-autoregressive Transformer, for the task of neural machine translation.\nMoreover, we also consider the intermediate latent alignment within target\nsentences to better learn the long-term token dependencies. Experimental\nresults on two real-world datasets (i.e., WMT14 En-De and WMT16 En-Ro) show\nthat our model achieves a significantly faster speed, as well as keeps the\ntranslation quality when compared with several state-of-the-art\nnon-autoregressive models.', 'Neural machine translation (NMT) is one of the best methods for understanding\nthe differences in semantic rules between two languages. Especially for\nIndo-European languages, subword-level models have achieved impressive results.\nHowever, when the translation task involves Chinese, semantic granularity\nremains at the word and character level, so there is still need more\nfine-grained translation model of Chinese. In this paper, we introduce a simple\nand effective method for Chinese translation at the sub-character level. Our\napproach uses the Wubi method to translate Chinese into English; byte-pair\nencoding (BPE) is then applied. Our method for Chinese-English translation\neliminates the need for a complicated word segmentation algorithm during\npreprocessing. Furthermore, our method allows for sub-character-level neural\ntranslation based on recurrent neural network (RNN) architecture, without\npreprocessing. The empirical results show that for Chinese-English translation\ntasks, our sub-character-level model has a comparable BLEU score to the subword\nmodel, despite having a much smaller vocabulary. Additionally, the small\nvocabulary is highly advantageous for NMT model compression.', 'Translation into morphologically-rich languages challenges neural machine\ntranslation (NMT) models with extremely sparse vocabularies where atomic\ntreatment of surface forms is unrealistic. This problem is typically addressed\nby either pre-processing words into subword units or performing translation\ndirectly at the level of characters. The former is based on word segmentation\nalgorithms optimized using corpus-level statistics with no regard to the\ntranslation task. The latter learns directly from translation data but requires\nrather deep architectures. In this paper, we propose to translate words by\nmodeling word formation through a hierarchical latent variable model which\nmimics the process of morphological inflection. Our model generates words one\ncharacter at a time by composing two latent representations: a continuous one,\naimed at capturing the lexical semantics, and a set of (approximately) discrete\nfeatures, aimed at capturing the morphosyntactic function, which are shared\namong different surface forms. Our model achieves better accuracy in\ntranslation into three morphologically-rich languages than conventional\nopen-vocabulary NMT methods, while also demonstrating a better generalization\ncapacity under low to mid-resource settings.', 'Recent studies on direct speech translation show continuous improvements by\nmeans of data augmentation techniques and bigger deep learning models. While\nthese methods are helping to close the gap between this new approach and the\nmore traditional cascaded one, there are many incongruities among different\nstudies that make it difficult to assess the state of the art. Surprisingly,\none point of discussion is the segmentation of the target text. Character-level\nsegmentation has been initially proposed to obtain an open vocabulary, but it\nresults on long sequences and long training time. Then, subword-level\nsegmentation became the state of the art in neural machine translation as it\nproduces shorter sequences that reduce the training time, while being superior\nto word-level models. As such, recent works on speech translation started using\ntarget subwords despite the initial use of characters and some recent claims of\nbetter results at the character level. In this work, we perform an extensive\ncomparison of the two methods on three benchmarks covering 8 language\ndirections and multilingual training. Subword-level segmentation compares\nfavorably in all settings, outperforming its character-level counterpart in a\nrange of 1 to 3 BLEU points.', 'Dealing with the complex word forms in morphologically rich languages is an\nopen problem in language processing, and is particularly important in\ntranslation. In contrast to most modern neural systems of translation, which\ndiscard the identity for rare words, in this paper we propose several\narchitectures for learning word representations from character and morpheme\nlevel word decompositions. We incorporate these representations in a novel\nmachine translation model which jointly learns word alignments and translations\nvia a hard attention mechanism. Evaluating on translating from several\nmorphologically rich languages into English, we show consistent improvements\nover strong baseline methods, of between 1 and 1.5 BLEU points.', 'A great proportion of sequence-to-sequence (Seq2Seq) models for Neural\nMachine Translation (NMT) adopt Recurrent Neural Network (RNN) to generate\ntranslation word by word following a sequential order. As the studies of\nlinguistics have proved that language is not linear word sequence but sequence\nof complex structure, translation at each step should be conditioned on the\nwhole target-side context. To tackle the problem, we propose a new NMT model\nthat decodes the sequence with the guidance of its structural prediction of the\ncontext of the target sequence. Our model generates translation based on the\nstructural prediction of the target-side context so that the translation can be\nfreed from the bind of sequential order. Experimental results demonstrate that\nour model is more competitive compared with the state-of-the-art methods, and\nthe analysis reflects that our model is also robust to translating sentences of\ndifferent lengths and it also reduces repetition with the instruction from the\ntarget-side context for decoding.', 'Neural machine translation (NMT) takes deterministic sequences for source\nrepresentations. However, either word-level or subword-level segmentations have\nmultiple choices to split a source sequence with different word segmentors or\ndifferent subword vocabulary sizes. We hypothesize that the diversity in\nsegmentations may affect the NMT performance. To integrate different\nsegmentations with the state-of-the-art NMT model, Transformer, we propose\nlattice-based encoders to explore effective word or subword representation in\nan automatic way during training. We propose two methods: 1) lattice positional\nencoding and 2) lattice-aware self-attention. These two methods can be used\ntogether and show complementary to each other to further improve translation\nperformance. Experiment results show superiorities of lattice-based encoders in\nword-level and subword-level representations over conventional Transformer\nencoder.', 'Neural Machine Translation (NMT) is an end-to-end learning approach for\nautomated translation, with the potential to overcome many of the weaknesses of\nconventional phrase-based translation systems. Unfortunately, NMT systems are\nknown to be computationally expensive both in training and in translation\ninference. Also, most NMT systems have difficulty with rare words. These issues\nhave hindered NMT\'s use in practical deployments and services, where both\naccuracy and speed are essential. In this work, we present GNMT, Google\'s\nNeural Machine Translation system, which attempts to address many of these\nissues. Our model consists of a deep LSTM network with 8 encoder and 8 decoder\nlayers using attention and residual connections. To improve parallelism and\ntherefore decrease training time, our attention mechanism connects the bottom\nlayer of the decoder to the top layer of the encoder. To accelerate the final\ntranslation speed, we employ low-precision arithmetic during inference\ncomputations. To improve handling of rare words, we divide words into a limited\nset of common sub-word units (""wordpieces"") for both input and output. This\nmethod provides a good balance between the flexibility of ""character""-delimited\nmodels and the efficiency of ""word""-delimited models, naturally handles\ntranslation of rare words, and ultimately improves the overall accuracy of the\nsystem. Our beam search technique employs a length-normalization procedure and\nuses a coverage penalty, which encourages generation of an output sentence that\nis most likely to cover all the words in the source sentence. On the WMT\'14\nEnglish-to-French and English-to-German benchmarks, GNMT achieves competitive\nresults to state-of-the-art. Using a human side-by-side evaluation on a set of\nisolated simple sentences, it reduces translation errors by an average of 60%\ncompared to Google\'s phrase-based production system.', ""Recently, neural machine translation (NMT) has emerged as a powerful\nalternative to conventional statistical approaches. However, its performance\ndrops considerably in the presence of morphologically rich languages (MRLs).\nNeural engines usually fail to tackle the large vocabulary and high\nout-of-vocabulary (OOV) word rate of MRLs. Therefore, it is not suitable to\nexploit existing word-based models to translate this set of languages. In this\npaper, we propose an extension to the state-of-the-art model of Chung et al.\n(2016), which works at the character level and boosts the decoder with\ntarget-side morphological information. In our architecture, an additional\nmorphology table is plugged into the model. Each time the decoder samples from\na target vocabulary, the table sends auxiliary signals from the most relevant\naffixes in order to enrich the decoder's current state and constrain it to\nprovide better predictions. We evaluated our model to translate English into\nGerman, Russian, and Turkish as three MRLs and observed significant\nimprovements."", ""Neural machine translation (NMT) is nowadays commonly applied at the subword\nlevel, using byte-pair encoding. A promising alternative approach focuses on\ncharacter-level translation, which simplifies processing pipelines in NMT\nconsiderably. This approach, however, must consider relatively longer\nsequences, rendering the training process prohibitively expensive. In this\npaper, we discuss a novel, Transformer-based approach, that we compare, both in\nspeed and in quality to the Transformer at subword and character levels, as\nwell as previously developed character-level models. We evaluate our models on\n4 language pairs from WMT'15: DE-EN, CS-EN, FI-EN and RU-EN. The proposed novel\narchitecture can be trained on a single GPU and is 34% percent faster than the\ncharacter-level Transformer; still, the obtained results are at least on par\nwith it. In addition, our proposed model outperforms the subword-level model in\nFI-EN and shows close results in CS-EN. To stimulate further research in this\narea and close the gap with subword-level NMT, we make all our code and models\npublicly available."", 'In recent years, Neural Machine Translation (NMT) has been shown to be more\neffective than phrase-based statistical methods, thus quickly becoming the\nstate of the art in machine translation (MT). However, NMT systems are limited\nin translating low-resourced languages, due to the significant amount of\nparallel data that is required to learn useful mappings between languages. In\nthis work, we show how the so-called multilingual NMT can help to tackle the\nchallenges associated with low-resourced language translation. The underlying\nprinciple of multilingual NMT is to force the creation of hidden\nrepresentations of words in a shared semantic space across multiple languages,\nthus enabling a positive parameter transfer across languages. Along this\ndirection, we present multilingual translation experiments with three languages\n(English, Italian, Romanian) covering six translation directions, utilizing\nboth recurrent neural networks and transformer (or self-attentive) neural\nnetworks. We then focus on the zero-shot translation problem, that is how to\nleverage multi-lingual data in order to learn translation directions that are\nnot covered by the available training material. To this aim, we introduce our\nrecently proposed iterative self-training method, which incrementally improves\na multilingual NMT on a zero-shot direction by just relying on monolingual\ndata. Our results on TED talks data show that multilingual NMT outperforms\nconventional bilingual NMT, that the transformer NMT outperforms recurrent NMT,\nand that zero-shot NMT outperforms conventional pivoting methods and even\nmatches the performance of a fully-trained bilingual system.', 'Subword units are an effective way to alleviate the open vocabulary problems\nin neural machine translation (NMT). While sentences are usually converted into\nunique subword sequences, subword segmentation is potentially ambiguous and\nmultiple segmentations are possible even with the same vocabulary. The question\naddressed in this paper is whether it is possible to harness the segmentation\nambiguity as a noise to improve the robustness of NMT. We present a simple\nregularization method, subword regularization, which trains the model with\nmultiple subword segmentations probabilistically sampled during training. In\naddition, for better subword sampling, we propose a new subword segmentation\nalgorithm based on a unigram language model. We experiment with multiple\ncorpora and report consistent improvements especially on low resource and\nout-of-domain settings.', 'Applying the Transformer architecture on the character level usually requires\nvery deep architectures that are difficult and slow to train. These problems\ncan be partially overcome by incorporating a segmentation into tokens in the\nmodel. We show that by initially training a subword model and then finetuning\nit on characters, we can obtain a neural machine translation model that works\nat the character level without requiring token segmentation. We use only the\nvanilla 6-layer Transformer Base architecture. Our character-level models\nbetter capture morphological phenomena and show more robustness to noise at the\nexpense of somewhat worse overall translation quality. Our study is a\nsignificant step towards high-performance and easy to train character-based\nmodels that are not extremely large.', 'Neural machine translation (NMT) has achieved notable success in recent\ntimes, however it is also widely recognized that this approach has limitations\nwith handling infrequent words and word pairs. This paper presents a novel\nmemory-augmented NMT (M-NMT) architecture, which stores knowledge about how\nwords (usually infrequently encountered ones) should be translated in a memory\nand then utilizes them to assist the neural model. We use this memory mechanism\nto combine the knowledge learned from a conventional statistical machine\ntranslation system and the rules learned by an NMT system, and also propose a\nsolution for out-of-vocabulary (OOV) words based on this framework. Our\nexperiments on two Chinese-English translation tasks demonstrated that the\nM-NMT architecture outperformed the NMT baseline by $9.0$ and $2.7$ BLEU points\non the two tasks, respectively. Additionally, we found this architecture\nresulted in a much more effective OOV treatment compared to competitive\nmethods.', 'Phrases play an important role in natural language understanding and machine\ntranslation (Sag et al., 2002; Villavicencio et al., 2005). However, it is\ndifficult to integrate them into current neural machine translation (NMT) which\nreads and generates sentences word by word. In this work, we propose a method\nto translate phrases in NMT by integrating a phrase memory storing target\nphrases from a phrase-based statistical machine translation (SMT) system into\nthe encoder-decoder architecture of NMT. At each decoding step, the phrase\nmemory is first re-written by the SMT model, which dynamically generates\nrelevant target phrases with contextual information provided by the NMT model.\nThen the proposed model reads the phrase memory to make probability estimations\nfor all phrases in the phrase memory. If phrase generation is carried on, the\nNMT decoder selects an appropriate phrase from the memory to perform phrase\ntranslation and updates its decoding state by consuming the words in the\nselected phrase. Otherwise, the NMT decoder generates a word from the\nvocabulary as the general NMT decoder does. Experiment results on the Chinese\nto English translation show that the proposed model achieves significant\nimprovements over the baseline on various test sets.', 'Document-level machine translation incorporates inter-sentential dependencies\ninto the translation of a source sentence. In this paper, we propose a new\nframework to model cross-sentence dependencies by training neural machine\ntranslation (NMT) to predict both the target translation and surrounding\nsentences of a source sentence. By enforcing the NMT model to predict source\ncontext, we want the model to learn ""contextualized"" source sentence\nrepresentations that capture document-level dependencies on the source side. We\nfurther propose two different methods to learn and integrate such\ncontextualized sentence embeddings into NMT: a joint training method that\njointly trains an NMT model with the source context prediction model and a\npre-training & fine-tuning method that pretrains the source context prediction\nmodel on a large-scale monolingual document corpus and then fine-tunes it with\nthe NMT model. Experiments on Chinese-English and English-German translation\nshow that both methods can substantially improve the translation quality over a\nstrong document-level Transformer baseline.', 'The necessity of using a fixed-size word vocabulary in order to control the\nmodel complexity in state-of-the-art neural machine translation (NMT) systems\nis an important bottleneck on performance, especially for morphologically rich\nlanguages. Conventional methods that aim to overcome this problem by using\nsub-word or character-level representations solely rely on statistics and\ndisregard the linguistic properties of words, which leads to interruptions in\nthe word structure and causes semantic and syntactic losses. In this paper, we\npropose a new vocabulary reduction method for NMT, which can reduce the\nvocabulary of a given input corpus at any rate while also considering the\nmorphological properties of the language. Our method is based on unsupervised\nmorphology learning and can be, in principle, used for pre-processing any\nlanguage pair. We also present an alternative word segmentation method based on\nsupervised morphological analysis, which aids us in measuring the accuracy of\nour model. We evaluate our method in Turkish-to-English NMT task where the\ninput language is morphologically rich and agglutinative. We analyze different\nrepresentation methods in terms of translation accuracy as well as the semantic\nand syntactic properties of the generated output. Our method obtains a\nsignificant improvement of 2.3 BLEU points over the conventional vocabulary\nreduction technique, showing that it can provide better accuracy in open\nvocabulary translation of morphologically rich languages.', 'Deep Learning techniques are powerful in mimicking humans in a particular set\nof problems. They have achieved a remarkable performance in complex learning\ntasks. Deep learning inspired Neural Machine Translation (NMT) is a proficient\ntechnique that outperforms traditional machine translation. Performing\nmachine-aided translation on Indic languages has always been a challenging task\nconsidering their rich and diverse grammar. The neural machine translation has\nshown quality results compared to the traditional machine translation\napproaches. The fully automatic machine translation becomes problematic when it\ncomes to low-resourced languages, especially with Sanskrit. This paper presents\nattention mechanism based neural machine translation by selectively focusing on\na particular part of language sentences during translation. The work shows the\nconstruction of Sanskrit to Hindi bilingual parallel corpus with nearly 10K\nsamples and having 178,000 tokens. The neural translation model equipped with\nan attention mechanism has been trained on Sanskrit to Hindi parallel corpus.\nThe approach has shown the significance of attention mechanisms to overcome\nlong-term dependencies, primarily associated with low resources Indic\nlanguages. The paper shows the attention plots on testing data to demonstrate\nthe alignment between source and translated words. For the evaluation of the\ntranslated sentences, manual score based human evaluation and automatic\nevaluation metric based techniques have been adopted. The attention mechanism\nbased neural translation has achieved 88% accuracy in human evaluation and a\nBLEU score of 0.92 on Sanskrit to Hindi translation.', ""Most existing machine translation systems operate at the level of words,\nrelying on explicit segmentation to extract tokens. We introduce a neural\nmachine translation (NMT) model that maps a source character sequence to a\ntarget character sequence without any segmentation. We employ a character-level\nconvolutional network with max-pooling at the encoder to reduce the length of\nsource representation, allowing the model to be trained at a speed comparable\nto subword-level models while capturing local regularities. Our\ncharacter-to-character model outperforms a recently proposed baseline with a\nsubword-level encoder on WMT'15 DE-EN and CS-EN, and gives comparable\nperformance on FI-EN and RU-EN. We then demonstrate that it is possible to\nshare a single character-level encoder across multiple languages by training a\nmodel on a many-to-one translation task. In this multilingual setting, the\ncharacter-level encoder significantly outperforms the subword-level encoder on\nall the language pairs. We observe that on CS-EN, FI-EN and RU-EN, the quality\nof the multilingual character-level translation even surpasses the models\nspecifically trained on that language pair alone, both in terms of BLEU score\nand human judgment."", 'The authors of (Cho et al., 2014a) have shown that the recently introduced\nneural network translation systems suffer from a significant drop in\ntranslation quality when translating long sentences, unlike existing\nphrase-based translation systems. In this paper, we propose a way to address\nthis issue by automatically segmenting an input sentence into phrases that can\nbe easily translated by the neural network translation model. Once each segment\nhas been independently translated by the neural machine translation model, the\ntranslated clauses are concatenated to form a final translation. Empirical\nresults show a significant improvement in translation quality for long\nsentences.', 'We propose a language-independent approach for improving statistical machine\ntranslation for morphologically rich languages using a hybrid morpheme-word\nrepresentation where the basic unit of translation is the morpheme, but word\nboundaries are respected at all stages of the translation process. Our model\nextends the classic phrase-based model by means of (1) word boundary-aware\nmorpheme-level phrase extraction, (2) minimum error-rate training for a\nmorpheme-level translation model using word-level BLEU, and (3) joint scoring\nwith morpheme- and word-level language models. Further improvements are\nachieved by combining our model with the classic one. The evaluation on English\nto Finnish using Europarl (714K sentence pairs; 15.5M English words) shows\nstatistically significant improvements over the classic model based on BLEU and\nhuman judgments.', 'In neural machine translation (NMT), generation of a target word depends on\nboth source and target contexts. We find that source contexts have a direct\nimpact on the adequacy of a translation while target contexts affect the\nfluency. Intuitively, generation of a content word should rely more on the\nsource context and generation of a functional word should rely more on the\ntarget context. Due to the lack of effective control over the influence from\nsource and target contexts, conventional NMT tends to yield fluent but\ninadequate translations. To address this problem, we propose context gates\nwhich dynamically control the ratios at which source and target contexts\ncontribute to the generation of target words. In this way, we can enhance both\nthe adequacy and fluency of NMT with more careful control of the information\nflow from contexts. Experiments show that our approach significantly improves\nupon a standard attention-based NMT system by +2.3 BLEU points.', ""Nearly all previous work on neural machine translation (NMT) has used quite\nrestricted vocabularies, perhaps with a subsequent method to patch in unknown\nwords. This paper presents a novel word-character solution to achieving open\nvocabulary NMT. We build hybrid systems that translate mostly at the word level\nand consult the character components for rare words. Our character-level\nrecurrent neural networks compute source word representations and recover\nunknown target words when needed. The twofold advantage of such a hybrid\napproach is that it is much faster and easier to train than character-based\nones; at the same time, it never produces unknown words as in the case of\nword-based models. On the WMT'15 English to Czech translation task, this hybrid\napproach offers an addition boost of +2.1-11.4 BLEU points over models that\nalready handle unknown words. Our best system achieves a new state-of-the-art\nresult with 20.7 BLEU score. We demonstrate that our character models can\nsuccessfully learn to not only generate well-formed words for Czech, a\nhighly-inflected language with a very complex vocabulary, but also build\ncorrect representations for English source words."", 'It has been shown that increasing model depth improves the quality of neural\nmachine translation. However, different architectural variants to increase\nmodel depth have been proposed, and so far, there has been no thorough\ncomparative study.\n  In this work, we describe and evaluate several existing approaches to\nintroduce depth in neural machine translation. Additionally, we explore novel\narchitectural variants, including deep transition RNNs, and we vary how\nattention is used in the deep decoder. We introduce a novel ""BiDeep"" RNN\narchitecture that combines deep transition RNNs and stacked RNNs.\n  Our evaluation is carried out on the English to German WMT news translation\ndataset, using a single-GPU machine for both training and inference. We find\nthat several of our proposed architectures improve upon existing approaches in\nterms of speed and translation quality. We obtain best improvements with a\nBiDeep RNN of combined depth 8, obtaining an average improvement of 1.5 BLEU\nover a strong shallow baseline.\n  We release our code for ease of adoption.', ""Most of the existing Neural Machine Translation (NMT) models focus on the\nconversion of sequential data and do not directly use syntactic information. We\npropose a novel end-to-end syntactic NMT model, extending a\nsequence-to-sequence model with the source-side phrase structure. Our model has\nan attention mechanism that enables the decoder to generate a translated word\nwhile softly aligning it with phrases as well as words of the source sentence.\nExperimental results on the WAT'15 English-to-Japanese dataset demonstrate that\nour proposed model considerably outperforms sequence-to-sequence attentional\nNMT models and compares favorably with the state-of-the-art tree-to-string SMT\nsystem."", 'Exposing diverse subword segmentations to neural machine translation (NMT)\nmodels often improves the robustness of machine translation as NMT models can\nexperience various subword candidates. However, the diversification of subword\nsegmentations mostly relies on the pre-trained subword language models from\nwhich erroneous segmentations of unseen words are less likely to be sampled. In\nthis paper, we present adversarial subword regularization (ADVSR) to study\nwhether gradient signals during training can be a substitute criterion for\nexposing diverse subword segmentations. We experimentally show that our\nmodel-based adversarial samples effectively encourage NMT models to be less\nsensitive to segmentation errors and improve the performance of NMT models in\nlow-resource and out-domain datasets.', ""We present IntelliCAT, an interactive translation interface with neural\nmodels that streamline the post-editing process on machine translation output.\nWe leverage two quality estimation (QE) models at different granularities:\nsentence-level QE, to predict the quality of each machine-translated sentence,\nand word-level QE, to locate the parts of the machine-translated sentence that\nneed correction. Additionally, we introduce a novel translation suggestion\nmodel conditioned on both the left and right contexts, providing alternatives\nfor specific words or phrases for correction. Finally, with word alignments,\nIntelliCAT automatically preserves the original document's styles in the\ntranslated document. The experimental results show that post-editing based on\nthe proposed QE and translation suggestions can significantly improve\ntranslation quality. Furthermore, a user study reveals that three features\nprovided in IntelliCAT significantly accelerate the post-editing task,\nachieving a 52.9\\% speedup in translation time compared to translating from\nscratch. The interface is publicly available at\nhttps://intellicat.beringlab.com/."", 'Recurrent sequence generators conditioned on input data through an attention\nmechanism have recently shown very good performance on a range of tasks in-\ncluding machine translation, handwriting synthesis and image caption gen-\neration. We extend the attention-mechanism with features needed for speech\nrecognition. We show that while an adaptation of the model used for machine\ntranslation in reaches a competitive 18.7% phoneme error rate (PER) on the\nTIMIT phoneme recognition task, it can only be applied to utterances which are\nroughly as long as the ones it was trained on. We offer a qualitative\nexplanation of this failure and propose a novel and generic method of adding\nlocation-awareness to the attention mechanism to alleviate this issue. The new\nmethod yields a model that is robust to long inputs and achieves 18% PER in\nsingle utterances and 20% in 10-times longer (repeated) utterances. Finally, we\npropose a change to the at- tention mechanism that prevents it from\nconcentrating too much on single frames, which further reduces PER to 17.6%\nlevel.', 'Most of the Neural Machine Translation (NMT) models are based on the\nsequence-to-sequence (Seq2Seq) model with an encoder-decoder framework equipped\nwith the attention mechanism. However, the conventional attention mechanism\ntreats the decoding at each time step equally with the same matrix, which is\nproblematic since the softness of the attention for different types of words\n(e.g. content words and function words) should differ. Therefore, we propose a\nnew model with a mechanism called Self-Adaptive Control of Temperature (SACT)\nto control the softness of attention by means of an attention temperature.\nExperimental results on the Chinese-English translation and English-Vietnamese\ntranslation demonstrate that our model outperforms the baseline models, and the\nanalysis and the case study show that our model can attend to the most relevant\nelements in the source-side contexts and generate the translation of high\nquality.', 'Neural machine translation (NMT) has achieved impressive performance on\nmachine translation task in recent years. However, in consideration of\nefficiency, a limited-size vocabulary that only contains the top-N highest\nfrequency words are employed for model training, which leads to many rare and\nunknown words. It is rather difficult when translating from the low-resource\nand morphologically-rich agglutinative languages, which have complex morphology\nand large vocabulary. In this paper, we propose a morphological word\nsegmentation method on the source-side for NMT that incorporates morphology\nknowledge to preserve the linguistic and semantic information in the word\nstructure while reducing the vocabulary size at training time. It can be\nutilized as a preprocessing tool to segment the words in agglutinative\nlanguages for other natural language processing (NLP) tasks. Experimental\nresults show that our morphologically motivated word segmentation method is\nbetter suitable for the NMT model, which achieves significant improvements on\nTurkish-English and Uyghur-Chinese machine translation tasks on account of\nreducing data sparseness and language complexity.', 'Neural machine translation (NMT) systems aim to map text from one language\ninto another. While there are a wide variety of applications of NMT, one of the\nmost important is translation of natural language. A distinguishing factor of\nnatural language is that words are typically ordered according to the rules of\nthe grammar of a given language. Although many advances have been made in\ndeveloping NMT systems for translating natural language, little research has\nbeen done on understanding how the word ordering of and lexical similarity\nbetween the source and target language affect translation performance. Here, we\ninvestigate these relationships on a variety of low-resource language pairs\nfrom the OpenSubtitles2016 database, where the source language is English, and\nfind that the more similar the target language is to English, the greater the\ntranslation performance. In addition, we study the impact of providing NMT\nmodels with part of speech of words (POS) in the English sequence and find\nthat, for Transformer-based models, the more dissimilar the target language is\nfrom English, the greater the benefit provided by POS.', 'Neural machine translation (MT) models obtain state-of-the-art performance\nwhile maintaining a simple, end-to-end architecture. However, little is known\nabout what these models learn about source and target languages during the\ntraining process. In this work, we analyze the representations learned by\nneural MT models at various levels of granularity and empirically evaluate the\nquality of the representations for learning morphology through extrinsic\npart-of-speech and morphological tagging tasks. We conduct a thorough\ninvestigation along several parameters: word-based vs. character-based\nrepresentations, depth of the encoding layer, the identity of the target\nlanguage, and encoder vs. decoder representations. Our data-driven,\nquantitative evaluation sheds light on important aspects in the neural MT\nsystem and its ability to capture word structure.', ""Neural Machine Translation (NMT) can be used to generate fluent output. As\nsuch, language models have been investigated for incorporation with NMT. In\nprior investigations, two models have been used: a translation model and a\nlanguage model. The translation model's predictions are weighted by the\nlanguage model with a hand-crafted ratio in advance. However, these approaches\nfail to adopt the language model weighting with regard to the translation\nhistory. In another line of approach, language model prediction is incorporated\ninto the translation model by jointly considering source and target\ninformation. However, this line of approach is limited because it largely\nignores the adequacy of the translation output.\n  Accordingly, this work employs two mechanisms, the translation model and the\nlanguage model, with an attentive architecture to the language model as an\nauxiliary element of the translation model. Compared with previous work in\nEnglish--Japanese machine translation using a language model, the experimental\nresults obtained with the proposed Dynamic Fusion mechanism improve BLEU and\nRank-based Intuitive Bilingual Evaluation Scores (RIBES) scores. Additionally,\nin the analyses of the attention and predictivity of the language model, the\nDynamic Fusion mechanism allows predictive language modeling that conforms to\nthe appropriate grammatical structure."", 'Standard neural machine translation (NMT) is on the assumption of\ndocument-level context independent. Most existing document-level NMT methods\nare satisfied with a smattering sense of brief document-level information,\nwhile this work focuses on exploiting detailed document-level context in terms\nof multiple forms of document embeddings, which is capable of sufficiently\nmodeling deeper and richer document-level context. The proposed document-aware\nNMT is implemented to enhance the Transformer baseline by introducing both\nglobal and local document-level clues on the source end. Experiments show that\nthe proposed method significantly improves the translation performance over\nstrong baselines and other related studies.', ""Neural machine translation (NMT) aims at solving machine translation (MT)\nproblems using neural networks and has exhibited promising results in recent\nyears. However, most of the existing NMT models are shallow and there is still\na performance gap between a single NMT model and the best conventional MT\nsystem. In this work, we introduce a new type of linear connections, named\nfast-forward connections, based on deep Long Short-Term Memory (LSTM) networks,\nand an interleaved bi-directional architecture for stacking the LSTM layers.\nFast-forward connections play an essential role in propagating the gradients\nand building a deep topology of depth 16. On the WMT'14 English-to-French task,\nwe achieve BLEU=37.7 with a single attention model, which outperforms the\ncorresponding single shallow model by 6.2 BLEU points. This is the first time\nthat a single NMT model achieves state-of-the-art performance and outperforms\nthe best conventional model by 0.7 BLEU points. We can still achieve BLEU=36.3\neven without using an attention mechanism. After special handling of unknown\nwords and model ensembling, we obtain the best score reported to date on this\ntask with BLEU=40.4. Our models are also validated on the more difficult WMT'14\nEnglish-to-German task."", 'Translating characters instead of words or word-fragments has the potential\nto simplify the processing pipeline for neural machine translation (NMT), and\nimprove results by eliminating hyper-parameters and manual feature engineering.\nHowever, it results in longer sequences in which each symbol contains less\ninformation, creating both modeling and computational challenges. In this\npaper, we show that the modeling problem can be solved by standard\nsequence-to-sequence architectures of sufficient depth, and that deep models\noperating at the character level outperform identical models operating over\nword fragments. This result implies that alternative architectures for handling\ncharacter input are better viewed as methods for reducing computation time than\nas improved ways of modeling longer sequences. From this perspective, we\nevaluate several techniques for character-level NMT, verify that they do not\nmatch the performance of our deep character baseline model, and evaluate the\nperformance versus computation time tradeoffs they offer. Within this\nframework, we also perform the first evaluation for NMT of conditional\ncomputation over time, in which the model learns which timesteps can be\nskipped, rather than having them be dictated by a fixed schedule specified\nbefore training begins.', 'Neural Machine Translation (NMT) for low-resource languages suffers from low\nperformance because of the lack of large amounts of parallel data and language\ndiversity. To contribute to ameliorating this problem, we built a baseline\nmodel for English-Hausa machine translation, which is considered a task for\nlow-resource language. The Hausa language is the second largest Afro-Asiatic\nlanguage in the world after Arabic and it is the third largest language for\ntrading across a larger swath of West Africa countries, after English and\nFrench. In this paper, we curated different datasets containing Hausa-English\nparallel corpus for our translation. We trained baseline models and evaluated\nthe performance of our models using the Recurrent and Transformer\nencoder-decoder architecture with two tokenization approaches: standard\nword-level tokenization and Byte Pair Encoding (BPE) subword tokenization.', 'In this paper we demonstrate that $\\textit{context free grammar (CFG) based\nmethods for grammar induction benefit from modeling lexical dependencies}$.\nThis contrasts to the most popular current methods for grammar induction, which\nfocus on discovering $\\textit{either}$ constituents $\\textit{or}$ dependencies.\nPrevious approaches to marry these two disparate syntactic formalisms (e.g.\nlexicalized PCFGs) have been plagued by sparsity, making them unsuitable for\nunsupervised grammar induction. However, in this work, we present novel neural\nmodels of lexicalized PCFGs which allow us to overcome sparsity problems and\neffectively induce both constituents and dependencies within a single model.\nExperiments demonstrate that this unified framework results in stronger results\non both representations than achieved when modeling either formalism alone.\nCode is available at https://github.com/neulab/neural-lpcfg.', 'Neural machine translation is a recently proposed approach to machine\ntranslation. Unlike the traditional statistical machine translation, the neural\nmachine translation aims at building a single neural network that can be\njointly tuned to maximize the translation performance. The models proposed\nrecently for neural machine translation often belong to a family of\nencoder-decoders and consists of an encoder that encodes a source sentence into\na fixed-length vector from which a decoder generates a translation. In this\npaper, we conjecture that the use of a fixed-length vector is a bottleneck in\nimproving the performance of this basic encoder-decoder architecture, and\npropose to extend this by allowing a model to automatically (soft-)search for\nparts of a source sentence that are relevant to predicting a target word,\nwithout having to form these parts as a hard segment explicitly. With this new\napproach, we achieve a translation performance comparable to the existing\nstate-of-the-art phrase-based system on the task of English-to-French\ntranslation. Furthermore, qualitative analysis reveals that the\n(soft-)alignments found by the model agree well with our intuition.', 'Most modern neural machine translation (NMT) systems rely on presegmented\ninputs. Segmentation granularity importantly determines the input and output\nsequence lengths, hence the modeling depth, and source and target vocabularies,\nwhich in turn determine model size, computational costs of softmax\nnormalization, and handling of out-of-vocabulary words. However, the current\npractice is to use static, heuristic-based segmentations that are fixed before\nNMT training. This begs the question whether the chosen segmentation is optimal\nfor the translation task. To overcome suboptimal segmentation choices, we\npresent an algorithm for dynamic segmentation based on the Adaptative\nComputation Time algorithm (Graves 2016), that is trainable end-to-end and\ndriven by the NMT objective. In an evaluation on four translation tasks we\nfound that, given the freedom to navigate between different segmentation\nlevels, the model prefers to operate on (almost) character level, providing\nsupport for purely character-level NMT models from a novel angle.', 'Multilingual neural machine translation (NMT), which translates multiple\nlanguages using a single model, is of great practical importance due to its\nadvantages in simplifying the training process, reducing online maintenance\ncosts, and enhancing low-resource and zero-shot translation. Given there are\nthousands of languages in the world and some of them are very different, it is\nextremely burdensome to handle them all in a single model or use a separate\nmodel for each language pair. Therefore, given a fixed resource budget, e.g.,\nthe number of models, how to determine which languages should be supported by\none model is critical to multilingual NMT, which, unfortunately, has been\nignored by previous work. In this work, we develop a framework that clusters\nlanguages into different groups and trains one multilingual model for each\ncluster. We study two methods for language clustering: (1) using prior\nknowledge, where we cluster languages according to language family, and (2)\nusing language embedding, in which we represent each language by an embedding\nvector and cluster them in the embedding space. In particular, we obtain the\nembedding vectors of all the languages by training a universal neural machine\ntranslation model. Our experiments on 23 languages show that the first\nclustering method is simple and easy to understand but leading to suboptimal\ntranslation accuracy, while the second method sufficiently captures the\nrelationship among languages well and improves the translation accuracy for\nalmost all the languages over baseline methods', 'Neural Machine Translation (NMT) is a new approach for Machine Translation\n(MT), and due to its success, it has absorbed the attention of many researchers\nin the field. In this paper, we study NMT model on Persian-English language\npairs, to analyze the model and investigate the appropriateness of the model\nfor scarce-resourced scenarios, the situation that exists for Persian-centered\ntranslation systems. We adjust the model for the Persian language and find the\nbest parameters and hyper parameters for two tasks: translation and\ntransliteration. We also apply some preprocessing task on the Persian dataset\nwhich yields to increase for about one point in terms of BLEU score. Also, we\nhave modified the loss function to enhance the word alignment of the model.\nThis new loss function yields a total of 1.87 point improvements in terms of\nBLEU score in the translation quality.', ""Neural machine translation (NMT) has arguably achieved human level parity\nwhen trained and evaluated at the sentence-level. Document-level neural machine\ntranslation has received less attention and lags behind its sentence-level\ncounterpart. The majority of the proposed document-level approaches investigate\nways of conditioning the model on several source or target sentences to capture\ndocument context. These approaches require training a specialized NMT model\nfrom scratch on parallel document-level corpora. We propose an approach that\ndoesn't require training a specialized model on parallel document-level corpora\nand is applied to a trained sentence-level NMT model at decoding time. We\nprocess the document from left to right multiple times and self-train the\nsentence-level model on pairs of source sentences and generated translations.\nOur approach reinforces the choices made by the model, thus making it more\nlikely that the same choices will be made in other sentences in the document.\nWe evaluate our approach on three document-level datasets: NIST\nChinese-English, WMT'19 Chinese-English and OpenSubtitles English-Russian. We\ndemonstrate that our approach has higher BLEU score and higher human preference\nthan the baseline. Qualitative analysis of our approach shows that choices made\nby model are consistent across the document."", 'Neural machine translation (NMT), a new approach to machine translation, has\nachieved promising results comparable to those of traditional approaches such\nas statistical machine translation (SMT). Despite its recent success, NMT\ncannot handle a larger vocabulary because the training complexity and decoding\ncomplexity proportionally increase with the number of target words. This\nproblem becomes even more serious when translating patent documents, which\ncontain many technical terms that are observed infrequently. In this paper, we\npropose to select phrases that contain out-of-vocabulary words using the\nstatistical approach of branching entropy. This allows the proposed NMT system\nto be applied to a translation task of any language pair without any\nlanguage-specific knowledge about technical term identification. The selected\nphrases are then replaced with tokens during training and post-translated by\nthe phrase translation table of SMT. Evaluation on Japanese-to-Chinese,\nChinese-to-Japanese, Japanese-to-English and English-to-Japanese patent\nsentence translation proved the effectiveness of phrases selected with\nbranching entropy, where the proposed NMT model achieves a substantial\nimprovement over a baseline NMT model without our proposed technique. Moreover,\nthe number of translation errors of under-translation by the baseline NMT model\nwithout our proposed technique reduces to around half by the proposed NMT\nmodel.', 'The quality of machine translation is rapidly evolving. Today one can find\nseveral machine translation systems on the web that provide reasonable\ntranslations, although the systems are not perfect. In some specific domains,\nthe quality may decrease. A recently proposed approach to this domain is neural\nmachine translation. It aims at building a jointly-tuned single neural network\nthat maximizes translation performance, a very different approach from\ntraditional statistical machine translation. Recently proposed neural machine\ntranslation models often belong to the encoder-decoder family in which a source\nsentence is encoded into a fixed length vector that is, in turn, decoded to\ngenerate a translation. The present research examines the effects of different\ntraining methods on a Polish-English Machine Translation system used for\nmedical data. The European Medicines Agency parallel text corpus was used as\nthe basis for training of neural and statistical network-based translation\nsystems. The main machine translation evaluation metrics have also been used in\nanalysis of the systems. A comparison and implementation of a real-time medical\ntranslator is the main focus of our experiments.', 'We introduce a neural machine translation model that views the input and\noutput sentences as sequences of characters rather than words. Since word-level\ninformation provides a crucial source of bias, our input model composes\nrepresentations of character sequences into representations of words (as\ndetermined by whitespace boundaries), and then these are translated using a\njoint attention/translation model. In the target language, the translation is\nmodeled as a sequence of word vectors, but each word is generated one character\nat a time, conditional on the previous character generations in each word. As\nthe representation and generation of words is performed at the character level,\nour model is capable of interpreting and generating unseen word forms. A\nsecondary benefit of this approach is that it alleviates much of the challenges\nassociated with preprocessing/tokenization of the source and target languages.\nWe show that our model can achieve translation results that are on par with\nconventional word-based models.', 'Zero pronouns (ZPs) are frequently omitted in pro-drop languages, but should\nbe recalled in non-pro-drop languages. This discourse phenomenon poses a\nsignificant challenge for machine translation (MT) when translating texts from\npro-drop to non-pro-drop languages. In this paper, we propose a unified and\ndiscourse-aware ZP translation approach for neural MT models. Specifically, we\njointly learn to predict and translate ZPs in an end-to-end manner, allowing\nboth components to interact with each other. In addition, we employ\nhierarchical neural networks to exploit discourse-level context, which is\nbeneficial for ZP prediction and thus translation. Experimental results on both\nChinese-English and Japanese-English data show that our approach significantly\nand accumulatively improves both translation performance and ZP prediction\naccuracy over not only baseline but also previous works using external ZP\nprediction models. Extensive analyses confirm that the performance improvement\ncomes from the alleviation of different kinds of errors especially caused by\nsubjective ZPs.', 'In this paper, we introduce a hybrid search for attention-based neural\nmachine translation (NMT). A target phrase learned with statistical MT models\nextends a hypothesis in the NMT beam search when the attention of the NMT model\nfocuses on the source words translated by this phrase. Phrases added in this\nway are scored with the NMT model, but also with SMT features including\nphrase-level translation probabilities and a target language model.\nExperimental results on German->English news domain and English->Russian\ne-commerce domain translation tasks show that using phrase-based models in NMT\nsearch improves MT quality by up to 2.3% BLEU absolute as compared to a strong\nNMT baseline.', 'Neural machine translation, a recently proposed approach to machine\ntranslation based purely on neural networks, has shown promising results\ncompared to the existing approaches such as phrase-based statistical machine\ntranslation. Despite its recent success, neural machine translation has its\nlimitation in handling a larger vocabulary, as training complexity as well as\ndecoding complexity increase proportionally to the number of target words. In\nthis paper, we propose a method that allows us to use a very large target\nvocabulary without increasing training complexity, based on importance\nsampling. We show that decoding can be efficiently done even with the model\nhaving a very large target vocabulary by selecting only a small subset of the\nwhole target vocabulary. The models trained by the proposed approach are\nempirically found to outperform the baseline models with a small vocabulary as\nwell as the LSTM-based neural machine translation models. Furthermore, when we\nuse the ensemble of a few models with very large target vocabularies, we\nachieve the state-of-the-art translation performance (measured by BLEU) on the\nEnglish->German translation and almost as high performance as state-of-the-art\nEnglish->French translation system.', 'Interest in larger-context neural machine translation, including\ndocument-level and multi-modal translation, has been growing. Multiple works\nhave proposed new network architectures or evaluation schemes, but potentially\nhelpful context is still sometimes ignored by larger-context translation\nmodels. In this paper, we propose a novel learning algorithm that explicitly\nencourages a neural translation model to take into account additional context\nusing a multilevel pair-wise ranking loss. We evaluate the proposed learning\nalgorithm with a transformer-based larger-context translation system on\ndocument-level translation. By comparing performance using actual and random\ncontexts, we show that a model trained with the proposed algorithm is more\nsensitive to the additional context.', ""Machine translation is one of the applications of natural language processing\nwhich has been explored in different languages. Recently researchers started\npaying attention towards machine translation for resource-poor languages and\nclosely related languages. A widespread and underlying problem for these\nmachine translation systems is the variation in orthographic conventions which\ncauses many issues to traditional approaches. Two languages written in two\ndifferent orthographies are not easily comparable, but orthographic information\ncan also be used to improve the machine translation system. This article offers\na survey of research regarding orthography's influence on machine translation\nof under-resourced languages. It introduces under-resourced languages in terms\nof machine translation and how orthographic information can be utilised to\nimprove machine translation. We describe previous work in this area, discussing\nwhat underlying assumptions were made, and showing how orthographic knowledge\nimproves the performance of machine translation of under-resourced languages.\nWe discuss different types of machine translation and demonstrate a recent\ntrend that seeks to link orthographic information with well-established machine\ntranslation methods. Considerable attention is given to current efforts of\ncognates information at different levels of machine translation and the lessons\nthat can be drawn from this. Additionally, multilingual neural machine\ntranslation of closely related languages is given a particular focus in this\nsurvey. This article ends with a discussion of the way forward in machine\ntranslation with orthographic information, focusing on multilingual settings\nand bilingual lexicon induction."", 'To simultaneously capture syntax and global semantics from a text corpus, we\npropose a new larger-context recurrent neural network (RNN) based language\nmodel, which extracts recurrent hierarchical semantic structure via a dynamic\ndeep topic model to guide natural language generation. Moving beyond a\nconventional RNN-based language model that ignores long-range word dependencies\nand sentence order, the proposed model captures not only intra-sentence word\ndependencies, but also temporal transitions between sentences and\ninter-sentence topic dependencies. For inference, we develop a hybrid of\nstochastic-gradient Markov chain Monte Carlo and recurrent autoencoding\nvariational Bayes. Experimental results on a variety of real-world text corpora\ndemonstrate that the proposed model not only outperforms larger-context\nRNN-based language models, but also learns interpretable recurrent multilayer\ntopics and generates diverse sentences and paragraphs that are syntactically\ncorrect and semantically coherent.', 'Sentences in a well-formed text are connected to each other via various links\nto form the cohesive structure of the text. Current neural machine translation\n(NMT) systems translate a text in a conventional sentence-by-sentence fashion,\nignoring such cross-sentence links and dependencies. This may lead to generate\nan incoherent target text for a coherent source text. In order to handle this\nissue, we propose a cache-based approach to modeling coherence for neural\nmachine translation by capturing contextual information either from recently\ntranslated sentences or the entire document. Particularly, we explore two types\nof caches: a dynamic cache, which stores words from the best translation\nhypotheses of preceding sentences, and a topic cache, which maintains a set of\ntarget-side topical words that are semantically related to the document to be\ntranslated. On this basis, we build a new layer to score target words in these\ntwo caches with a cache-based neural model. Here the estimated probabilities\nfrom the cache-based neural model are combined with NMT probabilities into the\nfinal word prediction probabilities via a gating mechanism. Finally, the\nproposed cache-based neural model is trained jointly with NMT system in an\nend-to-end manner. Experiments and analysis presented in this paper demonstrate\nthat the proposed cache-based model achieves substantial improvements over\nseveral state-of-the-art SMT and NMT baselines.', 'The basic concept in Neural Machine Translation (NMT) is to train a large\nNeural Network that maximizes the translation performance on a given parallel\ncorpus. NMT is then using a simple left-to-right beam-search decoder to\ngenerate new translations that approximately maximize the trained conditional\nprobability. The current beam search strategy generates the target sentence\nword by word from left-to- right while keeping a fixed amount of active\ncandidates at each time step. First, this simple search is less adaptive as it\nalso expands candidates whose scores are much worse than the current best.\nSecondly, it does not expand hypotheses if they are not within the best scoring\ncandidates, even if their scores are close to the best one. The latter one can\nbe avoided by increasing the beam size until no performance improvement can be\nobserved. While you can reach better performance, this has the draw- back of a\nslower decoding speed. In this paper, we concentrate on speeding up the decoder\nby applying a more flexible beam search strategy whose candidate size may vary\nat each time step depending on the candidate scores. We speed up the original\ndecoder by up to 43% for the two language pairs German-English and\nChinese-English without losing any translation quality.', 'How to effectively adapt neural machine translation (NMT) models according to\nemerging cases without retraining? Despite the great success of neural machine\ntranslation, updating the deployed models online remains a challenge. Existing\nnon-parametric approaches that retrieve similar examples from a database to\nguide the translation process are promising but are prone to overfit the\nretrieved examples. In this work, we propose to learn Kernel-Smoothed\nTranslation with Example Retrieval (KSTER), an effective approach to adapt\nneural machine translation models online. Experiments on domain adaptation and\nmulti-domain machine translation datasets show that even without expensive\nretraining, KSTER is able to achieve improvement of 1.1 to 1.5 BLEU scores over\nthe best existing online adaptation methods. The code and trained models are\nreleased at https://github.com/jiangqn/KSTER.', 'In this paper, we study a new learning paradigm for Neural Machine\nTranslation (NMT). Instead of maximizing the likelihood of the human\ntranslation as in previous works, we minimize the distinction between human\ntranslation and the translation given by an NMT model. To achieve this goal,\ninspired by the recent success of generative adversarial networks (GANs), we\nemploy an adversarial training architecture and name it as Adversarial-NMT. In\nAdversarial-NMT, the training of the NMT model is assisted by an adversary,\nwhich is an elaborately designed Convolutional Neural Network (CNN). The goal\nof the adversary is to differentiate the translation result generated by the\nNMT model from that by human. The goal of the NMT model is to produce high\nquality translations so as to cheat the adversary. A policy gradient method is\nleveraged to co-train the NMT model and the adversary. Experimental results on\nEnglish$\\rightarrow$French and German$\\rightarrow$English translation tasks\nshow that Adversarial-NMT can achieve significantly better translation quality\nthan several strong baselines.', 'In the past few years, attention mechanisms have become an indispensable\ncomponent of end-to-end neural machine translation models. However, previous\nattention models always refer to some source words when predicting a target\nword, which contradicts with the fact that some target words have no\ncorresponding source words. Motivated by this observation, we propose a novel\nattention model that has the capability of determining when a decoder should\nattend to source words and when it should not. Experimental results on NIST\nChinese-English translation tasks show that the new model achieves an\nimprovement of 0.8 BLEU score over a state-of-the-art baseline.', 'This paper describes the acquisition, preprocessing, segmentation, and\nalignment of an Amharic-English parallel corpus. It will be helpful for machine\ntranslation of a low-resource language, Amharic. We freely released the corpus\nfor research purposes. Furthermore, we developed baseline statistical and\nneural machine translation systems; we trained statistical and neural machine\ntranslation models using the corpus. In the experiments, we also used a large\nmonolingual corpus for the language model of statistical machine translation\nand back-translation of neural machine translation. In the automatic\nevaluation, neural machine translation models outperform statistical machine\ntranslation models by approximately six to seven Bilingual Evaluation\nUnderstudy (BLEU) points. Besides, among the neural machine translation models,\nthe subword models outperform the word-based models by three to four BLEU\npoints. Moreover, two other relevant automatic evaluation metrics, Translation\nEdit Rate on Character Level and Better Evaluation as Ranking, reflect\ncorresponding differences among the trained models.', 'Neural Machine Translation (NMT) is a new approach to machine translation\nthat has made great progress in recent years. However, recent studies show that\nNMT generally produces fluent but inadequate translations (Tu et al. 2016b; Tu\net al. 2016a; He et al. 2016; Tu et al. 2017). This is in contrast to\nconventional Statistical Machine Translation (SMT), which usually yields\nadequate but non-fluent translations. It is natural, therefore, to leverage the\nadvantages of both models for better translations, and in this work we propose\nto incorporate SMT model into NMT framework. More specifically, at each\ndecoding step, SMT offers additional recommendations of generated words based\non the decoding information from NMT (e.g., the generated partial translation\nand attention history). Then we employ an auxiliary classifier to score the SMT\nrecommendations and a gating function to combine the SMT recommendations with\nNMT generations, both of which are jointly trained within the NMT architecture\nin an end-to-end manner. Experimental results on Chinese-English translation\nshow that the proposed approach achieves significant and consistent\nimprovements over state-of-the-art NMT and SMT systems on multiple NIST test\nsets.', ""Neural Machine Translation (NMT) can be improved by including document-level\ncontextual information. For this purpose, we propose a hierarchical attention\nmodel to capture the context in a structured and dynamic manner. The model is\nintegrated in the original NMT architecture as another level of abstraction,\nconditioning on the NMT model's own previous hidden states. Experiments show\nthat hierarchical attention significantly improves the BLEU score over a strong\nNMT baseline with the state-of-the-art in context-aware methods, and that both\nthe encoder and decoder benefit from context in complementary ways."", 'A large number of significant assets are available online in English, which\nis frequently translated into native languages to ease the information sharing\namong local people who are not much familiar with English. However, manual\ntranslation is a very tedious, costly, and time-taking process. To this end,\nmachine translation is an effective approach to convert text to a different\nlanguage without any human involvement. Neural machine translation (NMT) is one\nof the most proficient translation techniques amongst all existing machine\ntranslation systems. In this paper, we have applied NMT on two of the most\nmorphological rich Indian languages, i.e. English-Tamil and English-Malayalam.\nWe proposed a novel NMT model using Multihead self-attention along with\npre-trained Byte-Pair-Encoded (BPE) and MultiBPE embeddings to develop an\nefficient translation system that overcomes the OOV (Out Of Vocabulary) problem\nfor low resourced morphological rich Indian languages which do not have much\ntranslation available online. We also collected corpus from different sources,\naddressed the issues with these publicly available data and refined them for\nfurther uses. We used the BLEU score for evaluating our system performance.\nExperimental results and survey confirmed that our proposed translator (24.34\nand 9.78 BLEU score) outperforms Google translator (9.40 and 5.94 BLEU score)\nrespectively.', 'Standard neural machine translation (NMT) is on the assumption that the\ndocument-level context is independent. Most existing document-level NMT\napproaches are satisfied with a smattering sense of global document-level\ninformation, while this work focuses on exploiting detailed document-level\ncontext in terms of a memory network. The capacity of the memory network that\ndetecting the most relevant part of the current sentence from memory renders a\nnatural solution to model the rich document-level context. In this work, the\nproposed document-aware memory network is implemented to enhance the\nTransformer NMT baseline. Experiments on several tasks show that the proposed\nmethod significantly improves the NMT performance over strong Transformer\nbaselines and other related studies.', 'In machine translation (MT) that involves translating between two languages\nwith significant differences in word order, determining the correct word order\nof translated words is a major challenge. The dependency parse tree of a source\nsentence can help to determine the correct word order of the translated words.\nIn this paper, we present a novel reordering approach utilizing a neural\nnetwork and dependency-based embeddings to predict whether the translations of\ntwo source words linked by a dependency relation should remain in the same\norder or should be swapped in the translated sentence. Experiments on\nChinese-to-English translation show that our approach yields a statistically\nsignificant improvement of 0.57 BLEU point on benchmark NIST test sets,\ncompared to our prior state-of-the-art statistical MT system that uses sparse\ndependency-based reordering features.', 'Compared to traditional statistical machine translation (SMT), neural machine\ntranslation (NMT) often sacrifices adequacy for the sake of fluency. We propose\na method to combine the advantages of traditional SMT and NMT by exploiting an\nexisting phrase-based SMT model to compute the phrase-based decoding cost for\nan NMT output and then using this cost to rerank the n-best NMT outputs. The\nmain challenge in implementing this approach is that NMT outputs may not be in\nthe search space of the standard phrase-based decoding algorithm, because the\nsearch space of phrase-based SMT is limited by the phrase-based translation\nrule table. We propose a soft forced decoding algorithm, which can always\nsuccessfully find a decoding path for any NMT output. We show that using the\nforced decoding cost to rerank the NMT outputs can successfully improve\ntranslation quality on four different language pairs.', 'Modern statistical machine translation (SMT) systems usually use a linear\ncombination of features to model the quality of each translation hypothesis.\nThe linear combination assumes that all the features are in a linear\nrelationship and constrains that each feature interacts with the rest features\nin an linear manner, which might limit the expressive power of the model and\nlead to a under-fit model on the current data. In this paper, we propose a\nnon-linear modeling for the quality of translation hypotheses based on neural\nnetworks, which allows more complex interaction between features. A learning\nframework is presented for training the non-linear models. We also discuss\npossible heuristics in designing the network structure which may improve the\nnon-linear learning performance. Experimental results show that with the basic\nfeatures of a hierarchical phrase-based machine translation system, our method\nproduce translations that are better than a linear model.', 'This paper presents an empirical study of two machine translation-based\napproaches for Vietnamese diacritic restoration problem, including phrase-based\nand neural-based machine translation models. This is the first work that\napplies neural-based machine translation method to this problem and gives a\nthorough comparison to the phrase-based machine translation method which is the\ncurrent state-of-the-art method for this problem. On a large dataset, the\nphrase-based approach has an accuracy of 97.32% while that of the neural-based\napproach is 96.15%. While the neural-based method has a slightly lower\naccuracy, it is about twice faster than the phrase-based method in terms of\ninference speed. Moreover, neural-based machine translation method has much\nroom for future improvement such as incorporating pre-trained word embeddings\nand collecting more training data.', 'We first observe a potential weakness of continuous vector representations of\nsymbols in neural machine translation. That is, the continuous vector\nrepresentation, or a word embedding vector, of a symbol encodes multiple\ndimensions of similarity, equivalent to encoding more than one meaning of the\nword. This has the consequence that the encoder and decoder recurrent networks\nin neural machine translation need to spend substantial amount of their\ncapacity in disambiguating source and target words based on the context which\nis defined by a source sentence. Based on this observation, in this paper we\npropose to contextualize the word embedding vectors using a nonlinear\nbag-of-words representation of the source sentence. Additionally, we propose to\nrepresent special tokens (such as numbers, proper nouns and acronyms) with\ntyped symbols to facilitate translating those words that are not well-suited to\nbe translated via continuous vectors. The experiments on En-Fr and En-De reveal\nthat the proposed approaches of contextualization and symbolization improves\nthe translation quality of neural machine translation systems significantly.', 'Recently, the development of neural machine translation (NMT) has\nsignificantly improved the translation quality of automatic machine\ntranslation. While most sentences are more accurate and fluent than\ntranslations by statistical machine translation (SMT)-based systems, in some\ncases, the NMT system produces translations that have a completely different\nmeaning. This is especially the case when rare words occur.\n  When using statistical machine translation, it has already been shown that\nsignificant gains can be achieved by simplifying the input in a preprocessing\nstep. A commonly used example is the pre-reordering approach.\n  In this work, we used phrase-based machine translation to pre-translate the\ninput into the target language. Then a neural machine translation system\ngenerates the final hypothesis using the pre-translation. Thereby, we use\neither only the output of the phrase-based machine translation (PBMT) system or\na combination of the PBMT output and the source sentence.\n  We evaluate the technique on the English to German translation task. Using\nthis approach we are able to outperform the PBMT system as well as the baseline\nneural MT system by up to 2 BLEU points. We analyzed the influence of the\nquality of the initial system on the final result.', 'What are the units of text that we want to model? From bytes to multi-word\nexpressions, text can be analyzed and generated at many granularities. Until\nrecently, most natural language processing (NLP) models operated over words,\ntreating those as discrete and atomic tokens, but starting with byte-pair\nencoding (BPE), subword-based approaches have become dominant in many areas,\nenabling small vocabularies while still allowing for fast inference. Is the end\nof the road character-level model or byte-level processing? In this survey, we\nconnect several lines of work from the pre-neural and neural era, by showing\nhow hybrid approaches of words and characters as well as subword-based\napproaches based on learned segmentation have been proposed and evaluated. We\nconclude that there is and likely will never be a silver bullet singular\nsolution for all applications and that thinking seriously about tokenization\nremains important for many applications.', 'Neural machine translation (NMT) becomes a new state-of-the-art and achieves\npromising translation results using a simple encoder-decoder neural network.\nThis neural network is trained once on the parallel corpus and the fixed\nnetwork is used to translate all the test sentences. We argue that the general\nfixed network cannot best fit the specific test sentences. In this paper, we\npropose the dynamic NMT which learns a general network as usual, and then\nfine-tunes the network for each test sentence. The fine-tune work is done on a\nsmall set of the bilingual training data that is obtained through similarity\nsearch according to the test sentence. Extensive experiments demonstrate that\nthis method can significantly improve the translation performance, especially\nwhen highly similar sentences are available.', ""We propose a simple solution to use a single Neural Machine Translation (NMT)\nmodel to translate between multiple languages. Our solution requires no change\nin the model architecture from our base system but instead introduces an\nartificial token at the beginning of the input sentence to specify the required\ntarget language. The rest of the model, which includes encoder, decoder and\nattention, remains unchanged and is shared across all languages. Using a shared\nwordpiece vocabulary, our approach enables Multilingual NMT using a single\nmodel without any increase in parameters, which is significantly simpler than\nprevious proposals for Multilingual NMT. Our method often improves the\ntranslation quality of all involved language pairs, even while keeping the\ntotal number of model parameters constant. On the WMT'14 benchmarks, a single\nmultilingual model achieves comparable performance for\nEnglish$\\rightarrow$French and surpasses state-of-the-art results for\nEnglish$\\rightarrow$German. Similarly, a single multilingual model surpasses\nstate-of-the-art results for French$\\rightarrow$English and\nGerman$\\rightarrow$English on WMT'14 and WMT'15 benchmarks respectively. On\nproduction corpora, multilingual models of up to twelve language pairs allow\nfor better translation of many individual pairs. In addition to improving the\ntranslation quality of language pairs that the model was trained with, our\nmodels can also learn to perform implicit bridging between language pairs never\nseen explicitly during training, showing that transfer learning and zero-shot\ntranslation is possible for neural translation. Finally, we show analyses that\nhints at a universal interlingua representation in our models and show some\ninteresting examples when mixing languages."", 'Differently from the traditional statistical MT that decomposes the\ntranslation task into distinct separately learned components, neural machine\ntranslation uses a single neural network to model the entire translation\nprocess. Despite neural machine translation being de-facto standard, it is\nstill not clear how NMT models acquire different competences over the course of\ntraining, and how this mirrors the different models in traditional SMT. In this\nwork, we look at the competences related to three core SMT components and find\nthat during training, NMT first focuses on learning target-side language\nmodeling, then improves translation quality approaching word-by-word\ntranslation, and finally learns more complicated reordering patterns. We show\nthat this behavior holds for several models and language pairs. Additionally,\nwe explain how such an understanding of the training process can be useful in\npractice and, as an example, show how it can be used to improve vanilla\nnon-autoregressive neural machine translation by guiding teacher model\nselection.', 'Neural machine translation (NMT) models generally adopt an encoder-decoder\narchitecture for modeling the entire translation process. The encoder\nsummarizes the representation of input sentence from scratch, which is\npotentially a problem if the sentence is ambiguous. When translating a text,\nhumans often create an initial understanding of the source sentence and then\nincrementally refine it along the translation on the target side. Starting from\nthis intuition, we propose a novel encoder-refiner-decoder framework, which\ndynamically refines the source representations based on the generated\ntarget-side information at each decoding step. Since the refining operations\nare time-consuming, we propose a strategy, leveraging the power of\nreinforcement learning models, to decide when to refine at specific decoding\nsteps. Experimental results on both Chinese-English and English-German\ntranslation tasks show that the proposed approach significantly and\nconsistently improves translation performance over the standard encoder-decoder\nframework. Furthermore, when refining strategy is applied, results still show\nreasonable improvement over the baseline without much decrease in decoding\nspeed.', 'Neural language models learn word representations, or embeddings, that\ncapture rich linguistic and conceptual information. Here we investigate the\nembeddings learned by neural machine translation models, a recently-developed\nclass of neural language model. We show that embeddings from translation models\noutperform those learned by monolingual models at tasks that require knowledge\nof both conceptual similarity and lexical-syntactic role. We further show that\nthese effects hold when translating from both English to French and English to\nGerman, and argue that the desirable properties of translation embeddings\nshould emerge largely independently of the source and target languages.\nFinally, we apply a new method for training neural translation models with very\nlarge vocabularies, and show that this vocabulary expansion algorithm results\nin minimal degradation of embedding quality. Our embedding spaces can be\nqueried in an online demo and downloaded from our web page. Overall, our\nanalyses indicate that translation-based embeddings should be used in\napplications that require concepts to be organised according to similarity\nand/or lexical function, while monolingual embeddings are better suited to\nmodelling (nonspecific) inter-word relatedness.', 'In a pipeline speech translation system, automatic speech recognition (ASR)\nsystem will transmit errors in recognition to the downstream machine\ntranslation (MT) system. A standard machine translation system is usually\ntrained on parallel corpus composed of clean text and will perform poorly on\ntext with recognition noise, a gap well known in speech translation community.\nIn this paper, we propose a training architecture which aims at making a neural\nmachine translation model more robust against speech recognition errors. Our\napproach addresses the encoder and the decoder simultaneously using adversarial\nlearning and data augmentation, respectively. Experimental results on IWSLT2018\nspeech translation task show that our approach can bridge the gap between the\nASR output and the MT input, outperforms the baseline by up to 2.83 BLEU on\nnoisy ASR output, while maintaining close performance on clean text.', 'Neural machine translation (NMT) often makes mistakes in translating\nlow-frequency content words that are essential to understanding the meaning of\nthe sentence. We propose a method to alleviate this problem by augmenting NMT\nsystems with discrete translation lexicons that efficiently encode translations\nof these low-frequency words. We describe a method to calculate the lexicon\nprobability of the next word in the translation candidate by using the\nattention vector of the NMT model to select which source word lexical\nprobabilities the model should focus on. We test two methods to combine this\nprobability with the standard NMT probability: (1) using it as a bias, and (2)\nlinear interpolation. Experiments on two corpora show an improvement of 2.0-2.3\nBLEU and 0.13-0.44 NIST score, and faster convergence time.', 'Every person speaks or writes their own flavor of their native language,\ninfluenced by a number of factors: the content they tend to talk about, their\ngender, their social status, or their geographical origin.\n  When attempting to perform Machine Translation (MT), these variations have a\nsignificant effect on how the system should perform translation, but this is\nnot captured well by standard one-size-fits-all models.\n  In this paper, we propose a simple and parameter-efficient adaptation\ntechnique that only requires adapting the bias of the output softmax to each\nparticular user of the MT system, either directly or through a factored\napproximation.\n  Experiments on TED talks in three languages demonstrate improvements in\ntranslation accuracy, and better reflection of speaker traits in the target\ntext.', 'Neural machine translation has become a major alternative to widely used\nphrase-based statistical machine translation. We notice however that much of\nresearch on neural machine translation has focused on European languages\ndespite its language agnostic nature. In this paper, we apply neural machine\ntranslation to the task of Arabic translation (Ar<->En) and compare it against\na standard phrase-based translation system. We run extensive comparison using\nvarious configurations in preprocessing Arabic script and show that the\nphrase-based and neural translation systems perform comparably to each other\nand that proper preprocessing of Arabic script has a similar effect on both of\nthe systems. We however observe that the neural machine translation\nsignificantly outperform the phrase-based system on an out-of-domain test set,\nmaking it attractive for real-world deployment.', 'System combination is an important technique for combining the hypotheses of\ndifferent machine translation systems to improve translation performance.\nAlthough early statistical approaches to system combination have been proven\neffective in analyzing the consensus between hypotheses, they suffer from the\nerror propagation problem due to the use of pipelines. While this problem has\nbeen alleviated by end-to-end training of multi-source sequence-to-sequence\nmodels recently, these neural models do not explicitly analyze the relations\nbetween hypotheses and fail to capture their agreement because the attention to\na word in a hypothesis is calculated independently, ignoring the fact that the\nword might occur in multiple hypotheses. In this work, we propose an approach\nto modeling voting for system combination in machine translation. The basic\nidea is to enable words in hypotheses from different systems to vote on words\nthat are representative and should get involved in the generation process. This\ncan be done by quantifying the influence of each voter and its preference for\neach candidate. Our approach combines the advantages of statistical and neural\nmethods since it can not only analyze the relations between hypotheses but also\nallow for end-to-end training. Experiments show that our approach is capable of\nbetter taking advantage of the consensus between hypotheses and achieves\nsignificant improvements over state-of-the-art baselines on Chinese-English and\nEnglish-German machine translation tasks.', 'For different language pairs, word-level neural machine translation (NMT)\nmodels with a fixed-size vocabulary suffer from the same problem of\nrepresenting out-of-vocabulary (OOV) words. The common practice usually\nreplaces all these rare or unknown words with a <UNK> token, which limits the\ntranslation performance to some extent. Most of recent work handled such a\nproblem by splitting words into characters or other specially extracted subword\nunits to enable open-vocabulary translation. Byte pair encoding (BPE) is one of\nthe successful attempts that has been shown extremely competitive by providing\neffective subword segmentation for NMT systems. In this paper, we extend the\nBPE style segmentation to a general unsupervised framework with three\nstatistical measures: frequency (FRQ), accessor variety (AV) and description\nlength gain (DLG). We test our approach on two translation tasks: German to\nEnglish and Chinese to English. The experimental results show that AV and DLG\nenhanced systems outperform the FRQ baseline in the frequency weighted schemes\nat different significant levels.', 'Multilingual machine translation, which translates multiple languages with a\nsingle model, has attracted much attention due to its efficiency of offline\ntraining and online serving. However, traditional multilingual translation\nusually yields inferior accuracy compared with the counterpart using individual\nmodels for each language pair, due to language diversity and model capacity\nlimitations. In this paper, we propose a distillation-based approach to boost\nthe accuracy of multilingual machine translation. Specifically, individual\nmodels are first trained and regarded as teachers, and then the multilingual\nmodel is trained to fit the training data and match the outputs of individual\nmodels simultaneously through knowledge distillation. Experiments on IWSLT, WMT\nand Ted talk translation datasets demonstrate the effectiveness of our method.\nParticularly, we show that one model is enough to handle multiple languages (up\nto 44 languages in our experiment), with comparable or even better accuracy\nthan individual models.', 'We study the application of active learning techniques to the translation of\nunbounded data streams via interactive neural machine translation. The main\nidea is to select, from an unbounded stream of source sentences, those worth to\nbe supervised by a human agent. The user will interactively translate those\nsamples. Once validated, these data is useful for adapting the neural machine\ntranslation model.\n  We propose two novel methods for selecting the samples to be validated. We\nexploit the information from the attention mechanism of a neural machine\ntranslation system. Our experiments show that the inclusion of active learning\ntechniques into this pipeline allows to reduce the effort required during the\nprocess, while increasing the quality of the translation system. Moreover, it\nenables to balance the human effort required for achieving a certain\ntranslation quality. Moreover, our neural system outperforms classical\napproaches by a large margin.', 'Neural machine translation (NMT) has recently become popular in the field of\nmachine translation. However, NMT suffers from the problem of repeating or\nmissing words in the translation. To address this problem, Tu et al. (2017)\nproposed an encoder-decoder-reconstructor framework for NMT using\nback-translation. In this method, they selected the best forward translation\nmodel in the same manner as Bahdanau et al. (2015), and then trained a\nbi-directional translation model as fine-tuning. Their experiments show that it\noffers significant improvement in BLEU scores in Chinese-English translation\ntask. We confirm that our re-implementation also shows the same tendency and\nalleviates the problem of repeating and missing words in the translation on a\nEnglish-Japanese task too. In addition, we evaluate the effectiveness of\npre-training by comparing it with a jointly-trained model of forward\ntranslation and back-translation.', 'The manuscript presents an experiment at implementation of a Machine\nTranslation system in a MapReduce model. The empirical evaluation was done\nusing fully implemented translation systems embedded into the MapReduce\nprogramming model. Two machine translation paradigms were studied: shallow\ntransfer Rule Based Machine Translation and Statistical Machine Translation.\n  The results show that the MapReduce model can be successfully used to\nincrease the throughput of a machine translation system. Furthermore this\nmethod enhances the throughput of a machine translation system without\ndecreasing the quality of the translation output.\n  Thus, the present manuscript also represents a contribution to the seminal\nwork in natural language processing, specifically Machine Translation. It first\npoints toward the importance of the definition of the metric of throughput of\ntranslation system and, second, the applicability of the machine translation\ntask to the MapReduce paradigm.', 'Recently, neural machine translation (NMT) has been extended to\nmultilinguality, that is to handle more than one translation direction with a\nsingle system. Multilingual NMT showed competitive performance against pure\nbilingual systems. Notably, in low-resource settings, it proved to work\neffectively and efficiently, thanks to shared representation space that is\nforced across languages and induces a sort of transfer-learning. Furthermore,\nmultilingual NMT enables so-called zero-shot inference across language pairs\nnever seen at training time. Despite the increasing interest in this framework,\nan in-depth analysis of what a multilingual NMT model is capable of and what it\nis not is still missing. Motivated by this, our work (i) provides a\nquantitative and comparative analysis of the translations produced by\nbilingual, multilingual and zero-shot systems; (ii) investigates the\ntranslation quality of two of the currently dominant neural architectures in\nMT, which are the Recurrent and the Transformer ones; and (iii) quantitatively\nexplores how the closeness between languages influences the zero-shot\ntranslation. Our analysis leverages multiple professional post-edits of\nautomatic translations by several different systems and focuses both on\nautomatic standard metrics (BLEU and TER) and on widely used error categories,\nwhich are lexical, morphology, and word order errors.', 'Neural Machine Translation (NMT) leverages one or more trained neural\nnetworks for the translation of phrases. Sutskever introduced a sequence to\nsequence based encoder-decoder model which became the standard for NMT based\nsystems. Attention mechanisms were later introduced to address the issues with\nthe translation of long sentences and improving overall accuracy. In this\npaper, we propose a singular model for Neural Machine Translation based on\nencoder-decoder models. Most translation models are trained as one model for\none translation. We introduce a neutral/universal model representation that can\nbe used to predict more than one language depending on the source and a\nprovided target. Secondly, we introduce an attention model by adding an overall\nlearning vector to the multiplicative model. With these two changes, by using\nthe novel universal model the number of models needed for multiple language\ntranslation applications are reduced.', 'The interest in statistical machine translation systems increases currently\ndue to political and social events in the world. A proposed Statistical Machine\nTranslation (SMT) based model that can be used to translate a sentence from the\nsource Language (English) to the target language (Arabic) automatically through\nefficiently incorporating different statistical and Natural Language Processing\n(NLP) models such as language model, alignment model, phrase based model,\nreordering model, and translation model. These models are combined to enhance\nthe performance of statistical machine translation (SMT). Many implementation\ntools have been used in this work such as Moses, Gizaa++, IRSTLM, KenLM, and\nBLEU. Based on the implementation, evaluation of this model, and comparing the\ngenerated translation with other implemented machine translation systems like\nGoogle Translate, it was proved that this proposed model has enhanced the\nresults of the statistical machine translation, and forms a reliable and\nefficient model in this field of research.', 'An effective method to generate a large number of parallel sentences for\ntraining improved neural machine translation (NMT) systems is the use of the\nback-translations of the target-side monolingual data. The standard\nback-translation method has been shown to be unable to efficiently utilize the\navailable huge amount of existing monolingual data because of the inability of\ntranslation models to differentiate between the authentic and synthetic\nparallel data during training. Tagging, or using gates, has been used to enable\ntranslation models to distinguish between synthetic and authentic data,\nimproving standard back-translation and also enabling the use of iterative\nback-translation on language pairs that underperformed using standard\nback-translation. In this work, we approach back-translation as a domain\nadaptation problem, eliminating the need for explicit tagging. In the approach\n-- \\emph{tag-less back-translation} -- the synthetic and authentic parallel\ndata are treated as out-of-domain and in-domain data respectively and, through\npre-training and fine-tuning, the translation model is shown to be able to\nlearn more efficiently from them during training. Experimental results have\nshown that the approach outperforms the standard and tagged back-translation\napproaches on low resource English-Vietnamese and English-German neural machine\ntranslation.', 'We introduce our efforts towards building a universal neural machine\ntranslation (NMT) system capable of translating between any language pair. We\nset a milestone towards this goal by building a single massively multilingual\nNMT model handling 103 languages trained on over 25 billion examples. Our\nsystem demonstrates effective transfer learning ability, significantly\nimproving translation quality of low-resource languages, while keeping\nhigh-resource language translation quality on-par with competitive bilingual\nbaselines. We provide in-depth analysis of various aspects of model building\nthat are crucial to achieving quality and practicality in universal NMT. While\nwe prototype a high-quality universal translation system, our extensive\nempirical analysis exposes issues that need to be further addressed, and we\nsuggest directions for future research.', 'We investigate the potential of attention-based neural machine translation in\nsimultaneous translation. We introduce a novel decoding algorithm, called\nsimultaneous greedy decoding, that allows an existing neural machine\ntranslation model to begin translating before a full source sentence is\nreceived. This approach is unique from previous works on simultaneous\ntranslation in that segmentation and translation are done jointly to maximize\nthe translation quality and that translating each segment is strongly\nconditioned on all the previous segments. This paper presents a first step\ntoward building a full simultaneous translation system based on neural machine\ntranslation.', 'Retrieving information from correlative paragraphs or documents to answer\nopen-domain multi-hop questions is very challenging. To deal with this\nchallenge, most of the existing works consider paragraphs as nodes in a graph\nand propose graph-based methods to retrieve them. However, in this paper, we\npoint out the intrinsic defect of such methods. Instead, we propose a new\narchitecture that models paragraphs as sequential data and considers multi-hop\ninformation retrieval as a kind of sequence labeling task. Specifically, we\ndesign a rewritable external memory to model the dependency among paragraphs.\nMoreover, a threshold gate mechanism is proposed to eliminate the distraction\nof noise paragraphs. We evaluate our method on both full wiki and distractor\nsubtask of HotpotQA, a public textual multi-hop QA dataset requiring multi-hop\ninformation retrieval. Experiments show that our method achieves significant\nimprovement over the published state-of-the-art method in retrieval and\ndownstream QA task performance.', 'Neural machine translation is a relatively new approach to statistical\nmachine translation based purely on neural networks. The neural machine\ntranslation models often consist of an encoder and a decoder. The encoder\nextracts a fixed-length representation from a variable-length input sentence,\nand the decoder generates a correct translation from this representation. In\nthis paper, we focus on analyzing the properties of the neural machine\ntranslation using two models; RNN Encoder--Decoder and a newly proposed gated\nrecursive convolutional neural network. We show that the neural machine\ntranslation performs relatively well on short sentences without unknown words,\nbut its performance degrades rapidly as the length of the sentence and the\nnumber of unknown words increase. Furthermore, we find that the proposed gated\nrecursive convolutional network learns a grammatical structure of a sentence\nautomatically.', 'Neural machine translation (NMT) has significantly improved the quality of\nautomatic translation models. One of the main challenges in current systems is\nthe translation of rare words. We present a generic approach to address this\nweakness by having external models annotate the training data as Experts, and\ncontrol the model-expert interaction with a pointer network and reinforcement\nlearning. Our experiments using phrase-based models to simulate Experts to\ncomplement neural machine translation models show that the model can be trained\nto copy the annotations into the output consistently. We demonstrate the\nbenefit of our proposed framework in outof-domain translation scenarios with\nonly lexical resources, improving more than 1.0 BLEU point in both translation\ndirections English to Spanish and German to English', 'Neural Machine Translation (NMT) is a new approach to machine translation\nthat has shown promising results that are comparable to traditional approaches.\nA significant weakness in conventional NMT systems is their inability to\ncorrectly translate very rare words: end-to-end NMTs tend to have relatively\nsmall vocabularies with a single unk symbol that represents every possible\nout-of-vocabulary (OOV) word. In this paper, we propose and implement an\neffective technique to address this problem. We train an NMT system on data\nthat is augmented by the output of a word alignment algorithm, allowing the NMT\nsystem to emit, for each OOV word in the target sentence, the position of its\ncorresponding word in the source sentence. This information is later utilized\nin a post-processing step that translates every OOV word using a dictionary.\nOur experiments on the WMT14 English to French translation task show that this\nmethod provides a substantial improvement of up to 2.8 BLEU points over an\nequivalent NMT system that does not use this technique. With 37.5 BLEU points,\nour NMT system is the first to surpass the best result achieved on a WMT14\ncontest task.', 'As the quality of machine translation rises and neural machine translation\n(NMT) is moving from sentence to document level translations, it is becoming\nincreasingly difficult to evaluate the output of translation systems.\n  We provide a test suite for WMT19 aimed at assessing discourse phenomena of\nMT systems participating in the News Translation Task. We have manually checked\nthe outputs and identified types of translation errors that are relevant to\ndocument-level translation.', 'By using a trigram model and fine-tuning a pretrained BERT model for sequence\nclassification, we show that machine translation and human translation can be\nclassified with an accuracy above chance level, which suggests that machine\ntranslation and human translation are different in a systematic way. The\nclassification accuracy of machine translation is much higher than of human\ntranslation. We show that this may be explained by the difference in lexical\ndiversity between machine translation and human translation. If machine\ntranslation has independent patterns from human translation, automatic metrics\nwhich measure the deviation of machine translation from human translation may\nconflate difference with quality. Our experiment with two different types of\nautomatic metrics shows correlation with the result of the classification task.\nTherefore, we suggest the difference in lexical diversity between machine\ntranslation and human translation be given more attention in machine\ntranslation evaluation.', ""In this paper, we evaluate the translation of negation both automatically and\nmanually, in English--German (EN--DE) and English--Chinese (EN--ZH). We show\nthat the ability of neural machine translation (NMT) models to translate\nnegation has improved with deeper and more advanced networks, although the\nperformance varies between language pairs and translation directions. The\naccuracy of manual evaluation in EN-DE, DE-EN, EN-ZH, and ZH-EN is 95.7%,\n94.8%, 93.4%, and 91.7%, respectively. In addition, we show that\nunder-translation is the most significant error type in NMT, which contrasts\nwith the more diverse error profile previously observed for statistical machine\ntranslation. To better understand the root of the under-translation of\nnegation, we study the model's information flow and training data. While our\ninformation flow analysis does not reveal any deficiencies that could be used\nto detect or fix the under-translation of negation, we find that negation is\noften rephrased during training, which could make it more difficult for the\nmodel to learn a reliable link between source and target negation. We finally\nconduct intrinsic analysis and extrinsic probing tasks on negation, showing\nthat NMT models can distinguish negation and non-negation tokens very well and\nencode a lot of information about negation in hidden states but nevertheless\nleave room for improvement."", ""We study a formalization of the grammar induction problem that models\nsentences as being generated by a compound probabilistic context-free grammar.\nIn contrast to traditional formulations which learn a single stochastic\ngrammar, our grammar's rule probabilities are modulated by a per-sentence\ncontinuous latent variable, which induces marginal dependencies beyond the\ntraditional context-free assumptions. Inference in this grammar is performed by\ncollapsed variational inference, in which an amortized variational posterior is\nplaced on the continuous variable, and the latent trees are marginalized out\nwith dynamic programming. Experiments on English and Chinese show the\neffectiveness of our approach compared to recent state-of-the-art methods when\nevaluated on unsupervised parsing."", 'We present a literature and empirical survey that critically assesses the\nstate of the art in character-level modeling for machine translation (MT).\nDespite evidence in the literature that character-level systems are comparable\nwith subword systems, they are virtually never used in competitive setups in\nWMT competitions. We empirically show that even with recent modeling\ninnovations in character-level natural language processing, character-level MT\nsystems still struggle to match their subword-based counterparts.\nCharacter-level MT systems show neither better domain robustness, nor better\nmorphological generalization, despite being often so motivated. However, we are\nable to show robustness towards source side noise and that translation quality\ndoes not degrade with increasing beam size at decoding time.', 'Evaluating translation models is a trade-off between effort and detail. On\nthe one end of the spectrum there are automatic count-based methods such as\nBLEU, on the other end linguistic evaluations by humans, which arguably are\nmore informative but also require a disproportionately high effort. To narrow\nthe spectrum, we propose a general approach on how to automatically expose\nsystematic differences between human and machine translations to human experts.\nInspired by adversarial settings, we train a neural text classifier to\ndistinguish human from machine translations. A classifier that performs and\ngeneralizes well after training should recognize systematic differences between\nthe two classes, which we uncover with neural explainability methods. Our\nproof-of-concept implementation, DiaMaT, is open source. Applied to a dataset\ntranslated by a state-of-the-art neural Transformer model, DiaMaT achieves a\nclassification accuracy of 75% and exposes meaningful differences between\nhumans and the Transformer, amidst the current discussion about human parity.', 'Media coverage has a substantial effect on the public perception of events.\nNevertheless, media outlets are often biased. One way to bias news articles is\nby altering the word choice. The automatic identification of bias by word\nchoice is challenging, primarily due to the lack of gold-standard data sets and\nhigh context dependencies. In this research project, I aim to devise data sets\nand methods to identify media bias. To achieve this, I plan to research methods\nusing natural language processing and deep learning while employing models and\nusing analysis concepts from psychology and linguistics. The first results\nindicate the effectiveness of an interdisciplinary research approach. My vision\nis to devise a system that helps news readers become aware of media coverage\ndifferences caused by bias. So far, my best performing BERT-based model is\npre-trained on a larger corpus consisting of distant labels, indicating that\ndistant supervision has the potential to become a solution for the difficult\ntask of bias detection.', ""We introduce and describe the results of a novel shared task on bandit\nlearning for machine translation. The task was organized jointly by Amazon and\nHeidelberg University for the first time at the Second Conference on Machine\nTranslation (WMT 2017). The goal of the task is to encourage research on\nlearning machine translation from weak user feedback instead of human\nreferences or post-edits. On each of a sequence of rounds, a machine\ntranslation system is required to propose a translation for an input, and\nreceives a real-valued estimate of the quality of the proposed translation for\nlearning. This paper describes the shared task's learning and evaluation setup,\nusing services hosted on Amazon Web Services (AWS), the data and evaluation\nmetrics, and the results of various machine translation architectures and\nlearning protocols.""]"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am working on a natural language processing project focused on entity recognition, which necessitates the comprehension and processing of vast amounts of text data. My goal is to integrate real-world knowledge during the training phase to bolster the model's performance across various tasks, thus bridging the gap between raw text and meaningful entities. Therefore, I am in search of models that can generate and learn embeddings for words and entities by leveraging extensive real-world knowledge bases. Naturally, the model needs to be able to process knowledge and concept in some differentiable way. Ideally, the trained model should excel in performance on entity-centric datasets and produce competitive results on other standard benchmark datasets.","['The inception of modeling contextual information using models such as BERT,\nELMo, and Flair has significantly improved representation learning for words.\nIt has also given SOTA results in almost every NLP task - Machine Translation,\nText Summarization and Named Entity Recognition, to name a few. In this work,\nin addition to using these dominant context-aware representations, we propose a\nKnowledge Aware Representation Learning (KARL) Network for Named Entity\nRecognition (NER). We discuss the challenges of using existing methods in\nincorporating world knowledge for NER and show how our proposed methods could\nbe leveraged to overcome those challenges. KARL is based on a Transformer\nEncoder that utilizes large knowledge bases represented as fact triplets,\nconverts them to a graph context, and extracts essential entity information\nresiding inside to generate contextualized triplet representation for feature\naugmentation. Experimental results show that the augmentation done using KARL\ncan considerably boost the performance of our NER system and achieve\nsignificantly better results than existing approaches in the literature on\nthree publicly available NER datasets, namely CoNLL 2003, CoNLL++, and\nOntoNotes v5. We also observe better generalization and application to a\nreal-world setting from KARL on unseen entities.', ""Contextual word representations, typically trained on unstructured, unlabeled\ntext, do not contain any explicit grounding to real world entities and are\noften unable to remember facts about those entities. We propose a general\nmethod to embed multiple knowledge bases (KBs) into large scale models, and\nthereby enhance their representations with structured, human-curated knowledge.\nFor each KB, we first use an integrated entity linker to retrieve relevant\nentity embeddings, then update contextual word representations via a form of\nword-to-entity attention. In contrast to previous approaches, the entity\nlinkers and self-supervised language modeling objective are jointly trained\nend-to-end in a multitask setting that combines a small amount of entity\nlinking supervision with a large amount of raw text. After integrating WordNet\nand a subset of Wikipedia into BERT, the knowledge enhanced BERT (KnowBert)\ndemonstrates improved perplexity, ability to recall facts as measured in a\nprobing task and downstream performance on relationship extraction, entity\ntyping, and word sense disambiguation. KnowBert's runtime is comparable to\nBERT's and it scales to large KBs."", 'Pretrained language models have shown success in many natural language\nprocessing tasks. Many works explore incorporating knowledge into language\nmodels. In the biomedical domain, experts have taken decades of effort on\nbuilding large-scale knowledge bases. For example, the Unified Medical Language\nSystem (UMLS) contains millions of entities with their synonyms and defines\nhundreds of relations among entities. Leveraging this knowledge can benefit a\nvariety of downstream tasks such as named entity recognition and relation\nextraction. To this end, we propose KeBioLM, a biomedical pretrained language\nmodel that explicitly leverages knowledge from the UMLS knowledge bases.\nSpecifically, we extract entities from PubMed abstracts and link them to UMLS.\nWe then train a knowledge-aware language model that firstly applies a text-only\nencoding layer to learn entity representation and applies a text-entity fusion\nencoding to aggregate entity representation. Besides, we add two training\nobjectives as entity detection and entity linking. Experiments on the named\nentity recognition and relation extraction from the BLURB benchmark demonstrate\nthe effectiveness of our approach. Further analysis on a collected probing\ndataset shows that our model has better ability to model medical knowledge.', 'Learned vector representations of words are useful tools for many information\nretrieval and natural language processing tasks due to their ability to capture\nlexical semantics. However, while many such tasks involve or even rely on named\nentities as central components, popular word embedding models have so far\nfailed to include entities as first-class citizens. While it seems intuitive\nthat annotating named entities in the training corpus should result in more\nintelligent word features for downstream tasks, performance issues arise when\npopular embedding approaches are naively applied to entity annotated corpora.\nNot only are the resulting entity embeddings less useful than expected, but one\nalso finds that the performance of the non-entity word embeddings degrades in\ncomparison to those trained on the raw, unannotated corpus. In this paper, we\ninvestigate approaches to jointly train word and entity embeddings on a large\ncorpus with automatically annotated and linked entities. We discuss two\ndistinct approaches to the generation of such embeddings, namely the training\nof state-of-the-art embeddings on raw-text and annotated versions of the\ncorpus, as well as node embeddings of a co-occurrence graph representation of\nthe annotated corpus. We compare the performance of annotated embeddings and\nclassical word embeddings on a variety of word similarity, analogy, and\nclustering evaluation tasks, and investigate their performance in\nentity-specific tasks. Our findings show that it takes more than training\npopular word embedding models on an annotated corpus to create entity\nembeddings with acceptable performance on common test cases. Based on these\nresults, we discuss how and when node embeddings of the co-occurrence graph\nrepresentation of the text can restore the performance.', 'This paper addresses the problem of mapping natural language text to\nknowledge base entities. The mapping process is approached as a composition of\na phrase or a sentence into a point in a multi-dimensional entity space\nobtained from a knowledge graph. The compositional model is an LSTM equipped\nwith a dynamic disambiguation mechanism on the input word embeddings (a\nMulti-Sense LSTM), addressing polysemy issues. Further, the knowledge base\nspace is prepared by collecting random walks from a graph enhanced with textual\nfeatures, which act as a set of semantic bridges between text and knowledge\nbase entities. The ideas of this work are demonstrated on large-scale\ntext-to-entity mapping and entity classification tasks, with state of the art\nresults.', 'Existing state of the art neural entity linking models employ attention-based\nbag-of-words context model and pre-trained entity embeddings bootstrapped from\nword embeddings to assess topic level context compatibility. However, the\nlatent entity type information in the immediate context of the mention is\nneglected, which causes the models often link mentions to incorrect entities\nwith incorrect type. To tackle this problem, we propose to inject latent entity\ntype information into the entity embeddings based on pre-trained BERT. In\naddition, we integrate a BERT-based entity similarity score into the local\ncontext model of a state-of-the-art model to better capture latent entity type\ninformation. Our model significantly outperforms the state-of-the-art entity\nlinking models on standard benchmark (AIDA-CoNLL). Detailed experiment analysis\ndemonstrates that our model corrects most of the type errors produced by the\ndirect baseline.', 'Knowledge graphs (KGs) contain rich information about world knowledge,\nentities and relations. Thus, they can be great supplements to existing\npre-trained language models. However, it remains a challenge to efficiently\nintegrate information from KG into language modeling. And the understanding of\na knowledge graph requires related context. We propose a novel joint\npre-training framework, JAKET, to model both the knowledge graph and language.\nThe knowledge module and language module provide essential information to\nmutually assist each other: the knowledge module produces embeddings for\nentities in text while the language module generates context-aware initial\nembeddings for entities and relations in the graph. Our design enables the\npre-trained model to easily adapt to unseen knowledge graphs in new domains.\nExperimental results on several knowledge-aware NLP tasks show that our\nproposed framework achieves superior performance by effectively leveraging\nknowledge in language understanding.', 'Due to the lack of structured knowledge applied in learning distributed\nrepresentation of cate- gories, existing work cannot incorporate category\nhierarchies into entity information. We propose a framework that embeds\nentities and categories into a semantic space by integrating structured\nknowledge and taxonomy hierarchy from large knowledge bases. The framework\nallows to com- pute meaningful semantic relatedness between entities and\ncategories. Our framework can han- dle both single-word concepts and\nmultiple-word concepts with superior performance on concept categorization and\nyield state of the art results on dataless hierarchical classification.', 'We describe a neural network model that jointly learns distributed\nrepresentations of texts and knowledge base (KB) entities. Given a text in the\nKB, we train our proposed model to predict entities that are relevant to the\ntext. Our model is designed to be generic with the ability to address various\nNLP tasks with ease. We train the model using a large corpus of texts and their\nentity annotations extracted from Wikipedia. We evaluated the model on three\nimportant NLP tasks (i.e., sentence textual similarity, entity linking, and\nfactoid question answering) involving both unsupervised and supervised\nsettings. As a result, we achieved state-of-the-art results on all three of\nthese tasks. Our code and trained models are publicly available for further\nacademic research.', 'Information Extraction (IE) from text refers to the task of extracting\nstructured knowledge from unstructured text. The task typically consists of a\nseries of sub-tasks such as Named Entity Recognition and Relation Extraction.\nSourcing entity and relation type specific training data is a major bottleneck\nin domains with limited resources such as biomedicine. In this work we present\na slot filling approach to the task of biomedical IE, effectively replacing the\nneed for entity and relation-specific training data, allowing us to deal with\nzero-shot settings. We follow the recently proposed paradigm of coupling a\nTranformer-based bi-encoder, Dense Passage Retrieval, with a Transformer-based\nreading comprehension model to extract relations from biomedical text. We\nassemble a biomedical slot filling dataset for both retrieval and reading\ncomprehension and conduct a series of experiments demonstrating that our\napproach outperforms a number of simpler baselines. We also evaluate our\napproach end-to-end for standard as well as zero-shot settings. Our work\nprovides a fresh perspective on how to solve biomedical IE tasks, in the\nabsence of relevant training data. Our code, models and datasets are available\nat https://github.com/ypapanik/biomedical-slot-filling.', 'The challenge of recognizing named entities in a given text has been a very\ndynamic field in recent years. This is due to the advances in neural network\narchitectures, increase of computing power and the availability of diverse\nlabeled datasets, which deliver pre-trained, highly accurate models. These\ntasks are generally focused on tagging common entities, but domain-specific\nuse-cases require tagging custom entities which are not part of the pre-trained\nmodels. This can be solved by either fine-tuning the pre-trained models, or by\ntraining custom models. The main challenge lies in obtaining reliable labeled\ntraining and test datasets, and manual labeling would be a highly tedious task.\n  In this paper we present PharmKE, a text analysis platform focused on the\npharmaceutical domain, which applies deep learning through several stages for\nthorough semantic analysis of pharmaceutical articles. It performs text\nclassification using state-of-the-art transfer learning models, and thoroughly\nintegrates the results obtained through a proposed methodology. The methodology\nis used to create accurately labeled training and test datasets, which are then\nused to train models for custom entity labeling tasks, centered on the\npharmaceutical domain. The obtained results are compared to the fine-tuned BERT\nand BioBERT models trained on the same dataset. Additionally, the PharmKE\nplatform integrates the results obtained from named entity recognition tasks to\nresolve co-references of entities and analyze the semantic relations in every\nsentence, thus setting up a baseline for additional text analysis tasks, such\nas question answering and fact extraction. The recognized entities are also\nused to expand the knowledge graph generated by DBpedia Spotlight for a given\npharmaceutical text.', 'Background: Finding biomedical named entities is one of the most essential\ntasks in biomedical text mining. Recently, deep learning-based approaches have\nbeen applied to biomedical named entity recognition (BioNER) and showed\npromising results. However, as deep learning approaches need an abundant amount\nof training data, a lack of data can hinder performance. BioNER datasets are\nscarce resources and each dataset covers only a small subset of entity types.\nFurthermore, many bio entities are polysemous, which is one of the major\nobstacles in named entity recognition. Results: To address the lack of data and\nthe entity type misclassification problem, we propose CollaboNet which utilizes\na combination of multiple NER models. In CollaboNet, models trained on a\ndifferent dataset are connected to each other so that a target model obtains\ninformation from other collaborator models to reduce false positives. Every\nmodel is an expert on their target entity type and takes turns serving as a\ntarget and a collaborator model during training time. The experimental results\nshow that CollaboNet can be used to greatly reduce the number of false\npositives and misclassified entities including polysemous words. CollaboNet\nachieved state-of-the-art performance in terms of precision, recall and F1\nscore. Conclusions: We demonstrated the benefits of combining multiple models\nfor BioNER. Our model has successfully reduced the number of misclassified\nentities and improved the performance by leveraging multiple datasets annotated\nfor different entity types. Given the state-of-the-art performance of our\nmodel, we believe that CollaboNet can improve the accuracy of downstream\nbiomedical text mining applications such as bio-entity relation extraction.', 'This study proposes a Neural Attentive Bag-of-Entities model, which is a\nneural network model that performs text classification using entities in a\nknowledge base. Entities provide unambiguous and relevant semantic signals that\nare beneficial for capturing semantics in texts. We combine simple high-recall\nentity detection based on a dictionary, to detect entities in a document, with\na novel neural attention mechanism that enables the model to focus on a small\nnumber of unambiguous and relevant entities. We tested the effectiveness of our\nmodel using two standard text classification datasets (i.e., the 20 Newsgroups\nand R8 datasets) and a popular factoid question answering dataset based on a\ntrivia quiz game. As a result, our model achieved state-of-the-art results on\nall datasets. The source code of the proposed model is available online at\nhttps://github.com/wikipedia2vec/wikipedia2vec.', 'Text representations using neural word embeddings have proven effective in\nmany NLP applications. Recent researches adapt the traditional word embedding\nmodels to learn vectors of multiword expressions (concepts/entities). However,\nthese methods are limited to textual knowledge bases (e.g., Wikipedia). In this\npaper, we propose a novel and simple technique for integrating the knowledge\nabout concepts from two large scale knowledge bases of different structure\n(Wikipedia and Probase) in order to learn concept representations. We adapt the\nefficient skip-gram model to seamlessly learn from the knowledge in Wikipedia\ntext and Probase concept graph. We evaluate our concept embedding models on two\ntasks: (1) analogical reasoning, where we achieve a state-of-the-art\nperformance of 91% on semantic analogies, (2) concept categorization, where we\nachieve a state-of-the-art performance on two benchmark datasets achieving\ncategorization accuracy of 100% on one and 98% on the other. Additionally, we\npresent a case study to evaluate our model on unsupervised argument type\nidentification for neural semantic parsing. We demonstrate the competitive\naccuracy of our unsupervised method and its ability to better generalize to out\nof vocabulary entity mentions compared to the tedious and error prone methods\nwhich depend on gazetteers and regular expressions.', 'Joint representation learning of text and knowledge within a unified semantic\nspace enables us to perform knowledge graph completion more accurately. In this\nwork, we propose a novel framework to embed words, entities and relations into\nthe same continuous vector space. In this model, both entity and relation\nembeddings are learned by taking knowledge graph and plain text into\nconsideration. In experiments, we evaluate the joint learning model on three\ntasks including entity prediction, relation prediction and relation\nclassification from text. The experiment results show that our model can\nsignificantly and consistently improve the performance on the three tasks as\ncompared with other baselines.', 'Extracting entities and relations for types of interest from text is\nimportant for understanding massive text corpora. Traditionally, systems of\nentity relation extraction have relied on human-annotated corpora for training\nand adopted an incremental pipeline. Such systems require additional human\nexpertise to be ported to a new domain, and are vulnerable to errors cascading\ndown the pipeline. In this paper, we investigate joint extraction of typed\nentities and relations with labeled data heuristically obtained from knowledge\nbases (i.e., distant supervision). As our algorithm for type labeling via\ndistant supervision is context-agnostic, noisy training data poses unique\nchallenges for the task. We propose a novel domain-independent framework,\ncalled CoType, that runs a data-driven text segmentation algorithm to extract\nentity mentions, and jointly embeds entity mentions, relation mentions, text\nfeatures and type labels into two low-dimensional spaces (for entity and\nrelation mentions respectively), where, in each space, objects whose types are\nclose will also have similar representations. CoType, then using these learned\nembeddings, estimates the types of test (unlinkable) mentions. We formulate a\njoint optimization problem to learn embeddings from text corpora and knowledge\nbases, adopting a novel partial-label loss function for noisy labeled data and\nintroducing an object ""translation"" function to capture the cross-constraints\nof entities and relations on each other. Experiments on three public datasets\ndemonstrate the effectiveness of CoType across different domains (e.g., news,\nbiomedical), with an average of 25% improvement in F1 score compared to the\nnext best method.', ""In standard methodology for natural language processing, entities in text are\ntypically embedded in dense vector spaces with pre-trained models. The\nembeddings produced this way are effective when fed into downstream models, but\nthey require end-task fine-tuning and are fundamentally difficult to interpret.\nIn this paper, we present an approach to creating entity representations that\nare human readable and achieve high performance on entity-related tasks out of\nthe box. Our representations are vectors whose values correspond to posterior\nprobabilities over fine-grained entity types, indicating the confidence of a\ntyping model's decision that the entity belongs to the corresponding type. We\nobtain these representations using a fine-grained entity typing model, trained\neither on supervised ultra-fine entity typing data (Choi et al. 2018) or\ndistantly-supervised examples from Wikipedia. On entity probing tasks involving\nrecognizing entity identity, our embeddings used in parameter-free downstream\nmodels achieve competitive performance with ELMo- and BERT-based embeddings in\ntrained models. We also show that it is possible to reduce the size of our type\nset in a learning-based way for particular domains. Finally, we show that these\nembeddings can be post-hoc modified through a small number of rules to\nincorporate domain knowledge and improve performance."", 'Named entity recognition (NER) is one of the tasks in natural language\nprocessing that can greatly benefit from the use of external knowledge sources.\nWe propose a named entity recognition framework composed of knowledge-based\nfeature extractors and a deep learning model including contextual word\nembeddings, long short-term memory (LSTM) layers and conditional random fields\n(CRF) inference layer. We use an entity linking module to integrate our system\nwith Wikipedia. The combination of effective neural architecture and external\nresources allows us to obtain state-of-the-art results on recognition of Polish\nproper names. We evaluate our model on data from PolEval 2018 NER challenge on\nwhich it outperforms other methods, reducing the error rate by 22.4% compared\nto the winning solution. Our work shows that combining neural NER model and\nentity linking model with a knowledge base is more effective in recognizing\nnamed entities than using NER model alone.', 'Entities are at the center of how we represent and aggregate knowledge. For\ninstance, Encyclopedias such as Wikipedia are structured by entities (e.g., one\nper Wikipedia article). The ability to retrieve such entities given a query is\nfundamental for knowledge-intensive tasks such as entity linking and\nopen-domain question answering. Current approaches can be understood as\nclassifiers among atomic labels, one for each entity. Their weight vectors are\ndense entity representations produced by encoding entity meta information such\nas their descriptions. This approach has several shortcomings: (i) context and\nentity affinity is mainly captured through a vector dot product, potentially\nmissing fine-grained interactions; (ii) a large memory footprint is needed to\nstore dense representations when considering large entity sets; (iii) an\nappropriately hard set of negative data has to be subsampled at training time.\nIn this work, we propose GENRE, the first system that retrieves entities by\ngenerating their unique names, left to right, token-by-token in an\nautoregressive fashion. This mitigates the aforementioned technical issues\nsince: (i) the autoregressive formulation directly captures relations between\ncontext and entity name, effectively cross encoding both; (ii) the memory\nfootprint is greatly reduced because the parameters of our encoder-decoder\narchitecture scale with vocabulary size, not entity count; (iii) the softmax\nloss is computed without subsampling negative data. We experiment with more\nthan 20 datasets on entity disambiguation, end-to-end entity linking and\ndocument retrieval tasks, achieving new state-of-the-art or very competitive\nresults while using a tiny fraction of the memory footprint of competing\nsystems. Finally, we demonstrate that new entities can be added by simply\nspecifying their names. Code and pre-trained models at\nhttps://github.com/facebookresearch/GENRE.', 'Deep neural network models have helped named entity (NE) recognition achieve\namazing performance without handcrafting features. However, existing systems\nrequire large amounts of human annotated training data. Efforts have been made\nto replace human annotations with external knowledge (e.g., NE dictionary,\npart-of-speech tags), while it is another challenge to obtain such effective\nresources. In this work, we propose a fully unsupervised NE recognition model\nwhich only needs to take informative clues from pre-trained word embeddings. We\nfirst apply Gaussian Hidden Markov Model and Deep Autoencoding Gaussian Mixture\nModel on word embeddings for entity span detection and type prediction, and\nthen further design an instance selector based on reinforcement learning to\ndistinguish positive sentences from noisy sentences and refine these\ncoarse-grained annotations through neural networks. Extensive experiments on\nCoNLL benchmark datasets demonstrate that our proposed light NE recognition\nmodel achieves remarkable performance without using any annotated lexicon or\ncorpus.', 'Much of biomedical and healthcare data is encoded in discrete, symbolic form\nsuch as text and medical codes. There is a wealth of expert-curated biomedical\ndomain knowledge stored in knowledge bases and ontologies, but the lack of\nreliable methods for learning knowledge representation has limited their\nusefulness in machine learning applications. While text-based representation\nlearning has significantly improved in recent years through advances in natural\nlanguage processing, attempts to learn biomedical concept embeddings so far\nhave been lacking. A recent family of models called knowledge graph embeddings\nhave shown promising results on general domain knowledge graphs, and we explore\ntheir capabilities in the biomedical domain. We train several state-of-the-art\nknowledge graph embedding models on the SNOMED-CT knowledge graph, provide a\nbenchmark with comparison to existing methods and in-depth discussion on best\npractices, and make a case for the importance of leveraging the\nmulti-relational nature of knowledge graphs for learning biomedical knowledge\nrepresentation. The embeddings, code, and materials will be made available to\nthe communitY.', 'Knowledge bases (KBs) and text often contain complementary knowledge: KBs\nstore structured knowledge that can support long range reasoning, while text\nstores more comprehensive and timely knowledge in an unstructured way.\nSeparately embedding the individual knowledge sources into vector spaces has\ndemonstrated tremendous successes in encoding the respective knowledge, but how\nto jointly embed and reason with both knowledge sources to fully leverage the\ncomplementary information is still largely an open problem. We conduct a\nlarge-scale, systematic investigation of aligning KB and text embeddings for\njoint reasoning. We set up a novel evaluation framework with two evaluation\ntasks, few-shot link prediction and analogical reasoning, and evaluate an array\nof KB-text embedding alignment methods. We also demonstrate how such alignment\ncan infuse textual information into KB embeddings for more accurate link\nprediction on emerging entities and events, using COVID-19 as a case study.', ""A typical architecture for end-to-end entity linking systems consists of\nthree steps: mention detection, candidate generation and entity disambiguation.\nIn this study we investigate the following questions: (a) Can all those steps\nbe learned jointly with a model for contextualized text-representations, i.e.\nBERT (Devlin et al., 2019)? (b) How much entity knowledge is already contained\nin pretrained BERT? (c) Does additional entity knowledge improve BERT's\nperformance in downstream tasks? To this end, we propose an extreme\nsimplification of the entity linking setup that works surprisingly well: simply\ncast it as a per token classification over the entire entity vocabulary (over\n700K classes in our case). We show on an entity linking benchmark that (i) this\nmodel improves the entity representations over plain BERT, (ii) that it\noutperforms entity linking architectures that optimize the tasks separately and\n(iii) that it only comes second to the current state-of-the-art that does\nmention detection and entity disambiguation jointly. Additionally, we\ninvestigate the usefulness of entity-aware token-representations in the\ntext-understanding benchmark GLUE, as well as the question answering benchmarks\nSQUAD V2 and SWAG and also the EN-DE WMT14 machine translation benchmark. To\nour surprise, we find that most of those benchmarks do not benefit from\nadditional entity knowledge, except for a task with very small training data,\nthe RTE task in GLUE, which improves by 2%."", ""Entity synonyms discovery is crucial for entity-leveraging applications.\nHowever, existing studies suffer from several critical issues: (1) the input\nmentions may be out-of-vocabulary (OOV) and may come from a different semantic\nspace of the entities; (2) the connection between mentions and entities may be\nhidden and cannot be established by surface matching; and (3) some entities\nrarely appear due to the long-tail effect. To tackle these challenges, we\nfacilitate knowledge graphs and propose a novel entity synonyms discovery\nframework, named \\emph{KGSynNet}. Specifically, we pre-train subword embeddings\nfor mentions and entities using a large-scale domain-specific corpus while\nlearning the knowledge embeddings of entities via a joint TransC-TransE model.\nMore importantly, to obtain a comprehensive representation of entities, we\nemploy a specifically designed \\emph{fusion gate} to adaptively absorb the\nentities' knowledge information into their semantic features. We conduct\nextensive experiments to demonstrate the effectiveness of our \\emph{KGSynNet}\nin leveraging the knowledge graph. The experimental results show that the\n\\emph{KGSynNet} improves the state-of-the-art methods by 14.7\\% in terms of\nhits@3 in the offline evaluation and outperforms the BERT model by 8.3\\% in the\npositive feedback rate of an online A/B test on the entity linking module of a\nquestion answering system."", 'Traditional language models are unable to efficiently model entity names\nobserved in text. All but the most popular named entities appear infrequently\nin text providing insufficient context. Recent efforts have recognized that\ncontext can be generalized between entity names that share the same type (e.g.,\n\\emph{person} or \\emph{location}) and have equipped language models with access\nto an external knowledge base (KB). Our Knowledge-Augmented Language Model\n(KALM) continues this line of work by augmenting a traditional model with a KB.\nUnlike previous methods, however, we train with an end-to-end predictive\nobjective optimizing the perplexity of text. We do not require any additional\ninformation such as named entity tags. In addition to improving language\nmodeling performance, KALM learns to recognize named entities in an entirely\nunsupervised way by using entity type information latent in the model. On a\nNamed Entity Recognition (NER) task, KALM achieves performance comparable with\nstate-of-the-art supervised models. Our work demonstrates that named entities\n(and possibly other types of world knowledge) can be modeled successfully using\npredictive learning and training on large corpora of text without any\nadditional information.', 'In this paper, we describe TextEnt, a neural network model that learns\ndistributed representations of entities and documents directly from a knowledge\nbase (KB). Given a document in a KB consisting of words and entity annotations,\nwe train our model to predict the entity that the document describes and map\nthe document and its target entity close to each other in a continuous vector\nspace. Our model is trained using a large number of documents extracted from\nWikipedia. The performance of the proposed model is evaluated using two tasks,\nnamely fine-grained entity typing and multiclass text classification. The\nresults demonstrate that our model achieves state-of-the-art performance on\nboth tasks. The code and the trained representations are made available online\nfor further academic research.', 'We propose a novel deep learning model for joint document-level entity\ndisambiguation, which leverages learned neural representations. Key components\nare entity embeddings, a neural attention mechanism over local context windows,\nand a differentiable joint inference stage for disambiguation. Our approach\nthereby combines benefits of deep learning with more traditional approaches\nsuch as graphical models and probabilistic mention-entity maps. Extensive\nexperiments show that we are able to obtain competitive or state-of-the-art\naccuracy at moderate computational costs.', 'Neural language models are a powerful tool to embed words into semantic\nvector spaces. However, learning such models generally relies on the\navailability of abundant and diverse training examples. In highly specialised\ndomains this requirement may not be met due to difficulties in obtaining a\nlarge corpus, or the limited range of expression in average use. Such domains\nmay encode prior knowledge about entities in a knowledge base or ontology. We\npropose a generative model which integrates evidence from diverse data sources,\nenabling the sharing of semantic information. We achieve this by generalising\nthe concept of co-occurrence from distributional semantics to include other\nrelationships between entities or words, which we model as affine\ntransformations on the embedding space. We demonstrate the effectiveness of\nthis approach by outperforming recent models on a link prediction task and\ndemonstrating its ability to profit from partially or fully unobserved data\ntraining labels. We further demonstrate the usefulness of learning from\ndifferent data sources with overlapping vocabularies.', 'Information extraction is an important task in NLP, enabling the automatic\nextraction of data for relational database filling. Historically, research and\ndata was produced for English text, followed in subsequent years by datasets in\nArabic, Chinese (ACE/OntoNotes), Dutch, Spanish, German (CoNLL evaluations),\nand many others. The natural tendency has been to treat each language as a\ndifferent dataset and build optimized models for each. In this paper we\ninvestigate a single Named Entity Recognition model, based on a multilingual\nBERT, that is trained jointly on many languages simultaneously, and is able to\ndecode these languages with better accuracy than models trained only on one\nlanguage. To improve the initial model, we study the use of regularization\nstrategies such as multitask learning and partial gradient updates. In addition\nto being a single model that can tackle multiple languages (including code\nswitch), the model could be used to make zero-shot predictions on a new\nlanguage, even ones for which training data is not available, out of the box.\nThe results show that this model not only performs competitively with\nmonolingual models, but it also achieves state-of-the-art results on the\nCoNLL02 Dutch and Spanish datasets, OntoNotes Arabic and Chinese datasets.\nMoreover, it performs reasonably well on unseen languages, achieving\nstate-of-the-art for zero-shot on three CoNLL languages.', ""Named entity recognition (NER) and relation extraction (RE) are two important\ntasks in information extraction and retrieval (IE \\& IR). Recent work has\ndemonstrated that it is beneficial to learn these tasks jointly, which avoids\nthe propagation of error inherent in pipeline-based systems and improves\nperformance. However, state-of-the-art joint models typically rely on external\nnatural language processing (NLP) tools, such as dependency parsers, limiting\ntheir usefulness to domains (e.g. news) where those tools perform well. The few\nneural, end-to-end models that have been proposed are trained almost completely\nfrom scratch. In this paper, we propose a neural, end-to-end model for jointly\nextracting entities and their relations which does not rely on external NLP\ntools and which integrates a large, pre-trained language model. Because the\nbulk of our model's parameters are pre-trained and we eschew recurrence for\nself-attention, our model is fast to train. On 5 datasets across 3 domains, our\nmodel matches or exceeds state-of-the-art performance, sometimes by a large\nmargin."", 'An image caption should fluently present the essential information in a given\nimage, including informative, fine-grained entity mentions and the manner in\nwhich these entities interact. However, current captioning models are usually\ntrained to generate captions that only contain common object names, thus\nfalling short on an important ""informativeness"" dimension. We present a\nmechanism for integrating image information together with fine-grained labels\n(assumed to be generated by some upstream models) into a caption that describes\nthe image in a fluent and informative manner. We introduce a multimodal,\nmulti-encoder model based on Transformer that ingests both image features and\nmultiple sources of entity labels. We demonstrate that we can learn to control\nthe appearance of these entity labels in the output, resulting in captions that\nare both fluent and informative.', ""The amount of archaeological literature is growing rapidly. Until recently,\nthese data were only accessible through metadata search. We implemented a text\nretrieval engine for a large archaeological text collection ($\\sim 658$ Million\nwords). In archaeological IR, domain-specific entities such as locations, time\nperiods, and artefacts, play a central role. This motivated the development of\na named entity recognition (NER) model to annotate the full collection with\narchaeological named entities. In this paper, we present ArcheoBERTje, a BERT\nmodel pre-trained on Dutch archaeological texts. We compare the model's quality\nand output on a Named Entity Recognition task to a generic multilingual model\nand a generic Dutch model. We also investigate ensemble methods for combining\nmultiple BERT models, and combining the best BERT model with a domain thesaurus\nusing Conditional Random Fields (CRF). We find that ArcheoBERTje outperforms\nboth the multilingual and Dutch model significantly with a smaller standard\ndeviation between runs, reaching an average F1 score of 0.735. The model also\noutperforms ensemble methods combining the three models. Combining ArcheoBERTje\npredictions and explicit domain knowledge from the thesaurus did not increase\nthe F1 score. We quantitatively and qualitatively analyse the differences\nbetween the vocabulary and output of the BERT models on the full collection and\nprovide some valuable insights in the effect of fine-tuning for specific\ndomains. Our results indicate that for a highly specific text domain such as\narchaeology, further pre-training on domain-specific data increases the model's\nquality on NER by a much larger margin than shown for other domains in the\nliterature, and that domain-specific pre-training makes the addition of domain\nknowledge from a thesaurus unnecessary."", 'Most benchmark datasets targeting commonsense reasoning focus on everyday\nscenarios: physical knowledge like knowing that you could fill a cup under a\nwaterfall [Talmor et al., 2019], social knowledge like bumping into someone is\nawkward [Sap et al., 2019], and other generic situations. However, there is a\nrich space of commonsense inferences anchored to knowledge about specific\nentities: for example, deciding the truthfulness of a claim ""Harry Potter can\nteach classes on how to fly on a broomstick."" Can models learn to combine\nentity knowledge with commonsense reasoning in this fashion? We introduce\nCREAK, a testbed for commonsense reasoning about entity knowledge, bridging\nfact-checking about entities (Harry Potter is a wizard and is skilled at riding\na broomstick) with commonsense inferences (if you\'re good at a skill you can\nteach others how to do it). Our dataset consists of 13k human-authored English\nclaims about entities that are either true or false, in addition to a small\ncontrast set. Crowdworkers can easily come up with these statements and human\nperformance on the dataset is high (high 90s); we argue that models should be\nable to blend entity knowledge and commonsense reasoning to do well here. In\nour experiments, we focus on the closed-book setting and observe that a\nbaseline model finetuned on existing fact verification benchmark struggles on\nCREAK. Training a model on CREAK improves accuracy by a substantial margin, but\nstill falls short of human performance. Our benchmark provides a unique probe\ninto natural language understanding models, testing both its ability to\nretrieve facts (e.g., who teaches at the University of Chicago?) and unstated\ncommonsense knowledge (e.g., butlers do not yell at guests).', ""Entity Linking (EL) seeks to align entity mentions in text to entries in a\nknowledge-base and is usually comprised of two phases: candidate generation and\ncandidate ranking. While most methods focus on the latter, it is the candidate\ngeneration phase that sets an upper bound to both time and accuracy performance\nof the overall EL system. This work's contribution is a significant improvement\nin candidate generation which thus raises the performance threshold for EL, by\ngenerating candidates that include the gold entity in the least candidate set\n(top-K). We propose a simple approach that efficiently embeds mention-entity\npairs in dense space through a BERT-based bi-encoder. Specifically, we extend\n(Wu et al., 2020) by introducing a new pooling function and incorporating\nentity type side-information. We achieve a new state-of-the-art 84.28% accuracy\non top-50 candidates on the Zeshel dataset, compared to the previous 82.06% on\nthe top-64 of (Wu et al., 2020). We report the results from extensive\nexperimentation using our proposed model on both seen and unseen entity\ndatasets. Our results suggest that our method could be a useful complement to\nexisting EL approaches."", 'Entity embeddings, which represent different aspects of each entity with a\nsingle vector like word embeddings, are a key component of neural entity\nlinking models. Existing entity embeddings are learned from canonical Wikipedia\narticles and local contexts surrounding target entities. Such entity embeddings\nare effective, but too distinctive for linking models to learn contextual\ncommonality. We propose a simple yet effective method, FGS2EE, to inject\nfine-grained semantic information into entity embeddings to reduce the\ndistinctiveness and facilitate the learning of contextual commonality. FGS2EE\nfirst uses the embeddings of semantic type words to generate semantic\nembeddings, and then combines them with existing entity embeddings through\nlinear aggregation. Extensive experiments show the effectiveness of such\nembeddings. Based on our entity embeddings, we achieved new sate-of-the-art\nperformance on entity linking.', 'Knowledge is captured in the form of entities and their relationships and\nstored in knowledge graphs. Knowledge graphs enhance the capabilities of\napplications in many different areas including Web search, recommendation, and\nnatural language understanding. This is mainly because, entities enable\nmachines to understand things that go beyond simple tokens. Many modern\nalgorithms use learned entity embeddings from these structured representations.\nHowever, building a knowledge graph takes time and effort, hence very costly\nand nontrivial. On the other hand, many Web sources describe entities in some\nstructured format and therefore, finding ways to get them into useful entity\nknowledge is advantageous. We propose an approach that processes entity centric\ntextual knowledge sources to learn entity embeddings and in turn avoids the\nneed for a traditional knowledge graph. We first extract triples into the new\nrepresentation format that does not use traditional complex triple extraction\nmethods defined by pre-determined relationship labels. Then we learn entity\nembeddings through this new type of triples. We show that the embeddings\nlearned from our approach are: (i) high quality and comparable to a known\nknowledge graph-based embeddings and can be used to improve them further, (ii)\nbetter than a contextual language model-based entity embeddings, and (iii) easy\nto compute and versatile in domain-specific applications where a knowledge\ngraph is not readily available', 'A growing number of applications users daily interact with have to operate in\n(near) real-time: chatbots, digital companions, knowledge work support systems\n-- just to name a few. To perform the services desired by the user, these\nsystems have to analyze user activity logs or explicit user input extremely\nfast. In particular, text content (e.g. in form of text snippets) needs to be\nprocessed in an information extraction task. Regarding the aforementioned\ntemporal requirements, this has to be accomplished in just a few milliseconds,\nwhich limits the number of methods that can be applied. Practically, only very\nfast methods remain, which on the other hand deliver worse results than slower\nbut more sophisticated Natural Language Processing (NLP) pipelines. In this\npaper, we investigate and propose methods for real-time capable Named Entity\nRecognition (NER). As a first improvement step we address are word variations\ninduced by inflection, for example present in the German language. Our approach\nis ontology-based and makes use of several language information sources like\nWiktionary. We evaluated it using the German Wikipedia (about 9.4B characters),\nfor which the whole NER process took considerably less than an hour. Since\nprecision and recall are higher than with comparably fast methods, we conclude\nthat the quality gap between high speed methods and sophisticated NLP pipelines\ncan be narrowed a bit more without losing too much runtime performance.']","['Over the years, Twitter has become one of the largest communication platforms\nproviding key data to various applications such as brand monitoring, trend\ndetection, among others. Entity linking is one of the major tasks in natural\nlanguage understanding from tweets and it associates entity mentions in text to\ncorresponding entries in knowledge bases in order to provide unambiguous\ninterpretation and additional con- text. State-of-the-art techniques have\nfocused on linking explicitly mentioned entities in tweets with reasonable\nsuccess. However, we argue that in addition to explicit mentions i.e. The movie\nGravity was more ex- pensive than the mars orbiter mission entities (movie\nGravity) can also be mentioned implicitly i.e. This new space movie is crazy.\nyou must watch it!. This paper introduces the problem of implicit entity\nlinking in tweets. We propose an approach that models the entities by\nexploiting their factual and contextual knowledge. We demonstrate how to use\nthese models to perform implicit entity linking on a ground truth dataset with\n397 tweets from two domains, namely, Movie and Book. Specifically, we show: 1)\nthe importance of linking implicit entities and its value addition to the\nstandard entity linking task, and 2) the importance of exploiting contextual\nknowledge associated with an entity for linking their implicit mentions. We\nalso make the ground truth dataset publicly available to foster the research in\nthis new research area.', 'Named Entity Recognition (NER) is a fundamental task in Natural Language\nProcessing, concerned with identifying spans of text expressing references to\nentities. NER research is often focused on flat entities only (flat NER),\nignoring the fact that entity references can be nested, as in [Bank of [China]]\n(Finkel and Manning, 2009). In this paper, we use ideas from graph-based\ndependency parsing to provide our model a global view on the input via a\nbiaffine model (Dozat and Manning, 2017). The biaffine model scores pairs of\nstart and end tokens in a sentence which we use to explore all spans, so that\nthe model is able to predict named entities accurately. We show that the model\nworks well for both nested and flat NER through evaluation on 8 corpora and\nachieving SoTA performance on all of them, with accuracy gains of up to 2.2\npercentage points.', 'As a natural language generation task, it is challenging to generate\ninformative and coherent review text. In order to enhance the informativeness\nof the generated text, existing solutions typically learn to copy entities or\ntriples from knowledge graphs (KGs). However, they lack overall consideration\nto select and arrange the incorporated knowledge, which tends to cause text\nincoherence.\n  To address the above issue, we focus on improving entity-centric coherence of\nthe generated reviews by leveraging the semantic structure of KGs. In this\npaper, we propose a novel Coherence Enhanced Text Planning model (CETP) based\non knowledge graphs (KGs) to improve both global and local coherence for review\ngeneration. The proposed model learns a two-level text plan for generating a\ndocument: (1) the document plan is modeled as a sequence of sentence plans in\norder, and (2) the sentence plan is modeled as an entity-based subgraph from\nKG. Local coherence can be naturally enforced by KG subgraphs through\nintra-sentence correlations between entities. For global coherence, we design a\nhierarchical self-attentive architecture with both subgraph- and node-level\nattention to enhance the correlations between subgraphs. To our knowledge, we\nare the first to utilize a KG-based text planning model to enhance text\ncoherence for review generation. Extensive experiments on three datasets\nconfirm the effectiveness of our model on improving the content coherence of\ngenerated texts.', 'In this work, we aim at equipping pre-trained language models with structured\nknowledge. We present two self-supervised tasks learning over raw text with the\nguidance from knowledge graphs. Building upon entity-level masked language\nmodels, our first contribution is an entity masking scheme that exploits\nrelational knowledge underlying the text. This is fulfilled by using a linked\nknowledge graph to select informative entities and then masking their mentions.\nIn addition we use knowledge graphs to obtain distractors for the masked\nentities, and propose a novel distractor-suppressed ranking objective which is\noptimized jointly with masked language model. In contrast to existing\nparadigms, our approach uses knowledge graphs implicitly, only during\npre-training, to inject language models with structured knowledge via learning\nfrom raw text. It is more efficient than retrieval-based methods that perform\nentity linking and integration during finetuning and inference, and generalizes\nmore effectively than the methods that directly learn from concatenated graph\ntriples. Experiments show that our proposed model achieves improved performance\non five benchmark datasets, including question answering and knowledge base\ncompletion tasks.', 'Much research effort has been put to multilingual knowledge graph (KG)\nembedding methods to address the entity alignment task, which seeks to match\nentities in different languagespecific KGs that refer to the same real-world\nobject. Such methods are often hindered by the insufficiency of seed alignment\nprovided between KGs. Therefore, we propose an incidentally supervised model,\nJEANS , which jointly represents multilingual KGs and text corpora in a shared\nembedding scheme, and seeks to improve entity alignment with incidental\nsupervision signals from text. JEANS first deploys an entity grounding process\nto combine each KG with the monolingual text corpus. Then, two learning\nprocesses are conducted: (i) an embedding learning process to encode the KG and\ntext of each language in one embedding space, and (ii) a selflearning based\nalignment learning process to iteratively induce the matching of entities and\nthat of lexemes between embeddings. Experiments on benchmark datasets show that\nJEANS leads to promising improvement on entity alignment with incidental\nsupervision, and significantly outperforms state-of-the-art methods that solely\nrely on internal information of KGs.', 'In this paper, we propose a new strategy for the task of named entity\nrecognition (NER). We cast the task as a query-based machine reading\ncomprehension task: e.g., the task of extracting entities with PER is\nformalized as answering the question of ""which person is mentioned in the text\n?"". Such a strategy comes with the advantage that it solves the long-standing\nissue of handling overlapping or nested entities (the same token that\nparticipates in more than one entity categories) with sequence-labeling\ntechniques for NER. Additionally, since the query encodes informative prior\nknowledge, this strategy facilitates the process of entity extraction, leading\nto better performances. We experiment the proposed model on five widely used\nNER datasets on English and Chinese, including MSRA, Resume, OntoNotes, ACE04\nand ACE05. The proposed model sets new SOTA results on all of these datasets.', 'Recently, end-to-end (E2E) trained models for question answering over\nknowledge graphs (KGQA) have delivered promising results using only a weakly\nsupervised dataset. However, these models are trained and evaluated in a\nsetting where hand-annotated question entities are supplied to the model,\nleaving the important and non-trivial task of entity resolution (ER) outside\nthe scope of E2E learning. In this work, we extend the boundaries of E2E\nlearning for KGQA to include the training of an ER component. Our model only\nneeds the question text and the answer entities to train, and delivers a\nstand-alone QA model that does not require an additional ER component to be\nsupplied during runtime. Our approach is fully differentiable, thanks to its\nreliance on a recent method for building differentiable KGs (Cohen et al.,\n2020). We evaluate our E2E trained model on two public datasets and show that\nit comes close to baseline models that use hand-annotated entities.', 'Knowledge graph learning plays a critical role in integrating domain specific\nknowledge bases when deploying machine learning and data mining models in\npractice. Existing methods on knowledge graph learning primarily focus on\nmodeling the relations among entities as translations among the relations and\nentities, and many of these methods are not able to handle zero-shot problems,\nwhen new entities emerge. In this paper, we present a new convolutional neural\nnetwork (CNN)-based dual-chain model. Different from translation based methods,\nin our model, interactions among relations and entities are directly captured\nvia CNN over their embeddings. Moreover, a secondary chain of learning is\nconducted simultaneously to incorporate additional information and to enable\nbetter performance. We also present an extension of this model, which\nincorporates descriptions of entities and learns a second set of entity\nembeddings from the descriptions. As a result, the extended model is able to\neffectively handle zero-shot problems. We conducted comprehensive experiments,\ncomparing our methods with 15 methods on 8 benchmark datasets. Extensive\nexperimental results demonstrate that our proposed methods achieve or\noutperform the state-of-the-art results on knowledge graph learning, and\noutperform other methods on zero-shot problems. In addition, our methods\napplied to real-world biomedical data are able to produce results that conform\nto expert domain knowledge.', 'We present a multilingual bag-of-entities model that effectively boosts the\nperformance of zero-shot cross-lingual text classification by extending a\nmultilingual pre-trained language model (e.g., M-BERT). It leverages the\nmultilingual nature of Wikidata: entities in multiple languages representing\nthe same concept are defined with a unique identifier. This enables entities\ndescribed in multiple languages to be represented using shared embeddings. A\nmodel trained on entity features in a resource-rich language can thus be\ndirectly applied to other languages. Our experimental results on cross-lingual\ntopic classification (using the MLDoc and TED-CLDC datasets) and entity typing\n(using the SHINRA2020-ML dataset) show that the proposed model consistently\noutperforms state-of-the-art models.', 'The state-of-the-art named entity recognition (NER) systems are statistical\nmachine learning models that have strong generalization capability (i.e., can\nrecognize unseen entities that do not appear in training data) based on lexical\nand contextual information. However, such a model could still make mistakes if\nits features favor a wrong entity type. In this paper, we utilize Wikipedia as\nan open knowledge base to improve multilingual NER systems. Central to our\napproach is the construction of high-accuracy, high-coverage multilingual\nWikipedia entity type mappings. These mappings are built from weakly annotated\ndata and can be extended to new languages with no human annotation or\nlanguage-dependent knowledge involved. Based on these mappings, we develop\nseveral approaches to improve an NER system. We evaluate the performance of the\napproaches via experiments on NER systems trained for 6 languages. Experimental\nresults show that the proposed approaches are effective in improving the\naccuracy of such systems on unseen entities, especially when a system is\napplied to a new domain or it is trained with little training data (up to 18.3\nF1 score improvement).', ""We address the problem of calibrating prediction confidence for output\nentities of interest in natural language processing (NLP) applications. It is\nimportant that NLP applications such as named entity recognition and question\nanswering produce calibrated confidence scores for their predictions,\nespecially if the system is to be deployed in a safety-critical domain such as\nhealthcare. However, the output space of such structured prediction models is\noften too large to adapt binary or multi-class calibration methods directly. In\nthis study, we propose a general calibration scheme for output entities of\ninterest in neural-network based structured prediction models. Our proposed\nmethod can be used with any binary class calibration scheme and a neural\nnetwork model. Additionally, we show that our calibration method can also be\nused as an uncertainty-aware, entity-specific decoding step to improve the\nperformance of the underlying model at no additional training cost or data\nrequirements. We show that our method outperforms current calibration\ntechniques for named-entity-recognition, part-of-speech and question answering.\nWe also improve our model's performance from our decoding step across several\ntasks and benchmark datasets. Our method improves the calibration and model\nperformance on out-of-domain test scenarios as well."", 'The first stage of every knowledge base question answering approach is to\nlink entities in the input question. We investigate entity linking in the\ncontext of a question answering task and present a jointly optimized neural\narchitecture for entity mention detection and entity disambiguation that models\nthe surrounding context on different levels of granularity. We use the Wikidata\nknowledge base and available question answering datasets to create benchmarks\nfor entity linking on question answering data. Our approach outperforms the\nprevious state-of-the-art system on this data, resulting in an average 8%\nimprovement of the final score. We further demonstrate that our model delivers\na strong performance across different entity categories.', 'Natural language understanding involves reading between the lines with\nimplicit background knowledge. Current systems either rely on pre-trained\nlanguage models as the sole implicit source of world knowledge, or resort to\nexternal knowledge bases (KBs) to incorporate additional relevant knowledge. We\npropose an unsupervised framework based on self-talk as a novel alternative to\nmultiple-choice commonsense tasks. Inspired by inquiry-based discovery learning\n(Bruner, 1961), our approach inquires language models with a number of\ninformation seeking questions such as ""$\\textit{what is the definition of\n...}$"" to discover additional background knowledge. Empirical results\ndemonstrate that the self-talk procedure substantially improves the performance\nof zero-shot language model baselines on four out of six commonsense\nbenchmarks, and competes with models that obtain knowledge from external KBs.\nWhile our approach improves performance on several benchmarks, the self-talk\ninduced knowledge even when leading to correct answers is not always seen as\nuseful by human judges, raising interesting questions about the inner-workings\nof pre-trained language models for commonsense reasoning.', 'We focus on the problem of learning distributed representations for entity\nsearch queries, named entities, and their short descriptions. With our\nrepresentation learning models, the entity search query, named entity and\ndescription can be represented as low-dimensional vectors. Our goal is to\ndevelop a simple but effective model that can make the distributed\nrepresentations of query related entities similar to the query in the vector\nspace. Hence, we propose three kinds of learning strategies, and the difference\nbetween them mainly lies in how to deal with the relationship between an entity\nand its description. We analyze the strengths and weaknesses of each learning\nstrategy and validate our methods on public datasets which contain four kinds\nof named entities, i.e., movies, TV shows, restaurants and celebrities. The\nexperimental results indicate that our proposed methods can adapt to different\ntypes of entity search queries, and outperform the current state-of-the-art\nmethods based on keyword matching and vanilla word2vec models. Besides, the\nproposed methods can be trained fast and be easily extended to other similar\ntasks.', ""Current systems of fine-grained entity typing use distant supervision in\nconjunction with existing knowledge bases to assign categories (type labels) to\nentity mentions. However, the type labels so obtained from knowledge bases are\noften noisy (i.e., incorrect for the entity mention's local context). We define\na new task, Label Noise Reduction in Entity Typing (LNR), to be the automatic\nidentification of correct type labels (type-paths) for training examples, given\nthe set of candidate type labels obtained by distant supervision with a given\ntype hierarchy. The unknown type labels for individual entity mentions and the\nsemantic similarity between entity types pose unique challenges for solving the\nLNR task. We propose a general framework, called PLE, to jointly embed entity\nmentions, text features and entity types into the same low-dimensional space\nwhere, in that space, objects whose types are semantically close have similar\nrepresentations. Then we estimate the type-path for each training example in a\ntop-down manner using the learned embeddings. We formulate a global objective\nfor learning the embeddings from text corpora and knowledge bases, which adopts\na novel margin-based loss that is robust to noisy labels and faithfully models\ntype correlation derived from knowledge bases. Our experiments on three public\ntyping datasets demonstrate the effectiveness and robustness of PLE, with an\naverage of 25% improvement in accuracy compared to next best method."", 'Entity representations are useful in natural language tasks involving\nentities. In this paper, we propose new pretrained contextualized\nrepresentations of words and entities based on the bidirectional transformer.\nThe proposed model treats words and entities in a given text as independent\ntokens, and outputs contextualized representations of them. Our model is\ntrained using a new pretraining task based on the masked language model of\nBERT. The task involves predicting randomly masked words and entities in a\nlarge entity-annotated corpus retrieved from Wikipedia. We also propose an\nentity-aware self-attention mechanism that is an extension of the\nself-attention mechanism of the transformer, and considers the types of tokens\n(words or entities) when computing attention scores. The proposed model\nachieves impressive empirical performance on a wide range of entity-related\ntasks. In particular, it obtains state-of-the-art results on five well-known\ndatasets: Open Entity (entity typing), TACRED (relation classification),\nCoNLL-2003 (named entity recognition), ReCoRD (cloze-style question answering),\nand SQuAD 1.1 (extractive question answering). Our source code and pretrained\nrepresentations are available at https://github.com/studio-ousia/luke.', 'This thesis presents new methods for unsupervised learning of distributed\nrepresentations of words and entities from text and knowledge bases. The first\nalgorithm presented in the thesis is a multi-view algorithm for learning\nrepresentations of words called Multiview Latent Semantic Analysis (MVLSA). By\nincorporating up to 46 different types of co-occurrence statistics for the same\nvocabulary of english words, I show that MVLSA outperforms other\nstate-of-the-art word embedding models. Next, I focus on learning entity\nrepresentations for search and recommendation and present the second method of\nthis thesis, Neural Variational Set Expansion (NVSE). NVSE is also an\nunsupervised learning method, but it is based on the Variational Autoencoder\nframework. Evaluations with human annotators show that NVSE can facilitate\nbetter search and recommendation of information gathered from noisy, automatic\nannotation of unstructured natural language corpora. Finally, I move from\nunstructured data and focus on structured knowledge graphs. I present novel\napproaches for learning embeddings of vertices and edges in a knowledge graph\nthat obey logical constraints.', 'In biomedical literature, it is common for entity boundaries to not align\nwith word boundaries. Therefore, effective identification of entity spans\nrequires approaches capable of considering tokens that are smaller than words.\nWe introduce a novel, subword approach for named entity recognition (NER) that\nuses byte-pair encodings (BPE) in combination with convolutional and recurrent\nneural networks to produce byte-level tags of entities. We present experimental\nresults on several standard biomedical datasets, namely the BioCreative VI\nBio-ID, JNLPBA, and GENETAG datasets. We demonstrate competitive performance\nwhile bypassing the specialized domain expertise needed to create biomedical\ntext tokenization rules.', 'Rich entity representations are useful for a wide class of problems involving\nentities. Despite their importance, there is no standardized benchmark that\nevaluates the overall quality of entity representations. In this work, we\npropose EntEval: a test suite of diverse tasks that require nontrivial\nunderstanding of entities including entity typing, entity similarity, entity\nrelation prediction, and entity disambiguation. In addition, we develop\ntraining techniques for learning better entity representations by using natural\nhyperlink annotations in Wikipedia. We identify effective objectives for\nincorporating the contextual information in hyperlinks into state-of-the-art\npretrained language models and show that they improve strong baselines on\nmultiple EntEval tasks.', 'In this work, we formulate the NER task as a multi-answer knowledge guided QA\ntask (KGQA) which helps to predict entities only by assigning B, I and O tags\nwithout associating entity types with the tags. We provide different knowledge\ncontexts, such as, entity types, questions, definitions and examples along with\nthe text and train on a combined dataset of 18 biomedical corpora. This\nformulation (a) enables systems to jointly learn NER specific features from\nvaried NER datasets, (b) can use knowledge-text attention to identify words\nhaving higher similarity to provided knowledge, improving performance, (c)\nreduces system confusion by reducing the prediction classes to B, I, O only,\nand (d) makes detection of nested entities easier. We perform extensive\nexperiments of this KGQA formulation on 18 biomedical NER datasets, and through\nexperiments we note that knowledge helps in achieving better performance. Our\nproblem formulation is able to achieve state-of-the-art results in 12 datasets.', 'Learning representations for knowledge base entities and concepts is becoming\nincreasingly important for NLP applications. However, recent entity embedding\nmethods have relied on structured resources that are expensive to create for\nnew domains and corpora. We present a distantly-supervised method for jointly\nlearning embeddings of entities and text from an unnanotated corpus, using only\na list of mappings between entities and surface forms. We learn embeddings from\nopen-domain and biomedical corpora, and compare against prior methods that rely\non human-annotated text or large knowledge graph structure. Our embeddings\ncapture entity similarity and relatedness better than prior work, both in\nexisting biomedical datasets and a new Wikipedia-based dataset that we release\nto the community. Results on analogy completion and entity sense disambiguation\nindicate that entities and words capture complementary information that can be\neffectively combined for downstream use.', ""``Classical'' word embeddings, such as Word2Vec, have been shown to capture\nthe semantics of words based on their distributional properties. However, their\nability to represent the different meanings that a word may have is limited.\nSuch approaches also do not explicitly encode relations between entities, as\ndenoted by words. Embeddings of knowledge bases (KB) capture the explicit\nrelations between entities denoted by words, but are not able to directly\ncapture the syntagmatic properties of these words. To our knowledge, recent\nresearch have focused on representation learning that augment the strengths of\none with the other. In this work, we begin exploring another approach using\ncontextual and KB embeddings jointly at the same level and propose two tasks --\nan entity typing and a relation typing task -- that evaluate the performance of\ncontextual and KB embeddings. We also evaluated a concatenated model of\ncontextual and KB embeddings with these two tasks, and obtain conclusive\nresults on the first task. We hope our work may contribute as a basis for\nmodels and datasets that develop in the direction of this approach."", 'Named Entity Disambiguation (NED) refers to the task of resolving multiple\nnamed entity mentions in a document to their correct references in a knowledge\nbase (KB) (e.g., Wikipedia). In this paper, we propose a novel embedding method\nspecifically designed for NED. The proposed method jointly maps words and\nentities into the same continuous vector space. We extend the skip-gram model\nby using two models. The KB graph model learns the relatedness of entities\nusing the link structure of the KB, whereas the anchor context model aims to\nalign vectors such that similar words and entities occur close to one another\nin the vector space by leveraging KB anchors and their context words. By\ncombining contexts based on the proposed embedding with standard NED features,\nwe achieved state-of-the-art accuracy of 93.1% on the standard CoNLL dataset\nand 85.2% on the TAC 2010 dataset.', 'Named entity recognition is one of the core tasks in NLP. Although many\nimprovements have been made on this task during the last years, the\nstate-of-the-art systems do not explicitly take into account the recursive\nnature of language. Instead of only treating the text as a plain sequence of\nwords, we incorporate a linguistically-inspired way to recognize entities based\non syntax and tree structures. Our model exploits syntactic relationships among\nwords using a Tree-LSTM guided by dependency trees. Then, we enhance these\nfeatures by applying relative and global attention mechanisms. On the one hand,\nthe relative attention detects the most informative words in the sentence with\nrespect to the word being evaluated. On the other hand, the global attention\nspots the most relevant words in the sequence. Lastly, we linearly project the\nweighted vectors into the tagging space so that a conditional random field\nclassifier predicts the entity labels. Our findings show that the model detects\nwords that disclose the entity types based on their syntactic roles in a\nsentence (e.g., verbs such as speak and write are attended when the entity type\nis PERSON, whereas meet and travel strongly relate to LOCATION). We confirm our\nfindings and establish a new state of the art on two datasets.', ""The lack of labeled training data has limited the development of natural\nlanguage processing tools, such as named entity recognition, for many languages\nspoken in developing countries. Techniques such as distant and weak supervision\ncan be used to create labeled data in a (semi-) automatic way. Additionally, to\nalleviate some of the negative effects of the errors in automatic annotation,\nnoise-handling methods can be integrated. Pretrained word embeddings are\nanother key component of most neural named entity classifiers. With the advent\nof more complex contextual word embeddings, an interesting trade-off between\nmodel size and performance arises. While these techniques have been shown to\nwork well in high-resource settings, we want to study how they perform in\nlow-resource scenarios. In this work, we perform named entity recognition for\nHausa and Yor\\`ub\\'a, two languages that are widely spoken in several\ndeveloping countries. We evaluate different embedding approaches and show that\ndistant supervision can be successfully leveraged in a realistic low-resource\nscenario where it can more than double a classifier's performance."", 'Previous models for learning entity and relationship embeddings of knowledge\ngraphs such as TransE, TransH, and TransR aim to explore new links based on\nlearned representations. However, these models interpret relationships as\nsimple translations on entity embeddings. In this paper, we try to learn more\ncomplex connections between entities and relationships. In particular, we use a\nConvolutional Neural Network (CNN) to learn entity and relationship\nrepresentations in knowledge graphs. In our model, we treat entities and\nrelationships as one-dimensional numerical sequences with the same length.\nAfter that, we combine each triplet of head, relationship, and tail together as\na matrix with height 3. CNN is applied to the triplets to get confidence\nscores. Positive and manually corrupted negative triplets are used to train the\nembeddings and the CNN model simultaneously. Experimental results on public\nbenchmark datasets show that the proposed model outperforms state-of-the-art\nmodels on exploring unseen relationships, which proves that CNN is effective to\nlearn complex interactive patterns between entities and relationships.', ""Universal schema jointly embeds knowledge bases and textual patterns to\nreason about entities and relations for automatic knowledge base construction\nand information extraction. In the past, entity pairs and relations were\nrepresented as learned vectors with compatibility determined by a scoring\nfunction, limiting generalization to unseen text patterns and entities.\nRecently, 'column-less' versions of Universal Schema have used compositional\npattern encoders to generalize to all text patterns. In this work we take the\nnext step and propose a 'row-less' model of universal schema, removing explicit\nentity pair representations. Instead of learning vector representations for\neach entity pair in our training set, we treat an entity pair as a function of\nits relation types. In experimental results on the FB15k-237 benchmark we\ndemonstrate that we can match the performance of a comparable model with\nexplicit entity pair representations using a model of attention over relation\ntypes. We further demonstrate that the model per- forms with nearly the same\naccuracy on entity pairs never seen during training."", 'Fine-grained Named Entity Recognition is a task whereby we detect and\nclassify entity mentions to a large set of types. These types can span diverse\ndomains such as finance, healthcare, and politics. We observe that when the\ntype set spans several domains the accuracy of the entity detection becomes a\nlimitation for supervised learning models. The primary reason being the lack of\ndatasets where entity boundaries are properly annotated, whilst covering a\nlarge spectrum of entity types. Furthermore, many named entity systems suffer\nwhen considering the categorization of fine grained entity types. Our work\nattempts to address these issues, in part, by combining state-of-the-art deep\nlearning models (ELMo) with an expansive knowledge base (Wikidata). Using our\nframework, we cross-validate our model on the 112 fine-grained entity types\nbased on the hierarchy given from the Wiki(gold) dataset.', 'We study the problem of training named entity recognition (NER) models using\nonly distantly-labeled data, which can be automatically obtained by matching\nentity mentions in the raw text with entity types in a knowledge base. The\nbiggest challenge of distantly-supervised NER is that the distant supervision\nmay induce incomplete and noisy labels, rendering the straightforward\napplication of supervised learning ineffective. In this paper, we propose (1) a\nnoise-robust learning scheme comprised of a new loss function and a noisy label\nremoval step, for training NER models on distantly-labeled data, and (2) a\nself-training method that uses contextualized augmentations created by\npre-trained language models to improve the generalization ability of the NER\nmodel. On three benchmark datasets, our method achieves superior performance,\noutperforming existing distantly-supervised NER models by significant margins.', 'Fine-grained entity typing is a challenging problem since it usually involves\na relatively large tag set and may require to understand the context of the\nentity mention. In this paper, we use entity linking to help with the\nfine-grained entity type classification process. We propose a deep neural model\nthat makes predictions based on both the context and the information obtained\nfrom entity linking results. Experimental results on two commonly used datasets\ndemonstrates the effectiveness of our approach. On both datasets, it achieves\nmore than 5\\% absolute strict accuracy improvement over the state of the art.', 'This paper explores the problem of matching entities across different\nknowledge graphs. Given a query entity in one knowledge graph, we wish to find\nthe corresponding real-world entity in another knowledge graph. We formalize\nthis problem and present two large-scale datasets for this task based on\nexiting cross-ontology links between DBpedia and Wikidata, focused on several\nhundred thousand ambiguous entities. Using a classification-based approach, we\nfind that a simple multi-layered perceptron based on representations derived\nfrom RDF2Vec graph embeddings of entities in each knowledge graph is sufficient\nto achieve high accuracy, with only small amounts of training data. The\ncontributions of our work are datasets for examining this problem and strong\nbaselines on which future work can be based.', 'Entity retrieval, which aims at disambiguating mentions to canonical entities\nfrom massive KBs, is essential for many tasks in natural language processing.\nRecent progress in entity retrieval shows that the dual-encoder structure is a\npowerful and efficient framework to nominate candidates if entities are only\nidentified by descriptions. However, they ignore the property that meanings of\nentity mentions diverge in different contexts and are related to various\nportions of descriptions, which are treated equally in previous works. In this\nwork, we propose Multi-View Entity Representations (MuVER), a novel approach\nfor entity retrieval that constructs multi-view representations for entity\ndescriptions and approximates the optimal view for mentions via a heuristic\nsearching method. Our method achieves the state-of-the-art performance on\nZESHEL and improves the quality of candidates on three standard Entity Linking\ndatasets', 'This paper introduces SocialVec, a general framework for eliciting social\nworld knowledge from social networks, and applies this framework to Twitter.\nSocialVec learns low-dimensional embeddings of popular accounts, which\nrepresent entities of general interest, based on their co-occurrences patterns\nwithin the accounts followed by individual users, thus modeling entity\nsimilarity in socio-demographic terms. Similar to word embeddings, which\nfacilitate tasks that involve text processing, we expect social entity\nembeddings to benefit tasks of social flavor. We have learned social embeddings\nfor roughly 200,000 popular accounts from a sample of the Twitter network that\nincludes more than 1.3 million users and the accounts that they follow, and\nevaluate the resulting embeddings on two different tasks. The first task\ninvolves the automatic inference of personal traits of users from their social\nmedia profiles. In another study, we exploit SocialVec embeddings for gauging\nthe political bias of news sources in Twitter. In both cases, we prove\nSocialVec embeddings to be advantageous compared with existing entity embedding\nschemes. We will make the SocialVec entity embeddings publicly available to\nsupport further exploration of social world knowledge as reflected in Twitter.', ""Question-answering systems and voice assistants are becoming major part of\nclient service departments of many organizations, helping them to reduce the\nlabor costs of staff. In many such systems, there is always natural language\nunderstanding module that solves intent classification task. This task is\ncomplicated because of its case-dependency - every subject area has its own\nsemantic kernel. The state of art approaches for intent classification are\ndifferent machine learning and deep learning methods that use text vector\nrepresentations as input. The basic vector representation models such as Bag of\nwords and TF-IDF generate sparse matrixes, which are becoming very big as the\namount of input data grows. Modern methods such as word2vec and FastText use\nneural networks to evaluate word embeddings with fixed dimension size. As we\nare developing a question-answering system for students and enrollees of the\nPerm National Research Polytechnic University, we have faced the problem of\nuser's intent detection. The subject area of our system is very specific, that\nis why there is a lack of training data. This aspect makes intent\nclassification task more challenging for using state of the art deep learning\nmethods. In this paper, we propose an approach of the questions embeddings\nrepresentation based on calculation of Shannon entropy.The goal of the approach\nis to produce low dimensional question vectors as neural approaches do and to\noutperform related methods, described above in condition of small dataset. We\nevaluate and compare our model with existing ones using logistic regression and\ndataset that contains questions asked by students and enrollees. The data is\nlabeled into six classes. Experimental comparison of proposed approach and\nother models revealed that proposed model performed better in the given task."", ""Knowledge graphs (KGs) consisting of a large number of triples have become\nwidespread recently, and many knowledge graph embedding (KGE) methods are\nproposed to embed entities and relations of a KG into continuous vector spaces.\nSuch embedding methods simplify the operations of conducting various in-KG\ntasks (e.g., link prediction) and out-of-KG tasks (e.g., question answering).\nThey can be viewed as general solutions for representing KGs. However, existing\nKGE methods are not applicable to inductive settings, where a model trained on\nsource KGs will be tested on target KGs with entities unseen during model\ntraining. Existing works focusing on KGs in inductive settings can only solve\nthe inductive relation prediction task. They can not handle other out-of-KG\ntasks as general as KGE methods since they don't produce embeddings for\nentities. In this paper, to achieve inductive knowledge graph embedding, we\npropose a model MorsE, which does not learn embeddings for entities but learns\ntransferable meta-knowledge that can be used to produce entity embeddings. Such\nmeta-knowledge is modeled by entity-independent modules and learned by\nmeta-learning. Experimental results show that our model significantly\noutperforms corresponding baselines for in-KG and out-of-KG tasks in inductive\nsettings."", 'The Biocreative VII Track-2 challenge consists of named entity recognition,\nentity-linking (or entity-normalization), and topic indexing tasks -- with\nentities and topics limited to chemicals for this challenge. Named entity\nrecognition is a well-established problem and we achieve our best performance\nwith BERT-based BioMegatron models. We extend our BERT-based approach to the\nentity linking task. After the second stage of pretraining BioBERT with a\nmetric-learning loss strategy called self-alignment pretraining (SAP), we link\nentities based on the cosine similarity between their SAP-BioBERT word\nembeddings. Despite the success of our named entity recognition experiments, we\nfind the chemical indexing task generally more challenging.\n  In addition to conventional NER methods, we attempt both named entity\nrecognition and entity linking with a novel text-to-text or ""prompt"" based\nmethod that uses generative language models such as T5 and GPT. We achieve\nencouraging results with this new approach.', 'Much of human knowledge is encoded in text, available in scientific\npublications, books, and the web. Given the rapid growth of these resources, we\nneed automated methods to extract such knowledge into machine-processable\nstructures, such as knowledge graphs. An important task in this process is\nentity normalization, which consists of mapping noisy entity mentions in text\nto canonical entities in well-known reference sets. However, entity\nnormalization is a challenging problem; there often are many textual forms for\na canonical entity that may not be captured in the reference set, and entities\nmentioned in text may include many syntactic variations, or errors. The problem\nis particularly acute in scientific domains, such as biology. To address this\nproblem, we have developed a general, scalable solution based on a deep Siamese\nneural network model to embed the semantic information about the entities, as\nwell as their syntactic variations. We use these embeddings for fast mapping of\nnew entities to large reference sets, and empirically show the effectiveness of\nour framework in challenging bio-entity normalization datasets.', 'Entity alignment (EA) is the task to discover entities referring to the same\nreal-world object from different knowledge graphs (KGs), which is the most\ncrucial step in integrating multi-source KGs. The majority of the existing\nembeddings-based entity alignment methods embed entities and relations into a\nvector space based on relation triples of KGs for local alignment. As these\nmethods insufficiently consider the multiple relations between entities, the\nstructure information of KGs has not been fully leveraged. In this paper, we\npropose a novel framework based on Relation-aware Graph Attention Networks to\ncapture the interactions between entities and relations. Our framework adopts\nthe self-attention mechanism to spread entity information to the relations and\nthen aggregate relation information back to entities. Furthermore, we propose a\nglobal alignment algorithm to make one-to-one entity alignments with a\nfine-grained similarity matrix. Experiments on three real-world cross-lingual\ndatasets show that our framework outperforms the state-of-the-art methods.', 'Named entity recognition is a natural language processing task to recognize\nand extract spans of text associated with named entities and classify them in\nsemantic Categories.\n  Google BERT is a deep bidirectional language model, pre-trained on large\ncorpora that can be fine-tuned to solve many NLP tasks such as question\nanswering, named entity recognition, part of speech tagging and etc. In this\npaper, we use the pre-trained deep bidirectional network, BERT, to make a model\nfor named entity recognition in Persian.\n  We also compare the results of our model with the previous state of the art\nresults achieved on Persian NER. Our evaluation metric is CONLL 2003 score in\ntwo levels of word and phrase. This model achieved second place in NSURL-2019\ntask 7 competition which associated with NER for the Persian language. our\nresults in this competition are 83.5 and 88.4 f1 CONLL score respectively in\nphrase and word level evaluation.', 'The objective of knowledge graph embedding is to encode both entities and\nrelations of knowledge graphs into continuous low-dimensional vector spaces.\nPreviously, most works focused on symbolic representation of knowledge graph\nwith structure information, which can not handle new entities or entities with\nfew facts well. In this paper, we propose a novel deep architecture to utilize\nboth structural and textual information of entities. Specifically, we introduce\nthree neural models to encode the valuable information from text description of\nentity, among which an attentive model can select related information as\nneeded. Then, a gating mechanism is applied to integrate representations of\nstructure and text into a unified architecture. Experiments show that our\nmodels outperform baseline by margin on link prediction and triplet\nclassification tasks. Source codes of this paper will be available on Github.', ""Reading comprehension is a challenging task in natural language processing\nand requires a set of skills to be solved. While current approaches focus on\nsolving the task as a whole, in this paper, we propose to use a neural network\n`skill' transfer approach. We transfer knowledge from several lower-level\nlanguage tasks (skills) including textual entailment, named entity recognition,\nparaphrase detection and question type classification into the reading\ncomprehension model.\n  We conduct an empirical evaluation and show that transferring language skill\nknowledge leads to significant improvements for the task with much fewer steps\ncompared to the baseline model. We also show that the skill transfer approach\nis effective even with small amounts of training data. Another finding of this\nwork is that using token-wise deep label supervision for text classification\nimproves the performance of transfer learning."", 'Deep learning currently dominates the benchmarks for various NLP tasks and,\nat the basis of such systems, words are frequently represented as embeddings\n--vectors in a low dimensional space-- learned from large text corpora and\nvarious algorithms have been proposed to learn both word and concept\nembeddings. One of the claimed benefits of such embeddings is that they capture\nknowledge about semantic relations. Such embeddings are most often evaluated\nthrough tasks such as predicting human-rated similarity and analogy which only\ntest a few, often ill-defined, relations. In this paper, we propose a method\nfor (i) reliably generating word and concept pair datasets for a wide number of\nrelations by using a knowledge graph and (ii) evaluating to what extent\npre-trained embeddings capture those relations. We evaluate the approach\nagainst a proprietary and a public knowledge graph and analyze the results,\nshowing which lexico-semantic relational knowledge is captured by current\nembedding learning approaches.', 'Named entity recognition (NER) is a fundamental component in many\napplications, such as Web Search and Voice Assistants. Although deep neural\nnetworks greatly improve the performance of NER, due to the requirement of\nlarge amounts of training data, deep neural networks can hardly scale out to\nmany languages in an industry setting. To tackle this challenge, cross-lingual\nNER transfers knowledge from a rich-resource language to languages with low\nresources through pre-trained multilingual language models. Instead of using\ntraining data in target languages, cross-lingual NER has to rely on only\ntraining data in source languages, and optionally adds the translated training\ndata derived from source languages. However, the existing cross-lingual NER\nmethods do not make good use of rich unlabeled data in target languages, which\nis relatively easy to collect in industry applications. To address the\nopportunities and challenges, in this paper we describe our novel practice in\nMicrosoft to leverage such large amounts of unlabeled data in target languages\nin real production settings. To effectively extract weak supervision signals\nfrom the unlabeled data, we develop a novel approach based on the ideas of\nsemi-supervised learning and reinforcement learning. The empirical study on\nthree benchmark data sets verifies that our approach establishes the new\nstate-of-the-art performance with clear edges. Now, the NER techniques reported\nin this paper are on their way to become a fundamental component for Web\nranking, Entity Pane, Answers Triggering, and Question Answering in the\nMicrosoft Bing search engine. Moreover, our techniques will also serve as part\nof the Spoken Language Understanding module for a commercial voice assistant.\nWe plan to open source the code of the prototype framework after deployment.', 'Named entity recognition (NER) is a widely applicable natural language\nprocessing task and building block of question answering, topic modeling,\ninformation retrieval, etc. In the medical domain, NER plays a crucial role by\nextracting meaningful chunks from clinical notes and reports, which are then\nfed to downstream tasks like assertion status detection, entity resolution,\nrelation extraction, and de-identification. Reimplementing a Bi-LSTM-CNN-Char\ndeep learning architecture on top of Apache Spark, we present a single\ntrainable NER model that obtains new state-of-the-art results on seven public\nbiomedical benchmarks without using heavy contextual embeddings like BERT. This\nincludes improving BC4CHEMD to 93.72% (4.1% gain), Species800 to 80.91% (4.6%\ngain), and JNLPBA to 81.29% (5.2% gain). In addition, this model is freely\navailable within a production-grade code base as part of the open-source Spark\nNLP library; can scale up for training and inference in any Spark cluster; has\nGPU support and libraries for popular programming languages such as Python, R,\nScala and Java; and can be extended to support other human languages with no\ncode changes.', 'Most state-of-the-art models in natural language processing (NLP) are neural\nmodels built on top of large, pre-trained, contextual language models that\ngenerate representations of words in context and are fine-tuned for the task at\nhand. The improvements afforded by these ""contextual embeddings"" come with a\nhigh computational cost. In this work, we explore a simple technique that\nsubstantially and consistently improves performance over a strong baseline with\nnegligible increase in run time. We concatenate multiple pre-trained embeddings\nto strengthen our representation of words. We show that this concatenation\ntechnique works across many tasks, datasets, and model types. We analyze\naspects of pre-trained embedding similarity and vocabulary coverage and find\nthat the representational diversity between different pre-trained embeddings is\nthe driving force of why this technique works. We provide open source\nimplementations of our models in both TensorFlow and PyTorch.', 'This paper addresses the problem of corpus-level entity typing, i.e.,\ninferring from a large corpus that an entity is a member of a class such as\n""food"" or ""artist"". The application of entity typing we are interested in is\nknowledge base completion, specifically, to learn which classes an entity is a\nmember of. We propose FIGMENT to tackle this problem. FIGMENT is embedding-\nbased and combines (i) a global model that scores based on aggregated\ncontextual information of an entity and (ii) a context model that first scores\nthe individual occurrences of an entity and then aggregates the scores. Each of\nthe two proposed models has some specific properties. For the global model,\nlearning high quality entity representations is crucial because it is the only\nsource used for the predictions. Therefore, we introduce representations using\nname and contexts of entities on the three levels of entity, word, and\ncharacter. We show each has complementary information and a multi-level\nrepresentation is the best. For the context model, we need to use distant\nsupervision since the context-level labels are not available for entities.\nDistant supervised labels are noisy and this harms the performance of models.\nTherefore, we introduce and apply new algorithms for noise mitigation using\nmulti-instance learning. We show the effectiveness of our models in a large\nentity typing dataset, built from Freebase.', 'In many scenarios, named entity recognition (NER) models severely suffer from\nunlabeled entity problem, where the entities of a sentence may not be fully\nannotated. Through empirical studies performed on synthetic datasets, we find\ntwo causes of performance degradation. One is the reduction of annotated\nentities and the other is treating unlabeled entities as negative instances.\nThe first cause has less impact than the second one and can be mitigated by\nadopting pretraining language models. The second cause seriously misguides a\nmodel in training and greatly affects its performances. Based on the above\nobservations, we propose a general approach, which can almost eliminate the\nmisguidance brought by unlabeled entities. The key idea is to use negative\nsampling that, to a large extent, avoids training NER models with unlabeled\nentities. Experiments on synthetic datasets and real-world datasets show that\nour model is robust to unlabeled entity problem and surpasses prior baselines.\nOn well-annotated datasets, our model is competitive with the state-of-the-art\nmethod.', 'Semantic annotation, the process of identifying key-phrases in texts and\nlinking them to concepts in a knowledge base, is an important basis for\nsemantic information retrieval and the Semantic Web uptake. Despite the\nemergence of semantic annotation systems, very few comparative studies have\nbeen published on their performance. In this paper, we provide an evaluation of\nthe performance of existing systems over three tasks: full semantic annotation,\nnamed entity recognition, and keyword detection. More specifically, the\nspotting capability (recognition of relevant surface forms in text) is\nevaluated for all three tasks, whereas the disambiguation (correctly\nassociating an entity from Wikipedia or DBpedia to the spotted surface forms)\nis evaluated only for the first two tasks. Our evaluation is twofold: First, we\ncompute standard precision and recall on the output of semantic annotators on\ndiverse datasets, each best suited for one of the identified tasks. Second, we\nbuild a statistical model using logistic regression to identify significant\nperformance differences. Our results show that systems that provide full\nannotation perform better than named entities annotators and keyword\nextractors, for all three tasks. However, there is still much room for\nimprovement for the identification of the most relevant entities described in a\ntext.', 'Machine understanding of user utterances in conversational systems is of\nutmost importance for enabling engaging and meaningful conversations with\nusers. Entity Linking (EL) is one of the means of text understanding, with\nproven efficacy for various downstream tasks in information retrieval. In this\npaper, we study entity linking for conversational systems. To develop a better\nunderstanding of what EL in a conversational setting entails, we analyze a\nlarge number of dialogues from existing conversational datasets and annotate\nreferences to concepts, named entities, and personal entities using\ncrowdsourcing. Based on the annotated dialogues, we identify the main\ncharacteristics of conversational entity linking. Further, we report on the\nperformance of traditional EL systems on our Conversational Entity Linking\ndataset, ConEL, and present an extension to these methods to better fit the\nconversational setting. The resources released with this paper include\nannotated datasets, detailed descriptions of crowdsourcing setups, as well as\nthe annotations produced by various EL systems. These new resources allow for\nan investigation of how the role of entities in conversations is different from\nthat in documents or isolated short text utterances like queries and tweets,\nand complement existing conversational datasets.', 'Named entity recognition (NER) is a vital task in spoken language\nunderstanding, which aims to identify mentions of named entities in text e.g.,\nfrom transcribed speech. Existing neural models for NER rely mostly on\ndedicated word-level representations, which suffer from two main shortcomings.\nFirst, the vocabulary size is large, yielding large memory requirements and\ntraining time. Second, these models are not able to learn morphological or\nphonological representations. To remedy the above shortcomings, we adopt a\nneural solution based on bidirectional LSTMs and conditional random fields,\nwhere we rely on subword units, namely characters, phonemes, and bytes. For\neach word in an utterance, our model learns a representation from each of the\nsubword units. We conducted experiments in a real-world large-scale setting for\nthe use case of a voice-controlled device covering four languages with up to\n5.5M utterances per language. Our experiments show that (1) with increasing\ntraining data, performance of models trained solely on subword units becomes\ncloser to that of models with dedicated word-level embeddings (91.35 vs 93.92\nF1 for English), while using a much smaller vocabulary size (332 vs 74K), (2)\nsubword units enhance models with dedicated word-level embeddings, and (3)\ncombining different subword units improves performance.', 'The field of natural language understanding has experienced exponential\nprogress in the last few years, with impressive results in several tasks. This\nsuccess has motivated researchers to study the underlying knowledge encoded by\nthese models. Despite this, attempts to understand their semantic capabilities\nhave not been successful, often leading to non-conclusive, or contradictory\nconclusions among different works. Via a probing classifier, we extract the\nunderlying knowledge graph of nine of the most influential language models of\nthe last years, including word embeddings, text generators, and context\nencoders. This probe is based on concept relatedness, grounded on WordNet. Our\nresults reveal that all the models encode this knowledge, but suffer from\nseveral inaccuracies. Furthermore, we show that the different architectures and\ntraining strategies lead to different model biases. We conduct a systematic\nevaluation to discover specific factors that explain why some concepts are\nchallenging. We hope our insights will motivate the development of models that\ncapture concepts more precisely.', 'This paper deals with the entity extraction task (named entity recognition)\nof a text mining process that aims at unveiling non-trivial semantic\nstructures, such as relationships and interaction between entities or\ncommunities. In this paper we present a simple and efficient named entity\nextraction algorithm. The method, named PAMPO (PAttern Matching and POs tagging\nbased algorithm for NER), relies on flexible pattern matching, part-of-speech\ntagging and lexical-based rules. It was developed to process texts written in\nPortuguese, however it is potentially applicable to other languages as well.\n  We compare our approach with current alternatives that support Named Entity\nRecognition (NER) for content written in Portuguese. These are Alchemy, Zemanta\nand Rembrandt. Evaluation of the efficacy of the entity extraction method on\nseveral texts written in Portuguese indicates a considerable improvement on\n$recall$ and $F_1$ measures.', 'Humans use language to refer to entities in the external world. Motivated by\nthis, in recent years several models that incorporate a bias towards learning\nentity representations have been proposed. Such entity-centric models have\nshown empirical success, but we still know little about why. In this paper we\nanalyze the behavior of two recently proposed entity-centric models in a\nreferential task, Entity Linking in Multi-party Dialogue (SemEval 2018 Task 4).\nWe show that these models outperform the state of the art on this task, and\nthat they do better on lower frequency entities than a counterpart model that\nis not entity-centric, with the same model size. We argue that making models\nentity-centric naturally fosters good architectural decisions. However, we also\nshow that these models do not really build entity representations and that they\nmake poor use of linguistic context. These negative results underscore the need\nfor model analysis, to test whether the motivations for particular\narchitectures are borne out in how models behave when deployed.', 'In this research, we improve upon the current state of the art in entity\nretrieval by re-ranking the result list using graph embeddings. The paper shows\nthat graph embeddings are useful for entity-oriented search tasks. We\ndemonstrate empirically that encoding information from the knowledge graph into\n(graph) embeddings contributes to a higher increase in effectiveness of entity\nretrieval results than using plain word embeddings. We analyze the impact of\nthe accuracy of the entity linker on the overall retrieval effectiveness. Our\nanalysis further deploys the cluster hypothesis to explain the observed\nadvantages of graph embeddings over the more widely used word embeddings, for\nuser tasks involving ranking entities.', 'We introduce EventNarrative, a knowledge graph-to-text dataset from publicly\navailable open-world knowledge graphs. Given the recent advances in\nevent-driven Information Extraction (IE), and that prior research on\ngraph-to-text only focused on entity-driven KGs, this paper focuses on\nevent-centric data. However, our data generation system can still be adapted to\nother other types of KG data. Existing large-scale datasets in the\ngraph-to-text area are non-parallel, meaning there is a large disconnect\nbetween the KGs and text. The datasets that have a paired KG and text, are\nsmall scale and manually generated or generated without a rich ontology, making\nthe corresponding graphs sparse. Furthermore, these datasets contain many\nunlinked entities between their KG and text pairs. EventNarrative consists of\napproximately 230,000 graphs and their corresponding natural language text, 6\ntimes larger than the current largest parallel dataset. It makes use of a rich\nontology, all of the KGs entities are linked to the text, and our manual\nannotations confirm a high data quality. Our aim is two-fold: help break new\nground in event-centric research where data is lacking, and to give researchers\na well-defined, large-scale dataset in order to better evaluate existing and\nfuture knowledge graph-to-text models. We also evaluate two types of baseline\non EventNarrative: a graph-to-text specific model and two state-of-the-art\nlanguage models, which previous work has shown to be adaptable to the knowledge\ngraph-to-text domain.', 'One of the key obstacles in making learning protocols realistic in\napplications is the need to supervise them, a costly process that often\nrequires hiring domain experts. We consider the framework to use the world\nknowledge as indirect supervision. World knowledge is general-purpose\nknowledge, which is not designed for any specific domain. Then the key\nchallenges are how to adapt the world knowledge to domains and how to represent\nit for learning. In this paper, we provide an example of using world knowledge\nfor domain dependent document clustering. We provide three ways to specify the\nworld knowledge to domains by resolving the ambiguity of the entities and their\ntypes, and represent the data with world knowledge as a heterogeneous\ninformation network. Then we propose a clustering algorithm that can cluster\nmultiple types and incorporate the sub-type information as constraints. In the\nexperiments, we use two existing knowledge bases as our sources of world\nknowledge. One is Freebase, which is collaboratively collected knowledge about\nentities and their organizations. The other is YAGO2, a knowledge base\nautomatically extracted from Wikipedia and maps knowledge to the linguistic\nknowledge base, WordNet. Experimental results on two text benchmark datasets\n(20newsgroups and RCV1) show that incorporating world knowledge as indirect\nsupervision can significantly outperform the state-of-the-art clustering\nalgorithms as well as clustering algorithms enhanced with world knowledge\nfeatures.', 'Knowledge bases of real-world facts about entities and their relationships\nare useful resources for a variety of natural language processing tasks.\nHowever, because knowledge bases are typically incomplete, it is useful to be\nable to perform link prediction or knowledge base completion, i.e., predict\nwhether a relationship not in the knowledge base is likely to be true. This\npaper combines insights from several previous link prediction models into a new\nembedding model STransE that represents each entity as a low-dimensional\nvector, and each relation by two matrices and a translation vector. STransE is\na simple combination of the SE and TransE models, but it obtains better link\nprediction performance on two benchmark datasets than previous embedding\nmodels. Thus, STransE can serve as a new baseline for the more complex models\nin the link prediction task.', 'Pre-trained word embeddings learned from unlabeled text have become a\nstandard component of neural network architectures for NLP tasks. However, in\nmost cases, the recurrent network that operates on word-level representations\nto produce context sensitive representations is trained on relatively little\nlabeled data. In this paper, we demonstrate a general semi-supervised approach\nfor adding pre- trained context embeddings from bidirectional language models\nto NLP systems and apply it to sequence labeling tasks. We evaluate our model\non two standard datasets for named entity recognition (NER) and chunking, and\nin both cases achieve state of the art results, surpassing previous systems\nthat use other forms of transfer or joint learning with additional labeled data\nand task specific gazetteers.', 'Answering complex questions over knowledge bases (KB-QA) faces huge input\ndata with billions of facts, involving millions of entities and thousands of\npredicates. For efficiency, QA systems first reduce the answer search space by\nidentifying a set of facts that is likely to contain all answers and relevant\ncues. The most common technique for doing this is to apply named entity\ndisambiguation (NED) systems to the question, and retrieve KB facts for the\ndisambiguated entities. This work presents CLOCQ, an efficient method that\nprunes irrelevant parts of the search space using KB-aware signals. CLOCQ uses\na top-k query processor over score-ordered lists of KB items that combine\nsignals about lexical matching, relevance to the question, coherence among\ncandidate items, and connectivity in the KB graph. Experiments with two recent\nQA benchmarks for complex questions demonstrate the superiority of CLOCQ over\nstate-of-the-art baselines with respect to answer presence, size of the search\nspace, and runtimes.', ""Effectively modeling text-rich fresh content such as news articles at\ndocument-level is a challenging problem. To ensure a content-based model\ngeneralize well to a broad range of applications, it is critical to have a\ntraining dataset that is large beyond the scale of human labels while achieving\ndesired quality. In this work, we address those two challenges by proposing a\nnovel approach to mine semantically-relevant fresh documents, and their topic\nlabels, with little human supervision. Meanwhile, we design a multitask model\ncalled NewsEmbed that alternatively trains a contrastive learning with a\nmulti-label classification to derive a universal document encoder. We show that\nthe proposed approach can provide billions of high quality organic training\nexamples and can be naturally extended to multilingual setting where texts in\ndifferent languages are encoded in the same semantic space. We experimentally\ndemonstrate NewsEmbed's competitive performance across multiple natural\nlanguage understanding tasks, both supervised and unsupervised."", ""Using deep learning for different machine learning tasks such as image\nclassification and word embedding has recently gained many attentions. Its\nappealing performance reported across specific Natural Language Processing\n(NLP) tasks in comparison with other approaches is the reason for its\npopularity. Word embedding is the task of mapping words or phrases to a low\ndimensional numerical vector. In this paper, we use deep learning to embed\nWikipedia Concepts and Entities. The English version of Wikipedia contains more\nthan five million pages, which suggest its capability to cover many English\nEntities, Phrases, and Concepts. Each Wikipedia page is considered as a\nconcept. Some concepts correspond to entities, such as a person's name, an\norganization or a place. Contrary to word embedding, Wikipedia Concepts\nEmbedding is not ambiguous, so there are different vectors for concepts with\nsimilar surface form but different mentions. We proposed several approaches and\nevaluated their performance based on Concept Analogy and Concept Similarity\ntasks. The results show that proposed approaches have the performance\ncomparable and in some cases even higher than the state-of-the-art methods."", ""Named Entity Recognition (NER) has been mostly studied in the context of\nwritten text. Specifically, NER is an important step in de-identification\n(de-ID) of medical records, many of which are recorded conversations between a\npatient and a doctor. In such recordings, audio spans with personal information\nshould be redacted, similar to the redaction of sensitive character spans in\nde-ID for written text. The application of NER in the context of audio\nde-identification has yet to be fully investigated. To this end, we define the\ntask of audio de-ID, in which audio spans with entity mentions should be\ndetected. We then present our pipeline for this task, which involves Automatic\nSpeech Recognition (ASR), NER on the transcript text, and text-to-audio\nalignment. Finally, we introduce a novel metric for audio de-ID and a new\nevaluation benchmark consisting of a large labeled segment of the Switchboard\nand Fisher audio datasets and detail our pipeline's results on it."", 'Given that in practice training data is scarce for all but a small set of\nproblems, a core question is how to incorporate prior knowledge into a model.\nIn this paper, we consider the case of prior procedural knowledge for neural\nnetworks, such as knowing how a program should traverse a sequence, but not\nwhat local actions should be performed at each step. To this end, we present an\nend-to-end differentiable interpreter for the programming language Forth which\nenables programmers to write program sketches with slots that can be filled\nwith behaviour trained from program input-output data. We can optimise this\nbehaviour directly through gradient descent techniques on user-specified\nobjectives, and also integrate the program into any larger neural computation\ngraph. We show empirically that our interpreter is able to effectively leverage\ndifferent levels of prior program structure and learn complex behaviours such\nas sequence sorting and addition. When connected to outputs of an LSTM and\ntrained jointly, our interpreter achieves state-of-the-art accuracy for\nend-to-end reasoning about quantities expressed in natural language stories.', 'This paper addresses the problem of corpus-level entity typing, i.e.,\ninferring from a large corpus that an entity is a member of a class such as\n""food"" or ""artist"". The application of entity typing we are interested in is\nknowledge base completion, specifically, to learn which classes an entity is a\nmember of. We propose FIGMENT to tackle this problem. FIGMENT is\nembedding-based and combines (i) a global model that scores based on aggregated\ncontextual information of an entity and (ii) a context model that first scores\nthe individual occurrences of an entity and then aggregates the scores. In our\nevaluation, FIGMENT strongly outperforms an approach to entity typing that\nrelies on relations obtained by an open information extraction system.', 'This survey presents a comprehensive description of recent neural entity\nlinking (EL) systems developed since 2015 as a result of the ""deep learning\nrevolution"" in natural language processing. Its goal is to systemize design\nfeatures of neural entity linking systems and compare their performance to the\nremarkable classic methods on common benchmarks. This work distills a generic\narchitecture of a neural EL system and discusses its components, such as\ncandidate generation, mention-context encoding, and entity ranking, summarizing\nprominent methods for each of them. The vast variety of modifications of this\ngeneral architecture are grouped by several common themes: joint entity mention\ndetection and disambiguation, models for global linking, domain-independent\ntechniques including zero-shot and distant supervision methods, and\ncross-lingual approaches. Since many neural models take advantage of entity and\nmention/context embeddings to represent their meaning, this work also overviews\nprominent entity embedding techniques. Finally, the survey touches on\napplications of entity linking, focusing on the recently emerged use-case of\nenhancing deep pre-trained masked language models based on the Transformer\narchitecture.', 'We address the task of Named Entity Disambiguation (NED) for noisy text. We\npresent WikilinksNED, a large-scale NED dataset of text fragments from the web,\nwhich is significantly noisier and more challenging than existing news-based\ndatasets. To capture the limited and noisy local context surrounding each\nmention, we design a neural model and train it with a novel method for sampling\ninformative negative examples. We also describe a new way of initializing word\nand entity embeddings that significantly improves performance. Our model\nsignificantly outperforms existing state-of-the-art methods on WikilinksNED\nwhile achieving comparable performance on a smaller newswire dataset.', 'Named entity recognition is an important task in natural language processing.\nIt is very well studied for rich language, but still under explored for\nlow-resource languages. The main reason is that the existing techniques\nrequired a lot of annotated data to reach good performance. Recently, a new\ndistributional representation of words has been proposed to project named\nentities from a rich language to a low-resource one. This representation has\nbeen coupled to a neural network in order to project named entities from\nEnglish to Ewondo, a Bantu language spoken in Cameroon. Although the proposed\nmethod reached appreciable results, the size of the used neural network was too\nlarge compared to the size of the dataset. Furthermore the impact of the model\nparameters has not been studied. In this paper, we show experimentally that the\nsame results can be obtained using a smaller neural network. We also emphasize\nthe parameters that are highly correlated to the network performance. This work\nis a step forward to build a reliable and robust network architecture for named\nentity projection in low resource languages.', 'Named Entity Recognition (NER) is a fundamental task in the fields of natural\nlanguage processing and information extraction. NER has been widely used as a\nstandalone tool or an essential component in a variety of applications such as\nquestion answering, dialogue assistants and knowledge graphs development.\nHowever, training reliable NER models requires a large amount of labelled data\nwhich is expensive to obtain, particularly in specialized domains. This paper\ndescribes a method to learn a domain-specific NER model for an arbitrary set of\nnamed entities when domain-specific supervision is not available. We assume\nthat the supervision can be obtained with no human effort, and neural models\ncan learn from each other. The code, data and models are publicly available.', 'This paper presents a discovery that the length of the entities in various\ndatasets follows a family of scale-free power law distributions. The concept of\nentity here broadly includes the named entity, entity mention, time expression,\naspect term, and domain-specific entity that are well investigated in natural\nlanguage processing and related areas. The entity length denotes the number of\nwords in an entity. The power law distributions in entity length possess the\nscale-free property and have well-defined means and finite variances. We\nexplain the phenomenon of power laws in entity length by the principle of least\neffort in communication and the preferential mechanism.', 'Constraint-based clustering algorithms exploit background knowledge to\nconstruct clusterings that are aligned with the interests of a particular user.\nThis background knowledge is often obtained by allowing the clustering system\nto pose pairwise queries to the user: should these two elements be in the same\ncluster or not? Active clustering methods aim to minimize the number of queries\nneeded to obtain a good clustering by querying the most informative pairs\nfirst. Ideally, a user should be able to answer a couple of these queries,\ninspect the resulting clustering, and repeat these two steps until a\nsatisfactory result is obtained. We present COBRAS, an approach to active\nclustering with pairwise constraints that is suited for such an interactive\nclustering process. A core concept in COBRAS is that of a super-instance: a\nlocal region in the data in which all instances are assumed to belong to the\nsame cluster. COBRAS constructs such super-instances in a top-down manner to\nproduce high-quality results early on in the clustering process, and keeps\nrefining these super-instances as more pairwise queries are given to get more\ndetailed clusterings later on. We experimentally demonstrate that COBRAS\nproduces good clusterings at fast run times, making it an excellent candidate\nfor the iterative clustering scenario outlined above.', 'Entity relatedness has emerged as an important feature in a plethora of\napplications such as information retrieval, entity recommendation and entity\nlinking. Given an entity, for instance a person or an organization, entity\nrelatedness measures can be exploited for generating a list of highly-related\nentities. However, the relation of an entity to some other entity depends on\nseveral factors, with time and context being two of the most important ones\n(where, in our case, context is determined by a particular corpus). For\nexample, the entities related to the International Monetary Fund are different\nnow compared to some years ago, while these entities also may highly differ in\nthe context of a USA news portal compared to a Greek news portal. In this\npaper, we propose a simple but flexible model for entity relatedness which\nconsiders time and entity aware word embeddings by exploiting the underlying\ncorpus. The proposed model does not require external knowledge and is language\nindependent, which makes it widely useful in a variety of applications.', 'Continual learning is the problem of sequentially learning new tasks or\nknowledge while protecting previously acquired knowledge. However, catastrophic\nforgetting poses a grand challenge for neural networks performing such learning\nprocess. Thus, neural networks that are deployed in the real world often\nstruggle in scenarios where the data distribution is non-stationary (concept\ndrift), imbalanced, or not always fully available, i.e., rare edge cases. We\npropose a Differentiable Hebbian Consolidation model which is composed of a\nDifferentiable Hebbian Plasticity (DHP) Softmax layer that adds a rapid\nlearning plastic component (compressed episodic memory) to the fixed (slow\nchanging) parameters of the softmax output layer; enabling learned\nrepresentations to be retained for a longer timescale. We demonstrate the\nflexibility of our method by integrating well-known task-specific synaptic\nconsolidation methods to penalize changes in the slow weights that are\nimportant for each target task. We evaluate our approach on the Permuted MNIST,\nSplit MNIST and Vision Datasets Mixture benchmarks, and introduce an imbalanced\nvariant of Permuted MNIST -- a dataset that combines the challenges of class\nimbalance and concept drift. Our proposed model requires no additional\nhyperparameters and outperforms comparable baselines by reducing forgetting.', 'Is pushing numbers on a single benchmark valuable in automatic speech\nrecognition? Research results in acoustic modeling are typically evaluated\nbased on performance on a single dataset. While the research community has\ncoalesced around various benchmarks, we set out to understand generalization\nperformance in acoustic modeling across datasets - in particular, if models\ntrained on a single dataset transfer to other (possibly out-of-domain)\ndatasets. We show that, in general, reverberative and additive noise\naugmentation improves generalization performance across domains. Further, we\ndemonstrate that when a large enough set of benchmarks is used, average word\nerror rate (WER) performance over them provides a good proxy for performance on\nreal-world noisy data. Finally, we show that training a single acoustic model\non the most widely-used datasets - combined - reaches competitive performance\non both research and real-world benchmarks.', ""Recent advances in natural language processing (NLP) have the ability to\ntransform how classroom learning takes place. Combined with the increasing\nintegration of technology in today's classrooms, NLP systems leveraging\nquestion answering and dialog processing techniques can serve as private tutors\nor participants in classroom discussions to increase student engagement and\nlearning. To progress towards this goal, we use the classroom discourse\nframework of academically productive talk (APT) to learn strategies that make\nfor the best learning experience. In this paper, we introduce a new task,\ncalled future talk move prediction (FTMP): it consists of predicting the next\ntalk move -- an utterance strategy from APT -- given a conversation history\nwith its corresponding talk moves. We further introduce a neural network model\nfor this task, which outperforms multiple baselines by a large margin. Finally,\nwe compare our model's performance on FTMP to human performance and show\nseveral similarities between the two."", 'In this work, we analyze how human gaze during reading comprehension is\nconditioned on the given reading comprehension question, and whether this\nsignal can be beneficial for machine reading comprehension. To this end, we\ncollect a new eye-tracking dataset with a large number of participants engaging\nin a multiple choice reading comprehension task. Our analysis of this data\nreveals increased fixation times over parts of the text that are most relevant\nfor answering the question. Motivated by this finding, we propose making\nautomated reading comprehension more human-like by mimicking human\ninformation-seeking reading behavior during reading comprehension. We\ndemonstrate that this approach leads to performance gains on multiple choice\nquestion answering in English for a state-of-the-art reading comprehension\nmodel.', 'Data processing is an important step in various natural language processing\ntasks. As the commonly used datasets in named entity recognition contain only a\nlimited number of samples, it is important to obtain additional labeled data in\nan efficient and reliable manner. A common practice is to utilize large\nmonolingual unlabeled corpora. Another popular technique is to create synthetic\ndata from the original labeled data (data augmentation). In this work, we\ninvestigate the impact of these two methods on the performance of three\ndifferent named entity recognition tasks.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'In this paper, we pose the question: do people talk about women and men in\ndifferent ways? We introduce two datasets and a novel integration of approaches\nfor automatically inferring gender associations from language, discovering\ncoherent word clusters, and labeling the clusters for the semantic concepts\nthey represent. The datasets allow us to compare how people write about women\nand men in two different settings - one set draws from celebrity news and the\nother from student reviews of computer science professors. We demonstrate that\nthere are large-scale differences in the ways that people talk about women and\nmen and that these differences vary across domains. Human evaluations show that\nour methods significantly outperform strong baselines.', 'We address the problem of solving mixed random linear equations. We have\nunlabeled observations coming from multiple linear regressions, and each\nobservation corresponds to exactly one of the regression models. The goal is to\nlearn the linear regressors from the observations. Classically, Alternating\nMinimization (AM) (which is a variant of Expectation Maximization (EM)) is used\nto solve this problem. AM iteratively alternates between the estimation of\nlabels and solving the regression problems with the estimated labels.\nEmpirically, it is observed that, for a large variety of non-convex problems\nincluding mixed linear regression, AM converges at a much faster rate compared\nto gradient based algorithms. However, the existing theory suggests similar\nrate of convergence for AM and gradient based methods, failing to capture this\nempirical behavior. In this paper, we close this gap between theory and\npractice for the special case of a mixture of $2$ linear regressions. We show\nthat, provided initialized properly, AM enjoys a \\emph{super-linear} rate of\nconvergence in certain parameter regimes. To the best of our knowledge, this is\nthe first work that theoretically establishes such rate for AM. Hence, if we\nwant to recover the unknown regressors upto an error (in $\\ell_2$ norm) of\n$\\epsilon$, AM only takes $\\mathcal{O}(\\log \\log (1/\\epsilon))$ iterations.\nFurthermore, we compare AM with a gradient based heuristic algorithm\nempirically and show that AM dominates in iteration complexity as well as\nwall-clock time.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : As a linguist, I aim to efficiently categorize various semantic elements within a sentence. For instance, I want to pinpoint the predicate-argument structure. I plan to create a model based on neural networks that enhances traditional methods which utilize syntactic trees. Rather than directly employing syntactic trees, my model will learn to convert syntax and grammatical structure into vectorized representations. These hidden representations will allow my model to be differentiable and converge rapidly. I'm also curious to see if using a semantic representation of my text from a large pre-trained language model could boost the accuracy of my labeling. Ultimately, I intend to test my model against a wide range of linguistic benchmarks.","['Semantic role labeling (SRL) is dedicated to recognizing the semantic\npredicate-argument structure of a sentence. Previous studies in terms of\ntraditional models have shown syntactic information can make remarkable\ncontributions to SRL performance; however, the necessity of syntactic\ninformation was challenged by a few recent neural SRL studies that demonstrate\nimpressive performance without syntactic backbones and suggest that syntax\ninformation becomes much less important for neural semantic role labeling,\nespecially when paired with recent deep neural network and large-scale\npre-trained language models. Despite this notion, the neural SRL field still\nlacks a systematic and full investigation on the relevance of syntactic\ninformation in SRL, for both dependency and both monolingual and multilingual\nsettings. This paper intends to quantify the importance of syntactic\ninformation for neural SRL in the deep learning framework. We introduce three\ntypical SRL frameworks (baselines), sequence-based, tree-based, and\ngraph-based, which are accompanied by two categories of exploiting syntactic\ninformation: syntax pruning-based and syntax feature-based. Experiments are\nconducted on the CoNLL-2005, 2009, and 2012 benchmarks for all languages\navailable, and results show that neural SRL models can still benefit from\nsyntactic information under certain conditions. Furthermore, we show the\nquantitative significance of syntax to neural SRL models together with a\nthorough empirical survey using existing models.', 'Semantic role labeling (SRL), also known as shallow semantic parsing, is an\nimportant yet challenging task in NLP. Motivated by the close correlation\nbetween syntactic and semantic structures, traditional discrete-feature-based\nSRL approaches make heavy use of syntactic features. In contrast,\ndeep-neural-network-based approaches usually encode the input sentence as a\nword sequence without considering the syntactic structures. In this work, we\ninvestigate several previous approaches for encoding syntactic trees, and make\na thorough study on whether extra syntax-aware representations are beneficial\nfor neural SRL models. Experiments on the benchmark CoNLL-2005 dataset show\nthat syntax-aware SRL approaches can effectively improve performance over a\nstrong baseline with external word representations from ELMo. With the extra\nsyntax-aware representations, our approaches achieve new state-of-the-art 85.6\nF1 (single model) and 86.6 F1 (ensemble) on the test data, outperforming the\ncorresponding strong baselines with ELMo by 0.8 and 1.0, respectively. Detailed\nerror analysis are conducted to gain more insights on the investigated\napproaches.', 'Current state-of-the-art semantic role labeling (SRL) uses a deep neural\nnetwork with no explicit linguistic features. However, prior work has shown\nthat gold syntax trees can dramatically improve SRL decoding, suggesting the\npossibility of increased accuracy from explicit modeling of syntax. In this\nwork, we present linguistically-informed self-attention (LISA): a neural\nnetwork model that combines multi-head self-attention with multi-task learning\nacross dependency parsing, part-of-speech tagging, predicate detection and SRL.\nUnlike previous models which require significant pre-processing to prepare\nlinguistic features, LISA can incorporate syntax using merely raw tokens as\ninput, encoding the sequence only once to simultaneously perform parsing,\npredicate detection and role labeling for all predicates. Syntax is\nincorporated by training one attention head to attend to syntactic parents for\neach token. Moreover, if a high-quality syntactic parse is already available,\nit can be beneficially injected at test time without re-training our SRL model.\nIn experiments on CoNLL-2005 SRL, LISA achieves new state-of-the-art\nperformance for a model using predicted predicates and standard word\nembeddings, attaining 2.5 F1 absolute higher than the previous state-of-the-art\non newswire and more than 3.5 F1 on out-of-domain data, nearly 10% reduction in\nerror. On ConLL-2012 English SRL we also show an improvement of more than 2.5\nF1. LISA also out-performs the state-of-the-art with contextually-encoded\n(ELMo) word representations, by nearly 1.0 F1 on news and more than 2.0 F1 on\nout-of-domain text.', 'The addition of syntax-aware decoding in Neural Machine Translation (NMT)\nsystems requires an effective tree-structured neural network, a syntax-aware\nattention model and a language generation model that is sensitive to sentence\nstructure. We exploit a top-down tree-structured model called DRNN\n(Doubly-Recurrent Neural Networks) first proposed by Alvarez-Melis and Jaakola\n(2017) to create an NMT model called Seq2DRNN that combines a sequential\nencoder with tree-structured decoding augmented with a syntax-aware attention\nmodel. Unlike previous approaches to syntax-based NMT which use dependency\nparsing models our method uses constituency parsing which we argue provides\nuseful information for translation. In addition, we use the syntactic structure\nof the sentence to add new connections to the tree-structured decoder neural\nnetwork (Seq2DRNN+SynC). We compare our NMT model with sequential and state of\nthe art syntax-based NMT models and show that our model produces more fluent\ntranslations with better reordering. Since our model is capable of doing\ntranslation and constituency parsing at the same time we also compare our\nparsing accuracy against other neural parsing models.', 'For natural language processing systems, two kinds of evidence support the\nuse of text representations from neural language models ""pretrained"" on large\nunannotated corpora: performance on application-inspired benchmarks (Peters et\nal., 2018, inter alia), and the emergence of syntactic abstractions in those\nrepresentations (Tenney et al., 2019, inter alia). On the other hand, the lack\nof grounded supervision calls into question how well these representations can\never capture meaning (Bender and Koller, 2020). We apply novel probes to recent\nlanguage models -- specifically focusing on predicate-argument structure as\noperationalized by semantic dependencies (Ivanova et al., 2012) -- and find\nthat, unlike syntax, semantics is not brought to the surface by today\'s\npretrained models. We then use convolutional graph encoders to explicitly\nincorporate semantic parses into task-specific finetuning, yielding benefits to\nnatural language understanding (NLU) tasks in the GLUE benchmark. This approach\ndemonstrates the potential for general-purpose (rather than task-specific)\nlinguistic supervision, above and beyond conventional pretraining and\nfinetuning. Several diagnostics help to localize the benefits of our approach.', 'Many natural language understanding (NLU) tasks, such as shallow parsing\n(i.e., text chunking) and semantic slot filling, require the assignment of\nrepresentative labels to the meaningful chunks in a sentence. Most of the\ncurrent deep neural network (DNN) based methods consider these tasks as a\nsequence labeling problem, in which a word, rather than a chunk, is treated as\nthe basic unit for labeling. These chunks are then inferred by the standard IOB\n(Inside-Outside-Beginning) labels. In this paper, we propose an alternative\napproach by investigating the use of DNN for sequence chunking, and propose\nthree neural models so that each chunk can be treated as a complete unit for\nlabeling. Experimental results show that the proposed neural sequence chunking\nmodels can achieve start-of-the-art performance on both the text chunking and\nslot filling tasks.', ""Tree-based Long short term memory (LSTM) network has become state-of-the-art\nfor modeling the meaning of language texts as they can effectively exploit the\ngrammatical syntax and thereby non-linear dependencies among words of the\nsentence. However, most of these models cannot recognize the difference in\nmeaning caused by a change in semantic roles of words or phrases because they\ndo not acknowledge the type of grammatical relations, also known as typed\ndependencies, in sentence structure. This paper proposes an enhanced LSTM\narchitecture, called relation gated LSTM, which can model the relationship\nbetween two inputs of a sequence using a control input. We also introduce a\nTree-LSTM model called Typed Dependency Tree-LSTM that uses the sentence\ndependency parse structure as well as the dependency type to embed sentence\nmeaning into a dense vector. The proposed model outperformed its type-unaware\ncounterpart in two typical NLP tasks - Semantic Relatedness Scoring and\nSentiment Analysis, in a lesser number of training epochs. The results were\ncomparable or competitive with other state-of-the-art models. Qualitative\nanalysis showed that changes in the voice of sentences had little effect on the\nmodel's predicted scores, while changes in nominal (noun) words had a more\nsignificant impact. The model recognized subtle semantic relationships in\nsentence pairs. The magnitudes of learned typed dependencies embeddings were\nalso in agreement with human intuitions. The research findings imply the\nsignificance of grammatical relations in sentence modeling. The proposed models\nwould serve as a base for future researches in this direction."", 'Deep neural networks for machine comprehension typically utilizes only word\nor character embeddings without explicitly taking advantage of structured\nlinguistic information such as constituency trees and dependency trees. In this\npaper, we propose structural embedding of syntactic trees (SEST), an algorithm\nframework to utilize structured information and encode them into vector\nrepresentations that can boost the performance of algorithms for the machine\ncomprehension. We evaluate our approach using a state-of-the-art neural\nattention model on the SQuAD dataset. Experimental results demonstrate that our\nmodel can accurately identify the syntactic boundaries of the sentences and\nextract answers that are syntactically coherent over the baseline methods.', ""Progress in pre-trained language models has led to a surge of impressive\nresults on downstream tasks for natural language understanding. Recent work on\nprobing pre-trained language models uncovered a wide range of linguistic\nproperties encoded in their contextualized representations. However, it is\nunclear whether they encode semantic knowledge that is crucial to symbolic\ninference methods. We propose a methodology for probing linguistic information\nfor logical inference in pre-trained language model representations. Our\nprobing datasets cover a list of linguistic phenomena required by major\nsymbolic inference systems. We find that (i) pre-trained language models do\nencode several types of linguistic information for inference, but there are\nalso some types of information that are weakly encoded, (ii) language models\ncan effectively learn missing linguistic information through fine-tuning.\nOverall, our findings provide insights into which aspects of linguistic\ninformation for logical inference do language models and their pre-training\nprocedures capture. Moreover, we have demonstrated language models' potential\nas semantic and background knowledge bases for supporting symbolic inference\nmethods."", 'Due to the increasing amount of data on the internet, finding a\nhighly-informative, low-dimensional representation for text is one of the main\nchallenges for efficient natural language processing tasks including text\nclassification. This representation should capture the semantic information of\nthe text while retaining their relevance level for document classification.\nThis approach maps the documents with similar topics to a similar space in\nvector space representation. To obtain representation for large text, we\npropose the utilization of deep Siamese neural networks. To embed document\nrelevance in topics in the distributed representation, we use a Siamese neural\nnetwork to jointly learn document representations. Our Siamese network consists\nof two sub-network of multi-layer perceptron. We examine our representation for\nthe text categorization task on BBC news dataset. The results show that the\nproposed representations outperform the conventional and state-of-the-art\nrepresentations in the text classification task on this dataset.', 'We propose a generative model for a sentence that uses two latent variables,\nwith one intended to represent the syntax of the sentence and the other to\nrepresent its semantics. We show we can achieve better disentanglement between\nsemantic and syntactic representations by training with multiple losses,\nincluding losses that exploit aligned paraphrastic sentences and word-order\ninformation. We also investigate the effect of moving from bag-of-words to\nrecurrent neural network modules. We evaluate our models as well as several\npopular pretrained embeddings on standard semantic similarity tasks and novel\nsyntactic similarity tasks. Empirically, we find that the model with the best\nperforming syntactic and semantic representations also gives rise to the most\ndisentangled representations.', 'Neural networks trained on natural language processing tasks capture syntax\neven though it is not provided as a supervision signal. This indicates that\nsyntactic analysis is essential to the understating of language in artificial\nintelligence systems. This overview paper covers approaches of evaluating the\namount of syntactic information included in the representations of words for\ndifferent neural network architectures. We mainly summarize re-search on\nEnglish monolingual data on language modeling tasks and multilingual data for\nneural machine translation systems and multilingual language models. We\ndescribe which pre-trained models and representations of language are best\nsuited for transfer to syntactic tasks.', 'We study the problem of leveraging the syntactic structure of text to enhance\npre-trained models such as BERT and RoBERTa. Existing methods utilize syntax of\ntext either in the pre-training stage or in the fine-tuning stage, so that they\nsuffer from discrepancy between the two stages. Such a problem would lead to\nthe necessity of having human-annotated syntactic information, which limits the\napplication of existing methods to broader scenarios. To address this, we\npresent a model that utilizes the syntax of text in both pre-training and\nfine-tuning stages. Our model is based on Transformer with a syntax-aware\nattention layer that considers the dependency tree of the text. We further\nintroduce a new pre-training task of predicting the syntactic distance among\ntokens in the dependency tree. We evaluate the model on three downstream tasks,\nincluding relation classification, entity typing, and question answering.\nResults show that our model achieves state-of-the-art performance on six public\nbenchmark datasets. We have two major findings. First, we demonstrate that\ninfusing automatically produced syntax of text improves pre-trained models.\nSecond, global syntactic distances among tokens bring larger performance gains\ncompared to local head relations between contiguous tokens.', 'The latest developments in neural semantic role labeling (SRL) have shown\ngreat performance improvements with both the dependency and span\nformalisms/styles. Although the two styles share many similarities in\nlinguistic meaning and computation, most previous studies focus on a single\nstyle. In this paper, we define a new cross-style semantic role label\nconvention and propose a new cross-style joint optimization model designed\naround the most basic linguistic meaning of a semantic role, providing a\nsolution to make the results of the two styles more comparable and allowing\nboth formalisms of SRL to benefit from their natural connections in both\nlinguistics and computation. Our model learns a general semantic argument\nstructure and is capable of outputting in either style. Additionally, we\npropose a syntax-aided method to uniformly enhance the learning of both\ndependency and span representations. Experiments show that the proposed methods\nare effective on both span and dependency SRL benchmarks.', 'Sense representations have gone beyond word representations like Word2Vec,\nGloVe and FastText and achieved innovative performance on a wide range of\nnatural language processing tasks. Although very useful in many applications,\nthe traditional approaches for generating word embeddings have a strict\ndrawback: they produce a single vector representation for a given word ignoring\nthe fact that ambiguous words can assume different meanings. In this paper, we\nexplore unsupervised sense representations which, different from traditional\nword embeddings, are able to induce different senses of a word by analyzing its\ncontextual semantics in a text. The unsupervised sense representations\ninvestigated in this paper are: sense embeddings and deep neural language\nmodels. We present the first experiments carried out for generating sense\nembeddings for Portuguese. Our experiments show that the sense embedding model\n(Sense2vec) outperformed traditional word embeddings in syntactic and semantic\nanalogies task, proving that the language resource generated here can improve\nthe performance of NLP tasks in Portuguese. We also evaluated the performance\nof pre-trained deep neural language models (ELMo and BERT) in two transfer\nlearning approaches: feature based and fine-tuning, in the semantic textual\nsimilarity task. Our experiments indicate that the fine tuned Multilingual and\nPortuguese BERT language models were able to achieve better accuracy than the\nELMo model and baselines.', 'Existing neural semantic parsers mainly utilize a sequence encoder, i.e., a\nsequential LSTM, to extract word order features while neglecting other valuable\nsyntactic information such as dependency graph or constituent trees. In this\npaper, we first propose to use the \\textit{syntactic graph} to represent three\ntypes of syntactic information, i.e., word order, dependency and constituency\nfeatures. We further employ a graph-to-sequence model to encode the syntactic\ngraph and decode a logical form. Experimental results on benchmark datasets\nshow that our model is comparable to the state-of-the-art on Jobs640, ATIS and\nGeo880. Experimental results on adversarial examples demonstrate the robustness\nof the model is also improved by encoding more syntactic information.', 'Semantic role labeling (SRL) is the task of identifying the\npredicate-argument structure of a sentence. It is typically regarded as an\nimportant step in the standard NLP pipeline. As the semantic representations\nare closely related to syntactic ones, we exploit syntactic information in our\nmodel. We propose a version of graph convolutional networks (GCNs), a recent\nclass of neural networks operating on graphs, suited to model syntactic\ndependency graphs. GCNs over syntactic dependency trees are used as sentence\nencoders, producing latent feature representations of words in a sentence. We\nobserve that GCN layers are complementary to LSTM ones: when we stack both GCN\nand LSTM layers, we obtain a substantial improvement over an already\nstate-of-the-art LSTM SRL model, resulting in the best reported scores on the\nstandard benchmark (CoNLL-2009) both for Chinese and English.', 'Syntactic structure of sentences in a document substantially informs about\nits authorial writing style. Sentence representation learning has been widely\nexplored in recent years and it has been shown that it improves the\ngeneralization of different downstream tasks across many domains. Even though\nutilizing probing methods in several studies suggests that these learned\ncontextual representations implicitly encode some amount of syntax, explicit\nsyntactic information further improves the performance of deep neural models in\nthe domain of authorship attribution. These observations have motivated us to\ninvestigate the explicit representation learning of syntactic structure of\nsentences. In this paper, we propose a self-supervised framework for learning\nstructural representations of sentences. The self-supervised network contains\ntwo components; a lexical sub-network and a syntactic sub-network which take\nthe sequence of words and their corresponding structural labels as the input,\nrespectively. Due to the n-to-1 mapping of words to their structural labels,\neach word will be embedded into a vector representation which mainly carries\nstructural information. We evaluate the learned structural representations of\nsentences using different probing tasks, and subsequently utilize them in the\nauthorship attribution task. Our experimental results indicate that the\nstructural embeddings significantly improve the classification tasks when\nconcatenated with the existing pre-trained word embeddings.', 'Machine translation systems require semantic knowledge and grammatical\nunderstanding. Neural machine translation (NMT) systems often assume this\ninformation is captured by an attention mechanism and a decoder that ensures\nfluency. Recent work has shown that incorporating explicit syntax alleviates\nthe burden of modeling both types of knowledge. However, requiring parses is\nexpensive and does not explore the question of what syntax a model needs during\ntranslation. To address both of these issues we introduce a model that\nsimultaneously translates while inducing dependency trees. In this way, we\nleverage the benefits of structure while investigating what syntax NMT must\ninduce to maximize performance. We show that our dependency trees are 1.\nlanguage pair dependent and 2. improve translation quality.', 'Sentence semantic understanding is a key topic in the field of natural\nlanguage processing. Recently, contextualized word representations derived from\npre-trained language models such as ELMO and BERT have shown significant\nimprovements for a wide range of semantic tasks, e.g. question answering, text\nclassification and sentiment analysis. However, how to add external knowledge\nto further improve the semantic modeling capability of model is worth probing.\nIn this paper, we propose a novel approach to combining syntax information with\na pre-trained language model. In order to evaluate the effect of the\npre-training model, first, we introduce RNN-based and Transformer-based\npre-trained language models; secondly, to better integrate external knowledge,\nsuch as syntactic information integrate with the pre-training model, we propose\na dependency syntax expansion (DSE) model. For evaluation, we have selected two\nsubtasks: sentence completion task and biological relation extraction task. The\nexperimental results show that our model achieves 91.2\\% accuracy,\noutperforming the baseline model by 37.8\\% on sentence completion task. And it\nalso gets competitive performance by 75.1\\% $F_{1}$ score on relation\nextraction task.', 'While a lot of analysis has been carried to demonstrate linguistic knowledge\ncaptured by the representations learned within deep NLP models, very little\nattention has been paid towards individual neurons.We carry outa neuron-level\nanalysis using core linguistic tasks of predicting morphology, syntax and\nsemantics, on pre-trained language models, with questions like: i) do\nindividual neurons in pre-trained models capture linguistic information? ii)\nwhich parts of the network learn more about certain linguistic phenomena? iii)\nhow distributed or focused is the information? and iv) how do various\narchitectures differ in learning these properties? We found small subsets of\nneurons to predict linguistic tasks, with lower level tasks (such as\nmorphology) localized in fewer neurons, compared to higher level task of\npredicting syntax. Our study also reveals interesting cross architectural\ncomparisons. For example, we found neurons in XLNet to be more localized and\ndisjoint when predicting properties compared to BERT and others, where they are\nmore distributed and coupled.', 'Writing style is a combination of consistent decisions at different levels of\nlanguage production including lexical, syntactic, and structural associated to\na specific author (or author groups). While lexical-based models have been\nwidely explored in style-based text classification, relying on content makes\nthe model less scalable when dealing with heterogeneous data comprised of\nvarious topics. On the other hand, syntactic models which are\ncontent-independent, are more robust against topic variance. In this paper, we\nintroduce a syntactic recurrent neural network to encode the syntactic patterns\nof a document in a hierarchical structure. The model first learns the syntactic\nrepresentation of sentences from the sequence of part-of-speech tags. For this\npurpose, we exploit both convolutional filters and long short-term memories to\ninvestigate the short-term and long-term dependencies of part-of-speech tags in\nthe sentences. Subsequently, the syntactic representations of sentences are\naggregated into document representation using recurrent neural networks. Our\nexperimental results on PAN 2012 dataset for authorship attribution task shows\nthat syntactic recurrent neural network outperforms the lexical model with the\nidentical architecture by approximately 14% in terms of accuracy.', ""This paper investigates whether the power of the models pre-trained on text\ndata, such as BERT, can be transferred to general token sequence classification\napplications. To verify pre-trained models' transferability, we test the\npre-trained models on text classification tasks with meanings of tokens\nmismatches, and real-world non-text token sequence classification data,\nincluding amino acid, DNA, and music. We find that even on non-text data, the\nmodels pre-trained on text converge faster, perform better than the randomly\ninitialized models, and only slightly worse than the models using task-specific\nknowledge. We also find that the representations of the text and non-text\npre-trained models share non-trivial similarities."", 'We study the problem of integrating syntactic information from constituency\ntrees into a neural model in Frame-semantic parsing sub-tasks, namely Target\nIdentification (TI), FrameIdentification (FI), and Semantic Role Labeling\n(SRL). We use a Graph Convolutional Network to learn specific representations\nof constituents, such that each constituent is profiled as the production\ngrammar rule it corresponds to. We leverage these representations to build\nsyntactic features for each word in a sentence, computed as the sum of all the\nconstituents on the path between a word and a task-specific node in the tree,\ne.g. the target predicate for SRL. Our approach improves state-of-the-art\nresults on the TI and SRL of ~1%and~3.5% points, respectively (+2.5% additional\npoints are gained with BERT as input), when tested on FrameNet 1.5, while\nyielding comparable results on the CoNLL05 dataset to other syntax-aware\nsystems.', 'We study methods for learning sentence embeddings with syntactic structure.\nWe focus on methods of learning syntactic sentence-embeddings by using a\nmultilingual parallel-corpus augmented by Universal Parts-of-Speech tags. We\nevaluate the quality of the learned embeddings by examining sentence-level\nnearest neighbours and functional dissimilarity in the embedding space. We also\nevaluate the ability of the method to learn syntactic sentence-embeddings for\nlow-resource languages and demonstrate strong evidence for transfer learning.\nOur results show that syntactic sentence-embeddings can be learned while using\nless training data, fewer model parameters, and resulting in better evaluation\nmetrics than state-of-the-art language models.', 'Neuroscientists evaluate deep neural networks for natural language processing\nas possible candidate models for how language is processed in the brain. These\nmodels are often trained without explicit linguistic supervision, but have been\nshown to learn some linguistic structure in the absence of such supervision\n(Manning et al., 2020), potentially questioning the relevance of symbolic\nlinguistic theories in modeling such cognitive processes (Warstadt and Bowman,\n2020). We evaluate across two fMRI datasets whether language models align\nbetter with brain recordings, if their attention is biased by annotations from\nsyntactic or semantic formalisms. Using structure from dependency or minimal\nrecursion semantic annotations, we find alignments improve significantly for\none of the datasets. For another dataset, we see more mixed results. We present\nan extensive analysis of these results. Our proposed approach enables the\nevaluation of more targeted hypotheses about the composition of meaning in the\nbrain, expanding the range of possible scientific inferences a neuroscientist\ncould make, and opens up new opportunities for cross-pollination between\ncomputational neuroscience and linguistics.', 'Text classification is one of the most widely studied tasks in natural\nlanguage processing. Motivated by the principle of compositionality, large\nmultilayer neural network models have been employed for this task in an attempt\nto effectively utilize the constituent expressions. Almost all of the reported\nwork train large networks using discriminative approaches, which come with a\ncaveat of no proper capacity control, as they tend to latch on to any signal\nthat may not generalize. Using various recent state-of-the-art approaches for\ntext classification, we explore whether these models actually learn to compose\nthe meaning of the sentences or still just focus on some keywords or lexicons\nfor classifying the document. To test our hypothesis, we carefully construct\ndatasets where the training and test splits have no direct overlap of such\nlexicons, but overall language structure would be similar. We study various\ntext classifiers and observe that there is a big performance drop on these\ndatasets. Finally, we show that even simple models with our proposed\nregularization techniques, which disincentivize focusing on key lexicons, can\nsubstantially improve classification accuracy.', 'In natural languages, words are used in association to construct sentences.\nIt is not words in isolation, but the appropriate combination of hierarchical\nstructures that conveys the meaning of the whole sentence. Neural networks can\ncapture expressive language features; however, insights into the link between\nwords and sentences are difficult to acquire automatically. In this work, we\ndesign a deep neural network architecture that explicitly wires lower and\nhigher linguistic components; we then evaluate its ability to perform the same\ntask at different hierarchical levels. Settling on broad text classification\ntasks, we show that our model, MHAL, learns to simultaneously solve them at\ndifferent levels of granularity by fluidly transferring knowledge between\nhierarchies. Using a multi-head attention mechanism to tie the representations\nbetween single words and full sentences, MHAL systematically outperforms\nequivalent models that are not incentivized towards developing compositional\nrepresentations. Moreover, we demonstrate that, with the proposed architecture,\nthe sentence information flows naturally to individual words, allowing the\nmodel to behave like a sequence labeller (which is a lower, word-level task)\neven without any word supervision, in a zero-shot fashion.', 'We propose a neural language model capable of unsupervised syntactic\nstructure induction. The model leverages the structure information to form\nbetter semantic representations and better language modeling. Standard\nrecurrent neural networks are limited by their structure and fail to\nefficiently use syntactic information. On the other hand, tree-structured\nrecursive networks usually require additional structural supervision at the\ncost of human expert annotation. In this paper, We propose a novel neural\nlanguage model, called the Parsing-Reading-Predict Networks (PRPN), that can\nsimultaneously induce the syntactic structure from unannotated sentences and\nleverage the inferred structure to learn a better language model. In our model,\nthe gradient can be directly back-propagated from the language model loss into\nthe neural parsing network. Experiments show that the proposed model can\ndiscover the underlying syntactic structure and achieve state-of-the-art\nperformance on word/character-level language model tasks.', 'This paper proposes a neural semantic parsing approach -- Sequence-to-Action,\nwhich models semantic parsing as an end-to-end semantic graph generation\nprocess. Our method simultaneously leverages the advantages from two recent\npromising directions of semantic parsing. Firstly, our model uses a semantic\ngraph to represent the meaning of a sentence, which has a tight-coupling with\nknowledge bases. Secondly, by leveraging the powerful representation learning\nand prediction ability of neural network models, we propose a RNN model which\ncan effectively map sentences to action sequences for semantic graph\ngeneration. Experiments show that our method achieves state-of-the-art\nperformance on OVERNIGHT dataset and gets competitive performance on GEO and\nATIS datasets.', 'We present novel methods for analyzing the activation patterns of RNNs from a\nlinguistic point of view and explore the types of linguistic structure they\nlearn. As a case study, we use a multi-task gated recurrent network\narchitecture consisting of two parallel pathways with shared word embeddings\ntrained on predicting the representations of the visual scene corresponding to\nan input sentence, and predicting the next word in the same sentence. Based on\nour proposed method to estimate the amount of contribution of individual tokens\nin the input to the final prediction of the networks we show that the image\nprediction pathway: a) is sensitive to the information structure of the\nsentence b) pays selective attention to lexical categories and grammatical\nfunctions that carry semantic information c) learns to treat the same input\ntoken differently depending on its grammatical functions in the sentence. In\ncontrast the language model is comparatively more sensitive to words with a\nsyntactic function. Furthermore, we propose methods to ex- plore the function\nof individual hidden units in RNNs and show that the two pathways of the\narchitecture in our case study contain specialized units tuned to patterns\ninformative for the task, some of which can carry activations to later time\nsteps to encode long-term dependencies.', 'This paper describes a neural semantic parser that maps natural language\nutterances onto logical forms which can be executed against a task-specific\nenvironment, such as a knowledge base or a database, to produce a response. The\nparser generates tree-structured logical forms with a transition-based approach\nwhich combines a generic tree-generation algorithm with domain-general\noperations defined by the logical language. The generation process is modeled\nby structured recurrent neural networks, which provide a rich encoding of the\nsentential context and generation history for making predictions. To tackle\nmismatches between natural language and logical form tokens, various attention\nmechanisms are explored. Finally, we consider different training settings for\nthe neural semantic parser, including a fully supervised training where\nannotated logical forms are given, weakly-supervised training where denotations\nare provided, and distant supervision where only unlabeled sentences and a\nknowledge base are available. Experiments across a wide range of datasets\ndemonstrate the effectiveness of our parser.', 'While state-of-the-art neural network models continue to achieve lower\nperplexity scores on language modeling benchmarks, it remains unknown whether\noptimizing for broad-coverage predictive performance leads to human-like\nsyntactic knowledge. Furthermore, existing work has not provided a clear\npicture about the model properties required to produce proper syntactic\ngeneralizations. We present a systematic evaluation of the syntactic knowledge\nof neural language models, testing 20 combinations of model types and data\nsizes on a set of 34 English-language syntactic test suites. We find\nsubstantial differences in syntactic generalization performance by model\narchitecture, with sequential models underperforming other architectures.\nFactorially manipulating model architecture and training dataset size (1M--40M\nwords), we find that variability in syntactic generalization performance is\nsubstantially greater by architecture than by dataset size for the corpora\ntested in our experiments. Our results also reveal a dissociation between\nperplexity and syntactic generalization performance.']","['We present a set of experiments to demonstrate that deep recurrent neural\nnetworks (RNNs) learn internal representations that capture soft hierarchical\nnotions of syntax from highly varied supervision. We consider four syntax tasks\nat different depths of the parse tree; for each word, we predict its part of\nspeech as well as the first (parent), second (grandparent) and third level\n(great-grandparent) constituent labels that appear above it. These predictions\nare made from representations produced at different depths in networks that are\npretrained with one of four objectives: dependency parsing, semantic role\nlabeling, machine translation, or language modeling. In every case, we find a\ncorrespondence between network depth and syntactic depth, suggesting that a\nsoft syntactic hierarchy emerges. This effect is robust across all conditions,\nindicating that the models encode significant amounts of syntax even in the\nabsence of an explicit syntactic training supervision.', 'Grammatical Error Correction (GEC) is a task of detecting and correcting\ngrammatical errors in sentences. Recently, neural machine translation systems\nhave become popular approaches for this task. However, these methods lack the\nuse of syntactic knowledge which plays an important role in the correction of\ngrammatical errors. In this work, we propose a syntax-guided GEC model (SG-GEC)\nwhich adopts the graph attention mechanism to utilize the syntactic knowledge\nof dependency trees. Considering the dependency trees of the grammatically\nincorrect source sentences might provide incorrect syntactic knowledge, we\npropose a dependency tree correction task to deal with it. Combining with data\naugmentation method, our model achieves strong performances without using any\nlarge pre-trained models. We evaluate our model on public benchmarks of GEC\ntask and it achieves competitive results.', 'Building robust natural language understanding systems will require a clear\ncharacterization of whether and how various linguistic meaning representations\ncomplement each other. To perform a systematic comparative analysis, we\nevaluate the mapping between meaning representations from different frameworks\nusing two complementary methods: (i) a rule-based converter, and (ii) a\nsupervised delexicalized parser that parses to one framework using only\ninformation from the other as features. We apply these methods to convert the\nSTREUSLE corpus (with syntactic and lexical semantic annotations) to UCCA (a\ngraph-structured full-sentence meaning representation). Both methods yield\nsurprisingly accurate target representations, close to fully supervised UCCA\nparser quality---indicating that UCCA annotations are partially redundant with\nSTREUSLE annotations. Despite this substantial convergence between frameworks,\nwe find several important areas of divergence.', ""The success of long short-term memory (LSTM) neural networks in language\nprocessing is typically attributed to their ability to capture long-distance\nstatistical regularities. Linguistic regularities are often sensitive to\nsyntactic structure; can such dependencies be captured by LSTMs, which do not\nhave explicit structural representations? We begin addressing this question\nusing number agreement in English subject-verb dependencies. We probe the\narchitecture's grammatical competence both using training objectives with an\nexplicit grammatical target (number prediction, grammaticality judgments) and\nusing language models. In the strongly supervised settings, the LSTM achieved\nvery high overall accuracy (less than 1% errors), but errors increased when\nsequential and structural information conflicted. The frequency of such errors\nrose sharply in the language-modeling setting. We conclude that LSTMs can\ncapture a non-trivial amount of grammatical structure given targeted\nsupervision, but stronger architectures may be required to further reduce\nerrors; furthermore, the language modeling signal is insufficient for capturing\nsyntax-sensitive dependencies, and should be supplemented with more direct\nsupervision if such dependencies need to be captured."", ""We introduce a new test of how well language models capture meaning in\nchildren's books. Unlike standard language modelling benchmarks, it\ndistinguishes the task of predicting syntactic function words from that of\npredicting lower-frequency words, which carry greater semantic content. We\ncompare a range of state-of-the-art models, each with a different way of\nencoding what has been previously read. We show that models which store\nexplicit representations of long-term contexts outperform state-of-the-art\nneural language models at predicting semantic content words, although this\nadvantage is not observed for syntactic function words. Interestingly, we find\nthat the amount of text encoded in a single memory representation is highly\ninfluential to the performance: there is a sweet-spot, not too big and not too\nsmall, between single words and full sentences that allows the most meaningful\ninformation in a text to be effectively retained and recalled. Further, the\nattention over such window-based memories can be trained effectively through\nself-supervision. We then assess the generality of this principle by applying\nit to the CNN QA benchmark, which involves identifying named entities in\nparaphrased summaries of news articles, and achieve state-of-the-art\nperformance."", 'In this dissertation we report results of our research on dense distributed\nrepresentations of text data. We propose two novel neural models for learning\nsuch representations. The first model learns representations at the document\nlevel, while the second model learns word-level representations.\n  For document-level representations we propose Binary Paragraph Vector: a\nneural network models for learning binary representations of text documents,\nwhich can be used for fast document retrieval. We provide a thorough evaluation\nof these models and demonstrate that they outperform the seminal method in the\nfield in the information retrieval task. We also report strong results in\ntransfer learning settings, where our models are trained on a generic text\ncorpus and then used to infer codes for documents from a domain-specific\ndataset. In contrast to previously proposed approaches, Binary Paragraph Vector\nmodels learn embeddings directly from raw text data.\n  For word-level representations we propose Disambiguated Skip-gram: a neural\nnetwork model for learning multi-sense word embeddings. Representations learned\nby this model can be used in downstream tasks, like part-of-speech tagging or\nidentification of semantic relations. In the word sense induction task\nDisambiguated Skip-gram outperforms state-of-the-art models on three out of\nfour benchmarks datasets. Our model has an elegant probabilistic\ninterpretation. Furthermore, unlike previous models of this kind, it is\ndifferentiable with respect to all its parameters and can be trained with\nbackpropagation. In addition to quantitative results, we present qualitative\nevaluation of Disambiguated Skip-gram, including two-dimensional visualisations\nof selected word-sense embeddings.', 'Question Answering with NLP has progressed through the evolution of advanced\nmodel architectures like BERT and BiDAF and earlier word, character, and\ncontext-based embeddings. As BERT has leapfrogged the accuracy of models, an\nelement of the next frontier can be the introduction of deep networks and an\neffective way to train them. In this context, I explored the effectiveness of\ndeep networks focussing on the model encoder layer of BiDAF. BiDAF with its\nheterogeneous layers provides the opportunity not only to explore the\neffectiveness of deep networks but also to evaluate whether the refinements\nmade in lower layers are additive to the refinements made in the upper layers\nof the model architecture. I believe the next greatest model in NLP will in\nfact fold in a solid language modeling like BERT with a composite architecture\nwhich will bring in refinements in addition to generic language modeling and\nwill have a more extensive layered architecture. I experimented with the Bypass\nnetwork, Residual Highway network, and DenseNet architectures. In addition, I\nevaluated the effectiveness of ensembling the last few layers of the network. I\nalso studied the difference character embeddings make in adding them to the\nword embeddings, and whether the effects are additive with deep networks. My\nstudies indicate that deep networks are in fact effective in giving a boost.\nAlso, the refinements in the lower layers like embeddings are passed on\nadditively to the gains made through deep networks.', 'We propose the Neural Vector Space Model (NVSM), a method that learns\nrepresentations of documents in an unsupervised manner for news article\nretrieval. In the NVSM paradigm, we learn low-dimensional representations of\nwords and documents from scratch using gradient descent and rank documents\naccording to their similarity with query representations that are composed from\nword representations. We show that NVSM performs better at document ranking\nthan existing latent semantic vector space methods. The addition of NVSM to a\nmixture of lexical language models and a state-of-the-art baseline vector space\nmodel yields a statistically significant increase in retrieval effectiveness.\nConsequently, NVSM adds a complementary relevance signal. Next to semantic\nmatching, we find that NVSM performs well in cases where lexical matching is\nneeded.\n  NVSM learns a notion of term specificity directly from the document\ncollection without feature engineering. We also show that NVSM learns\nregularities related to Luhn significance. Finally, we give advice on how to\ndeploy NVSM in situations where model selection (e.g., cross-validation) is\ninfeasible. We find that an unsupervised ensemble of multiple models trained\nwith different hyperparameter values performs better than a single\ncross-validated model. Therefore, NVSM can safely be used for ranking documents\nwithout supervised relevance judgments.', 'Recent advances in language modeling using deep neural networks have shown\nthat these models learn representations, that vary with the network depth from\nmorphology to semantic relationships like co-reference. We apply pre-trained\nlanguage models to low-resource named entity recognition for Historic German.\nWe show on a series of experiments that character-based pre-trained language\nmodels do not run into trouble when faced with low-resource datasets. Our\npre-trained character-based language models improve upon classical CRF-based\nmethods and previous work on Bi-LSTMs by boosting F1 score performance by up to\n6%. Our pre-trained language and NER models are publicly available under\nhttps://github.com/stefan-it/historic-ner .', 'Syntax has been demonstrated highly effective in neural machine translation\n(NMT). Previous NMT models integrate syntax by representing 1-best tree outputs\nfrom a well-trained parsing system, e.g., the representative Tree-RNN and\nTree-Linearization methods, which may suffer from error propagation. In this\nwork, we propose a novel method to integrate source-side syntax implicitly for\nNMT. The basic idea is to use the intermediate hidden representations of a\nwell-trained end-to-end dependency parser, which are referred to as\nsyntax-aware word representations (SAWRs). Then, we simply concatenate such\nSAWRs with ordinary word embeddings to enhance basic NMT models. The method can\nbe straightforwardly integrated into the widely-used sequence-to-sequence\n(Seq2Seq) NMT models. We start with a representative RNN-based Seq2Seq baseline\nsystem, and test the effectiveness of our proposed method on two benchmark\ndatasets of the Chinese-English and English-Vietnamese translation tasks,\nrespectively. Experimental results show that the proposed approach is able to\nbring significant BLEU score improvements on the two datasets compared with the\nbaseline, 1.74 points for Chinese-English translation and 0.80 point for\nEnglish-Vietnamese translation, respectively. In addition, the approach also\noutperforms the explicit Tree-RNN and Tree-Linearization methods.', 'Semantic parsing is the process of mapping a natural language sentence into a\nformal representation of its meaning. In this work we use the neural network\napproach to transform natural language sentence into a query to an ontology\ndatabase in the SPARQL language. This method does not rely on handcraft-rules,\nhigh-quality lexicons, manually-built templates or other handmade complex\nstructures. Our approach is based on vector space model and neural networks.\nThe proposed model is based in two learning steps. The first step generates a\nvector representation for the sentence in natural language and SPARQL query.\nThe second step uses this vector representation as input to a neural network\n(LSTM with attention mechanism) to generate a model able to encode natural\nlanguage and decode SPARQL.', 'Recursive neural network models and their accompanying vector representations\nfor words have seen success in an array of increasingly semantically\nsophisticated tasks, but almost nothing is known about their ability to\naccurately capture the aspects of linguistic meaning that are necessary for\ninterpretation or reasoning. To evaluate this, I train a recursive model on a\nnew corpus of constructed examples of logical reasoning in short sentences,\nlike the inference of ""some animal walks"" from ""some dog walks"" or ""some cat\nwalks,"" given that dogs and cats are animals. This model learns representations\nthat generalize well to new types of reasoning pattern in all but a few cases,\na result which is promising for the ability of learned representation models to\ncapture logical reasoning.', ""Semantic sentence embedding models encode natural language sentences into\nvectors, such that closeness in embedding space indicates closeness in the\nsemantics between the sentences. Bilingual data offers a useful signal for\nlearning such embeddings: properties shared by both sentences in a translation\npair are likely semantic, while divergent properties are likely stylistic or\nlanguage-specific. We propose a deep latent variable model that attempts to\nperform source separation on parallel sentences, isolating what they have in\ncommon in a latent semantic vector, and explaining what is left over with\nlanguage-specific latent vectors. Our proposed approach differs from past work\non semantic sentence encoding in two ways. First, by using a variational\nprobabilistic framework, we introduce priors that encourage source separation,\nand can use our model's posterior to predict sentence embeddings for\nmonolingual data at test time. Second, we use high-capacity transformers as\nboth data generating distributions and inference networks -- contrasting with\nmost past work on sentence embeddings. In experiments, our approach\nsubstantially outperforms the state-of-the-art on a standard suite of\nunsupervised semantic similarity evaluations. Further, we demonstrate that our\napproach yields the largest gains on more difficult subsets of these\nevaluations where simple word overlap is not a good indicator of similarity."", 'Much information available to applied researchers is contained within written\nlanguage or spoken text. Deep language models such as BERT have achieved\nunprecedented success in many applications of computational linguistics.\nHowever, much less is known about how these models can be used to analyze\nexisting text. We propose a novel method that combines transformer models with\nnetwork analysis to form a self-referential representation of language use\nwithin a corpus of interest. Our approach produces linguistic relations\nstrongly consistent with the underlying model as well as mathematically\nwell-defined operations on them, while reducing the amount of discretionary\nchoices of representation and distance measures. It represents, to the best of\nour knowledge, the first unsupervised method to extract semantic networks\ndirectly from deep language models. We illustrate our approach in a semantic\nanalysis of the term ""founder"". Using the entire corpus of Harvard Business\nReview from 1980 to 2020, we find that ties in our network track the semantics\nof discourse over time, and across contexts, identifying and relating clusters\nof semantic and syntactic relations. Finally, we discuss how this method can\nalso complement and inform analyses of the behavior of deep learning models.', 'Semantic representations have long been argued as potentially useful for\nenforcing meaning preservation and improving generalization performance of\nmachine translation methods. In this work, we are the first to incorporate\ninformation about predicate-argument structure of source sentences (namely,\nsemantic-role representations) into neural machine translation. We use Graph\nConvolutional Networks (GCNs) to inject a semantic bias into sentence encoders\nand achieve improvements in BLEU scores over the linguistic-agnostic and\nsyntax-aware versions on the English--German language pair.', ""Humans can learn structural properties about a word from minimal experience,\nand deploy their learned syntactic representations uniformly in different\ngrammatical contexts. We assess the ability of modern neural language models to\nreproduce this behavior in English and evaluate the effect of structural\nsupervision on learning outcomes. First, we assess few-shot learning\ncapabilities by developing controlled experiments that probe models' syntactic\nnominal number and verbal argument structure generalizations for tokens seen as\nfew as two times during training. Second, we assess invariance properties of\nlearned representation: the ability of a model to transfer syntactic\ngeneralizations from a base context (e.g., a simple declarative active-voice\nsentence) to a transformed context (e.g., an interrogative sentence). We test\nfour models trained on the same dataset: an n-gram baseline, an LSTM, and two\nLSTM-variants trained with explicit structural supervision (Dyer et al.,2016;\nCharniak et al., 2016). We find that in most cases, the neural models are able\nto induce the proper syntactic generalizations after minimal exposure, often\nfrom just two examples during training, and that the two structurally\nsupervised models generalize more accurately than the LSTM model. All neural\nmodels are able to leverage information learned in base contexts to drive\nexpectations in transformed contexts, indicating that they have learned some\ninvariance properties of syntax."", 'We extend the SKIP-GRAM model of Mikolov et al. (2013a) by taking visual\ninformation into account. Like SKIP-GRAM, our multimodal models (MMSKIP-GRAM)\nbuild vector-based word representations by learning to predict linguistic\ncontexts in text corpora. However, for a restricted set of words, the models\nare also exposed to visual representations of the objects they denote\n(extracted from natural images), and must predict linguistic and visual\nfeatures jointly. The MMSKIP-GRAM models achieve good performance on a variety\nof semantic benchmarks. Moreover, since they propagate visual information to\nall words, we use them to improve image labeling and retrieval in the zero-shot\nsetup, where the test concepts are never seen during model training. Finally,\nthe MMSKIP-GRAM models discover intriguing visual properties of abstract words,\npaving the way to realistic implementations of embodied theories of meaning.', 'This paper introduces a sentence to vector encoding framework suitable for\nadvanced natural language processing. Our latent representation is shown to\nencode sentences with common semantic information with similar vector\nrepresentations. The vector representation is extracted from an encoder-decoder\nmodel which is trained on sentence paraphrase pairs. We demonstrate the\napplication of the sentence representations for two different tasks -- sentence\nparaphrasing and paragraph summarization, making it attractive for commonly\nused recurrent frameworks that process text. Experimental results help gain\ninsight how vector representations are suitable for advanced language\nembedding.', 'The goal of semantic role labelling (SRL) is to recognise the\npredicate-argument structure of a sentence. Recent models have shown that\nsyntactic information can enhance the SRL performance, but other\nsyntax-agnostic approaches achieved reasonable performance. The best way to\nencode syntactic information for the SRL task is still an open question. In\nthis paper, we propose the Syntax-aware Graph-to-Graph Transformer (SynG2G-Tr)\narchitecture, which encodes the syntactic structure with a novel way to input\ngraph relations as embeddings directly into the self-attention mechanism of\nTransformer. This approach adds a soft bias towards attention patterns that\nfollow the syntactic structure but also allows the model to use this\ninformation to learn alternative patterns. We evaluate our model on both\ndependency-based and span-based SRL datasets, and outperform all previous\nsyntax-aware and syntax-agnostic models in both in-domain and out-of-domain\nsettings, on the CoNLL 2005 and CoNLL 2009 datasets. Our architecture is\ngeneral and can be applied to encode any graph information for a desired\ndownstream task.', 'Inspired by recent advances in multimodal learning and machine translation,\nwe introduce an encoder-decoder pipeline that learns (a): a multimodal joint\nembedding space with images and text and (b): a novel language model for\ndecoding distributed representations from our space. Our pipeline effectively\nunifies joint image-text embedding models with multimodal neural language\nmodels. We introduce the structure-content neural language model that\ndisentangles the structure of a sentence to its content, conditioned on\nrepresentations produced by the encoder. The encoder allows one to rank images\nand sentences while the decoder can generate novel descriptions from scratch.\nUsing LSTM to encode sentences, we match the state-of-the-art performance on\nFlickr8K and Flickr30K without using object detections. We also set new best\nresults when using the 19-layer Oxford convolutional network. Furthermore we\nshow that with linear encoders, the learned embedding space captures multimodal\nregularities in terms of vector space arithmetic e.g. *image of a blue car* -\n""blue"" + ""red"" is near images of red cars. Sample captions generated for 800\nimages are made available for comparison.', ""A conversational agent (chatbot) is a piece of software that is able to\ncommunicate with humans using natural language. Modeling conversation is an\nimportant task in natural language processing and artificial intelligence.\nWhile chatbots can be used for various tasks, in general they have to\nunderstand users' utterances and provide responses that are relevant to the\nproblem at hand.\n  In my work, I conduct an in-depth survey of recent literature, examining over\n70 publications related to chatbots published in the last 3 years. Then, I\nproceed to make the argument that the very nature of the general conversation\ndomain demands approaches that are different from current state-of-of-the-art\narchitectures. Based on several examples from the literature I show why current\nchatbot models fail to take into account enough priors when generating\nresponses and how this affects the quality of the conversation. In the case of\nchatbots, these priors can be outside sources of information that the\nconversation is conditioned on like the persona or mood of the conversers. In\naddition to presenting the reasons behind this problem, I propose several ideas\non how it could be remedied.\n  The next section focuses on adapting the very recent Transformer model to the\nchatbot domain, which is currently state-of-the-art in neural machine\ntranslation. I first present experiments with the vanilla model, using\nconversations extracted from the Cornell Movie-Dialog Corpus. Secondly, I\naugment the model with some of my ideas regarding the issues of encoder-decoder\narchitectures. More specifically, I feed additional features into the model\nlike mood or persona together with the raw conversation data. Finally, I\nconduct a detailed analysis of how the vanilla model performs on conversational\ndata by comparing it to previous chatbot models and how the additional features\naffect the quality of the generated responses."", ""In the last half-decade, the field of natural language processing (NLP) has\nundergone two major transitions: the switch to neural networks as the primary\nmodeling paradigm and the homogenization of the training regime (pre-train,\nthen fine-tune). Amidst this process, language models have emerged as NLP's\nworkhorse, displaying increasingly fluent generation capabilities and proving\nto be an indispensable means of knowledge transfer downstream. Due to the\notherwise opaque, black-box nature of such models, researchers have employed\naspects of linguistic theory in order to characterize their behavior. Questions\ncentral to syntax -- the study of the hierarchical structure of language --\nhave factored heavily into such work, shedding invaluable insights about\nmodels' inherent biases and their ability to make human-like generalizations.\nIn this paper, we attempt to take stock of this growing body of literature. In\ndoing so, we observe a lack of clarity across numerous dimensions, which\ninfluences the hypotheses that researchers form, as well as the conclusions\nthey draw from their findings. To remedy this, we urge researchers make careful\nconsiderations when investigating coding properties, selecting representations,\nand evaluating via downstream tasks. Furthermore, we outline the implications\nof the different types of research questions exhibited in studies on syntax, as\nwell as the inherent pitfalls of aggregate metrics. Ultimately, we hope that\nour discussion adds nuance to the prospect of studying language models and\npaves the way for a less monolithic perspective on syntax in this context."", 'Incorporating syntax into neural approaches in NLP has a multitude of\npractical and scientific benefits. For instance, a language model that is\nsyntax-aware is likely to be able to produce better samples; even a\ndiscriminative model like BERT with a syntax module could be used for core NLP\ntasks like unsupervised syntactic parsing. Rapid progress in recent years was\narguably spurred on by the empirical success of the Parsing-Reading-Predict\narchitecture of (Shen et al., 2018a), later simplified by the Order Neuron LSTM\nof (Shen et al., 2019). Most notably, this is the first time neural approaches\nwere able to successfully perform unsupervised syntactic parsing (evaluated by\nvarious metrics like F-1 score).\n  However, even heuristic (much less fully mathematical) understanding of why\nand when these architectures work is lagging severely behind. In this work, we\nanswer representational questions raised by the architectures in (Shen et al.,\n2018a, 2019), as well as some transition-based syntax-aware language models\n(Dyer et al., 2016): what kind of syntactic structure can current neural\napproaches to syntax represent? Concretely, we ground this question in the\nsandbox of probabilistic context-free-grammars (PCFGs), and identify a key\naspect of the representational power of these approaches: the amount and\ndirectionality of context that the predictor has access to when forced to make\nparsing decision. We show that with limited context (either bounded, or\nunidirectional), there are PCFGs, for which these approaches cannot represent\nthe max-likelihood parse; conversely, if the context is unlimited, they can\nrepresent the max-likelihood parse of any PCFG.', 'The field of natural language understanding has experienced exponential\nprogress in the last few years, with impressive results in several tasks. This\nsuccess has motivated researchers to study the underlying knowledge encoded by\nthese models. Despite this, attempts to understand their semantic capabilities\nhave not been successful, often leading to non-conclusive, or contradictory\nconclusions among different works. Via a probing classifier, we extract the\nunderlying knowledge graph of nine of the most influential language models of\nthe last years, including word embeddings, text generators, and context\nencoders. This probe is based on concept relatedness, grounded on WordNet. Our\nresults reveal that all the models encode this knowledge, but suffer from\nseveral inaccuracies. Furthermore, we show that the different architectures and\ntraining strategies lead to different model biases. We conduct a systematic\nevaluation to discover specific factors that explain why some concepts are\nchallenging. We hope our insights will motivate the development of models that\ncapture concepts more precisely.', 'At present, different deep learning models are presenting high accuracy on\npopular inference datasets such as SNLI, MNLI, and SciTail. However, there are\ndifferent indicators that those datasets can be exploited by using some simple\nlinguistic patterns. This fact poses difficulties to our understanding of the\nactual capacity of machine learning models to solve the complex task of textual\ninference. We propose a new set of syntactic tasks focused on contradiction\ndetection that require specific capacities over linguistic logical forms such\nas: Boolean coordination, quantifiers, definite description, and counting\noperators. We evaluate two kinds of deep learning models that implicitly\nexploit language structure: recurrent models and the Transformer network BERT.\nWe show that although BERT is clearly more efficient to generalize over most\nlogical forms, there is space for improvement when dealing with counting\noperators. Since the syntactic tasks can be implemented in different languages,\nwe show a successful case of cross-lingual transfer learning between English\nand Portuguese.', 'We propose transfer learning as a method for analyzing the encoding of\ngrammatical structure in neural language models. We train LSTMs on\nnon-linguistic data and evaluate their performance on natural language to\nassess which kinds of data induce generalizable structural features that LSTMs\ncan use for natural language. We find that training on non-linguistic data with\nlatent structure (MIDI music or Java code) improves test performance on natural\nlanguage, despite no overlap in surface form or vocabulary. To pinpoint the\nkinds of abstract structure that models may be encoding to lead to this\nimprovement, we run similar experiments with two artificial parentheses\nlanguages: one which has a hierarchical recursive structure, and a control\nwhich has paired tokens but no recursion. Surprisingly, training a model on\neither of these artificial languages leads to the same substantial gains when\ntesting on natural language. Further experiments on transfer between natural\nlanguages controlling for vocabulary overlap show that zero-shot performance on\na test language is highly correlated with typological syntactic similarity to\nthe training language, suggesting that representations induced by pre-training\ncorrespond to the cross-linguistic syntactic properties. Our results provide\ninsights into the ways that neural models represent abstract syntactic\nstructure, and also about the kind of structural inductive biases which allow\nfor natural language acquisition.', 'In this work we propose a simple and efficient framework for learning\nsentence representations from unlabelled data. Drawing inspiration from the\ndistributional hypothesis and recent work on learning sentence representations,\nwe reformulate the problem of predicting the context in which a sentence\nappears as a classification problem. Given a sentence and its context, a\nclassifier distinguishes context sentences from other contrastive sentences\nbased on their vector representations. This allows us to efficiently learn\ndifferent types of encoding functions, and we show that the model learns\nhigh-quality sentence representations. We demonstrate that our sentence\nrepresentations outperform state-of-the-art unsupervised and supervised\nrepresentation learning methods on several downstream NLP tasks that involve\nunderstanding sentence semantics while achieving an order of magnitude speedup\nin training time.', ""Long-range semantic coherence remains a challenge in automatic language\ngeneration and understanding. We demonstrate that large language models have\ninsufficiently learned the effect of distant words on next-token prediction. We\npresent coherence boosting, an inference procedure that increases a LM's focus\non a long context. We show the benefits of coherence boosting with pretrained\nmodels by distributional analyses of generated ordinary text and dialog\nresponses. It is also found that coherence boosting with state-of-the-art\nmodels for various zero-shot NLP tasks yields performance gains with no\nadditional training."", ""In this paper we investigate the linguistic knowledge learned by a Neural\nLanguage Model (NLM) before and after a fine-tuning process and how this\nknowledge affects its predictions during several classification problems. We\nuse a wide set of probing tasks, each of which corresponds to a distinct\nsentence-level feature extracted from different levels of linguistic\nannotation. We show that BERT is able to encode a wide range of linguistic\ncharacteristics, but it tends to lose this information when trained on specific\ndownstream tasks. We also find that BERT's capacity to encode different kind of\nlinguistic properties has a positive influence on its predictions: the more it\nstores readable linguistic information of a sentence, the higher will be its\ncapacity of predicting the expected label assigned to that sentence."", 'Measuring text complexity is an essential task in several fields and\napplications (such as NLP, semantic web, smart education, etc.). The semantic\nlayer of text is more tacit than its syntactic structure and, as a result,\ncalculation of semantic complexity is more difficult than syntactic complexity.\nWhile there are famous and powerful academic and commercial syntactic\ncomplexity measures, the problem of measuring semantic complexity is still a\nchallenging one. In this paper, we introduce the DAST model, which stands for\nDeciding About Semantic Complexity of a Text. DAST proposes an intuitionistic\napproach to semantics that lets us have a well-defined model for the semantics\nof a text and its complexity: semantic is considered as a lattice of intuitions\nand, as a result, semantic complexity is defined as the result of a calculation\non this lattice. A set theoretic formal definition of semantic complexity, as a\n6-tuple formal system, is provided. By using this formal system, a method for\nmeasuring semantic complexity is presented. The evaluation of the proposed\napproach is done by a set of three human-judgment experiments. The results show\nthat DAST model is capable of deciding about semantic complexity of text.\nFurthermore, the analysis of the results leads us to introduce a Markovian\nmodel for the process of common-sense, multiple-steps and semantic-complexity\nreasoning in people. The results of Experiments demonstrate that our method\noutperforms the random baseline with improvement in better precision and\ncompetes with other methods by less error percentage.', 'Statistical methods have been widely employed in recent years to grasp many\nlanguage properties. The application of such techniques have allowed an\nimprovement of several linguistic applications, which encompasses machine\ntranslation, automatic summarization and document classification. In the\nlatter, many approaches have emphasized the semantical content of texts, as it\nis the case of bag-of-word language models. This approach has certainly yielded\nreasonable performance. However, some potential features such as the structural\norganization of texts have been used only on a few studies. In this context, we\nprobe how features derived from textual structure analysis can be effectively\nemployed in a classification task. More specifically, we performed a supervised\nclassification aiming at discriminating informative from imaginative documents.\nUsing a networked model that describes the local topological/dynamical\nproperties of function words, we achieved an accuracy rate of up to 95%, which\nis much higher than similar networked approaches. A systematic analysis of\nfeature relevance revealed that symmetry and accessibility measurements are\namong the most prominent network measurements. Our results suggest that these\nmeasurements could be used in related language applications, as they play a\ncomplementary role in characterizing texts.', 'A desideratum of high-quality translation systems is that they preserve\nmeaning, in the sense that two sentences with different meanings should not\ntranslate to one and the same sentence in another language. However,\nstate-of-the-art systems often fail in this regard, particularly in cases where\nthe source and target languages partition the ""meaning space"" in different\nways. For instance, ""I cut my finger."" and ""I cut my finger off."" describe\ndifferent states of the world but are translated to French (by both Fairseq and\nGoogle Translate) as ""Je me suis coupe le doigt."", which is ambiguous as to\nwhether the finger is detached. More generally, translation systems are\ntypically many-to-one (non-injective) functions from source to target language,\nwhich in many cases results in important distinctions in meaning being lost in\ntranslation. Building on Bayesian models of informative utterance production,\nwe present a method to define a less ambiguous translation system in terms of\nan underlying pre-trained neural sequence-to-sequence model. This method\nincreases injectivity, resulting in greater preservation of meaning as measured\nby improvement in cycle-consistency, without impeding translation quality\n(measured by BLEU score).', 'Pre-trained language models have achieved huge success on a wide range of NLP\ntasks. However, contextual representations from pre-trained models contain\nentangled semantic and syntactic information, and therefore cannot be directly\nused to derive useful semantic sentence embeddings for some tasks. Paraphrase\npairs offer an effective way of learning the distinction between semantics and\nsyntax, as they naturally share semantics and often vary in syntax. In this\nwork, we present ParaBART, a semantic sentence embedding model that learns to\ndisentangle semantics and syntax in sentence embeddings obtained by pre-trained\nlanguage models. ParaBART is trained to perform syntax-guided paraphrasing,\nbased on a source sentence that shares semantics with the target paraphrase,\nand a parse tree that specifies the target syntax. In this way, ParaBART learns\ndisentangled semantic and syntactic representations from their respective\ninputs with separate encoders. Experiments in English show that ParaBART\noutperforms state-of-the-art sentence embedding models on unsupervised semantic\nsimilarity tasks. Additionally, we show that our approach can effectively\nremove syntactic information from semantic sentence embeddings, leading to\nbetter robustness against syntactic variation on downstream semantic tasks.', 'We propose two methods of learning vector representations of words and\nphrases that each combine sentence context with structural features extracted\nfrom dependency trees. Using several variations of neural network classifier,\nwe show that these combined methods lead to improved performance when used as\ninput features for supervised term-matching.', 'Neural network-based generative language models like ELMo and BERT can work\neffectively as general purpose sentence encoders in text classification without\nfurther fine-tuning. Is it possible to adapt them in a similar way for use as\ngeneral-purpose decoders? For this to be possible, it would need to be the case\nthat for any target sentence of interest, there is some continuous\nrepresentation that can be passed to the language model to cause it to\nreproduce that sentence. We set aside the difficult problem of designing an\nencoder that can produce such representations and, instead, ask directly\nwhether such representations exist at all. To do this, we introduce a pair of\neffective, complementary methods for feeding representations into pretrained\nunconditional language models and a corresponding set of methods to map\nsentences into and out of this representation space, the reparametrized\nsentence space. We then investigate the conditions under which a language model\ncan be made to generate a sentence through the identification of a point in\nsuch a space and find that it is possible to recover arbitrary sentences nearly\nperfectly with language models and representations of moderate size without\nmodifying any model parameters.', 'One central mystery of neural NLP is what neural models ""know"" about their\nsubject matter. When a neural machine translation system learns to translate\nfrom one language to another, does it learn the syntax or semantics of the\nlanguages? Can this knowledge be extracted from the system to fill holes in\nhuman scientific knowledge? Existing typological databases contain relatively\nfull feature specifications for only a few hundred languages. Exploiting the\nexistence of parallel texts in more than a thousand languages, we build a\nmassive many-to-one neural machine translation (NMT) system from 1017 languages\ninto English, and use this to predict information missing from typological\ndatabases. Experiments show that the proposed method is able to infer not only\nsyntactic, but also phonological and phonetic inventory features, and improves\nover a baseline that has access to information about the languages\' geographic\nand phylogenetic neighbors.', ""Understanding the semantic relationships between terms is a fundamental task\nin natural language processing applications. While structured resources that\ncan express those relationships in a formal way, such as ontologies, are still\nscarce, a large number of linguistic resources gathering dictionary definitions\nis becoming available, but understanding the semantic structure of natural\nlanguage definitions is fundamental to make them useful in semantic\ninterpretation tasks. Based on an analysis of a subset of WordNet's glosses, we\npropose a set of semantic roles that compose the semantic structure of a\ndictionary definition, and show how they are related to the definition's\nsyntactic configuration, identifying patterns that can be used in the\ndevelopment of information extraction frameworks and semantic models."", 'One of the most complex syntactic representations used in computational\nlinguistics and NLP are discontinuous constituent trees, crucial for\nrepresenting all grammatical phenomena of languages such as German. Recent\nadvances in dependency parsing have shown that Pointer Networks excel in\nefficiently parsing syntactic relations between words in a sentence. This kind\nof sequence-to-sequence models achieve outstanding accuracies in building\nnon-projective dependency trees, but its potential has not been proved yet on a\nmore difficult task. We propose a novel neural network architecture that, by\nmeans of Pointer Networks, is able to generate the most accurate discontinuous\nconstituent representations to date, even without the need of Part-of-Speech\ntagging information. To do so, we internally model discontinuous constituent\nstructures as augmented non-projective dependency structures. The proposed\napproach achieves state-of-the-art results on the two widely-used NEGRA and\nTIGER benchmarks, outperforming previous work by a wide margin.', 'Current approaches to learning vector representations of text that are\ncompatible between different languages usually require some amount of parallel\ntext, aligned at word, sentence or at least document level. We hypothesize\nhowever, that different natural languages share enough semantic structure that\nit should be possible, in principle, to learn compatible vector representations\njust by analyzing the monolingual distribution of words.\n  In order to evaluate this hypothesis, we propose a scheme to map word vectors\ntrained on a source language to vectors semantically compatible with word\nvectors trained on a target language using an adversarial autoencoder.\n  We present preliminary qualitative results and discuss possible future\ndevelopments of this technique, such as applications to cross-lingual sentence\nrepresentations.', 'There is growing evidence that the classical notion of adversarial robustness\noriginally introduced for images has been adopted as a de facto standard by a\nlarge part of the NLP research community. We show that this notion is\nproblematic in the context of NLP as it considers a narrow spectrum of\nlinguistic phenomena. In this paper, we argue for semantic robustness, which is\nbetter aligned with the human concept of linguistic fidelity. We characterize\nsemantic robustness in terms of biases that it is expected to induce in a\nmodel. We study semantic robustness of a range of vanilla and robustly trained\narchitectures using a template-based generative test bed. We complement the\nanalysis with empirical evidence that, despite being harder to implement,\nsemantic robustness can improve performance %gives guarantees for on complex\nlinguistic phenomena where models robust in the classical sense fail.', ""We investigated the impact of noisy linguistic features on the performance of\na Japanese speech synthesis system based on neural network that uses WaveNet\nvocoder. We compared an ideal system that uses manually corrected linguistic\nfeatures including phoneme and prosodic information in training and test sets\nagainst a few other systems that use corrupted linguistic features. Both\nsubjective and objective results demonstrate that corrupted linguistic\nfeatures, especially those in the test set, affected the ideal system's\nperformance significantly in a statistical sense due to a mismatched condition\nbetween the training and test sets. Interestingly, while an utterance-level\nTuring test showed that listeners had a difficult time differentiating\nsynthetic speech from natural speech, it further indicated that adding noise to\nthe linguistic features in the training set can partially reduce the effect of\nthe mismatch, regularize the model, and help the system perform better when\nlinguistic features of the test set are noisy."", 'The task of linearization is to find a grammatical order given a set of\nwords. Traditional models use statistical methods. Syntactic linearization\nsystems, which generate a sentence along with its syntactic tree, have shown\nstate-of-the-art performance. Recent work shows that a multi-layer LSTM\nlanguage model outperforms competitive statistical syntactic linearization\nsystems without using syntax. In this paper, we study neural syntactic\nlinearization, building a transition-based syntactic linearizer leveraging a\nfeed-forward neural network, observing significantly better results compared to\nLSTM language models on this task.', 'Labeling explicit discourse relations is one of the most challenging\nsub-tasks of the shallow discourse parsing where the goal is to identify the\ndiscourse connectives and the boundaries of their arguments. The\nstate-of-the-art models achieve slightly above 45% of F-score by using\nhand-crafted features. The current paper investigates the efficacy of the\npre-trained language models in this task. We find that the pre-trained language\nmodels, when finetuned, are powerful enough to replace the linguistic features.\nWe evaluate our model on PDTB 2.0 and report the state-of-the-art results in\nthe extraction of the full relation. This is the first time when a model\noutperforms the knowledge intensive models without employing any linguistic\nfeatures.', 'We introduce the syntactic scaffold, an approach to incorporating syntactic\ninformation into semantic tasks. Syntactic scaffolds avoid expensive syntactic\nprocessing at runtime, only making use of a treebank during training, through a\nmultitask objective. We improve over strong baselines on PropBank semantics,\nframe semantics, and coreference resolution, achieving competitive performance\non all three tasks.', 'Semantic parsing, i.e., the automatic derivation of meaning representation\nsuch as an instantiated predicate-argument structure for a sentence, plays a\ncritical role in deep processing of natural language. Unlike all other top\nsystems of semantic dependency parsing that have to rely on a pipeline\nframework to chain up a series of submodels each specialized for a specific\nsubtask, the one presented in this article integrates everything into one\nmodel, in hopes of achieving desirable integrity and practicality for real\napplications while maintaining a competitive performance. This integrative\napproach tackles semantic parsing as a word pair classification problem using a\nmaximum entropy classifier. We leverage adaptive pruning of argument candidates\nand large-scale feature selection engineering to allow the largest feature\nspace ever in use so far in this field, it achieves a state-of-the-art\nperformance on the evaluation data set for CoNLL-2008 shared task, on top of\nall but one top pipeline system, confirming its feasibility and effectiveness.', 'We introduce semantic form mid-tuning, an approach for transferring semantic\nknowledge from semantic meaning representations into transformer-based language\nencoders. In mid-tuning, we learn to align the text of general sentences -- not\ntied to any particular inference task -- and structured semantic\nrepresentations of those sentences. Our approach does not require gold\nannotated semantic representations. Instead, it makes use of automatically\ngenerated semantic representations, such as from off-the-shelf PropBank and\nFrameNet semantic parsers. We show that this alignment can be learned\nimplicitly via classification or directly via triplet loss. Our method yields\nlanguage encoders that demonstrate improved predictive performance across\ninference, reading comprehension, textual similarity, and other semantic tasks\ndrawn from the GLUE, SuperGLUE, and SentEval benchmarks. We evaluate our\napproach on three popular baseline models, where our experimental results and\nanalysis concludes that current pre-trained language models can further benefit\nfrom structured semantic frames with the proposed mid-tuning method, as they\ninject additional task-agnostic knowledge to the encoder, improving the\ngenerated embeddings as well as the linguistic properties of the given model,\nas evident from improvements on a popular sentence embedding toolkit and a\nvariety of probing tasks.', 'As a fundamental NLP task, semantic role labeling (SRL) aims to discover the\nsemantic roles for each predicate within one sentence. This paper investigates\nhow to incorporate syntactic knowledge into the SRL task effectively. We\npresent different approaches of encoding the syntactic information derived from\ndependency trees of different quality and representations; we propose a\nsyntax-enhanced self-attention model and compare it with other two strong\nbaseline methods; and we conduct experiments with newly published deep\ncontextualized word representations as well. The experiment results demonstrate\nthat with proper incorporation of the high quality syntactic information, our\nmodel achieves a new state-of-the-art performance for the Chinese SRL task on\nthe CoNLL-2009 dataset.', 'Answer sentence selection is the task of identifying sentences that contain\nthe answer to a given question. This is an important problem in its own right\nas well as in the larger context of open domain question answering. We propose\na novel approach to solving this task via means of distributed representations,\nand learn to match questions with answers by considering their semantic\nencoding. This contrasts prior work on this task, which typically relies on\nclassifiers with large numbers of hand-crafted syntactic and semantic features\nand various external resources. Our approach does not require any feature\nengineering nor does it involve specialist linguistic data, making this model\neasily applicable to a wide range of domains and languages. Experimental\nresults on a standard benchmark dataset from TREC demonstrate that---despite\nits simplicity---our model matches state of the art performance on the answer\nsentence selection task.', 'In many multilingual text classification problems, the documents in different\nlanguages often share the same set of categories. To reduce the labeling cost\nof training a classification model for each individual language, it is\nimportant to transfer the label knowledge gained from one language to another\nlanguage by conducting cross language classification. In this paper we develop\na novel subspace co-regularized multi-view learning method for cross language\ntext classification. This method is built on parallel corpora produced by\nmachine translation. It jointly minimizes the training error of each classifier\nin each language while penalizing the distance between the subspace\nrepresentations of parallel documents. Our empirical study on a large set of\ncross language text classification tasks shows the proposed method consistently\noutperforms a number of inductive methods, domain adaptation methods, and\nmulti-view learning methods.', ""This paper investigates the ability of artificial neural networks to judge\nthe grammatical acceptability of a sentence, with the goal of testing their\nlinguistic competence. We introduce the Corpus of Linguistic Acceptability\n(CoLA), a set of 10,657 English sentences labeled as grammatical or\nungrammatical from published linguistics literature. As baselines, we train\nseveral recurrent neural network models on acceptability classification, and\nfind that our models outperform unsupervised models by Lau et al (2016) on\nCoLA. Error-analysis on specific grammatical phenomena reveals that both Lau et\nal.'s models and ours learn systematic generalizations like subject-verb-object\norder. However, all models we test perform far below human level on a wide\nrange of grammatical constructions."", 'Semantic role labeling (SRL) is to recognize the predicate-argument structure\nof a sentence, including subtasks of predicate disambiguation and argument\nlabeling. Previous studies usually formulate the entire SRL problem into two or\nmore subtasks. For the first time, this paper introduces an end-to-end neural\nmodel which unifiedly tackles the predicate disambiguation and the argument\nlabeling in one shot. Using a biaffine scorer, our model directly predicts all\nsemantic role labels for all given word pairs in the sentence without relying\non any syntactic parse information. Specifically, we augment the BiLSTM encoder\nwith a non-linear transformation to further distinguish the predicate and the\nargument in a given sentence, and model the semantic role labeling process as a\nword pair classification task by employing the biaffine attentional mechanism.\nThough the proposed model is syntax-agnostic with local decoder, it outperforms\nthe state-of-the-art syntax-aware SRL systems on the CoNLL-2008, 2009\nbenchmarks for both English and Chinese. To our best knowledge, we report the\nfirst syntax-agnostic SRL model that surpasses all known syntax-aware models.', ""Neural language models learn, to varying degrees of accuracy, the grammatical\nproperties of natural languages. In this work, we investigate whether there are\nsystematic sources of variation in the language models' accuracy. Focusing on\nsubject-verb agreement and reflexive anaphora, we find that certain nouns are\nsystematically understood better than others, an effect which is robust across\ngrammatical tasks and different language models. Surprisingly, we find that\nacross four orders of magnitude, corpus frequency is unrelated to a noun's\nperformance on grammatical tasks. Finally, we find that a novel noun's\ngrammatical properties can be few-shot learned from various types of training\ndata. The results present a paradox: there should be less variation in\ngrammatical performance than is actually observed."", 'Document categorization, which aims to assign a topic label to each document,\nplays a fundamental role in a wide variety of applications. Despite the success\nof existing studies in conventional supervised document classification, they\nare less concerned with two real problems: (1) the presence of metadata: in\nmany domains, text is accompanied by various additional information such as\nauthors and tags. Such metadata serve as compelling topic indicators and should\nbe leveraged into the categorization framework; (2) label scarcity: labeled\ntraining samples are expensive to obtain in some cases, where categorization\nneeds to be performed using only a small set of annotated data. In recognition\nof these two challenges, we propose MetaCat, a minimally supervised framework\nto categorize text with metadata. Specifically, we develop a generative process\ndescribing the relationships between words, documents, labels, and metadata.\nGuided by the generative model, we embed text and metadata into the same\nsemantic space to encode heterogeneous signals. Then, based on the same\ngenerative process, we synthesize training samples to address the bottleneck of\nlabel scarcity. We conduct a thorough evaluation on a wide range of datasets.\nExperimental results prove the effectiveness of MetaCat over many competitive\nbaselines.', 'Organizing complex peer production projects and advancing scientific\nknowledge of open collaboration each depend on the ability to measure quality.\nArticle quality ratings on English language Wikipedia have been widely used by\nboth Wikipedia community members and academic researchers for purposes like\ntracking knowledge gaps and studying how political polarization shapes\ncollaboration. Even so, measuring quality presents many methodological\nchallenges. The most widely used systems use labels on discrete ordinal scales\nwhen assessing quality, but such labels can be inconvenient for statistics and\nmachine learning. Prior work handles this by assuming that different levels of\nquality are ""evenly spaced"" from one another. This assumption runs counter to\nintuitions about the relative degrees of effort needed to raise Wikipedia\nencyclopedia articles to different quality levels. Furthermore, models from\nprior work are fit to datasets that oversample high-quality articles. This\nlimits their accuracy for representative samples of articles or revisions. I\ndescribe a technique extending the Wikimedia Foundations\' ORES article quality\nmodel to address these limitations. My method uses weighted ordinal regression\nmodels to construct one-dimensional continuous measures of quality. While\nscores from my technique and from prior approaches are correlated, my approach\nimproves accuracy for research datasets and provides evidence that the ""evenly\nspaced"" assumption is unfounded in practice on English Wikipedia. I conclude\nwith recommendations for using quality scores in future research and include\nthe full code, data, and models.', ""My doctoral research focuses on understanding semantic knowledge in neural\nnetwork models trained solely to predict natural language (referred to as\nlanguage models, or LMs), by drawing on insights from the study of concepts and\ncategories grounded in cognitive science. I propose a framework inspired by\n'inductive reasoning,' a phenomenon that sheds light on how humans utilize\nbackground knowledge to make inductive leaps and generalize from new pieces of\ninformation about concepts and their properties. Drawing from experiments that\nstudy inductive reasoning, I propose to analyze semantic inductive\ngeneralization in LMs using phenomena observed in human-induction literature,\ninvestigate inductive behavior on tasks such as implicit reasoning and emergent\nfeature recognition, and analyze and relate induction dynamics to the learned\nconceptual representation space."", 'Media coverage has a substantial effect on the public perception of events.\nNevertheless, media outlets are often biased. One way to bias news articles is\nby altering the word choice. The automatic identification of bias by word\nchoice is challenging, primarily due to the lack of gold-standard data sets and\nhigh context dependencies. In this research project, I aim to devise data sets\nand methods to identify media bias. To achieve this, I plan to research methods\nusing natural language processing and deep learning while employing models and\nusing analysis concepts from psychology and linguistics. The first results\nindicate the effectiveness of an interdisciplinary research approach. My vision\nis to devise a system that helps news readers become aware of media coverage\ndifferences caused by bias. So far, my best performing BERT-based model is\npre-trained on a larger corpus consisting of distant labels, indicating that\ndistant supervision has the potential to become a solution for the difficult\ntask of bias detection.', ""Vector representations have become a central element in semantic language\nmodelling, leading to mathematical overlaps with many fields including quantum\ntheory. Compositionality is a core goal for such representations: given\nrepresentations for 'wet' and 'fish', how should the concept 'wet fish' be\nrepresented?\n  This position paper surveys this question from two points of view. The first\nconsiders the question of whether an explicit mathematical representation can\nbe successful using only tools from within linear algebra, or whether other\nmathematical tools are needed. The second considers whether semantic vector\ncomposition should be explicitly described mathematically, or whether it can be\na model-internal side-effect of training a neural network.\n  A third and newer question is whether a compositional model can be\nimplemented on a quantum computer. Given the fundamentally linear nature of\nquantum mechanics, we propose that these questions are related, and that this\nsurvey may help to highlight candidate operations for future quantum\nimplementation."", 'We reduce the task of (span-based) PropBank-style semantic role labeling\n(SRL) to syntactic dependency parsing. Our approach is motivated by our\nempirical analysis that shows three common syntactic patterns account for over\n98% of the SRL annotations for both English and Chinese data. Based on this\nobservation, we present a conversion scheme that packs SRL annotations into\ndependency tree representations through joint labels that permit highly\naccurate recovery back to the original format. This representation allows us to\ntrain statistical dependency parsers to tackle SRL and achieve competitive\nperformance with the current state of the art. Our findings show the promise of\nsyntactic dependency trees in encoding semantic role relations within their\nsyntactic domain of locality, and point to potential further integration of\nsyntactic methods into semantic role labeling in the future.', 'Neural network based models are a very powerful tool for creating word\nembeddings, the objective of these models is to group similar words together.\nThese embeddings have been used as features to improve results in various\napplications such as document classification, named entity recognition, etc.\nNeural language models are able to learn word representations which have been\nused to capture semantic shifts across time and geography. The objective of\nthis paper is to first identify and then visualize how words change meaning in\ndifferent text corpus. We will train a neural language model on texts from a\ndiverse set of disciplines philosophy, religion, fiction etc. Each text will\nalter the embeddings of the words to represent the meaning of the word inside\nthat text. We will present a computational technique to detect words that\nexhibit significant linguistic shift in meaning and usage. We then use enhanced\nscatterplots and storyline visualization to visualize the linguistic shift.', 'Predicting the structure of a discourse is challenging because relations\nbetween discourse segments are often implicit and thus hard to distinguish\ncomputationally. I extend previous work to classify implicit discourse\nrelations by introducing a novel set of features on the level of semantic\nroles. My results demonstrate that such features are helpful, yielding results\ncompetitive with other feature-rich approaches on the PDTB. My main\ncontribution is an analysis of improvements that can be traced back to\nrole-based features, providing insights into why and when role semantics is\nhelpful.', 'To investigate the evolution of syntax, we need to ascertain the evolutionary\nr\\^ole of syntax and, before that, the very nature of syntax. Here, we will\nassume that syntax is computing. And then, since we are computationally Turing\ncomplete, we meet an evolutionary anomaly, the anomaly of sytax: we are\nsyntactically too competent for syntax. Assuming that problem solving is\ncomputing, and realizing that the evolutionary advantage of Turing completeness\nis full problem solving and not syntactic proficiency, we explain the anomaly\nof syntax by postulating that syntax and problem solving co-evolved in humans\ntowards Turing completeness. Examining the requirements that full problem\nsolving impose on language, we find firstly that semantics is not sufficient\nand that syntax is necessary to represent problems. Our final conclusion is\nthat full problem solving requires a functional semantics on an infinite\ntree-structured syntax. Besides these results, the introduction of Turing\ncompleteness and problem solving to explain the evolution of syntax should help\nus to fit the evolution of language within the evolution of cognition, giving\nus some new clues to understand the elusive relation between language and\nthinking.', 'The growing popularity of Virtual Assistants poses new challenges for Entity\nResolution, the task of linking mentions in text to their referent entities in\na knowledge base. Specifically, in the shopping domain, customers tend to use\nimplicit utterances (e.g., ""organic milk"") rather than explicit names, leading\nto a large number of candidate products. Meanwhile, for the same query,\ndifferent customers may expect different results. For example, with ""add milk\nto my cart"", a customer may refer to a certain organic product, while some\ncustomers may want to re-order products they regularly purchase. To address\nthese issues, we propose a new framework that leverages personalized features\nto improve the accuracy of product ranking. We first build a cross-source\nheterogeneous knowledge graph from customer purchase history and product\nknowledge graph to jointly learn customer and product embeddings. After that,\nwe incorporate product, customer, and history representations into a neural\nreranking model to predict which candidate is most likely to be purchased for a\nspecific customer. Experiments show that our model substantially improves the\naccuracy of the top ranked candidates by 24.6% compared to the state-of-the-art\nproduct search model.', ""In Computer-Supported learning, monitoring and engaging a group of learners\nis a complex task for teachers, especially when learners are working\ncollaboratively: Are my students motivated? What kind of progress are they\nmaking? Should I intervene? Is my communication and the didactic design adapted\nto my students? Our hypothesis is that the analysis of natural language\ninteractions between students, and between students and teachers, provide very\nvaluable information and could be used to produce qualitative indicators to\nhelp teachers' decisions. We develop an automatic approach in three steps (1)\nto explore the discursive functions of messages in a CSCL platform, (2) to\nclassify the messages automatically and (3) to evaluate correlations between\ndiscursive attitudes and other variables linked to the learning activity.\nResults tend to show that some types of discourse are correlated with a notion\nof Progress on the learning activities and the importance of emotive\nparticipation from the Teacher."", 'Semantic measures are widely used today to estimate the strength of the\nsemantic relationship between elements of various types: units of language\n(e.g., words, sentences, documents), concepts or even instances semantically\ncharacterized (e.g., diseases, genes, geographical locations). Semantic\nmeasures play an important role to compare such elements according to semantic\nproxies: texts and knowledge representations, which support their meaning or\ndescribe their nature. Semantic measures are therefore essential for designing\nintelligent agents which will for example take advantage of semantic analysis\nto mimic human ability to compare abstract or concrete objects. This paper\nproposes a comprehensive survey of the broad notion of semantic measure for the\ncomparison of units of language, concepts or instances based on semantic proxy\nanalyses. Semantic measures generalize the well-known notions of semantic\nsimilarity, semantic relatedness and semantic distance, which have been\nextensively studied by various communities over the last decades (e.g.,\nCognitive Sciences, Linguistics, and Artificial Intelligence to mention a few).', 'Recursive noun phrases (NPs) have interesting semantic properties. For\nexample, ""my favorite new movie"" is not necessarily my favorite movie, whereas\n""my new favorite movie"" is. This is common sense to humans, yet it is unknown\nwhether language models have such knowledge. We introduce the Recursive Noun\nPhrase Challenge (RNPC), a dataset of three textual inference tasks involving\ntextual entailment and event plausibility comparison, precisely targeting the\nunderstanding of recursive NPs. When evaluated on RNPC, state-of-the-art\nTransformer models only perform around chance. Still, we show that such\nknowledge is learnable with appropriate data. We further probe the models for\nrelevant linguistic features that can be learned from our tasks, including\nmodifier semantic category and modifier scope. Finally, models trained on RNPC\nachieve strong zero-shot performance on an extrinsic Harm Detection evaluation\ntask, showing the usefulness of the understanding of recursive NPs in\ndownstream applications.', 'Emotion stimulus extraction is a fine-grained subtask of emotion analysis\nthat focuses on identifying the description of the cause behind an emotion\nexpression from a text passage (e.g., in the sentence ""I am happy that I passed\nmy exam"" the phrase ""passed my exam"" corresponds to the stimulus.). Previous\nwork mainly focused on Mandarin and English, with no resources or models for\nGerman. We fill this research gap by developing a corpus of 2006 German news\nheadlines annotated with emotions and 811 instances with annotations of\nstimulus phrases. Given that such corpus creation efforts are time-consuming\nand expensive, we additionally work on an approach for projecting the existing\nEnglish GoodNewsEveryone (GNE) corpus to a machine-translated German version.\nWe compare the performance of a conditional random field (CRF) model (trained\nmonolingually on German and cross-lingually via projection) with a multilingual\nXLM-RoBERTa (XLM-R) model. Our results show that training with the German\ncorpus achieves higher F1 scores than projection. Experiments with XLM-R\noutperform their respective CRF counterparts.', 'The question of how to probe contextual word representations for linguistic\nstructure in a way that is both principled and useful has seen significant\nattention recently in the NLP literature. In our contribution to this\ndiscussion, we argue for a probe metric that reflects the fundamental trade-off\nbetween probe complexity and performance: the Pareto hypervolume. To measure\ncomplexity, we present a number of parametric and non-parametric metrics. Our\nexperiments using Pareto hypervolume as an evaluation metric show that probes\noften do not conform to our expectations---e.g., why should the non-contextual\nfastText representations encode more morpho-syntactic information than the\ncontextual BERT representations? These results suggest that common, simplistic\nprobing tasks, such as part-of-speech labeling and dependency arc labeling, are\ninadequate to evaluate the linguistic structure encoded in contextual word\nrepresentations. This leads us to propose full dependency parsing as a probing\ntask. In support of our suggestion that harder probing tasks are necessary, our\nexperiments with dependency parsing reveal a wide gap in syntactic knowledge\nbetween contextual and non-contextual representations.', ""We present a dataset for evaluating the grammaticality of the predictions of\na language model. We automatically construct a large number of minimally\ndifferent pairs of English sentences, each consisting of a grammatical and an\nungrammatical sentence. The sentence pairs represent different variations of\nstructure-sensitive phenomena: subject-verb agreement, reflexive anaphora and\nnegative polarity items. We expect a language model to assign a higher\nprobability to the grammatical sentence than the ungrammatical one. In an\nexperiment using this data set, an LSTM language model performed poorly on many\nof the constructions. Multi-task training with a syntactic objective (CCG\nsupertagging) improved the LSTM's accuracy, but a large gap remained between\nits performance and the accuracy of human participants recruited online. This\nsuggests that there is considerable room for improvement over LSTMs in\ncapturing syntax in a language model."", 'My notes on Deep Learning for NLP.', 'Basic body-part names (BBPNs) were defined as body-part names in Swadesh\nbasic 200 words. Non-Mayan cognates of Mayan (MY) BBPNs were extensively\nsearched for, by comparing with non-MY vocabulary, including ca.1300 basic\nwords of 82 AN languages listed by Tryon (1985), etc. Thus found cognates (CGs)\nin non-MY are listed in Table 1, as classified by language groups to which most\nsimilar cognates (MSCs) of MY BBPNs belong. CGs of MY are classified to 23\nmutually unrelated CG-items, of which 17.5 CG-items have their MSCs in\nAustronesian (AN), giving its closest similarity score (CSS), CSS(AN) = 17.5,\nwhich consists of 10.33 MSCs in Formosan, 1.83 MSCs in Western\nMalayo-Polynesian (W.MP), 0.33 in Central MP, 0.0 in SHWNG, and 5.0 in Oceanic\n[i.e., CSS(FORM)= 10.33, CSS(W.MP) = 1.88, ..., CSS(OC)= 5.0]. These CSSs for\nlanguage (sub)groups are also listed in the underline portion of every section\nof (Section1 - Section 6) in Table 1. Chi-squar test (degree of freedom = 1)\nusing [Eq 1] and [Eqs.2] revealed that MSCs of MY BBPNs are distributed in\nFormosan in significantly higher frequency (P < 0.001) than in other subgroups\nof AN, as well as than in non-AN languages. MY is thus concluded to have been\nderived from Formosan of AN. Eskimo shows some BBPN similarities to FORM and\nMY.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am interested in investigating linguistics from a statistical and mathematical standpoint. My focus is on modeling linguistic patterns and phenomena as components of a complex system. I aim to compute statistical and information-theoretic features from extensive streams of natural language texts. Ideally, these features would provide insights into modeling the long-term syntactical and semantic dependencies between various linguistic elements within a lengthy text stream. Moreover, I am keen on modeling human language as a complex communication system regulated by the principles of information theory. The ultimate objective is to reveal hidden patterns and enhance or consolidate existing patterns using the language of information theory and statistics. Therefore, I need a thorough review of recent advancements at the intersection of linguistics and complex systems.","['Traditional linguistic theories have largely regard language as a formal\nsystem composed of rigid rules. However, their failures in processing real\nlanguage, the recent successes in statistical natural language processing, and\nthe findings of many psychological experiments have suggested that language may\nbe more a probabilistic system than a formal system, and thus cannot be\nfaithfully modeled with the either/or rules of formal linguistic theory. The\npresent study, based on authentic language data, confirmed that those important\nlinguistic issues, such as linguistic universal, diachronic drift, and language\nvariations can be translated into probability and frequency patterns in parole.\nThese findings suggest that human language may well be probabilistic systems by\nnature, and that statistical may well make inherent properties of human\nlanguages.', 'As is the case of many signals produced by complex systems, language presents\na statistical structure that is balanced between order and disorder. Here we\nreview and extend recent results from quantitative characterisations of the\ndegree of order in linguistic sequences that give insights into two relevant\naspects of language: the presence of statistical universals in word ordering,\nand the link between semantic information and the statistical linguistic\nstructure. We first analyse a measure of relative entropy that assesses how\nmuch the ordering of words contributes to the overall statistical structure of\nlanguage. This measure presents an almost constant value close to 3.5 bits/word\nacross several linguistic families. Then, we show that a direct application of\ninformation theory leads to an entropy measure that can quantify and extract\nsemantic structures from linguistic samples, even without prior knowledge of\nthe underlying language.', 'We review computational and robotics models of early language learning and\ndevelopment. We first explain why and how these models are used to understand\nbetter how children learn language. We argue that they provide concrete\ntheories of language learning as a complex dynamic system, complementing\ntraditional methods in psychology and linguistics. We review different modeling\nformalisms, grounded in techniques from machine learning and artificial\nintelligence such as Bayesian and neural network approaches. We then discuss\ntheir role in understanding several key mechanisms of language development:\ncross-situational statistical learning, embodiment, situated social\ninteraction, intrinsically motivated learning, and cultural evolution. We\nconclude by discussing future challenges for research, including modeling of\nlarge-scale empirical data about language acquisition in real-world\nenvironments.\n  Keywords: Early language learning, Computational and robotic models, machine\nlearning, development, embodiment, social interaction, intrinsic motivation,\nself-organization, dynamical systems, complexity.', 'The availability of an abundance of knowledge sources has spurred a large\namount of effort in the development and enhancement of Information Retrieval\ntechniques. Users information needs are expressed in natural language and\nsuccessful retrieval is very much dependent on the effective communication of\nthe intended purpose. Natural language queries consist of multiple linguistic\nfeatures which serve to represent the intended search goal. Linguistic\ncharacteristics that cause semantic ambiguity and misinterpretation of queries\nas well as additional factors such as the lack of familiarity with the search\nenvironment affect the users ability to accurately represent their information\nneeds, coined by the concept intention gap. The latter directly affects the\nrelevance of the returned search results which may not be to the users\nsatisfaction and therefore is a major issue impacting the effectiveness of\ninformation retrieval systems. Central to our discussion is the identification\nof the significant constituents that characterize the query intent and their\nenrichment through the addition of meaningful terms, phrases or even latent\nrepresentations, either manually or automatically to capture their intended\nmeaning. Specifically, we discuss techniques to achieve the enrichment and in\nparticular those utilizing the information gathered from statistical processing\nof term dependencies within a document corpus or from external knowledge\nsources such as ontologies. We lay down the anatomy of a generic linguistic\nbased query expansion framework and propose its module-based decomposition,\ncovering topical issues from query processing, information retrieval,\ncomputational linguistics and ontology engineering. For each of the modules we\nreview state-of-the-art solutions in the literature categorized and analyzed\nunder the light of the techniques used.', 'While the use of statistical physics methods to analyze large corpora has\nbeen useful to unveil many patterns in texts, no comprehensive investigation\nhas been performed investigating the properties of statistical measurements\nacross different languages and texts. In this study we propose a framework that\naims at determining if a text is compatible with a natural language and which\nlanguages are closest to it, without any knowledge of the meaning of the words.\nThe approach is based on three types of statistical measurements, i.e. obtained\nfrom first-order statistics of word properties in a text, from the topology of\ncomplex networks representing text, and from intermittency concepts where text\nis treated as a time series. Comparative experiments were performed with the\nNew Testament in 15 different languages and with distinct books in English and\nPortuguese in order to quantify the dependency of the different measurements on\nthe language and on the story being told in the book. The metrics found to be\ninformative in distinguishing real texts from their shuffled versions include\nassortativity, degree and selectivity of words. As an illustration, we analyze\nan undeciphered medieval manuscript known as the Voynich Manuscript. We show\nthat it is mostly compatible with natural languages and incompatible with\nrandom texts. We also obtain candidates for key-words of the Voynich Manuscript\nwhich could be helpful in the effort of deciphering it. Because we were able to\nidentify statistical measurements that are more dependent on the syntax than on\nthe semantics, the framework may also serve for text analysis in\nlanguage-dependent applications.', ""In this article, we evaluate computational models of natural language with\nrespect to the universal statistical behaviors of natural language. Statistical\nmechanical analyses have revealed that natural language text is characterized\nby scaling properties, which quantify the global structure in the vocabulary\npopulation and the long memory of a text. We study whether five scaling\nproperties (given by Zipf's law, Heaps' law, Ebeling's method, Taylor's law,\nand long-range correlation analysis) can serve for evaluation of computational\nmodels. Specifically, we test $n$-gram language models, a probabilistic\ncontext-free grammar (PCFG), language models based on Simon/Pitman-Yor\nprocesses, neural language models, and generative adversarial networks (GANs)\nfor text generation. Our analysis reveals that language models based on\nrecurrent neural networks (RNNs) with a gating mechanism (i.e., long short-term\nmemory, LSTM; a gated recurrent unit, GRU; and quasi-recurrent neural networks,\nQRNNs) are the only computational models that can reproduce the long memory\nbehavior of natural language. Furthermore, through comparison with recently\nproposed model-based evaluation methods, we find that the exponent of Taylor's\nlaw is a good indicator of model quality."", ""We propose a new computational approach for tracking and detecting\nstatistically significant linguistic shifts in the meaning and usage of words.\nSuch linguistic shifts are especially prevalent on the Internet, where the\nrapid exchange of ideas can quickly change a word's meaning. Our meta-analysis\napproach constructs property time series of word usage, and then uses\nstatistically sound change point detection algorithms to identify significant\nlinguistic shifts.\n  We consider and analyze three approaches of increasing complexity to generate\nsuch linguistic property time series, the culmination of which uses\ndistributional characteristics inferred from word co-occurrences. Using\nrecently proposed deep neural language models, we first train vector\nrepresentations of words for each time period. Second, we warp the vector\nspaces into one unified coordinate system. Finally, we construct a\ndistance-based distributional time series for each word to track it's\nlinguistic displacement over time.\n  We demonstrate that our approach is scalable by tracking linguistic change\nacross years of micro-blogging using Twitter, a decade of product reviews using\na corpus of movie reviews from Amazon, and a century of written books using the\nGoogle Book-ngrams. Our analysis reveals interesting patterns of language usage\nchange commensurate with each medium."", 'In this work we extend previous analyses of linguistic networks by adopting a\nmulti-layer network framework for modelling the human mental lexicon, i.e. an\nabstract mental repository where words and concepts are stored together with\ntheir linguistic patterns. Across a three-layer linguistic multiplex, we model\nEnglish words as nodes and connect them according to (i) phonological\nsimilarities, (ii) synonym relationships and (iii) free word associations. Our\nmain aim is to exploit this multi-layered structure to explore the influence of\nphonological and semantic relationships on lexicon assembly over time. We\npropose a model of lexicon growth which is driven by the phonological layer:\nwords are suggested according to different orderings of insertion (e.g. shorter\nword length, highest frequency, semantic multiplex features) and accepted or\nrejected subject to constraints. We then measure times of network assembly and\ncompare these to empirical data about the age of acquisition of words. In\nagreement with empirical studies in psycholinguistics, our results provide\nquantitative evidence for the hypothesis that word acquisition is driven by\nfeatures at multiple levels of organisation within language.', ""A family of information theoretic models of communication was introduced more\nthan a decade ago to explain the origins of Zipf's law for word frequencies.\nThe family is a based on a combination of two information theoretic principles:\nmaximization of mutual information between forms and meanings and minimization\nof form entropy. The family also sheds light on the origins of three other\npatterns: the principle of contrast, a related vocabulary learning bias and the\nmeaning-frequency law. Here two important components of the family, namely the\ninformation theoretic principles and the energy function that combines them\nlinearly, are reviewed from the perspective of psycholinguistics, language\nlearning, information theory and synergetic linguistics. The minimization of\nthis linear function is linked to the problem of compression of standard\ninformation theory and might be tuned by self-organization."", ""Progress in pre-trained language models has led to a surge of impressive\nresults on downstream tasks for natural language understanding. Recent work on\nprobing pre-trained language models uncovered a wide range of linguistic\nproperties encoded in their contextualized representations. However, it is\nunclear whether they encode semantic knowledge that is crucial to symbolic\ninference methods. We propose a methodology for probing linguistic information\nfor logical inference in pre-trained language model representations. Our\nprobing datasets cover a list of linguistic phenomena required by major\nsymbolic inference systems. We find that (i) pre-trained language models do\nencode several types of linguistic information for inference, but there are\nalso some types of information that are weakly encoded, (ii) language models\ncan effectively learn missing linguistic information through fine-tuning.\nOverall, our findings provide insights into which aspects of linguistic\ninformation for logical inference do language models and their pre-training\nprocedures capture. Moreover, we have demonstrated language models' potential\nas semantic and background knowledge bases for supporting symbolic inference\nmethods."", 'Many real systems have been modelled in terms of network concepts, and\nwritten texts are a particular example of information networks. In recent\nyears, the use of network methods to analyze language has allowed the discovery\nof several interesting findings, including the proposition of novel models to\nexplain the emergence of fundamental universal patterns. While syntactical\nnetworks, one of the most prevalent networked models of written texts, display\nboth scale-free and small-world properties, such representation fails in\ncapturing other textual features, such as the organization in topics or\nsubjects. In this context, we propose a novel network representation whose main\npurpose is to capture the semantical relationships of words in a simple way. To\ndo so, we link all words co-occurring in the same semantic context, which is\ndefined in a threefold way. We show that the proposed representations favours\nthe emergence of communities of semantically related words, and this feature\nmay be used to identify relevant topics. The proposed methodology to detect\ntopics was applied to segment selected Wikipedia articles. We have found that,\nin general, our methods outperform traditional bag-of-words representations,\nwhich suggests that a high-level textual representation may be useful to study\nsemantical features of texts.', 'Recently, the focus of complex networks research has shifted from the\nanalysis of isolated properties of a system toward a more realistic modeling of\nmultiple phenomena - multilayer networks. Motivated by the prosperity of\nmultilayer approach in social, transport or trade systems, we propose the\nintroduction of multilayer networks for language. The multilayer network of\nlanguage is a unified framework for modeling linguistic subsystems and their\nstructural properties enabling the exploration of their mutual interactions.\nVarious aspects of natural language systems can be represented as complex\nnetworks, whose vertices depict linguistic units, while links model their\nrelations. The multilayer network of language is defined by three aspects: the\nnetwork construction principle, the linguistic subsystem and the language of\ninterest. More precisely, we construct a word-level (syntax, co-occurrence and\nits shuffled counterpart) and a subword level (syllables and graphemes) network\nlayers, from five variations of original text (in the modeled language). The\nobtained results suggest that there are substantial differences between the\nnetworks structures of different language subsystems, which are hidden during\nthe exploration of an isolated layer. The word-level layers share structural\nproperties regardless of the language (e.g. Croatian or English), while the\nsyllabic subword level expresses more language dependent structural properties.\nThe preserved weighted overlap quantifies the similarity of word-level layers\nin weighted and directed networks. Moreover, the analysis of motifs reveals a\nclose topological structure of the syntactic and syllabic layers for both\nlanguages. The findings corroborate that the multilayer network framework is a\npowerful, consistent and systematic approach to model several linguistic\nsubsystems simultaneously and hence to provide a more unified view on language.', 'Much information available to applied researchers is contained within written\nlanguage or spoken text. Deep language models such as BERT have achieved\nunprecedented success in many applications of computational linguistics.\nHowever, much less is known about how these models can be used to analyze\nexisting text. We propose a novel method that combines transformer models with\nnetwork analysis to form a self-referential representation of language use\nwithin a corpus of interest. Our approach produces linguistic relations\nstrongly consistent with the underlying model as well as mathematically\nwell-defined operations on them, while reducing the amount of discretionary\nchoices of representation and distance measures. It represents, to the best of\nour knowledge, the first unsupervised method to extract semantic networks\ndirectly from deep language models. We illustrate our approach in a semantic\nanalysis of the term ""founder"". Using the entire corpus of Harvard Business\nReview from 1980 to 2020, we find that ties in our network track the semantics\nof discourse over time, and across contexts, identifying and relating clusters\nof semantic and syntactic relations. Finally, we discuss how this method can\nalso complement and inform analyses of the behavior of deep learning models.', ""Complex systems, such as life and languages, are governed by principles of\nevolution. The analogy and comparison between biology and\nlinguistics\\cite{alphafold2, RoseTTAFold, lang_virus, cell language, faculty1,\nlanguage of gene, Protein linguistics, dictionary, Grammar of pro_dom,\ncomplexity, genomics_nlp, InterPro, language modeling, Protein language\nmodeling} provide a computational foundation for characterizing and analyzing\nprotein sequences, human corpora, and their evolution. However, no general\nmathematical formula has been proposed so far to illuminate the origin of\nquantitative hallmarks shared by life and language. Here we show several new\nstatistical relationships shared by proteins and words, which inspire us to\nestablish a general mechanism of evolution with explicit formulations that can\nincorporate both old and new characteristics. We found natural selection can be\nquantified via the entropic formulation by the principle of least effort to\ndetermine the sequence variation that survives in evolution. Besides, the\norigin of power law behavior and how changes in the environment stimulate the\nemergence of new proteins and words can also be explained via the introduction\nof function connection network. Our results demonstrate not only the\ncorrespondence between genetics and linguistics over their different\nhierarchies but also new fundamental physical properties for the evolution of\ncomplex adaptive systems. We anticipate our statistical tests can function as\nquantitative criteria to examine whether an evolution theory of sequence is\nconsistent with the regularity of real data. In the meantime, their\ncorrespondence broadens the bridge to exchange existing knowledge, spurs new\ninterpretations, and opens Pandora's box to release several potentially\nrevolutionary challenges. For example, does linguistic arbitrariness conflict\nwith the dogma that structure determines function?"", 'Statistical methods have been widely employed to study the fundamental\nproperties of language. In recent years, methods from complex and dynamical\nsystems proved useful to create several language models. Despite the large\namount of studies devoted to represent texts with physical models, only a\nlimited number of studies have shown how the properties of the underlying\nphysical systems can be employed to improve the performance of natural language\nprocessing tasks. In this paper, I address this problem by devising complex\nnetworks methods that are able to improve the performance of current\nstatistical methods. Using a fuzzy classification strategy, I show that the\ntopological properties extracted from texts complement the traditional textual\ndescription. In several cases, the performance obtained with hybrid approaches\noutperformed the results obtained when only traditional or networked methods\nwere used. Because the proposed model is generic, the framework devised here\ncould be straightforwardly used to study similar textual applications where the\ntopology plays a pivotal role in the description of the interacting agents.', ""Human beings possess the most sophisticated computational machinery in the\nknown universe. We can understand language of rich descriptive power, and\ncommunicate in the same environment with astonishing clarity. Two of the many\ncontributors to the interest in natural language - the properties of\nCompositionality and Structure Dependence, are well documented, and offer a\nvast space to ask interesting modelling questions. The first step to begin\nanswering these questions is to ground verbal theory in formal terms. Drawing\non linguistics and set theory, a formalisation of these ideas is presented in\nthe first half of this thesis. We see how cognitive systems that process\nlanguage need to have certain functional constraints, viz. time based,\nincremental operations that rely on a structurally defined domain. The\nobservations that result from analysing this formal setup are examined as part\nof a modelling exercise. Using the advances of word embedding techniques, a\nmodel of relational learning is simulated with a custom dataset to demonstrate\nhow a time based role-filler binding mechanism satisfies some of the\nconstraints described in the first section. The model's ability to map\nstructure, along with its symbolic-connectionist architecture makes for a\ncognitively plausible implementation. The formalisation and simulation are\ntogether an attempt to recognise the constraints imposed by linguistic theory,\nand explore the opportunities presented by a cognitive model of relation\nlearning to realise these constraints."", 'Language is a social phenomenon and variation is inherent to its social\nnature. Recently, there has been a surge of interest within the computational\nlinguistics (CL) community in the social dimension of language. In this article\nwe present a survey of the emerging field of ""Computational Sociolinguistics""\nthat reflects this increased interest. We aim to provide a comprehensive\noverview of CL research on sociolinguistic themes, featuring topics such as the\nrelation between language and social identity, language use in social\ninteraction and multilingual communication. Moreover, we demonstrate the\npotential for synergy between the research communities involved, by showing how\nthe large-scale data-driven methods that are widely used in CL can complement\nexisting sociolinguistic studies, and how sociolinguistics can inform and\nchallenge the methods and assumptions employed in CL studies. We hope to convey\nthe possible benefits of a closer collaboration between the two communities and\nconclude with a discussion of open challenges.']","['Recent research in computational linguistics has developed algorithms which\nassociate matrices with adjectives and verbs, based on the distribution of\nwords in a corpus of text. These matrices are linear operators on a vector\nspace of context words. They are used to construct the meaning of composite\nexpressions from that of the elementary constituents, forming part of a\ncompositional distributional approach to semantics. We propose a Matrix Theory\napproach to this data, based on permutation symmetry along with Gaussian\nweights and their perturbations. A simple Gaussian model is tested against word\nmatrices created from a large corpus of text. We characterize the cubic and\nquartic departures from the model, which we propose, alongside the Gaussian\nparameters, as signatures for comparison of linguistic corpora. We propose that\nperturbed Gaussian models with permutation symmetry provide a promising\nframework for characterizing the nature of universality in the statistical\nproperties of word matrices. The matrix theory framework developed here\nexploits the view of statistics as zero dimensional perturbative quantum field\ntheory. It perceives language as a physical system realizing a universality\nclass of matrix statistics characterized by permutation symmetry.', 'Statistical methods have been widely employed in recent years to grasp many\nlanguage properties. The application of such techniques have allowed an\nimprovement of several linguistic applications, which encompasses machine\ntranslation, automatic summarization and document classification. In the\nlatter, many approaches have emphasized the semantical content of texts, as it\nis the case of bag-of-word language models. This approach has certainly yielded\nreasonable performance. However, some potential features such as the structural\norganization of texts have been used only on a few studies. In this context, we\nprobe how features derived from textual structure analysis can be effectively\nemployed in a classification task. More specifically, we performed a supervised\nclassification aiming at discriminating informative from imaginative documents.\nUsing a networked model that describes the local topological/dynamical\nproperties of function words, we achieved an accuracy rate of up to 95%, which\nis much higher than similar networked approaches. A systematic analysis of\nfeature relevance revealed that symmetry and accessibility measurements are\namong the most prominent network measurements. Our results suggest that these\nmeasurements could be used in related language applications, as they play a\ncomplementary role in characterizing texts.', 'The sequential structure of language, and the order of words in a sentence\nspecifically, plays a central role in human language processing. Consequently,\nin designing computational models of language, the de facto approach is to\npresent sentences to machines with the words ordered in the same order as in\nthe original human-authored sentence. The very essence of this work is to\nquestion the implicit assumption that this is desirable and inject theoretical\nsoundness into the consideration of word order in natural language processing.\nIn this thesis, we begin by uniting the disparate treatments of word order in\ncognitive science, psycholinguistics, computational linguistics, and natural\nlanguage processing under a flexible algorithmic framework. We proceed to use\nthis heterogeneous theoretical foundation as the basis for exploring new word\norders with an undercurrent of psycholinguistic optimality. In particular, we\nfocus on notions of dependency length minimization given the difficulties in\nhuman and computational language processing in handling long-distance\ndependencies. We then discuss algorithms for finding optimal word orders\nefficiently in spite of the combinatorial space of possibilities. We conclude\nby addressing the implications of these word orders on human language and their\ndownstream impacts when integrated in computational models.', ""Many task domains require robots to interpret and act upon natural language\ncommands which are given by people and which refer to the robot's physical\nsurroundings. Such interpretation is known variously as the symbol grounding\nproblem, grounded semantics and grounded language acquisition. This problem is\nchallenging because people employ diverse vocabulary and grammar, and because\nrobots have substantial uncertainty about the nature and contents of their\nsurroundings, making it difficult to associate the constitutive language\nelements (principally noun phrases and spatial relations) of the command text\nto elements of those surroundings. Symbolic models capture linguistic structure\nbut have not scaled successfully to handle the diverse language produced by\nuntrained users. Existing statistical approaches can better handle diversity,\nbut have not to date modeled complex linguistic structure, limiting achievable\naccuracy. Recent hybrid approaches have addressed limitations in scaling and\ncomplexity, but have not effectively associated linguistic and perceptual\nfeatures. Our framework, called Generalized Grounding Graphs (G^3), addresses\nthese issues by defining a probabilistic graphical model dynamically according\nto the linguistic parse structure of a natural language command. This approach\nscales effectively, handles linguistic diversity, and enables the system to\nassociate parts of a command with the specific objects, places, and events in\nthe external world to which they refer. We show that robots can learn word\nmeanings and use those learned meanings to robustly follow natural language\ncommands produced by untrained users. We demonstrate our approach for both\nmobility commands and mobile manipulation commands involving a variety of\nsemi-autonomous robotic platforms, including a wheelchair, a micro-air vehicle,\na forklift, and the Willow Garage PR2."", 'Distributional semantics provides multi-dimensional, graded, empirically\ninduced word representations that successfully capture many aspects of meaning\nin natural languages, as shown in a large body of work in computational\nlinguistics; yet, its impact in theoretical linguistics has so far been\nlimited. This review provides a critical discussion of the literature on\ndistributional semantics, with an emphasis on methods and results that are of\nrelevance for theoretical linguistics, in three areas: semantic change,\npolysemy and composition, and the grammar-semantics interface (specifically,\nthe interface of semantics with syntax and with derivational morphology). The\nreview aims at fostering greater cross-fertilization of theoretical and\ncomputational approaches to language, as a means to advance our collective\nknowledge of how it works.', 'In this paper, an unsupervised and cognitively driven weighted-entropy method\nfor embedding semantic categories in hyperbolic geometry is proposed. The model\nis driven by two fields of research in cognitive linguistics: the first is the\nstatistical learning theory of language acquisition and the proposal of using\nhigh-dimensional networks to represent semantic knowledge in cognition, and the\nsecond is the domain-specific approach to semantic communication. Weighted\nconditional entropy of word co-occurrence is proposed as the embedding metric,\nand the two weighting parameters are collocation diversity and conditional\nprobability ranking in the corresponding statistical distribution. The\nBoltzmann distribution is then used on the weighted-entropy metric and embedded\ninto a hyperbolic Poincare disk model. Testing has been in particular performed\nin the domains of basic color and kinship words, which belong to the classes\nthat domain-specificity focused research in cognitive semantics has most\nintensively investigated. Results show that this new approach can successfully\nmodel and map the semantic relationships of popularity and similarity for most\nof the basic color and kinship words in English and have potential to be\ngeneralized to other semantic domains and different languages. Generally, this\npaper contributes to both computational cognitive semantics and the research on\nnetwork and geometry-driven language embedding in computational linguistics and\nNLP.', 'In this paper, we propose to study language modelling as a multi-task\nproblem, bringing together three strands of research: multi-task learning,\nlinguistics, and interpretability. Based on hypotheses derived from linguistic\ntheory, we investigate whether language models adhere to learning principles of\nmulti-task learning during training. To showcase the idea, we analyse the\ngeneralisation behaviour of language models as they learn the linguistic\nconcept of Negative Polarity Items (NPIs). Our experiments demonstrate that a\nmulti-task setting naturally emerges within the objective of the more general\ntask of language modelling.We argue that this insight is valuable for\nmulti-task learning, linguistics and interpretability research and can lead to\nexciting new findings in all three domains.', 'Neuroscientists evaluate deep neural networks for natural language processing\nas possible candidate models for how language is processed in the brain. These\nmodels are often trained without explicit linguistic supervision, but have been\nshown to learn some linguistic structure in the absence of such supervision\n(Manning et al., 2020), potentially questioning the relevance of symbolic\nlinguistic theories in modeling such cognitive processes (Warstadt and Bowman,\n2020). We evaluate across two fMRI datasets whether language models align\nbetter with brain recordings, if their attention is biased by annotations from\nsyntactic or semantic formalisms. Using structure from dependency or minimal\nrecursion semantic annotations, we find alignments improve significantly for\none of the datasets. For another dataset, we see more mixed results. We present\nan extensive analysis of these results. Our proposed approach enables the\nevaluation of more targeted hypotheses about the composition of meaning in the\nbrain, expanding the range of possible scientific inferences a neuroscientist\ncould make, and opens up new opportunities for cross-pollination between\ncomputational neuroscience and linguistics.', ""In the last half-decade, the field of natural language processing (NLP) has\nundergone two major transitions: the switch to neural networks as the primary\nmodeling paradigm and the homogenization of the training regime (pre-train,\nthen fine-tune). Amidst this process, language models have emerged as NLP's\nworkhorse, displaying increasingly fluent generation capabilities and proving\nto be an indispensable means of knowledge transfer downstream. Due to the\notherwise opaque, black-box nature of such models, researchers have employed\naspects of linguistic theory in order to characterize their behavior. Questions\ncentral to syntax -- the study of the hierarchical structure of language --\nhave factored heavily into such work, shedding invaluable insights about\nmodels' inherent biases and their ability to make human-like generalizations.\nIn this paper, we attempt to take stock of this growing body of literature. In\ndoing so, we observe a lack of clarity across numerous dimensions, which\ninfluences the hypotheses that researchers form, as well as the conclusions\nthey draw from their findings. To remedy this, we urge researchers make careful\nconsiderations when investigating coding properties, selecting representations,\nand evaluating via downstream tasks. Furthermore, we outline the implications\nof the different types of research questions exhibited in studies on syntax, as\nwell as the inherent pitfalls of aggregate metrics. Ultimately, we hope that\nour discussion adds nuance to the prospect of studying language models and\npaves the way for a less monolithic perspective on syntax in this context."", ""With a growing interest in modeling inherent subjectivity in natural\nlanguage, we present a linguistically-motivated process to understand and\nanalyze the writing style of individuals from three perspectives: lexical,\nsyntactic, and semantic. We discuss the stylistically expressive elements\nwithin each of these levels and use existing methods to quantify the linguistic\nintuitions related to some of these elements. We show that such a multi-level\nanalysis is useful for developing a well-knit understanding of style - which is\nindependent of the natural language task at hand, and also demonstrate its\nvalue in solving three downstream tasks: authors' style analysis, authorship\nattribution, and emotion prediction. We conduct experiments on a variety of\ndatasets, comprising texts from social networking sites, user reviews, legal\ndocuments, literary books, and newswire. The results on the aforementioned\ntasks and datasets illustrate that such a multi-level understanding of style,\nwhich has been largely ignored in recent works, models style-related\nsubjectivity in text and can be leveraged to improve performance on multiple\ndownstream tasks both qualitatively and quantitatively."", 'Nowadays, with the booming development of the Internet, people benefit from\nits convenience due to its open and sharing nature. A large volume of natural\nlanguage texts is being generated by users in various forms, such as search\nqueries, documents, and social media posts. As the unstructured text corpus is\nusually noisy and messy, it becomes imperative to correctly identify and\naccurately annotate structured information in order to obtain meaningful\ninsights or better understand unstructured texts. On the other hand, the\nexisting structured information, which embodies our knowledge such as entity or\nconcept relations, often suffers from incompleteness or quality-related issues.\nGiven a gigantic collection of texts which offers rich semantic information, it\nis also important to harness the massiveness of the unannotated text corpus to\nexpand and refine existing structured knowledge with fewer annotation efforts.\n  In this dissertation, I will introduce principles, models, and algorithms for\neffective structured knowledge discovery from the massive text corpus. We are\ngenerally interested in obtaining insights and better understanding\nunstructured texts with the help of structured annotations or by\nstructure-aware modeling. Also, given the existing structured knowledge, we are\ninterested in expanding its scale and improving its quality harnessing the\nmassiveness of the text corpus. In particular, four problems are studied in\nthis dissertation: Structured Intent Detection for Natural Language\nUnderstanding, Structure-aware Natural Language Modeling, Generative Structured\nKnowledge Expansion, and Synonym Refinement on Structured Knowledge.', 'There is an ongoing debate in the NLP community whether modern language\nmodels contain linguistic knowledge, recovered through so-called probes. In\nthis paper, we study whether linguistic knowledge is a necessary condition for\nthe good performance of modern language models, which we call the\n\\textit{rediscovery hypothesis}. In the first place, we show that language\nmodels that are significantly compressed but perform well on their pretraining\nobjectives retain good scores when probed for linguistic structures. This\nresult supports the rediscovery hypothesis and leads to the second contribution\nof our paper: an information-theoretic framework that relates language modeling\nobjectives with linguistic information. This framework also provides a metric\nto measure the impact of linguistic information on the word prediction task. We\nreinforce our analytical results with various experiments, both on synthetic\nand on real NLP tasks in English.', 'New machine translations (MT) technologies are emerging rapidly and with\nthem, bold claims of achieving human parity such as: (i) the results produced\napproach ""accuracy achieved by average bilingual human translators"" (Wu et al.,\n2017b) or (ii) the ""translation quality is at human parity when compared to\nprofessional human translators"" (Hassan et al., 2018) have seen the light of\nday (Laubli et al., 2018). Aside from the fact that many of these papers craft\ntheir own definition of human parity, these sensational claims are often not\nsupported by a complete analysis of all aspects involved in translation.\nEstablishing the discrepancies between the strengths of statistical approaches\nto MT and the way humans translate has been the starting point of our research.\nBy looking at MT output and linguistic theory, we were able to identify some\nremaining issues. The problems range from simple number and gender agreement\nerrors to more complex phenomena such as the correct translation of aspectual\nvalues and tenses. Our experiments confirm, along with other studies\n(Bentivogli et al., 2016), that neural MT has surpassed statistical MT in many\naspects. However, some problems remain and others have emerged. We cover a\nseries of problems related to the integration of specific linguistic features\ninto statistical and neural MT, aiming to analyse and provide a solution to\nsome of them. Our work focuses on addressing three main research questions that\nrevolve around the complex relationship between linguistics and MT in general.\nWe identify linguistic information that is lacking in order for automatic\ntranslation systems to produce more accurate translations and integrate\nadditional features into the existing pipelines. We identify overgeneralization\nor \'algorithmic bias\' as a potential drawback of neural MT and link it to many\nof the remaining linguistic issues.', 'The learning trajectories of linguistic phenomena in humans provide insight\ninto linguistic representation, beyond what can be gleaned from inspecting the\nbehavior of an adult speaker. To apply a similar approach to analyze neural\nlanguage models (NLM), it is first necessary to establish that different models\nare similar enough in the generalizations they make. In this paper, we show\nthat NLMs with different initialization, architecture, and training data\nacquire linguistic phenomena in a similar order, despite their different end\nperformance. These findings suggest that there is some mutual inductive bias\nthat underlies these models\' learning of linguistic phenomena. Taking\ninspiration from psycholinguistics, we argue that studying this inductive bias\nis an opportunity to study the linguistic representation implicit in NLMs.\n  Leveraging these findings, we compare the relative performance on different\nphenomena at varying learning stages with simpler reference models. Results\nsuggest that NLMs exhibit consistent ""developmental"" stages. Moreover, we find\nthe learning trajectory to be approximately one-dimensional: given an NLM with\na certain overall performance, it is possible to predict what linguistic\ngeneralizations it has already acquired. Initial analysis of these stages\npresents phenomena clusters (notably morphological ones), whose performance\nprogresses in unison, suggesting a potential link between the generalizations\nbehind them.', 'Statistical methods have been widely employed in many practical natural\nlanguage processing applications. More specifically, complex networks concepts\nand methods from dynamical systems theory have been successfully applied to\nrecognize stylistic patterns in written texts. Despite the large amount of\nstudies devoted to represent texts with physical models, only a few studies\nhave assessed the relevance of attributes derived from the analysis of\nstylistic fluctuations. Because fluctuations represent a pivotal factor for\ncharacterizing a myriad of real systems, this study focused on the analysis of\nthe properties of stylistic fluctuations in texts via topological analysis of\ncomplex networks and intermittency measurements. The results showed that\ndifferent authors display distinct fluctuation patterns. In particular, it was\nfound that it is possible to identify the authorship of books using the\nintermittency of specific words. Taken together, the results described here\nsuggest that the patterns found in stylistic fluctuations could be used to\nanalyze other related complex systems. Furthermore, the discovery of novel\npatterns related to textual stylistic fluctuations indicates that these\npatterns could be useful to improve the state of the art of many\nstylistic-based natural language processing tasks.', ""Words in some natural languages can have a composite structure. Elements of\nthis structure include the root (that could also be composite), prefixes and\nsuffixes with which various nuances and relations to other words can be\nexpressed. Thus, in order to build a proper word representation one must take\ninto account its internal structure. From a corpus of texts we extract a set of\nfrequent subwords and from the latter set we select patterns, i.e. subwords\nwhich encapsulate information on character $n$-gram regularities. The selection\nis made using the pattern-based Conditional Random Field model with $l_1$\nregularization. Further, for every word we construct a new sequence over an\nalphabet of patterns. The new alphabet's symbols confine a local statistical\ncontext stronger than the characters, therefore they allow better\nrepresentations in ${\\mathbb{R}}^n$ and are better building blocks for word\nrepresentation. In the task of subword-aware language modeling, pattern-based\nmodels outperform character-based analogues by 2-20 perplexity points. Also, a\nrecurrent neural network in which a word is represented as a sum of embeddings\nof its patterns is on par with a competitive and significantly more\nsophisticated character-based convolutional architecture."", 'We analyze the information provided by the word embeddings about the\ngrammatical gender in Swedish. We wish that this paper may serve as one of the\nbridges to connect the methods of computational linguistics and general\nlinguistics. Taking nominal classification in Swedish as a case study, we first\nshow how the information about grammatical gender in language can be captured\nby word embedding models and artificial neural networks. Then, we match our\nresults with previous linguistic hypotheses on assignment and usage of\ngrammatical gender in Swedish and analyze the errors made by the computational\nmodel from a linguistic perspective.', ""The mathematical formalism of quantum theory has been successfully used in\nhuman cognition to model decision processes and to deliver representations of\nhuman knowledge. As such, quantum cognition inspired tools have improved\ntechnologies for Natural Language Processing and Information Retrieval. In this\npaper, we overview the quantum cognition approach developed in our Brussels\nteam during the last two decades, specifically our identification of quantum\nstructures in human concepts and language, and the modeling of data from\npsychological and corpus-text-based experiments. We discuss our\nquantum-theoretic framework for concepts and their conjunctions/disjunctions in\na Fock-Hilbert space structure, adequately modeling a large amount of data\ncollected on concept combinations. Inspired by this modeling, we put forward\nelements for a quantum contextual and meaning-based approach to information\ntechnologies in which 'entities of meaning' are inversely reconstructed from\ntexts, which are considered as traces of these entities' states."", 'Social media platforms provide a goldmine for mining public opinion on issues\nof wide societal interest and impact. Opinion mining is a problem that can be\noperationalised by capturing and aggregating the stance of individual social\nmedia posts as supporting, opposing or being neutral towards the issue at hand.\nWhile most prior work in stance detection has investigated datasets that cover\nshort periods of time, interest in investigating longitudinal datasets has\nrecently increased. Evolving dynamics in linguistic and behavioural patterns\nobserved in new data require adapting stance detection systems to deal with the\nchanges. In this survey paper, we investigate the intersection between\ncomputational linguistics and the temporal evolution of human communication in\ndigital media. We perform a critical review of emerging research considering\ndynamics, exploring different semantic and pragmatic factors that impact\nlinguistic data in general, and stance in particular. We further discuss\ncurrent directions in capturing stance dynamics in social media. We discuss the\nchallenges encountered when dealing with stance dynamics, identify open\nchallenges and discuss future directions in three key dimensions: utterance,\ncontext and influence.', 'Citation content analysis seeks to understand citations based on the language\nused during the making of a citation. A key issue in citation content analysis\nis looking for linguistic structures that characterize distinct classes of\ncitations for the purposes of understanding the intent and function of a\ncitation. Previous works have focused on modeling linguistic features first and\ndrawn conclusions on the language structures unique to each class of citation\nfunction based on the performance of a classification task or inter-annotator\nagreement. In this study, we start with a large sample of a pre-classified\ncitation corpus, 2 million citations from each class of the scite Smart\nCitation dataset (supporting, disputing, and mentioning citations), and analyze\nits corpus linguistics in order to reveal the unique and statistically\nsignificant language structures belonging to each type of citation. By\ngenerating comparison tables for each citation type we present a number of\ninteresting linguistic features that uniquely characterize citation type. What\nwe find is that within citation collocates, there is very low correlation\nbetween citation type and sentiment. Additionally, we find that the\nsubjectivity of citation collocates across classes is very low. These findings\nsuggest that the sentiment of collocates is not a predictor of citation\nfunction and that due to their low subjectivity, an opinion-expressing mode of\nunderstanding citations, implicit in previous citation sentiment analysis\nliterature, is inappropriate. Instead, we suggest that citations can be better\nunderstood as claims-making devices where the citation type can be explained by\nunderstanding how two claims are being compared. By presenting this approach,\nwe hope to inspire similar corpus linguistic studies on citations that derive a\nmore robust theory of citation from an empirical basis using citation corpora', 'It is often stated that human languages, as other biological systems, are\nshaped by cost-cutting pressures but, to what extent? Attempts to quantify the\ndegree of optimality of languages by means of an optimality score have been\nscarce and focused mostly on English. Here we recast the problem of the\noptimality of the word order of a sentence as an optimization problem on a\nspatial network where the vertices are words, arcs indicate syntactic\ndependencies and the space is defined by the linear order of the words in the\nsentence. We introduce a new score to quantify the cognitive pressure to reduce\nthe distance between linked words in a sentence. The analysis of sentences from\n93 languages representing 19 linguistic families reveals that half of languages\nare optimized to a 70% or more. The score indicates that distances are not\nsignificantly reduced in a few languages and confirms two theoretical\npredictions, i.e. that longer sentences are more optimized and that distances\nare more likely to be longer than expected by chance in short sentences. We\npresent a new hierarchical ranking of languages by their degree of\noptimization. The new score has implications for various fields of language\nresearch (dependency linguistics, typology, historical linguistics, clinical\nlinguistics and cognitive science). Finally, the principles behind the design\nof the score have implications for network science.', ""Human language is recognized as a very complex domain since decades. No\ncomputer system has been able to reach human levels of performance so far. The\nonly known computational system capable of proper language processing is the\nhuman brain. While we gather more and more data about the brain, its\nfundamental computational processes still remain obscure. The lack of a sound\ncomputational brain theory also prevents the fundamental understanding of\nNatural Language Processing. As always when science lacks a theoretical\nfoundation, statistical modeling is applied to accommodate as many sampled\nreal-world data as possible. An unsolved fundamental issue is the actual\nrepresentation of language (data) within the brain, denoted as the\nRepresentational Problem. Starting with Jeff Hawkins' Hierarchical Temporal\nMemory (HTM) theory, a consistent computational theory of the human cortex, we\nhave developed a corresponding theory of language data representation: The\nSemantic Folding Theory. The process of encoding words, by using a topographic\nsemantic space as distributional reference frame into a sparse binary\nrepresentational vector is called Semantic Folding and is the central topic of\nthis document. Semantic Folding describes a method of converting language from\nits symbolic representation (text) into an explicit, semantically grounded\nrepresentation that can be generically processed by Hawkins' HTM networks. As\nit turned out, this change in representation, by itself, can solve many complex\nNLP problems by applying Boolean operators and a generic similarity function\nlike the Euclidian Distance. Many practical problems of statistical NLP\nsystems, like the high cost of computation, the fundamental incongruity of\nprecision and recall , the complex tuning procedures etc., can be elegantly\novercome by applying Semantic Folding."", 'Formal Semantics and Distributional Semantics are two important semantic\nframeworks in Natural Language Processing (NLP). Cognitive Semantics belongs to\nthe movement of Cognitive Linguistics, which is based on contemporary cognitive\nscience. Each framework could deal with some meaning phenomena, but none of\nthem fulfills all requirements proposed by applications. A unified semantic\ntheory characterizing all important language phenomena has both theoretical and\npractical significance; however, although many attempts have been made in\nrecent years, no existing theory has achieved this goal yet.\n  This article introduces a new semantic theory that has the potential to\ncharacterize most of the important meaning phenomena of natural language and to\nfulfill most of the necessary requirements for philosophical analysis and for\nNLP applications. The theory is based on a unified representation of\ninformation, and constructs a kind of mathematical model called cognitive model\nto interpret natural language expressions in a compositional manner. It accepts\nthe empirical assumption of Cognitive Semantics, and overcomes most\nshortcomings of Formal Semantics and of Distributional Semantics. The theory,\nhowever, is not a simple combination of existing theories, but an extensive\ngeneralization of classic logic and Formal Semantics. It inherits nearly all\nadvantages of Formal Semantics, and also provides descriptive contents for\nobjects and events as fine-gram as possible, descriptive contents which\nrepresent the results of human cognition.', 'Purpose: The purpose of this paper is to identify the factors influencing the\ngrowth of healthcare virtual communities of practice (VCoPs) through a\nseven-year longitudinal study conducted using metrics from social-network and\nsemantic analysis. By studying online communication along the three dimensions\nof social interactions (connectivity, interactivity and language use), the\nauthors aim to provide VCoP managers with valuable insights to improve the\nsuccess of their communities. Design/methodology/approach: Communications over\na period of seven years (April 2008 to April 2015) and between 14,000 members\nof 16 different healthcare VCoPs coexisting on the same web platform were\nanalysed. Multilevel regression models were used to reveal the main\ndeterminants of community growth over time. Independent variables were derived\nfrom social network and semantic analysis measures. Findings: Results show that\nstructural and content-based variables predict the growth of the community.\nProgressively, more people will join a community if its structure is more\ncentralised, leaders are more dynamic (they rotate more) and the language used\nin the posts is less complex. Research limitations/implications: The available\ndata set included one Web platform and a limited number of control variables.\nTo consolidate the findings of the present study, the experiment should be\nreplicated on other healthcare VCoPs. Originality/value: The study provides\nuseful recommendations for setting up and nurturing the growth of professional\ncommunities, considering, at the same time, the interaction patterns among the\ncommunity members, the dynamic evolution of these interactions and the use of\nlanguage. New analytical tools are presented, together with the use of\ninnovative interaction metrics, that can significantly influence community\ngrowth, such as rotating leadership.', ""We present a study of the relationship between gender, linguistic style, and\nsocial networks, using a novel corpus of 14,000 Twitter users. Prior\nquantitative work on gender often treats this social variable as a female/male\nbinary; we argue for a more nuanced approach. By clustering Twitter users, we\nfind a natural decomposition of the dataset into various styles and topical\ninterests. Many clusters have strong gender orientations, but their use of\nlinguistic resources sometimes directly conflicts with the population-level\nlanguage statistics. We view these clusters as a more accurate reflection of\nthe multifaceted nature of gendered language styles. Previous corpus-based work\nhas also had little to say about individuals whose linguistic styles defy\npopulation-level gender patterns. To identify such individuals, we train a\nstatistical classifier, and measure the classifier confidence for each\nindividual in the dataset. Examining individuals whose language does not match\nthe classifier's model for their gender, we find that they have social networks\nthat include significantly fewer same-gender social connections and that, in\ngeneral, social network homophily is correlated with the use of same-gender\nlanguage markers. Pairing computational methods and social theory thus offers a\nnew perspective on how gender emerges as individuals position themselves\nrelative to audiences, topics, and mainstream gender norms."", 'Do state-of-the-art models for language understanding already have, or can\nthey easily learn, abilities such as boolean coordination, quantification,\nconditionals, comparatives, and monotonicity reasoning (i.e., reasoning about\nword substitutions in sentential contexts)? While such phenomena are involved\nin natural language inference (NLI) and go beyond basic linguistic\nunderstanding, it is unclear the extent to which they are captured in existing\nNLI benchmarks and effectively learned by models. To investigate this, we\npropose the use of semantic fragments---systematically generated datasets that\neach target a different semantic phenomenon---for probing, and efficiently\nimproving, such capabilities of linguistic models. This approach to creating\nchallenge datasets allows direct control over the semantic diversity and\ncomplexity of the targeted linguistic phenomena, and results in a more precise\ncharacterization of a model\'s linguistic behavior. Our experiments, using a\nlibrary of 8 such semantic fragments, reveal two remarkable findings: (a)\nState-of-the-art models, including BERT, that are pre-trained on existing NLI\nbenchmark datasets perform poorly on these new fragments, even though the\nphenomena probed here are central to the NLI task. (b) On the other hand, with\nonly a few minutes of additional fine-tuning---with a carefully selected\nlearning rate and a novel variation of ""inoculation""---a BERT-based model can\nmaster all of these logic and monotonicity fragments while retaining its\nperformance on established NLI benchmarks.', 'The use of linguistic typological resources in natural language processing\nhas been steadily gaining more popularity. It has been observed that the use of\ntypological information, often combined with distributed language\nrepresentations, leads to significantly more powerful models. While linguistic\ntypology representations from various resources have mostly been used for\nconditioning the models, there has been relatively little attention on\npredicting features from these resources from the input data. In this paper we\ninvestigate whether the various linguistic features from World Atlas of\nLanguage Structures (WALS) can be reliably inferred from multi-lingual text.\nSuch a predictor can be used to infer structural features for a language never\nobserved in training data. We frame this task as a multi-label classification\ninvolving predicting the set of non-mutually exclusive and extremely sparse\nmulti-valued labels (WALS features). We construct a recurrent neural network\npredictor based on byte embeddings and convolutional layers and test its\nperformance on 556 languages, providing analysis for various linguistic types,\nmacro-areas, language families and individual features. We show that some\nfeatures from various linguistic types can be predicted reliably.', ""It is widely accepted that information derived from analyzing speech (the\nacoustic signal) and language production (words and sentences) serves as a\nuseful window into the health of an individual's cognitive ability. In fact,\nmost neuropsychological testing batteries have a component related to speech\nand language where clinicians elicit speech from patients for subjective\nevaluation across a broad set of dimensions. With advances in speech signal\nprocessing and natural language processing, there has been recent interest in\ndeveloping tools to detect more subtle changes in cognitive-linguistic\nfunction. This work relies on extracting a set of features from recorded and\ntranscribed speech for objective assessments of speech and language, early\ndiagnosis of neurological disease, and tracking of disease after diagnosis.\nWith an emphasis on cognitive and thought disorders, in this paper we provide a\nreview of existing speech and language features used in this domain, discuss\ntheir clinical application, and highlight their advantages and disadvantages.\nBroadly speaking, the review is split into two categories: language features\nbased on natural language processing and speech features based on speech signal\nprocessing. Within each category, we consider features that aim to measure\ncomplementary dimensions of cognitive-linguistics, including language\ndiversity, syntactic complexity, semantic coherence, and timing. We conclude\nthe review with a proposal of new research directions to further advance the\nfield."", 'The goal of language modeling techniques is to capture the statistical and\nstructural properties of natural languages from training corpora. This task\ntypically involves the learning of short range dependencies, which generally\nmodel the syntactic properties of a language and/or long range dependencies,\nwhich are semantic in nature. We propose in this paper a new multi-span\narchitecture, which separately models the short and long context information\nwhile it dynamically merges them to perform the language modeling task. This is\ndone through a novel recurrent Long-Short Range Context (LSRC) network, which\nexplicitly models the local (short) and global (long) context using two\nseparate hidden states that evolve in time. This new architecture is an\nadaptation of the Long-Short Term Memory network (LSTM) to take into account\nthe linguistic properties. Extensive experiments conducted on the Penn Treebank\n(PTB) and the Large Text Compression Benchmark (LTCB) corpus showed a\nsignificant reduction of the perplexity when compared to state-of-the-art\nlanguage modeling techniques.', 'Exploiting the fact that natural languages are complex systems, the present\nexploratory article proposes a direct method based on frequency distributions\nthat may be useful when making a decision on the status of problematic\nphonemes, an open problem in linguistics. The main notion is that natural\nlanguages, which can be considered from a complex outlook as information\nprocessing machines, and which somehow manage to set appropriate levels of\nredundancy, already ""made the choice"" whether a linguistic unit is a phoneme or\nnot, and this would be reflected in a greater smoothness in a frequency versus\nrank graph. For the particular case we chose to study, we conclude that it is\nreasonable to consider the Spanish semiconsonant /w/ as a separate phoneme from\nits vowel counterpart /u/, on the one hand, and possibly also the semiconsonant\n/j/ as a separate phoneme from its vowel counterpart /i/, on the other. As\nlanguage has been so central a topic in the study of complexity, this\ndiscussion grants us, in addition, an opportunity to gain insight into emerging\nproperties in the broader complex systems debate.', 'A short review of similarities between dolphins and humans with the help of\nquantitative linguistics and information theory.', ""We investigated the impact of noisy linguistic features on the performance of\na Japanese speech synthesis system based on neural network that uses WaveNet\nvocoder. We compared an ideal system that uses manually corrected linguistic\nfeatures including phoneme and prosodic information in training and test sets\nagainst a few other systems that use corrupted linguistic features. Both\nsubjective and objective results demonstrate that corrupted linguistic\nfeatures, especially those in the test set, affected the ideal system's\nperformance significantly in a statistical sense due to a mismatched condition\nbetween the training and test sets. Interestingly, while an utterance-level\nTuring test showed that listeners had a difficult time differentiating\nsynthetic speech from natural speech, it further indicated that adding noise to\nthe linguistic features in the training set can partially reduce the effect of\nthe mismatch, regularize the model, and help the system perform better when\nlinguistic features of the test set are noisy."", 'Natural language descriptions sometimes accompany visualizations to better\ncommunicate and contextualize their insights, and to improve their\naccessibility for readers with disabilities. However, it is difficult to\nevaluate the usefulness of these descriptions, and how effectively they improve\naccess to meaningful information, because we have little understanding of the\nsemantic content they convey, and how different readers receive this content.\nIn response, we introduce a conceptual model for the semantic content conveyed\nby natural language descriptions of visualizations. Developed through a\ngrounded theory analysis of 2,147 sentences, our model spans four levels of\nsemantic content: enumerating visualization construction properties (e.g.,\nmarks and encodings); reporting statistical concepts and relations (e.g.,\nextrema and correlations); identifying perceptual and cognitive phenomena\n(e.g., complex trends and patterns); and elucidating domain-specific insights\n(e.g., social and political context). To demonstrate how our model can be\napplied to evaluate the effectiveness of visualization descriptions, we conduct\na mixed-methods evaluation with 30 blind and 90 sighted readers, and find that\nthese reader groups differ significantly on which semantic content they rank as\nmost useful. Together, our model and findings suggest that access to meaningful\ninformation is strongly reader-specific, and that research in automatic\nvisualization captioning should orient toward descriptions that more richly\ncommunicate overall trends and statistics, sensitive to reader preferences. Our\nwork further opens a space of research on natural language as a data interface\ncoequal with visualization.', 'Linguistic laws constitute one of the quantitative cornerstones of modern\ncognitive sciences and have been routinely investigated in written corpora, or\nin the equivalent transcription of oral corpora. This means that inferences of\nstatistical patterns of language in acoustics are biased by the arbitrary,\nlanguage-dependent segmentation of the signal, and virtually precludes the\npossibility of making comparative studies between human voice and other animal\ncommunication systems. Here we bridge this gap by proposing a method that\nallows to measure such patterns in acoustic signals of arbitrary origin,\nwithout needs to have access to the language corpus underneath. The method has\nbeen applied to six different human languages, recovering successfully some\nwell-known laws of human communication at timescales even below the phoneme and\nfinding yet another link between complexity and criticality in a biological\nsystem. These methods further pave the way for new comparative studies in\nanimal communication or the analysis of signals of unknown code.', 'This project explores the nature of language acquisition in computers, guided\nby techniques similar to those used in children. While existing natural\nlanguage processing methods are limited in scope and understanding, our system\naims to gain an understanding of language from first principles and hence\nminimal initial input. The first portion of our system was implemented in Java\nand is focused on understanding the morphology of language using bigrams. We\nuse frequency distributions and differences between them to define and\ndistinguish languages. English and French texts were analyzed to determine a\ndifference threshold of 55 before the texts are considered to be in different\nlanguages, and this threshold was verified using Spanish texts. The second\nportion of our system focuses on gaining an understanding of the syntax of a\nlanguage using a recursive method. The program uses one of two possible methods\nto analyze given sentences based on either sentence patterns or surrounding\nwords. Both methods have been implemented in C++. The program is able to\nunderstand the structure of simple sentences and learn new words. In addition,\nwe have provided some suggestions regarding future work and potential\nextensions of the existing program.', ""My doctoral research focuses on understanding semantic knowledge in neural\nnetwork models trained solely to predict natural language (referred to as\nlanguage models, or LMs), by drawing on insights from the study of concepts and\ncategories grounded in cognitive science. I propose a framework inspired by\n'inductive reasoning,' a phenomenon that sheds light on how humans utilize\nbackground knowledge to make inductive leaps and generalize from new pieces of\ninformation about concepts and their properties. Drawing from experiments that\nstudy inductive reasoning, I propose to analyze semantic inductive\ngeneralization in LMs using phenomena observed in human-induction literature,\ninvestigate inductive behavior on tasks such as implicit reasoning and emergent\nfeature recognition, and analyze and relate induction dynamics to the learned\nconceptual representation space."", 'In recent years, automated approaches to assessing linguistic complexity in\nsecond language (L2) writing have made significant progress in gauging learner\nperformance, predicting human ratings of the quality of learner productions,\nand benchmarking L2 development. In contrast, there is comparatively little\nwork in the area of speaking, particularly with respect to fully automated\napproaches to assessing L2 spontaneous speech. While the importance of a\nwell-performing ASR system is widely recognized, little research has been\nconducted to investigate the impact of its performance on subsequent automatic\ntext analysis. In this paper, we focus on this issue and examine the impact of\nusing a state-of-the-art ASR system for subsequent automatic analysis of\nlinguistic complexity in spontaneously produced L2 speech. A set of 30 selected\nmeasures were considered, falling into four categories: syntactic, lexical,\nn-gram frequency, and information-theoretic measures. The agreement between the\nscores for these measures obtained on the basis of ASR-generated vs. manual\ntranscriptions was determined through correlation analysis. A more differential\neffect of ASR performance on specific types of complexity measures when\ncontrolling for task type effects is also presented.', 'This paper aims to catalyze the discussions about text feature extraction\ntechniques using neural network architectures. The research questions discussed\nin the paper focus on the state-of-the-art neural network techniques that have\nproven to be useful tools for language processing, language generation, text\nclassification and other computational linguistics tasks.', 'Neural models excel at extracting statistical patterns from large amounts of\ndata, but struggle to learn patterns or reason about language from only a few\nexamples. In this paper, we ask: Can we learn explicit rules that generalize\nwell from only a few examples? We explore this question using program\nsynthesis. We develop a synthesis model to learn phonology rules as programs in\na domain-specific language. We test the ability of our models to generalize\nfrom few training examples using our new dataset of problems from the\nLinguistics Olympiad, a challenging set of tasks that require strong linguistic\nreasoning ability. In addition to being highly sample-efficient, our approach\ngenerates human-readable programs, and allows control over the generalizability\nof the learnt programs.', 'More and more languages in the world are under study nowadays, as a result,\nthe traditional way of historical linguistics study is facing some challenges.\nFor example, the linguistic comparative research among languages needs manual\nannotation, which becomes more and more impossible with the increasing amount\nof language data coming out all around the world. Although it could hardly\nreplace linguists work, the automatic computational methods have been taken\ninto consideration and it can help people reduce their workload. One of the\nmost important work in historical linguistics is word comparison from different\nlanguages and find the cognate words for them, which means people try to figure\nout if the two languages are related to each other or not. In this paper, I am\ngoing to use computational method to cluster the languages and use Markov Chain\nMonte Carlo (MCMC) method to build the language typology relationship tree\nbased on the clusters.', 'Distributional semantics has had enormous empirical success in Computational\nLinguistics and Cognitive Science in modeling various semantic phenomena, such\nas semantic similarity, and distributional models are widely used in\nstate-of-the-art Natural Language Processing systems. However, the theoretical\nstatus of distributional semantics within a broader theory of language and\ncognition is still unclear: What does distributional semantics model? Can it\nbe, on its own, a fully adequate model of the meanings of linguistic\nexpressions? The standard answer is that distributional semantics is not fully\nadequate in this regard, because it falls short on some of the central aspects\nof formal semantic approaches: truth conditions, entailment, reference, and\ncertain aspects of compositionality. We argue that this standard answer rests\non a misconception: These aspects do not belong in a theory of expression\nmeaning, they are instead aspects of speaker meaning, i.e., communicative\nintentions in a particular context. In a slogan: words do not refer, speakers\ndo. Clearing this up enables us to argue that distributional semantics on its\nown is an adequate model of expression meaning. Our proposal sheds light on the\nrole of distributional semantics in a broader theory of language and cognition,\nits relationship to formal semantics, and its place in computational models.', ""Linguistic typology aims to capture structural and semantic variation across\nthe world's languages. A large-scale typology could provide excellent guidance\nfor multilingual Natural Language Processing (NLP), particularly for languages\nthat suffer from the lack of human labeled resources. We present an extensive\nliterature survey on the use of typological information in the development of\nNLP techniques. Our survey demonstrates that to date, the use of information in\nexisting typological databases has resulted in consistent but modest\nimprovements in system performance. We show that this is due to both intrinsic\nlimitations of databases (in terms of coverage and feature granularity) and\nunder-employment of the typological features included in them. We advocate for\na new approach that adapts the broad and discrete nature of typological\ncategories to the contextual and continuous nature of machine learning\nalgorithms used in contemporary NLP. In particular, we suggest that such\napproach could be facilitated by recent developments in data-driven induction\nof typological knowledge."", ""Qualitative analysis of verbal data is of central importance in the learning\nsciences. It is labor-intensive and time-consuming, however, which limits the\namount of data researchers can include in studies. This work is a step towards\nbuilding a statistical machine learning (ML) method for achieving an automated\nsupport for qualitative analyses of students' writing, here specifically in\nscore laboratory reports in introductory biology for sophistication of\nargumentation and reasoning. We start with a set of lab reports from an\nundergraduate biology course, scored by a four-level scheme that considers the\ncomplexity of argument structure, the scope of evidence, and the care and\nnuance of conclusions. Using this set of labeled data, we show that a popular\nnatural language modeling processing pipeline, namely vector representation of\nwords, a.k.a word embeddings, followed by Long Short Term Memory (LSTM) model\nfor capturing language generation as a state-space model, is able to\nquantitatively capture the scoring, with a high Quadratic Weighted Kappa (QWK)\nprediction score, when trained in via a novel contrastive learning set-up. We\nshow that the ML algorithm approached the inter-rater reliability of human\nanalysis. Ultimately, we conclude, that machine learning (ML) for natural\nlanguage processing (NLP) holds promise for assisting learning sciences\nresearchers in conducting qualitative studies at much larger scales than is\ncurrently possible."", 'Modelling concept representation is a foundational problem in the study of\ncognition and linguistics. This work builds on the confluence of conceptual\ntools from G\\""ardenfors semantic spaces, categorical compositional linguistics,\nand applied category theory to present a domain-independent and categorical\nformalism of \'concept\'.', 'This research offers a new interdisciplinary approach to the field of\nLinguistics by using Computational Linguistics, NLP, Bayesian Statistics and\nSociolinguistics methods. This thesis investigates word order change in\ninfinitival clauses from Object-Verb (OV) to Verb-Object (VO) in the history of\nLatin and Old French. By applying a variationist approach, I examine a\nsynchronic word order variation in each stage of language change, from which I\ninfer the character, periodization and constraints of diachronic variation. I\nalso show that in discourse-configurational languages, such as Latin and Early\nOld French, it is possible to identify pragmatically neutral contexts by using\ninformation structure annotation. I further argue that by mapping pragmatic\ncategories into a syntactic structure, we can detect how word order change\nunfolds. For this investigation, the data are extracted from annotated corpora\nspanning several centuries of Latin and Old French and from additional\nresources created by using computational linguistic methods. The data are then\nfurther codified for various pragmatic, semantic, syntactic and sociolinguistic\nfactors. This study also evaluates previous factors proposed to account for\nword order alternation and change. I show how information structure and\nsyntactic constraints change over time and propose a method that allows\nresearchers to differentiate a stable word order alternation from alternation\nindicating a change. Finally, I present a three-stage probabilistic model of\nword order change, which also conforms to traditional language change patterns.', 'Counting letters in written texts is a very ancient practice. It has\naccompanied the development of Cryptology, Quantitative Linguistics, and\nStatistics. In Cryptology, counting frequencies of the different characters in\nan encrypted message is the basis of the so called frequency analysis method.\nIn Quantitative Linguistics, the proportion of vowels to consonants in\ndifferent languages was studied long before authorship attribution. In\nStatistics, the alternation vowel-consonants was the only example that Markov\never gave of his theory of chained events. A short history of letter counting\nis presented. The three domains, Cryptology, Quantitative Linguistics, and\nStatistics, are then examined, focusing on the interactions with the other two\nfields through letter counting. As a conclusion, the eclectism of past\ncenturies scholars, their background in humanities, and their familiarity with\ncryptograms, are identified as contributing factors to the mutual enrichment\nprocess which is described here.', ""Speech-controlled user interfaces facilitate the operation of devices and\nhousehold functions to laymen. State-of-the-art language technology scans the\nacoustically analyzed speech signal for relevant keywords that are subsequently\ninserted into semantic slots to interpret the user's intent. In order to\ndevelop proper cognitive information and communication technologies, simple\nslot-filling should be replaced by utterance meaning transducers (UMT) that are\nbased on semantic parsers and a \\emph{mental lexicon}, comprising syntactic,\nphonetic and semantic features of the language under consideration. This\nlexicon must be acquired by a cognitive agent during interaction with its\nusers. We outline a reinforcement learning algorithm for the acquisition of the\nsyntactic morphology and arithmetic semantics of English numerals, based on\nminimalist grammar (MG), a recent computational implementation of generative\nlinguistics. Number words are presented to the agent by a teacher in form of\nutterance meaning pairs (UMP) where the meanings are encoded as arithmetic\nterms from a suitable term algebra. Since MG encodes universal linguistic\ncompetence through inference rules, thereby separating innate linguistic\nknowledge from the contingently acquired lexicon, our approach unifies\ngenerative grammar and reinforcement learning, hence potentially resolving the\nstill pending Chomsky-Skinner controversy."", ""Word meaning changes over time, depending on linguistic and extra-linguistic\nfactors. Associating a word's correct meaning in its historical context is a\ncentral challenge in diachronic research, and is relevant to a range of NLP\ntasks, including information retrieval and semantic search in historical texts.\nBayesian models for semantic change have emerged as a powerful tool to address\nthis challenge, providing explicit and interpretable representations of\nsemantic change phenomena. However, while corpora typically come with rich\nmetadata, existing models are limited by their inability to exploit contextual\ninformation (such as text genre) beyond the document time-stamp. This is\nparticularly critical in the case of ancient languages, where lack of data and\nlong diachronic span make it harder to draw a clear distinction between\npolysemy (the fact that a word has several senses) and semantic change (the\nprocess of acquiring, losing, or changing senses), and current systems perform\npoorly on these languages. We develop GASC, a dynamic semantic change model\nthat leverages categorical metadata about the texts' genre to boost inference\nand uncover the evolution of meanings in Ancient Greek corpora. In a new\nevaluation framework, our model achieves improved predictive performance\ncompared to the state of the art."", 'For building question answering systems and natural language interfaces,\nsemantic parsing has emerged as an important and powerful paradigm. Semantic\nparsers map natural language into logical forms, the classic representation for\nmany important linguistic phenomena. The modern twist is that we are interested\nin learning semantic parsers from data, which introduces a new layer of\nstatistical and computational issues. This article lays out the components of a\nstatistical semantic parser, highlighting the key challenges. We will see that\nsemantic parsing is a rich fusion of the logical and the statistical world, and\nthat this fusion will play an integral role in the future of natural language\nunderstanding systems.', 'The annotation of textual information is a fundamental activity in\nLinguistics and Computational Linguistics. This article presents various\nobservations on annotations. It approaches the topic from several angles\nincluding Hypertext, Computational Linguistics and Language Technology,\nArtificial Intelligence and Open Science. Annotations can be examined along\ndifferent dimensions. In terms of complexity, they can range from trivial to\nhighly sophisticated, in terms of maturity from experimental to standardised.\nAnnotations can be annotated themselves using more abstract annotations.\nPrimary research data such as, e.g., text documents can be annotated on\ndifferent layers concurrently, which are independent but can be exploited using\nmulti-layer querying. Standards guarantee interoperability and reusability of\ndata sets. The chapter concludes with four final observations, formulated as\nresearch questions or rather provocative remarks on the current state of\nannotation research.', 'With the increasing number of texts made available on the Internet, many\napplications have relied on text mining tools to tackle a diversity of\nproblems. A relevant model to represent texts is the so-called word adjacency\n(co-occurrence) representation, which is known to capture mainly syntactical\nfeatures of texts.In this study, we introduce a novel network representation\nthat considers the semantic similarity between paragraphs. Two main properties\nof paragraph networks are considered: (i) their ability to incorporate\ncharacteristics that can discriminate real from artificial, shuffled\nmanuscripts and (ii) their ability to capture syntactical and semantic textual\nfeatures. Our results revealed that real texts are organized into communities,\nwhich turned out to be an important feature for discriminating them from\nartificial texts. Interestingly, we have also found that, differently from\ntraditional co-occurrence networks, the adopted representation is able to\ncapture semantic features. Additionally, the proposed framework was employed to\nanalyze the Voynich manuscript, which was found to be compatible with texts\nwritten in natural languages. Taken together, our findings suggest that the\nproposed methodology can be combined with traditional network models to improve\ntext classification tasks.', ""Representation of linguistic phenomena in computational language models is\ntypically assessed against the predictions of existing linguistic theories of\nthese phenomena. Using the notion of polarity as a case study, we show that\nthis is not always the most adequate set-up. We probe polarity via so-called\n'negative polarity items' (in particular, English 'any') in two pre-trained\nTransformer-based models (BERT and GPT-2). We show that - at least for polarity\n- metrics derived from language models are more consistent with data from\npsycholinguistic experiments than linguistic theory predictions. Establishing\nthis allows us to more adequately evaluate the performance of language models\nand also to use language models to discover new insights into natural language\ngrammar beyond existing linguistic theories. This work contributes to\nestablishing closer ties between psycholinguistic experiments and experiments\nwith language models."", 'In recent years, central components of a new approach to linguistics, the\nMinimalist Program (MP) have come closer to physics. Features of the Minimalist\nProgram, such as the unconstrained nature of recursive Merge, the operation of\nthe Labeling Algorithm that only operates at the interface of Narrow Syntax\nwith the Conceptual-Intentional and the Sensory-Motor interfaces, the\ndifference between pronounced and un-pronounced copies of elements in a\nsentence and the build-up of the Fibonacci sequence in the syntactic derivation\nof sentence structures, are directly accessible to representation in terms of\nalgebraic formalism. Although in our scheme linguistic structures are classical\nones, we find that an interesting and productive isomorphism can be established\nbetween the MP structure, algebraic structures and many-body field theory\nopening new avenues of inquiry on the dynamics underlying some central aspects\nof linguistics.', 'This special issue is dedicated to get a better picture of the relationships\nbetween computational linguistics and cognitive science. It specifically raises\ntwo questions: ""what is the potential contribution of computational language\nmodeling to cognitive science?"" and conversely: ""what is the influence of\ncognitive science in contemporary computational linguistics?""', 'Significant advances have been made in artificial systems by using biological\nsystems as a guide. However, there is often little interaction between\ncomputational models for emergent communication and biological models of the\nemergence of language. Many researchers in language origins and emergent\ncommunication take compositionality as their primary target for explaining how\nsimple communication systems can become more like natural language. However,\nthere is reason to think that compositionality is the wrong target on the\nbiological side, and so too the wrong target on the machine-learning side. As\nsuch, the purpose of this paper is to explore this claim. This has theoretical\nimplications for language origins research more generally, but the focus here\nwill be the implications for research on emergent communication in computer\nscience and machine learning---specifically regarding the types of programmes\nthat might be expected to work and those which will not. I further suggest an\nalternative approach for future research which focuses on reflexivity, rather\nthan compositionality, as a target for explaining how simple communication\nsystems may become more like natural language. I end by providing some\nreference to the language origins literature that may be of some use to\nresearchers in machine learning.', 'Linguistic steganography studies how to hide secret messages in natural\nlanguage cover texts. Traditional methods aim to transform a secret message\ninto an innocent text via lexical substitution or syntactical modification.\nRecently, advances in neural language models (LMs) enable us to directly\ngenerate cover text conditioned on the secret message. In this study, we\npresent a new linguistic steganography method which encodes secret messages\nusing self-adjusting arithmetic coding based on a neural language model. We\nformally analyze the statistical imperceptibility of this method and\nempirically show it outperforms the previous state-of-the-art methods on four\ndatasets by 15.3% and 38.9% in terms of bits/word and KL metrics, respectively.\nFinally, human evaluations show that 51% of generated cover texts can indeed\nfool eavesdroppers.', 'Speech rhythms have been dealt with in three main ways: from the\nintrospective analyses of rhythm as a correlate of syllable and foot timing in\nlinguistics and applied linguistics, through analyses of durations of segments\nof utterances associated with consonantal and vocalic properties, syllables,\nfeet and words, to models of rhythms in speech production and perception as\nphysical oscillations. The present study avoids introspection and\nhuman-filtered annotation methods and extends the signal processing paradigm of\namplitude envelope spectrum analysis by adding an additional analytic step of\nedge detection, and postulating the co-existence of multiple speech rhythms in\nrhythm zones marked by identifiable edges (Rhythm Zone Theory, RZT). An\nexploratory investigation of the utility of RZT is conducted, suggesting that\nnative and non-native readings of the same text are distinct sub-genres of read\nspeech: a reading by a US native speaker and non-native readings by relatively\nlow-performing Cantonese adult learners of English. The study concludes by\nnoting that with the methods used, RZT can distinguish between the speech\nrhythms of well-defined sub-genres of native speaker reading vs. non-native\nlearner reading, but needs further refinement in order to be applied to the\nparadoxically more complex speech of low-performing language learners, whose\nspeech rhythms are co-determined by non-fluency and disfluency factors in\naddition to well-known linguistic factors of grammar, vocabulary and discourse\nconstraints.', 'Due to their inherent capability in semantic alignment of aspects and their\ncontext words, attention mechanism and Convolutional Neural Networks (CNNs) are\nwidely applied for aspect-based sentiment classification. However, these models\nlack a mechanism to account for relevant syntactical constraints and long-range\nword dependencies, and hence may mistakenly recognize syntactically irrelevant\ncontextual words as clues for judging aspect sentiment. To tackle this problem,\nwe propose to build a Graph Convolutional Network (GCN) over the dependency\ntree of a sentence to exploit syntactical information and word dependencies.\nBased on it, a novel aspect-specific sentiment classification framework is\nraised. Experiments on three benchmarking collections illustrate that our\nproposed model has comparable effectiveness to a range of state-of-the-art\nmodels, and further demonstrate that both syntactical information and\nlong-range word dependencies are properly captured by the graph convolution\nstructure.', 'Interaction between caregivers and children plays a critical role in human\nlanguage acquisition and development. Given this observation, it is remarkable\nthat explicit interaction plays little to no role in artificial language\nmodeling -- which also targets the acquisition of human language, yet by\nartificial models. Moreover, an interactive approach to language modeling has\nthe potential to make language models substantially more versatile and to\nconsiderably impact downstream applications. Motivated by these considerations,\nwe pioneer the space of interactive language modeling. As a first contribution\nwe present a road map in which we detail the steps that need to be taken\ntowards interactive language modeling. We then lead by example and take the\nfirst steps on this road map, showing the initial feasibility of our approach.\nAs such, this work aims to be the start of a larger research agenda on\ninteractive language modeling.', 'Drawing causal conclusions from observational real-world data is a very much\ndesired but challenging task. In this paper we present mixed-method analyses to\ninvestigate causal influences of publication trends and behavior on the\nadoption, persistence, and retirement of certain research foci --\nmethodologies, materials, and tasks that are of interest to the computational\nlinguistics (CL) community. Our key findings highlight evidence of the\ntransition to rapidly emerging methodologies in the research community (e.g.,\nadoption of bidirectional LSTMs influencing the retirement of LSTMs), the\npersistent engagement with trending tasks and techniques (e.g., deep learning,\nembeddings, generative, and language models), the effect of scientist location\nfrom outside the US, e.g., China on propensity of researching languages beyond\nEnglish, and the potential impact of funding for large-scale research programs.\nWe anticipate this work to provide useful insights about publication trends and\nbehavior and raise the awareness about the potential for causal inference in\nthe computational linguistics and a broader scientific community.', ""The question of how an effective and efficient communication system can\nemerge in a population of agents that need to solve a particular task attracts\nmore and more attention from researchers in many fields, including artificial\nintelligence, linguistics and statistical physics. A common methodology for\nstudying this question consists of carrying out multi-agent experiments in\nwhich a population of agents takes part in a series of scripted and\ntask-oriented communicative interactions, called 'language games'. While each\nindividual language game is typically played by two agents in the population, a\nlarge series of games allows the population to converge on a shared\ncommunication system. Setting up an experiment in which a rich system for\ncommunicating about the real world emerges is a major enterprise, as it\nrequires a variety of software components for running multi-agent experiments,\nfor interacting with sensors and actuators, for conceptualising and\ninterpreting semantic structures, and for mapping between these semantic\nstructures and linguistic utterances. The aim of this paper is twofold. On the\none hand, it introduces a high-level robot interface that extends the Babel\nsoftware system, presenting for the first time a toolkit that provides flexible\nmodules for dealing with each subtask involved in running advanced grounded\nlanguage game experiments. On the other hand, it provides a practical guide to\nusing the toolkit for implementing such experiments, taking a grounded colour\nnaming game experiment as a didactic example."", 'Matrix syntax is a formal model of syntactic relations in language. The\npurpose of this paper is to explain its mathematical foundations, for an\naudience with some formal background. We make an axiomatic presentation,\nmotivating each axiom on linguistic and practical grounds. The resulting\nmathematical structure resembles some aspects of quantum mechanics. Matrix\nsyntax allows us to describe a number of language phenomena that are otherwise\nvery difficult to explain, such as linguistic chains, and is arguably a more\neconomical theory of language than most of the theories proposed in the context\nof the minimalist program in linguistics. In particular, sentences are\nnaturally modelled as vectors in a Hilbert space with a tensor product\nstructure, built from 2x2 matrices belonging to some specific group.', 'The development of intelligent machines is one of the biggest unsolved\nchallenges in computer science. In this paper, we propose some fundamental\nproperties these machines should have, focusing in particular on communication\nand learning. We discuss a simple environment that could be used to\nincrementally teach a machine the basics of natural-language-based\ncommunication, as a prerequisite to more complex interaction with human users.\nWe also present some conjectures on the sort of algorithms the machine should\nsupport in order to profitably learn from the environment.', 'This entry introduces the topic of machine learning and provides an overview\nof its relevance for applied linguistics and language learning. The discussion\nwill focus on giving an introduction to the methods and applications of machine\nlearning in applied linguistics, and will provide references for further study.', 'Sorani Kurdish, also known as Central Kurdish, has a complex morphology,\nparticularly due to the patterns in which morphemes appear. Although several\naspects of Kurdish morphology have been studied, such as pronominal endoclitics\nand Izafa constructions, Sorani Kurdish morphology has received trivial\nattention in computational linguistics. Moreover, some morphemes, such as the\nemphasis endoclitic =\\^i\\c{s}, and derivational morphemes have not been\npreviously studied. To tackle the complex morphology of Sorani, we provide a\nthorough description of Sorani Kurdish morphological and morphophonological\nconstructions in a formal way such that they can be used as finite-state\ntransducers for morphological analysis and synthesis.', 'Although Natural Language Processing (NLP) is at the core of many tools young\npeople use in their everyday life, high school curricula (in Italy) do not\ninclude any computational linguistics education. This lack of exposure makes\nthe use of such tools less responsible than it could be and makes choosing\ncomputational linguistics as a university degree unlikely. To raise awareness,\ncuriosity, and longer-term interest in young people, we have developed an\ninteractive workshop designed to illustrate the basic principles of NLP and\ncomputational linguistics to high school Italian students aged between 13 and\n18 years. The workshop takes the form of a game in which participants play the\nrole of machines needing to solve some of the most common problems a computer\nfaces in understanding language: from voice recognition to Markov chains to\nsyntactic parsing. Participants are guided through the workshop with the help\nof instructors, who present the activities and explain core concepts from\ncomputational linguistics. The workshop was presented at numerous outlets in\nItaly between 2019 and 2021, both face-to-face and online.', 'Word reordering is one of the most difficult aspects of statistical machine\ntranslation (SMT), and an important factor of its quality and efficiency.\nDespite the vast amount of research published to date, the interest of the\ncommunity in this problem has not decreased, and no single method appears to be\nstrongly dominant across language pairs. Instead, the choice of the optimal\napproach for a new translation task still seems to be mostly driven by\nempirical trials. To orientate the reader in this vast and complex research\narea, we present a comprehensive survey of word reordering viewed as a\nstatistical modeling challenge and as a natural language phenomenon. The survey\ndescribes in detail how word reordering is modeled within different\nstring-based and tree-based SMT frameworks and as a stand-alone task, including\nsystematic overviews of the literature in advanced reordering modeling. We then\nquestion why some approaches are more successful than others in different\nlanguage pairs. We argue that, besides measuring the amount of reordering, it\nis important to understand which kinds of reordering occur in a given language\npair. To this end, we conduct a qualitative analysis of word reordering\nphenomena in a diverse sample of language pairs, based on a large collection of\nlinguistic knowledge. Empirical results in the SMT literature are shown to\nsupport the hypothesis that a few linguistic facts can be very useful to\nanticipate the reordering characteristics of a language pair and to select the\nSMT framework that best suits them.', 'Neural network-based approaches have become the driven forces for Natural\nLanguage Processing (NLP) tasks. Conventionally, there are two mainstream\nneural architectures for NLP tasks: the recurrent neural network (RNN) and the\nconvolution neural network (ConvNet). RNNs are good at modeling long-term\ndependencies over input texts, but preclude parallel computation. ConvNets do\nnot have memory capability and it has to model sequential data as un-ordered\nfeatures. Therefore, ConvNets fail to learn sequential dependencies over the\ninput texts, but it is able to carry out high-efficient parallel computation.\nAs each neural architecture, such as RNN and ConvNets, has its own pro and con,\nintegration of different architectures is assumed to be able to enrich the\nsemantic representation of texts, thus enhance the performance of NLP tasks.\nHowever, few investigation explores the reconciliation of these seemingly\nincompatible architectures. To address this issue, we propose a hybrid\narchitecture based on a novel hierarchical multi-granularity attention\nmechanism, named Multi-granularity Attention-based Hybrid Neural Network\n(MahNN). The attention mechanism is to assign different weights to different\nparts of the input sequence to increase the computation efficiency and\nperformance of neural models. In MahNN, two types of attentions are introduced:\nthe syntactical attention and the semantical attention. The syntactical\nattention computes the importance of the syntactic elements (such as words or\nsentence) at the lower symbolic level and the semantical attention is used to\ncompute the importance of the embedded space dimension corresponding to the\nupper latent semantics. We adopt the text classification as an exemplifying way\nto illustrate the ability of MahNN to understand texts.', 'Natural language is one of the most fundamental features that distinguish\npeople from other living things and enable people to communicate each other.\nLanguage is a tool that enables people to express their feelings and thoughts\nand to transfers cultures through generations. Texts and audio are examples of\nnatural language in daily life. In the natural language, many words disappear\nin time, on the other hand new words are derived. Therefore, while the process\nof natural language processing (NLP) is complex even for human, it is difficult\nto process in computer system. The area of linguistics examines how people use\nlanguage. NLP, which requires the collaboration of linguists and computer\nscientists, plays an important role in human computer interaction. Studies in\nNLP have increased with the use of artificial intelligence technologies in the\nfield of linguistics. With the deep learning methods which are one of the\nartificial intelligence study areas, platforms close to natural language are\nbeing developed. Developed platforms for language comprehension, machine\ntranslation and part of speech (POS) tagging benefit from deep learning\nmethods. Recurrent Neural Network (RNN), one of the deep learning\narchitectures, is preferred for processing sequential data such as text or\naudio data. In this study, Turkish POS tagging model has been proposed by using\nBidirectional Long-Short Term Memory (BLSTM) which is an RNN type. The proposed\nPOS tagging model is provided to natural language researchers with a platform\nthat allows them to perform and use their own analysis. In the development\nphase of the platform developed by using BLSTM, the error rate of the POS\ntagger has been reduced by taking feedback with expert opinion.', 'The goal of this project is to (i) accumulate annotated informal/formal\nmathematical corpora suitable for training semi-automated translation between\ninformal and formal mathematics by statistical machine-translation methods,\n(ii) to develop such methods oriented at the formalization task, and in\nparticular (iii) to combine such methods with learning-assisted automated\nreasoning that will serve as a strong semantic component. We describe these\nideas, the initial set of corpora, and some initial experiments done over them.', ""Virtual Personal Assistants like Siri have great potential but such\ndevelopments hit the fundamental problem of how to make computational devices\nthat understand human speech. Natural language understanding is one of the more\ndisappointing failures of AI research and it seems there is something we\ncomputer scientists don't get about the nature of language. Of course\nphilosophers and linguists think quite differently about language and this\npaper describes how we have taken ideas from other disciplines and implemented\nthem. The background to the work is to take seriously the notion of language as\naction and look at what people actually do with language using the techniques\nof Conversation Analysis. The observation has been that human communication is\n(behind the scenes) about the management of social relations as well as the\n(foregrounded) passing of information. To claim this is one thing but to\nimplement it requires a mechanism. The mechanism described here is based on the\nnotion of language being intentional - we think intentionally, talk about them\nand recognise them in others - and cooperative in that we are compelled to help\nout. The way we are compelled points to a solution to the ever present problem\nof keeping the human on topic. The approach has led to a recent success in\nwhich we significantly improve user satisfaction independent of task\ncompletion. Talk Markup Language (TalkML) is a draft alternative to VoiceXML\nthat, we propose, greatly simplifies the scripting of interaction by providing\ndefault behaviours for no input and not recognised speech events."", 'Recent years have seen a number of proposals for performing Natural Language\nGeneration (NLG) based in large part on statistical techniques. Despite having\nmany attractive features, we argue that these existing approaches nonetheless\nhave some important drawbacks, sometimes because the approach in question is\nnot fully statistical (i.e., relies on a certain amount of handcrafting),\nsometimes because the approach in question lacks transparency. Focussing on\nsome of the key NLG tasks (namely Content Selection, Lexical Choice, and\nLinguistic Realisation), we propose a novel approach, called the Text\nReassembling approach to NLG (TRG), which approaches the ideal of a purely\nstatistical approach very closely, and which is at the same time highly\ntransparent. We evaluate the TRG approach and discuss how TRG may be extended\nto deal with other NLG tasks, such as Document Structuring, and Aggregation. We\ndiscuss the strengths and limitations of TRG, concluding that the method may\nhold particular promise for domain experts who want to build an NLG system\ndespite having little expertise in linguistics and NLG.', ""With the explosive growth of online social media, the ancient problem of\ninformation disorders interfering with news diffusion has surfaced with a\nrenewed intensity threatening our democracies, public health, and news outlets'\ncredibility. Therefore, thousands of scientific papers have been published in a\nrelatively short period, making researchers of different disciplines struggle\nwith an information overload problem. The aim of this survey is threefold: (1)\nwe present the results of a network-based analysis of the existing\nmultidisciplinary literature to support the search for relevant trends and\ncentral publications; (2) we describe the main results and necessary background\nto attack the problem under a computational perspective; (3) we review selected\ncontributions using network science as a unifying framework and computational\nlinguistics as the tool to make sense of the shared content. Despite scholars\nworking on computational linguistics and networks traditionally belong to\ndifferent scientific communities, we expect that those interested in the area\nof fake news should be aware of crucial aspects of both disciplines."", 'In natural language processing, most models try to learn semantic\nrepresentations merely from texts. The learned representations encode the\ndistributional semantics but fail to connect to any knowledge about the\nphysical world. In contrast, humans learn language by grounding concepts in\nperception and action and the brain encodes grounded semantics for cognition.\nInspired by this notion and recent work in vision-language learning, we design\na two-stream model for grounding language learning in vision. The model\nincludes a VGG-based visual stream and a Bert-based language stream. The two\nstreams merge into a joint representational space. Through cross-modal\ncontrastive learning, the model first learns to align visual and language\nrepresentations with the MS COCO dataset. The model further learns to retrieve\nvisual objects with language queries through a cross-modal attention module and\nto infer the visual relations between the retrieved objects through a bilinear\noperator with the Visual Genome dataset. After training, the language stream of\nthis model is a stand-alone language model capable of embedding concepts in a\nvisually grounded semantic space. This semantic space manifests principal\ndimensions explainable with human intuition and neurobiological knowledge. Word\nembeddings in this semantic space are predictive of human-defined norms of\nsemantic features and are segregated into perceptually distinctive clusters.\nFurthermore, the visually grounded language model also enables compositional\nlanguage understanding based on visual knowledge and multimodal image search\nwith queries based on images, texts, or their combinations.', 'In vision and linguistics; the main input modalities are facial expressions,\nspeech patterns, and the words uttered. The issue with analysis of any one mode\nof expression (Visual, Verbal or Vocal) is that lot of contextual information\ncan get lost. This asks researchers to inspect multiple modalities to get a\nthorough understanding of the cross-modal dependencies and temporal context of\nthe situation to analyze the expression. This work attempts at preserving the\nlong-range dependencies within and across different modalities, which would be\nbottle-necked by the use of recurrent networks and adds the concept of\ndelta-attention to focus on local differences per modality to capture the\nidiosyncrasy of different people. We explore a cross-attention fusion technique\nto get the global view of the emotion expressed through these\ndelta-self-attended modalities, in order to fuse all the local nuances and\nglobal context together. The addition of attention is new to the multi-modal\nfusion field and currently being scrutinized for on what stage the attention\nmechanism should be used, this work achieves competitive accuracy for overall\nand per-class classification which is close to the current state-of-the-art\nwith almost half number of parameters.', ""This paper outlines the results of sentence level linguistics based rules for\nimproving part-of-speech tagging. It is well known that the performance of\ncomplex NLP systems is negatively affected if one of the preliminary stages is\nless than perfect. Errors in the initial stages in the pipeline have a\nsnowballing effect on the pipeline's end performance. We have created a set of\nlinguistics based rules at the sentence level which adjust part-of-speech tags\nfrom state-of-the-art taggers. Comparison with state-of-the-art taggers on\nwidely used benchmarks demonstrate significant improvements in tagging accuracy\nand consequently in the quality and accuracy of NLP systems."", 'Linguistics holds unique characteristics of generality, stability, and\nnationality, which will affect the formulation of extraction strategies and\nshould be incorporated into the relation extraction. Chinese open relation\nextraction is not well-established, because of the complexity of Chinese\nlinguistics makes it harder to operate, and the methods for English are not\ncompatible with that for Chinese. The diversities between Chinese and English\nlinguistics are mainly reflected in morphology and syntax.', 'Ideas from forensic linguistics are now being used frequently in Natural\nLanguage Processing (NLP), using machine learning techniques. While the role of\nforensic linguistics was more benign earlier, it is now being used for purposes\nwhich are questionable. Certain methods from forensic linguistics are employed,\nwithout considering their scientific limitations and ethical concerns. While we\ntake the specific case of forensic linguistics as an example of such trends in\nNLP and machine learning, the issue is a larger one and present in many other\nscientific and data-driven domains. We suggest that such trends indicate that\nsome of the applied sciences are exceeding their legal and scientific briefs.\nWe highlight how carelessly implemented practices are serving to short-circuit\nthe due processes of law as well breach ethical codes.', 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', 'Traditional learning-based coreference resolvers operate by training the\nmention-pair model for determining whether two mentions are coreferent or not.\nThough conceptually simple and easy to understand, the mention-pair model is\nlinguistically rather unappealing and lags far behind the heuristic-based\ncoreference models proposed in the pre-statistical NLP era in terms of\nsophistication. Two independent lines of recent research have attempted to\nimprove the mention-pair model, one by acquiring the mention-ranking model to\nrank preceding mentions for a given anaphor, and the other by training the\nentity-mention model to determine whether a preceding cluster is coreferent\nwith a given mention. We propose a cluster-ranking approach to coreference\nresolution, which combines the strengths of the mention-ranking model and the\nentity-mention model, and is therefore theoretically more appealing than both\nof these models. In addition, we seek to improve cluster rankers via two\nextensions: (1) lexicalization and (2) incorporating knowledge of anaphoricity\nby jointly modeling anaphoricity determination and coreference resolution.\nExperimental results on the ACE data sets demonstrate the superior performance\nof cluster rankers to competing approaches as well as the effectiveness of our\ntwo extensions.', 'Neural sequence-to-sequence text-to-speech synthesis (TTS) can produce\nhigh-quality speech directly from text or simple linguistic features such as\nphonemes. Unlike traditional pipeline TTS, the neural sequence-to-sequence TTS\ndoes not require manually annotated and complicated linguistic features such as\npart-of-speech tags and syntactic structures for system training. However, it\nmust be carefully designed and well optimized so that it can implicitly extract\nuseful linguistic features from the input features. In this paper we\ninvestigate under what conditions the neural sequence-to-sequence TTS can work\nwell in Japanese and English along with comparisons with deep neural network\n(DNN) based pipeline TTS systems. Unlike past comparative studies, the pipeline\nsystems also use autoregressive probabilistic modeling and a neural vocoder. We\ninvestigated systems from three aspects: a) model architecture, b) model\nparameter size, and c) language. For the model architecture aspect, we adopt\nmodified Tacotron systems that we previously proposed and their variants using\nan encoder from Tacotron or Tacotron2. For the model parameter size aspect, we\ninvestigate two model parameter sizes. For the language aspect, we conduct\nlistening tests in both Japanese and English to see if our findings can be\ngeneralized across languages. Our experiments suggest that a) a neural\nsequence-to-sequence TTS system should have a sufficient number of model\nparameters to produce high quality speech, b) it should also use a powerful\nencoder when it takes characters as inputs, and c) the encoder still has a room\nfor improvement and needs to have an improved architecture to learn\nsupra-segmental features more appropriately.', ""We investigate neural techniques for end-to-end computational argumentation\nmining (AM). We frame AM both as a token-based dependency parsing and as a\ntoken-based sequence tagging problem, including a multi-task learning setup.\nContrary to models that operate on the argument component level, we find that\nframing AM as dependency parsing leads to subpar performance results. In\ncontrast, less complex (local) tagging models based on BiLSTMs perform robustly\nacross classification scenarios, being able to catch long-range dependencies\ninherent to the AM problem. Moreover, we find that jointly learning 'natural'\nsubtasks, in a multi-task learning setup, improves performance."", 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently developing an application that can efficiently search for images relevant to given text descriptions by integrating both visual and textual data. Although current state-of-the-art pre-trained models are effective, they are limited by slow inference speeds and substantial computational costs. These limitations are primarily due to the attention mechanisms in the Transformer architecture, making them impractical for real-world applications where low latency and computational efficiency are crucial. Consequently, I am seeking a solution that can enhance inference speeds without sacrificing model accuracy. One potential approach is to represent texts and images in a structured, easily searchable embedding format. This format should remain a vectorized embedding but could incorporate certain structures that facilitate faster search and ranking. All these processes should be conducted offline, and during online inference, only highly efficient cosine similarity matching should be performed from their embeddings. Ultimately, my goal is for this solution to outperform existing pre-trained models in both speed and accuracy across various image and text retrieval benchmarks.","[""Conventional approaches to image-text retrieval mainly focus on indexing\nvisual objects appearing in pictures but ignore the interactions between these\nobjects. Such objects occurrences and interactions are equivalently useful and\nimportant in this field as they are usually mentioned in the text. Scene graph\npresentation is a suitable method for the image-text matching challenge and\nobtained good results due to its ability to capture the inter-relationship\ninformation. Both images and text are represented in scene graph levels and\nformulate the retrieval challenge as a scene graph matching challenge. In this\npaper, we introduce the Local and Global Scene Graph Matching (LGSGM) model\nthat enhances the state-of-the-art method by integrating an extra graph\nconvolution network to capture the general information of a graph.\nSpecifically, for a pair of scene graphs of an image and its caption, two\nseparate models are used to learn the features of each graph's nodes and edges.\nThen a Siamese-structure graph convolution model is employed to embed graphs\ninto vector forms. We finally combine the graph-level and the vector-level to\ncalculate the similarity of this image-text pair. The empirical experiments\nshow that our enhancement with the combination of levels can improve the\nperformance of the baseline method by increasing the recall by more than 10% on\nthe Flickr30k dataset."", 'Text-to-image retrieval is an essential task in cross-modal information\nretrieval, i.e., retrieving relevant images from a large and unlabelled dataset\ngiven textual queries. In this paper, we propose VisualSparta, a novel\n(Visual-text Sparse Transformer Matching) model that shows significant\nimprovement in terms of both accuracy and efficiency. VisualSparta is capable\nof outperforming previous state-of-the-art scalable methods in MSCOCO and\nFlickr30K. We also show that it achieves substantial retrieving speed\nadvantages, i.e., for a 1 million image index, VisualSparta using CPU gets\n~391X speedup compared to CPU vector search and ~5.4X speedup compared to\nvector search with GPU acceleration. Experiments show that this speed advantage\neven gets bigger for larger datasets because VisualSparta can be efficiently\nimplemented as an inverted index. To the best of our knowledge, VisualSparta is\nthe first transformer-based text-to-image retrieval model that can achieve\nreal-time searching for large-scale datasets, with significant accuracy\nimprovement compared to previous state-of-the-art methods.', 'At Pinterest, we utilize image embeddings throughout our search and\nrecommendation systems to help our users navigate through visual content by\npowering experiences like browsing of related content and searching for exact\nproducts for shopping. In this work we describe a multi-task deep metric\nlearning system to learn a single unified image embedding which can be used to\npower our multiple visual search products. The solution we present not only\nallows us to train for multiple application objectives in a single deep neural\nnetwork architecture, but takes advantage of correlated information in the\ncombination of all training data from each application to generate a unified\nembedding that outperforms all specialized embeddings previously deployed for\neach product. We discuss the challenges of handling images from different\ndomains such as camera photos, high quality web images, and clean product\ncatalog images. We also detail how to jointly train for multiple product\nobjectives and how to leverage both engagement data and human labeled data. In\naddition, our trained embeddings can also be binarized for efficient storage\nand retrieval without compromising precision and recall. Through comprehensive\nevaluations on offline metrics, user studies, and online A/B experiments, we\ndemonstrate that our proposed unified embedding improves both relevance and\nengagement of our visual search products for both browsing and searching\npurposes when compared to existing specialized embeddings. Finally, the\ndeployment of the unified embedding at Pinterest has drastically reduced the\noperation and engineering cost of maintaining multiple embeddings while\nimproving quality.', 'Inspired by recent advances in multimodal learning and machine translation,\nwe introduce an encoder-decoder pipeline that learns (a): a multimodal joint\nembedding space with images and text and (b): a novel language model for\ndecoding distributed representations from our space. Our pipeline effectively\nunifies joint image-text embedding models with multimodal neural language\nmodels. We introduce the structure-content neural language model that\ndisentangles the structure of a sentence to its content, conditioned on\nrepresentations produced by the encoder. The encoder allows one to rank images\nand sentences while the decoder can generate novel descriptions from scratch.\nUsing LSTM to encode sentences, we match the state-of-the-art performance on\nFlickr8K and Flickr30K without using object detections. We also set new best\nresults when using the 19-layer Oxford convolutional network. Furthermore we\nshow that with linear encoders, the learned embedding space captures multimodal\nregularities in terms of vector space arithmetic e.g. *image of a blue car* -\n""blue"" + ""red"" is near images of red cars. Sample captions generated for 800\nimages are made available for comparison.', 'The ability to describe images with natural language sentences is the\nhallmark for image and language understanding. Such a system has wide ranging\napplications such as annotating images and using natural sentences to search\nfor images.In this project we focus on the task of bidirectional image\nretrieval: such asystem is capable of retrieving an image based on a sentence\n(image search) andretrieve sentence based on an image query (image annotation).\nWe present asystem based on a global ranking objective function which uses a\ncombinationof convolutional neural networks (CNN) and multi layer perceptrons\n(MLP).It takes a pair of image and sentence and processes them in different\nchannels,finally embedding it into a common multimodal vector space. These\nembeddingsencode abstract semantic information about the two inputs and can be\ncomparedusing traditional information retrieval approaches. For each such pair,\nthe modelreturns a score which is interpretted as a similarity metric. If this\nscore is high,the image and sentence are likely to convey similar meaning, and\nif the score is low then they are likely not to.\n  The visual input is modeled via deep convolutional neural network. On\ntheother hand we explore three models for the textual module. The first one\nisbag of words with an MLP. The second one uses n-grams (bigram, trigrams,and a\ncombination of trigram & skip-grams) with an MLP. The third is morespecialized\ndeep network specific for modeling variable length sequences (SSE).We report\ncomparable performance to recent work in the field, even though ouroverall\nmodel is simpler. We also show that the training time choice of how wecan\ngenerate our negative samples has a significant impact on performance, and can\nbe used to specialize the bi-directional system in one particular task.', 'Current state-of-the-art approaches to cross-modal retrieval process text and\nvisual input jointly, relying on Transformer-based architectures with\ncross-attention mechanisms that attend over all words and objects in an image.\nWhile offering unmatched retrieval performance, such models: 1) are typically\npretrained from scratch and thus less scalable, 2) suffer from huge retrieval\nlatency and inefficiency issues, which makes them impractical in realistic\napplications. To address these crucial gaps towards both improved and efficient\ncross-modal retrieval, we propose a novel fine-tuning framework that turns any\npretrained text-image multi-modal model into an efficient retrieval model. The\nframework is based on a cooperative retrieve-and-rerank approach which\ncombines: 1) twin networks (i.e., a bi-encoder) to separately encode all items\nof a corpus, enabling efficient initial retrieval, and 2) a cross-encoder\ncomponent for a more nuanced (i.e., smarter) ranking of the retrieved small set\nof items. We also propose to jointly fine-tune the two components with shared\nweights, yielding a more parameter-efficient model. Our experiments on a series\nof standard cross-modal retrieval benchmarks in monolingual, multilingual, and\nzero-shot setups, demonstrate improved accuracy and huge efficiency benefits\nover the state-of-the-art cross-encoders.', 'In the task of near similar image search, features from Deep Neural Network\nis often used to compare images and measure similarity. In the past, we only\nfocused visual search in image dataset without text data. However, since deep\nneural network emerged, the performance of visual search becomes high enough to\napply it in many industries from 3D data to multimodal data. Compared to the\nneeds of multimodal search, there has not been sufficient researches.\n  In this paper, we present a method of near similar search with image and text\nmultimodal dataset. Earlier time, similar image search, especially when\nsearching shopping items, treated image and text separately to search similar\nitems and reorder the results. This regards two tasks of image search and text\nmatching as two different tasks. Our method, however, explore the vast data to\ncompute k-nearest neighbors using both image and text.\n  In our experiment of similar item search, our system using multimodal data\nshows better performance than single data while it only increases minute\ncomputing time. For the experiment, we collected more than 15 million of\naccessory and six million of digital product items from online shopping\nwebsites, in which the product item comprises item images, titles, categories,\nand descriptions. Then we compare the performance of multimodal searching to\nsingle space searching in these datasets.', 'Image-text matching is an interesting and fascinating task in modern AI\nresearch. Despite the evolution of deep-learning-based image and text\nprocessing systems, multi-modal matching remains a challenging problem. In this\nwork, we consider the problem of accurate image-text matching for the task of\nmulti-modal large-scale information retrieval. State-of-the-art results in\nimage-text matching are achieved by inter-playing image and text features from\nthe two different processing pipelines, usually using mutual attention\nmechanisms. However, this invalidates any chance to extract separate visual and\ntextual features needed for later indexing steps in large-scale retrieval\nsystems. In this regard, we introduce the Transformer Encoder Reasoning Network\n(TERN), an architecture built upon one of the modern relationship-aware\nself-attentive architectures, the Transformer Encoder (TE). This architecture\nis able to separately reason on the two different modalities and to enforce a\nfinal common abstract concept space by sharing the weights of the deeper\ntransformer layers. Thanks to this design, the implemented network is able to\nproduce compact and very rich visual and textual features available for the\nsuccessive indexing step. Experiments are conducted on the MS-COCO dataset, and\nwe evaluate the results using a discounted cumulative gain metric with\nrelevance computed exploiting caption similarities, in order to assess possibly\nnon-exact but relevant search results. We demonstrate that on this metric we\nare able to achieve state-of-the-art results in the image retrieval task. Our\ncode is freely available at https://github.com/mesnico/TERN.', 'Cross-model retrieval has emerged as one of the most important upgrades for\ntext-only search engines (SE). Recently, with powerful representation for\npairwise text-image inputs via early interaction, the accuracy of\nvision-language (VL) transformers has outperformed existing methods for\ntext-image retrieval. However, when the same paradigm is used for inference,\nthe efficiency of the VL transformers is still too low to be applied in a real\ncross-modal SE. Inspired by the mechanism of human learning and using\ncross-modal knowledge, this paper presents a novel Vision-Language Decomposed\nTransformer (VLDeformer), which greatly increases the efficiency of VL\ntransformers while maintaining their outstanding accuracy. By the proposed\nmethod, the cross-model retrieval is separated into two stages: the VL\ntransformer learning stage, and the VL decomposition stage. The latter stage\nplays the role of single modal indexing, which is to some extent like the term\nindexing of a text SE. The model learns cross-modal knowledge from\nearly-interaction pre-training and is then decomposed into an individual\nencoder. The decomposition requires only small target datasets for supervision\nand achieves both $1000+$ times acceleration and less than $0.6$\\% average\nrecall drop. VLDeformer also outperforms state-of-the-art visual-semantic\nembedding methods on COCO and Flickr30k.', 'Joint image-text embedding is the bedrock for most Vision-and-Language (V+L)\ntasks, where multimodality inputs are simultaneously processed for joint visual\nand textual understanding. In this paper, we introduce UNITER, a UNiversal\nImage-TExt Representation, learned through large-scale pre-training over four\nimage-text datasets (COCO, Visual Genome, Conceptual Captions, and SBU\nCaptions), which can power heterogeneous downstream V+L tasks with joint\nmultimodal embeddings. We design four pre-training tasks: Masked Language\nModeling (MLM), Masked Region Modeling (MRM, with three variants), Image-Text\nMatching (ITM), and Word-Region Alignment (WRA). Different from previous work\nthat applies joint random masking to both modalities, we use conditional\nmasking on pre-training tasks (i.e., masked language/region modeling is\nconditioned on full observation of image/text). In addition to ITM for global\nimage-text alignment, we also propose WRA via the use of Optimal Transport (OT)\nto explicitly encourage fine-grained alignment between words and image regions\nduring pre-training. Comprehensive analysis shows that both conditional masking\nand OT-based WRA contribute to better pre-training. We also conduct a thorough\nablation study to find an optimal combination of pre-training tasks. Extensive\nexperiments show that UNITER achieves new state of the art across six V+L tasks\n(over nine datasets), including Visual Question Answering, Image-Text\nRetrieval, Referring Expression Comprehension, Visual Commonsense Reasoning,\nVisual Entailment, and NLVR$^2$. Code is available at\nhttps://github.com/ChenRocks/UNITER.', 'Our objective is language-based search of large-scale image and video\ndatasets. For this task, the approach that consists of independently mapping\ntext and vision to a joint embedding space, a.k.a. dual encoders, is attractive\nas retrieval scales and is efficient for billions of images using approximate\nnearest neighbour search. An alternative approach of using vision-text\ntransformers with cross-attention gives considerable improvements in accuracy\nover the joint embeddings, but is often inapplicable in practice for\nlarge-scale retrieval given the cost of the cross-attention mechanisms required\nfor each sample at test time. This work combines the best of both worlds. We\nmake the following three contributions. First, we equip transformer-based\nmodels with a new fine-grained cross-attention architecture, providing\nsignificant improvements in retrieval accuracy whilst preserving scalability.\nSecond, we introduce a generic approach for combining a Fast dual encoder model\nwith our Slow but accurate transformer-based model via distillation and\nre-ranking. Finally, we validate our approach on the Flickr30K image dataset\nwhere we show an increase in inference speed by several orders of magnitude\nwhile having results competitive to the state of the art. We also extend our\nmethod to the video domain, improving the state of the art on the VATEX\ndataset.', 'In this paper we propose to learn a multimodal image and text embedding from\nWeb and Social Media data, aiming to leverage the semantic knowledge learnt in\nthe text domain and transfer it to a visual model for semantic image retrieval.\nWe demonstrate that the pipeline can learn from images with associated text\nwithout supervision and perform a thourough analysis of five different text\nembeddings in three different benchmarks. We show that the embeddings learnt\nwith Web and Social Media data have competitive performances over supervised\nmethods in the text based image retrieval task, and we clearly outperform state\nof the art in the MIRFlickr dataset when training in the target data. Further\nwe demonstrate how semantic multimodal image retrieval can be performed using\nthe learnt embeddings, going beyond classical instance-level retrieval\nproblems. Finally, we present a new dataset, InstaCities1M, composed by\nInstagram images and their associated texts that can be used for fair\ncomparison of image-text embeddings.', 'Self-Supervised learning from multimodal image and text data allows deep\nneural networks to learn powerful features with no need of human annotated\ndata. Web and Social Media platforms provide a virtually unlimited amount of\nthis multimodal data. In this work we propose to exploit this free available\ndata to learn a multimodal image and text embedding, aiming to leverage the\nsemantic knowledge learnt in the text domain and transfer it to a visual model\nfor semantic image retrieval. We demonstrate that the proposed pipeline can\nlearn from images with associated textwithout supervision and analyze the\nsemantic structure of the learnt joint image and text embedding space. We\nperform a thorough analysis and performance comparison of five different state\nof the art text embeddings in three different benchmarks. We show that the\nembeddings learnt with Web and Social Media data have competitive performances\nover supervised methods in the text based image retrieval task, and we clearly\noutperform state of the art in the MIRFlickr dataset when training in the\ntarget data. Further, we demonstrate how semantic multimodal image retrieval\ncan be performed using the learnt embeddings, going beyond classical\ninstance-level retrieval problems. Finally, we present a new dataset,\nInstaCities1M, composed by Instagram images and their associated texts that can\nbe used for fair comparison of image-text embeddings.', 'We propose Pixel-BERT to align image pixels with text by deep multi-modal\ntransformers that jointly learn visual and language embedding in a unified\nend-to-end framework. We aim to build a more accurate and thorough connection\nbetween image pixels and language semantics directly from image and sentence\npairs instead of using region-based image features as the most recent vision\nand language tasks. Our Pixel-BERT which aligns semantic connection in pixel\nand text level solves the limitation of task-specific visual representation for\nvision and language tasks. It also relieves the cost of bounding box\nannotations and overcomes the unbalance between semantic labels in visual task\nand language semantic. To provide a better representation for down-stream\ntasks, we pre-train a universal end-to-end model with image and sentence pairs\nfrom Visual Genome dataset and MS-COCO dataset. We propose to use a random\npixel sampling mechanism to enhance the robustness of visual representation and\nto apply the Masked Language Model and Image-Text Matching as pre-training\ntasks. Extensive experiments on downstream tasks with our pre-trained model\nshow that our approach makes the most state-of-the-arts in downstream tasks,\nincluding Visual Question Answering (VQA), image-text retrieval, Natural\nLanguage for Visual Reasoning for Real (NLVR). Particularly, we boost the\nperformance of a single model in VQA task by 2.17 points compared with SOTA\nunder fair comparison.', 'A major challenge in matching images and text is that they have intrinsically\ndifferent data distributions and feature representations. Most existing\napproaches are based either on embedding or classification, the first one\nmapping image and text instances into a common embedding space for distance\nmeasuring, and the second one regarding image-text matching as a binary\nclassification problem. Neither of these approaches can, however, balance the\nmatching accuracy and model complexity well. We propose a novel framework that\nachieves remarkable matching performance with acceptable model complexity.\nSpecifically, in the training stage, we propose a novel Multi-modal Tensor\nFusion Network (MTFN) to explicitly learn an accurate image-text similarity\nfunction with rank-based tensor fusion rather than seeking a common embedding\nspace for each image-text instance. Then, during testing, we deploy a generic\nCross-modal Re-ranking (RR) scheme for refinement without requiring additional\ntraining procedure. Extensive experiments on two datasets demonstrate that our\nMTFN-RR consistently achieves the state-of-the-art matching performance with\nmuch less time complexity. The implementation code is available at\nhttps://github.com/Wangt-CN/MTFN-RR-PyTorch-Code.', 'Textual-visual cross-modal retrieval has been a hot research topic in both\ncomputer vision and natural language processing communities. Learning\nappropriate representations for multi-modal data is crucial for the cross-modal\nretrieval performance. Unlike existing image-text retrieval approaches that\nembed image-text pairs as single feature vectors in a common representational\nspace, we propose to incorporate generative processes into the cross-modal\nfeature embedding, through which we are able to learn not only the global\nabstract features but also the local grounded features. Extensive experiments\nshow that our framework can well match images and sentences with complex\ncontent, and achieve the state-of-the-art cross-modal retrieval results on\nMSCOCO dataset.', 'In this work, we propose an approach to index Deep Convolutional Neural\nNetwork Features to support efficient content-based retrieval on large image\ndatabases. To this aim, we have converted the these features into a textual\nform, to index them into an inverted index by means of Lucene. In this way, we\nwere able to set up a robust retrieval system that combines full-text search\nwith content-based image retrieval capabilities. We evaluated different\nstrategies of textual representation in order to optimize the index occupation\nand the query response time. In order to show that our approach is able to\nhandle large datasets, we have developed a web-based prototype that provides an\ninterface for combined textual and visual searching into a dataset of about 100\nmillion of images.', 'In this paper, we describe our end-to-end content-based image retrieval\nsystem built upon Elasticsearch, a well-known and popular textual search\nengine. As far as we know, this is the first time such a system has been\nimplemented in eCommerce, and our efforts have turned out to be highly\nworthwhile. We end up with a novel and exciting visual search solution that is\nextremely easy to be deployed, distributed, scaled and monitored in a\ncost-friendly manner. Moreover, our platform is intrinsically flexible in\nsupporting multimodal searches, where visual and textual information can be\njointly leveraged in retrieval.\n  The core idea is to encode image feature vectors into a collection of string\ntokens in a way such that closer vectors will share more string tokens in\ncommon. By doing that, we can utilize Elasticsearch to efficiently retrieve\nsimilar images based on similarities within encoded sting tokens. As part of\nthe development, we propose a novel vector to string encoding method, which is\nshown to substantially outperform the previous ones in terms of both precision\nand latency.\n  First-hand experiences in implementing this Elasticsearch-based platform are\nextensively addressed, which should be valuable to practitioners also\ninterested in building visual search engine on top of Elasticsearch.', 'This paper considers the task of matching images and sentences by learning a\nvisual-textual embedding space for cross-modal retrieval. Finding such a space\nis a challenging task since the features and representations of text and image\nare not comparable. In this work, we introduce an end-to-end deep multimodal\nconvolutional-recurrent network for learning both vision and language\nrepresentations simultaneously to infer image-text similarity. The model learns\nwhich pairs are a match (positive) and which ones are a mismatch (negative)\nusing a hinge-based triplet ranking. To learn about the joint representations,\nwe leverage our newly extracted collection of tweets from Twitter. The main\ncharacteristic of our dataset is that the images and tweets are not\nstandardized the same as the benchmarks. Furthermore, there can be a higher\nsemantic correlation between the pictures and tweets contrary to benchmarks in\nwhich the descriptions are well-organized. Experimental results on MS-COCO\nbenchmark dataset show that our model outperforms certain methods presented\npreviously and has competitive performance compared to the state-of-the-art.\nThe code and dataset have been made available publicly.', 'Generative adversarial networks conditioned on textual image descriptions are\ncapable of generating realistic-looking images. However, current methods still\nstruggle to generate images based on complex image captions from a\nheterogeneous domain. Furthermore, quantitatively evaluating these\ntext-to-image models is challenging, as most evaluation metrics only judge\nimage quality but not the conformity between the image and its caption. To\naddress these challenges we introduce a new model that explicitly models\nindividual objects within an image and a new evaluation metric called Semantic\nObject Accuracy (SOA) that specifically evaluates images given an image\ncaption. The SOA uses a pre-trained object detector to evaluate if a generated\nimage contains objects that are mentioned in the image caption, e.g. whether an\nimage generated from ""a car driving down the street"" contains a car. We perform\na user study comparing several text-to-image models and show that our SOA\nmetric ranks the models the same way as humans, whereas other metrics such as\nthe Inception Score do not. Our evaluation also shows that models which\nexplicitly model objects outperform models which only model global image\ncharacteristics.', 'This paper investigates the problem of modeling Internet images and\nassociated text or tags for tasks such as image-to-image search, tag-to-image\nsearch, and image-to-tag search (image annotation). We start with canonical\ncorrelation analysis (CCA), a popular and successful approach for mapping\nvisual and textual features to the same latent space, and incorporate a third\nview capturing high-level image semantics, represented either by a single\ncategory or multiple non-mutually-exclusive concepts. We present two ways to\ntrain the three-view embedding: supervised, with the third view coming from\nground-truth labels or search keywords; and unsupervised, with semantic themes\nautomatically obtained by clustering the tags. To ensure high accuracy for\nretrieval tasks while keeping the learning process scalable, we combine\nmultiple strong visual features and use explicit nonlinear kernel mappings to\nefficiently approximate kernel CCA. To perform retrieval, we use a specially\ndesigned similarity function in the embedded space, which substantially\noutperforms the Euclidean distance. The resulting system produces compelling\nqualitative results and outperforms a number of two-view baselines on retrieval\ntasks on three large-scale Internet image datasets.', 'The use of deep learning models within scientific experimental facilities\nfrequently requires low-latency inference, so that, for example, quality\ncontrol operations can be performed while data are being collected. Edge\ncomputing devices can be useful in this context, as their low cost and compact\nform factor permit them to be co-located with the experimental apparatus. Can\nsuch devices, with their limited resources, can perform neural network\nfeed-forward computations efficiently and effectively? We explore this question\nby evaluating the performance and accuracy of a scientific image restoration\nmodel, for which both model input and output are images, on edge computing\ndevices. Specifically, we evaluate deployments of TomoGAN, an image-denoising\nmodel based on generative adversarial networks developed for low-dose x-ray\nimaging, on the Google Edge TPU and NVIDIA Jetson. We adapt TomoGAN for edge\nexecution, evaluate model inference performance, and propose methods to address\nthe accuracy drop caused by model quantization. We show that these edge\ncomputing devices can deliver accuracy comparable to that of a full-fledged CPU\nor GPU model, at speeds that are more than adequate for use in the intended\ndeployments, denoising a 1024 x 1024 image in less than a second. Our\nexperiments also show that the Edge TPU models can provide 3x faster inference\nresponse than a CPU-based model and 1.5x faster than an edge GPU-based model.\nThis combination of high speed and low cost permits image restoration anywhere.', 'Text-visual (or called semantic-visual) embedding is a central problem in\nvision-language research. It typically involves mapping of an image and a text\ndescription to a common feature space through a CNN image encoder and a RNN\nlanguage encoder. In this paper, we propose a new method for learning\ntext-visual embedding using both image titles and click-through data from an\nimage search engine. We also propose a new triplet loss function by modeling\npositive awareness of the embedding, and introduce a novel mini-batch-based\nhard negative sampling approach for better data efficiency in the learning\nprocess. Experimental results show that our proposed method outperforms\nexisting methods, and is also effective for real-world text-to-visual\nretrieval.', 'Zero-shot learning (ZSL) models rely on learning a joint embedding space\nwhere both textual/semantic description of object classes and visual\nrepresentation of object images can be projected to for nearest neighbour\nsearch. Despite the success of deep neural networks that learn an end-to-end\nmodel between text and images in other vision problems such as image\ncaptioning, very few deep ZSL model exists and they show little advantage over\nZSL models that utilise deep feature representations but do not learn an\nend-to-end embedding. In this paper we argue that the key to make deep ZSL\nmodels succeed is to choose the right embedding space. Instead of embedding\ninto a semantic space or an intermediate space, we propose to use the visual\nspace as the embedding space. This is because that in this space, the\nsubsequent nearest neighbour search would suffer much less from the hubness\nproblem and thus become more effective. This model design also provides a\nnatural mechanism for multiple semantic modalities (e.g., attributes and\nsentence descriptions) to be fused and optimised jointly in an end-to-end\nmanner. Extensive experiments on four benchmarks show that our model\nsignificantly outperforms the existing models. Code is available at\nhttps://github.com/lzrobots/DeepEmbeddingModel_ZSL', 'Multilingual (or cross-lingual) embeddings represent several languages in a\nunique vector space. Using a common embedding space enables for a shared\nsemantic between words from different languages. In this paper, we propose to\nembed images and texts into a unique distributional vector space, enabling to\nsearch images by using text queries expressing information needs related to the\n(visual) content of images, as well as using image similarity. Our framework\nforces the representation of an image to be similar to the representation of\nthe text that describes it. Moreover, by using multilingual embeddings we\nensure that words from two different languages have close descriptors and thus\nare attached to similar images. We provide experimental evidence of the\nefficiency of our approach by experimenting it on two datasets: Common Objects\nin COntext (COCO) [19] and Multi30K [7].', 'Researchers often spend weeks sifting through decades of unlabeled satellite\nimagery(on NASA Worldview) in order to develop datasets on which they can start\nconducting research. We developed an interactive, scalable and fast image\nsimilarity search engine (which can take one or more images as the query image)\nthat automatically sifts through the unlabeled dataset reducing dataset\ngeneration time from weeks to minutes. In this work, we describe key components\nof the end to end pipeline. Our similarity search system was created to be able\nto identify similar images from a potentially petabyte scale database that are\nsimilar to an input image, and for this we had to break down each query image\ninto its features, which were generated by a classification layer stripped CNN\ntrained in a supervised manner. To store and search these features efficiently,\nwe had to make several scalability improvements. To improve the speed, reduce\nthe storage, and shrink memory requirements for embedding search, we add a\nfully connected layer to our CNN make all images into a 128 length vector\nbefore entering the classification layers. This helped us compress the size of\nour image features from 2048 (for ResNet, which was initially tried as our\nfeaturizer) to 128 for our new custom model. Additionally, we utilize existing\napproximate nearest neighbor search libraries to significantly speed up\nembedding search. Our system currently searches over our entire database of\nimages at 5 seconds per query on a single virtual machine in the cloud. In the\nfuture, we would like to incorporate a SimCLR based featurizing model which\ncould be trained without any labelling by a human (since the classification\naspect of the model is irrelevant to this use case).']","['Deep Learning is increasingly being adopted by industry for computer vision\napplications running on embedded devices. While Convolutional Neural Networks\'\naccuracy has achieved a mature and remarkable state, inference latency and\nthroughput are a major concern especially when targeting low-cost and low-power\nembedded platforms. CNNs\' inference latency may become a bottleneck for Deep\nLearning adoption by industry, as it is a crucial specification for many\nreal-time processes. Furthermore, deployment of CNNs across heterogeneous\nplatforms presents major compatibility issues due to vendor-specific technology\nand acceleration libraries. In this work, we present QS-DNN, a fully automatic\nsearch based on Reinforcement Learning which, combined with an inference engine\noptimizer, efficiently explores through the design space and empirically finds\nthe optimal combinations of libraries and primitives to speed up the inference\nof CNNs on heterogeneous embedded devices. We show that, an optimized\ncombination can achieve 45x speedup in inference latency on CPU compared to a\ndependency-free baseline and 2x on average on GPGPU compared to the best vendor\nlibrary. Further, we demonstrate that, the quality of results and time\n""to-solution"" is much better than with Random Search and achieves up to 15x\nbetter results for a short-time search.', 'Recently, Multi-modal Named Entity Recognition (MNER) has attracted a lot of\nattention. Most of the work utilizes image information through region-level\nvisual representations obtained from a pretrained object detector and relies on\nan attention mechanism to model the interactions between image and text\nrepresentations. However, it is difficult to model such interactions as image\nand text representations are trained separately on the data of their respective\nmodality and are not aligned in the same space. As text representations take\nthe most important role in MNER, in this paper, we propose {\\bf I}mage-{\\bf\nt}ext {\\bf A}lignments (ITA) to align image features into the textual space, so\nthat the attention mechanism in transformer-based pretrained textual embeddings\ncan be better utilized. ITA first aligns the image into regional object tags,\nimage-level captions and optical characters as visual contexts, concatenates\nthem with the input texts as a new cross-modal input, and then feeds it into a\npretrained textual embedding model. This makes it easier for the attention\nmodule of a pretrained textual embedding model to model the interaction between\nthe two modalities since they are both represented in the textual space. ITA\nfurther aligns the output distributions predicted from the cross-modal input\nand textual input views so that the MNER model can be more practical in dealing\nwith text-only inputs and robust to noises from images. In our experiments, we\nshow that ITA models can achieve state-of-the-art accuracy on multi-modal Named\nEntity Recognition datasets, even without image information.', 'Text-based person search is a sub-task in the field of image retrieval, which\naims to retrieve target person images according to a given textual description.\nThe significant feature gap between two modalities makes this task very\nchallenging. Many existing methods attempt to utilize local alignment to\naddress this problem in the fine-grained level. However, most relevant methods\nintroduce additional models or complicated training and evaluation strategies,\nwhich are hard to use in realistic scenarios. In order to facilitate the\npractical application, we propose a simple but effective end-to-end learning\nframework for text-based person search named TIPCB (i.e., Text-Image Part-based\nConvolutional Baseline). Firstly, a novel dual-path local alignment network\nstructure is proposed to extract visual and textual local representations, in\nwhich images are segmented horizontally and texts are aligned adaptively. Then,\nwe propose a multi-stage cross-modal matching strategy, which eliminates the\nmodality gap from three feature levels, including low level, local level and\nglobal level. Extensive experiments are conducted on the widely-used benchmark\ndataset (CUHK-PEDES) and verify that our method outperforms the\nstate-of-the-art methods by 3.69%, 2.95% and 2.31% in terms of Top-1, Top-5 and\nTop-10. Our code has been released in https://github.com/OrangeYHChen/TIPCB.', 'In this paper, we study the task of image retrieval, where the input query is\nspecified in the form of an image plus some text that describes desired\nmodifications to the input image. For example, we may present an image of the\nEiffel tower, and ask the system to find images which are visually similar but\nare modified in small ways, such as being taken at nighttime instead of during\nthe day. To tackle this task, we learn a similarity metric between a target\nimage and a source image plus source text, an embedding and composing function\nsuch that target image feature is close to the source image plus text\ncomposition feature. We propose a new way to combine image and text using such\nfunction that is designed for the retrieval task. We show this outperforms\nexisting approaches on 3 different datasets, namely Fashion-200k, MIT-States\nand a new synthetic dataset we create based on CLEVR. We also show that our\napproach can be used to classify input queries, in addition to image retrieval.', 'With the prolification of multimodal interaction in various domains, recently\nthere has been much interest in text based image retrieval in the computer\nvision community. However most of the state of the art techniques model this\nproblem in a purely neural way, which makes it difficult to incorporate\npragmatic strategies in searching a large scale catalog especially when the\nsearch requirements are insufficient and the model needs to resort to an\ninteractive retrieval process through multiple iterations of\nquestion-answering. Motivated by this, we propose a neural-symbolic approach\nfor a one-shot retrieval of images from a large scale catalog, given the\ncaption description. To facilitate this, we represent the catalog and caption\nas scene-graphs and model the retrieval task as a learnable graph matching\nproblem, trained end-to-end with a REINFORCE algorithm. Further, we briefly\ndescribe an extension of this pipeline to an iterative retrieval framework,\nbased on interactive questioning and answering.', ""Current Visual Question Answering (VQA) systems can answer intelligent\nquestions about `Known' visual content. However, their performance drops\nsignificantly when questions about visually and linguistically `Unknown'\nconcepts are presented during inference (`Open-world' scenario). A practical\nVQA system should be able to deal with novel concepts in real world settings.\nTo address this problem, we propose an exemplar-based approach that transfers\nlearning (i.e., knowledge) from previously `Known' concepts to answer questions\nabout the `Unknown'. We learn a highly discriminative joint embedding space,\nwhere visual and semantic features are fused to give a unified representation.\nOnce novel concepts are presented to the model, it looks for the closest match\nfrom an exemplar set in the joint embedding space. This auxiliary information\nis used alongside the given Image-Question pair to refine visual attention in a\nhierarchical fashion. Since handling the high dimensional exemplars on large\ndatasets can be a significant challenge, we introduce an efficient matching\nscheme that uses a compact feature description for search and retrieval. To\nevaluate our model, we propose a new split for VQA, separating Unknown visual\nand semantic concepts from the training set. Our approach shows significant\nimprovements over state-of-the-art VQA models on the proposed Open-World VQA\ndataset and standard VQA datasets."", 'Visual-semantic embedding enables various tasks such as image-text retrieval,\nimage captioning, and visual question answering. The key to successful\nvisual-semantic embedding is to express visual and textual data properly by\naccounting for their intricate relationship. While previous studies have\nachieved much advance by encoding the visual and textual data into a joint\nspace where similar concepts are closely located, they often represent data by\na single vector ignoring the presence of multiple important components in an\nimage or text. Thus, in addition to the joint embedding space, we propose a\nnovel multi-head self-attention network to capture various components of visual\nand textual data by attending to important parts in data. Our approach achieves\nthe new state-of-the-art results in image-text retrieval tasks on MS-COCO and\nFlicker30K datasets. Through the visualization of the attention maps that\ncapture distinct semantic components at multiple positions in the image and the\ntext, we demonstrate that our method achieves an effective and interpretable\nvisual-semantic joint space.', ""The convolutional neural network (CNN) features can give a good description\nof image content, which usually represent images with unique global vectors.\nAlthough they are compact compared to local descriptors, they still cannot\nefficiently deal with large-scale image retrieval due to the cost of the linear\nincremental computation and storage. To address this issue, we build a simple\nbut effective indexing framework based on inverted table, which significantly\ndecreases both the search time and memory usage. In addition, several\nstrategies are fully investigated under an indexing framework to adapt it to\nCNN features and compensate for quantization errors. First, we use multiple\nassignment for the query and database images to increase the probability of\nrelevant images' co-existing in the same Voronoi cells obtained via the\nclustering algorithm. Then, we introduce embedding codes to further improve\nprecision by removing false matches during a search. We demonstrate that by\nusing hashing schemes to calculate the embedding codes and by changing the\nranking rule, indexing framework speeds can be greatly improved. Extensive\nexperiments conducted on several unsupervised and supervised benchmarks support\nthese results and the superiority of the proposed indexing framework. We also\nprovide a fair comparison between the popular CNN features."", ""The inclusion of semantic information in any similarity measures improves the\nefficiency of the similarity measure and provides human interpretable results\nfor further analysis. The similarity calculation method that focuses on\nfeatures related to the text's words only, will give less accurate results.\nThis paper presents three different methods that not only focus on the text's\nwords but also incorporates semantic information of texts in their feature\nvector and computes semantic similarities. These methods are based on\ncorpus-based and knowledge-based methods, which are: cosine similarity using\ntf-idf vectors, cosine similarity using word embedding and soft cosine\nsimilarity using word embedding. Among these three, cosine similarity using\ntf-idf vectors performed best in finding similarities between short news texts.\nThe similar texts given by the method are easy to interpret and can be used\ndirectly in other information retrieval applications."", 'Our objective is video retrieval based on natural language queries. In\naddition, we consider the analogous problem of retrieving sentences or\ngenerating descriptions given an input video. Recent work has addressed the\nproblem by embedding visual and textual inputs into a common space where\nsemantic similarities correlate to distances. We also adopt the embedding\napproach, and make the following contributions: First, we utilize web image\nsearch in sentence embedding process to disambiguate fine-grained visual\nconcepts. Second, we propose embedding models for sentence, image, and video\ninputs whose parameters are learned simultaneously. Finally, we show how the\nproposed model can be applied to description generation. Overall, we observe a\nclear improvement over the state-of-the-art methods in the video and sentence\nretrieval tasks. In description generation, the performance level is comparable\nto the current state-of-the-art, although our embeddings were trained for the\nretrieval tasks.', ""This paper describes Designovel's systems which are submitted to the Fashion\nIQ Challenge 2019. Goal of the challenge is building an image retrieval system\nwhere input query is a candidate image plus two text phrases describe user's\nfeedback about visual differences between the candidate image and the search\ntarget. We built the systems by combining methods from recent work on deep\nmetric learning, multi-modal retrieval and natual language processing. First,\nwe encode both candidate and target images with CNNs into high-level\nrepresentations, and encode text descriptions to a single text vector using\nTransformer-based encoder. Then we compose candidate image vector and text\nrepresentation into a single vector which is exptected to be biased toward\ntarget image vector. Finally, we compute cosine similarities between composed\nvector and encoded vectors of whole dataset, and rank them in desceding order\nto get ranked list. We experimented with Fashion IQ 2019 dataset in various\nsettings of hyperparameters, achieved 39.12% average recall by a single model\nand 43.67% average recall by an ensemble of 16 models on test dataset."", 'Text in natural images contains rich semantics that are often highly relevant\nto objects or scene. In this paper, we focus on the problem of fully exploiting\nscene text for visual understanding. The main idea is combining word\nrepresentations and deep visual features into a globally trainable deep\nconvolutional neural network. First, the recognized words are obtained by a\nscene text reading system. Then, we combine the word embedding of the\nrecognized words and the deep visual features into a single representation,\nwhich is optimized by a convolutional neural network for fine-grained image\nclassification. In our framework, the attention mechanism is adopted to reveal\nthe relevance between each recognized word and the given image, which further\nenhances the recognition performance. We have performed experiments on two\ndatasets: Con-Text dataset and Drink Bottle dataset, that are proposed for\nfine-grained classification of business places and drink bottles, respectively.\nThe experimental results consistently demonstrate that the proposed method\ncombining textual and visual cues significantly outperforms classification with\nonly visual representations. Moreover, we have shown that the learned\nrepresentation improves the retrieval performance on the drink bottle images by\na large margin, making it potentially useful in product search.', 'Conventional methods for the image-text generation tasks mainly tackle the\nnaturally bidirectional generation tasks separately, focusing on designing\ntask-specific frameworks to improve the quality and fidelity of the generated\nsamples. Recently, Vision-Language Pre-training models have greatly improved\nthe performance of the image-to-text generation tasks, but large-scale\npre-training models for text-to-image synthesis task are still under-developed.\nIn this paper, we propose ERNIE-ViLG, a unified generative pre-training\nframework for bidirectional image-text generation with transformer model. Based\non the image quantization models, we formulate both image generation and text\ngeneration as autoregressive generative tasks conditioned on the text/image\ninput. The bidirectional image-text generative modeling eases the semantic\nalignments across vision and language. For the text-to-image generation\nprocess, we further propose an end-to-end training method to jointly learn the\nvisual sequence generator and the image reconstructor. To explore the landscape\nof large-scale pre-training for bidirectional text-image generation, we train a\n10-billion parameter ERNIE-ViLG model on a large-scale dataset of 145 million\n(Chinese) image-text pairs which achieves state-of-the-art performance for both\ntext-to-image and image-to-text tasks, obtaining an FID of 7.9 on MS-COCO for\ntext-to-image synthesis and best results on COCO-CN and AIC-ICC for image\ncaptioning.', ""Internet search affects people's cognition of the world, so mitigating biases\nin search results and learning fair models is imperative for social good. We\nstudy a unique gender bias in image search in this work: the search images are\noften gender-imbalanced for gender-neutral natural language queries. We\ndiagnose two typical image search models, the specialized model trained on\nin-domain datasets and the generalized representation model pre-trained on\nmassive image and text data across the internet. Both models suffer from severe\ngender bias. Therefore, we introduce two novel debiasing approaches: an\nin-processing fair sampling method to address the gender imbalance issue for\ntraining models, and a post-processing feature clipping method base on mutual\ninformation to debias multimodal representations of pre-trained models.\nExtensive experiments on MS-COCO and Flickr30K benchmarks show that our methods\nsignificantly reduce the gender bias in image search models."", 'Achieving good speed and accuracy trade-off on a target platform is very\nimportant in deploying deep neural networks in real world scenarios. However,\nmost existing automatic architecture search approaches only concentrate on high\nperformance. In this work, we propose an algorithm that can offer better\nspeed/accuracy trade-off of searched networks, which is termed ""Partial Order\nPruning"". It prunes the architecture search space with a partial order\nassumption to automatically search for the architectures with the best speed\nand accuracy trade-off. Our algorithm explicitly takes profile information\nabout the inference speed on the target platform into consideration. With the\nproposed algorithm, we present several Dongfeng (DF) networks that provide high\naccuracy and fast inference speed on various application GPU platforms. By\nfurther searching decoder architectures, our DF-Seg real-time segmentation\nnetworks yield state-of-the-art speed/accuracy trade-off on both the target\nembedded device and the high-end GPU.', 'Representation learning of textual networks poses a significant challenge as\nit involves capturing amalgamated information from two modalities: (i)\nunderlying network structure, and (ii) node textual attributes. For this, most\nexisting approaches learn embeddings of text and network structure by enforcing\nembeddings of connected nodes to be similar. Then for achieving a modality\nfusion they use the similarities between text embedding of a node with the\nstructure embedding of its connected node and vice versa. This implies that\nthese approaches require edge information for learning embeddings and they\ncannot learn embeddings of unseen nodes. In this paper we propose an approach\nthat achieves both modality fusion and the capability to learn embeddings of\nunseen nodes. The main feature of our model is that it uses an adversarial\nmechanism between text embedding based discriminator, and structure embedding\nbased generator to learn efficient representations. Then for learning\nembeddings of unseen nodes, we use the supervision provided by the text\nembedding based discriminator. In addition this, we propose a novel\narchitecture for learning text embedding that can combine both mutual attention\nand topological attention mechanism, which give more flexible text embeddings.\nThrough extensive experiments on real-world datasets, we demonstrate that our\nmodel makes substantial gains over several state-of-the-art benchmarks. In\ncomparison with previous state-of-the-art, it gives up to 7% improvement in\nperformance in predicting links among nodes seen in the training and up to 12%\nimprovement in performance in predicting links involving nodes not seen in\ntraining. Further, in the node classification task, it gives up to 2%\nimprovement in performance.', 'We propose a cross-modal attention distillation framework to train a\ndual-encoder model for vision-language understanding tasks, such as visual\nreasoning and visual question answering. Dual-encoder models have a faster\ninference speed than fusion-encoder models and enable the pre-computation of\nimages and text during inference. However, the shallow interaction module used\nin dual-encoder models is insufficient to handle complex vision-language\nunderstanding tasks. In order to learn deep interactions of images and text, we\nintroduce cross-modal attention distillation, which uses the image-to-text and\ntext-to-image attention distributions of a fusion-encoder model to guide the\ntraining of our dual-encoder model. In addition, we show that applying the\ncross-modal attention distillation for both pre-training and fine-tuning stages\nachieves further improvements. Experimental results demonstrate that the\ndistilled dual-encoder model achieves competitive performance for visual\nreasoning, visual entailment and visual question answering tasks while enjoying\na much faster inference speed than fusion-encoder models. Our code and models\nwill be publicly available at https://github.com/kugwzk/Distilled-DualEncoder.', 'Text-image cross-modal retrieval is a challenging task in the field of\nlanguage and vision. Most previous approaches independently embed images and\nsentences into a joint embedding space and compare their similarities. However,\nprevious approaches rarely explore the interactions between images and\nsentences before calculating similarities in the joint space. Intuitively, when\nmatching between images and sentences, human beings would alternatively attend\nto regions in images and words in sentences, and select the most salient\ninformation considering the interaction between both modalities. In this paper,\nwe propose Cross-modal Adaptive Message Passing (CAMP), which adaptively\ncontrols the information flow for message passing across modalities. Our\napproach not only takes comprehensive and fine-grained cross-modal interactions\ninto account, but also properly handles negative pairs and irrelevant\ninformation with an adaptive gating scheme. Moreover, instead of conventional\njoint embedding approaches for text-image matching, we infer the matching score\nbased on the fused features, and propose a hardest negative binary\ncross-entropy loss for training. Results on COCO and Flickr30k significantly\nsurpass state-of-the-art methods, demonstrating the effectiveness of our\napproach.', ""Contextual embeddings derived from transformer-based neural language models\nhave shown state-of-the-art performance for various tasks such as question\nanswering, sentiment analysis, and textual similarity in recent years.\nExtensive work shows how accurately such models can represent abstract,\nsemantic information present in text. In this expository work, we explore a\ntangent direction and analyze such models' performance on tasks that require a\nmore granular level of representation. We focus on the problem of textual\nsimilarity from two perspectives: matching documents on a granular level\n(requiring embeddings to capture fine-grained attributes in the text), and an\nabstract level (requiring embeddings to capture overall textual semantics). We\nempirically demonstrate, across two datasets from different domains, that\ndespite high performance in abstract document matching as expected, contextual\nembeddings are consistently (and at times, vastly) outperformed by simple\nbaselines like TF-IDF for more granular tasks. We then propose a simple but\neffective method to incorporate TF-IDF into models that use contextual\nembeddings, achieving relative improvements of up to 36% on granular tasks."", 'Large-scale pretraining of visual representations has led to state-of-the-art\nperformance on a range of benchmark computer vision tasks, yet the benefits of\nthese techniques at extreme scale in complex production systems has been\nrelatively unexplored. We consider the case of a popular visual discovery\nproduct, where these representations are trained with multi-task learning, from\nuse-case specific visual understanding (e.g. skin tone classification) to\ngeneral representation learning for all visual content (e.g. embeddings for\nretrieval). In this work, we describe how we (1) generate a dataset with over a\nbillion images via large weakly-supervised pretraining to improve the\nperformance of these visual representations, and (2) leverage Transformers to\nreplace the traditional convolutional backbone, with insights into both system\nand performance improvements, especially at 1B+ image scale. To support this\nbackbone model, we detail a systematic approach to deriving weakly-supervised\nimage annotations from heterogenous text signals, demonstrating the benefits of\nclustering techniques to handle the long-tail distribution of image labels.\nThrough a comprehensive study of offline and online evaluation, we show that\nlarge-scale Transformer-based pretraining provides significant benefits to\nindustry computer vision applications. The model is deployed in a production\nvisual shopping system, with 36% improvement in top-1 relevance and 23%\nimprovement in click-through volume. We conduct extensive experiments to better\nunderstand the empirical relationships between Transformer-based architectures,\ndataset scale, and the performance of production vision systems.', 'We tackle the problem of visual search under resource constraints. Existing\nsystems use the same embedding model to compute representations (embeddings)\nfor the query and gallery images. Such systems inherently face a hard\naccuracy-efficiency trade-off: the embedding model needs to be large enough to\nensure high accuracy, yet small enough to enable query-embedding computation on\nresource-constrained platforms. This trade-off could be mitigated if gallery\nembeddings are generated from a large model and query embeddings are extracted\nusing a compact model. The key to building such a system is to ensure\nrepresentation compatibility between the query and gallery models. In this\npaper, we address two forms of compatibility: One enforced by modifying the\nparameters of each model that computes the embeddings. The other by modifying\nthe architectures that compute the embeddings, leading to compatibility-aware\nneural architecture search (CMP-NAS). We test CMP-NAS on challenging retrieval\ntasks for fashion images (DeepFashion2), and face images (IJB-C). Compared to\nordinary (homogeneous) visual search using the largest embedding model\n(paragon), CMP-NAS achieves 80-fold and 23-fold cost reduction while\nmaintaining accuracy within 0.3% and 1.6% of the paragon on DeepFashion2 and\nIJB-C respectively.', 'We present a multi-modal dialog system to assist online shoppers in visually\nbrowsing through large catalogs. Visual browsing is different from visual\nsearch in that it allows the user to explore the wide range of products in a\ncatalog, beyond the exact search matches. We focus on a slightly asymmetric\nversion of the complete multi-modal dialog where the system can understand both\ntext and image queries but responds only in images. We formulate our problem of\n""showing $k$ best images to a user"" based on the dialog context so far, as\nsampling from a Gaussian Mixture Model in a high dimensional joint multi-modal\nembedding space, that embed both the text and the image queries. Our system\nremembers the context of the dialog and uses an exploration-exploitation\nparadigm to assist in visual browsing. We train and evaluate the system on a\nmulti-modal dialog dataset that we generate from large catalog data. Our\nexperiments are promising and show that the agent is capable of learning and\ncan display relevant results with an average cosine similarity of 0.85 to the\nground truth. Our preliminary human evaluation also corroborates the fact that\nsuch a multi-modal dialog system for visual browsing is well-received and is\ncapable of engaging human users.', 'Point cloud architecture design has become a crucial problem for 3D deep\nlearning. Several efforts exist to manually design architectures with high\naccuracy in point cloud tasks such as classification, segmentation, and\ndetection. Recent progress in automatic Neural Architecture Search (NAS)\nminimizes the human effort in network design and optimizes high performing\narchitectures. However, these efforts fail to consider important factors such\nas latency during inference. Latency is of high importance in time critical\napplications like self-driving cars, robot navigation, and mobile applications,\nthat are generally bound by the available hardware. In this paper, we introduce\na new NAS framework, dubbed LC-NAS, where we search for point cloud\narchitectures that are constrained to a target latency. We implement a novel\nlatency constraint formulation to trade-off between accuracy and latency in our\narchitecture search. Contrary to previous works, our latency loss guarantees\nthat the final network achieves latency under a specified target value. This is\ncrucial when the end task is to be deployed in a limited hardware setting.\nExtensive experiments show that LC-NAS is able to find state-of-the-art\narchitectures for point cloud classification in ModelNet40 with minimal\ncomputational cost. We also show how our searched architectures achieve any\ndesired latency with a reasonably low drop in accuracy. Finally, we show how\nour searched architectures easily transfer to a different task, part\nsegmentation on PartNet, where we achieve state-of-the-art results while\nlowering latency by a factor of 10.', 'In this paper, we propose a single UniFied transfOrmer (UFO), which is\ncapable of processing either unimodal inputs (e.g., image or language) or\nmultimodal inputs (e.g., the concatenation of the image and the question), for\nvision-language (VL) representation learning. Existing approaches typically\ndesign an individual network for each modality and/or a specific fusion network\nfor multimodal tasks. To simplify the network architecture, we use a single\ntransformer network and enforce multi-task learning during VL pre-training,\nwhich includes the image-text contrastive loss, image-text matching loss, and\nmasked language modeling loss based on the bidirectional and the seq2seq\nattention mask. The same transformer network is used as the image encoder, the\ntext encoder, or the fusion network in different pre-training tasks.\nEmpirically, we observe less conflict among different tasks and achieve new\nstate of the arts on visual question answering, COCO image captioning\n(cross-entropy optimization) and nocaps (in SPICE). On other downstream tasks,\ne.g., image-text retrieval, we also achieve competitive performance.', 'In this paper, we study the problem of image-text matching. Inferring the\nlatent semantic alignment between objects or other salient stuff (e.g. snow,\nsky, lawn) and the corresponding words in sentences allows to capture\nfine-grained interplay between vision and language, and makes image-text\nmatching more interpretable. Prior work either simply aggregates the similarity\nof all possible pairs of regions and words without attending differentially to\nmore and less important words or regions, or uses a multi-step attentional\nprocess to capture limited number of semantic alignments which is less\ninterpretable. In this paper, we present Stacked Cross Attention to discover\nthe full latent alignments using both image regions and words in a sentence as\ncontext and infer image-text similarity. Our approach achieves the\nstate-of-the-art results on the MS-COCO and Flickr30K datasets. On Flickr30K,\nour approach outperforms the current best methods by 22.1% relatively in text\nretrieval from image query, and 18.2% relatively in image retrieval with text\nquery (based on Recall@1). On MS-COCO, our approach improves sentence retrieval\nby 17.8% relatively and image retrieval by 16.6% relatively (based on Recall@1\nusing the 5K test set). Code has been made available at:\nhttps://github.com/kuanghuei/SCAN.', 'Although transformers have become the neural architectures of choice for\nnatural language processing, they require orders of magnitude more training\ndata, GPU memory, and computations in order to compete with convolutional\nneural networks for computer vision. The attention mechanism of transformers\nscales quadratically with the length of the input sequence, and unrolled images\nhave long sequence lengths. Plus, transformers lack an inductive bias that is\nappropriate for images. We tested three modifications to vision transformer\n(ViT) architectures that address these shortcomings. Firstly, we alleviate the\nquadratic bottleneck by using linear attention mechanisms, called X-formers\n(such that, X in {Performer, Linformer, Nystr\\""omformer}), thereby creating\nVision X-formers (ViXs). This resulted in up to a seven times reduction in the\nGPU memory requirement. We also compared their performance with FNet and\nmulti-layer perceptron mixers, which further reduced the GPU memory\nrequirement. Secondly, we introduced an inductive bias for images by replacing\nthe initial linear embedding layer by convolutional layers in ViX, which\nsignificantly increased classification accuracy without increasing the model\nsize. Thirdly, we replaced the learnable 1D position embeddings in ViT with\nRotary Position Embedding (RoPE), which increases the classification accuracy\nfor the same model size. We believe that incorporating such changes can\ndemocratize transformers by making them accessible to those with limited data\nand computing resources.', 'The existing methods for image search reranking suffer from the\nunfaithfulness of the assumptions under which the text-based images search\nresult. The resulting images contain more irrelevant images. Hence the re\nranking concept arises to re rank the retrieved images based on the text around\nthe image and data of data of image and visual feature of image. A number of\nmethods are differentiated for this re-ranking. The high ranked images are used\nas noisy data and a k means algorithm for classification is learned to rectify\nthe ranking further. We are study the affect ability of the cross validation\nmethod to this training data. The pre eminent originality of the overall method\nis in collecting text/metadata of image and visual features in order to achieve\nan automatic ranking of the images. Supervision is initiated to learn the model\nweights offline, previous to reranking process. While model learning needs\nmanual labeling of the results for a some limited queries, the resulting model\nis query autonomous and therefore applicable to any other query .Examples are\ngiven for a selection of other classes like vehicles, animals and other\nclasses.', 'The state of the art on many NLP tasks is currently achieved by large\npre-trained language models, which require a considerable amount of\ncomputation. We explore a setting where many different predictions are made on\na single piece of text. In that case, some of the computational cost during\ninference can be amortized over the different tasks using a shared text\nencoder. We compare approaches for training such an encoder and show that\nencoders pre-trained over multiple tasks generalize well to unseen tasks. We\nalso compare ways of extracting fixed- and limited-size representations from\nthis encoder, including different ways of pooling features extracted from\nmultiple layers or positions. Our best approach compares favorably to knowledge\ndistillation, achieving higher accuracy and lower computational cost once the\nsystem is handling around 7 tasks. Further, we show that through binary\nquantization, we can reduce the size of the extracted representations by a\nfactor of 16 making it feasible to store them for later use. The resulting\nmethod offers a compelling solution for using large-scale pre-trained models at\na fraction of the computational cost when multiple tasks are performed on the\nsame text.', ""We study the problem of recognizing visual entities from the textual\ndescriptions of their classes. Specifically, given birds' images with free-text\ndescriptions of their species, we learn to classify images of previously-unseen\nspecies based on specie descriptions. This setup has been studied in the vision\ncommunity under the name zero-shot learning from text, focusing on learning to\ntransfer knowledge about visual aspects of birds from seen classes to\npreviously-unseen ones. Here, we suggest focusing on the textual description\nand distilling from the description the most relevant information to\neffectively match visual features to the parts of the text that discuss them.\nSpecifically, (1) we propose to leverage the similarity between species,\nreflected in the similarity between text descriptions of the species. (2) we\nderive visual summaries of the texts, i.e., extractive summaries that focus on\nthe visual features that tend to be reflected in images. We propose a simple\nattention-based model augmented with the similarity and visual summaries\ncomponents. Our empirical results consistently and significantly outperform the\nstate-of-the-art on the largest benchmarks for text-based zero-shot learning,\nillustrating the critical importance of texts for zero-shot image-recognition."", 'Cross-modal similarity search is a problem about designing a search system\nsupporting querying across content modalities, e.g., using an image to search\nfor texts or using a text to search for images. This paper presents a compact\ncoding solution for efficient search, with a focus on the quantization approach\nwhich has already shown the superior performance over the hashing solutions in\nthe single-modal similarity search. We propose a cross-modal quantization\napproach, which is among the early attempts to introduce quantization into\ncross-modal search. The major contribution lies in jointly learning the\nquantizers for both modalities through aligning the quantized representations\nfor each pair of image and text belonging to a document. In addition, our\napproach simultaneously learns the common space for both modalities in which\nquantization is conducted to enable efficient and effective search using the\nEuclidean distance computed in the common space with fast distance table\nlookup. Experimental results compared with several competitive algorithms over\nthree benchmark datasets demonstrate that the proposed approach achieves the\nstate-of-the-art performance.', 'The accuracy of deep convolutional neural networks (CNNs) generally improves\nwhen fueled with high resolution images. However, this often comes at a high\ncomputational cost and high memory footprint. Inspired by the fact that not all\nregions in an image are task-relevant, we propose a novel framework that\nperforms efficient image classification by processing a sequence of relatively\nsmall inputs, which are strategically selected from the original image with\nreinforcement learning. Such a dynamic decision process naturally facilitates\nadaptive inference at test time, i.e., it can be terminated once the model is\nsufficiently confident about its prediction and thus avoids further redundant\ncomputation. Notably, our framework is general and flexible as it is compatible\nwith most of the state-of-the-art light-weighted CNNs (such as MobileNets,\nEfficientNets and RegNets), which can be conveniently deployed as the backbone\nfeature extractor. Experiments on ImageNet show that our method consistently\nimproves the computational efficiency of a wide variety of deep models. For\nexample, it further reduces the average latency of the highly efficient\nMobileNet-V3 on an iPhone XS Max by 20% without sacrificing accuracy. Code and\npre-trained models are available at\nhttps://github.com/blackfeather-wang/GFNet-Pytorch.', ""Our objective in this work is video-text retrieval - in particular a joint\nembedding that enables efficient text-to-video retrieval. The challenges in\nthis area include the design of the visual architecture and the nature of the\ntraining data, in that the available large scale video-text training datasets,\nsuch as HowTo100M, are noisy and hence competitive performance is achieved only\nat scale through large amounts of compute. We address both these challenges in\nthis paper. We propose an end-to-end trainable model that is designed to take\nadvantage of both large-scale image and video captioning datasets. Our model is\nan adaptation and extension of the recent ViT and Timesformer architectures,\nand consists of attention in both space and time. The model is flexible and can\nbe trained on both image and video text datasets, either independently or in\nconjunction. It is trained with a curriculum learning schedule that begins by\ntreating images as 'frozen' snapshots of video, and then gradually learns to\nattend to increasing temporal context when trained on video datasets. We also\nprovide a new video-text pretraining dataset WebVid-2M, comprised of over two\nmillion videos with weak captions scraped from the internet. Despite training\non datasets that are an order of magnitude smaller, we show that this approach\nyields state-of-the-art results on standard downstream video-retrieval\nbenchmarks including MSR-VTT, MSVD, DiDeMo and LSMDC."", 'Search engine has become a fundamental component in various web and mobile\napplications. Retrieving relevant documents from the massive datasets is\nchallenging for a search engine system, especially when faced with verbose or\ntail queries. In this paper, we explore a vector space search framework for\ndocument retrieval. Specifically, we trained a deep semantic matching model so\nthat each query and document can be encoded as a low dimensional embedding. Our\nmodel was trained based on BERT architecture. We deployed a fast\nk-nearest-neighbor index service for online serving. Both offline and online\nmetrics demonstrate that our method improved retrieval performance and search\nquality considerably, particularly for tail', ""LiveSketch is a novel algorithm for searching large image collections using\nhand-sketched queries. LiveSketch tackles the inherent ambiguity of sketch\nsearch by creating visual suggestions that augment the query as it is drawn,\nmaking query specification an iterative rather than one-shot process that helps\ndisambiguate users' search intent. Our technical contributions are: a triplet\nconvnet architecture that incorporates an RNN based variational autoencoder to\nsearch for images using vector (stroke-based) queries; real-time clustering to\nidentify likely search intents (and so, targets within the search embedding);\nand the use of backpropagation from those targets to perturb the input stroke\nsequence, so suggesting alterations to the query in order to guide the search.\nWe show improvements in accuracy and time-to-task over contemporary baselines\nusing a 67M image corpus."", 'Text-based person search aims to retrieve images of a certain pedestrian by a\ntextual description. The key challenge of this task is to eliminate the\ninter-modality gap and achieve the feature alignment across modalities. In this\npaper, we propose a semantic-aligned embedding method for text-based person\nsearch, in which the feature alignment across modalities is achieved by\nautomatically learning the semantic-aligned visual features and textual\nfeatures. First, we introduce two Transformer-based backbones to encode robust\nfeature representations of the images and texts. Second, we design a\nsemantic-aligned feature aggregation network to adaptively select and aggregate\nfeatures with the same semantics into part-aware features, which is achieved by\na multi-head attention module constrained by a cross-modality part alignment\nloss and a diversity loss. Experimental results on the CUHK-PEDES and Flickr30K\ndatasets show that our method achieves state-of-the-art performances.', 'Open-domain extractive question answering works well on textual data by first\nretrieving candidate texts and then extracting the answer from those\ncandidates. However, some questions cannot be answered by text alone but\nrequire information stored in tables. In this paper, we present an approach for\nretrieving both texts and tables relevant to a question by jointly encoding\ntexts, tables and questions into a single vector space. To this end, we create\na new multi-modal dataset based on text and table datasets from related work\nand compare the retrieval performance of different encoding schemata. We find\nthat dense vector embeddings of transformer models outperform sparse embeddings\non four out of six evaluation datasets. Comparing different dense embedding\nmodels, tri-encoders with one encoder for each question, text and table,\nincrease retrieval performance compared to bi-encoders with one encoder for the\nquestion and one for both text and tables. We release the newly created\nmulti-modal dataset to the community so that it can be used for training and\nevaluation.', 'Predicting patient mortality is an important and challenging problem in the\nhealthcare domain, especially for intensive care unit (ICU) patients.\nElectronic health notes serve as a rich source for learning patient\nrepresentations, that can facilitate effective risk assessment. However, a\nlarge portion of clinical notes are unstructured and also contain domain\nspecific terminologies, from which we need to extract structured information.\nIn this paper, we introduce an embedding framework to learn\nsemantically-plausible distributed representations of clinical notes that\nexploits the semantic correspondence between the unstructured texts and their\ncorresponding structured knowledge, known as semantic frame, in a hierarchical\nfashion. Our approach integrates text modeling and semantic correspondence\nlearning into a single model that comprises 1) an unstructured embedding module\nthat makes use of self-similarity matrix representations in order to inject\nstructural regularities of different segments inherent in clinical texts to\npromote local coherence, 2) a structured embedding module to embed the semantic\nframes (e.g., UMLS semantic types) with deep ConvNet and 3) a hierarchical\nsemantic correspondence module that embeds by enhancing the interactions\nbetween text-semantic frame embedding pairs at multiple levels (i.e., words,\nsentence, note). Evaluations on multiple embedding benchmarks on post discharge\nintensive care patient mortality prediction tasks demonstrate its effectiveness\ncompared to approaches that do not exploit the semantic interactions between\nstructured and unstructured information present in clinical notes.', 'Large-scale pre-training methods of learning cross-modal representations on\nimage-text pairs are becoming popular for vision-language tasks. While existing\nmethods simply concatenate image region features and text features as input to\nthe model to be pre-trained and use self-attention to learn image-text semantic\nalignments in a brute force manner, in this paper, we propose a new learning\nmethod Oscar (Object-Semantics Aligned Pre-training), which uses object tags\ndetected in images as anchor points to significantly ease the learning of\nalignments. Our method is motivated by the observation that the salient objects\nin an image can be accurately detected, and are often mentioned in the paired\ntext. We pre-train an Oscar model on the public corpus of 6.5 million\ntext-image pairs, and fine-tune it on downstream tasks, creating new\nstate-of-the-arts on six well-established vision-language understanding and\ngeneration tasks.', 'Latent text representations exhibit geometric regularities, such as the\nfamous analogy: queen is to king what woman is to man. Such structured semantic\nrelations were not demonstrated on image representations. Recent works aiming\nat bridging this semantic gap embed images and text into a multimodal space,\nenabling the transfer of text-defined transformations to the image modality. We\nintroduce the SIMAT dataset to evaluate the task of Image Retrieval with\nMultimodal queries. SIMAT contains 6k images and 18k textual transformation\nqueries that aim at either replacing scene elements or changing pairwise\nrelationships between scene elements. The goal is to retrieve an image\nconsistent with the (source image, text transformation) query. We use an\nimage/text matching oracle (OSCAR) to assess whether the image transformation\nis successful. The SIMAT dataset will be publicly available. We use SIMAT to\nevaluate the geometric properties of multimodal embedding spaces trained with\nan image/text matching objective, like CLIP. We show that vanilla CLIP\nembeddings are not very well suited to transform images with delta vectors, but\nthat a simple finetuning on the COCO dataset can bring dramatic improvements.\nWe also study whether it is beneficial to leverage pretrained universal\nsentence encoders (FastText, LASER and LaBSE).', 'Web search engines focus on serving highly relevant results within hundreds\nof milliseconds. Pre-trained language transformer models such as BERT are\ntherefore hard to use in this scenario due to their high computational demands.\nWe present our real-time approach to the document ranking problem leveraging a\nBERT-based siamese architecture. The model is already deployed in a commercial\nsearch engine and it improves production performance by more than 3%. For\nfurther research and evaluation, we release DaReCzech, a unique data set of 1.6\nmillion Czech user query-document pairs with manually assigned relevance\nlevels. We also release Small-E-Czech, an Electra-small language model\npre-trained on a large Czech corpus. We believe this data will support\nendeavours both of search relevance and multilingual-focused research\ncommunities.', 'This work addresses approximate nearest neighbor search applied in the domain\nof large-scale image retrieval. Within the group testing framework we propose\nan efficient off-line construction of the search structures. The linear-time\ncomplexity orthogonal grouping increases the probability that at most one\nelement from each group is matching to a given query. Non-maxima suppression\nwith each group efficiently reduces the number of false positive results at no\nextra cost. Unlike in other well-performing approaches, all processing is\nlocal, fast, and suitable to process data in batches and in parallel. We\nexperimentally show that the proposed method achieves search accuracy of the\nexhaustive search with significant reduction in the search complexity. The\nmethod can be naturally combined with existing embedding methods.', 'Reasoning with knowledge expressed in natural language and Knowledge Bases\n(KBs) is a major challenge for Artificial Intelligence, with applications in\nmachine reading, dialogue, and question answering. General neural architectures\nthat jointly learn representations and transformations of text are very\ndata-inefficient, and it is hard to analyse their reasoning process. These\nissues are addressed by end-to-end differentiable reasoning systems such as\nNeural Theorem Provers (NTPs), although they can only be used with small-scale\nsymbolic KBs. In this paper we first propose Greedy NTPs (GNTPs), an extension\nto NTPs addressing their complexity and scalability limitations, thus making\nthem applicable to real-world datasets. This result is achieved by dynamically\nconstructing the computation graph of NTPs and including only the most\npromising proof paths during inference, thus obtaining orders of magnitude more\nefficient models. Then, we propose a novel approach for jointly reasoning over\nKBs and textual mentions, by embedding logic facts and natural language\nsentences in a shared embedding space. We show that GNTPs perform on par with\nNTPs at a fraction of their cost while achieving competitive link prediction\nresults on large datasets, providing explanations for predictions, and inducing\ninterpretable models. Source code, datasets, and supplementary material are\navailable online at https://github.com/uclnlp/gntp.', 'In this paper, we introduce an attribute-based interactive image search which\ncan leverage human-in-the-loop feedback to iteratively refine image search\nresults. We study active image search where human feedback is solicited\nexclusively in visual form, without using relative attribute annotations used\nby prior work which are not typically found in many datasets. In order to\noptimize the image selection strategy, a deep reinforcement model is trained to\nlearn what images are informative rather than rely on hand-crafted measures\ntypically leveraged in prior work. Additionally, we extend the recently\nintroduced Conditional Similarity Network to incorporate global similarity in\ntraining visual embeddings, which results in more natural transitions as the\nuser explores the learned similarity embeddings. Our experiments demonstrate\nthe effectiveness of our approach, producing compelling results on both active\nimage search and image attribute representation tasks.', 'Person search targets to search the probe person from the unconstrainted\nscene images, which can be treated as the combination of person detection and\nperson matching. However, the existing methods based on the Detection-Matching\nframework ignore the person objectness and repulsion (OR) which are both\nbeneficial to reduce the effect of distractor images. In this paper, we propose\nan OR similarity by jointly considering the objectness and repulsion\ninformation. Besides the traditional visual similarity term, the OR similarity\nalso contains an objectness term and a repulsion term. The objectness term can\nreduce the similarity of distractor images that not contain a person and boost\nthe performance of person search by improving the ranking of positive samples.\nBecause the probe person has a different person ID with its \\emph{neighbors},\nthe gallery images having a higher similarity with the \\emph{neighbors of\nprobe} should have a lower similarity with the probe person. Based on this\nrepulsion constraint, the repulsion term is proposed to reduce the similarity\nof distractor images that are not most similar to the probe person. Treating\nthe Faster R-CNN as the person detector, the OR similarity is evaluated on PRW\nand CUHK-SYSU datasets by the Detection-Matching framework with six description\nmodels. The extensive experiments demonstrate that the proposed OR similarity\ncan effectively reduce the similarity of distractor samples and further boost\nthe performance of person search, e.g., improve the mAP from 92.32% to 93.23%\nfor CUHK-SYSY dataset, and from 50.91% to 52.30% for PRW datasets.', 'We present a universal framework to model contextualized sentence\nrepresentations with visual awareness that is motivated to overcome the\nshortcomings of the multimodal parallel data with manual annotations. For each\nsentence, we first retrieve a diversity of images from a shared cross-modal\nembedding space, which is pre-trained on a large-scale of text-image pairs.\nThen, the texts and images are respectively encoded by transformer encoder and\nconvolutional neural network. The two sequences of representations are further\nfused by a simple and effective attention layer. The architecture can be easily\napplied to text-only natural language processing tasks without manually\nannotating multimodal parallel corpora. We apply the proposed method on three\ntasks, including neural machine translation, natural language inference and\nsequence labeling and experimental results verify the effectiveness.', 'The goal of text ranking is to generate an ordered list of texts retrieved\nfrom a corpus in response to a query. Although the most common formulation of\ntext ranking is search, instances of the task can also be found in many natural\nlanguage processing applications. This survey provides an overview of text\nranking with neural network architectures known as transformers, of which BERT\nis the best-known example. The combination of transformers and self-supervised\npretraining has been responsible for a paradigm shift in natural language\nprocessing (NLP), information retrieval (IR), and beyond. In this survey, we\nprovide a synthesis of existing work as a single point of entry for\npractitioners who wish to gain a better understanding of how to apply\ntransformers to text ranking problems and researchers who wish to pursue work\nin this area. We cover a wide range of modern techniques, grouped into two\nhigh-level categories: transformer models that perform reranking in multi-stage\narchitectures and dense retrieval techniques that perform ranking directly.\nThere are two themes that pervade our survey: techniques for handling long\ndocuments, beyond typical sentence-by-sentence processing in NLP, and\ntechniques for addressing the tradeoff between effectiveness (i.e., result\nquality) and efficiency (e.g., query latency, model and index size). Although\ntransformer architectures and pretraining techniques are recent innovations,\nmany aspects of how they are applied to text ranking are relatively well\nunderstood and represent mature techniques. However, there remain many open\nresearch questions, and thus in addition to laying out the foundations of\npretrained transformers for text ranking, this survey also attempts to\nprognosticate where the field is heading.', 'Neural text matching models have been widely used in community question\nanswering, information retrieval, and dialogue. However, these models designed\nfor short texts cannot well address the long-form text matching problem,\nbecause there are many contexts in long-form texts can not be directly aligned\nwith each other, and it is difficult for existing models to capture the key\nmatching signals from such noisy data. Besides, these models are\ncomputationally expensive for simply use all textual data indiscriminately. To\ntackle the effectiveness and efficiency problem, we propose a novel\nhierarchical noise filtering model, namely Match-Ignition. The main idea is to\nplug the well-known PageRank algorithm into the Transformer, to identify and\nfilter both sentence and word level noisy information in the matching process.\nNoisy sentences are usually easy to detect because previous work has shown that\ntheir similarity can be explicitly evaluated by the word overlapping, so we\ndirectly use PageRank to filter such information based on a sentence similarity\ngraph. Unlike sentences, words rely on their contexts to express concrete\nmeanings, so we propose to jointly learn the filtering and matching process, to\nwell capture the critical word-level matching signals. Specifically, a word\ngraph is first built based on the attention scores in each self-attention block\nof Transformer, and key words are then selected by applying PageRank on this\ngraph. In this way, noisy words will be filtered out layer by layer in the\nmatching process. Experimental results show that Match-Ignition outperforms\nboth SOTA short text matching models and recent long-form text matching models.\nWe also conduct detailed analysis to show that Match-Ignition efficiently\ncaptures important sentences and words, to facilitate the long-form text\nmatching process.', 'Image captioning models generally lack the capability to take into account\nuser interest, and usually default to global descriptions that try to balance\nreadability, informativeness, and information overload. On the other hand, VQA\nmodels generally lack the ability to provide long descriptive answers, while\nexpecting the textual question to be quite precise. We present a method to\ncontrol the concepts that an image caption should focus on, using an additional\ninput called the guiding text that refers to either groundable or ungroundable\nconcepts in the image. Our model consists of a Transformer-based multimodal\nencoder that uses the guiding text together with global and object-level image\nfeatures to derive early-fusion representations used to generate the guided\ncaption. While models trained on Visual Genome data have an in-domain advantage\nof fitting well when guided with automatic object labels, we find that guided\ncaptioning models trained on Conceptual Captions generalize better on\nout-of-domain images and guiding texts. Our human-evaluation results indicate\nthat attempting in-the-wild guided image captioning requires access to large,\nunrestricted-domain training datasets, and that increased style diversity (even\nwithout increasing the number of unique tokens) is a key factor for improved\nperformance.', 'Action recognition, early prediction, and online action detection are\ncomplementary disciplines that are often studied independently. Most online\naction detection networks use a pre-trained feature extractor, which might not\nbe optimal for its new task. We address the task-specific feature extraction\nwith a teacher-student framework between the aforementioned disciplines, and a\nnovel training strategy. Our network, Online Knowledge Distillation Action\nDetection network (OKDAD), embeds online early prediction and online temporal\nsegment proposal subnetworks in parallel. Low interclass and high intraclass\nsimilarity are encouraged during teacher training. Knowledge distillation to\nthe OKDAD network is ensured via layer reuse and cosine similarity between\nteacher-student feature vectors. Layer reuse and similarity learning\nsignificantly improve our baseline which uses a generic feature extractor. We\nevaluate our framework on infrared videos from two popular datasets, NTU RGB+D\n(action recognition, early prediction) and PKU MMD (action detection). Unlike\nprevious attempts on those datasets, our student networks perform without any\nknowledge of the future. Even with this added difficulty, we achieve\nstate-of-the-art results on both datasets. Moreover, our networks use infrared\nfrom RGB-D cameras, which we are the first to use for online action detection,\nto our knowledge.', 'We consider the large-scale query-document retrieval problem: given a query\n(e.g., a question), return the set of relevant documents (e.g., paragraphs\ncontaining the answer) from a large document corpus. This problem is often\nsolved in two steps. The retrieval phase first reduces the solution space,\nreturning a subset of candidate documents. The scoring phase then re-ranks the\ndocuments. Critically, the retrieval algorithm not only desires high recall but\nalso requires to be highly efficient, returning candidates in time sublinear to\nthe number of documents. Unlike the scoring phase witnessing significant\nadvances recently due to the BERT-style pre-training tasks on cross-attention\nmodels, the retrieval phase remains less well studied. Most previous works rely\non classic Information Retrieval (IR) methods such as BM-25 (token matching +\nTF-IDF weights). These models only accept sparse handcrafted features and can\nnot be optimized for different downstream tasks of interest. In this paper, we\nconduct a comprehensive study on the embedding-based retrieval models. We show\nthat the key ingredient of learning a strong embedding-based Transformer model\nis the set of pre-training tasks. With adequately designed paragraph-level\npre-training tasks, the Transformer models can remarkably improve over the\nwidely-used BM-25 as well as embedding models without Transformers. The\nparagraph-level pre-training tasks we studied are Inverse Cloze Task (ICT),\nBody First Selection (BFS), Wiki Link Prediction (WLP), and the combination of\nall three.', 'The massive scale and growth of textual biomedical data have made its\nindexing and classification increasingly important. However, existing research\non this topic mainly utilized convolutional and recurrent neural networks,\nwhich generally achieve inferior performance than the novel transformers. On\nthe other hand, systems that apply transformers only focus on the target\ndocuments, overlooking the rich semantic information that label descriptions\ncontain. To address this gap, we develop a transformer-based biomedical text\nclassifier that considers label information. The system achieves this with a\nlabel attention module incorporated into the fine-tuning process of pretrained\nlanguage models (PTMs). Our results on two public medical datasets show that\nthe proposed fine-tuning scheme outperforms the vanilla PTMs and\nstate-of-the-art models.', 'In this paper we propose a framework for predicting kernelized classifiers in\nthe visual domain for categories with no training images where the knowledge\ncomes from textual description about these categories. Through our optimization\nframework, the proposed approach is capable of embedding the class-level\nknowledge from the text domain as kernel classifiers in the visual domain. We\nalso proposed a distributional semantic kernel between text descriptions which\nis shown to be effective in our setting. The proposed framework is not\nrestricted to textual descriptions, and can also be applied to other forms\nknowledge representations. Our approach was applied for the challenging task of\nzero-shot learning of fine-grained categories from text descriptions of these\ncategories.', 'We propose an accurate and efficient scene text detection framework, termed\nFAST (i.e., faster arbitrarily-shaped text detector). Different from recent\nadvanced text detectors that used complicated post-processing and hand-crafted\nnetwork architectures, resulting in low inference speed, FAST has two new\ndesigns. (1) We design a minimalist kernel representation (only has 1-channel\noutput) to model text with arbitrary shape, as well as a GPU-parallel\npost-processing to efficiently assemble text lines with a negligible time\noverhead. (2) We search the network architecture tailored for text detection,\nleading to more powerful features than most networks that are searched for\nimage classification. Benefiting from these two designs, FAST achieves an\nexcellent trade-off between accuracy and efficiency on several challenging\ndatasets, including Total Text, CTW1500, ICDAR 2015, and MSRA-TD500. For\nexample, FAST-T yields 81.6% F-measure at 152 FPS on Total-Text, outperforming\nthe previous fastest method by 1.7 points and 70 FPS in terms of accuracy and\nspeed. With TensorRT optimization, the inference speed can be further\naccelerated to over 600 FPS. Code and models will be released at\nhttps://github.com/czczup/FAST.', 'For some images, descriptions written by multiple people are consistent with\neach other. But for other images, descriptions across people vary considerably.\nIn other words, some images are specific $-$ they elicit consistent\ndescriptions from different people $-$ while other images are ambiguous.\nApplications involving images and text can benefit from an understanding of\nwhich images are specific and which ones are ambiguous. For instance, consider\ntext-based image retrieval. If a query description is moderately similar to the\ncaption (or reference description) of an ambiguous image, that query may be\nconsidered a decent match to the image. But if the image is very specific, a\nmoderate similarity between the query and the reference description may not be\nsufficient to retrieve the image.\n  In this paper, we introduce the notion of image specificity. We present two\nmechanisms to measure specificity given multiple descriptions of an image: an\nautomated measure and a measure that relies on human judgement. We analyze\nimage specificity with respect to image content and properties to better\nunderstand what makes an image specific. We then train models to automatically\npredict the specificity of an image from image features alone without requiring\ntextual descriptions of the image. Finally, we show that modeling image\nspecificity leads to improvements in a text-based image retrieval application.', 'The goal of our industrial ticketing system is to retrieve a relevant\nsolution for an input query, by matching with historical tickets stored in\nknowledge base. A query is comprised of subject and description, while a\nhistorical ticket consists of subject, description and solution. To retrieve a\nrelevant solution, we use textual similarity paradigm to learn similarity in\nthe query and historical tickets. The task is challenging due to significant\nterm mismatch in the query and ticket pairs of asymmetric lengths, where\nsubject is a short text but description and solution are multi-sentence texts.\nWe present a novel Replicated Siamese LSTM model to learn similarity in\nasymmetric text pairs, that gives 22% and 7% gain (Accuracy@10) for retrieval\ntask, respectively over unsupervised and supervised baselines. We also show\nthat the topic and distributed semantic features for short and long texts\nimproved both similarity learning and retrieval.', 'Modern convolutional neural networks apply the same operations on every pixel\nin an image. However, not all image regions are equally important. To address\nthis inefficiency, we propose a method to dynamically apply convolutions\nconditioned on the input image. We introduce a residual block where a small\ngating branch learns which spatial positions should be evaluated. These\ndiscrete gating decisions are trained end-to-end using the Gumbel-Softmax\ntrick, in combination with a sparsity criterion. Our experiments on CIFAR,\nImageNet and MPII show that our method has better focus on the region of\ninterest and better accuracy than existing methods, at a lower computational\ncomplexity. Moreover, we provide an efficient CUDA implementation of our\ndynamic convolutions using a gather-scatter approach, achieving a significant\nimprovement in inference speed with MobileNetV2 residual blocks. On human pose\nestimation, a task that is inherently spatially sparse, the processing speed is\nincreased by 60% with no loss in accuracy.', 'Text-to-image person re-identification (ReID) aims to search for images\ncontaining a person of interest using textual descriptions. However, due to the\nsignificant modality gap and the large intra-class variance in textual\ndescriptions, text-to-image ReID remains a challenging problem. Accordingly, in\nthis paper, we propose a Semantically Self-Aligned Network (SSAN) to handle the\nabove problems. First, we propose a novel method that automatically extracts\nsemantically aligned part-level features from the two modalities. Second, we\ndesign a multi-view non-local network that captures the relationships between\nbody parts, thereby establishing better correspondences between body parts and\nnoun phrases. Third, we introduce a Compound Ranking (CR) loss that makes use\nof textual descriptions for other images of the same identity to provide extra\nsupervision, thereby effectively reducing the intra-class variance in textual\nfeatures. Finally, to expedite future research in text-to-image ReID, we build\na new database named ICFG-PEDES. Extensive experiments demonstrate that SSAN\noutperforms state-of-the-art approaches by significant margins. Both the new\nICFG-PEDES database and the SSAN code are available at\nhttps://github.com/zifyloo/SSAN.', 'The abundance of multimodal data (e.g. social media posts) has inspired\ninterest in cross-modal retrieval methods. Popular approaches rely on a variety\nof metric learning losses, which prescribe what the proximity of image and text\nshould be, in the learned space. However, most prior methods have focused on\nthe case where image and text convey redundant information; in contrast,\nreal-world image-text pairs convey complementary information with little\noverlap. Further, images in news articles and media portray topics in a\nvisually diverse fashion; thus, we need to take special care to ensure a\nmeaningful image representation. We propose novel within-modality losses which\nencourage semantic coherency in both the text and image subspaces, which does\nnot necessarily align with visual coherency. Our method ensures that not only\nare paired images and texts close, but the expected image-image and text-text\nrelationships are also observed. Our approach improves the results of\ncross-modal retrieval on four datasets compared to five baselines.', 'Conducting text retrieval in a dense learned representation space has many\nintriguing advantages over sparse retrieval. Yet the effectiveness of dense\nretrieval (DR) often requires combination with sparse retrieval. In this paper,\nwe identify that the main bottleneck is in the training mechanisms, where the\nnegative instances used in training are not representative of the irrelevant\ndocuments in testing. This paper presents Approximate nearest neighbor Negative\nContrastive Estimation (ANCE), a training mechanism that constructs negatives\nfrom an Approximate Nearest Neighbor (ANN) index of the corpus, which is\nparallelly updated with the learning process to select more realistic negative\ntraining instances. This fundamentally resolves the discrepancy between the\ndata distribution used in the training and testing of DR. In our experiments,\nANCE boosts the BERT-Siamese DR model to outperform all competitive dense and\nsparse retrieval baselines. It nearly matches the accuracy of\nsparse-retrieval-and-BERT-reranking using dot-product in the ANCE-learned\nrepresentation space and provides almost 100x speed-up.', 'We propose Unicoder-VL, a universal encoder that aims to learn joint\nrepresentations of vision and language in a pre-training manner. Borrow ideas\nfrom cross-lingual pre-trained models, such as XLM and Unicoder, both visual\nand linguistic contents are fed into a multi-layer Transformer for the\ncross-modal pre-training, where three pre-trained tasks are employed, including\nMasked Language Modeling (MLM), Masked Object Classification (MOC) and\nVisual-linguistic Matching (VLM). The first two tasks learn context-aware\nrepresentations for input tokens based on linguistic and visual contents\njointly. The last task tries to predict whether an image and a text describe\neach other. After pretraining on large-scale image-caption pairs, we transfer\nUnicoder-VL to caption-based image-text retrieval and visual commonsense\nreasoning, with just one additional output layer. We achieve state-of-the-art\nor comparable results on both two tasks and show the powerful ability of the\ncross-modal pre-training.', ""This paper investigates whether the power of the models pre-trained on text\ndata, such as BERT, can be transferred to general token sequence classification\napplications. To verify pre-trained models' transferability, we test the\npre-trained models on text classification tasks with meanings of tokens\nmismatches, and real-world non-text token sequence classification data,\nincluding amino acid, DNA, and music. We find that even on non-text data, the\nmodels pre-trained on text converge faster, perform better than the randomly\ninitialized models, and only slightly worse than the models using task-specific\nknowledge. We also find that the representations of the text and non-text\npre-trained models share non-trivial similarities."", 'Image to image translation aims to learn a mapping that transforms an image\nfrom one visual domain to another. Recent works assume that images descriptors\ncan be disentangled into a domain-invariant content representation and a\ndomain-specific style representation. Thus, translation models seek to preserve\nthe content of source images while changing the style to a target visual\ndomain. However, synthesizing new images is extremely challenging especially in\nmulti-domain translations, as the network has to compose content and style to\ngenerate reliable and diverse images in multiple domains. In this paper we\npropose the use of an image retrieval system to assist the image-to-image\ntranslation task. First, we train an image-to-image translation model to map\nimages to multiple domains. Then, we train an image retrieval model using real\nand generated images to find images similar to a query one in content but in a\ndifferent domain. Finally, we exploit the image retrieval system to fine-tune\nthe image-to-image translation model and generate higher quality images. Our\nexperiments show the effectiveness of the proposed solution and highlight the\ncontribution of the retrieval network, which can benefit from additional\nunlabeled data and help image-to-image translation models in the presence of\nscarce data.', 'In this paper we propose the Structured Deep Neural Network (structured DNN)\nas a structured and deep learning framework. This approach can learn to find\nthe best structured object (such as a label sequence) given a structured input\n(such as a vector sequence) by globally considering the mapping relationships\nbetween the structures rather than item by item.\n  When automatic speech recognition is viewed as a special case of such a\nstructured learning problem, where we have the acoustic vector sequence as the\ninput and the phoneme label sequence as the output, it becomes possible to\ncomprehensively learn utterance by utterance as a whole, rather than frame by\nframe.\n  Structured Support Vector Machine (structured SVM) was proposed to perform\nASR with structured learning previously, but limited by the linear nature of\nSVM. Here we propose structured DNN to use nonlinear transformations in\nmulti-layers as a structured and deep learning approach. This approach was\nshown to beat structured SVM in preliminary experiments on TIMIT.', 'Lattices are an efficient and effective method to encode ambiguity of\nupstream systems in natural language processing tasks, for example to compactly\ncapture multiple speech recognition hypotheses, or to represent multiple\nlinguistic analyses. Previous work has extended recurrent neural networks to\nmodel lattice inputs and achieved improvements in various tasks, but these\nmodels suffer from very slow computation speeds. This paper extends the\nrecently proposed paradigm of self-attention to handle lattice inputs.\nSelf-attention is a sequence modeling technique that relates inputs to one\nanother by computing pairwise similarities and has gained popularity for both\nits strong results and its computational efficiency. To extend such models to\nhandle lattices, we introduce probabilistic reachability masks that incorporate\nlattice structure into the model and support lattice scores if available. We\nalso propose a method for adapting positional embeddings to lattice structures.\nWe apply the proposed model to a speech translation task and find that it\noutperforms all examined baselines while being much faster to compute than\nprevious neural lattice models during both training and inference.', 'Image manipulation can be considered a special case of image generation where\nthe image to be produced is a modification of an existing image. Image\ngeneration and manipulation have been, for the most part, tasks that operate on\nraw pixels. However, the remarkable progress in learning rich image and object\nrepresentations has opened the way for tasks such as text-to-image or\nlayout-to-image generation that are mainly driven by semantics. In our work, we\naddress the novel problem of image manipulation from scene graphs, in which a\nuser can edit images by merely applying changes in the nodes or edges of a\nsemantic graph that is generated from the image. Our goal is to encode image\ninformation in a given constellation and from there on generate new\nconstellations, such as replacing objects or even changing relationships\nbetween objects, while respecting the semantics and style from the original\nimage. We introduce a spatio-semantic scene graph network that does not require\ndirect supervision for constellation changes or image edits. This makes it\npossible to train the system from existing real-world datasets with no\nadditional annotation effort.', 'Measuring visual similarity is critical for image understanding. But what\nmakes two images similar? Most existing work on visual similarity assumes that\nimages are similar because they contain the same object instance or category.\nHowever, the reason why images are similar is much more complex. For example,\nfrom the perspective of category, a black dog image is similar to a white dog\nimage. However, in terms of color, a black dog image is more similar to a black\nhorse image than the white dog image. This example serves to illustrate that\nvisual similarity is ambiguous but can be made precise when given an explicit\ncontextual perspective. Based on this observation, we propose the concept of\ncontextual visual similarity. To be concrete, we examine the concept of\ncontextual visual similarity in the application domain of image search. Instead\nof providing only a single image for image similarity search (\\eg, Google image\nsearch), we require three images. Given a query image, a second positive image\nand a third negative image, dissimilar to the first two images, we define a\ncontextualized similarity search criteria. In particular, we learn feature\nweights over all the feature dimensions of each image such that the distance\nbetween the query image and the positive image is small and their distances to\nthe negative image are large after reweighting their features. The learned\nfeature weights encode the contextualized visual similarity specified by the\nuser and can be used for attribute specific image search. We also show the\nusefulness of our contextualized similarity weighting scheme for different\ntasks, such as answering visual analogy questions and unsupervised attribute\ndiscovery.', 'In an automated search system, similarity is a key concept in solving a human\ntask. Indeed, human process is usually a natural categorization that underlies\nmany natural abilities such as image recovery, language comprehension, decision\nmaking, or pattern recognition. In the image search axis, there are several\nways to measure the similarity between images in an image database, to a query\nimage. Image search by content is based on the similarity of the visual\ncharacteristics of the images. The distance function used to evaluate the\nsimilarity between images depends on the criteria of the search but also on the\nrepresentation of the characteristics of the image; this is the main idea of\nthe near and fuzzy sets approaches. In this article, we introduce a new\ncategory of beta type-2 fuzzy sets for the description of image characteristics\nas well as the near sets approach for image recovery. Finally, we illustrate\nour work with examples of image recovery problems used in the real world.', 'The growing deluge of scientific publications demands text analysis tools\nthat can help scientists and policy-makers navigate, forecast and beneficially\nguide scientific research. Recent advances in natural language understanding\ndriven by deep transformer networks offer new possibilities for mapping\nscience. Because the same surface text can take on multiple and sometimes\ncontradictory specialized senses across distinct research communities,\nsensitivity to context is critical for infometric applications. Transformer\nembedding models such as BERT capture shades of association and connotation\nthat vary across the different linguistic contexts of any particular word or\nspan of text. Here we report a procedure for encoding scientific documents with\nthese tools, measuring their improvement over static word embeddings in a\nnearest-neighbor retrieval task. We find discriminability of contextual\nrepresentations is strongly influenced by choice of pooling strategy for\nsummarizing the high-dimensional network activations. Importantly, we note that\nfundamentals such as domain-matched training data are more important than\nstate-of-the-art NLP tools. Yet state-of-the-art models did offer significant\ngains. The best approach we investigated combined domain-matched pretraining,\nsound pooling, and state-of-the-art deep transformer network encoders. Finally,\nwith the goal of leveraging contextual representations from deep encoders, we\npresent a range of measurements for understanding and forecasting research\ncommunities in science.', 'We present a new method for image feature-extraction which is based on\nrepresenting an image by a finite-dimensional vector of distances that measure\nhow different the image is from a set of image prototypes. We use the recently\nintroduced Universal Image Distance (UID) \\cite{RatsabyChesterIEEE2012} to\ncompare the similarity between an image and a prototype image. The advantage in\nusing the UID is the fact that no domain knowledge nor any image analysis need\nto be done. Each image is represented by a finite dimensional feature vector\nwhose components are the UID values between the image and a finite set of image\nprototypes from each of the feature categories. The method is automatic since\nonce the user selects the prototype images, the feature vectors are\nautomatically calculated without the need to do any image analysis. The\nprototype images can be of different size, in particular, different than the\nimage size. Based on a collection of such cases any supervised or unsupervised\nlearning algorithm can be used to train and produce an image classifier or\nimage cluster analysis. In this paper we present the image feature-extraction\nmethod and use it on several supervised and unsupervised learning experiments\nfor satellite image data.', 'Matching model is essential for Image-Text Retrieval framework. Existing\nresearch usually train the model with a triplet loss and explore various\nstrategy to retrieve hard negative sentences in the dataset. We argue that\ncurrent retrieval-based negative sample construction approach is limited in the\nscale of the dataset thus fail to identify negative sample of high difficulty\nfor every image. We propose our TAiloring neGative Sentences with\nDiscrimination and Correction (TAGS-DC) to generate synthetic sentences\nautomatically as negative samples. TAGS-DC is composed of masking and refilling\nto generate synthetic negative sentences with higher difficulty. To keep the\ndifficulty during training, we mutually improve the retrieval and generation\nthrough parameter sharing. To further utilize fine-grained semantic of mismatch\nin the negative sentence, we propose two auxiliary tasks, namely word\ndiscrimination and word correction to improve the training. In experiments, we\nverify the effectiveness of our model on MS-COCO and Flickr30K compared with\ncurrent state-of-the-art models and demonstrates its robustness and\nfaithfulness in the further analysis. Our code is available in\nhttps://github.com/LibertFan/TAGS.', 'Recent progress in pretrained Transformer-based language models has shown\ngreat success in learning contextual representation of text. However, due to\nthe quadratic self-attention complexity, most of the pretrained Transformers\nmodels can only handle relatively short text. It is still a challenge when it\ncomes to modeling very long documents. In this work, we propose to use a graph\nattention network on top of the available pretrained Transformers model to\nlearn document embeddings. This graph attention network allows us to leverage\nthe high-level semantic structure of the document. In addition, based on our\ngraph document model, we design a simple contrastive learning strategy to\npretrain our models on a large amount of unlabeled corpus. Empirically, we\ndemonstrate the effectiveness of our approaches in document classification and\ndocument retrieval tasks.', 'Are we using the right potential functions in the Conditional Random Field\nmodels that are popular in the Vision community? Semantic segmentation and\nother pixel-level labelling tasks have made significant progress recently due\nto the deep learning paradigm. However, most state-of-the-art structured\nprediction methods also include a random field model with a hand-crafted\nGaussian potential to model spatial priors, label consistencies and\nfeature-based image conditioning.\n  In this paper, we challenge this view by developing a new inference and\nlearning framework which can learn pairwise CRF potentials restricted only by\ntheir dependence on the image pixel values and the size of the support. Both\nstandard spatial and high-dimensional bilateral kernels are considered. Our\nframework is based on the observation that CRF inference can be achieved via\nprojected gradient descent and consequently, can easily be integrated in deep\nneural networks to allow for end-to-end training. It is empirically\ndemonstrated that such learned potentials can improve segmentation accuracy and\nthat certain label class interactions are indeed better modelled by a\nnon-Gaussian potential. In addition, we compare our inference method to the\ncommonly used mean-field algorithm. Our framework is evaluated on several\npublic benchmarks for semantic segmentation with improved performance compared\nto previous state-of-the-art CNN+CRF models.', 'Internal learning for single-image generation is a framework, where a\ngenerator is trained to produce novel images based on a single image. Since\nthese models are trained on a single image, they are limited in their scale and\napplication. To overcome these issues, we propose a meta-learning approach that\nenables training over a collection of images, in order to model the internal\nstatistics of the sample image more effectively. In the presented meta-learning\napproach, a single-image GAN model is generated given an input image, via a\nconvolutional feedforward hypernetwork $f$. This network is trained over a\ndataset of images, allowing for feature sharing among different models, and for\ninterpolation in the space of generative models. The generated single-image\nmodel contains a hierarchy of multiple generators and discriminators. It is\ntherefore required to train the meta-learner in an adversarial manner, which\nrequires careful design choices that we justify by a theoretical analysis. Our\nresults show that the models obtained are as suitable as single-image GANs for\nmany common image applications, significantly reduce the training time per\nimage without loss in performance, and introduce novel capabilities, such as\ninterpolation and feedforward modeling of novel images.', 'Similarity searching of molecular structure has been an important application\nin the Chemoinformatics, especially in drug discovery. Similarity searching is\na common method used for identification of molecular structure. It involve\nthree main principal component of similarity searching: structure\nrepresentation; weighting scheme; and similarity coefficient. In this paper, we\nintroduces Weighted Tanimoto Coefficient based on weighted Euclidean distance\nin order to investigate the effect of weight function on the result for\nsimilarity searching. The Tanimoto coefficient is one of the popular similarity\ncoefficients used to measure the similarity between pairs of the molecule. The\nmost of research area found that the similarity searching is based on binary or\nfingerprint data. Meanwhile, we used non-binary data and was set amphetamine\nstructure as a reference or targeted structure and the rest of the dataset\nbecomes a database structure. Throughout this study, it showed that there is\ndefinitely gives a different result between a similarity searching with and\nwithout weight.', 'Understanding images without explicit supervision has become an important\nproblem in computer vision. In this paper, we address image captioning by\ngenerating language descriptions of scenes without learning from annotated\npairs of images and their captions. The core component of our approach is a\nshared latent space that is structured by visual concepts. In this space, the\ntwo modalities should be indistinguishable. A language model is first trained\nto encode sentences into semantically structured embeddings. Image features\nthat are translated into this embedding space can be decoded into descriptions\nthrough the same language model, similarly to sentence embeddings. This\ntranslation is learned from weakly paired images and text using a loss robust\nto noisy assignments and a conditional adversarial component. Our approach\nallows to exploit large text corpora outside the annotated distributions of\nimage/caption data. Our experiments show that the proposed domain alignment\nlearns a semantically meaningful representation which outperforms previous\nwork.', 'Deep learning models have significantly improved the visual quality and\naccuracy on compressive sensing recovery. In this paper, we propose an\nalgorithm for signal reconstruction from compressed measurements with image\npriors captured by a generative model. We search and constrain on latent\nvariable space to make the method stable when the number of compressed\nmeasurements is extremely limited. We show that, by exploiting certain\nstructures of the latent variables, the proposed method produces improved\nreconstruction accuracy and preserves realistic and non-smooth features in the\nimage. Our algorithm achieves high computation speed by projecting between the\noriginal signal space and the latent variable space in an alternating fashion.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'Everyone knows that thousand of words are represented by a single image. As a\nresult image search has become a very popular mechanism for the Web searchers.\nImage search means, the search results are produced by the search engine should\nbe a set of images along with their Web page Unified Resource Locator. Now Web\nsearcher can perform two types of image search, they are Text to Image and\nImage to Image search. In Text to Image search, search query should be a text.\nBased on the input text data system will generate a set of images along with\ntheir Web page URL as an output. On the other hand, in Image to Image search,\nsearch query should be an image and based on this image system will generate a\nset of images along with their Web page URL as an output. According to the\ncurrent scenarios, Text to Image search mechanism always not returns perfect\nresult. It matches the text data and then displays the corresponding images as\nan output, which is not always perfect. To resolve this problem, Web\nresearchers have introduced the Image to Image search mechanism. In this paper,\nwe have also proposed an alternate approach of Image to Image search mechanism\nusing Histogram.', 'The objective of deep metric learning (DML) is to learn embeddings that can\ncapture semantic similarity and dissimilarity information among data points.\nExisting pairwise or tripletwise loss functions used in DML are known to suffer\nfrom slow convergence due to a large proportion of trivial pairs or triplets as\nthe model improves. To improve this, ranking-motivated structured losses are\nproposed recently to incorporate multiple examples and exploit the structured\ninformation among them. They converge faster and achieve state-of-the-art\nperformance. In this work, we unveil two limitations of existing\nranking-motivated structured losses and propose a novel ranked list loss to\nsolve both of them. First, given a query, only a fraction of data points is\nincorporated to build the similarity structure. Consequently, some useful\nexamples are ignored and the structure is less informative. To address this, we\npropose to build a set-based similarity structure by exploiting all instances\nin the gallery. The learning setting can be interpreted as few-shot retrieval:\ngiven a mini-batch, every example is iteratively used as a query, and the rest\nones compose the gallery to search, i.e., the support set in few-shot setting.\nThe rest examples are split into a positive set and a negative set. For every\nmini-batch, the learning objective of ranked list loss is to make the query\ncloser to the positive set than to the negative set by a margin. Second,\nprevious methods aim to pull positive pairs as close as possible in the\nembedding space. As a result, the intraclass data distribution tends to be\nextremely compressed. In contrast, we propose to learn a hypersphere for each\nclass in order to preserve useful similarity structure inside it, which\nfunctions as regularisation. Extensive experiments demonstrate the superiority\nof our proposal by comparing with the state-of-the-art methods.', 'The Flickr30k dataset has become a standard benchmark for sentence-based\nimage description. This paper presents Flickr30k Entities, which augments the\n158k captions from Flickr30k with 244k coreference chains, linking mentions of\nthe same entities across different captions for the same image, and associating\nthem with 276k manually annotated bounding boxes. Such annotations are\nessential for continued progress in automatic image description and grounded\nlanguage understanding. They enable us to define a new benchmark for\nlocalization of textual entity mentions in an image. We present a strong\nbaseline for this task that combines an image-text embedding, detectors for\ncommon objects, a color classifier, and a bias towards selecting larger\nobjects. While our baseline rivals in accuracy more complex state-of-the-art\nmodels, we show that its gains cannot be easily parlayed into improvements on\nsuch tasks as image-sentence retrieval, thus underlining the limitations of\ncurrent methods and the need for further research.', 'We introduce a new task, visual sense disambiguation for verbs: given an\nimage and a verb, assign the correct sense of the verb, i.e., the one that\ndescribes the action depicted in the image. Just as textual word sense\ndisambiguation is useful for a wide range of NLP tasks, visual sense\ndisambiguation can be useful for multimodal tasks such as image retrieval,\nimage description, and text illustration. We introduce VerSe, a new dataset\nthat augments existing multimodal datasets (COCO and TUHOI) with sense labels.\nWe propose an unsupervised algorithm based on Lesk which performs visual sense\ndisambiguation using textual, visual, or multimodal embeddings. We find that\ntextual embeddings perform well when gold-standard textual annotations (object\nlabels and image descriptions) are available, while multimodal embeddings\nperform well on unannotated images. We also verify our findings by using the\ntextual and multimodal embeddings as features in a supervised setting and\nanalyse the performance of visual sense disambiguation task. VerSe is made\npublicly available and can be downloaded at:\nhttps://github.com/spandanagella/verse.', 'Two competing file formats have become the de facto standards for\ndistributing pre-trained word embeddings. Both are named after the most popular\npre-trained embeddings that are distributed in that format. The GloVe format is\nan entirely text based format that suffers from huge file sizes and slow reads,\nand the word2vec format is a smaller binary format that mixes a textual\nrepresentation of words with a binary representation of the vectors themselves.\nBoth formats have problems that we solve with a new format we call the Leader\nformat. We include a word length prefix for faster reads while maintaining the\nsmaller file size a binary format offers. We also created a minimalist library\nto facilitate the reading and writing of various word vector formats, as well\nas tools for converting pre-trained embeddings to our new Leader format.', 'Image decomposition is a crucial subject in the field of image processing. It\ncan extract salient features from the source image. We propose a new image\ndecomposition method based on convolutional neural network. This method can be\napplied to many image processing tasks. In this paper, we apply the image\ndecomposition network to the image fusion task. We input infrared image and\nvisible light image and decompose them into three high-frequency feature images\nand a low-frequency feature image respectively. The two sets of feature images\nare fused using a specific fusion strategy to obtain fusion feature images.\nFinally, the feature images are reconstructed to obtain the fused image.\nCompared with the state-of-the-art fusion methods, this method has achieved\nbetter performance in both subjective and objective evaluation.', 'The multi-decoder (MD) end-to-end speech translation model has demonstrated\nhigh translation quality by searching for better intermediate automatic speech\nrecognition (ASR) decoder states as hidden intermediates (HI). It is a two-pass\ndecoding model decomposing the overall task into ASR and machine translation\nsub-tasks. However, the decoding speed is not fast enough for real-world\napplications because it conducts beam search for both sub-tasks during\ninference. We propose Fast-MD, a fast MD model that generates HI by\nnon-autoregressive (NAR) decoding based on connectionist temporal\nclassification (CTC) outputs followed by an ASR decoder. We investigated two\ntypes of NAR HI: (1) parallel HI by using an autoregressive Transformer ASR\ndecoder and (2) masked HI by using Mask-CTC, which combines CTC and the\nconditional masked language model. To reduce a mismatch in the ASR decoder\nbetween teacher-forcing during training and conditioning on CTC outputs during\ntesting, we also propose sampling CTC outputs during training. Experimental\nevaluations on three corpora show that Fast-MD achieved about 2x and 4x faster\ndecoding speed than that of the na\\""ive MD model on GPU and CPU with comparable\ntranslation quality. Adopting the Conformer encoder and intermediate CTC loss\nfurther boosts its quality without sacrificing decoding speed.', 'Neural architecture search (NAS) has become an important approach to\nautomatically find effective architectures. To cover all possible good\narchitectures, we need to search in an extremely large search space with\nbillions of candidate architectures. More critically, given a large search\nspace, we may face a very challenging issue of space explosion. However, due to\nthe limitation of computational resources, we can only sample a very small\nproportion of the architectures, which provides insufficient information for\nthe training. As a result, existing methods may often produce suboptimal\narchitectures. To alleviate this issue, we propose a curriculum search method\nthat starts from a small search space and gradually incorporates the learned\nknowledge to guide the search in a large space. With the proposed search\nstrategy, our Curriculum Neural Architecture Search (CNAS) method significantly\nimproves the search efficiency and finds better architectures than existing NAS\nmethods. Extensive experiments on CIFAR-10 and ImageNet demonstrate the\neffectiveness of the proposed method.', 'Recently, the vulnerability of deep image classification models to\nadversarial attacks has been investigated. However, such an issue has not been\nthoroughly studied for image-to-image tasks that take an input image and\ngenerate an output image (e.g., colorization, denoising, deblurring, etc.) This\npaper presents comprehensive investigations into the vulnerability of deep\nimage-to-image models to adversarial attacks. For five popular image-to-image\ntasks, 16 deep models are analyzed from various standpoints such as output\nquality degradation due to attacks, transferability of adversarial examples\nacross different tasks, and characteristics of perturbations. We show that\nunlike image classification tasks, the performance degradation on\nimage-to-image tasks largely differs depending on various factors, e.g., attack\nmethods and task objectives. In addition, we analyze the effectiveness of\nconventional defense methods used for classification models in improving the\nrobustness of the image-to-image models.', 'Edit-based approaches have recently shown promising results on multiple\nmonolingual sequence transduction tasks. In contrast to conventional\nsequence-to-sequence (Seq2Seq) models, which learn to generate text from\nscratch as they are trained on parallel corpora, these methods have proven to\nbe much more effective since they are able to learn to make fast and accurate\ntransformations while leveraging powerful pre-trained language models. Inspired\nby these ideas, we present TST, a simple and efficient Text Simplification\nsystem based on sequence Tagging, leveraging pre-trained Transformer-based\nencoders. Our system makes simplistic data augmentations and tweaks in training\nand inference on a pre-existing system, which makes it less reliant on large\namounts of parallel training data, provides more control over the outputs and\nenables faster inference speeds. Our best model achieves near state-of-the-art\nperformance on benchmark test datasets for the task. Since it is fully\nnon-autoregressive, it achieves faster inference speeds by over 11 times than\nthe current state-of-the-art text simplification system.', 'Link prediction is the task of predicting missing connections between\nentities in the knowledge graph (KG). While various forms of models are\nproposed for the link prediction task, most of them are designed based on a few\nknown relation patterns in several well-known datasets. Due to the diversity\nand complexity nature of the real-world KGs, it is inherently difficult to\ndesign a model that fits all datasets well. To address this issue, previous\nwork has tried to use Automated Machine Learning (AutoML) to search for the\nbest model for a given dataset. However, their search space is limited only to\nbilinear model families. In this paper, we propose a novel Neural Architecture\nSearch (NAS) framework for the link prediction task. First, the embeddings of\nthe input triplet are refined by the Representation Search Module. Then, the\nprediction score is searched within the Score Function Search Module. This\nframework entails a more general search space, which enables us to take\nadvantage of several mainstream model families, and thus it can potentially\nachieve better performance. We relax the search space to be continuous so that\nthe architecture can be optimized efficiently using gradient-based search\nstrategies. Experimental results on several benchmark datasets demonstrate the\neffectiveness of our method compared with several state-of-the-art approaches.', 'The requirement of large amounts of annotated images has become one grand\nchallenge while training deep neural network models for various visual\ndetection and recognition tasks. This paper presents a novel image synthesis\ntechnique that aims to generate a large amount of annotated scene text images\nfor training accurate and robust scene text detection and recognition models.\nThe proposed technique consists of three innovative designs. First, it realizes\n""semantic coherent"" synthesis by embedding texts at semantically sensible\nregions within the background image, where the semantic coherence is achieved\nby leveraging the semantic annotations of objects and image regions that have\nbeen created in the prior semantic segmentation research. Second, it exploits\nvisual saliency to determine the embedding locations within each semantic\nsensible region, which coincides with the fact that texts are often placed\naround homogeneous regions for better visibility in scenes. Third, it designs\nan adaptive text appearance model that determines the color and brightness of\nembedded texts by learning from the feature of real scene text images\nadaptively. The proposed technique has been evaluated over five public datasets\nand the experiments show its superior performance in training accurate and\nrobust scene text detection and recognition models.', 'Image-to-image translation is a long-established and a difficult problem in\ncomputer vision. In this paper we propose an adversarial based model for\nimage-to-image translation. The regular deep neural-network based methods\nperform the task of image-to-image translation by comparing gram matrices and\nusing image segmentation which requires human intervention. Our generative\nadversarial network based model works on a conditional probability approach.\nThis approach makes the image translation independent of any local, global and\ncontent or style features. In our approach we use a bidirectional\nreconstruction model appended with the affine transform factor that helps in\nconserving the content and photorealism as compared to other models. The\nadvantage of using such an approach is that the image-to-image translation is\nsemi-supervised, independant of image segmentation and inherits the properties\nof generative adversarial networks tending to produce realistic. This method\nhas proven to produce better results than Multimodal Unsupervised\nImage-to-image translation.', ""We study the behavior of several black-box search algorithms used for\ngenerating adversarial examples for natural language processing (NLP) tasks. We\nperform a fine-grained analysis of three elements relevant to search: search\nalgorithm, search space, and search budget. When new search algorithms are\nproposed in past work, the attack search space is often modified alongside the\nsearch algorithm. Without ablation studies benchmarking the search algorithm\nchange with the search space held constant, one cannot tell if an increase in\nattack success rate is a result of an improved search algorithm or a less\nrestrictive search space. Additionally, many previous studies fail to properly\nconsider the search algorithms' run-time cost, which is essential for\ndownstream tasks like adversarial training. Our experiments provide a\nreproducible benchmark of search algorithms across a variety of search spaces\nand query budgets to guide future research in adversarial NLP. Based on our\nexperiments, we recommend greedy attacks with word importance ranking when\nunder a time constraint or attacking long inputs, and either beam search or\nparticle swarm optimization otherwise. Code implementation shared via\nhttps://github.com/QData/TextAttack-Search-Benchmark"", 'Learning text representation is crucial for text classification and other\nlanguage related tasks. There are a diverse set of text representation networks\nin the literature, and how to find the optimal one is a non-trivial problem.\nRecently, the emerging Neural Architecture Search (NAS) techniques have\ndemonstrated good potential to solve the problem. Nevertheless, most of the\nexisting works of NAS focus on the search algorithms and pay little attention\nto the search space. In this paper, we argue that the search space is also an\nimportant human prior to the success of NAS in different applications. Thus, we\npropose a novel search space tailored for text representation. Through\nautomatic search, the discovered network architecture outperforms\nstate-of-the-art models on various public datasets on text classification and\nnatural language inference tasks. Furthermore, some of the design principles\nfound in the automatic network agree well with human intuition.', 'One of the major challenges in training text-to-image generation models is\nthe need of a large number of high-quality image-text pairs. While image\nsamples are often easily accessible, the associated text descriptions typically\nrequire careful human captioning, which is particularly time- and\ncost-consuming. In this paper, we propose the first work to train text-to-image\ngeneration models without any text data. Our method leverages the well-aligned\nmulti-modal semantic space of the powerful pre-trained CLIP model: the\nrequirement of text-conditioning is seamlessly alleviated via generating text\nfeatures from image features. Extensive experiments are conducted to illustrate\nthe effectiveness of the proposed method. We obtain state-of-the-art results in\nthe standard text-to-image generation tasks. Importantly, the proposed\nlanguage-free model outperforms most existing models trained with full\nimage-text pairs. Furthermore, our method can be applied in fine-tuning\npre-trained models, which saves both training time and cost in training\ntext-to-image generation models. Our pre-trained model obtains competitive\nresults in zero-shot text-to-image generation on the MS-COCO dataset, yet with\naround only 1% of the model size and training data size relative to the\nrecently proposed large DALL-E model.', 'A concept for a novel CMOS image sensor suited for analog image\npre-processing is presented in this paper. As an example, an image restoration\nalgorithm for reducing image noise is applied as image pre-processing in the\nanalog domain. To supply low-latency data input for analog image preprocessing,\nthe proposed concept for a CMOS image sensor offers a new sensor signal\nacquisition method in 2D. In comparison to image pre-processing in the digital\ndomain, the proposed analog image pre-processing promises an improved image\nquality. Furthermore, the image noise at the stage of analog sensor signal\nacquisition can be used to select the most effective restoration algorithm\napplied to the analog circuit due to image processing prior to the A/D\nconverter.', 'Language model pre-training has been shown to capture a surprising amount of\nworld knowledge, crucial for NLP tasks such as question answering. However,\nthis knowledge is stored implicitly in the parameters of a neural network,\nrequiring ever-larger networks to cover more facts.\n  To capture knowledge in a more modular and interpretable way, we augment\nlanguage model pre-training with a latent knowledge retriever, which allows the\nmodel to retrieve and attend over documents from a large corpus such as\nWikipedia, used during pre-training, fine-tuning and inference. For the first\ntime, we show how to pre-train such a knowledge retriever in an unsupervised\nmanner, using masked language modeling as the learning signal and\nbackpropagating through a retrieval step that considers millions of documents.\n  We demonstrate the effectiveness of Retrieval-Augmented Language Model\npre-training (REALM) by fine-tuning on the challenging task of Open-domain\nQuestion Answering (Open-QA). We compare against state-of-the-art models for\nboth explicit and implicit knowledge storage on three popular Open-QA\nbenchmarks, and find that we outperform all previous methods by a significant\nmargin (4-16% absolute accuracy), while also providing qualitative benefits\nsuch as interpretability and modularity.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently working on a project aimed at developing a chatbot capable of handling open-domain Q&A for technical queries within specific domains. To fine-tune a language model within this domain, I require a suitable dataset for domain adaptation and model fine-tuning. My interest lies in a dataset comprising real-world questions from users on technical forums, as opposed to artificially generated queries. The dataset should ideally be sufficiently large to train a model. Additionally, I am seeking an auxiliary resource for pretraining and learning representations of domain-specific terminologies. This resource could potentially be an external knowledge base, such as Wikipedia. Any academic papers that propose or utilize such a dataset would be of great assistance to my research, as they would aid in the development of a more effective and accurate chatbot system. Furthermore, I am interested in standard methodologies for creating and curating such a dataset.","['With the rise of large-scale pre-trained language models, open-domain\nquestion-answering (ODQA) has become an important research topic in NLP. Based\non the popular pre-training fine-tuning approach, we posit that an additional\nin-domain pre-training stage using a large-scale, natural, and diverse\nquestion-answering (QA) dataset can be beneficial for ODQA. Consequently, we\npropose a novel QA dataset based on the Common Crawl project in this paper.\nUsing the readily available schema.org annotation, we extract around 130\nmillion multilingual question-answer pairs, including about 60 million English\ndata-points. With this previously unseen number of natural QA pairs, we\npre-train popular language models to show the potential of large-scale\nin-domain pre-training for the task of question-answering. In our experiments,\nwe find that pre-training question-answering models on our Common Crawl\nQuestion Answering dataset (CCQA) achieves promising results in zero-shot, low\nresource and fine-tuned settings across multiple tasks, models and benchmarks.', 'Domain adaptive pretraining, i.e. the continued unsupervised pretraining of a\nlanguage model on domain-specific text, improves the modelling of text for\ndownstream tasks within the domain. Numerous real-world applications are based\non domain-specific text, e.g. working with financial or biomedical documents,\nand these applications often need to support multiple languages. However,\nlarge-scale domain-specific multilingual pretraining data for such scenarios\ncan be difficult to obtain, due to regulations, legislation, or simply a lack\nof language- and domain-specific text. One solution is to train a single\nmultilingual model, taking advantage of the data available in as many languages\nas possible. In this work, we explore the benefits of domain adaptive\npretraining with a focus on adapting to multiple languages within a specific\ndomain. We propose different techniques to compose pretraining corpora that\nenable a language model to both become domain-specific and multilingual.\nEvaluation on nine domain-specific datasets-for biomedical named entity\nrecognition and financial sentence classification-covering seven different\nlanguages show that a single multilingual domain-specific model can outperform\nthe general multilingual model, and performs close to its monolingual\ncounterpart. This finding holds across two different pretraining methods,\nadapter-based pretraining and full model pretraining.', ""While self-supervised learning has made rapid advances in natural language\nprocessing, it remains unclear when researchers should engage in\nresource-intensive domain-specific pretraining (domain pretraining). The law,\npuzzlingly, has yielded few documented instances of substantial gains to domain\npretraining in spite of the fact that legal language is widely seen to be\nunique. We hypothesize that these existing results stem from the fact that\nexisting legal NLP tasks are too easy and fail to meet conditions for when\ndomain pretraining can help. To address this, we first present CaseHOLD (Case\nHoldings On Legal Decisions), a new dataset comprised of over 53,000+ multiple\nchoice questions to identify the relevant holding of a cited case. This dataset\npresents a fundamental task to lawyers and is both legally meaningful and\ndifficult from an NLP perspective (F1 of 0.4 with a BiLSTM baseline). Second,\nwe assess performance gains on CaseHOLD and existing legal NLP datasets. While\na Transformer architecture (BERT) pretrained on a general corpus (Google Books\nand Wikipedia) improves performance, domain pretraining (using corpus of\napproximately 3.5M decisions across all courts in the U.S. that is larger than\nBERT's) with a custom legal vocabulary exhibits the most substantial\nperformance gains with CaseHOLD (gain of 7.2% on F1, representing a 12%\nimprovement on BERT) and consistent performance gains across two other legal\ntasks. Third, we show that domain pretraining may be warranted when the task\nexhibits sufficient similarity to the pretraining corpus: the level of\nperformance increase in three legal tasks was directly tied to the domain\nspecificity of the task. Our findings inform when researchers should engage\nresource-intensive pretraining and show that Transformer-based architectures,\ntoo, learn embeddings suggestive of distinct legal language."", 'The goal of this work is to build conversational Question Answering (QA)\ninterfaces for the large body of domain-specific information available in FAQ\nsites. We present DoQA, a dataset with 2,437 dialogues and 10,917 QA pairs. The\ndialogues are collected from three Stack Exchange sites using the Wizard of Oz\nmethod with crowdsourcing. Compared to previous work, DoQA comprises\nwell-defined information needs, leading to more coherent and natural\nconversations with less factoid questions and is multi-domain. In addition, we\nintroduce a more realistic information retrieval(IR) scenario where the system\nneeds to find the answer in any of the FAQ documents. The results of an\nexisting, strong, system show that, thanks to transfer learning from a\nWikipedia QA dataset and fine tuning on a single FAQ domain, it is possible to\nbuild high quality conversational QA systems for FAQs without in-domain\ntraining data. The good results carry over into the more challenging IR\nscenario. In both cases, there is still ample room for improvement, as\nindicated by the higher human upperbound.', 'Question Answering (QA) is a natural language processing task that aims at\nobtaining relevant answers to user questions. While some progress has been made\nin this area, biomedical questions are still a challenge to most QA approaches,\ndue to the complexity of the domain and limited availability of training sets.\nWe present a method to automatically extract question-article pairs from Q\\&A\nweb forums, which can be used for document retrieval, a crucial step of most QA\nsystems. The proposed framework extracts from selected forums the questions and\nthe respective answers that contain citations. This way, QA systems based on\ndocument retrieval can be developed and evaluated using the question-article\npairs annotated by users of these forums. We generated the BiQA corpus by\napplying our framework to three forums, obtaining 7,453 questions and 14,239\nquestion-article pairs. We evaluated how the number of articles associated with\neach question and the number of votes on each answer affects the performance of\nbaseline document retrieval approaches. Also, we demonstrated that the articles\ngiven as answers are significantly similar to the questions and trained a\nstate-of-the-art deep learning model that obtained similar performance to using\na dataset manually annotated by experts. The proposed framework can be used to\nupdate the BiQA corpus from the same forums as new posts are made, and from\nother forums that support their answers with documents. The BiQA corpus and the\nframework used to generate it are available at\n\\url{https://github.com/lasigeBioTM/BiQA}.']","['Domain-specific community question answering is becoming an integral part of\nprofessions. Finding related questions and answers in these communities can\nsignificantly improve the effectiveness and efficiency of information seeking.\nStack Overflow is one of the most popular communities that is being used by\nmillions of programmers. In this paper, we analyze the problem of predicting\nknowledge unit (question thread) relatedness in Stack Overflow. In particular,\nwe formulate the question relatedness task as a multi-class classification\nproblem with four degrees of relatedness. We present a large-scale dataset with\nmore than 300K pairs. To the best of our knowledge, this dataset is the largest\ndomain-specific dataset for Question-Question relatedness. We present the steps\nthat we took to collect, clean, process, and assure the quality of the dataset.\nThe proposed dataset Stack Overflow is a useful resource to develop novel\nsolutions, specifically data-hungry neural network models, for the prediction\nof relatedness in technical community question-answering forums. We adopt a\nneural network architecture and a traditional model for this task that\neffectively utilize information from different parts of knowledge units to\ncompute the relatedness between them. These models can be used to benchmark\nnovel models, as they perform well in our task and in a closely similar task.', 'The recent success of neural machine translation models relies on the\navailability of high quality, in-domain data. Domain adaptation is required\nwhen domain-specific data is scarce or nonexistent. Previous unsupervised\ndomain adaptation strategies include training the model with in-domain copied\nmonolingual or back-translated data. However, these methods use generic\nrepresentations for text regardless of domain shift, which makes it infeasible\nfor translation models to control outputs conditional on a specific domain. In\nthis work, we propose an approach that adapts models with domain-aware feature\nembeddings, which are learned via an auxiliary language modeling task. Our\napproach allows the model to assign domain-specific representations to words\nand output sentences in the desired domain. Our empirical results demonstrate\nthe effectiveness of the proposed strategy, achieving consistent improvements\nin multiple experimental settings. In addition, we show that combining our\nmethod with back translation can further improve the performance of the model.', 'Question answering has seen significant advances in recent times, especially\nwith the introduction of increasingly bigger transformer-based models\npre-trained on massive amounts of data. While achieving impressive results on\nmany benchmarks, their performances appear to be proportional to the amount of\ntraining data available in the target domain. In this work, we explore the\nability of current question-answering models to generalize - to both other\ndomains as well as with restricted training data. We find that large amounts of\ntraining data are necessary, both for pre-training as well as fine-tuning to a\ntask, for the models to perform well on the designated task. We introduce a\nnovel abductive reasoning approach based on Grenander\'s Pattern Theory\nframework to provide self-supervised domain adaptation cues or ""pseudo-labels,""\nwhich can be used instead of expensive human annotations. The proposed\nself-supervised training regimen allows for effective domain adaptation without\nlosing performance compared to fully supervised baselines. Extensive\nexperiments on two publicly available benchmarks show the efficacy of the\nproposed approach. We show that neural networks models trained using\nself-labeled data can retain up to $75\\%$ of the performance of models trained\non large amounts of human-annotated training data.', 'Question Answering (QA) is in increasing demand as the amount of information\navailable online and the desire for quick access to this content grows. A\ncommon approach to QA has been to fine-tune a pretrained language model on a\ntask-specific labeled dataset. This paradigm, however, relies on scarce, and\ncostly to obtain, large-scale human-labeled data. We propose an unsupervised\napproach to training QA models with generated pseudo-training data. We show\nthat generating questions for QA training by applying a simple template on a\nrelated, retrieved sentence rather than the original context sentence improves\ndownstream QA performance by allowing the model to learn more complex\ncontext-question relationships. Training a QA model on this data gives a\nrelative improvement over a previous unsupervised model in F1 score on the\nSQuAD dataset by about 14%, and 20% when the answer is a named entity,\nachieving state-of-the-art performance on SQuAD for unsupervised QA.', 'This paper introduces QAConv, a new question answering (QA) dataset that uses\nconversations as a knowledge source. We focus on informative conversations,\nincluding business emails, panel discussions, and work channels. Unlike\nopen-domain and task-oriented dialogues, these conversations are usually long,\ncomplex, asynchronous, and involve strong domain knowledge. In total, we\ncollect 34,608 QA pairs from 10,259 selected conversations with both\nhuman-written and machine-generated questions. We use a question generator and\na dialogue summarizer as auxiliary tools to collect and recommend questions.\nThe dataset has two testing scenarios: chunk mode and full mode, depending on\nwhether the grounded partial conversation is provided or retrieved.\nExperimental results show that state-of-the-art pretrained QA systems have\nlimited zero-shot performance and tend to predict our questions as\nunanswerable. Our dataset provides a new training and evaluation testbed to\nfacilitate QA on conversations research.', 'In a conversational question answering scenario, a questioner seeks to\nextract information about a topic through a series of interdependent questions\nand answers. As the conversation progresses, they may switch to related topics,\na phenomenon commonly observed in information-seeking search sessions. However,\ncurrent datasets for conversational question answering are limiting in two\nways: 1) they do not contain topic switches; and 2) they assume the reference\ntext for the conversation is given, i.e., the setting is not open-domain. We\nintroduce TopiOCQA (pronounced Tapioca), an open-domain conversational dataset\nwith topic switches on Wikipedia. TopiOCQA contains 3,920 conversations with\ninformation-seeking questions and free-form answers. On average, a conversation\nin our dataset spans 13 question-answer turns and involves four topics\n(documents). TopiOCQA poses a challenging test-bed for models, where efficient\nretrieval is required on multiple turns of the same conversation, in\nconjunction with constructing valid responses using conversational history. We\nevaluate several baselines, by combining state-of-the-art document retrieval\nmethods with neural reader models. Our best model achieves F1 of 55.8, falling\nshort of human performance by 14.2 points, indicating the difficulty of our\ndataset. Our dataset and code is available at\nhttps://mcgill-nlp.github.io/topiocqa', 'Most real world language problems require learning from heterogenous corpora,\nraising the problem of learning robust models which generalise well to both\nsimilar (in domain) and dissimilar (out of domain) instances to those seen in\ntraining. This requires learning an underlying task, while not learning\nirrelevant signals and biases specific to individual domains. We propose a\nnovel method to optimise both in- and out-of-domain accuracy based on joint\nlearning of a structured neural model with domain-specific and domain-general\ncomponents, coupled with adversarial training for domain. Evaluating on\nmulti-domain language identification and multi-domain sentiment analysis, we\nshow substantial improvements over standard domain adaptation techniques, and\ndomain-adversarial training.', 'Conversational modeling is an important task in natural language\nunderstanding and machine intelligence. Although previous approaches exist,\nthey are often restricted to specific domains (e.g., booking an airline ticket)\nand require hand-crafted rules. In this paper, we present a simple approach for\nthis task which uses the recently proposed sequence to sequence framework. Our\nmodel converses by predicting the next sentence given the previous sentence or\nsentences in a conversation. The strength of our model is that it can be\ntrained end-to-end and thus requires much fewer hand-crafted rules. We find\nthat this straightforward model can generate simple conversations given a large\nconversational training dataset. Our preliminary results suggest that, despite\noptimizing the wrong objective function, the model is able to converse well. It\nis able extract knowledge from both a domain specific dataset, and from a\nlarge, noisy, and general domain dataset of movie subtitles. On a\ndomain-specific IT helpdesk dataset, the model can find a solution to a\ntechnical problem via conversations. On a noisy open-domain movie transcript\ndataset, the model can perform simple forms of common sense reasoning. As\nexpected, we also find that the lack of consistency is a common failure mode of\nour model.', 'Learning high-quality domain word embeddings is important for achieving good\nperformance in many NLP tasks. General-purpose embeddings trained on\nlarge-scale corpora are often sub-optimal for domain-specific applications.\nHowever, domain-specific tasks often do not have large in-domain corpora for\ntraining high-quality domain embeddings. In this paper, we propose a novel\nlifelong learning setting for domain embedding. That is, when performing the\nnew domain embedding, the system has seen many past domains, and it tries to\nexpand the new in-domain corpus by exploiting the corpora from the past domains\nvia meta-learning. The proposed meta-learner characterizes the similarities of\nthe contexts of the same word in many domain corpora, which helps retrieve\nrelevant data from the past domains to expand the new domain corpus.\nExperimental results show that domain embeddings produced from such a process\nimprove the performance of the downstream tasks.', ""Language models pretrained on text from a wide variety of sources form the\nfoundation of today's NLP. In light of the success of these broad-coverage\nmodels, we investigate whether it is still helpful to tailor a pretrained model\nto the domain of a target task. We present a study across four domains\n(biomedical and computer science publications, news, and reviews) and eight\nclassification tasks, showing that a second phase of pretraining in-domain\n(domain-adaptive pretraining) leads to performance gains, under both high- and\nlow-resource settings. Moreover, adapting to the task's unlabeled data\n(task-adaptive pretraining) improves performance even after domain-adaptive\npretraining. Finally, we show that adapting to a task corpus augmented using\nsimple data selection strategies is an effective alternative, especially when\nresources for domain-adaptive pretraining might be unavailable. Overall, we\nconsistently find that multi-phase adaptive pretraining offers large gains in\ntask performance."", 'Alongside huge volumes of research on deep learning models in NLP in the\nrecent years, there has been also much work on benchmark datasets needed to\ntrack modeling progress. Question answering and reading comprehension have been\nparticularly prolific in this regard, with over 80 new datasets appearing in\nthe past two years. This study is the largest survey of the field to date. We\nprovide an overview of the various formats and domains of the current\nresources, highlighting the current lacunae for future work. We further discuss\nthe current classifications of ""skills"" that question answering/reading\ncomprehension systems are supposed to acquire, and propose a new taxonomy. The\nsupplementary materials survey the current multilingual resources and\nmonolingual resources for languages other than English, and we discuss the\nimplications of over-focusing on English. The study is aimed at both\npractitioners looking for pointers to the wealth of existing data, and at\nresearchers working on new resources.', 'A long-term goal of machine learning is to build intelligent conversational\nagents. One recent popular approach is to train end-to-end models on a large\namount of real dialog transcripts between humans (Sordoni et al., 2015; Vinyals\n& Le, 2015; Shang et al., 2015). However, this approach leaves many questions\nunanswered as an understanding of the precise successes and shortcomings of\neach model is hard to assess. A contrasting recent proposal are the bAbI tasks\n(Weston et al., 2015b) which are synthetic data that measure the ability of\nlearning machines at various reasoning tasks over toy language. Unfortunately,\nthose tests are very small and hence may encourage methods that do not scale.\nIn this work, we propose a suite of new tasks of a much larger scale that\nattempt to bridge the gap between the two regimes. Choosing the domain of\nmovies, we provide tasks that test the ability of models to answer factual\nquestions (utilizing OMDB), provide personalization (utilizing MovieLens),\ncarry short conversations about the two, and finally to perform on natural\ndialogs from Reddit. We provide a dataset covering 75k movie entities and with\n3.5M training examples. We present results of various models on these tasks,\nand evaluate their performance.', 'Question answering and conversational systems are often baffled and need help\nclarifying certain ambiguities. However, limitations of existing datasets\nhinder the development of large-scale models capable of generating and\nutilising clarification questions. In order to overcome these limitations, we\ndevise a novel bootstrapping framework (based on self-supervision) that assists\nin the creation of a diverse, large-scale dataset of clarification questions\nbased on post-comment tuples extracted from stackexchange. The framework\nutilises a neural network based architecture for classifying clarification\nquestions. It is a two-step method where the first aims to increase the\nprecision of the classifier and second aims to increase its recall. We\nquantitatively demonstrate the utility of the newly created dataset by applying\nit to the downstream task of question-answering. The final dataset, ClarQ,\nconsists of ~2M examples distributed across 173 domains of stackexchange. We\nrelease this dataset in order to foster research into the field of\nclarification question generation with the larger goal of enhancing dialog and\nquestion answering systems.', 'In the last few years, open-domain question answering (ODQA) has advanced\nrapidly due to the development of deep learning techniques and the availability\nof large-scale QA datasets. However, the current datasets are essentially\ndesigned for synchronic document collections (e.g., Wikipedia). Temporal news\ncollections such as long-term news archives spanning several decades, are\nrarely used in training the models despite they are quite valuable for our\nsociety. To foster the research in the field of ODQA on such historical\ncollections, we present ArchivalQA, a large question answering dataset\nconsisting of 532,444 question-answer pairs which is designed for temporal news\nQA. We divide our dataset into four subparts based on the question difficulty\nlevels and the containment of temporal expressions, which we believe are useful\nfor training and testing ODQA systems characterized by different strengths and\nabilities. The novel QA dataset-constructing framework that we introduce can be\nalso applied to generate non-ambiguous questions of good quality over other\ntypes of temporal document collections.', ""Natural language dialogue systems raise great attention recently. As many\ndialogue models are data-driven, high-quality datasets are essential to these\nsystems. In this paper, we introduce Pchatbot, a large-scale dialogue dataset\nthat contains two subsets collected from Weibo and Judicial forums\nrespectively. To adapt the raw dataset to dialogue systems, we elaborately\nnormalize the raw dataset via processes such as anonymization, deduplication,\nsegmentation, and filtering. The scale of Pchatbot is significantly larger than\nexisting Chinese datasets, which might benefit the data-driven models. Besides,\ncurrent dialogue datasets for personalized chatbot usually contain several\npersona sentences or attributes. Different from existing datasets, Pchatbot\nprovides anonymized user IDs and timestamps for both posts and responses. This\nenables the development of personalized dialogue models that directly learn\nimplicit user personality from the user's dialogue history. Our preliminary\nexperimental study benchmarks several state-of-the-art dialogue models to\nprovide a comparison for future work. The dataset can be publicly accessed at\nGithub."", ""Many users communicate with chatbots and AI assistants in order to help them\nwith various tasks. A key component of the assistant is the ability to\nunderstand and answer a user's natural language questions for\nquestion-answering (QA). Because data can be usually stored in a structured\nmanner, an essential step involves turning a natural language question into its\ncorresponding query language. However, in order to train most natural\nlanguage-to-query-language state-of-the-art models, a large amount of training\ndata is needed first. In most domains, this data is not available and\ncollecting such datasets for various domains can be tedious and time-consuming.\nIn this work, we propose a novel method for accelerating the training dataset\ncollection for developing the natural language-to-query-language machine\nlearning models. Our system allows one to generate conversational multi-term\ndata, where multiple turns define a dialogue session, enabling one to better\nutilize chatbot interfaces. We train two current state-of-the-art NL-to-QL\nmodels, on both an SQL and SPARQL-based datasets in order to showcase the\nadaptability and efficacy of our created data."", 'Expert-layman text style transfer technologies have the potential to improve\ncommunication between members of scientific communities and the general public.\nHigh-quality information produced by experts is often filled with difficult\njargon laypeople struggle to understand. This is a particularly notable issue\nin the medical domain, where layman are often confused by medical text online.\nAt present, two bottlenecks interfere with the goal of building high-quality\nmedical expert-layman style transfer systems: a dearth of pretrained\nmedical-domain language models spanning both expert and layman terminologies\nand a lack of parallel corpora for training the transfer task itself. To\nmitigate the first issue, we propose a novel language model (LM) pretraining\ntask, Knowledge Base Assimilation, to synthesize pretraining data from the\nedges of a graph of expert- and layman-style medical terminology terms into an\nLM during self-supervised learning. To mitigate the second issue, we build a\nlarge-scale parallel corpus in the medical expert-layman domain using a\nmargin-based criterion. Our experiments show that transformer-based models\npretrained on knowledge base assimilation and other well-established\npretraining tasks fine-tuning on our new parallel corpus leads to considerable\nimprovement against expert-layman transfer benchmarks, gaining an average\nrelative improvement of our human evaluation, the Overall Success Rate (OSR),\nby 106%. We release our code and parallel corpus for future research.', 'While most previous work has focused on different pretraining objectives and\narchitectures for transfer learning, we ask how to best adapt the pretrained\nmodel to a given target task. We focus on the two most common forms of\nadaptation, feature extraction (where the pretrained weights are frozen), and\ndirectly fine-tuning the pretrained model. Our empirical results across diverse\nNLP tasks with two state-of-the-art models show that the relative performance\nof fine-tuning vs. feature extraction depends on the similarity of the\npretraining and target tasks. We explore possible explanations for this finding\nand provide a set of adaptation guidelines for the NLP practitioner.', 'The meaning of a word often varies depending on its usage in different\ndomains. The standard word embedding models struggle to represent this\nvariation, as they learn a single global representation for a word. We propose\na method to learn domain-specific word embeddings, from text organized into\nhierarchical domains, such as reviews in an e-commerce website, where products\nfollow a taxonomy. Our structured probabilistic model allows vector\nrepresentations for the same word to drift away from each other for distant\ndomains in the taxonomy, to accommodate its domain-specific meanings. By\nlearning sets of domain-specific word representations jointly, our model can\nleverage domain relationships, and it scales well with the number of domains.\nUsing large real-world review datasets, we demonstrate the effectiveness of our\nmodel compared to state-of-the-art approaches, in learning domain-specific word\nembeddings that are both intuitive to humans and benefit downstream NLP tasks.', 'Neural machine translation often adopts the fine-tuning approach to adapt to\nspecific domains. However, nonrestricted fine-tuning can easily degrade on the\ngeneral domain and over-fit to the target domain. To mitigate the issue, we\npropose Prune-Tune, a novel domain adaptation method via gradual pruning. It\nlearns tiny domain-specific sub-networks during fine-tuning on new domains.\nPrune-Tune alleviates the over-fitting and the degradation problem without\nmodel modification. Furthermore, Prune-Tune is able to sequentially learn a\nsingle network with multiple disjoint domain-specific sub-networks for multiple\ndomains. Empirical experiment results show that Prune-Tune outperforms several\nstrong competitors in the target domain test set without sacrificing the\nquality on the general domain in both single and multi-domain settings. The\nsource code and data are available at https://github.com/ohlionel/Prune-Tune.', 'Contextual embedding-based language models trained on large data sets, such\nas BERT and RoBERTa, provide strong performance across a wide range of tasks\nand are ubiquitous in modern NLP. It has been observed that fine-tuning these\nmodels on tasks involving data from domains different from that on which they\nwere pretrained can lead to suboptimal performance. Recent work has explored\napproaches to adapt pretrained language models to new domains by incorporating\nadditional pretraining using domain-specific corpora and task data. We propose\nan alternative approach for transferring pretrained language models to new\ndomains by adapting their tokenizers. We show that domain-specific subword\nsequences can be efficiently determined directly from divergences in the\nconditional token distributions of the base and domain-specific corpora. In\ndatasets from four disparate domains, we find adaptive tokenization on a\npretrained RoBERTa model provides >97% of the performance benefits of domain\nspecific pretraining. Our approach produces smaller models and less training\nand inference time than other approaches using tokenizer augmentation. While\nadaptive tokenization incurs a 6% increase in model parameters in our\nexperimentation, due to the introduction of 10k new domain-specific tokens, our\napproach, using 64 vCPUs, is 72x faster than further pretraining the language\nmodel on domain-specific corpora on 8 TPUs.', 'Domain Adaptation arises when we aim at learning from source domain a model\nthat can per- form acceptably well on a different target domain. It is\nespecially crucial for Natural Language Generation (NLG) in Spoken Dialogue\nSystems when there are sufficient annotated data in the source domain, but\nthere is a limited labeled data in the target domain. How to effectively\nutilize as much of existing abilities from source domains is a crucial issue in\ndomain adaptation. In this paper, we propose an adversarial training procedure\nto train a Variational encoder-decoder based language generator via multiple\nadaptation steps. In this procedure, a model is first trained on a source\ndomain data and then fine-tuned on a small set of target domain utterances\nunder the guidance of two proposed critics. Experimental results show that the\nproposed method can effec- tively leverage the existing knowledge in the source\ndomain to adapt to another related domain by using only a small amount of\nin-domain data.', 'We propose a simple domain adaptation method for neural networks in a\nsupervised setting. Supervised domain adaptation is a way of improving the\ngeneralization performance on the target domain by using the source domain\ndataset, assuming that both of the datasets are labeled. Recently, recurrent\nneural networks have been shown to be successful on a variety of NLP tasks such\nas caption generation; however, the existing domain adaptation techniques are\nlimited to (1) tune the model parameters by the target dataset after the\ntraining by the source dataset, or (2) design the network to have dual output,\none for the source domain and the other for the target domain. Reformulating\nthe idea of the domain adaptation technique proposed by Daume (2007), we\npropose a simple domain adaptation method, which can be applied to neural\nnetworks trained with a cross-entropy loss. On captioning datasets, we show\nperformance improvements over other domain adaptation methods.', ""This work presents biomedical and clinical language models for Spanish by\nexperimenting with different pretraining choices, such as masking at word and\nsubword level, varying the vocabulary size and testing with domain data,\nlooking for better language representations. Interestingly, in the absence of\nenough clinical data to train a model from scratch, we applied mixed-domain\npretraining and cross-domain transfer approaches to generate a performant\nbio-clinical model suitable for real-world clinical data. We evaluated our\nmodels on Named Entity Recognition (NER) tasks for biomedical documents and\nchallenging hospital discharge reports. When compared against the competitive\nmBERT and BETO models, we outperform them in all NER tasks by a significant\nmargin. Finally, we studied the impact of the model's vocabulary on the NER\nperformances by offering an interesting vocabulary-centric analysis. The\nresults confirm that domain-specific pretraining is fundamental to achieving\nhigher performances in downstream NER tasks, even within a mid-resource\nscenario. To the best of our knowledge, we provide the first biomedical and\nclinical transformer-based pretrained language models for Spanish, intending to\nboost native Spanish NLP applications in biomedicine. Our best models are\nfreely available in the HuggingFace hub: https://huggingface.co/BSC-TeMU."", 'Building automatic technical support system is an important yet challenge\ntask. Conceptually, to answer a user question on a technical forum, a human\nexpert has to first retrieve relevant documents, and then read them carefully\nto identify the answer snippet. Despite huge success the researchers have\nachieved in coping with general domain question answering (QA), much less\nattentions have been paid for investigating technical QA. Specifically,\nexisting methods suffer from several unique challenges (i) the question and\nanswer rarely overlaps substantially and (ii) very limited data size. In this\npaper, we propose a novel framework of deep transfer learning to effectively\naddress technical QA across tasks and domains. To this end, we present an\nadjustable joint learning approach for document retrieval and reading\ncomprehension tasks. Our experiments on the TechQA demonstrates superior\nperformance compared with state-of-the-art methods.', 'Research in question answering datasets and models has gained a lot of\nattention in the research community. Many of them release their own question\nanswering datasets as well as the models. There is tremendous progress that we\nhave seen in this area of research. The aim of this survey is to recognize,\nsummarize and analyze the existing datasets that have been released by many\nresearchers, especially in non-English datasets as well as resources such as\nresearch code, and evaluation metrics. In this paper, we review question\nanswering datasets that are available in common languages other than English\nsuch as French, German, Japanese, Chinese, Arabic, Russian, as well as the\nmultilingual and cross-lingual question-answering datasets.', 'One of the difficulties in training dialogue systems is the lack of training\ndata. We explore the possibility of creating dialogue data through the\ninteraction between a dialogue system and a user simulator. Our goal is to\ndevelop a modelling framework that can incorporate new dialogue scenarios\nthrough self-play between the two agents. In this framework, we first pre-train\nthe two agents on a collection of source domain dialogues, which equips the\nagents to converse with each other via natural language. With further\nfine-tuning on a small amount of target domain data, the agents continue to\ninteract with the aim of improving their behaviors using reinforcement learning\nwith structured reward functions. In experiments on the MultiWOZ dataset, two\npractical transfer learning problems are investigated: 1) domain adaptation and\n2) single-to-multiple domain transfer. We demonstrate that the proposed\nframework is highly effective in bootstrapping the performance of the two\nagents in transfer learning. We also show that our method leads to improvements\nin dialogue system performance on complete datasets.', 'Numerous new dialog domains are being created every day while collecting data\nfor these domains is extremely costly since it involves human interactions.\nTherefore, it is essential to develop algorithms that can adapt to different\ndomains efficiently when building data-driven dialog models. The most recent\nresearches on domain adaption focus on giving the model a better\ninitialization, rather than optimizing the adaptation process. We propose an\nefficient domain adaptive task-oriented dialog system model, which incorporates\na meta-teacher model to emphasize the different impacts between generated\ntokens with respect to the context. We first train our base dialog model and\nmeta-teacher model adversarially in a meta-learning setting on rich-resource\ndomains. The meta-teacher learns to quantify the importance of tokens under\ndifferent contexts across different domains. During adaptation, the\nmeta-teacher guides the dialog model to focus on important tokens in order to\nachieve better adaptation efficiency. We evaluate our model on two multi-domain\ndatasets, MultiWOZ and Google Schema-Guided Dialogue, and achieve\nstate-of-the-art performance.', 'To bridge the gap between the capabilities of the state-of-the-art in factoid\nquestion answering (QA) and what users ask, we need large datasets of real user\nquestions that capture the various question phenomena users are interested in,\nand the diverse ways in which these questions are formulated. We introduce\nComQA, a large dataset of real user questions that exhibit different\nchallenging aspects such as compositionality, temporal reasoning, and\ncomparisons. ComQA questions come from the WikiAnswers community QA platform,\nwhich typically contains questions that are not satisfactorily answerable by\nexisting search engine technology. Through a large crowdsourcing effort, we\nclean the question dataset, group questions into paraphrase clusters, and\nannotate clusters with their answers. ComQA contains 11,214 questions grouped\ninto 4,834 paraphrase clusters. We detail the process of constructing ComQA,\nincluding the measures taken to ensure its high quality while making effective\nuse of crowdsourcing. We also present an extensive analysis of the dataset and\nthe results achieved by state-of-the-art systems on ComQA, demonstrating that\nour dataset can be a driver of future research on QA.', 'Recently multi-domain recommender systems have received much attention from\nresearchers because they can solve cold-start problem as well as support for\ncross-selling. However, when applying into multi-domain items, although\nalgorithms specifically addressing a single domain have many difficulties in\ncapturing the specific characteristics of each domain, multi-domain algorithms\nhave less opportunity to obtain similar features among domains. Because both\nsimilarities and differences exist among domains, multi-domain models must\ncapture both to achieve good performance. Other studies of multi-domain systems\nmerely transfer knowledge from the source domain to the target domain, so the\nsource domain usually comes from external factors such as the search query or\nsocial network, which is sometimes impossible to obtain. To handle the two\nproblems, we propose a model that can extract both homogeneous and divergent\nfeatures among domains and extract data in a domain can support for other\ndomain equally: a so-called Domain-to-Domain Translation Model (D2D-TM). It is\nbased on generative adversarial networks (GANs), Variational Autoencoders\n(VAEs), and Cycle-Consistency (CC) for weight-sharing. We use the user\ninteraction history of each domain as input and extract latent features through\na VAE-GAN-CC network. Experiments underscore the effectiveness of the proposed\nsystem over state-of-the-art methods by a large margin.', 'Contextualized word embeddings such as ELMo and BERT provide a foundation for\nstrong performance across a wide range of natural language processing tasks by\npretraining on large corpora of unlabeled text. However, the applicability of\nthis approach is unknown when the target domain varies substantially from the\npretraining corpus. We are specifically interested in the scenario in which\nlabeled data is available in only a canonical source domain such as newstext,\nand the target domain is distinct from both the labeled and pretraining texts.\nTo address this scenario, we propose domain-adaptive fine-tuning, in which the\ncontextualized embeddings are adapted by masked language modeling on text from\nthe target domain. We test this approach on sequence labeling in two\nchallenging domains: Early Modern English and Twitter. Both domains differ\nsubstantially from existing pretraining corpora, and domain-adaptive\nfine-tuning yields substantial improvements over strong BERT baselines, with\nparticularly impressive results on out-of-vocabulary words. We conclude that\ndomain-adaptive fine-tuning offers a simple and effective approach for the\nunsupervised adaptation of sequence labeling to difficult new domains.', 'Domain adaptation of neural networks commonly relies on three training\nphases: pretraining, selected data training and then fine tuning. Data\nselection improves target domain generalization by training further on\npretraining data identified by relying on a small sample of target domain data.\nThis work examines the benefit of data selection for language modeling and\nmachine translation. Our experiments assess the complementarity of selection\nwith fine tuning and result in practical recommendations: (i) selected data\nmust be similar to the fine-tuning domain but not so much as to erode the\ncomplementary effect of fine-tuning; (ii) there is a trade-off between\nselecting little data for fast but limited progress or much data for slow but\nlong lasting progress; (iii) data selection can be applied early during\npretraining, with performance gains comparable to long pretraining session;\n(iv) data selection from domain classifiers is often more effective than the\npopular contrastive data selection method.', 'The key challenge of multi-domain translation lies in simultaneously encoding\nboth the general knowledge shared across domains and the particular knowledge\ndistinctive to each domain in a unified model. Previous work shows that the\nstandard neural machine translation (NMT) model, trained on mixed-domain data,\ngenerally captures the general knowledge, but misses the domain-specific\nknowledge. In response to this problem, we augment NMT model with additional\ndomain transformation networks to transform the general representations to\ndomain-specific representations, which are subsequently fed to the NMT decoder.\nTo guarantee the knowledge transformation, we also propose two complementary\nsupervision signals by leveraging the power of knowledge distillation and\nadversarial learning. Experimental results on several language pairs, covering\nboth balanced and unbalanced multi-domain translation, demonstrate the\neffectiveness and universality of the proposed approach. Encouragingly, the\nproposed unified model achieves comparable results with the fine-tuning\napproach that requires multiple models to preserve the particular knowledge.\nFurther analyses reveal that the domain transformation networks successfully\ncapture the domain-specific knowledge as expected.', 'Named entity recognition (NER) for identifying proper nouns in unstructured\ntext is one of the most important and fundamental tasks in natural language\nprocessing. However, despite the widespread use of NER models, they still\nrequire a large-scale labeled data set, which incurs a heavy burden due to\nmanual annotation. Domain adaptation is one of the most promising solutions to\nthis problem, where rich labeled data from the relevant source domain are\nutilized to strengthen the generalizability of a model based on the target\ndomain. However, the mainstream cross-domain NER models are still affected by\nthe following two challenges (1) Extracting domain-invariant information such\nas syntactic information for cross-domain transfer. (2) Integrating\ndomain-specific information such as semantic information into the model to\nimprove the performance of NER. In this study, we present a semi-supervised\nframework for transferable NER, which disentangles the domain-invariant latent\nvariables and domain-specific latent variables. In the proposed framework, the\ndomain-specific information is integrated with the domain-specific latent\nvariables by using a domain predictor. The domain-specific and domain-invariant\nlatent variables are disentangled using three mutual information regularization\nterms, i.e., maximizing the mutual information between the domain-specific\nlatent variables and the original embedding, maximizing the mutual information\nbetween the domain-invariant latent variables and the original embedding, and\nminimizing the mutual information between the domain-specific and\ndomain-invariant latent variables. Extensive experiments demonstrated that our\nmodel can obtain state-of-the-art performance with cross-domain and\ncross-lingual NER benchmark data sets.', 'One weakness of machine-learned NLP models is that they typically perform\npoorly on out-of-domain data. In this work, we study the task of identifying\nproducts being bought and sold in online cybercrime forums, which exhibits\nparticularly challenging cross-domain effects. We formulate a task that\nrepresents a hybrid of slot-filling information extraction and named entity\nrecognition and annotate data from four different forums. Each of these forums\nconstitutes its own ""fine-grained domain"" in that the forums cover different\nmarket sectors with different properties, even though all forums are in the\nbroad domain of cybercrime. We characterize these domain differences in the\ncontext of a learning-based system: supervised models see decreased accuracy\nwhen applied to new forums, and standard techniques for semi-supervised\nlearning and domain adaptation have limited effectiveness on this data, which\nsuggests the need to improve these techniques. We release a dataset of 1,938\nannotated posts from across the four forums.', 'Linking clinical narratives to standardized vocabularies and coding systems\nis a key component of unlocking the information in medical text for analysis.\nHowever, many domains of medical concepts lack well-developed terminologies\nthat can support effective coding of medical text. We present a framework for\ndeveloping natural language processing (NLP) technologies for automated coding\nof under-studied types of medical information, and demonstrate its\napplicability via a case study on physical mobility function. Mobility is a\ncomponent of many health measures, from post-acute care and surgical outcomes\nto chronic frailty and disability, and is coded in the International\nClassification of Functioning, Disability, and Health (ICF). However, mobility\nand other types of functional activity remain under-studied in medical\ninformatics, and neither the ICF nor commonly-used medical terminologies\ncapture functional status terminology in practice. We investigated two\ndata-driven paradigms, classification and candidate selection, to link\nnarrative observations of mobility to standardized ICF codes, using a dataset\nof clinical narratives from physical therapy encounters. Recent advances in\nlanguage modeling and word embedding were used as features for established\nmachine learning models and a novel deep learning approach, achieving a macro\nF-1 score of 84% on linking mobility activity reports to ICF codes. Both\nclassification and candidate selection approaches present distinct strengths\nfor automated coding in under-studied domains, and we highlight that the\ncombination of (i) a small annotated data set; (ii) expert definitions of codes\nof interest; and (iii) a representative text corpus is sufficient to produce\nhigh-performing automated coding systems. This study has implications for the\nongoing growth of NLP tools for a variety of specialized applications in\nclinical care and research.', ""Neural rankers based on deep pretrained language models (LMs) have been shown\nto improve many information retrieval benchmarks. However, these methods are\naffected by their the correlation between pretraining domain and target domain\nand rely on massive fine-tuning relevance labels. Directly applying pretraining\nmethods to specific domains may result in suboptimal search quality because\nspecific domains may have domain adaption problems, such as the COVID domain.\nThis paper presents a search system to alleviate the special domain adaption\nproblem. The system utilizes the domain-adaptive pretraining and few-shot\nlearning technologies to help neural rankers mitigate the domain discrepancy\nand label scarcity problems. Besides, we also integrate dense retrieval to\nalleviate traditional sparse retrieval's vocabulary mismatch obstacle. Our\nsystem performs the best among the non-manual runs in Round 2 of the TREC-COVID\ntask, which aims to retrieve useful information from scientific literature\nrelated to COVID-19. Our code is publicly available at\nhttps://github.com/thunlp/OpenMatch."", 'Numerous recent works have proposed pretraining generic visio-linguistic\nrepresentations and then finetuning them for downstream vision and language\ntasks. While architecture and objective function design choices have received\nattention, the choice of pretraining datasets has received little attention. In\nthis work, we question some of the default choices made in literature. For\ninstance, we systematically study how varying similarity between the\npretraining dataset domain (textual and visual) and the downstream domain\naffects performance. Surprisingly, we show that automatically generated data in\na domain closer to the downstream task (e.g., VQA v2) is a better choice for\npretraining than ""natural"" data but of a slightly different domain (e.g.,\nConceptual Captions). On the other hand, some seemingly reasonable choices of\npretraining datasets were found to be entirely ineffective for some downstream\ntasks. This suggests that despite the numerous recent efforts, vision &\nlanguage pretraining does not quite work ""out of the box"" yet. Overall, as a\nby-product of our study, we find that simple design choices in pretraining can\nhelp us achieve close to state-of-art results on downstream tasks without any\narchitectural changes.', 'In this paper, we propose a novel domain adaptation method named ""mixed fine\ntuning"" for neural machine translation (NMT). We combine two existing\napproaches namely fine tuning and multi domain NMT. We first train an NMT model\non an out-of-domain parallel corpus, and then fine tune it on a parallel corpus\nwhich is a mix of the in-domain and out-of-domain corpora. All corpora are\naugmented with artificial tags to indicate specific domains. We empirically\ncompare our proposed method against fine tuning and multi domain methods and\ndiscuss its benefits and shortcomings.', 'Spoken language understanding has been addressed as a supervised learning\nproblem, where a set of training data is available for each domain. However,\nannotating data for each domain is both financially costly and non-scalable so\nwe should fully utilize information across all domains. One existing approach\nsolves the problem by conducting multi-domain learning, using shared parameters\nfor joint training across domains. We propose to improve the parameterization\nof this method by using domain-specific and task-specific model parameters to\nimprove knowledge learning and transfer. Experiments on 5 domains show that our\nmodel is more effective for multi-domain SLU and obtain the best results. In\naddition, we show its transferability by outperforming the prior best model by\n12.4\\% when adapting to a new domain with little data.', 'In this paper we present the results of our experiments in training and\ndeploying a self-supervised retrieval-based chatbot trained with contrastive\nlearning for assisting customer support agents. In contrast to most existing\nresearch papers in this area where the focus is on solving just one component\nof a deployable chatbot, we present an end-to-end set of solutions to take the\nreader from an unlabelled chatlogs to a deployed chatbot. This set of solutions\nincludes creating a self-supervised dataset and a weakly labelled dataset from\nchatlogs, as well as a systematic approach to selecting a fixed list of canned\nresponses. We present a hierarchical-based RNN architecture for the response\nselection model, chosen for its ability to cache intermediate utterance\nembeddings, which helped to meet deployment inference speed requirements. We\ncompare the performance of this architecture across 3 different learning\nobjectives: self-supervised contrastive learning, binary classification, and\nmulti-class classification. We find that using a self-supervised contrastive\nlearning model outperforms training the binary and multi-class classification\nmodels on a weakly labelled dataset. Our results validate that the\nself-supervised contrastive learning approach can be effectively used for a\nreal-world chatbot scenario.', 'Neural network methods exhibit strong performance only in a few resource-rich\ndomains. Practitioners, therefore, employ domain adaptation from resource-rich\ndomains that are, in most cases, distant from the target domain. Domain\nadaptation between distant domains (e.g., movie subtitles and research papers),\nhowever, cannot be performed effectively due to mismatches in vocabulary; it\nwill encounter many domain-specific words (e.g., ""angstrom"") and words whose\nmeanings shift across domains(e.g., ""conductor""). In this study, aiming to\nsolve these vocabulary mismatches in domain adaptation for neural machine\ntranslation (NMT), we propose vocabulary adaptation, a simple method for\neffective fine-tuning that adapts embedding layers in a given pre-trained NMT\nmodel to the target domain. Prior to fine-tuning, our method replaces the\nembedding layers of the NMT model by projecting general word embeddings induced\nfrom monolingual data in a target domain onto a source-domain embedding space.\nExperimental results indicate that our method improves the performance of\nconventional fine-tuning by 3.86 and 3.28 BLEU points in En-Ja and De-En\ntranslation, respectively.', 'Frequently asked question (FAQ) retrieval, with the purpose of providing\ninformation on frequent questions or concerns, has far-reaching applications in\nmany areas, where a collection of question-answer (Q-A) pairs compiled a priori\ncan be employed to retrieve an appropriate answer in response to a user\\u2019s\nquery that is likely to reoccur frequently. To this end, predominant approaches\nto FAQ retrieval typically rank question-answer pairs by considering either the\nsimilarity between the query and a question (q-Q), the relevance between the\nquery and the associated answer of a question (q-A), or combining the clues\ngathered from the q-Q similarity measure and the q-A relevance measure. In this\npaper, we extend this line of research by combining the clues gathered from the\nq-Q similarity measure and the q-A relevance measure and meanwhile injecting\nextra word interaction information, distilled from a generic (open domain)\nknowledge base, into a contextual language model for inferring the q-A\nrelevance. Furthermore, we also explore to capitalize on domain-specific\ntopically-relevant relations between words in an unsupervised manner, acting as\na surrogate to the supervised domain-specific knowledge base information. As\nsuch, it enables the model to equip sentence representations with the knowledge\nabout domain-specific and topically-relevant relations among words, thereby\nproviding a better q-A relevance measure. We evaluate variants of our approach\non a publicly-available Chinese FAQ dataset, and further apply and\ncontextualize it to a large-scale question-matching task, which aims to search\nquestions from a QA dataset that have a similar intent as an input query.\nExtensive experimental results on these two datasets confirm the promising\nperformance of the proposed approach in relation to some state-of-the-art ones.', 'We develop a chatbot using Deep Bidirectional Transformer models (BERT) to\nhandle client questions in financial investment customer service. The bot can\nrecognize 381 intents, and decides when to say ""I don\'t know"" and escalates\nirrelevant/uncertain questions to human operators. Our main novel contribution\nis the discussion about uncertainty measure for BERT, where three different\napproaches are systematically compared on real problems. We investigated two\nuncertainty metrics, information entropy and variance of dropout sampling in\nBERT, followed by mixed-integer programming to optimize decision thresholds.\nAnother novel contribution is the usage of BERT as a language model in\nautomatic spelling correction. Inputs with accidental spelling errors can\nsignificantly decrease intent classification performance. The proposed approach\ncombines probabilities from masked language model and word edit distances to\nfind the best corrections for misspelled words. The chatbot and the entire\nconversational AI system are developed using open-source tools, and deployed\nwithin our company\'s intranet. The proposed approach can be useful for\nindustries seeking similar in-house solutions in their specific business\ndomains. We share all our code and a sample chatbot built on a public dataset\non Github.', 'With social media becoming increasingly pop-ular on which lots of news and\nreal-time eventsare reported, developing automated questionanswering systems is\ncritical to the effective-ness of many applications that rely on real-time\nknowledge. While previous datasets haveconcentrated on question answering (QA)\nforformal text like news and Wikipedia, wepresent the first large-scale dataset\nfor QA oversocial media data. To ensure that the tweetswe collected are useful,\nwe only gather tweetsused by journalists to write news articles. Wethen ask\nhuman annotators to write questionsand answers upon these tweets. Unlike\notherQA datasets like SQuAD in which the answersare extractive, we allow the\nanswers to be ab-stractive. We show that two recently proposedneural models\nthat perform well on formaltexts are limited in their performance when ap-plied\nto our dataset. In addition, even the fine-tuned BERT model is still lagging\nbehind hu-man performance with a large margin. Our re-sults thus point to the\nneed of improved QAsystems targeting social media text.', 'Much of NLP research has focused on crowdsourced static datasets and the\nsupervised learning paradigm of training once and then evaluating test\nperformance. As argued in de Vries et al. (2020), crowdsourced data has the\nissues of lack of naturalness and relevance to real-world use cases, while the\nstatic dataset paradigm does not allow for a model to learn from its\nexperiences of using language (Silver et al., 2013). In contrast, one might\nhope for machine learning systems that become more useful as they interact with\npeople. In this work, we build and deploy a role-playing game, whereby human\nplayers converse with learning agents situated in an open-domain fantasy world.\nWe show that by training models on the conversations they have with humans in\nthe game the models progressively improve, as measured by automatic metrics and\nonline engagement scores. This learning is shown to be more efficient than\ncrowdsourced data when applied to conversations with real users, as well as\nbeing far cheaper to collect.', 'Continuously-growing data volumes lead to larger generic models. Specific\nuse-cases are usually left out, since generic models tend to perform poorly in\ndomain-specific cases. Our work addresses this gap with a method for selecting\nin-domain data from generic-domain (parallel text) corpora, for the task of\nmachine translation. The proposed method ranks sentences in parallel\ngeneral-domain data according to their cosine similarity with a monolingual\ndomain-specific data set. We then select the top K sentences with the highest\nsimilarity score to train a new machine translation system tuned to the\nspecific in-domain data. Our experimental results show that models trained on\nthis in-domain data outperform models trained on generic or a mixture of\ngeneric and domain data. That is, our method selects high-quality\ndomain-specific training instances at low computational cost and data size.', 'Chatbot is a technology that is used to mimic human behavior using natural\nlanguage. There are different types of Chatbot that can be used as\nconversational agent in various business domains in order to increase the\ncustomer service and satisfaction. For any business domain, it requires a\nknowledge base to be built for that domain and design an information retrieval\nbased system that can respond the user with a piece of documentation or\ngenerated sentences. The core component of a Chatbot is Natural Language\nUnderstanding (NLU) which has been impressively improved by deep learning\nmethods. But we often lack such properly built NLU modules and requires more\ntime to build it from scratch for high quality conversations. This may\nencourage fresh learners to build a Chatbot from scratch with simple\narchitecture and using small dataset, although it may have reduced\nfunctionality, rather than building high quality data driven methods. This\nresearch focuses on Named Entity Recognition (NER) and Intent Classification\nmodels which can be integrated into NLU service of a Chatbot. Named entities\nwill be inserted manually in the knowledge base and automatically detected in a\ngiven sentence. The NER model in the proposed architecture is based on\nartificial neural network which is trained on manually created entities and\nevaluated using CoNLL-2003 dataset.', 'Aspect-based summarization is the task of generating focused summaries based\non specific points of interest. Such summaries aid efficient analysis of text,\nsuch as quickly understanding reviews or opinions from different angles.\nHowever, due to large differences in the type of aspects for different domains\n(e.g., sentiment, product features), the development of previous models has\ntended to be domain-specific. In this paper, we propose WikiAsp, a large-scale\ndataset for multi-domain aspect-based summarization that attempts to spur\nresearch in the direction of open-domain aspect-based summarization.\nSpecifically, we build the dataset using Wikipedia articles from 20 different\ndomains, using the section titles and boundaries of each article as a proxy for\naspect annotation. We propose several straightforward baseline models for this\ntask and conduct experiments on the dataset. Results highlight key challenges\nthat existing summarization models face in this setting, such as proper pronoun\nhandling of quoted sources and consistent explanation of time-sensitive events.', 'Fine-tuning pre-trained Neural Machine Translation (NMT) models is the\ndominant approach for adapting to new languages and domains. However,\nfine-tuning requires adapting and maintaining a separate model for each target\ntask. We propose a simple yet efficient approach for adaptation in NMT. Our\nproposed approach consists of injecting tiny task specific adapter layers into\na pre-trained model. These lightweight adapters, with just a small fraction of\nthe original model size, adapt the model to multiple individual tasks\nsimultaneously. We evaluate our approach on two tasks: (i) Domain Adaptation\nand (ii) Massively Multilingual NMT. Experiments on domain adaptation\ndemonstrate that our proposed approach is on par with full fine-tuning on\nvarious domains, dataset sizes and model capacities. On a massively\nmultilingual dataset of 103 languages, our adaptation approach bridges the gap\nbetween individual bilingual models and one massively multilingual model for\nmost language pairs, paving the way towards universal machine translation.', 'Given appropriate representations of the semantic relations between carpenter\nand wood and between mason and stone (for example, vectors in a vector space\nmodel), a suitable algorithm should be able to recognize that these relations\nare highly similar (carpenter is to wood as mason is to stone; the relations\nare analogous). Likewise, with representations of dog, house, and kennel, an\nalgorithm should be able to recognize that the semantic composition of dog and\nhouse, dog house, is highly similar to kennel (dog house and kennel are\nsynonymous). It seems that these two tasks, recognizing relations and\ncompositions, are closely connected. However, up to now, the best models for\nrelations are significantly different from the best models for compositions. In\nthis paper, we introduce a dual-space model that unifies these two tasks. This\nmodel matches the performance of the best previous models for relations and\ncompositions. The dual-space model consists of a space for measuring domain\nsimilarity and a space for measuring function similarity. Carpenter and wood\nshare the same domain, the domain of carpentry. Mason and stone share the same\ndomain, the domain of masonry. Carpenter and mason share the same function, the\nfunction of artisans. Wood and stone share the same function, the function of\nmaterials. In the composition dog house, kennel has some domain overlap with\nboth dog and house (the domains of pets and buildings). The function of kennel\nis similar to the function of house (the function of shelters). By combining\ndomain and function similarities in various ways, we can model relations,\ncompositions, and other aspects of semantics.', 'GNNs and chemical fingerprints are the predominant approaches to representing\nmolecules for property prediction. However, in NLP, transformers have become\nthe de-facto standard for representation learning thanks to their strong\ndownstream task transfer. In parallel, the software ecosystem around\ntransformers is maturing rapidly, with libraries like HuggingFace and BertViz\nenabling streamlined training and introspection. In this work, we make one of\nthe first attempts to systematically evaluate transformers on molecular\nproperty prediction tasks via our ChemBERTa model. ChemBERTa scales well with\npretraining dataset size, offering competitive downstream performance on\nMoleculeNet and useful attention-based visualization modalities. Our results\nsuggest that transformers offer a promising avenue of future work for molecular\nrepresentation learning and property prediction. To facilitate these efforts,\nwe release a curated dataset of 77M SMILES from PubChem suitable for\nlarge-scale self-supervised pretraining.', 'We present BART, a denoising autoencoder for pretraining sequence-to-sequence\nmodels. BART is trained by (1) corrupting text with an arbitrary noising\nfunction, and (2) learning a model to reconstruct the original text. It uses a\nstandard Tranformer-based neural machine translation architecture which,\ndespite its simplicity, can be seen as generalizing BERT (due to the\nbidirectional encoder), GPT (with the left-to-right decoder), and many other\nmore recent pretraining schemes. We evaluate a number of noising approaches,\nfinding the best performance by both randomly shuffling the order of the\noriginal sentences and using a novel in-filling scheme, where spans of text are\nreplaced with a single mask token. BART is particularly effective when fine\ntuned for text generation but also works well for comprehension tasks. It\nmatches the performance of RoBERTa with comparable training resources on GLUE\nand SQuAD, achieves new state-of-the-art results on a range of abstractive\ndialogue, question answering, and summarization tasks, with gains of up to 6\nROUGE. BART also provides a 1.1 BLEU increase over a back-translation system\nfor machine translation, with only target language pretraining. We also report\nablation experiments that replicate other pretraining schemes within the BART\nframework, to better measure which factors most influence end-task performance.', ""The goal of semi-supervised learning is to utilize the unlabeled, in-domain\ndataset U to improve models trained on the labeled dataset D. Under the context\nof large-scale language-model (LM) pretraining, how we can make the best use of\nU is poorly understood: is semi-supervised learning still beneficial with the\npresence of large-scale pretraining? should U be used for in-domain LM\npretraining or pseudo-label generation? how should the pseudo-label based\nsemi-supervised model be actually implemented? how different semi-supervised\nstrategies affect performances regarding D of different sizes, U of different\nsizes, etc. In this paper, we conduct comprehensive studies on semi-supervised\nlearning in the task of text classification under the context of large-scale LM\npretraining. Our studies shed important lights on the behavior of\nsemi-supervised learning methods: (1) with the presence of in-domain\npretraining LM on U, open-domain LM pretraining is unnecessary; (2) both the\nin-domain pretraining strategy and the pseudo-label based strategy introduce\nsignificant performance boosts, with the former performing better with larger\nU, the latter performing better with smaller U, and the combination leading to\nthe largest performance boost; (3) self-training (pretraining first on pseudo\nlabels D' and then fine-tuning on D) yields better performances when D is\nsmall, while joint training on the combination of pseudo labels D' and the\noriginal dataset D yields better performances when D is large. Using\nsemi-supervised learning strategies, we are able to achieve a performance of\naround 93.8% accuracy with only 50 training data points on the IMDB dataset,\nand a competitive performance of 96.6% with the full IMDB dataset. Our work\nmarks an initial step in understanding the behavior of semi-supervised learning\nmodels under the context of large-scale pretraining."", 'We study the problem of recommending relevant products to users in relatively\nresource-scarce markets by leveraging data from similar, richer in resource\nauxiliary markets. We hypothesize that data from one market can be used to\nimprove performance in another. Only a few studies have been conducted in this\narea, partly due to the lack of publicly available experimental data. To this\nend, we collect and release XMarket, a large dataset covering 18 local markets\non 16 different product categories, featuring 52.5 million user-item\ninteractions. We introduce and formalize the problem of cross-market product\nrecommendation, i.e., market adaptation. We explore different market-adaptation\ntechniques inspired by state-of-the-art domain-adaptation and meta-learning\napproaches and propose a novel neural approach for market adaptation, named\nFOREC. Our model follows a three-step procedure -- pre-training, forking, and\nfine-tuning -- in order to fully utilize the data from an auxiliary market as\nwell as the target market. We conduct extensive experiments studying the impact\nof market adaptation on different pairs of markets. Our proposed approach\ndemonstrates robust effectiveness, consistently improving the performance on\ntarget markets compared to competitive baselines selected for our analysis. In\nparticular, FOREC improves on average 24% and up to 50% in terms of nDCG@10,\ncompared to the NMF baseline. Our analysis and experiments suggest specific\nfuture directions in this research area. We release our data and code for\nacademic purposes.', 'Cross-domain sentiment classification has drawn much attention in recent\nyears. Most existing approaches focus on learning domain-invariant\nrepresentations in both the source and target domains, while few of them pay\nattention to the domain-specific information. Despite the non-transferability\nof the domain-specific information, simultaneously learning domain-dependent\nrepresentations can facilitate the learning of domain-invariant\nrepresentations. In this paper, we focus on aspect-level cross-domain sentiment\nclassification, and propose to distill the domain-invariant sentiment features\nwith the help of an orthogonal domain-dependent task, i.e. aspect detection,\nwhich is built on the aspects varying widely in different domains. We conduct\nextensive experiments on three public datasets and the experimental results\ndemonstrate the effectiveness of our method.', 'The aim of this work is to create a framework for synthetically generating\nquestion/query pairs with as little human input as possible. These datasets can\nbe used to train machine translation systems to convert natural language\nquestions into queries, a useful tool that could allow for more natural access\nto database information. Existing methods of dataset generation require human\ninput that scales linearly with the size of the dataset, resulting in small\ndatasets. Aside from a short initial configuration task, no human input is\nrequired during the query generation process of our system. We leverage\nWikiData, a knowledge base of RDF triples, as a source for generating the main\ncontent of questions and queries. Using multiple layers of question templating\nwe are able to sidestep some of the most challenging parts of query generation\nthat have been handled by humans in previous methods; humans never have to\nmodify, aggregate, inspect, annotate, or generate any questions or queries at\nany step in the process. Our system is easily configurable to multiple domains\nand can be modified to generate queries in natural languages other than\nEnglish. We also present an example dataset of 110,000 question/query pairs\nacross four WikiData domains. We then present a baseline model that we train\nusing the dataset which shows promise in a commercial QA setting.', 'Recently, semantic parsing has attracted much attention in the community.\nAlthough many neural modeling efforts have greatly improved the performance, it\nstill suffers from the data scarcity issue. In this paper, we propose a novel\nsemantic parser for domain adaptation, where we have much fewer annotated data\nin the target domain compared to the source domain. Our semantic parser\nbenefits from a two-stage coarse-to-fine framework, thus can provide different\nand accurate treatments for the two stages, i.e., focusing on domain invariant\nand domain specific information, respectively. In the coarse stage, our novel\ndomain discrimination component and domain relevance attention encourage the\nmodel to learn transferable domain general structures. In the fine stage, the\nmodel is guided to concentrate on domain related details. Experiments on a\nbenchmark dataset show that our method consistently outperforms several popular\ndomain adaptation strategies. Additionally, we show that our model can well\nexploit limited target data to capture the difference between the source and\ntarget domain, even when the target domain has far fewer training instances.', 'Previous studies on the domain adaptation for neural machine translation\n(NMT) mainly focus on the one-pass transferring out-of-domain translation\nknowledge to in-domain NMT model. In this paper, we argue that such a strategy\nfails to fully extract the domain-shared translation knowledge, and repeatedly\nutilizing corpora of different domains can lead to better distillation of\ndomain-shared translation knowledge. To this end, we propose an iterative dual\ndomain adaptation framework for NMT. Specifically, we first pre-train in-domain\nand out-of-domain NMT models using their own training corpora respectively, and\nthen iteratively perform bidirectional translation knowledge transfer (from\nin-domain to out-of-domain and then vice versa) based on knowledge distillation\nuntil the in-domain NMT model convergences. Furthermore, we extend the proposed\nframework to the scenario of multiple out-of-domain training corpora, where the\nabove-mentioned transfer is performed sequentially between the in-domain and\neach out-of-domain NMT models in the ascending order of their domain\nsimilarities. Empirical results on Chinese-English and English-German\ntranslation tasks demonstrate the effectiveness of our framework.', ""Large, human-annotated datasets are central to the development of natural\nlanguage processing models. Collecting these datasets can be the most\nchallenging part of the development process. We address this problem by\nintroducing a general purpose technique for ``simulation-to-real'' transfer in\nlanguage understanding problems with a delimited set of target behaviors,\nmaking it possible to develop models that can interpret natural utterances\nwithout natural training data. We begin with a synthetic data generation\nprocedure, and train a model that can accurately interpret utterances produced\nby the data generator. To generalize to natural utterances, we automatically\nfind projections of natural language utterances onto the support of the\nsynthetic language, using learned sentence embeddings to define a distance\nmetric. With only synthetic training data, our approach matches or outperforms\nstate-of-the-art models trained on natural language data in several domains.\nThese results suggest that simulation-to-real transfer is a practical framework\nfor developing NLP applications, and that improved models for transfer might\nprovide wide-ranging improvements in downstream tasks."", ""With a lot of work about context-free question answering systems, there is an\nemerging trend of conversational question answering models in the natural\nlanguage processing field. Thanks to the recently collected datasets, including\nQuAC and CoQA, there has been more work on conversational question answering,\nand recent work has achieved competitive performance on both datasets. However,\nto best of our knowledge, two important questions for conversational\ncomprehension research have not been well studied: 1) How well can the\nbenchmark dataset reflect models' content understanding? 2) Do the models well\nutilize the conversation content when answering questions? To investigate these\nquestions, we design different training settings, testing settings, as well as\nan attack to verify the models' capability of content understanding on QuAC and\nCoQA. The experimental results indicate some potential hazards in the benchmark\ndatasets, QuAC and CoQA, for conversational comprehension research. Our\nanalysis also sheds light on both what models may learn and how datasets may\nbias the models. With deep investigation of the task, it is believed that this\nwork can benefit the future progress of conversation comprehension. The source\ncode is available at https://github.com/MiuLab/CQA-Study."", 'Making accurate recommendations for cold-start users has been a longstanding\nand critical challenge for recommender systems (RS). Cross-domain\nrecommendations (CDR) offer a solution to tackle such a cold-start problem when\nthere is no sufficient data for the users who have rarely used the system. An\neffective approach in CDR is to leverage the knowledge (e.g., user\nrepresentations) learned from a related but different domain and transfer it to\nthe target domain. Fine-tuning works as an effective transfer learning\ntechnique for this objective, which adapts the parameters of a pre-trained\nmodel from the source domain to the target domain. However, current methods are\nmainly based on the global fine-tuning strategy: the decision of which layers\nof the pre-trained model to freeze or fine-tune is taken for all users in the\ntarget domain. In this paper, we argue that users in RS are personalized and\nshould have their own fine-tuning policies for better preference transfer\nlearning. As such, we propose a novel User-specific Adaptive Fine-tuning method\n(UAF), selecting which layers of the pre-trained network to fine-tune, on a\nper-user basis. Specifically, we devise a policy network with three alternative\nstrategies to automatically decide which layers to be fine-tuned and which\nlayers to have their parameters frozen for each user. Extensive experiments\nshow that the proposed UAF exhibits significantly better and more robust\nperformance for user cold-start recommendation.', 'In this paper, we study transfer learning for the PI and NLI problems, aiming\nto propose a general framework, which can effectively and efficiently adapt the\nshared knowledge learned from a resource-rich source domain to a resource- poor\ntarget domain. Specifically, since most existing transfer learning methods only\nfocus on learning a shared feature space across domains while ignoring the\nrelationship between the source and target domains, we propose to\nsimultaneously learn shared representations and domain relationships in a\nunified framework. Furthermore, we propose an efficient and effective hybrid\nmodel by combining a sentence encoding- based method and a sentence\ninteraction-based method as our base model. Extensive experiments on both\nparaphrase identification and natural language inference demonstrate that our\nbase model is efficient and has promising performance compared to the competing\nmodels, and our transfer learning method can help to significantly boost the\nperformance. Further analysis shows that the inter-domain and intra-domain\nrelationship captured by our model are insightful. Last but not least, we\ndeploy our transfer learning model for PI into our online chatbot system, which\ncan bring in significant improvements over our existing system. Finally, we\nlaunch our new system on the chatbot platform Eva in our E-commerce site\nAliExpress.', 'Answering questions on scholarly knowledge comprising text and other\nartifacts is a vital part of any research life cycle. Querying scholarly\nknowledge and retrieving suitable answers is currently hardly possible due to\nthe following primary reason: machine inactionable, ambiguous and unstructured\ncontent in publications. We present JarvisQA, a BERT based system to answer\nquestions on tabular views of scholarly knowledge graphs. Such tables can be\nfound in a variety of shapes in the scholarly literature (e.g., surveys,\ncomparisons or results). Our system can retrieve direct answers to a variety of\ndifferent questions asked on tabular data in articles. Furthermore, we present\na preliminary dataset of related tables and a corresponding set of natural\nlanguage questions. This dataset is used as a benchmark for our system and can\nbe reused by others. Additionally, JarvisQA is evaluated on two datasets\nagainst other baselines and shows an improvement of two to three folds in\nperformance compared to related methods.', 'Multi-modal pretraining for learning high-level multi-modal representation is\na further step towards deep learning and artificial intelligence. In this work,\nwe propose a novel model, namely InterBERT (BERT for Interaction), which is the\nfirst model of our series of multimodal pretraining methods M6\n(MultiModality-to-MultiModality Multitask Mega-transformer). The model owns\nstrong capability of modeling interaction between the information flows of\ndifferent modalities. The single-stream interaction module is capable of\neffectively processing information of multiple modalilties, and the two-stream\nmodule on top preserves the independence of each modality to avoid performance\ndowngrade in single-modal tasks. We pretrain the model with three pretraining\ntasks, including masked segment modeling (MSM), masked region modeling (MRM)\nand image-text matching (ITM); and finetune the model on a series of\nvision-and-language downstream tasks. Experimental results demonstrate that\nInterBERT outperforms a series of strong baselines, including the most recent\nmulti-modal pretraining methods, and the analysis shows that MSM and MRM are\neffective for pretraining and our method can achieve performances comparable to\nBERT in single-modal tasks. Besides, we propose a large-scale dataset for\nmulti-modal pretraining in Chinese, and we develop the Chinese InterBERT which\nis the first Chinese multi-modal pretrained model. We pretrain the Chinese\nInterBERT on our proposed dataset of 3.1M image-text pairs from the mobile\nTaobao, the largest Chinese e-commerce platform. We finetune the model for\ntext-based image retrieval, and recently we deployed the model online for\ntopic-based recommendation.', 'Many domain adaptation approaches rely on learning cross domain shared\nrepresentations to transfer the knowledge learned in one domain to other\ndomains. Traditional domain adaptation only considers adapting for one task. In\nthis paper, we explore multi-task representation learning under the domain\nadaptation scenario. We propose a neural network framework that supports domain\nadaptation for multiple tasks simultaneously, and learns shared representations\nthat better generalize for domain adaptation. We apply the proposed framework\nto domain adaptation for sequence tagging problems considering two tasks:\nChinese word segmentation and named entity recognition. Experiments show that\nmulti-task domain adaptation works better than disjoint domain adaptation for\neach task, and achieves the state-of-the-art results for both tasks in the\nsocial media domain.', ""Various forums and question answering (Q&A) sites are available online that\nallow Ubuntu users to find results similar to their queries. However, searching\nfor a result is often time consuming as it requires the user to find a specific\nproblem instance relevant to his/her query from a large set of questions. In\nthis paper, we present an automated question answering system for Ubuntu users\ncalled Dr. Tux that is designed to answer user's queries by selecting the most\nsimilar question from an online database. The prototype was implemented in\nPython and uses NLTK and CoreNLP tools for Natural Language Processing. The\ndata for the prototype was taken from the AskUbuntu website which contains\nabout 150k questions. The results obtained from the manual evaluation of the\nprototype were promising while also presenting some interesting opportunities\nfor improvement."", 'Neural network pretraining is gaining attention due to its outstanding\nperformance in natural language processing applications. However, pretraining\nusually leverages predefined task sequences to learn general linguistic clues.\nThe lack of mechanisms in choosing proper tasks during pretraining makes the\nlearning and knowledge encoding inefficient. We thus propose using\nmeta-learning to select tasks that provide the most informative learning\nsignals in each episode of pretraining. With the proposed method, we aim to\nachieve better efficiency in computation and memory usage for the pretraining\nprocess and resulting networks while maintaining the performance. In this\npreliminary work, we discuss the algorithm of the method and its two variants,\ndownstream-aware and downstream-agnostic pretraining. Our experiment plan is\nalso summarized, while empirical results will be shared in our future works.', 'In sequence labeling, previous domain adaptation methods focus on the\nadaptation from the source domain to the entire target domain without\nconsidering the diversity of individual target domain samples, which may lead\nto negative transfer results for certain samples. Besides, an important\ncharacteristic of sequence labeling tasks is that different elements within a\ngiven sample may also have diverse domain relevance, which requires further\nconsideration. To take the multi-level domain relevance discrepancy into\naccount, in this paper, we propose a fine-grained knowledge fusion model with\nthe domain relevance modeling scheme to control the balance between learning\nfrom the target domain data and learning from the source domain model.\nExperiments on three sequence labeling tasks show that our fine-grained\nknowledge fusion model outperforms strong baselines and other state-of-the-art\nsequence labeling domain adaptation methods.', 'Supervised models of NLP rely on large collections of text which closely\nresemble the intended testing setting. Unfortunately matching text is often not\navailable in sufficient quantity, and moreover, within any domain of text, data\nis often highly heterogenous. In this paper we propose a method to distill the\nimportant domain signal as part of a multi-domain learning system, using a\nlatent variable model in which parts of a neural model are stochastically gated\nbased on the inferred domain. We compare the use of discrete versus continuous\nlatent variables, operating in a domain-supervised or a domain semi-supervised\nsetting, where the domain is known only for a subset of training inputs. We\nshow that our model leads to substantial performance improvements over\ncompetitive benchmark domain adaptation methods, including methods using\nadversarial learning.', 'Named entity recognition (NER) is a fundamental task of natural language\nprocessing (NLP). However, most state-of-the-art research is mainly oriented to\nhigh-resource languages such as English and has not been widely applied to\nlow-resource languages. In Malay language, relevant NER resources are limited.\nIn this work, we propose a dataset construction framework, which is based on\nlabeled datasets of homologous languages and iterative optimization, to build a\nMalay NER dataset (MYNER) comprising 28,991 sentences (over 384 thousand\ntokens). Additionally, to better integrate boundary information for NER, we\npropose a multi-task (MT) model with a bidirectional revision (Bi-revision)\nmechanism for Malay NER task. Specifically, an auxiliary task, boundary\ndetection, is introduced to improve NER training in both explicit and implicit\nways. Furthermore, a gated ignoring mechanism is proposed to conduct\nconditional label transfer and alleviate error propagation by the auxiliary\ntask. Experimental results demonstrate that our model achieves comparable\nresults over baselines on MYNER. The dataset and the model in this paper would\nbe publicly released as a benchmark dataset.', ""The volume of academic paper submissions and publications is growing at an\never increasing rate. While this flood of research promises progress in various\nfields, the sheer volume of output inherently increases the amount of noise. We\npresent a system to automatically separate papers with a high from those with a\nlow likelihood of gaining citations as a means to quickly find high impact,\nhigh quality research. Our system uses both a visual classifier, useful for\nsurmising a document's overall appearance, and a text classifier, for making\ncontent-informed decisions. Current work in the field focuses on small datasets\ncomposed of papers from individual conferences. Attempts to use similar\ntechniques on larger datasets generally only considers excerpts of the\ndocuments such as the abstract, potentially throwing away valuable data. We\nrectify these issues by providing a dataset composed of PDF documents and\ncitation counts spanning a decade of output within two separate academic\ndomains: computer science and medicine. This new dataset allows us to expand on\ncurrent work in the field by generalizing across time and academic domain.\nMoreover, we explore inter-domain prediction models - evaluating a classifier's\nperformance on a domain it was not trained on - to shed further insight on this\nimportant problem."", 'We study the problem of domain adaptation for neural abstractive\nsummarization. We make initial efforts in investigating what information can be\ntransferred to a new domain. Experimental results on news stories and opinion\narticles indicate that neural summarization model benefits from pre-training\nbased on extractive summaries. We also find that the combination of in-domain\nand out-of-domain setup yields better summaries when in-domain data is\ninsufficient. Further analysis shows that, the model is capable to select\nsalient content even trained on out-of-domain data, but requires in-domain data\nto capture the style for a target domain.', 'Wikipedia is the largest online encyclopedia, used by algorithms and web\nusers as a central hub of reliable information on the web. The quality and\nreliability of Wikipedia content is maintained by a community of volunteer\neditors. Machine learning and information retrieval algorithms could help scale\nup editors\' manual efforts around Wikipedia content reliability. However, there\nis a lack of large-scale data to support the development of such research. To\nfill this gap, in this paper, we propose Wiki-Reliability, the first dataset of\nEnglish Wikipedia articles annotated with a wide set of content reliability\nissues. To build this dataset, we rely on Wikipedia ""templates"". Templates are\ntags used by expert Wikipedia editors to indicate content issues, such as the\npresence of ""non-neutral point of view"" or ""contradictory articles"", and serve\nas a strong signal for detecting reliability issues in a revision. We select\nthe 10 most popular reliability-related templates on Wikipedia, and propose an\neffective method to label almost 1M samples of Wikipedia article revisions as\npositive or negative with respect to each template. Each positive/negative\nexample in the dataset comes with the full article text and 20 features from\nthe revision\'s metadata. We provide an overview of the possible downstream\ntasks enabled by such data, and show that Wiki-Reliability can be used to train\nlarge-scale models for content reliability prediction. We release all data and\ncode for public use.', 'As open-ended human-chatbot interaction becomes commonplace, sensitive\ncontent detection gains importance. In this work, we propose a two stage\nsemi-supervised approach to bootstrap large-scale data for automatic sensitive\nlanguage detection from publicly available web resources. We explore various\ndata selection methods including 1) using a blacklist to rank online discussion\nforums by the level of their sensitiveness followed by randomly sampling\nutterances and 2) training a weakly supervised model in conjunction with the\nblacklist for scoring sentences from online discussion forums to curate a\ndataset. Our data collection strategy is flexible and allows the models to\ndetect implicit sensitive content for which manual annotations may be\ndifficult. We train models using publicly available annotated datasets as well\nas using the proposed large-scale semi-supervised datasets. We evaluate the\nperformance of all the models on Twitter and Toxic Wikipedia comments testsets\nas well as on a manually annotated spoken language dataset collected during a\nlarge scale chatbot competition. Results show that a model trained on this\ncollected data outperforms the baseline models by a large margin on both\nin-domain and out-of-domain testsets, achieving an F1 score of 95.5% on an\nout-of-domain testset compared to a score of 75% for models trained on public\ndatasets. We also showcase that large scale two stage semi-supervision\ngeneralizes well across multiple classes of sensitivities such as hate speech,\nracism, sexual and pornographic content, etc. without even providing explicit\nlabels for these classes, leading to an average recall of 95.5% versus the\nmodels trained using annotated public datasets which achieve an average recall\nof 73.2% across seven sensitive classes on out-of-domain testsets.', 'Large datasets are essential for neural modeling of many NLP tasks. Current\npublicly available open-domain dialogue datasets offer a trade-off between\nquality (e.g., DailyDialog) and size (e.g., Opensubtitles). We narrow this gap\nby building a high-quality dataset of 14.8M utterances in English, and smaller\ndatasets in German, Dutch, Spanish, Portuguese, Italian, and Hungarian. We\nextract and process dialogues from public-domain books made available by\nProject Gutenberg. We describe our dialogue extraction pipeline, analyze the\neffects of the various heuristics used, and present an error analysis of\nextracted dialogues. Finally, we conduct experiments showing that better\nresponse quality can be achieved in zero-shot and finetuning settings by\ntraining on our data than on the larger but much noisier Opensubtitles dataset.\nOur open-source pipeline (https://github.com/ricsinaruto/gutenberg-dialog) can\nbe extended to further languages with little additional effort. Researchers can\nalso build their versions of existing datasets by adjusting various trade-off\nparameters. We also built a web demo for interacting with our models:\nhttps://ricsinaruto.github.io/chatbot.html.', 'We propose a new architecture for adapting a sentence-level\nsequence-to-sequence transformer by incorporating multiple pretrained document\ncontext signals and assess the impact on translation performance of (1)\ndifferent pretraining approaches for generating these signals, (2) the quantity\nof parallel data for which document context is available, and (3) conditioning\non source, target, or source and target contexts. Experiments on the NIST\nChinese-English, and IWSLT and WMT English-German tasks support four general\nconclusions: that using pretrained context representations markedly improves\nsample efficiency, that adequate parallel data resources are crucial for\nlearning to use document context, that jointly conditioning on multiple context\nrepresentations outperforms any single representation, and that source context\nis more valuable for translation performance than target side context. Our best\nmulti-context model consistently outperforms the best existing context-aware\ntransformers.', ""Using chatbots to deliver recommendations is increasingly popular. The design\nof recommendation chatbots has primarily been taking an information-centric\napproach by focusing on the recommended content per se. Limited attention is on\nhow social connection and relational strategies, such as self-disclosure from a\nchatbot, may influence users' perception and acceptance of the recommendation.\nIn this work, we designed, implemented, and evaluated a social chatbot capable\nof performing three different levels of self-disclosure: factual information\n(low), cognitive opinions (medium), and emotions (high). In the evaluation, we\nrecruited 372 participants to converse with the chatbot on two topics: movies\nand COVID-19 experiences. In each topic, the chatbot performed small talks and\nmade recommendations relevant to the topic. Participants were randomly assigned\nto four experimental conditions where the chatbot used factual, cognitive,\nemotional, and adaptive strategies to perform self-disclosures. By training a\ntext classifier to identify users' level of self-disclosure in real-time, the\nadaptive chatbot can dynamically match its self-disclosure to the level of\ndisclosure exhibited by the users. Our results show that users reciprocate with\nhigher-level self-disclosure when a recommendation chatbot consistently\ndisplays emotions throughout the conversation. Chatbot's emotional disclosure\nalso led to increased interactional enjoyment and more positive interpersonal\nperception towards the bot, fostering a stronger human-chatbot relationship and\nthus leading to increased recommendation effectiveness, including a higher\ntendency to accept the recommendation. We discuss the understandings obtained\nand implications to future design."", 'General purpose Search Engines (SEs) crawl all domains (e.g., Sports, News,\nEntertainment) of the Web, but sometimes the informational need of a query is\nrestricted to a particular domain (e.g., Medical). We leverage the work of SEs\nas part of our effort to route domain specific queries to local Digital\nLibraries (DLs). SEs are often used even if they are not the ""best"" source for\ncertain types of queries. Rather than tell users to ""use this DL for this kind\nof query"", we intend to automatically detect when a query could be better\nserved by a local DL (such as a private, access-controlled DL that is not\ncrawlable via SEs). This is not an easy task because Web queries are short,\nambiguous, and there is lack of quality labeled training data (or it is\nexpensive to create). To detect queries that should be routed to local,\nspecialized DLs, we first send the queries to Google and then examine the\nfeatures in the resulting Search Engine Result Pages (SERPs), and then classify\nthe query as belonging to either the scholar or non-scholar domain. Using\n400,000 AOL queries for the non-scholar domain and 400,000 queries from the\nNASA Technical Report Server (NTRS) for the scholar domain, our classifier\nachieved a precision of 0.809 and F-measure of 0.805.', ""We investigate neural techniques for end-to-end computational argumentation\nmining (AM). We frame AM both as a token-based dependency parsing and as a\ntoken-based sequence tagging problem, including a multi-task learning setup.\nContrary to models that operate on the argument component level, we find that\nframing AM as dependency parsing leads to subpar performance results. In\ncontrast, less complex (local) tagging models based on BiLSTMs perform robustly\nacross classification scenarios, being able to catch long-range dependencies\ninherent to the AM problem. Moreover, we find that jointly learning 'natural'\nsubtasks, in a multi-task learning setup, improves performance."", 'This paper examines various unsupervised pretraining objectives for learning\ndialog context representations. Two novel methods of pretraining dialog context\nencoders are proposed, and a total of four methods are examined. Each\npretraining objective is fine-tuned and evaluated on a set of downstream dialog\ntasks using the MultiWoz dataset and strong performance improvement is\nobserved. Further evaluation shows that our pretraining objectives result in\nnot only better performance, but also better convergence, models that are less\ndata hungry and have better domain generalizability.', 'Creating open-domain chatbots requires large amounts of conversational data\nand related benchmark tasks to evaluate them. Standardized evaluation tasks are\ncrucial for creating automatic evaluation metrics for model development;\notherwise, comparing the models would require resource-expensive human\nevaluation. While chatbot challenges have recently managed to provide a\nplethora of such resources for English, resources in other languages are not\nyet available. In this work, we provide a starting point for Finnish\nopen-domain chatbot research. We describe our collection efforts to create the\nFinnish chat conversation corpus FinChat, which is made available publicly.\nFinChat includes unscripted conversations on seven topics from people of\ndifferent ages. Using this corpus, we also construct a retrieval-based\nevaluation task for Finnish chatbot development. We observe that off-the-shelf\nchatbot models trained on conversational corpora do not perform better than\nchance at choosing the right answer based on automatic metrics, while humans\ncan do the same task almost perfectly. Similarly, in a human evaluation,\nresponses to questions from the evaluation set generated by the chatbots are\npredominantly marked as incoherent. Thus, FinChat provides a challenging\nevaluation set, meant to encourage chatbot development in Finnish.', 'We propose a simple method for automatic speech recognition (ASR) by\nfine-tuning BERT, which is a language model (LM) trained on large-scale\nunlabeled text data and can generate rich contextual representations. Our\nassumption is that given a history context sequence, a powerful LM can narrow\nthe range of possible choices and the speech signal can be used as a simple\nclue. Hence, comparing to conventional ASR systems that train a powerful\nacoustic model (AM) from scratch, we believe that speech recognition is\npossible by simply fine-tuning a BERT model. As an initial study, we\ndemonstrate the effectiveness of the proposed idea on the AISHELL dataset and\nshow that stacking a very simple AM on top of BERT can yield reasonable\nperformance.', 'In this paper, we explore the ways to improve POS-tagging using various types\nof auxiliary losses and different word representations. As a baseline, we\nutilized a BiLSTM tagger, which is able to achieve state-of-the-art results on\nthe sequence labelling tasks. We developed a new method for character-level\nword representation using feedforward neural network. Such representation gave\nus better results in terms of speed and performance of the model. We also\napplied a novel technique of pretraining such word representations with\nexisting word vectors. Finally, we designed a new variant of auxiliary loss for\nsequence labelling tasks: an additional prediction of the neighbour labels.\nSuch loss forces a model to learn the dependencies in-side a sequence of labels\nand accelerates the process of training. We test these methods on English and\nRussian languages.', 'In this paper, we propose two novel methods for domain adaptation for the\nattention-only neural machine translation (NMT) model, i.e., the Transformer.\nOur methods focus on training a single translation model for multiple domains\nby either learning domain specialized hidden state representations or predictor\nbiases for each domain. We combine our methods with a previously proposed\nblack-box method called mixed fine tuning, which is known to be highly\neffective for domain adaptation. In addition, we incorporate multilingualism\ninto the domain adaptation framework. Experiments show that multilingual\nmulti-domain adaptation can significantly improve both resource-poor in-domain\nand resource-rich out-of-domain translations, and the combination of our\nmethods with mixed fine tuning achieves the best performance.', 'Pretraining Neural Language Models (NLMs) over a large corpus involves\nchunking the text into training examples, which are contiguous text segments of\nsizes processable by the neural architecture. We highlight a bias introduced by\nthis common practice: we prove that the pretrained NLM can model much stronger\ndependencies between text segments that appeared in the same training example,\nthan it can between text segments that appeared in different training examples.\nThis intuitive result has a twofold role. First, it formalizes the motivation\nbehind a broad line of recent successful NLM training heuristics, proposed for\nthe pretraining and fine-tuning stages, which do not necessarily appear related\nat first glance. Second, our result clearly indicates further improvements to\nbe made in NLM pretraining for the benefit of Natural Language Understanding\ntasks. As an example, we propose ""kNN-Pretraining"": we show that including\nsemantically related non-neighboring sentences in the same pretraining example\nyields improved sentence representations and open domain question answering\nabilities. This theoretically motivated degree of freedom for pretraining\nexample design indicates new training schemes for self-improving\nrepresentations.', 'In this work, we propose BertGCN, a model that combines large scale\npretraining and transductive learning for text classification. BertGCN\nconstructs a heterogeneous graph over the dataset and represents documents as\nnodes using BERT representations. By jointly training the BERT and GCN modules\nwithin BertGCN, the proposed model is able to leverage the advantages of both\nworlds: large-scale pretraining which takes the advantage of the massive amount\nof raw data and transductive learning which jointly learns representations for\nboth training data and unlabeled test data by propagating label influence\nthrough graph convolution. Experiments show that BertGCN achieves SOTA\nperformances on a wide range of text classification datasets. Code is available\nat https://github.com/ZeroRin/BertGCN.', 'Recent work using auxiliary prediction task classifiers to investigate the\nproperties of LSTM representations has begun to shed light on why pretrained\nrepresentations, like ELMo (Peters et al., 2018) and CoVe (McCann et al.,\n2017), are so beneficial for neural language understanding models. We still,\nthough, do not yet have a clear understanding of how the choice of pretraining\nobjective affects the type of linguistic information that models learn. With\nthis in mind, we compare four objectives---language modeling, translation,\nskip-thought, and autoencoding---on their ability to induce syntactic and\npart-of-speech information. We make a fair comparison between the tasks by\nholding constant the quantity and genre of the training data, as well as the\nLSTM architecture. We find that representations from language models\nconsistently perform best on our syntactic auxiliary prediction tasks, even\nwhen trained on relatively small amounts of data. These results suggest that\nlanguage modeling may be the best data-rich pretraining task for transfer\nlearning applications requiring syntactic information. We also find that the\nrepresentations from randomly-initialized, frozen LSTMs perform strikingly well\non our syntactic auxiliary tasks, but this effect disappears when the amount of\ntraining data for the auxiliary tasks is reduced.', 'We present a CLSRIL-23, a self supervised learning based audio pre-trained\nmodel which learns cross lingual speech representations from raw audio across\n23 Indic languages. It is built on top of wav2vec 2.0 which is solved by\ntraining a contrastive task over masked latent speech representations and\njointly learns the quantization of latents shared across all languages. We\ncompare the language wise loss during pretraining to compare effects of\nmonolingual and multilingual pretraining. Performance on some downstream\nfine-tuning tasks for speech recognition is also compared and our experiments\nshow that multilingual pretraining outperforms monolingual training, in terms\nof learning speech representations which encodes phonetic similarity of\nlanguages and also in terms of performance on down stream tasks. A decrease of\n5% is observed in WER and 9.5% in CER when a multilingual pretrained model is\nused for finetuning in Hindi. All the code models are also open sourced.\nCLSRIL-23 is a model trained on $23$ languages and almost 10,000 hours of audio\ndata to facilitate research in speech recognition for Indic languages. We hope\nthat new state of the art systems will be created using the self supervised\napproach, especially for low resources Indic languages.', 'State-of-the-art speech recognition systems rely heavily on three basic\ncomponents: an acoustic model, a pronunciation lexicon and a language model. To\nbuild these components, a researcher needs linguistic as well as technical\nexpertise, which is a barrier in low-resource domains. Techniques to construct\nthese three components without having expert domain knowledge are in great\ndemand. Urdu, despite having millions of speakers all over the world, is a\nlow-resource language in terms of standard publically available linguistic\nresources. In this paper, we present a grapheme-to-phoneme conversion tool for\nUrdu that generates a pronunciation lexicon in a form suitable for use with\nspeech recognition systems from a list of Urdu words. The tool predicts the\npronunciation of words using a LSTM-based model trained on a handcrafted expert\nlexicon of around 39,000 words and shows an accuracy of 64% upon internal\nevaluation. For external evaluation on a speech recognition task, we obtain a\nword error rate comparable to one achieved using a fully handcrafted expert\nlexicon.', 'Self-supervised pretraining for Automated Speech Recognition (ASR) has shown\nvaried degrees of success. In this paper, we propose to jointly learn\nrepresentations during pretraining from two different modalities: speech and\ntext. The proposed method, tts4pretrain complements the power of contrastive\nlearning in self-supervision with linguistic/lexical representations derived\nfrom synthesized speech, effectively learning from untranscribed speech and\nunspoken text. Lexical learning in the speech encoder is enforced through an\nadditional sequence loss term that is coupled with contrastive loss during\npretraining. We demonstrate that this novel pretraining method yields Word\nError Rate (WER) reductions of 10% relative on the well-benchmarked,\nLibrispeech task over a state-of-the-art baseline pretrained with wav2vec2.0\nonly. The proposed method also serves as an effective strategy to compensate\nfor the lack of transcribed speech, effectively matching the performance of\n5000 hours of transcribed speech with just 100 hours of transcribed speech on\nthe AMI meeting transcription task. Finally, we demonstrate WER reductions of\nup to 15% on an in-house Voice Search task over traditional pretraining.\nIncorporating text into encoder pretraining is complimentary to rescoring with\na larger or in-domain language model, resulting in additional 6% relative\nreduction in WER.', 'The continuous growth of scientific literature brings innovations and, at the\nsame time, raises new challenges. One of them is related to the fact that its\nanalysis has become difficult due to the high volume of published papers for\nwhich manual effort for annotations and management is required. Novel\ntechnological infrastructures are needed to help researchers, research policy\nmakers, and companies to time-efficiently browse, analyse, and forecast\nscientific research. Knowledge graphs i.e., large networks of entities and\nrelationships, have proved to be effective solution in this space. Scientific\nknowledge graphs focus on the scholarly domain and typically contain metadata\ndescribing research publications such as authors, venues, organizations,\nresearch topics, and citations. However, the current generation of knowledge\ngraphs lacks of an explicit representation of the knowledge presented in the\nresearch papers. As such, in this paper, we present a new architecture that\ntakes advantage of Natural Language Processing and Machine Learning methods for\nextracting entities and relationships from research publications and integrates\nthem in a large-scale knowledge graph. Within this research work, we i) tackle\nthe challenge of knowledge extraction by employing several state-of-the-art\nNatural Language Processing and Text Mining tools, ii) describe an approach for\nintegrating entities and relationships generated by these tools, iii) show the\nadvantage of such an hybrid system over alternative approaches, and vi) as a\nchosen use case, we generated a scientific knowledge graph including 109,105\ntriples, extracted from 26,827 abstracts of papers within the Semantic Web\ndomain. As our approach is general and can be applied to any domain, we expect\nthat it can facilitate the management, analysis, dissemination, and processing\nof scientific knowledge.', 'As the wide adoption of intelligent chatbot in human daily life, user demands\nfor such systems evolve from basic task-solving conversations to more casual\nand friend-like communication. To meet the user needs and build emotional bond\nwith users, it is essential for social chatbots to incorporate more human-like\nand advanced linguistic features. In this paper, we investigate the usage of a\ncommonly used rhetorical device by human -- metaphor for social chatbot. Our\nwork first designs a metaphor generation framework, which generates topic-aware\nand novel figurative sentences. By embedding the framework into a chatbot\nsystem, we then enables the chatbot to communicate with users using figurative\nlanguage. Human annotators validate the novelty and properness of the generated\nmetaphors. More importantly, we evaluate the effects of employing metaphors in\nhuman-chatbot conversations. Experiments indicate that our system effectively\narouses user interests in communicating with our chatbot, resulting in\nsignificantly longer human-chatbot conversations.', ""Email responses often contain items-such as a file or a hyperlink to an\nexternal document-that are attached to or included inline in the body of the\nmessage. Analysis of an enterprise email corpus reveals that 35% of the time\nwhen users include these items as part of their response, the attachable item\nis already present in their inbox or sent folder. A modern email client can\nproactively retrieve relevant attachable items from the user's past emails\nbased on the context of the current conversation, and recommend them for\ninclusion, to reduce the time and effort involved in composing the response. In\nthis paper, we propose a weakly supervised learning framework for recommending\nattachable items to the user. As email search systems are commonly available,\nwe constrain the recommendation task to formulating effective search queries\nfrom the context of the conversations. The query is submitted to an existing IR\nsystem to retrieve relevant items for attachment. We also present a novel\nstrategy for generating labels from an email corpus---without the need for\nmanual annotations---that can be used to train and evaluate the query\nformulation model. In addition, we describe a deep convolutional neural network\nthat demonstrates satisfactory performance on this query formulation task when\nevaluated on the publicly available Avocado dataset and a proprietary dataset\nof internal emails obtained through an employee participation program."", 'Contextual word representations derived from large-scale neural language\nmodels are successful across a diverse set of NLP tasks, suggesting that they\nencode useful and transferable features of language. To shed light on the\nlinguistic knowledge they capture, we study the representations produced by\nseveral recent pretrained contextualizers (variants of ELMo, the OpenAI\ntransformer language model, and BERT) with a suite of seventeen diverse probing\ntasks. We find that linear models trained on top of frozen contextual\nrepresentations are competitive with state-of-the-art task-specific models in\nmany cases, but fail on tasks requiring fine-grained linguistic knowledge\n(e.g., conjunct identification). To investigate the transferability of\ncontextual word representations, we quantify differences in the transferability\nof individual layers within contextualizers, especially between recurrent\nneural networks (RNNs) and transformers. For instance, higher layers of RNNs\nare more task-specific, while transformer layers do not exhibit the same\nmonotonic trend. In addition, to better understand what makes contextual word\nrepresentations transferable, we compare language model pretraining with eleven\nsupervised pretraining tasks. For any given task, pretraining on a closely\nrelated task yields better performance than language model pretraining (which\nis better on average) when the pretraining dataset is fixed. However, language\nmodel pretraining on more data gives the best results.', 'This work explores the application of textual entailment in news claim\nverification and stance prediction using a new corpus in Arabic. The publicly\navailable corpus comes in two perspectives: a version consisting of 4,547 true\nand false claims and a version consisting of 3,786 pairs (claim, evidence). We\ndescribe the methodology for creating the corpus and the annotation process.\nUsing the introduced corpus, we also develop two machine learning baselines for\ntwo proposed tasks: claim verification and stance prediction. Our best model\nutilizes pretraining (BERT) and achieves 76.7 F1 on the stance prediction task\nand 64.3 F1 on the claim verification task. Our preliminary experiments shed\nsome light on the limits of automatic claim verification that relies on claims\ntext only. Results hint that while the linguistic features and world knowledge\nlearned during pretraining are useful for stance prediction, such learned\nrepresentations from pretraining are insufficient for verifying claims without\naccess to context or evidence.', 'We introduce a set of nine challenge tasks that test for the understanding of\nfunction words. These tasks are created by structurally mutating sentences from\nexisting datasets to target the comprehension of specific types of function\nwords (e.g., prepositions, wh-words). Using these probing tasks, we explore the\neffects of various pretraining objectives for sentence encoders (e.g., language\nmodeling, CCG supertagging and natural language inference (NLI)) on the learned\nrepresentations. Our results show that pretraining on language modeling\nperforms the best on average across our probing tasks, supporting its\nwidespread use for pretraining state-of-the-art NLP models, and CCG\nsupertagging and NLI pretraining perform comparably. Overall, no pretraining\nobjective dominates across the board, and our function word probing tasks\nhighlight several intuitive differences between pretraining objectives, e.g.,\nthat NLI helps the comprehension of negation.', 'Language Models like ELMo and BERT have provided robust representations of\nnatural language, which serve as the language understanding component for a\ndiverse range of downstream tasks.Curriculum learning is a method that employs\na structured training regime instead, which has been leveraged in computer\nvision and machine translation to improve model training speed and model\nperformance. While language models have proven transformational for the natural\nlanguage processing community, these models have proven expensive,\nenergy-intensive, and challenging to train. In this work, we explore the effect\nof curriculum learning on language model pretraining using various\nlinguistically motivated curricula and evaluate transfer performance on the\nGLUE Benchmark. Despite a broad variety of training methodologies and\nexperiments we do not find compelling evidence that curriculum learning methods\nimprove language model training.', 'Users of machine translation (MT) may want to ensure the use of specific\nlexical terminologies. While there exist techniques for incorporating\nterminology constraints during inference for MT, current APE approaches cannot\nensure that they will appear in the final translation. In this paper, we\npresent both autoregressive and non-autoregressive models for lexically\nconstrained APE, demonstrating that our approach enables preservation of 95% of\nthe terminologies and also improves translation quality on English-German\nbenchmarks. Even when applied to lexically constrained MT output, our approach\nis able to improve preservation of the terminologies. However, we show that our\nmodels do not learn to copy constraints systematically and suggest a simple\ndata augmentation technique that leads to improved performance and robustness.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently engaged in a project that involves the generation of realistic text using large-scale, pre-trained language models. However, I am encountering difficulties in controlling the results of the generation process. Existing methods, such as prompting, do not offer the level of control I require.  I am in search of an innovative approach that can enhance the controllability of text generation while simultaneously preserving the quality of the text produced.  More specifically, I am considering the projection of certain constraints and control conditions into the latent space. By injecting these embeddings into the language model's inputs, I anticipate that it could regulate the output of the language model.  I am particularly interested in applying this method to tasks such as constrained text generation, where explicit constraints are present. My goal is for my proposed method to surpass the performance of other fine-tuned or prompting-based methods.","['Large pre-trained language models have repeatedly shown their ability to\nproduce fluent text. Yet even when starting from a prompt, generation can\ncontinue in many plausible directions. Current decoding methods with the goal\nof controlling generation, e.g., to ensure specific words are included, either\nrequire additional models or fine-tuning, or work poorly when the task at hand\nis semantically unconstrained, e.g., story generation. In this work, we present\na plug-and-play decoding method for controlled language generation that is so\nsimple and intuitive, it can be described in a single sentence: given a topic\nor keyword, we add a shift to the probability distribution over our vocabulary\ntowards semantically similar words. We show how annealing this distribution can\nbe used to impose hard constraints on language generation, something no other\nplug-and-play method is currently able to do with SOTA language generators.\nDespite the simplicity of this approach, we see it works incredibly well in\npractice: decoding from GPT-2 leads to diverse and fluent sentences while\nguaranteeing the appearance of given guide words. We perform two user studies,\nrevealing that (1) our method outperforms competing methods in human\nevaluations; and (2) forcing the guide words to appear in the generated text\nhas no impact on the fluency of the generated text.', ""Large transformer-based language models (LMs) trained on huge text corpora\nhave shown unparalleled generation capabilities. However, controlling\nattributes of the generated language (e.g. switching topic or sentiment) is\ndifficult without modifying the model architecture or fine-tuning on\nattribute-specific data and entailing the significant cost of retraining. We\npropose a simple alternative: the Plug and Play Language Model (PPLM) for\ncontrollable language generation, which combines a pretrained LM with one or\nmore simple attribute classifiers that guide text generation without any\nfurther training of the LM. In the canonical scenario we present, the attribute\nmodels are simple classifiers consisting of a user-specified bag of words or a\nsingle learned layer with 100,000 times fewer parameters than the LM. Sampling\nentails a forward and backward pass in which gradients from the attribute model\npush the LM's hidden activations and thus guide the generation. Model samples\ndemonstrate control over a range of topics and sentiment styles, and extensive\nautomated and human annotated evaluations show attribute alignment and fluency.\nPPLMs are flexible in that any combination of differentiable attribute models\nmay be used to steer text generation, which will allow for diverse and creative\napplications beyond the examples given in this paper."", 'Conditional text generation has been a challenging task that is yet to see\nhuman-level performance from state-of-the-art models. In this work, we\nspecifically focus on the Commongen benchmark, wherein the aim is to generate a\nplausible sentence for a given set of input concepts. Despite advances in other\ntasks, large pre-trained language models that are fine-tuned on this dataset\noften produce sentences that are syntactically correct but qualitatively\ndeviate from a human understanding of common sense. Furthermore, generated\nsequences are unable to fulfill such lexical requirements as matching\npart-of-speech and full concept coverage. In this paper, we explore how\ncommonsense knowledge graphs can enhance model performance, with respect to\ncommonsense reasoning and lexically-constrained decoding. We propose strategies\nfor enhancing the semantic correctness of the generated text, which we\naccomplish through: extracting commonsense relations from Conceptnet, injecting\nthese relations into the Unified Language Model (UniLM) through attention\nmechanisms, and enforcing the aforementioned lexical requirements through\noutput constraints. By performing several ablations, we find that commonsense\ninjection enables the generation of sentences that are more aligned with human\nunderstanding, while remaining compliant with lexical requirements.', 'Large pre-trained language models are capable of generating realistic text.\nHowever, controlling these models so that the generated text satisfies lexical\nconstraints, i.e., contains specific words, is a challenging problem. Given\nthat state-of-the-art language models are too large to be trained from scratch\nin a manageable time, it is desirable to control these models without\nre-training them. Methods capable of doing this are called plug-and-play.\nRecent plug-and-play methods have been successful in constraining small\nbidirectional language models as well as forward models in tasks with a\nrestricted search space, e.g., machine translation. However, controlling large\ntransformer-based models to meet lexical constraints without re-training them\nremains a challenge. In this work, we propose Directed Beam Search (DBS), a\nplug-and-play method for lexically constrained language generation. Our method\ncan be applied to any language model, is easy to implement and can be used for\ngeneral language generation. In our experiments we use DBS to control GPT-2. We\ndemonstrate its performance on keyword-to-phrase generation and we obtain\ncomparable results as a state-of-the-art non-plug-and-play model for lexically\nconstrained story generation.', ""Conditional text generation often requires lexical constraints, i.e., which\nwords should or shouldn't be included in the output text. While the dominant\nrecipe for conditional text generation has been large-scale pretrained language\nmodels that are finetuned on the task-specific training data, such models do\nnot learn to follow the underlying constraints reliably, even when supervised\nwith large amounts of task-specific examples.\n  We propose NeuroLogic Decoding, a simple yet effective algorithm that enables\nneural language models -- supervised or not -- to generate fluent text while\nsatisfying complex lexical constraints. Our approach is powerful yet efficient.\nIt handles any set of lexical constraints that is expressible under predicate\nlogic, while its asymptotic runtime is equivalent to conventional beam search.\n  Empirical results on four benchmarks show that NeuroLogic Decoding\noutperforms previous approaches, including algorithms that handle a subset of\nour constraints. Moreover, we find that unsupervised models with NeuroLogic\nDecoding often outperform supervised models with conventional decoding, even\nwhen the latter is based on considerably larger networks. Our results suggest\nthe limit of large-scale neural networks for fine-grained controllable\ngeneration and the promise of inference-time algorithms."", 'Large language models benefit from training with a large amount of unlabeled\ntext, which gives them increasingly fluent and diverse generation capabilities.\nHowever, using these models for text generation that takes into account target\nattributes, such as sentiment polarity or specific topics, remains a challenge.\nWe propose a simple and flexible method for controlling text generation by\naligning disentangled attribute representations. In contrast to recent efforts\non training a discriminator to perturb the token level distribution for an\nattribute, we use the same data to learn an alignment function to guide the\npre-trained, non-controlled language model to generate texts with the target\nattribute without changing the original language model parameters. We evaluate\nour method on sentiment- and topic-controlled generation, and show large\nperformance gains over previous methods while retaining fluency and diversity.', 'Large-scale language models show promising text generation capabilities, but\nusers cannot easily control particular aspects of the generated text. We\nrelease CTRL, a 1.63 billion-parameter conditional transformer language model,\ntrained to condition on control codes that govern style, content, and\ntask-specific behavior. Control codes were derived from structure that\nnaturally co-occurs with raw text, preserving the advantages of unsupervised\nlearning while providing more explicit control over text generation. These\ncodes also allow CTRL to predict which parts of the training data are most\nlikely given a sequence. This provides a potential method for analyzing large\namounts of data via model-based source attribution. We have released multiple\nfull-sized, pretrained versions of CTRL at https://github.com/salesforce/ctrl.', 'Large language models (LM) based on Transformers allow to generate plausible\nlong texts. In this paper, we explore how this generation can be further\ncontrolled at decoding time to satisfy certain constraints (e.g. being\nnon-toxic, conveying certain emotions, using a specific writing style, etc.)\nwithout fine-tuning the LM. Precisely, we formalize constrained generation as a\ntree exploration process guided by a discriminator that indicates how well the\nassociated sequence respects the constraint. This approach, in addition to\nbeing easier and cheaper to train than fine-tuning the LM, allows to apply the\nconstraint more finely and dynamically. We propose several original methods to\nsearch this generation tree, notably the Monte Carlo Tree Search (MCTS) which\nprovides theoretical guarantees on the search efficiency, but also simpler\nmethods based on re-ranking a pool of diverse sequences using the discriminator\nscores. These methods are evaluated, with automatic and human-based metrics, on\ntwo types of constraints and languages: review polarity and emotion control in\nFrench and English. We show that discriminator-guided MCTS decoding achieves\nstate-of-the-art results without having to tune the language model, in both\ntasks and languages. We also demonstrate that other proposed decoding methods\nbased on re-ranking can be really effective when diversity among the generated\npropositions is encouraged.', ""Nowadays, there exist powerful language models such as OpenAI's GPT-2 that\ncan generate readable text and can be fine-tuned to generate text for a\nspecific domain. Considering GPT-2, it cannot directly generate synthetic news\nwith respect to a given topic and the output of the language model cannot be\nexplicitly controlled. In this paper, we study the novel problem of\ntopic-preserving synthetic news generation. We propose a novel deep\nreinforcement learning-based method to control the output of GPT-2 with respect\nto a given news topic. When generating text using GPT-2, by default, the most\nprobable word is selected from the vocabulary. Instead of selecting the best\nword each time from GPT-2's output, an RL agent tries to select words that\noptimize the matching of a given topic. In addition, using a fake news detector\nas an adversary, we investigate generating realistic news using our proposed\nmethod. In this paper, we consider realistic news as news that cannot be easily\ndetected by a fake news classifier. Experimental results demonstrate the\neffectiveness of the proposed framework on generating topic-preserving news\ncontent than state-of-the-art baselines."", 'Sequence-to-Sequence (S2S) neural text generation models, especially the\npre-trained ones (e.g., BART and T5), have exhibited compelling performance on\nvarious natural language generation tasks. However, the black-box nature of\nthese models limits their application in tasks where specific rules (e.g.,\ncontrollable constraints, prior knowledge) need to be executed. Previous works\neither design specific model structure (e.g., Copy Mechanism corresponding to\nthe rule ""the generated output should include certain words in the source\ninput"") or implement specialized inference algorithm (e.g., Constrained Beam\nSearch) to execute particular rules through the text generation. These methods\nrequire careful design case-by-case and are difficult to support multiple rules\nconcurrently. In this paper, we propose a novel module named Neural\nRule-Execution Tracking Machine that can be equipped into various\ntransformer-based generators to leverage multiple rules simultaneously to guide\nthe neural generation model for superior generation performance in a unified\nand scalable way. Extensive experimental results on several benchmarks verify\nthe effectiveness of our proposed model in both controllable and general text\ngeneration.', 'Existing pre-trained large language models have shown unparalleled generative\ncapabilities. However, they are not controllable. In this paper, we propose\nMEGATRON-CNTRL, a novel framework that uses large-scale language models and\nadds control to text generation by incorporating an external knowledge base.\nOur framework consists of a keyword predictor, a knowledge retriever, a\ncontextual knowledge ranker, and a conditional text generator. As we do not\nhave access to ground-truth supervision for the knowledge ranker, we make use\nof weak supervision from sentence embedding. The empirical results show that\nour model generates more fluent, consistent, and coherent stories with less\nrepetition and higher diversity compared to prior work on the ROC story\ndataset. We showcase the controllability of our model by replacing the keywords\nused to generate stories and re-running the generation process. Human\nevaluation results show that 77.5% of these stories are successfully controlled\nby the new keywords. Furthermore, by scaling our model from 124 million to 8.3\nbillion parameters we demonstrate that larger models improve both the quality\nof generation (from 74.5% to 93.0% for consistency) and controllability (from\n77.5% to 91.5%).', 'We propose a Distributional Approach for addressing Controlled Text\nGeneration from pre-trained Language Models (LMs). This approach permits to\nspecify, in a single formal framework, both ""pointwise"" and ""distributional""\nconstraints over the target LM -- to our knowledge, the first model with such\ngenerality -- while minimizing KL divergence from the initial LM distribution.\nThe optimal target distribution is then uniquely determined as an explicit EBM\n(Energy-Based Model) representation. From that optimal representation we then\ntrain a target controlled Autoregressive LM through an adaptive distributional\nvariant of Policy Gradient. We conduct a first set of experiments over\npointwise constraints showing the advantages of our approach over a set of\nbaselines, in terms of obtaining a controlled LM balancing constraint\nsatisfaction with divergence from the initial LM. We then perform experiments\nover distributional constraints, a unique feature of our approach,\ndemonstrating its potential as a remedy to the problem of Bias in Language\nModels. Through an ablation study, we show the effectiveness of our adaptive\ntechnique for obtaining faster convergence. (Code available at\nhttps://github.com/naver/gdc)', 'Large-scale pretrained language models have led to dramatic improvements in\ntext generation. Impressive performance can be achieved by finetuning only on a\nsmall number of instances (few-shot setting). Nonetheless, almost all previous\nwork simply applies random sampling to select the few-shot training instances.\nLittle to no attention has been paid to the selection strategies and how they\nwould affect model performance. In this work, we present a study on training\ninstance selection in few-shot neural text generation. The selection decision\nis made based only on the unlabeled data so as to identify the most worthwhile\ndata points that should be annotated under some budget of labeling cost. Based\non the intuition that the few-shot training instances should be diverse and\nrepresentative of the entire data distribution, we propose a simple selection\nstrategy with K-means clustering. We show that even with the naive\nclustering-based approach, the generation models consistently outperform random\nsampling on three text generation tasks: data-to-text generation, document\nsummarization and question generation. We hope that this work will call for\nmore attention on this largely unexplored area.', 'We investigate large-scale latent variable models (LVMs) for neural story\ngeneration -- an under-explored application for open-domain long text -- with\nobjectives in two threads: generation effectiveness and controllability. LVMs,\nespecially the variational autoencoder (VAE), have achieved both effective and\ncontrollable generation through exploiting flexible distributional latent\nrepresentations. Recently, Transformers and its variants have achieved\nremarkable effectiveness without explicit latent representation learning, thus\nlack satisfying controllability in generation. In this paper, we advocate to\nrevive latent variable modeling, essentially the power of representation\nlearning, in the era of Transformers to enhance controllability without hurting\nstate-of-the-art generation effectiveness. Specifically, we integrate latent\nrepresentation vectors with a Transformer-based pre-trained architecture to\nbuild conditional variational autoencoder (CVAE). Model components such as\nencoder, decoder and the variational posterior are all built on top of\npre-trained language models -- GPT2 specifically in this paper. Experiments\ndemonstrate state-of-the-art conditional generation ability of our model, as\nwell as its excellent representation learning capability and controllability.', ""Large-scale transformer-based language models (LMs) demonstrate impressive\ncapabilities in open text generation. However, controlling the generated text's\nproperties such as the topic, style, and sentiment is challenging and often\nrequires significant changes to the model architecture or retraining and\nfine-tuning the model on new supervised data. This paper presents a novel\napproach for Topical Language Generation (TLG) by combining a pre-trained LM\nwith topic modeling information. We cast the problem using Bayesian probability\nformulation with topic probabilities as a prior, LM probabilities as the\nlikelihood, and topical language generation probability as the posterior. In\nlearning the model, we derive the topic probability distribution from the\nuser-provided document's natural structure. Furthermore, we extend our model by\nintroducing new parameters and functions to influence the quantity of the\ntopical features presented in the generated text. This feature would allow us\nto easily control the topical properties of the generated text. Our\nexperimental results demonstrate that our model outperforms the\nstate-of-the-art results on coherency, diversity, and fluency while being\nfaster in decoding."", 'Conditional story generation and contextual text continuation have become\nincreasingly popular topics in NLP community. Existing models are often prone\nto output paragraphs of texts that gradually diverge from the given prompt.\nAlthough the generated text may have a reasonable perplexity and diversity, it\ncould easily be identified by human as gibberish. The goal of our project is to\nimprove the coherence and consistency across sentences in a language-generation\nmodel. We aim to solve this issue by first training a sentence pair coherence\nclassifier with GPT-2 pretrained model, and then co-train the GPT-2 language\nmodel with this new coherence objective using a method analogous to the\nREINFORCE algorithm. This fine-tuned language model is able to generate lengthy\nparagraph conditioned on a given topic without diverging too much. The\nsimplicity of this model allows it to be applicable to a variety of underlying\nlanguage model architecture since it only modifies the final layer of the\npre-trained model.', 'When trained effectively, the Variational Autoencoder (VAE) can be both a\npowerful generative model and an effective representation learning framework\nfor natural language. In this paper, we propose the first large-scale language\nVAE model, Optimus. A universal latent embedding space for sentences is first\npre-trained on large text corpus, and then fine-tuned for various language\ngeneration and understanding tasks. Compared with GPT-2, Optimus enables guided\nlanguage generation from an abstract level using the latent vectors. Compared\nwith BERT, Optimus can generalize better on low-resource language understanding\ntasks due to the smooth latent space structure. Extensive experimental results\non a wide range of language tasks demonstrate the effectiveness of Optimus. It\nachieves new state-of-the-art on VAE language modeling benchmarks. We hope that\nour first pre-trained big VAE language model itself and results can help the\nNLP community renew the interests of deep generative models in the era of\nlarge-scale pre-training, and make these principled methods more practical.', 'When directly using existing text generation datasets for controllable\ngeneration, we are facing the problem of not having the domain knowledge and\nthus the aspects that could be controlled are limited. A typical example is\nwhen using CNN/Daily Mail dataset for controllable text summarization, there is\nno guided information on the emphasis of summary sentences. A more useful text\ngenerator should leverage both the input text and the control signal to guide\nthe generation, which can only be built with a deep understanding of the domain\nknowledge. Motivated by this vision, our paper introduces a new text generation\ndataset, named MReD. Our new dataset consists of 7,089 meta-reviews and all its\n45k meta-review sentences are manually annotated with one of the 9 carefully\ndefined categories, including abstract, strength, decision, etc. We present\nexperimental results on start-of-the-art summarization models, and propose\nmethods for structure-controlled generation with both extractive and\nabstractive models using our annotated data. By exploring various settings and\nanalyzing the model behavior with respect to the control signal, we demonstrate\nthe challenges of our proposed task and the values of our dataset MReD.\nMeanwhile, MReD also allows us to have a better understanding of the\nmeta-review domain.', 'There has been considerable progress made towards conversational models that\ngenerate coherent and fluent responses; however, this often involves training\nlarge language models on large dialogue datasets, such as Reddit. These large\nconversational models provide little control over the generated responses, and\nthis control is further limited in the absence of annotated conversational\ndatasets for attribute specific generation that can be used for fine-tuning the\nmodel. In this paper, we first propose and evaluate plug-and-play methods for\ncontrollable response generation, which does not require dialogue specific\ndatasets and does not rely on fine-tuning a large model. While effective, the\ndecoding procedure induces considerable computational overhead, rendering the\nconversational model unsuitable for interactive usage. To overcome this, we\nintroduce an approach that does not require further computation at decoding\ntime, while also does not require any fine-tuning of a large language model. We\ndemonstrate, through extensive automatic and human evaluation, a high degree of\ncontrol over the generated conversational responses with regard to multiple\ndesired attributes, while being fluent.', 'Large-scale language models (LMs) pretrained on massive corpora of text, such\nas GPT-2, are powerful open-domain text generators. However, as our systematic\nexamination reveals, it is still challenging for such models to generate\ncoherent long passages of text (e.g., 1000 tokens), especially when the models\nare fine-tuned to the target domain on a small corpus. Previous\nplanning-then-generation methods also fall short of producing such long text in\nvarious domains. To overcome the limitations, we propose a simple but effective\nmethod of generating text in a progressive manner, inspired by generating\nimages from low to high resolution. Our method first produces domain-specific\ncontent keywords and then progressively refines them into complete passages in\nmultiple stages. The simple design allows our approach to take advantage of\npretrained LMs at each stage and effectively adapt to any target domain given\nonly a small set of examples. We conduct a comprehensive empirical study with a\nbroad set of evaluation metrics, and show that our approach significantly\nimproves upon the fine-tuned large LMs and various planning-then-generation\nmethods in terms of quality and sample efficiency. Human evaluation also\nvalidates that our model generations are more coherent.', 'Neural text generation models are often autoregressive language models or\nseq2seq models. These models generate text by sampling words sequentially, with\neach word conditioned on the previous word, and are state-of-the-art for\nseveral machine translation and summarization benchmarks. These benchmarks are\noften defined by validation perplexity even though this is not a direct measure\nof the quality of the generated text. Additionally, these models are typically\ntrained via maxi- mum likelihood and teacher forcing. These methods are\nwell-suited to optimizing perplexity but can result in poor sample quality\nsince generating text requires conditioning on sequences of words that may have\nnever been observed at training time. We propose to improve sample quality\nusing Generative Adversarial Networks (GANs), which explicitly train the\ngenerator to produce high quality samples and have shown a lot of success in\nimage generation. GANs were originally designed to output differentiable\nvalues, so discrete language generation is challenging for them. We claim that\nvalidation perplexity alone is not indicative of the quality of text generated\nby a model. We introduce an actor-critic conditional GAN that fills in missing\ntext conditioned on the surrounding context. We show qualitatively and\nquantitatively, evidence that this produces more realistic conditional and\nunconditional text samples compared to a maximum likelihood trained model.', 'Transfer learning through large pre-trained models has changed the landscape\nof current applications in natural language processing (NLP). Recently Optimus,\na variational autoencoder (VAE) which combines two pre-trained models, BERT and\nGPT-2, has been released, and its combination with generative adversial\nnetworks (GANs) has been shown to produce novel, yet very human-looking text.\nThe Optimus and GANs combination avoids the troublesome application of GANs to\nthe discrete domain of text, and prevents the exposure bias of standard maximum\nlikelihood methods. We combine the training of GANs in the latent space, with\nthe finetuning of the decoder of Optimus for single word generation. This\napproach lets us model both the high-level features of the sentences, and the\nlow-level word-by-word generation. We finetune using reinforcement learning\n(RL) by exploiting the structure of GPT-2 and by adding entropy-based\nintrinsically motivated rewards to balance between quality and diversity. We\nbenchmark the results of the VAE-GAN model, and show the improvements brought\nby our RL finetuning on three widely used datasets for text generation, with\nresults that greatly surpass the current state-of-the-art for the quality of\nthe generated texts.', 'This work introduces Focused-Variation Network (FVN), a novel model to\ncontrol language generation. The main problems in previous controlled language\ngeneration models range from the difficulty of generating text according to the\ngiven attributes, to the lack of diversity of the generated texts. FVN\naddresses these issues by learning disjoint discrete latent spaces for each\nattribute inside codebooks, which allows for both controllability and\ndiversity, while at the same time generating fluent text. We evaluate FVN on\ntwo text generation datasets with annotated content and style, and show\nstate-of-the-art performance as assessed by automatic and human evaluations.', 'Large-scale pretrained language models have shown thrilling generation\ncapabilities, especially when they generate consistent long text in thousands\nof words with ease. However, users of these models can only control the prefix\nof sentences or certain global aspects of generated text. It is challenging to\nsimultaneously achieve fine-grained controllability and preserve the\nstate-of-the-art unconditional text generation capability. In this paper, we\nfirst propose a new task named ""Outline to Story"" (O2S) as a test bed for\nfine-grained controllable generation of long text, which generates a\nmulti-paragraph story from cascaded events, i.e. a sequence of outline events\nthat guide subsequent paragraph generation. We then create dedicate datasets\nfor future benchmarks, built by state-of-the-art keyword extraction techniques.\nFinally, we propose an extremely simple yet strong baseline method for the O2S\ntask, which fine tunes pre-trained language models on augmented sequences of\noutline-story pairs with simple language modeling objective. Our method does\nnot introduce any new parameters or perform any architecture modification,\nexcept several special tokens as delimiters to build augmented sequences.\nExtensive experiments on various datasets demonstrate state-of-the-art\nconditional story generation performance with our model, achieving better\nfine-grained controllability and user flexibility. Our paper is among the first\nones by our knowledge to propose a model and to create datasets for the task of\n""outline to story"". Our work also instantiates research interest of\nfine-grained controllable generation of open-domain long text, where\ncontrolling inputs are represented by short text.', 'Recent neural approaches to data-to-text generation have mostly focused on\nimproving content fidelity while lacking explicit control over writing styles\n(e.g., word choices, sentence structures). More traditional systems use\ntemplates to determine the realization of text. Yet manual or automatic\nconstruction of high-quality templates is difficult, and a template acting as\nhard constraints could harm content fidelity when it does not match the record\nperfectly. We study a new way of stylistic control by using existing sentences\nas soft templates. That is, the model learns to imitate the writing style of\nany given exemplar sentence, with automatic adaptions to faithfully describe\nthe content record. The problem is challenging due to the lack of parallel\ndata. We develop a neural approach that includes a hybrid attention-copy\nmechanism, learns with weak supervisions, and is enhanced with a new content\ncoverage constraint. We conduct experiments in restaurants and sports domains.\nResults show our approach achieves stronger performance than a range of\ncomparison methods. Our approach balances well between content fidelity and\nstyle control given exemplars that match the records to varying degrees.', ""Advances in language modeling architectures and the availability of large\ntext corpora have driven progress in automatic text generation. While this\nresults in models capable of generating coherent texts, it also prompts models\nto internalize social biases present in the training corpus. This paper aims to\nquantify and reduce a particular type of bias exhibited by language models:\nbias in the sentiment of generated text. Given a conditioning context (e.g., a\nwriting prompt) and a language model, we analyze if (and how) the sentiment of\nthe generated text is affected by changes in values of sensitive attributes\n(e.g., country names, occupations, genders) in the conditioning context using a\nform of counterfactual evaluation. We quantify sentiment bias by adopting\nindividual and group fairness metrics from the fair machine learning\nliterature, and demonstrate that large-scale models trained on two different\ncorpora (news articles, and Wikipedia) exhibit considerable levels of bias. We\nthen propose embedding and sentiment prediction-derived regularization on the\nlanguage model's latent representations. The regularizations improve fairness\nmetrics while retaining comparable levels of perplexity and semantic\nsimilarity."", 'Text generation has become one of the most important yet challenging tasks in\nnatural language processing (NLP). The resurgence of deep learning has greatly\nadvanced this field by neural generation models, especially the paradigm of\npretrained language models (PLMs). In this paper, we present an overview of the\nmajor advances achieved in the topic of PLMs for text generation. As the\npreliminaries, we present the general task definition and briefly describe the\nmainstream architectures of PLMs for text generation. As the core content, we\ndiscuss how to adapt existing PLMs to model different input data and satisfy\nspecial properties in the generated text. We further summarize several\nimportant fine-tuning strategies for text generation. Finally, we present\nseveral future directions and conclude this paper. Our survey aims to provide\ntext generation researchers a synthesis and pointer to related research.', 'Large-scale language models such as GPT-3 are excellent few-shot learners,\nallowing them to be controlled via natural text prompts. Recent studies report\nthat prompt-based direct classification eliminates the need for fine-tuning but\nlacks data and inference scalability. This paper proposes a novel data\naugmentation technique that leverages large-scale language models to generate\nrealistic text samples from a mixture of real samples. We also propose\nutilizing soft-labels predicted by the language models, effectively distilling\nknowledge from the large-scale language models and creating textual\nperturbations simultaneously. We perform data augmentation experiments on\ndiverse classification tasks and show that our method hugely outperforms\nexisting text augmentation methods. Ablation studies and a qualitative analysis\nprovide more insights into our approach.', 'Controllable text generation is an appealing but challenging task, which\nallows users to specify particular attributes of the generated outputs. In this\npaper, we propose a controllable dialogue generation model to steer response\ngeneration under multi-attribute constraints. Specifically, we define and\ncategorize the commonly used control attributes into global and local ones,\nwhich possess different granularities of effects on response generation. Then,\nwe significantly extend the conventional seq2seq framework by introducing a\nnovel two-stage decoder, which first uses a multi-grained style specification\nlayer to impose the stylistic constraints and determine word-level control\nstates of responses based on the attributes, and then employs a response\ngeneration layer to generate final responses maintaining both semantic\nrelevancy to the contexts and fidelity to the attributes. Furthermore, we train\nour model with an attribute consistency reward to promote response control with\nexplicit supervision signals. Extensive experiments and in-depth analyses on\ntwo datasets indicate that our model can significantly outperform competitive\nbaselines in terms of response quality, content diversity and controllability.', 'Many tasks, including language generation, benefit from learning the\nstructure of the output space, particularly when the space of output labels is\nlarge and the data is sparse. State-of-the-art neural language models\nindirectly capture the output space structure in their classifier weights since\nthey lack parameter sharing across output labels. Learning shared output label\nmappings helps, but existing methods have limited expressivity and are prone to\noverfitting. In this paper, we investigate the usefulness of more powerful\nshared mappings for output labels, and propose a deep residual output mapping\nwith dropout between layers to better capture the structure of the output space\nand avoid overfitting. Evaluations on three language generation tasks show that\nour output label mapping can match or improve state-of-the-art recurrent and\nself-attention architectures, and suggest that the classifier does not\nnecessarily need to be high-rank to better model natural language if it is\nbetter at capturing the structure of the output space.', 'Fine-tuning pre-trained generative language models to down-stream language\ngeneration tasks has shown promising results. However, this comes with the cost\nof having a single, large model for each task, which is not ideal in\nlow-memory/power scenarios (e.g., mobile). In this paper, we propose an\neffective way to fine-tune multiple down-stream generation tasks simultaneously\nusing a single, large pre-trained model. The experiments on five diverse\nlanguage generation tasks show that by just using an additional 2-3% parameters\nfor each task, our model can maintain or even improve the performance of\nfine-tuning the whole model.', 'We present a novel approach to data-to-text generation based on iterative\ntext editing. Our approach maximizes the completeness and semantic accuracy of\nthe output text while leveraging the abilities of recent pre-trained models for\ntext editing (LaserTagger) and language modeling (GPT-2) to improve the text\nfluency. To this end, we first transform data items to text using trivial\ntemplates, and then we iteratively improve the resulting text by a neural model\ntrained for the sentence fusion task. The output of the model is filtered by a\nsimple heuristic and reranked with an off-the-shelf pre-trained language model.\nWe evaluate our approach on two major data-to-text datasets (WebNLG, Cleaned\nE2E) and analyze its caveats and benefits. Furthermore, we show that our\nformulation of data-to-text generation opens up the possibility for zero-shot\ndomain adaptation using a general-domain dataset for sentence fusion.', 'Data-to-Text Generation (DTG) is a subfield of Natural Language Generation\naiming at transcribing structured data in natural language descriptions. The\nfield has been recently boosted by the use of neural-based generators which\nexhibit on one side great syntactic skills without the need of hand-crafted\npipelines; on the other side, the quality of the generated text reflects the\nquality of the training data, which in realistic settings only offer\nimperfectly aligned structure-text pairs. Consequently, state-of-art neural\nmodels include misleading statements - usually called hallucinations - in their\noutputs. The control of this phenomenon is today a major challenge for DTG, and\nis the problem addressed in the paper.\n  Previous work deal with this issue at the instance level: using an alignment\nscore for each table-reference pair. In contrast, we propose a finer-grained\napproach, arguing that hallucinations should rather be treated at the word\nlevel. Specifically, we propose a Multi-Branch Decoder which is able to\nleverage word-level labels to learn the relevant parts of each training\ninstance. These labels are obtained following a simple and efficient scoring\nprocedure based on co-occurrence analysis and dependency parsing. Extensive\nevaluations, via automated metrics and human judgment on the standard WikiBio\nbenchmark, show the accuracy of our alignment labels and the effectiveness of\nthe proposed Multi-Branch Decoder. Our model is able to reduce and control\nhallucinations, while keeping fluency and coherence in generated texts. Further\nexperiments on a degraded version of ToTTo show that our model could be\nsuccessfully used on very noisy settings.', 'In this paper, we present our submission for the English to Czech Text\nTranslation Task of IWSLT 2019. Our system aims to study how pre-trained\nlanguage models, used as input embeddings, can improve a specialized machine\ntranslation system trained on few data. Therefore, we implemented a\nTransformer-based encoder-decoder neural system which is able to use the output\nof a pre-trained language model as input embeddings, and we compared its\nperformance under three configurations: 1) without any pre-trained language\nmodel (constrained), 2) using a language model trained on the monolingual parts\nof the allowed English-Czech data (constrained), and 3) using a language model\ntrained on a large quantity of external monolingual data (unconstrained). We\nused BERT as external pre-trained language model (configuration 3), and BERT\narchitecture for training our own language model (configuration 2). Regarding\nthe training data, we trained our MT system on a small quantity of parallel\ntext: one set only consists of the provided MuST-C corpus, and the other set\nconsists of the MuST-C corpus and the News Commentary corpus from WMT. We\nobserved that using the external pre-trained BERT improves the scores of our\nsystem by +0.8 to +1.5 of BLEU on our development set, and +0.97 to +1.94 of\nBLEU on the test set. However, using our own language model trained only on the\nallowed parallel data seems to improve the machine translation performances\nonly when the system is trained on the smallest dataset.', 'Given a sentence (e.g., ""I like mangoes"") and a constraint (e.g., sentiment\nflip), the goal of controlled text generation is to produce a sentence that\nadapts the input sentence to meet the requirements of the constraint (e.g., ""I\nhate mangoes""). Going beyond such simple constraints, recent works have started\nexploring the incorporation of complex syntactic-guidance as constraints in the\ntask of controlled paraphrase generation. In these methods, syntactic-guidance\nis sourced from a separate exemplar sentence. However, these prior works have\nonly utilized limited syntactic information available in the parse tree of the\nexemplar sentence. We address this limitation in the paper and propose Syntax\nGuided Controlled Paraphraser (SGCP), an end-to-end framework for syntactic\nparaphrase generation. We find that SGCP can generate syntax conforming\nsentences while not compromising on relevance. We perform extensive automated\nand human evaluations over multiple real-world English language datasets to\ndemonstrate the efficacy of SGCP over state-of-the-art baselines. To drive\nfuture research, we have made SGCP\'s source code available', 'Despite considerable advancements with deep neural language models, the\nenigma of neural text degeneration persists when these models are tested as\ntext generators. The counter-intuitive empirical observation is that even\nthough the use of likelihood as training objective leads to high quality models\nfor a broad range of language understanding tasks, using likelihood as a\ndecoding objective leads to text that is bland and strangely repetitive.\n  In this paper, we reveal surprising distributional differences between human\ntext and machine text. In addition, we find that decoding strategies alone can\ndramatically effect the quality of machine text, even when generated from\nexactly the same neural language model. Our findings motivate Nucleus Sampling,\na simple but effective method to draw the best out of neural generation. By\nsampling text from the dynamic nucleus of the probability distribution, which\nallows for diversity while effectively truncating the less reliable tail of the\ndistribution, the resulting text better demonstrates the quality of human text,\nyielding enhanced diversity without sacrificing fluency and coherence.', 'Concept-to-text Natural Language Generation is the task of expressing an\ninput meaning representation in natural language. Previous approaches in this\ntask have been able to generalise to rare or unseen instances by relying on a\ndelexicalisation of the input. However, this often requires that the input\nappears verbatim in the output text. This poses challenges in multilingual\nsettings, where the task expands to generate the output text in multiple\nlanguages given the same input. In this paper, we explore the application of\nmultilingual models in concept-to-text and propose Language Agnostic\nDelexicalisation, a novel delexicalisation method that uses multilingual\npretrained embeddings, and employs a character-level post-editing model to\ninflect words in their correct form during relexicalisation. Our experiments\nacross five datasets and five languages show that multilingual models\noutperform monolingual models in concept-to-text and that our framework\noutperforms previous approaches, especially for low resource languages.', 'Large, pre-trained transformer-based language models such as BERT have\ndrastically changed the Natural Language Processing (NLP) field. We present a\nsurvey of recent work that uses these large language models to solve NLP tasks\nvia pre-training then fine-tuning, prompting, or text generation approaches. We\nalso present approaches that use pre-trained language models to generate data\nfor training augmentation or other purposes. We conclude with discussions on\nlimitations and suggested directions for future research.', 'Graph-to-text generation has benefited from pre-trained language models\n(PLMs) in achieving better performance than structured graph encoders. However,\nthey fail to fully utilize the structure information of the input graph. In\nthis paper, we aim to further improve the performance of the pre-trained\nlanguage model by proposing a structured graph-to-text model with a two-step\nfine-tuning mechanism which first fine-tunes the model on Wikipedia before\nadapting to the graph-to-text generation. In addition to using the traditional\ntoken and position embeddings to encode the knowledge graph (KG), we propose a\nnovel tree-level embedding method to capture the inter-dependency structures of\nthe input graph. This new approach has significantly improved the performance\nof all text generation metrics for the English WebNLG 2017 dataset.', 'Training generative models that can generate high-quality text with\nsufficient diversity is an important open problem for Natural Language\nGeneration (NLG) community. Recently, generative adversarial models have been\napplied extensively on text generation tasks, where the adversarially trained\ngenerators alleviate the exposure bias experienced by conventional maximum\nlikelihood approaches and result in promising generation quality. However, due\nto the notorious defect of mode collapse for adversarial training, the\nadversarially trained generators face a quality-diversity trade-off, i.e., the\ngenerator models tend to sacrifice generation diversity severely for increasing\ngeneration quality. In this paper, we propose a novel approach which aims to\nimprove the performance of adversarial text generation via efficiently\ndecelerating mode collapse of the adversarial training. To this end, we\nintroduce a cooperative training paradigm, where a language model is\ncooperatively trained with the generator and we utilize the language model to\nefficiently shape the data distribution of the generator against mode collapse.\nMoreover, instead of engaging the cooperative update for the generator in a\nprincipled way, we formulate a meta learning mechanism, where the cooperative\nupdate to the generator serves as a high level meta task, with an intuition of\nensuring the parameters of the generator after the adversarial update would\nstay resistant against mode collapse. In the experiment, we demonstrate our\nproposed approach can efficiently slow down the pace of mode collapse for the\nadversarial text generators. Overall, our proposed method is able to outperform\nthe baseline approaches with significant margins in terms of both generation\nquality and diversity in the testified domains.', 'Many neural network models nowadays have achieved promising performances in\nChit-chat settings. The majority of them rely on an encoder for understanding\nthe post and a decoder for generating the response. Without given assigned\nsemantics, the models lack the fine-grained control over responses as the\nsemantic mapping between posts and responses is hidden on the fly within the\nend-to-end manners. Some previous works utilize sampled latent words as a\ncontrollable semantic form to drive the generated response around the work, but\nfew works attempt to use more complex semantic patterns to guide the\ngeneration. In this paper, we propose to use more detailed semantic forms,\nincluding latent responses and part-of-speech sequences sampled from the\ncorresponding distributions, as the controllable semantics to guide the\ngeneration. Our results show that the richer semantics are not only able to\nprovide informative and diverse responses, but also increase the overall\nperformance of response quality, including fluency and coherence.']","[""This paper investigates whether the power of the models pre-trained on text\ndata, such as BERT, can be transferred to general token sequence classification\napplications. To verify pre-trained models' transferability, we test the\npre-trained models on text classification tasks with meanings of tokens\nmismatches, and real-world non-text token sequence classification data,\nincluding amino acid, DNA, and music. We find that even on non-text data, the\nmodels pre-trained on text converge faster, perform better than the randomly\ninitialized models, and only slightly worse than the models using task-specific\nknowledge. We also find that the representations of the text and non-text\npre-trained models share non-trivial similarities."", 'Representing a text as a graph for obtaining automatic text summarization has\nbeen investigated for over ten years. With the development of attention or\nTransformer on natural language processing (NLP), it is possible to make a\nconnection between the graph and attention structure for a text. In this paper,\nan attention matrix between the sentences of the whole text is adopted as a\nweighted adjacent matrix of a fully connected graph of the text, which can be\nproduced through the pre-training language model. The GCN is further applied to\nthe text graph model for classifying each node and finding out the salient\nsentences from the text. It is demonstrated by the experimental results on two\ntypical datasets that our proposed model can achieve a competitive result in\ncomparison with sate-of-the-art models.', 'This paper presents a systematic survey on recent development of neural text\ngeneration models. Specifically, we start from recurrent neural network\nlanguage models with the traditional maximum likelihood estimation training\nscheme and point out its shortcoming for text generation. We thus introduce the\nrecently proposed methods for text generation based on reinforcement learning,\nre-parametrization tricks and generative adversarial nets (GAN) techniques. We\ncompare different properties of these models and the corresponding techniques\nto handle their common problems such as gradient vanishing and generation\ndiversity. Finally, we conduct a benchmarking experiment with different types\nof neural text generation models on two well-known datasets and discuss the\nempirical results along with the aforementioned model properties.', 'The dominant language modeling paradigm handles text as a sequence of\ndiscrete tokens. While that approach can capture the latent structure of the\ntext, it is inherently constrained to sequential dynamics for text generation.\nWe propose a new paradigm for introducing a syntactic inductive bias into\nneural text generation, where the dependency parse tree is used to drive the\nTransformer model to generate sentences iteratively.\n  Our experiments show that this paradigm is effective at text generation, with\nquality between LSTMs and Transformers, and comparable diversity, requiring\nless than half their decoding steps, and its generation process allows direct\ncontrol over the syntactic constructions of the generated text, enabling the\ninduction of stylistic variations.', 'Recent trends in natural language processing using pretraining have shifted\nfocus towards pretraining and fine-tuning approaches for text generation. Often\nthe focus has been on task-agnostic approaches that generalize the language\nmodeling objective. We propose question generation as a pretraining method,\nwhich better aligns with the text generation objectives. Our text generation\nmodels pretrained with this method are better at understanding the essence of\nthe input and are better language models for the target task. When evaluated on\ntwo text generation tasks, abstractive summarization and answer-focused\nquestion generation, our models result in state-of-the-art performances in\nterms of automatic metrics. Human evaluators also found our summaries and\ngenerated questions to be more natural, concise and informative.', 'Controlled text perturbation is useful for evaluating and improving model\ngeneralizability. However, current techniques rely on training a model for\nevery target perturbation, which is expensive and hard to generalize. We\npresent Tailor, a semantically-controlled text generation system. Tailor builds\non a pretrained seq2seq model and produces textual outputs conditioned on\ncontrol codes derived from semantic representations. We craft a set of\noperations to modify the control codes, which in turn steer generation towards\ntargeted attributes. These operations can be further composed into higher-level\nones, allowing for flexible perturbation strategies. We demonstrate the\neffectiveness of these perturbations in multiple applications. First, we use\nTailor to automatically create high-quality contrast sets for four distinct\nnatural language processing (NLP) tasks. These contrast sets contain fewer\nspurious artifacts and are complementary to manually annotated ones in their\nlexical diversity. Second, we show that Tailor perturbations can improve model\ngeneralization through data augmentation. Perturbing just 2% of training data\nleads to a 5.8-point gain on an NLI challenge set measuring reliance on\nsyntactic heuristics.', 'Lexically constrained sentence generation allows the incorporation of prior\nknowledge such as lexical constraints into the output. This technique has been\napplied to machine translation, and dialog response generation. Previous work\nusually used Markov Chain Monte Carlo (MCMC) sampling to generate lexically\nconstrained sentences, but they randomly determined the position to be edited\nand the action to be taken, resulting in many invalid refinements. To overcome\nthis challenge, we used a classifier to instruct the MCMC-based models where\nand how to refine the candidate sentences. First, we developed two methods to\ncreate synthetic data on which the pre-trained model is fine-tuned to obtain a\nreliable classifier. Next, we proposed a two-step approach, ""Predict and\nRevise"", for constrained sentence generation. During the predict step, we\nleveraged the classifier to compute the learned prior for the candidate\nsentence. During the revise step, we resorted to MCMC sampling to revise the\ncandidate sentence by conducting a sampled action at a sampled position drawn\nfrom the learned prior. We compared our proposed models with many strong\nbaselines on two tasks, generating sentences with lexical constraints and text\ninfilling. Experimental results have demonstrated that our proposed model\nperforms much better than the previous work in terms of sentence fluency and\ndiversity. Our code and pre-trained models are available at\nhttps://github.com/NLPCode/MCMCXLNet.', 'In this work, we tackle the problem of structured text generation,\nspecifically academic paper generation in $\\LaTeX{}$, inspired by the\nsurprisingly good results of basic character-level language models. Our\nmotivation is using more recent and advanced methods of language modeling on a\nmore complex dataset of $\\LaTeX{}$ source files to generate realistic academic\npapers. Our first contribution is preparing a dataset with $\\LaTeX{}$ source\nfiles on recent open-source computer vision papers. Our second contribution is\nexperimenting with recent methods of language modeling and text generation such\nas Transformer and Transformer-XL to generate consistent $\\LaTeX{}$ code. We\nreport cross-entropy and bits-per-character (BPC) results of the trained\nmodels, and we also discuss interesting points on some examples of the\ngenerated $\\LaTeX{}$ code.', 'Conventional methods for the image-text generation tasks mainly tackle the\nnaturally bidirectional generation tasks separately, focusing on designing\ntask-specific frameworks to improve the quality and fidelity of the generated\nsamples. Recently, Vision-Language Pre-training models have greatly improved\nthe performance of the image-to-text generation tasks, but large-scale\npre-training models for text-to-image synthesis task are still under-developed.\nIn this paper, we propose ERNIE-ViLG, a unified generative pre-training\nframework for bidirectional image-text generation with transformer model. Based\non the image quantization models, we formulate both image generation and text\ngeneration as autoregressive generative tasks conditioned on the text/image\ninput. The bidirectional image-text generative modeling eases the semantic\nalignments across vision and language. For the text-to-image generation\nprocess, we further propose an end-to-end training method to jointly learn the\nvisual sequence generator and the image reconstructor. To explore the landscape\nof large-scale pre-training for bidirectional text-image generation, we train a\n10-billion parameter ERNIE-ViLG model on a large-scale dataset of 145 million\n(Chinese) image-text pairs which achieves state-of-the-art performance for both\ntext-to-image and image-to-text tasks, obtaining an FID of 7.9 on MS-COCO for\ntext-to-image synthesis and best results on COCO-CN and AIC-ICC for image\ncaptioning.', ""The purpose of unconditional text generation is to train a model with real\nsentences, then generate novel sentences of the same quality and diversity as\nthe training data. However, when different metrics are used for comparing the\nmethods of unconditional text generation, contradictory conclusions are drawn.\nThe difficulty is that both the diversity and quality of the sample should be\nconsidered simultaneously when the models are evaluated. To solve this problem,\na novel metric of distributional discrepancy (DD) is designed to evaluate\ngenerators based on the discrepancy between the generated and real training\nsentences. However, it cannot compute the DD directly because the distribution\nof real sentences is unavailable. Thus, we propose a method for estimating the\nDD by training a neural-network-based text classifier. For comparison, three\nexisting metrics, bi-lingual evaluation understudy (BLEU) versus self-BLEU,\nlanguage model score versus reverse language model score, and Fr\\'{e}chet\nembedding distance, along with the proposed DD, are used to evaluate two\npopular generative models of long short-term memory and generative pretrained\ntransformer 2 on both syntactic and real data. Experimental results show that\nDD is significantly better than the three existing metrics for ranking these\ngenerative models."", 'It has been shown for English that discrete and soft prompting perform\nstrongly in few-shot learning with pretrained language models (PLMs). In this\npaper, we show that discrete and soft prompting perform better than finetuning\nin multilingual cases: Crosslingual transfer and in-language training of\nmultilingual natural language inference. For example, with 48 English training\nexamples, finetuning obtains 33.74% accuracy in crosslingual transfer, barely\nsurpassing the majority baseline (33.33%). In contrast, discrete and soft\nprompting outperform finetuning, achieving 36.43% and 38.79%. We also\ndemonstrate good performance of prompting with training data in multiple\nlanguages other than English.', 'When fine-tuning pretrained models for classification, researchers either use\na generic model head or a task-specific prompt for prediction. Proponents of\nprompting have argued that prompts provide a method for injecting task-specific\nguidance, which is beneficial in low-data regimes. We aim to quantify this\nbenefit through rigorous testing of prompts in a fair setting: comparing\nprompted and head-based fine-tuning in equal conditions across many tasks and\ndata sizes. By controlling for many sources of advantage, we find that\nprompting does indeed provide a benefit, and that this benefit can be\nquantified per task. Results show that prompting is often worth 100s of data\npoints on average across classification tasks.', 'Lexically constrained text generation aims to control the generated text by\nincorporating some pre-specified keywords into the output. Previous work\ninjects lexical constraints into the output by controlling the decoding process\nor refining the candidate output iteratively, which tends to generate generic\nor ungrammatical sentences, and has high computational complexity. To address\nthese challenges, we propose Constrained BART (CBART) for lexically constrained\ntext generation. CBART leverages the pre-trained model BART and transfers part\nof the generation burden from the decoder to the encoder by decomposing this\ntask into two sub-tasks, thereby improving the sentence quality. Concretely, we\nextend BART by adding a token-level classifier over the encoder, aiming at\ninstructing the decoder where to replace and insert. Guided by the encoder, the\ndecoder refines multiple tokens of the input in one step by inserting tokens\nbefore specific positions and re-predicting tokens with low confidence. To\nfurther reduce the inference latency, the decoder predicts all tokens in\nparallel. Experiment results on One-Billion-Word and Yelp show that CBART can\ngenerate plausible text with high quality and diversity while significantly\naccelerating inference.', 'Recent advances in neural network-based generative modeling have reignited\nthe hopes in having computer systems capable of seamlessly conversing with\nhumans and able to understand natural language. Neural architectures have been\nemployed to generate text excerpts to various degrees of success, in a\nmultitude of contexts and tasks that fulfil various user needs. Notably, high\ncapacity deep learning models trained on large scale datasets demonstrate\nunparalleled abilities to learn patterns in the data even in the lack of\nexplicit supervision signals, opening up a plethora of new possibilities\nregarding producing realistic and coherent texts. While the field of natural\nlanguage generation is evolving rapidly, there are still many open challenges\nto address. In this survey we formally define and categorize the problem of\nnatural language generation. We review particular application tasks that are\ninstantiations of these general formulations, in which generating natural\nlanguage is of practical importance. Next we include a comprehensive outline of\nmethods and neural architectures employed for generating diverse texts.\nNevertheless, there is no standard way to assess the quality of text produced\nby these generative models, which constitutes a serious bottleneck towards the\nprogress of the field. To this end, we also review current approaches to\nevaluating natural language generation systems. We hope this survey will\nprovide an informative overview of formulations, methods, and assessments of\nneural natural language generation.', 'Machine learning approaches applied to NLP are often evaluated by summarizing\ntheir performance in a single number, for example accuracy. Since most test\nsets are constructed as an i.i.d. sample from the overall data, this approach\noverly simplifies the complexity of language and encourages overfitting to the\nhead of the data distribution. As such, rare language phenomena or text about\nunderrepresented groups are not equally included in the evaluation. To\nencourage more in-depth model analyses, researchers have proposed the use of\nmultiple test sets, also called challenge sets, that assess specific\ncapabilities of a model. In this paper, we develop a framework based on this\nidea which is able to generate controlled perturbations and identify subsets in\ntext-to-scalar, text-to-text, or data-to-text settings. By applying this\nframework to the GEM generation benchmark, we propose an evaluation suite made\nof 80 challenge sets, demonstrate the kinds of analyses that it enables and\nshed light onto the limits of current generation models.', 'Many important NLP problems can be posed as dual-sequence or\nsequence-to-sequence modeling tasks. Recent advances in building end-to-end\nneural architectures have been highly successful in solving such tasks. In this\nwork we propose a new architecture for dual-sequence modeling that is based on\nassociative memory. We derive AM-RNNs, a recurrent associative memory (AM)\nwhich augments generic recurrent neural networks (RNN). This architecture is\nextended to the Dual AM-RNN which operates on two AMs at once. Our models\nachieve very competitive results on textual entailment. A qualitative analysis\ndemonstrates that long range dependencies between source and target-sequence\ncan be bridged effectively using Dual AM-RNNs. However, an initial experiment\non auto-encoding reveals that these benefits are not exploited by the system\nwhen learning to solve sequence-to-sequence tasks which indicates that\nadditional supervision or regularization is needed.', 'We investigate the difficulty levels of questions in reading comprehension\ndatasets such as SQuAD, and propose a new question generation setting, named\nDifficulty-controllable Question Generation (DQG). Taking as input a sentence\nin the reading comprehension paragraph and some of its text fragments (i.e.,\nanswers) that we want to ask questions about, a DQG method needs to generate\nquestions each of which has a given text fragment as its answer, and meanwhile\nthe generation is under the control of specified difficulty labels---the output\nquestions should satisfy the specified difficulty as much as possible. To solve\nthis task, we propose an end-to-end framework to generate questions of\ndesignated difficulty levels by exploring a few important intuitions. For\nevaluation, we prepared the first dataset of reading comprehension questions\nwith difficulty labels. The results show that the question generated by our\nframework not only have better quality under the metrics like BLEU, but also\ncomply with the specified difficulty labels.', 'In this paper, we propose a novel controllable text-to-image generative\nadversarial network (ControlGAN), which can effectively synthesise high-quality\nimages and also control parts of the image generation according to natural\nlanguage descriptions. To achieve this, we introduce a word-level spatial and\nchannel-wise attention-driven generator that can disentangle different visual\nattributes, and allow the model to focus on generating and manipulating\nsubregions corresponding to the most relevant words. Also, a word-level\ndiscriminator is proposed to provide fine-grained supervisory feedback by\ncorrelating words with image regions, facilitating training an effective\ngenerator which is able to manipulate specific visual attributes without\naffecting the generation of other content. Furthermore, perceptual loss is\nadopted to reduce the randomness involved in the image generation, and to\nencourage the generator to manipulate specific attributes required in the\nmodified text. Extensive experiments on benchmark datasets demonstrate that our\nmethod outperforms existing state of the art, and is able to effectively\nmanipulate synthetic images using natural language descriptions. Code is\navailable at https://github.com/mrlibw/ControlGAN.', 'Current multilingual vision-language models either require a large number of\nadditional parameters for each supported language, or suffer performance\ndegradation as languages are added. In this paper, we propose a Scalable\nMultilingual Aligned Language Representation (SMALR) that supports many\nlanguages with few model parameters without sacrificing downstream task\nperformance. SMALR learns a fixed size language-agnostic representation for\nmost words in a multilingual vocabulary, keeping language-specific features for\njust a few. We use a masked cross-language modeling loss to align features with\ncontext from other languages. Additionally, we propose a cross-lingual\nconsistency module that ensures predictions made for a query and its machine\ntranslation are comparable. The effectiveness of SMALR is demonstrated with ten\ndiverse languages, over twice the number supported in vision-language tasks to\ndate. We evaluate on multilingual image-sentence retrieval and outperform prior\nwork by 3-4% with less than 1/5th the training parameters compared to other\nword embedding methods.', ""The softmax operator is one of the most important functions in machine\nlearning models. When applying neural networks to multi-category\nclassification, the correlations among different categories are often ignored.\nFor example, in text generation, a language model makes a choice of each new\nword based only on the former selection of its context. In this scenario, the\nlink statistics information of concurrent words based on a corpus (an analogy\nof the natural way of expression) is also valuable in choosing the next word,\nwhich can help to improve the sentence's fluency and smoothness. To fully\nexplore such important information, we propose a graph softmax function for\ntext generation. It is expected that the final classification result would be\ndominated by both the language model and graphical text relationships among\nwords. We use a graph total variation term to regularize softmax so as to\nincorporate the concurrent relationship into the language model. The total\nvariation of the generated words should be small locally. We apply the proposed\ngraph softmax to GPT2 for the text generation task. Experimental results\ndemonstrate that the proposed graph softmax achieves better BLEU and perplexity\nthan softmax. Human testers can also easily distinguish the text generated by\nthe graph softmax or softmax."", 'We improve the informativeness of models for conditional text generation\nusing techniques from computational pragmatics. These techniques formulate\nlanguage production as a game between speakers and listeners, in which a\nspeaker should generate output text that a listener can use to correctly\nidentify the original input that the text describes. While such approaches are\nwidely used in cognitive science and grounded language learning, they have\nreceived less attention for more standard language generation tasks. We\nconsider two pragmatic modeling methods for text generation: one where\npragmatics is imposed by information preservation, and another where pragmatics\nis imposed by explicit modeling of distractors. We find that these methods\nimprove the performance of strong existing systems for abstractive\nsummarization and generation from structured meaning representations.', 'Deep generative models are commonly used for generating images and text.\nInterpretability of these models is one important pursuit, other than the\ngeneration quality. Variational auto-encoder (VAE) with Gaussian distribution\nas prior has been successfully applied in text generation, but it is hard to\ninterpret the meaning of the latent variable. To enhance the controllability\nand interpretability, one can replace the Gaussian prior with a mixture of\nGaussian distributions (GM-VAE), whose mixture components could be related to\nhidden semantic aspects of data. In this paper, we generalize the practice and\nintroduce DEM-VAE, a class of models for text generation using VAEs with a\nmixture distribution of exponential family. Unfortunately, a standard\nvariational training algorithm fails due to the mode-collapse problem. We\ntheoretically identify the root cause of the problem and propose an effective\nalgorithm to train DEM-VAE. Our method penalizes the training with an extra\ndispersion term to induce a well-structured latent space. Experimental results\nshow that our approach does obtain a meaningful space, and it outperforms\nstrong baselines in text generation benchmarks. The code is available at\nhttps://github.com/wenxianxian/demvae.', 'The goal of text generation is to make machines express in human language. It\nis one of the most important yet challenging tasks in natural language\nprocessing (NLP). Since 2014, various neural encoder-decoder models pioneered\nby Seq2Seq have been proposed to achieve the goal by learning to map input text\nto output text. However, the input text alone often provides limited knowledge\nto generate the desired output, so the performance of text generation is still\nfar from satisfaction in many real-world scenarios. To address this issue,\nresearchers have considered incorporating various forms of knowledge beyond the\ninput text into the generation models. This research direction is known as\nknowledge-enhanced text generation. In this survey, we present a comprehensive\nreview of the research on knowledge enhanced text generation over the past five\nyears. The main content includes two parts: (i) general methods and\narchitectures for integrating knowledge into text generation; (ii) specific\ntechniques and applications according to different forms of knowledge data.\nThis survey can have broad audiences, researchers and practitioners, in\nacademia and industry.', 'Abstractive text summarization aims to shorten long text documents into a\nhuman readable form that contains the most important facts from the original\ndocument. However, the level of actual abstraction as measured by novel phrases\nthat do not appear in the source document remains low in existing approaches.\nWe propose two techniques to improve the level of abstraction of generated\nsummaries. First, we decompose the decoder into a contextual network that\nretrieves relevant parts of the source document, and a pretrained language\nmodel that incorporates prior knowledge about language generation. Second, we\npropose a novelty metric that is optimized directly through policy learning to\nencourage the generation of novel phrases. Our model achieves results\ncomparable to state-of-the-art models, as determined by ROUGE scores and human\nevaluations, while achieving a significantly higher level of abstraction as\nmeasured by n-gram overlap with the source document.', 'The advent of large pre-trained language models has made it possible to make\nhigh-quality predictions on how to add or change a sentence in a document.\nHowever, the high branching factor inherent to text generation impedes the\nability of even the strongest language models to offer useful editing\nsuggestions at a more global or document level. We introduce a new task,\ndocument sketching, which involves generating entire draft documents for the\nwriter to review and revise. These drafts are built from sets of documents that\noverlap in form - sharing large segments of potentially reusable text - while\ndiverging in content. To support this task, we introduce a Wikipedia-based\ndataset of analogous documents and investigate the application of weakly\nsupervised methods, including use of a transformer-based mixture of experts,\ntogether with reinforcement learning. We report experiments using automated and\nhuman evaluation methods and discuss relative merits of these models.', 'Recent deep generative models are able to provide photo-realistic images as\nwell as visual or textual content embeddings useful to address various tasks of\ncomputer vision and natural language processing. Their usefulness is\nnevertheless often limited by the lack of control over the generative process\nor the poor understanding of the learned representation. To overcome these\nmajor issues, very recent work has shown the interest of studying the semantics\nof the latent space of generative models. In this paper, we propose to advance\non the interpretability of the latent space of generative models by introducing\na new method to find meaningful directions in the latent space of any\ngenerative model along which we can move to control precisely specific\nproperties of the generated image like the position or scale of the object in\nthe image. Our method does not require human annotations and is particularly\nwell suited for the search of directions encoding simple transformations of the\ngenerated image, such as translation, zoom or color variations. We demonstrate\nthe effectiveness of our method qualitatively and quantitatively, both for GANs\nand variational auto-encoders.', ""The adoption of natural language generation (NLG) models can leave\nindividuals vulnerable to the generation of harmful information memorized by\nthe models, such as conspiracy theories. While previous studies examine\nconspiracy theories in the context of social media, they have not evaluated\ntheir presence in the new space of generative language models. In this work, we\ninvestigate the capability of language models to generate conspiracy theory\ntext. Specifically, we aim to answer: can we test pretrained generative\nlanguage models for the memorization and elicitation of conspiracy theories\nwithout access to the model's training data? We highlight the difficulties of\nthis task and discuss it in the context of memorization, generalization, and\nhallucination. Utilizing a new dataset consisting of conspiracy theory topics\nand machine-generated conspiracy theories helps us discover that many\nconspiracy theories are deeply rooted in the pretrained language models. Our\nexperiments demonstrate a relationship between model parameters such as size\nand temperature and their propensity to generate conspiracy theory text. These\nresults indicate the need for a more thorough review of NLG applications before\nrelease and an in-depth discussion of the drawbacks of memorization in\ngenerative language models."", 'Some consider large-scale language models that can generate long and coherent\npieces of text as dangerous, since they may be used in misinformation\ncampaigns. Here we formulate large-scale language model output detection as a\nhypothesis testing problem to classify text as genuine or generated. We show\nthat error exponents for particular language models are bounded in terms of\ntheir perplexity, a standard measure of language generation performance. Under\nthe assumption that human language is stationary and ergodic, the formulation\nis extended from considering specific language models to considering maximum\nlikelihood language models, among the class of k-order Markov approximations;\nerror probabilities are characterized. Some discussion of incorporating\nsemantic side information is also given.', 'Recent advances in natural language generation have introduced powerful\nlanguage models with high-quality output text. However, this raises concerns\nabout the potential misuse of such models for malicious purposes. In this\npaper, we study natural language watermarking as a defense to help better mark\nand trace the provenance of text. We introduce the Adversarial Watermarking\nTransformer (AWT) with a jointly trained encoder-decoder and adversarial\ntraining that, given an input text and a binary message, generates an output\ntext that is unobtrusively encoded with the given message. We further study\ndifferent training and inference strategies to achieve minimal changes to the\nsemantics and correctness of the input text.\n  AWT is the first end-to-end model to hide data in text by automatically\nlearning -- without ground truth -- word substitutions along with their\nlocations in order to encode the message. We empirically show that our model is\neffective in largely preserving text utility and decoding the watermark while\nhiding its presence against adversaries. Additionally, we demonstrate that our\nmethod is robust against a range of attacks.', 'Sentiment analysis is a task that may suffer from a lack of data in certain\ncases, as the datasets are often generated and annotated by humans. In cases\nwhere data is inadequate for training discriminative models, generate models\nmay aid training via data augmentation. Generative Adversarial Networks (GANs)\nare one such model that has advanced the state of the art in several tasks,\nincluding as image and text generation. In this paper, I train GAN models on\nlow resource datasets, then use them for the purpose of data augmentation\ntowards improving sentiment classifier generalization. Given the constraints of\nlimited data, I explore various techniques to train the GAN models. I also\npresent an analysis of the quality of generated GAN data as more training data\nfor the GAN is made available. In this analysis, the generated data is\nevaluated as a test set (against a model trained on real data points) as well\nas a training set to train classification models. Finally, I also conduct a\nvisual analysis by projecting the generated and the real data into a\ntwo-dimensional space using the t-Distributed Stochastic Neighbor Embedding\n(t-SNE) method.', 'The recent explosion of machine learning (ML) and artificial intelligence\n(AI) shows great potential in the breakthrough of metal additive manufacturing\n(AM) process modeling. However, the success of conventional machine learning\ntools in data science is primarily attributed to the unprecedented large amount\nof labeled data-sets (big data), which can be either obtained by experiments or\nfirst-principle simulations. Unfortunately, these labeled data-sets are\nexpensive to obtain in AM due to the high expense of the AM experiments and\nprohibitive computational cost of high-fidelity simulations.\n  We propose a physics-informed neural network (PINN) framework that fuses both\ndata and first physical principles, including conservation laws of momentum,\nmass, and energy, into the neural network to inform the learning processes. To\nthe best knowledge of the authors, this is the first application of PINN to\nthree dimensional AM processes modeling. Besides, we propose a hard-type\napproach for Dirichlet boundary conditions (BCs) based on a Heaviside function,\nwhich can not only enforce the BCs but also accelerate the learning process.\nThe PINN framework is applied to two representative metal manufacturing\nproblems, including the 2018 NIST AM-Benchmark test series. We carefully assess\nthe performance of the PINN model by comparing the predictions with available\nexperimental data and high-fidelity simulation results. The investigations show\nthat the PINN, owed to the additional physical knowledge, can accurately\npredict the temperature and melt pool dynamics during metal AM processes with\nonly a moderate amount of labeled data-sets. The foray of PINN to metal AM\nshows the great potential of physics-informed deep learning for broader\napplications to advanced manufacturing.', 'A lot of work has been done to build text-based language models for\nperforming different NLP tasks, but not much research has been done in the case\nof audio-based language models. This paper proposes a Convolutional Autoencoder\nbased neural architecture to model syntactically and semantically adequate\ncontextualized representations of varying length spoken words. The use of such\nrepresentations can not only lead to great advances in the audio-based NLP\ntasks but can also curtail the loss of information like tone, expression,\naccent, etc while converting speech to text to perform these tasks. The\nperformance of the proposed model is validated by (1) examining the generated\nvector space, and (2) evaluating its performance on three benchmark datasets\nfor measuring word similarities, against existing widely used text-based\nlanguage models that are trained on the transcriptions. The proposed model was\nable to demonstrate its robustness when compared to the other two\nlanguage-based models.', ""We investigate neural techniques for end-to-end computational argumentation\nmining (AM). We frame AM both as a token-based dependency parsing and as a\ntoken-based sequence tagging problem, including a multi-task learning setup.\nContrary to models that operate on the argument component level, we find that\nframing AM as dependency parsing leads to subpar performance results. In\ncontrast, less complex (local) tagging models based on BiLSTMs perform robustly\nacross classification scenarios, being able to catch long-range dependencies\ninherent to the AM problem. Moreover, we find that jointly learning 'natural'\nsubtasks, in a multi-task learning setup, improves performance."", 'The performance of text classification methods has improved greatly over the\nlast decade for text instances of less than 512 tokens. This limit has been\nadopted by most state-of-the-research transformer models due to the high\ncomputational cost of analyzing longer text instances. To mitigate this problem\nand to improve classification for longer texts, researchers have sought to\nresolve the underlying causes of the computational cost and have proposed\noptimizations for the attention mechanism, which is the key element of every\ntransformer model. In our study, we are not pursuing the ultimate goal of long\ntext classification, i.e., the ability to analyze entire text instances at one\ntime while preserving high performance at a reasonable computational cost.\nInstead, we propose a text truncation method called Text Guide, in which the\noriginal text length is reduced to a predefined limit in a manner that improves\nperformance over naive and semi-naive approaches while preserving low\ncomputational costs. Text Guide benefits from the concept of feature\nimportance, a notion from the explainable artificial intelligence domain. We\ndemonstrate that Text Guide can be used to improve the performance of recent\nlanguage models specifically designed for long text classification, such as\nLongformer. Moreover, we discovered that parameter optimization is the key to\nText Guide performance and must be conducted before the method is deployed.\nFuture experiments may reveal additional benefits provided by this new method.', 'Natural language processing (NLP) techniques have become mainstream in the\nrecent decade. Most of these advances are attributed to the processing of a\nsingle language. More recently, with the extensive growth of social media\nplatforms focus has shifted to code-mixed text. The code-mixed text comprises\ntext written in more than one language. People naturally tend to combine local\nlanguage with global languages like English. To process such texts, current NLP\ntechniques are not sufficient. As a first step, the text is processed to\nidentify the language of the words in the text. In this work, we focus on\nlanguage identification in code-mixed sentences for Hindi-English mixed text.\nThe task of language identification is formulated as a token classification\ntask. In the supervised setting, each word in the sentence has an associated\nlanguage label. We evaluate different deep learning models and input\nrepresentation combinations for this task. Mainly, character, sub-word, and\nword embeddings are considered in combination with CNN and LSTM based models.\nWe show that sub-word representation along with the LSTM model gives the best\nresults. In general sub-word representations perform significantly better than\nother input representations. We report the best accuracy of 94.52% using a\nsingle layer LSTM model on the standard SAIL ICON 2017 test set.', 'Vagueness and uncertainty management is counted among one of the challenges\nthat remain unresolved in systems that generate texts from non-linguistic data,\nknown as data-to-text systems. In the last decade, work in fuzzy linguistic\nsummarization and description of data has raised the interest of using fuzzy\nsets to model and manage the imprecision of human language in data-to-text\nsystems. However, despite some research in this direction, there has not been\nan actual clear discussion and justification on how fuzzy sets can contribute\nto data-to-text for modeling vagueness and uncertainty in words and\nexpressions. This paper intends to bridge this gap by answering the following\nquestions: What does vagueness mean in fuzzy sets theory? What does vagueness\nmean in data-to-text contexts? In what ways can fuzzy sets theory contribute to\nimprove data-to-text systems? What are the challenges that researchers from\nboth disciplines need to address for a successful integration of fuzzy sets\ninto data-to-text systems? In what cases should the use of fuzzy sets be\navoided in D2T? For this, we review and discuss the state of the art of\nvagueness modeling in natural language generation and data-to-text, describe\npotential and actual usages of fuzzy sets in data-to-text contexts, and provide\nsome additional insights about the engineering of data-to-text systems that\nmake use of fuzzy set-based techniques.', 'We investigate whether and where multi-task learning (MTL) can improve\nperformance on NLP problems related to argumentation mining (AM), in particular\nargument component identification. Our results show that MTL performs\nparticularly well (and better than single-task learning) when little training\ndata is available for the main task, a common scenario in AM. Our findings\nchallenge previous assumptions that conceptualizations across AM datasets are\ndivergent and that MTL is difficult for semantic or higher-level tasks.', 'Traditionally, text generation models take in a sequence of text as input,\nand iteratively generate the next most probable word using pre-trained\nparameters. In this work, we propose the architecture to use images instead of\ntext as the input of the text generation model, called StoryGen. In the\narchitecture, we design a Relational Text Data Generator algorithm that relates\ndifferent features from multiple images. The output samples from the model\ndemonstrate the ability to generate meaningful paragraphs of text containing\nthe extracted features from the input images. This is an undergraduate project\nreport. Completed Dec. 2019 at the Cooper Union.', 'A deep neural network (DNN)-based speech enhancement (SE) aiming to maximize\nthe performance of an automatic speech recognition (ASR) system is proposed in\nthis paper. In order to optimize the DNN-based SE model in terms of the\ncharacter error rate (CER), which is one of the metric to evaluate the ASR\nsystem and generally non-differentiable, our method uses two DNNs: one for\nspeech processing and one for mimicking the output CERs derived through an\nacoustic model (AM). Then both of DNNs are alternately optimized in the\ntraining phase. Even if the AM is a black-box, e.g., like one provided by a\nthird-party, the proposed method enables the DNN-based SE model to be optimized\nin terms of the CER since the DNN mimicking the AM is differentiable.\nConsequently, it becomes feasible to build CER-centric SE model that has no\nnegative effect, e.g., additional calculation cost and changing network\narchitecture, on the inference phase since our method is merely a training\nscheme for the existing DNN-based methods. Experimental results show that our\nmethod improved CER by 8.8% relative derived through a black-box AM although\ncertain noise levels are kept.', 'Most work in machine reading focuses on question answering problems where the\nanswer is directly expressed in the text to read. However, many real-world\nquestion answering problems require the reading of text not because it contains\nthe literal answer, but because it contains a recipe to derive an answer\ntogether with the reader\'s background knowledge. One example is the task of\ninterpreting regulations to answer ""Can I...?"" or ""Do I have to...?"" questions\nsuch as ""I am working in Canada. Do I have to carry on paying UK National\nInsurance?"" after reading a UK government website about this topic. This task\nrequires both the interpretation of rules and the application of background\nknowledge. It is further complicated due to the fact that, in practice, most\nquestions are underspecified, and a human assistant will regularly have to ask\nclarification questions such as ""How long have you been working abroad?"" when\nthe answer cannot be directly derived from the question and text. In this\npaper, we formalise this task and develop a crowd-sourcing strategy to collect\n32k task instances based on real-world rules and crowd-generated questions and\nscenarios. We analyse the challenges of this task and assess its difficulty by\nevaluating the performance of rule-based and machine-learning baselines. We\nobserve promising results when no background knowledge is necessary, and\nsubstantial room for improvement whenever background knowledge is needed.', 'We present an approach to analyze $C^1(\\mathbb{R}^m)$ functions that\naddresses limitations present in the Active Subspaces (AS) method of\nConstantine et al.(2015; 2014). Under appropriate hypotheses, our Active\nManifolds (AM) method identifies a 1-D curve in the domain (the active\nmanifold) on which nearly all values of the unknown function are attained, and\nwhich can be exploited for approximation or analysis, especially when $m$ is\nlarge (high-dimensional input space). We provide theorems justifying our AM\ntechnique and an algorithm permitting functional approximation and sensitivity\nanalysis. Using accessible, low-dimensional functions as initial examples, we\nshow AM reduces approximation error by an order of magnitude compared to AS, at\nthe expense of more computation. Following this, we revisit the sensitivity\nanalysis by Glaws et al. (2017), who apply AS to analyze a magnetohydrodynamic\npower generator model, and compare the performance of AM on the same data. Our\nanalysis provides detailed information not captured by AS, exhibiting the\ninfluence of each parameter individually along an active manifold. Overall, AM\nrepresents a novel technique for analyzing functional models with benefits\nincluding: reducing $m$-dimensional analysis to a 1-D analogue, permitting more\naccurate regression than AS (at more computational expense), enabling more\ninformative sensitivity analysis, and granting accessible visualizations(2-D\nplots) of parameter sensitivity along the AM.', 'This article describes an unsupervised language model adaptation approach\nthat can be used to enhance the performance of language identification methods.\nThe approach is applied to a current version of the HeLI language\nidentification method, which is now called HeLI 2.0. We describe the HeLI 2.0\nmethod in detail. The resulting system is evaluated using the datasets from the\nGerman dialect identification and Indo-Aryan language identification shared\ntasks of the VarDial workshops 2017 and 2018. The new approach with language\nidentification provides considerably higher F1-scores than the previous HeLI\nmethod or the other systems which participated in the shared tasks. The results\nindicate that unsupervised language model adaptation should be considered as an\noption in all language identification tasks, especially in those where\nencountering out-of-domain data is likely.', 'In the era of big data, a large number of text data generated by the Internet\nhas given birth to a variety of text representation methods. In natural\nlanguage processing (NLP), text representation transforms text into vectors\nthat can be processed by computer without losing the original semantic\ninformation. However, these methods are difficult to effectively extract the\nsemantic features among words and distinguish polysemy in language. Therefore,\na text feature representation model based on convolutional neural network (CNN)\nand variational autoencoder (VAE) is proposed to extract the text features and\napply the obtained text feature representation on the text classification\ntasks. CNN is used to extract the features of text vector to get the semantics\namong words and VAE is introduced to make the text feature space more\nconsistent with Gaussian distribution. In addition, the output of the improved\nword2vec model is employed as the input of the proposed model to distinguish\ndifferent meanings of the same word in different contexts. The experimental\nresults show that the proposed model outperforms in k-nearest neighbor (KNN),\nrandom forest (RF) and support vector machine (SVM) classification algorithms.', ""The focus of my PhD thesis is on exploring parallel approaches to efficiently\nsolve problems modeled by constraints and presenting a new proposal. Current\nsolvers are very advanced; they are carefully designed to effectively manage\nthe high-level problems' description and include refined strategies to avoid\nuseless work. Despite this, finding a solution can take an unacceptable amount\nof time. Parallelization can mitigate this problem when the instance of the\nproblem modeled is large, as it happens in real world problems. It is done by\npropagating constraints in parallel and concurrently exploring different parts\nof the search space. I am developing on a constraint solver that exploits the\nmany cores available on Graphics Processing Units (GPU) to speed up the search."", 'The ability to accurately quantify the performance an additively manufactured\n(AM) product is important for a widespread industry adoption of AM as the\ndesign is required to: (1) satisfy geometrical constraints, (2) satisfy\nstructural constraints dictated by its intended function, and (3) be cost\neffective compared to traditional manufacturing methods. Optimization\ntechniques offer design aids in creating cost-effective structures that meet\nthe prescribed structural objectives. The fundamental problem in existing\napproaches lies in the difficulty to quantify the structural performance as\neach unique design leads to a new set of analyses to determine the structural\nrobustness and such analyses can be very costly due to the complexity of in-use\nforces experienced by the structure. This work develops computationally\ntractable methods tailored to maximize the structural performance of AM\nproducts. A geometry preserving build orientation optimization method as well\nas data-driven shape optimization approaches to structural design are\npresented. Proposed methods greatly enhance the value of AM technology by\ntaking advantage of the design space enabled by it for a broad class of\nproblems involving complex in-use loads.', ""Hybrid manufacturing (HM) technologies combine additive and subtractive\nmanufacturing (AM/SM) capabilities, leveraging AM's strengths in fabricating\ncomplex geometries and SM's precision and quality to produce finished parts. We\npresent a systematic approach to automated computer-aided process planning\n(CAPP) for HM that can identify non-trivial, qualitatively distinct, and\ncost-optimal combinations of AM/SM modalities. A multimodal HM process plan is\nrepresented by a finite Boolean expression of AM and SM manufacturing\nprimitives, such that the expression evaluates to an 'as-manufactured'\nartifact. We show that primitives that respect spatial constraints such as\naccessibility and collision avoidance may be constructed by solving inverse\nconfiguration space problems on the 'as-designed' artifact and manufacturing\ninstruments. The primitives generate a finite Boolean algebra (FBA) that\nenumerates the entire search space for planning. The FBA's canonical\nintersection terms (i.e., 'atoms') provide the complete domain decomposition to\nreframe manufacturability analysis and process planning into purely symbolic\nreasoning, once a subcollection of atoms is found to be interchangeable with\nthe design target. The approach subsumes unimodal (all-AM or all-SM) process\nplanning as special cases. We demonstrate the practical potency of our\nframework and its computational efficiency when applied to process planning of\ncomplex 3D parts with dramatically different AM and SM instruments."", 'One of the major challenges in training text-to-image generation models is\nthe need of a large number of high-quality image-text pairs. While image\nsamples are often easily accessible, the associated text descriptions typically\nrequire careful human captioning, which is particularly time- and\ncost-consuming. In this paper, we propose the first work to train text-to-image\ngeneration models without any text data. Our method leverages the well-aligned\nmulti-modal semantic space of the powerful pre-trained CLIP model: the\nrequirement of text-conditioning is seamlessly alleviated via generating text\nfeatures from image features. Extensive experiments are conducted to illustrate\nthe effectiveness of the proposed method. We obtain state-of-the-art results in\nthe standard text-to-image generation tasks. Importantly, the proposed\nlanguage-free model outperforms most existing models trained with full\nimage-text pairs. Furthermore, our method can be applied in fine-tuning\npre-trained models, which saves both training time and cost in training\ntext-to-image generation models. Our pre-trained model obtains competitive\nresults in zero-shot text-to-image generation on the MS-COCO dataset, yet with\naround only 1% of the model size and training data size relative to the\nrecently proposed large DALL-E model.', 'Speech-driven visual speech synthesis involves mapping features extracted\nfrom acoustic speech to the corresponding lip animation controls for a face\nmodel. This mapping can take many forms, but a powerful approach is to use deep\nneural networks (DNNs). However, a limitation is the lack of synchronized\naudio, video, and depth data required to reliably train the DNNs, especially\nfor speaker-independent models. In this paper, we investigate adapting an\nautomatic speech recognition (ASR) acoustic model (AM) for the visual speech\nsynthesis problem. We train the AM on ten thousand hours of audio-only data.\nThe AM is then adapted to the visual speech synthesis domain using ninety hours\nof synchronized audio-visual speech. Using a subjective assessment test, we\ncompared the performance of the AM-initialized DNN to one with a random\ninitialization. The results show that viewers significantly prefer animations\ngenerated from the AM-initialized DNN than the ones generated using the\nrandomly initialized model. We conclude that visual speech synthesis can\nsignificantly benefit from the powerful representation of speech in the ASR\nacoustic models.', 'In realistic environments, speech is usually interfered by various noise and\nreverberation, which dramatically degrades the performance of automatic speech\nrecognition (ASR) systems. To alleviate this issue, the commonest way is to use\na well-designed speech enhancement approach as the front-end of ASR. However,\nmore complex pipelines, more computations and even higher hardware costs\n(microphone array) are additionally consumed for this kind of methods. In\naddition, speech enhancement would result in speech distortions and mismatches\nto training. In this paper, we propose an adversarial training method to\ndirectly boost noise robustness of acoustic model. Specifically, a jointly\ncompositional scheme of generative adversarial net (GAN) and neural\nnetwork-based acoustic model (AM) is used in the training phase. GAN is used to\ngenerate clean feature representations from noisy features by the guidance of a\ndiscriminator that tries to distinguish between the true clean signals and\ngenerated signals. The joint optimization of generator, discriminator and AM\nconcentrates the strengths of both GAN and AM for speech recognition.\nSystematic experiments on CHiME-4 show that the proposed method significantly\nimproves the noise robustness of AM and achieves the average relative error\nrate reduction of 23.38% and 11.54% on the development and test set,\nrespectively.', 'AM dependency parsing is a method for neural semantic graph parsing that\nexploits the principle of compositionality. While AM dependency parsers have\nbeen shown to be fast and accurate across several graphbanks, they require\nexplicit annotations of the compositional tree structures for training. In the\npast, these were obtained using complex graphbank-specific heuristics written\nby experts. Here we show how they can instead be trained directly on the graphs\nwith a neural latent-variable model, drastically reducing the amount and\ncomplexity of manual heuristics. We demonstrate that our model picks up on\nseveral linguistic phenomena on its own and achieves comparable accuracy to\nsupervised training, greatly facilitating the use of AM dependency parsing for\nnew sembanks.', 'Emotion stimulus extraction is a fine-grained subtask of emotion analysis\nthat focuses on identifying the description of the cause behind an emotion\nexpression from a text passage (e.g., in the sentence ""I am happy that I passed\nmy exam"" the phrase ""passed my exam"" corresponds to the stimulus.). Previous\nwork mainly focused on Mandarin and English, with no resources or models for\nGerman. We fill this research gap by developing a corpus of 2006 German news\nheadlines annotated with emotions and 811 instances with annotations of\nstimulus phrases. Given that such corpus creation efforts are time-consuming\nand expensive, we additionally work on an approach for projecting the existing\nEnglish GoodNewsEveryone (GNE) corpus to a machine-translated German version.\nWe compare the performance of a conditional random field (CRF) model (trained\nmonolingually on German and cross-lingually via projection) with a multilingual\nXLM-RoBERTa (XLM-R) model. Our results show that training with the German\ncorpus achieves higher F1 scores than projection. Experiments with XLM-R\noutperform their respective CRF counterparts.', 'Argumentation mining (AM) requires the identification of complex discourse\nstructures and has lately been applied with success monolingually. In this\nwork, we show that the existing resources are, however, not adequate for\nassessing cross-lingual AM, due to their heterogeneity or lack of complexity.\nWe therefore create suitable parallel corpora by (human and machine)\ntranslating a popular AM dataset consisting of persuasive student essays into\nGerman, French, Spanish, and Chinese. We then compare (i) annotation projection\nand (ii) bilingual word embeddings based direct transfer strategies for\ncross-lingual AM, finding that the former performs considerably better and\nalmost eliminates the loss from cross-lingual transfer. Moreover, we find that\nannotation projection works equally well when using either costly human or\ncheap machine translations. Our code and data are available at\n\\url{http://github.com/UKPLab/coling2018-xling_argument_mining}.', 'In this study, listeners of varied Indian nativities are asked to listen and\nrecognize TIMIT utterances spoken by American speakers. We have three kinds of\nresponses from each listener while they recognize an utterance: 1. Sentence\ndifficulty ratings, 2. Speaker difficulty ratings, and 3. Transcription of the\nutterance. From these transcriptions, word error rate (WER) is calculated and\nused as a metric to evaluate the similarity between the recognized and the\noriginal sentences.The sentences selected in this study are categorized into\nthree groups: Easy, Medium and Hard, based on the frequency ofoccurrence of the\nwords in them. We observe that the sentence, speaker difficulty ratings and the\nWERs increase from easy to hard categories of sentences. We also compare the\nhuman speech recognition performance with that using three automatic speech\nrecognition (ASR) under following three combinations of acoustic model (AM) and\nlanguage model(LM): ASR1) AM trained with recordings from speakers of Indian\norigin and LM built on TIMIT text, ASR2) AM using recordings from native\nAmerican speakers and LM built ontext from LIBRI speech corpus, and ASR3) AM\nusing recordings from native American speakers and LM build on LIBRI speech and\nTIMIT text. We observe that HSR performance is similar to that of ASR1 whereas\nASR3 achieves the best performance. Speaker nativity wise analysis shows that\nutterances from speakers of some nativity are more difficult to recognize by\nIndian listeners compared to few other nativities', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', 'The additive margin softmax (AM-Softmax) loss has delivered remarkable\nperformance in speaker verification. A supposed behavior of AM-Softmax is that\nit can shrink within-class variation by putting emphasis on target logits,\nwhich in turn improves margin between target and non-target classes. In this\npaper, we conduct a careful analysis on the behavior of AM-Softmax loss, and\nshow that this loss does not implement real max-margin training. Based on this\nobservation, we present a Real AM-Softmax loss which involves a true margin\nfunction in the softmax training. Experiments conducted on VoxCeleb1, SITW and\nCNCeleb demonstrated that the corrected AM-Softmax loss consistently\noutperforms the original one. The code has been released at\nhttps://gitlab.com/csltstu/sunine.', 'The present document is an excerpt of an essay that I wrote as part of my\napplication material to graduate school in Computer Science (with a focus on\nArtificial Intelligence), in 1986. I was not invited by any of the schools that\nreceived it, so I became a theoretical physicist instead. The essay\'s full\ntitle was ""Some Topics in Philosophy and Computer Science"". I am making this\ntext (unchanged from 1985, preserving the typesetting as much as possible)\navailable now in memory of Jerry Fodor, whose writings had influenced me\nsignificantly at the time (even though I did not always agree).', ""In this paper, I will discuss the work I am currently doing as a Ph.D.\nstudent at the University of Potsdam, under the tutoring of T. Schaub. I'm\ncurrently looking into action description in ASP. More precisely, my goal is to\nexplore how to represent actions with durations in ASP, in different contexts.\nRight now, I'm focused on Multi-Agent Path Finding (MAPF), looking at how to\nrepresent speeds for different agents and contexts.\n  Before tackling duration, I wanted to explore and compare different\nrepresentations of action taking in ASP. For this, I started comparing\ndifferent simple encodings tackling the MAPF problem. Even in simple code,\nchoices and assumptions have been made in their creations. The objective of my\nwork is to present the consequences of those design decisions in terms of\nperformance and knowledge representation. As far as I know, there is no current\nresearch on this topic.\n  Besides that, I'm also exploring different ways to represent duration and to\nsolve related problems. I planed to compare them the same way I described\nbefore. I also want this to help me find innovative and effective ways to solve\nproblems with duration."", 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'Additive Manufacturing (AM) is a manufacturing paradigm that builds\nthree-dimensional objects from a computer-aided design model by successively\nadding material layer by layer. AM has become very popular in the past decade\ndue to its utility for fast prototyping such as 3D printing as well as\nmanufacturing functional parts with complex geometries using processes such as\nlaser metal deposition that would be difficult to create using traditional\nmachining. As the process for creating an intricate part for an expensive metal\nsuch as Titanium is prohibitive with respect to cost, computational models are\nused to simulate the behavior of AM processes before the experimental run.\nHowever, as the simulations are computationally costly and time-consuming for\npredicting multiscale multi-physics phenomena in AM, physics-informed\ndata-driven machine-learning systems for predicting the behavior of AM\nprocesses are immensely beneficial. Such models accelerate not only multiscale\nsimulation tools but also empower real-time control systems using in-situ data.\nIn this paper, we design and develop essential components of a scientific\nframework for developing a data-driven model-based real-time control system.\nFinite element methods are employed for solving time-dependent heat equations\nand developing the database. The proposed framework uses extremely randomized\ntrees - an ensemble of bagged decision trees as the regression algorithm\niteratively using temperatures of prior voxels and laser information as inputs\nto predict temperatures of subsequent voxels. The models achieve mean absolute\npercentage errors below 1% for predicting temperature profiles for AM\nprocesses. The code is made available for the research community at\nhttps://github.com/paularindam/ml-iter-additive.', 'We address the problem of solving mixed random linear equations. We have\nunlabeled observations coming from multiple linear regressions, and each\nobservation corresponds to exactly one of the regression models. The goal is to\nlearn the linear regressors from the observations. Classically, Alternating\nMinimization (AM) (which is a variant of Expectation Maximization (EM)) is used\nto solve this problem. AM iteratively alternates between the estimation of\nlabels and solving the regression problems with the estimated labels.\nEmpirically, it is observed that, for a large variety of non-convex problems\nincluding mixed linear regression, AM converges at a much faster rate compared\nto gradient based algorithms. However, the existing theory suggests similar\nrate of convergence for AM and gradient based methods, failing to capture this\nempirical behavior. In this paper, we close this gap between theory and\npractice for the special case of a mixture of $2$ linear regressions. We show\nthat, provided initialized properly, AM enjoys a \\emph{super-linear} rate of\nconvergence in certain parameter regimes. To the best of our knowledge, this is\nthe first work that theoretically establishes such rate for AM. Hence, if we\nwant to recover the unknown regressors upto an error (in $\\ell_2$ norm) of\n$\\epsilon$, AM only takes $\\mathcal{O}(\\log \\log (1/\\epsilon))$ iterations.\nFurthermore, we compare AM with a gradient based heuristic algorithm\nempirically and show that AM dominates in iteration complexity as well as\nwall-clock time.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently developing a chatbot capable of engaging in open-ended conversations with users. However, I am dissatisfied with the conventional temperature-based stochastic sampling decoding procedure, as it results in derivative output that lacks creativity and diversity. I am seeking literature that explores the issue of diversity deficiency in language model decoding procedures. Additionally, I aim to create a decoding scheme that strikes a balance between output diversity and quality. I am also interested in research that examines the phenomenon of high-likelihood sentences exhibiting unexpectedly low quality. I wish to understand this occurrence and prevent it from affecting my decoding scheme. Ultimately, my goal is to develop a model that surpasses existing decoding algorithms in terms of balancing the quality and diversity of generated responses.","[""A conversational agent (chatbot) is a piece of software that is able to\ncommunicate with humans using natural language. Modeling conversation is an\nimportant task in natural language processing and artificial intelligence.\nWhile chatbots can be used for various tasks, in general they have to\nunderstand users' utterances and provide responses that are relevant to the\nproblem at hand.\n  In my work, I conduct an in-depth survey of recent literature, examining over\n70 publications related to chatbots published in the last 3 years. Then, I\nproceed to make the argument that the very nature of the general conversation\ndomain demands approaches that are different from current state-of-of-the-art\narchitectures. Based on several examples from the literature I show why current\nchatbot models fail to take into account enough priors when generating\nresponses and how this affects the quality of the conversation. In the case of\nchatbots, these priors can be outside sources of information that the\nconversation is conditioned on like the persona or mood of the conversers. In\naddition to presenting the reasons behind this problem, I propose several ideas\non how it could be remedied.\n  The next section focuses on adapting the very recent Transformer model to the\nchatbot domain, which is currently state-of-the-art in neural machine\ntranslation. I first present experiments with the vanilla model, using\nconversations extracted from the Cornell Movie-Dialog Corpus. Secondly, I\naugment the model with some of my ideas regarding the issues of encoder-decoder\narchitectures. More specifically, I feed additional features into the model\nlike mood or persona together with the raw conversation data. Finally, I\nconduct a detailed analysis of how the vanilla model performs on conversational\ndata by comparing it to previous chatbot models and how the additional features\naffect the quality of the generated responses."", 'Neural text generation models are likely to suffer from the low-diversity\nproblem. Various decoding strategies and training-based methods have been\nproposed to promote diversity only by exploiting contextual features, but\nrarely do they consider incorporating syntactic structure clues. In this work,\nwe propose using linguistic annotation, i.e., part-of-speech (POS), to guide\nthe text generation. In detail, we introduce POS Guided Softmax to explicitly\nmodel two posterior probabilities: (i) next-POS, and (ii) next-token from the\nvocabulary of the target POS. A POS Guided Sampling strategy is further\nproposed to address the low-diversity problem by enriching the diversity of\nPOS. Extensive experiments and human evaluations show that, compared with\nexisting state-of-the-art methods, our POS Guided Softmax and Sampling (POSG)\ncan generate more diverse text while maintaining comparable quality.', 'Large pre-trained language models have repeatedly shown their ability to\nproduce fluent text. Yet even when starting from a prompt, generation can\ncontinue in many plausible directions. Current decoding methods with the goal\nof controlling generation, e.g., to ensure specific words are included, either\nrequire additional models or fine-tuning, or work poorly when the task at hand\nis semantically unconstrained, e.g., story generation. In this work, we present\na plug-and-play decoding method for controlled language generation that is so\nsimple and intuitive, it can be described in a single sentence: given a topic\nor keyword, we add a shift to the probability distribution over our vocabulary\ntowards semantically similar words. We show how annealing this distribution can\nbe used to impose hard constraints on language generation, something no other\nplug-and-play method is currently able to do with SOTA language generators.\nDespite the simplicity of this approach, we see it works incredibly well in\npractice: decoding from GPT-2 leads to diverse and fluent sentences while\nguaranteeing the appearance of given guide words. We perform two user studies,\nrevealing that (1) our method outperforms competing methods in human\nevaluations; and (2) forcing the guide words to appear in the generated text\nhas no impact on the fluency of the generated text.', 'Many existing conversation models that are based on the encoder-decoder\nframework have focused on ways to make the encoder more complicated to enrich\nthe context vectors so as to increase the diversity and informativeness of\ngenerated responses. However, these approaches face two problems. First, the\ndecoder is too simple to effectively utilize the previously generated\ninformation and tends to generate duplicated and self-contradicting responses.\nSecond, the complex encoder tends to generate diverse but incoherent responses\nbecause the complex context vectors may deviate from the original semantics of\ncontext. In this work, we proposed a conversation model named ""THINK"" (Teamwork\ngeneration Hover around Impressive Noticeable Keywords) to make the decoder\nmore complicated and avoid generating duplicated and self-contradicting\nresponses. The model simplifies the context vectors and increases the coherence\nof generated responses in a reasonable way. For this model, we propose Teamwork\ngeneration framework and Semantics Extractor. Compared with other baselines,\nboth automatic and human evaluation showed the advantages of our model.', ""Deep-learning models for language generation tasks tend to produce repetitive\noutput. Various methods have been proposed to encourage lexical diversity\nduring decoding, but this often comes at a cost to the perceived fluency and\nadequacy of the output. In this work, we propose to ameliorate this cost by\nusing an Imitation Learning approach to explore the level of diversity that a\nlanguage generation model can reliably produce. Specifically, we augment the\ndecoding process with a meta-classifier trained to distinguish which words at\nany given timestep will lead to high-quality output. We focus our experiments\non concept-to-text generation where models are sensitive to the inclusion of\nirrelevant words due to the strict relation between input and output. Our\nanalysis shows that previous methods for diversity underperform in this\nsetting, while human evaluation suggests that our proposed method achieves a\nhigh level of diversity with minimal effect to the output's fluency and\nadequacy."", 'We study response generation for open domain conversation in chatbots.\nExisting methods assume that words in responses are generated from an identical\nvocabulary regardless of their inputs, which not only makes them vulnerable to\ngeneric patterns and irrelevant noise, but also causes a high cost in decoding.\nWe propose a dynamic vocabulary sequence-to-sequence (DVS2S) model which allows\neach input to possess their own vocabulary in decoding. In training, vocabulary\nconstruction and response generation are jointly learned by maximizing a lower\nbound of the true objective with a Monte Carlo sampling method. In inference,\nthe model dynamically allocates a small vocabulary for an input with the word\nprediction model, and conducts decoding only with the small vocabulary. Because\nof the dynamic vocabulary mechanism, DVS2S eludes many generic patterns and\nirrelevant words in generation, and enjoys efficient decoding at the same time.\nExperimental results on both automatic metrics and human annotations show that\nDVS2S can significantly outperform state-of-the-art methods in terms of\nresponse quality, but only requires 60% decoding time compared to the most\nefficient baseline.', 'Despite considerable advancements with deep neural language models, the\nenigma of neural text degeneration persists when these models are tested as\ntext generators. The counter-intuitive empirical observation is that even\nthough the use of likelihood as training objective leads to high quality models\nfor a broad range of language understanding tasks, using likelihood as a\ndecoding objective leads to text that is bland and strangely repetitive.\n  In this paper, we reveal surprising distributional differences between human\ntext and machine text. In addition, we find that decoding strategies alone can\ndramatically effect the quality of machine text, even when generated from\nexactly the same neural language model. Our findings motivate Nucleus Sampling,\na simple but effective method to draw the best out of neural generation. By\nsampling text from the dynamic nucleus of the probability distribution, which\nallows for diversity while effectively truncating the less reliable tail of the\ndistribution, the resulting text better demonstrates the quality of human text,\nyielding enhanced diversity without sacrificing fluency and coherence.', 'In open-domain dialogue systems, generative approaches have attracted much\nattention for response generation. However, existing methods are heavily\nplagued by generating safe responses and unnatural responses. To alleviate\nthese two problems, we propose a novel framework named Dual Adversarial\nLearning (DAL) for high-quality response generation. DAL is the first work to\ninnovatively utilizes the duality between query generation and response\ngeneration to avoid safe responses and increase the diversity of the generated\nresponses. Additionally, DAL uses adversarial learning to mimic human judges\nand guides the system to generate natural responses. Experimental results\ndemonstrate that DAL effectively improves both diversity and overall quality of\nthe generated responses. DAL outperforms the state-of-the-art methods regarding\nautomatic metrics and human evaluations.', 'While conditional language models have greatly improved in their ability to\noutput high-quality natural language, many NLP applications benefit from being\nable to generate a diverse set of candidate sequences. Diverse decoding\nstrategies aim to, within a given-sized candidate list, cover as much of the\nspace of high-quality outputs as possible, leading to improvements for tasks\nthat re-rank and combine candidate outputs. Standard decoding methods, such as\nbeam search, optimize for generating high likelihood sequences rather than\ndiverse ones, though recent work has focused on increasing diversity in these\nmethods. In this work, we perform an extensive survey of decoding-time\nstrategies for generating diverse outputs from conditional language models. We\nalso show how diversity can be improved without sacrificing quality by\nover-sampling additional candidates, then filtering to the desired number.', 'Recent advances in neural sequence-to-sequence models have led to promising\nresults for several language generation-based tasks, including dialogue\nresponse generation, summarization, and machine translation. However, these\nmodels are known to have several problems, especially in the context of\nchit-chat based dialogue systems: they tend to generate short and dull\nresponses that are often too generic. Furthermore, these models do not ground\nconversational responses on knowledge and facts, resulting in turns that are\nnot accurate, informative and engaging for the users. In this paper, we propose\nand experiment with a series of response generation models that aim to serve in\nthe general scenario where in addition to the dialogue context, relevant\nunstructured external knowledge in the form of text is also assumed to be\navailable for models to harness. Our proposed approach extends\npointer-generator networks (See et al., 2017) by allowing the decoder to\nhierarchically attend and copy from external knowledge in addition to the\ndialogue context. We empirically show the effectiveness of the proposed model\ncompared to several baselines including (Ghazvininejad et al., 2018; Zhang et\nal., 2018) through both automatic evaluation metrics and human evaluation on\nCONVAI2 dataset.', 'Diversity is a long-studied topic in information retrieval that usually\nrefers to the requirement that retrieved results should be non-repetitive and\ncover different aspects. In a conversational setting, an additional dimension\nof diversity matters: an engaging response generation system should be able to\noutput responses that are diverse and interesting. Sequence-to-sequence\n(Seq2Seq) models have been shown to be very effective for response generation.\nHowever, dialogue responses generated by Seq2Seq models tend to have low\ndiversity. In this paper, we review known sources and existing approaches to\nthis low-diversity problem. We also identify a source of low diversity that has\nbeen little studied so far, namely model over-confidence. We sketch several\ndirections for tackling model over-confidence and, hence, the low-diversity\nproblem, including confidence penalties and label smoothing.', 'How can we measure whether a natural language generation system produces both\nhigh quality and diverse outputs? Human evaluation captures quality but not\ndiversity, as it does not catch models that simply plagiarize from the training\nset. On the other hand, statistical evaluation (i.e., perplexity) captures\ndiversity but not quality, as models that occasionally emit low quality samples\nwould be insufficiently penalized. In this paper, we propose a unified\nframework which evaluates both diversity and quality, based on the optimal\nerror rate of predicting whether a sentence is human- or machine-generated. We\ndemonstrate that this error rate can be efficiently estimated by combining\nhuman and statistical evaluation, using an evaluation metric which we call\nHUSE. On summarization and chit-chat dialogue, we show that (i) HUSE detects\ndiversity defects which fool pure human evaluation and that (ii) techniques\nsuch as annealing for improving quality actually decrease HUSE due to decreased\ndiversity.', 'Many sequence-to-sequence dialogue models tend to generate safe,\nuninformative responses. There have been various useful efforts on trying to\neliminate them. However, these approaches either improve decoding algorithms\nduring inference, rely on hand-crafted features, or employ complex models. In\nour work, we build dialogue models that are dynamically aware of what\nutterances or tokens are dull without any feature-engineering. Specifically, we\nstart with a simple yet effective automatic metric, AvgOut, which calculates\nthe average output probability distribution of all time steps on the decoder\nside during training. This metric directly estimates which tokens are more\nlikely to be generated, thus making it a faithful evaluation of the model\ndiversity (i.e., for diverse models, the token probabilities should be more\nevenly distributed rather than peaked at a few dull tokens). We then leverage\nthis novel metric to propose three models that promote diversity without losing\nrelevance. The first model, MinAvgOut, directly maximizes the diversity score\nthrough the output distributions of each batch; the second model, Label\nFine-Tuning (LFT), prepends to the source sequence a label continuously scaled\nby the diversity score to control the diversity level; the third model, RL,\nadopts Reinforcement Learning and treats the diversity score as a reward\nsignal. Moreover, we experiment with a hybrid model by combining the loss terms\nof MinAvgOut and RL. All four models outperform their base LSTM-RNN model on\nboth diversity and relevance by a large margin, and are comparable to or better\nthan competitive baselines (also verified via human evaluation). Moreover, our\napproaches are orthogonal to the base model, making them applicable as an\nadd-on to other emerging better dialogue models in the future.', 'The wave of pre-training language models has been continuously improving the\nquality of the machine-generated conversations, however, some of the generated\nresponses still suffer from excessive repetition, sometimes repeating words\nfrom utterance, sometimes repeating words within self-generated responses, or\nboth. Inappropriate repetition of words can significantly degrade the quality\nof the generated texts. Penalized sampling is one popular solution, reducing\nthe sampling probability of existing words during inference, however, it is\nhighly vulnerable to the inappropriate setting of the static weight. Setting it\ntoo high can yield strange and unrealistic sentences while setting it too low\nmakes the task of suppressing repetition trivial. To remedy the shortcomings of\nthe above methods, we design a context-aware classifier to explicitly decide\nwhen to allow repetition and when to employ penalized sampling. Such a\nclassifier can be easily integrated with existing decoding methods, reducing\nrepetitions where appropriate while preserving the diversity of the text.\nExperimental results demonstrate that our method can generate higher quality\nand more authentic dialogues.', 'Recent progress in pre-trained language models led to systems that are able\nto generate text of an increasingly high quality. While several works have\ninvestigated the fluency and grammatical correctness of such models, it is\nstill unclear to which extent the generated text is consistent with factual\nworld knowledge. Here, we go beyond fluency and also investigate the\nverifiability of text generated by state-of-the-art pre-trained language\nmodels. A generated sentence is verifiable if it can be corroborated or\ndisproved by Wikipedia, and we find that the verifiability of generated text\nstrongly depends on the decoding strategy. In particular, we discover a\ntradeoff between factuality (i.e., the ability of generating Wikipedia\ncorroborated text) and repetitiveness. While decoding strategies such as top-k\nand nucleus sampling lead to less repetitive generations, they also produce\nless verifiable text. Based on these finding, we introduce a simple and\neffective decoding strategy which, in comparison to previously used decoding\nstrategies, produces less repetitive and more verifiable text.', 'A good open-domain chatbot should avoid presenting contradictory responses\nabout facts or opinions in a conversational session, known as its consistency\ncapacity. However, evaluating the consistency capacity of a chatbot is still\nchallenging. Employing human judges to interact with chatbots on purpose to\ncheck their capacities is costly and low-efficient, and difficult to get rid of\nsubjective bias. In this paper, we propose the Addressing Inquiries about\nHistory (AIH), an efficient and practical framework for the consistency\nevaluation. At the conversation stage, AIH attempts to address appropriate\ninquiries about the dialogue history to induce the chatbot to redeclare the\nhistorical facts or opinions. We carry out the conversation between chatbots,\nwhich is more efficient than the human-bot interaction and can also alleviate\nthe subjective bias. In this way, we manage to rapidly obtain a dialog session\nthat contains responses with high contradiction possibilities. At the\ncontradiction recognition stage, we can either employ human judges or a natural\nlanguage inference (NLI) model to recognize whether the answers to the\ninquiries are contradictory with history. Finally, we are able to rank chatbots\naccording to the contradiction statistics. Experiments on open-domain chatbots\nshow that our approach can efficiently and reliably assess the consistency\ncapacity of chatbots and achieve a high ranking correlation with the human\nevaluation. We release the framework and hope to help improve the consistency\ncapacity of chatbots. \\footnote{\\url{https://github.com/ictnlp/AIH}}', 'In this paper, we propose a simple, fast decoding algorithm that fosters\ndiversity in neural generation. The algorithm modifies the standard beam search\nalgorithm by adding an inter-sibling ranking penalty, favoring choosing\nhypotheses from diverse parents. We evaluate the proposed model on the tasks of\ndialogue response generation, abstractive summarization and machine\ntranslation. We find that diverse decoding helps across all tasks, especially\nthose for which reranking is needed.\n  We further propose a variation that is capable of automatically adjusting its\ndiversity decoding rates for different inputs using reinforcement learning\n(RL). We observe a further performance boost from this RL technique. This paper\nincludes material from the unpublished script ""Mutual Information and Diverse\nDecoding Improve Neural Machine Translation"" (Li and Jurafsky, 2016).', 'The neural text generation suffers from the text degeneration issue such as\nrepetition. Traditional stochastic sampling methods only focus on truncating\nthe unreliable ""tail"" of the distribution, and do not address the ""head"" part,\nwhich we show might contain tedious or even repetitive candidates with high\nprobability that lead to repetition loops. They also do not consider the issue\nthat human text does not always favor high-probability words. Inspired by\nthese, in this work we propose a heuristic sampling method. We propose to use\ninterquartile range of the predicted distribution to determine the ""head"" part,\nthen permutate and rescale the ""head"" with inverse probability. This aims at\ndecreasing the probability for the tedious and possibly repetitive candidates\nwith higher probability, and increasing the probability for the rational but\nmore surprising candidates with lower probability. The proposed algorithm\nprovides a reasonable permutation on the predicted distribution which enhances\ndiversity without compromising rationality of the distribution. We use\npre-trained language model to compare our algorithm with traditional methods.\nResults show that our algorithm can effectively increase the diversity of\ngenerated samples while achieving close resemblance to human text.', 'Many neural network models nowadays have achieved promising performances in\nChit-chat settings. The majority of them rely on an encoder for understanding\nthe post and a decoder for generating the response. Without given assigned\nsemantics, the models lack the fine-grained control over responses as the\nsemantic mapping between posts and responses is hidden on the fly within the\nend-to-end manners. Some previous works utilize sampled latent words as a\ncontrollable semantic form to drive the generated response around the work, but\nfew works attempt to use more complex semantic patterns to guide the\ngeneration. In this paper, we propose to use more detailed semantic forms,\nincluding latent responses and part-of-speech sequences sampled from the\ncorresponding distributions, as the controllable semantics to guide the\ngeneration. Our results show that the richer semantics are not only able to\nprovide informative and diverse responses, but also increase the overall\nperformance of response quality, including fluency and coherence.', 'Neural conversation models such as encoder-decoder models are easy to\ngenerate bland and generic responses. Some researchers propose to use the\nconditional variational autoencoder(CVAE) which maximizes the lower bound on\nthe conditional log-likelihood on a continuous latent variable. With different\nsampled la-tent variables, the model is expected to generate diverse responses.\nAlthough the CVAE-based models have shown tremendous potential, their\nimprovement of generating high-quality responses is still unsatisfactory. In\nthis paper, we introduce a discrete latent variable with an explicit semantic\nmeaning to improve the CVAE on short-text conversation. A major advantage of\nour model is that we can exploit the semantic distance between the latent\nvariables to maintain good diversity between the sampled latent variables.\nAccordingly, we pro-pose a two-stage sampling approach to enable efficient\ndiverse variable selection from a large latent space assumed in the short-text\nconversation task. Experimental results indicate that our model outperforms\nvarious kinds of generation models under both automatic and human evaluations\nand generates more diverse and in-formative responses.']","['Despite the effectiveness of sequence-to-sequence framework on the task of\nShort-Text Conversation (STC), the issue of under-exploitation of training data\n(i.e., the supervision signals from query text is \\textit{ignored}) still\nremains unresolved. Also, the adopted \\textit{maximization}-based decoding\nstrategies, inclined to generating the generic responses or responses with\nrepetition, are unsuited to the STC task. In this paper, we propose to\nformulate the STC task as a language modeling problem and tailor-make a\ntraining strategy to adapt a language model for response generation. To enhance\ngeneration performance, we design a relevance-promoting transformer language\nmodel, which performs additional supervised source attention after the\nself-attention to increase the importance of informative query tokens in\ncalculating the token-level representation. The model further refines the query\nrepresentation with relevance clues inferred from its multiple references\nduring training. In testing, we adopt a\n\\textit{randomization-over-maximization} strategy to reduce the generation of\ngeneric responses. Experimental results on a large Chinese STC dataset\ndemonstrate the superiority of the proposed model on relevance metrics and\ndiversity metrics.\\footnote{Code available at\nhttps://ai.tencent.com/ailab/nlp/dialogue/.', 'Neural text decoding is important for generating high-quality texts using\nlanguage models. To generate high-quality text, popular decoding algorithms\nlike top-k, top-p (nucleus), and temperature-based sampling truncate or distort\nthe unreliable low probability tail of the language model. Though these methods\ngenerate high-quality text after parameter tuning, they are ad hoc. Not much is\nknown about the control they provide over the statistics of the output, which\nis important since recent reports show text quality is highest for a specific\nrange of likelihoods. Here, first we provide a theoretical analysis of\nperplexity in top-k, top-p, and temperature sampling, finding that\ncross-entropy behaves approximately linearly as a function of p in top-p\nsampling whereas it is a nonlinear function of k in top-k sampling, under\nZipfian statistics. We use this analysis to design a feedback-based adaptive\ntop-k text decoding algorithm called mirostat that generates text (of any\nlength) with a predetermined value of perplexity, and thereby high-quality text\nwithout any tuning. Experiments show that for low values of k and p in top-k\nand top-p sampling, perplexity drops significantly with generated text length,\nwhich is also correlated with excessive repetitions in the text (the boredom\ntrap). On the other hand, for large values of k and p, we find that perplexity\nincreases with generated text length, which is correlated with incoherence in\nthe text (confusion trap). Mirostat avoids both traps: experiments show that\ncross-entropy has a near-linear relation with repetition in generated text.\nThis relation is almost independent of the sampling method but slightly\ndependent on the model used. Hence, for a given language model, control over\nperplexity also gives control over repetitions. Experiments with human raters\nfor fluency, coherence, and quality further verify our findings.', ""Having engaging and informative conversations with users is the utmost goal\nfor open-domain conversational systems. Recent advances in transformer-based\nlanguage models and their applications to dialogue systems have succeeded to\ngenerate fluent and human-like responses. However, they still lack control over\nthe generation process towards producing contentful responses and achieving\nengaging conversations. To achieve this goal, we present \\textbf{DiSCoL}\n(\\textbf{Di}alogue \\textbf{S}ystems through \\textbf{Co}versational\n\\textbf{L}ine guided response generation). DiSCoL is an open-domain dialogue\nsystem that leverages conversational lines (briefly \\textbf{convlines}) as\ncontrollable and informative content-planning elements to guide the generation\nmodel produce engaging and informative responses. Two primary modules in\nDiSCoL's pipeline are conditional generators trained for 1) predicting relevant\nand informative convlines for dialogue contexts and 2) generating high-quality\nresponses conditioned on the predicted convlines. Users can also change the\nreturned convlines to \\textit{control} the direction of the conversations\ntowards topics that are more interesting for them. Through automatic and human\nevaluations, we demonstrate the efficiency of the convlines in producing\nengaging conversations."", 'Endowing a chatbot with personality or an identity is quite challenging but\ncritical to deliver more realistic and natural conversations. In this paper, we\naddress the issue of generating responses that are coherent to a pre-specified\nagent profile. We design a model consisting of three modules: a profile\ndetector to decide whether a post should be responded using the profile and\nwhich key should be addressed, a bidirectional decoder to generate responses\nforward and backward starting from a selected profile value, and a position\ndetector that predicts a word position from which decoding should start given a\nselected profile value. We show that general conversation data from social\nmedia can be used to generate profile-coherent responses. Manual and automatic\nevaluation shows that our model can deliver more coherent, natural, and\ndiversified responses.', 'Quite surprisingly, exact maximum a posteriori (MAP) decoding of neural\nlanguage generators frequently leads to low-quality results. Rather, most\nstate-of-the-art results on language generation tasks are attained using beam\nsearch despite its overwhelmingly high search error rate. This implies that the\nMAP objective alone does not express the properties we desire in text, which\nmerits the question: if beam search is the answer, what was the question? We\nframe beam search as the exact solution to a different decoding objective in\norder to gain insights into why high probability under a model alone may not\nindicate adequacy. We find that beam search enforces uniform information\ndensity in text, a property motivated by cognitive science. We suggest a set of\ndecoding objectives that explicitly enforce this property and find that exact\ndecoding with these objectives alleviates the problems encountered when\ndecoding poorly calibrated language generation models. Additionally, we analyze\nthe text produced using various decoding strategies and see that, in our neural\nmachine translation experiments, the extent to which this property is adhered\nto strongly correlates with BLEU.', 'Chatbot is increasingly thriving in different domains, however, because of\nunexpected discourse complexity and training data sparseness, its potential\ndistrust hatches vital apprehension. Recently, Machine-Human Chatting Handoff\n(MHCH), predicting chatbot failure and enabling human-algorithm collaboration\nto enhance chatbot quality, has attracted increasing attention from industry\nand academia. In this study, we propose a novel model, Role-Selected Sharing\nNetwork (RSSN), which integrates both dialogue satisfaction estimation and\nhandoff prediction in one multi-task learning framework. Unlike prior efforts\nin dialog mining, by utilizing local user satisfaction as a bridge, global\nsatisfaction detector and handoff predictor can effectively exchange critical\ninformation. Specifically, we decouple the relation and interaction between the\ntwo tasks by the role information after the shared encoder. Extensive\nexperiments on two public datasets demonstrate the effectiveness of our model.', 'There has been considerable progress made towards conversational models that\ngenerate coherent and fluent responses; however, this often involves training\nlarge language models on large dialogue datasets, such as Reddit. These large\nconversational models provide little control over the generated responses, and\nthis control is further limited in the absence of annotated conversational\ndatasets for attribute specific generation that can be used for fine-tuning the\nmodel. In this paper, we first propose and evaluate plug-and-play methods for\ncontrollable response generation, which does not require dialogue specific\ndatasets and does not rely on fine-tuning a large model. While effective, the\ndecoding procedure induces considerable computational overhead, rendering the\nconversational model unsuitable for interactive usage. To overcome this, we\nintroduce an approach that does not require further computation at decoding\ntime, while also does not require any fine-tuning of a large language model. We\ndemonstrate, through extensive automatic and human evaluation, a high degree of\ncontrol over the generated conversational responses with regard to multiple\ndesired attributes, while being fluent.', 'Responses generated by neural conversational models tend to lack\ninformativeness and diversity. We present Adversarial Information Maximization\n(AIM), an adversarial learning strategy that addresses these two related but\ndistinct problems. To foster response diversity, we leverage adversarial\ntraining that allows distributional matching of synthetic and real responses.\nTo improve informativeness, our framework explicitly optimizes a variational\nlower bound on pairwise mutual information between query and response.\nEmpirical results from automatic and human evaluations demonstrate that our\nmethods significantly boost informativeness and diversity.', 'All AI models are susceptible to learning biases in data that they are\ntrained on. For generative dialogue models, being trained on real human\nconversations containing unbalanced gender and race/ethnicity references can\nlead to models that display learned biases, which we define here broadly as any\nmeasurable differences in the distributions of words or semantic content of\nconversations based on demographic groups. We measure the strength of such\nbiases by producing artificial conversations between two copies of a dialogue\nmodel, conditioning one conversational partner to state a name commonly\nassociated with a certain gender and/or race/ethnicity. We find that larger\ncapacity models tend to exhibit more gender bias and greater stereotyping of\noccupations by gender. We show that several methods of tuning these dialogue\nmodels, specifically name scrambling, controlled generation, and unlikelihood\ntraining, are effective in reducing bias in conversation, including on a\ndownstream conversational task. Name scrambling is also effective in lowering\ndifferences in token usage across conversations where partners have names\nassociated with different genders or races/ethnicities.', 'Neural Machine Translation (NMT) models often lack diversity in their\ngenerated translations, even when paired with search algorithm, like beam\nsearch. A challenge is that the diversity in translations are caused by the\nvariability in the target language, and cannot be inferred from the source\nsentence alone. In this paper, we propose to explicitly model this one-to-many\nmapping by conditioning the decoder of a NMT model on a latent variable that\nrepresents the domain of target sentences. The domain is a discrete variable\ngenerated by a target encoder that is jointly trained with the NMT model. The\npredicted domain of target sentences are given as input to the decoder during\ntraining. At inference, we can generate diverse translations by decoding with\ndifferent domains. Unlike our strongest baseline (Shen et al., 2019), our\nmethod can scale to any number of domains without affecting the performance or\nthe training time. We assess the quality and diversity of translations\ngenerated by our model with several metrics, on three different datasets.', 'Encoder-decoder based neural architectures serve as the basis of\nstate-of-the-art approaches in end-to-end open domain dialog systems. Since\nmost of such systems are trained with a maximum likelihood~(MLE) objective they\nsuffer from issues such as lack of generalizability and the generic response\nproblem, i.e., a system response that can be an answer to a large number of\nuser utterances, e.g., ""Maybe, I don\'t know."" Having explicit feedback on the\nrelevance and interestingness of a system response at each turn can be a useful\nsignal for mitigating such issues and improving system quality by selecting\nresponses from different approaches. Towards this goal, we present a system\nthat evaluates chatbot responses at each dialog turn for coherence and\nengagement. Our system provides explicit turn-level dialog quality feedback,\nwhich we show to be highly correlated with human evaluation. To show that\nincorporating this feedback in the neural response generation models improves\ndialog quality, we present two different and complementary mechanisms to\nincorporate explicit feedback into a neural response generation model:\nreranking and direct modification of the loss function during training. Our\nstudies show that a response generation model that incorporates these combined\nfeedback mechanisms produce more engaging and coherent responses in an\nopen-domain spoken dialog setting, significantly improving the response quality\nusing both automatic and human evaluation.', 'Conventional seq2seq chatbot models attempt only to find sentences with the\nhighest probabilities conditioned on the input sequences, without considering\nthe sentiment of the output sentences. In this paper, we investigate four\nmodels to scale or adjust the sentiment of the chatbot response: a\npersona-based model, reinforcement learning, a plug and play model, and\nCycleGAN, all based on the seq2seq model. We also develop machine-evaluated\nmetrics to estimate whether the responses are reasonable given the input. These\nmetrics, together with human evaluation, are used to analyze the performance of\nthe four models in terms of different aspects; reinforcement learning and\nCycleGAN are shown to be very attractive.', ""Using chatbots to deliver recommendations is increasingly popular. The design\nof recommendation chatbots has primarily been taking an information-centric\napproach by focusing on the recommended content per se. Limited attention is on\nhow social connection and relational strategies, such as self-disclosure from a\nchatbot, may influence users' perception and acceptance of the recommendation.\nIn this work, we designed, implemented, and evaluated a social chatbot capable\nof performing three different levels of self-disclosure: factual information\n(low), cognitive opinions (medium), and emotions (high). In the evaluation, we\nrecruited 372 participants to converse with the chatbot on two topics: movies\nand COVID-19 experiences. In each topic, the chatbot performed small talks and\nmade recommendations relevant to the topic. Participants were randomly assigned\nto four experimental conditions where the chatbot used factual, cognitive,\nemotional, and adaptive strategies to perform self-disclosures. By training a\ntext classifier to identify users' level of self-disclosure in real-time, the\nadaptive chatbot can dynamically match its self-disclosure to the level of\ndisclosure exhibited by the users. Our results show that users reciprocate with\nhigher-level self-disclosure when a recommendation chatbot consistently\ndisplays emotions throughout the conversation. Chatbot's emotional disclosure\nalso led to increased interactional enjoyment and more positive interpersonal\nperception towards the bot, fostering a stronger human-chatbot relationship and\nthus leading to increased recommendation effectiveness, including a higher\ntendency to accept the recommendation. We discuss the understandings obtained\nand implications to future design."", ""Although current state-of-the-art language models have achieved impressive\nresults in numerous natural language processing tasks, still they could not\nsolve the problem of producing repetitive, dull and sometimes inconsistent text\nin open-ended text generation. Studies often attribute this problem to the\nmaximum likelihood training objective, and propose alternative approaches by\nusing stochastic decoding methods or altering the training objective. However,\nthere is still a lack of consistent evaluation metrics to directly compare the\nefficacy of these solutions. In this work, we study different evaluation\nmetrics that have been proposed to evaluate quality, diversity and consistency\nof machine-generated text. From there, we propose a practical pipeline to\nevaluate language models in open-ended generation task, and research on how to\nimprove the model's performance in all dimensions by leveraging different\nauxiliary training objectives."", ""The purpose of unconditional text generation is to train a model with real\nsentences, then generate novel sentences of the same quality and diversity as\nthe training data. However, when different metrics are used for comparing the\nmethods of unconditional text generation, contradictory conclusions are drawn.\nThe difficulty is that both the diversity and quality of the sample should be\nconsidered simultaneously when the models are evaluated. To solve this problem,\na novel metric of distributional discrepancy (DD) is designed to evaluate\ngenerators based on the discrepancy between the generated and real training\nsentences. However, it cannot compute the DD directly because the distribution\nof real sentences is unavailable. Thus, we propose a method for estimating the\nDD by training a neural-network-based text classifier. For comparison, three\nexisting metrics, bi-lingual evaluation understudy (BLEU) versus self-BLEU,\nlanguage model score versus reverse language model score, and Fr\\'{e}chet\nembedding distance, along with the proposed DD, are used to evaluate two\npopular generative models of long short-term memory and generative pretrained\ntransformer 2 on both syntactic and real data. Experimental results show that\nDD is significantly better than the three existing metrics for ranking these\ngenerative models."", 'This paper addresses the question: Why do neural dialog systems generate\nshort and meaningless replies? We conjecture that, in a dialog system, an\nutterance may have multiple equally plausible replies, causing the deficiency\nof neural networks in the dialog application. We propose a systematic way to\nmimic the dialog scenario in a machine translation system, and manage to\nreproduce the phenomenon of generating short and less meaningful sentences in\nthe translation setting, showing evidence of our conjecture.', 'Recently, variational auto-encoder (VAE) based approaches have made\nimpressive progress on improving the diversity of generated responses. However,\nthese methods usually suffer the cost of decreased relevance accompanied by\ndiversity improvements. In this paper, we propose a novel multimodal response\ngeneration framework with exemplar augmentation and curriculum optimization to\nenhance relevance and diversity of generated responses. First, unlike existing\nVAE-based models that usually approximate a simple Gaussian posterior\ndistribution, we present a Gaussian mixture posterior distribution (i.e,\nmultimodal) to further boost response diversity, which helps capture complex\nsemantics of responses. Then, to ensure that relevance does not decrease while\ndiversity increases, we fully exploit similar examples (exemplars) retrieved\nfrom the training data into posterior distribution modeling to augment response\nrelevance. Furthermore, to facilitate the convergence of Gaussian mixture prior\nand posterior distributions, we devise a curriculum optimization strategy to\nprogressively train the model under multiple training criteria from easy to\nhard. Experimental results on widely used SwitchBoard and DailyDialog datasets\ndemonstrate that our model achieves significant improvements compared to strong\nbaselines in terms of diversity and relevance.', 'Sequence to sequence (SEQ2SEQ) models often lack diversity in their generated\ntranslations. This can be attributed to the limitation of SEQ2SEQ models in\ncapturing lexical and syntactic variations in a parallel corpus resulting from\ndifferent styles, genres, topics, or ambiguity of the translation process. In\nthis paper, we develop a novel sequence to sequence mixture (S2SMIX) model that\nimproves both translation diversity and quality by adopting a committee of\nspecialized translation models rather than a single translation model. Each\nmixture component selects its own training dataset via optimization of the\nmarginal loglikelihood, which leads to a soft clustering of the parallel\ncorpus. Experiments on four language pairs demonstrate the superiority of our\nmixture model compared to a SEQ2SEQ baseline with standard or diversity-boosted\nbeam search. Our mixture model uses negligible additional parameters and incurs\nno extra computation cost during decoding.', 'Neural sequence models are widely used to model time-series data. Equally\nubiquitous is the usage of beam search (BS) as an approximate inference\nalgorithm to decode output sequences from these models. BS explores the search\nspace in a greedy left-right fashion retaining only the top-B candidates -\nresulting in sequences that differ only slightly from each other. Producing\nlists of nearly identical sequences is not only computationally wasteful but\nalso typically fails to capture the inherent ambiguity of complex AI tasks. To\novercome this problem, we propose Diverse Beam Search (DBS), an alternative to\nBS that decodes a list of diverse outputs by optimizing for a\ndiversity-augmented objective. We observe that our method finds better top-1\nsolutions by controlling for the exploration and exploitation of the search\nspace - implying that DBS is a better search algorithm. Moreover, these gains\nare achieved with minimal computational or memory over- head as compared to\nbeam search. To demonstrate the broad applicability of our method, we present\nresults on image captioning, machine translation and visual question generation\nusing both standard quantitative metrics and qualitative human studies.\nFurther, we study the role of diversity for image-grounded language generation\ntasks as the complexity of the image changes. We observe that our method\nconsistently outperforms BS and previously proposed techniques for diverse\ndecoding from neural sequence models.', 'Exposure bias has been regarded as a central problem for auto-regressive\nlanguage models (LM). It claims that teacher forcing would cause the test-time\ngeneration to be incrementally distorted due to the training-generation\ndiscrepancy. Although a lot of algorithms have been proposed to avoid teacher\nforcing and therefore alleviate exposure bias, there is little work showing how\nserious the exposure bias problem actually is. In this work, we focus on the\ntask of open-ended language generation, propose metrics to quantify the impact\nof exposure bias in the aspects of quality, diversity, and consistency. Our key\nintuition is that if we feed ground-truth data prefixes (instead of prefixes\ngenerated by the model itself) into the model and ask it to continue the\ngeneration, the performance should become much better because the\ntraining-generation discrepancy in the prefix is removed. Both automatic and\nhuman evaluations are conducted in our experiments. On the contrary to the\npopular belief in exposure bias, we find that the the distortion induced by the\nprefix discrepancy is limited, and does not seem to be incremental during the\ngeneration. Moreover, our analysis reveals an interesting self-recovery ability\nof the LM, which we hypothesize to be countering the harmful effects from\nexposure bias.', 'Search result diversification is a beneficial approach to overcome\nunder-specified queries, such as those that are ambiguous or multi-faceted.\nExisting approaches often rely on massive query logs and interaction data to\ngenerate a variety of possible query intents, which then can be used to re-rank\ndocuments. However, relying on user interaction data is problematic because one\nfirst needs a massive user base to build a sufficient log; public query logs\nare insufficient on their own. Given the recent success of causal language\nmodels (such as the Text-To-Text Transformer (T5) model) at text generation\ntasks, we explore the capacity of these models to generate potential query\nintents. We find that to encourage diversity in the generated queries, it is\nbeneficial to adapt the model by including a new Distributional Causal Language\nModeling (DCLM) objective during fine-tuning and a representation replacement\nduring inference. Across six standard evaluation benchmarks, we find that our\nmethod (which we call IntenT5) improves search result diversity and attains\n(and sometimes exceeds) the diversity obtained when using query suggestions\nbased on a proprietary query log. Our analysis shows that our approach is most\neffective for multi-faceted queries and is able to generalize effectively to\nqueries that were unseen in training data.', ""In recent years, several high-performance conversational systems have been\nproposed based on the Transformer encoder-decoder model. Although previous\nstudies analyzed the effects of the model parameters and the decoding method on\nsubjective dialogue evaluations with overall metrics, they did not analyze how\nthe differences of fine-tuning datasets affect on user's detailed impression.\nIn addition, the Transformer-based approach has only been verified for English,\nnot for such languages with large inter-language distances as Japanese. In this\nstudy, we develop large-scale Transformer-based Japanese dialogue models and\nJapanese chit-chat datasets to examine the effectiveness of the\nTransformer-based approach for building chit-chat dialogue systems. We\nevaluated and analyzed the impressions of human dialogues in different\nfine-tuning datasets, model parameters, and the use of additional information."", 'Open-domain response generation is the task of generating sensible and\ninformative re-sponses to the source sentence. However, neural models tend to\ngenerate safe and mean-ingless responses. While cue-word introducing approaches\nencourage responses with concrete semantics and have shown tremendous\npotential, they still fail to explore di-verse responses during decoding. In\nthis paper, we propose a novel but natural approach that can produce multiple\ncue-words during decoding, and then uses the produced cue-words to drive\ndecoding and shrinks the decoding vocabulary. Thus the neural genera-tion model\ncan explore the full space of responses and discover informative ones with\nefficiency. Experimental results show that our approach significantly\noutperforms several strong baseline models with much lower decoding complexity.\nEspecially, our approach can converge to concrete semantics more efficiently\nduring decoding.', 'Building open-domain chatbots is a challenging area for machine learning\nresearch. While prior work has shown that scaling neural models in the number\nof parameters and the size of the data they are trained on gives improved\nresults, we show that other ingredients are important for a high-performing\nchatbot. Good conversation requires a number of skills that an expert\nconversationalist blends in a seamless way: providing engaging talking points\nand listening to their partners, and displaying knowledge, empathy and\npersonality appropriately, while maintaining a consistent persona. We show that\nlarge scale models can learn these skills when given appropriate training data\nand choice of generation strategy. We build variants of these recipes with 90M,\n2.7B and 9.4B parameter models, and make our models and code publicly\navailable. Human evaluations show our best models are superior to existing\napproaches in multi-turn dialogue in terms of engagingness and humanness\nmeasurements. We then discuss the limitations of this work by analyzing failure\ncases of our models.', 'Conventional seq2seq chatbot models only try to find the sentences with the\nhighest probabilities conditioned on the input sequences, without considering\nthe sentiment of the output sentences. Some research works trying to modify the\nsentiment of the output sequences were reported. In this paper, we propose five\nmodels to scale or adjust the sentiment of the chatbot response: persona-based\nmodel, reinforcement learning, plug and play model, sentiment transformation\nnetwork and cycleGAN, all based on the conventional seq2seq model. We also\ndevelop two evaluation metrics to estimate if the responses are reasonable\ngiven the input. These metrics together with other two popularly used metrics\nwere used to analyze the performance of the five proposed models on different\naspects, and reinforcement learning and cycleGAN were shown to be very\nattractive. The evaluation metrics were also found to be well correlated with\nhuman evaluation.', 'We develop a chatbot using Deep Bidirectional Transformer models (BERT) to\nhandle client questions in financial investment customer service. The bot can\nrecognize 381 intents, and decides when to say ""I don\'t know"" and escalates\nirrelevant/uncertain questions to human operators. Our main novel contribution\nis the discussion about uncertainty measure for BERT, where three different\napproaches are systematically compared on real problems. We investigated two\nuncertainty metrics, information entropy and variance of dropout sampling in\nBERT, followed by mixed-integer programming to optimize decision thresholds.\nAnother novel contribution is the usage of BERT as a language model in\nautomatic spelling correction. Inputs with accidental spelling errors can\nsignificantly decrease intent classification performance. The proposed approach\ncombines probabilities from masked language model and word edit distances to\nfind the best corrections for misspelled words. The chatbot and the entire\nconversational AI system are developed using open-source tools, and deployed\nwithin our company\'s intranet. The proposed approach can be useful for\nindustries seeking similar in-house solutions in their specific business\ndomains. We share all our code and a sample chatbot built on a public dataset\non Github.', ""Non-goal oriented dialog agents (i.e. chatbots) aim to produce varying and\nengaging conversations with a user; however, they typically exhibit either\ninconsistent personality across conversations or the average personality of all\nusers. This paper addresses these issues by controlling an agent's persona upon\ngeneration via conditioning on prior conversations of a target actor. In doing\nso, we are able to utilize more abstract patterns within a person's speech and\nbetter emulate them in generated responses. This work introduces the Generative\nConversation Control model, an augmented and fine-tuned GPT-2 language model\nthat conditions on past reference conversations to probabilistically model\nmulti-turn conversations in the actor's persona. We introduce an accompanying\ndata collection procedure to obtain 10.3M conversations from 6 months worth of\nReddit comments. We demonstrate that scaling model sizes from 117M to 8.3B\nparameters yields an improvement from 23.14 to 13.14 perplexity on 1.7M held\nout Reddit conversations. Increasing model scale yielded similar improvements\nin human evaluations that measure preference of model samples to the held out\ntarget distribution in terms of realism (31% increased to 37% preference),\nstyle matching (37% to 42%), grammar and content quality (29% to 42%), and\nconversation coherency (32% to 40%). We find that conditionally modeling past\nconversations improves perplexity by 0.47 in automatic evaluations. Through\nhuman trials we identify positive trends between conditional modeling and style\nmatching and outline steps to further improve persona control."", 'Technology for language generation has advanced rapidly, spurred by\nadvancements in pre-training large models on massive amounts of data and the\nneed for intelligent agents to communicate in a natural manner. While\ntechniques can effectively generate fluent text, they can also produce\nundesirable societal biases that can have a disproportionately negative impact\non marginalized populations. Language generation presents unique challenges for\nbiases in terms of direct user interaction and the structure of decoding\ntechniques. To better understand these challenges, we present a survey on\nsocietal biases in language generation, focusing on how data and techniques\ncontribute to biases and progress towards reducing biases. Motivated by a lack\nof studies on biases from decoding techniques, we also conduct experiments to\nquantify the effects of these techniques. By further discussing general trends\nand open challenges, we call to attention promising directions for research and\nthe importance of fairness and inclusivity considerations for language\ngeneration applications.', ""Generating high-quality text with sufficient diversity is essential for a\nwide range of Natural Language Generation (NLG) tasks. Maximum-Likelihood (MLE)\nmodels trained with teacher forcing have consistently been reported as weak\nbaselines, where poor performance is attributed to exposure bias (Bengio et\nal., 2015; Ranzato et al., 2015); at inference time, the model is fed its own\nprediction instead of a ground-truth token, which can lead to accumulating\nerrors and poor samples. This line of reasoning has led to an outbreak of\nadversarial based approaches for NLG, on the account that GANs do not suffer\nfrom exposure bias. In this work, we make several surprising observations which\ncontradict common beliefs. First, we revisit the canonical evaluation framework\nfor NLG, and point out fundamental flaws with quality-only evaluation: we show\nthat one can outperform such metrics using a simple, well-known temperature\nparameter to artificially reduce the entropy of the model's conditional\ndistributions. Second, we leverage the control over the quality / diversity\ntrade-off given by this parameter to evaluate models over the whole\nquality-diversity spectrum and find MLE models constantly outperform the\nproposed GAN variants over the whole quality-diversity space. Our results have\nseveral implications: 1) The impact of exposure bias on sample quality is less\nsevere than previously thought, 2) temperature tuning provides a better quality\n/ diversity trade-off than adversarial training while being easier to train,\neasier to cross-validate, and less computationally expensive. Code to reproduce\nthe experiments is available at github.com/pclucas14/GansFallingShort"", 'Machine translation is a popular test bed for research in neural\nsequence-to-sequence models but despite much recent research, there is still a\nlack of understanding of these models. Practitioners report performance\ndegradation with large beams, the under-estimation of rare words and a lack of\ndiversity in the final translations. Our study relates some of these issues to\nthe inherent uncertainty of the task, due to the existence of multiple valid\ntranslations for a single source sentence, and to the extrinsic uncertainty\ncaused by noisy training data. We propose tools and metrics to assess how\nuncertainty in the data is captured by the model distribution and how it\naffects search strategies that generate translations. Our results show that\nsearch works remarkably well but that models tend to spread too much\nprobability mass over the hypothesis space. Next, we propose tools to assess\nmodel calibration and show how to easily fix some shortcomings of current\nmodels. As part of this study, we release multiple human reference translations\nfor two popular benchmarks.', 'Neural text generation is a key tool in natural language applications, but it\nis well known there are major problems at its core. In particular, standard\nlikelihood training and decoding leads to dull and repetitive outputs. While\nsome post-hoc fixes have been proposed, in particular top-$k$ and nucleus\nsampling, they do not address the fact that the token-level probabilities\npredicted by the model are poor. In this paper we show that the likelihood\nobjective itself is at fault, resulting in a model that assigns too much\nprobability to sequences containing repeats and frequent words, unlike those\nfrom the human training distribution. We propose a new objective, unlikelihood\ntraining, which forces unlikely generations to be assigned lower probability by\nthe model. We show that both token and sequence level unlikelihood training\ngive less repetitive, less dull text while maintaining perplexity, giving\nsuperior generations using standard greedy or beam search. According to human\nevaluations, our approach with standard beam search also outperforms the\ncurrently popular decoding methods of nucleus sampling or beam blocking, thus\nproviding a strong alternative to existing techniques.', 'To build an open-domain multi-turn conversation system is one of the most\ninteresting and challenging tasks in Artificial Intelligence. Many research\nefforts have been dedicated to building such dialogue systems, yet few shed\nlight on modeling the conversation flow in an ongoing dialogue. Besides, it is\ncommon for people to talk about highly relevant aspects during a conversation.\nAnd the topics are coherent and drift naturally, which demonstrates the\nnecessity of dialogue flow modeling. To this end, we present the multi-turn\ncue-words driven conversation system with reinforcement learning method (RLCw),\nwhich strives to select an adaptive cue word with the greatest future credit,\nand therefore improve the quality of generated responses. We introduce a new\nreward to measure the quality of cue words in terms of effectiveness and\nrelevance. To further optimize the model for long-term conversations, a\nreinforcement approach is adopted in this paper. Experiments on real-life\ndataset demonstrate that our model consistently outperforms a set of\ncompetitive baselines in terms of simulated turns, diversity and human\nevaluation.', 'Systems powered by artificial intelligence are being developed to be more\nuser-friendly by communicating with users in a progressively human-like\nconversational way. Chatbots, also known as dialogue systems, interactive\nconversational agents, or virtual agents are an example of such systems used in\na wide variety of applications ranging from customer support in the business\ndomain to companionship in the healthcare sector. It is becoming increasingly\nimportant to develop chatbots that can best respond to the personalized needs\nof their users so that they can be as helpful to the user as possible in a real\nhuman way. This paper investigates and compares three popular existing chatbots\nAPI offerings and then propose and develop a voice interactive and multilingual\nchatbot that can effectively respond to users mood, tone, and language using\nIBM Watson Assistant, Tone Analyzer, and Language Translator. The chatbot was\nevaluated using a use case that was targeted at responding to users needs\nregarding exam stress based on university students survey data generated using\nGoogle Forms. The results of measuring the chatbot effectiveness at analyzing\nresponses regarding exam stress indicate that the chatbot responding\nappropriately to the user queries regarding how they are feeling about exams\n76.5%. The chatbot could also be adapted for use in other application areas\nsuch as student info-centers, government kiosks, and mental health support\nsystems.', ""The ability of a dialog system to express consistent language style during\nconversations has a direct, positive impact on its usability and on user\nsatisfaction. Although previous studies have demonstrated that style transfer\nis feasible with a large amount of parallel data, it is often impossible to\ncollect such data for different styles. In this paper, instead of manually\nconstructing conversation data with a certain style, we propose a flexible\nframework that adapts a generic retrieval-based dialogue system to mimic the\nlanguage style of a specified persona without any parallel data. Our approach\nis based on automatic generation of stylized data by learning the usage of\njargon, and then rewriting the generic conversations to a stylized one by\nincorporating the jargon. In experiments we implemented dialogue systems with\nfive distinct language styles, and the result shows our framework significantly\noutperforms baselines in terms of the average score of responses' relevance and\nstyle degree, and content diversity. A/B testing on a commercial chatbot shows\nthat users are more satisfied with our system. This study demonstrates the\nfeasibility of building stylistic dialogue systems by simple data augmentation."", ""Being able to generate informative and coherent dialogue responses is crucial\nwhen designing human-like open-domain dialogue systems. Encoder-decoder-based\ndialogue models tend to produce generic and dull responses during the decoding\nstep because the most predictable response is likely to be a non-informative\nresponse instead of the most suitable one. To alleviate this problem, we\npropose to train the generation model in a bidirectional manner by adding a\nbackward reasoning step to the vanilla encoder-decoder training. The proposed\nbackward reasoning step pushes the model to produce more informative and\ncoherent content because the forward generation step's output is used to infer\nthe dialogue context in the backward direction. The advantage of our method is\nthat the forward generation and backward reasoning steps are trained\nsimultaneously through the use of a latent variable to facilitate bidirectional\noptimization. Our method can improve response quality without introducing side\ninformation (e.g., a pre-trained topic model). The proposed bidirectional\nresponse generation method achieves state-of-the-art performance for response\nquality."", 'We study multi-turn response generation for open-domain dialogues. The\nexisting state-of-the-art addresses the problem with deep neural architectures.\nWhile these models improved response quality, their complexity also hinders the\napplication of the models in real systems. In this work, we pursue a model that\nhas a simple structure yet can effectively leverage conversation contexts for\nresponse generation. To this end, we propose four auxiliary tasks including\nword order recovery, utterance order recovery, masked word recovery, and masked\nutterance recovery, and optimize the objectives of these tasks together with\nmaximizing the likelihood of generation. By this means, the auxiliary tasks\nthat relate to context understanding can guide the learning of the generation\nmodel to achieve a better local optimum. Empirical studies with three\nbenchmarks indicate that our model can significantly outperform\nstate-of-the-art generation models in terms of response quality on both\nautomatic evaluation and human judgment, and at the same time enjoys a much\nfaster decoding process.', ""In recent times, a large number of people have been involved in establishing\ntheir own businesses. Unlike humans, chatbots can serve multiple customers at a\ntime, are available 24/7 and reply in less than a fraction of a second. Though\nchatbots perform well in task-oriented activities, in most cases they fail to\nunderstand personalized opinions, statements or even queries which later impact\nthe organization for poor service management. Lack of understanding\ncapabilities in bots disinterest humans to continue conversations with them.\nUsually, chatbots give absurd responses when they are unable to interpret a\nuser's text accurately. Extracting the client reviews from conversations by\nusing chatbots, organizations can reduce the major gap of understanding between\nthe users and the chatbot and improve their quality of products and\nservices.Thus, in our research we incorporated all the key elements that are\nnecessary for a chatbot to analyse and understand an input text precisely and\naccurately. We performed sentiment analysis, emotion detection, intent\nclassification and named-entity recognition using deep learning to develop\nchatbots with humanistic understanding and intelligence. The efficiency of our\napproach can be demonstrated accordingly by the detailed analysis."", 'Recent research in neural machine translation has largely focused on two\naspects; neural network architectures and end-to-end learning algorithms. The\nproblem of decoding, however, has received relatively little attention from the\nresearch community. In this paper, we solely focus on the problem of decoding\ngiven a trained neural machine translation model. Instead of trying to build a\nnew decoding algorithm for any specific decoding objective, we propose the idea\nof trainable decoding algorithm in which we train a decoding algorithm to find\na translation that maximizes an arbitrary decoding objective. More\nspecifically, we design an actor that observes and manipulates the hidden state\nof the neural machine translation decoder and propose to train it using a\nvariant of deterministic policy gradient. We extensively evaluate the proposed\nalgorithm using four language pairs and two decoding objectives and show that\nwe can indeed train a trainable greedy decoder that generates a better\ntranslation (in terms of a target decoding objective) with minimal\ncomputational overhead.', 'Generating diverse and relevant questions over text is a task with widespread\napplications. We argue that commonly-used evaluation metrics such as BLEU and\nMETEOR are not suitable for this task due to the inherent diversity of\nreference questions, and propose a scheme for extending conventional metrics to\nreflect diversity. We furthermore propose a variational encoder-decoder model\nfor this task. We show through automatic and human evaluation that our\nvariational model improves diversity without loss of quality, and demonstrate\nhow our evaluation scheme reflects this improvement.', 'Chatbot is a technology that is used to mimic human behavior using natural\nlanguage. There are different types of Chatbot that can be used as\nconversational agent in various business domains in order to increase the\ncustomer service and satisfaction. For any business domain, it requires a\nknowledge base to be built for that domain and design an information retrieval\nbased system that can respond the user with a piece of documentation or\ngenerated sentences. The core component of a Chatbot is Natural Language\nUnderstanding (NLU) which has been impressively improved by deep learning\nmethods. But we often lack such properly built NLU modules and requires more\ntime to build it from scratch for high quality conversations. This may\nencourage fresh learners to build a Chatbot from scratch with simple\narchitecture and using small dataset, although it may have reduced\nfunctionality, rather than building high quality data driven methods. This\nresearch focuses on Named Entity Recognition (NER) and Intent Classification\nmodels which can be integrated into NLU service of a Chatbot. Named entities\nwill be inserted manually in the knowledge base and automatically detected in a\ngiven sentence. The NER model in the proposed architecture is based on\nartificial neural network which is trained on manually created entities and\nevaluated using CoNLL-2003 dataset.', 'Traditional neural language models tend to generate generic replies with poor\nlogic and no emotion. In this paper, a syntactically constrained\nbidirectional-asynchronous approach for emotional conversation generation\n(E-SCBA) is proposed to address this issue. In our model, pre-generated emotion\nkeywords and topic keywords are asynchronously introduced into the process of\ndecoding. It is much different from most existing methods which generate\nreplies from the first word to the last. Through experiments, the results\nindicate that our approach not only improves the diversity of replies, but\ngains a boost on both logic and emotion compared with baselines.', 'Existing approaches to neural machine translation (NMT) generate the target\nlanguage sequence token by token from left to right. However, this kind of\nunidirectional decoding framework cannot make full use of the target-side\nfuture contexts which can be produced in a right-to-left decoding direction,\nand thus suffers from the issue of unbalanced outputs. In this paper, we\nintroduce a synchronous bidirectional neural machine translation (SB-NMT) that\npredicts its outputs using left-to-right and right-to-left decoding\nsimultaneously and interactively, in order to leverage both of the history and\nfuture information at the same time. Specifically, we first propose a new\nalgorithm that enables synchronous bidirectional decoding in a single model.\nThen, we present an interactive decoding model in which left-to-right\n(right-to-left) generation does not only depend on its previously generated\noutputs, but also relies on future contexts predicted by right-to-left\n(left-to-right) decoding. We extensively evaluate the proposed SB-NMT model on\nlarge-scale NIST Chinese-English, WMT14 English-German, and WMT18\nRussian-English translation tasks. Experimental results demonstrate that our\nmodel achieves significant improvements over the strong Transformer model by\n3.92, 1.49 and 1.04 BLEU points respectively, and obtains the state-of-the-art\nperformance on Chinese-English and English-German translation tasks.', 'In this paper, we study sequence-to-sequence (S2S) keyphrase generation\nmodels from the perspective of diversity. Recent advances in neural natural\nlanguage generation have made possible remarkable progress on the task of\nkeyphrase generation, demonstrated through improvements on quality metrics such\nas F1-score. However, the importance of diversity in keyphrase generation has\nbeen largely ignored. We first analyze the extent of information redundancy\npresent in the outputs generated by a baseline model trained using maximum\nlikelihood estimation (MLE). Our findings show that repetition of keyphrases is\na major issue with MLE training. To alleviate this issue, we adopt neural\nunlikelihood (UL) objective for training the S2S model. Our version of UL\ntraining operates at (1) the target token level to discourage the generation of\nrepeating tokens; (2) the copy token level to avoid copying repetitive tokens\nfrom the source text. Further, to encourage better model planning during the\ndecoding process, we incorporate K-step ahead token prediction objective that\ncomputes both MLE and UL losses on future tokens as well. Through extensive\nexperiments on datasets from three different domains we demonstrate that the\nproposed approach attains considerably large diversity gains, while maintaining\ncompetitive output quality.', 'Politically sensitive topics are still a challenge for open-domain chatbots.\nHowever, dealing with politically sensitive content in a responsible,\nnon-partisan, and safe behavior way is integral for these chatbots. Currently,\nthe main approach to handling political sensitivity is by simply changing such\na topic when it is detected. This is safe but evasive and results in a chatbot\nthat is less engaging. In this work, as a first step towards a politically safe\nchatbot, we propose a group of metrics for assessing their political prudence.\nWe then conduct political prudence analysis of various chatbots and discuss\ntheir behavior from multiple angles through our automatic metric and human\nevaluation metrics. The testsets and codebase are released to promote research\nin this area.', 'We present Meena, a multi-turn open-domain chatbot trained end-to-end on data\nmined and filtered from public domain social media conversations. This 2.6B\nparameter neural network is simply trained to minimize perplexity of the next\ntoken. We also propose a human evaluation metric called Sensibleness and\nSpecificity Average (SSA), which captures key elements of a human-like\nmulti-turn conversation. Our experiments show strong correlation between\nperplexity and SSA. The fact that the best perplexity end-to-end trained Meena\nscores high on SSA (72% on multi-turn evaluation) suggests that a human-level\nSSA of 86% is potentially within reach if we can better optimize perplexity.\nAdditionally, the full version of Meena (with a filtering mechanism and tuned\ndecoding) scores 79% SSA, 23% higher in absolute SSA than the existing chatbots\nwe evaluated.', 'Narrative generation is an open-ended NLP task in which a model generates a\nstory given a prompt. The task is similar to neural response generation for\nchatbots; however, innovations in response generation are often not applied to\nnarrative generation, despite the similarity between these tasks. We aim to\nbridge this gap by applying and evaluating advances in decoding methods for\nneural response generation to neural narrative generation. In particular, we\nemploy GPT-2 and perform ablations across nucleus sampling thresholds and\ndiverse decoding hyperparameters -- specifically, maximum mutual information --\nanalyzing results over multiple criteria with automatic and human evaluation.\nWe find that (1) nucleus sampling is generally best with thresholds between 0.7\nand 0.9; (2) a maximum mutual information objective can improve the quality of\ngenerated stories; and (3) established automatic metrics do not correlate well\nwith human judgments of narrative quality on any qualitative metric.', ""Diverse and accurate vision+language modeling is an important goal to retain\ncreative freedom and maintain user engagement. However, adequately capturing\nthe intricacies of diversity in language models is challenging. Recent works\ncommonly resort to latent variable models augmented with more or less\nsupervision from object detectors or part-of-speech tags. Common to all those\nmethods is the fact that the latent variable either only initializes the\nsentence generation process or is identical across the steps of generation.\nBoth methods offer no fine-grained control. To address this concern, we propose\nSeq-CVAE which learns a latent space for every word position. We encourage this\ntemporal latent space to capture the 'intention' about how to complete the\nsentence by mimicking a representation which summarizes the future. We\nillustrate the efficacy of the proposed approach to anticipate the sentence\ncontinuation on the challenging MSCOCO dataset, significantly improving\ndiversity metrics compared to baselines while performing on par w.r.t sentence\nquality."", 'Despite growing interest in natural language generation (NLG) models that\nproduce diverse outputs, there is currently no principled method for evaluating\nthe diversity of an NLG system. In this work, we propose a framework for\nevaluating diversity metrics. The framework measures the correlation between a\nproposed diversity metric and a diversity parameter, a single parameter that\ncontrols some aspect of diversity in generated text. For example, a diversity\nparameter might be a binary variable used to instruct crowdsourcing workers to\ngenerate text with either low or high content diversity. We demonstrate the\nutility of our framework by: (a) establishing best practices for eliciting\ndiversity judgments from humans, (b) showing that humans substantially\noutperform automatic metrics in estimating content diversity, and (c)\ndemonstrating that existing methods for controlling diversity by tuning a\n""decoding parameter"" mostly affect form but not meaning. Our framework can\nadvance the understanding of different diversity metrics, an essential step on\nthe road towards better NLG systems.', 'In neural machine translation, a source sequence of words is encoded into a\nvector from which a target sequence is generated in the decoding phase.\nDifferently from statistical machine translation, the associations between\nsource words and their possible target counterparts are not explicitly stored.\nSource and target words are at the two ends of a long information processing\nprocedure, mediated by hidden states at both the source encoding and the target\ndecoding phases. This makes it possible that a source word is incorrectly\ntranslated into a target word that is not any of its admissible equivalent\ncounterparts in the target language.\n  In this paper, we seek to somewhat shorten the distance between source and\ntarget words in that procedure, and thus strengthen their association, by means\nof a method we term bridging source and target word embeddings. We experiment\nwith three strategies: (1) a source-side bridging model, where source word\nembeddings are moved one step closer to the output target sequence; (2) a\ntarget-side bridging model, which explores the more relevant source word\nembeddings for the prediction of the target sequence; and (3) a direct bridging\nmodel, which directly connects source and target word embeddings seeking to\nminimize errors in the translation of ones by the others.\n  Experiments and analysis presented in this paper demonstrate that the\nproposed bridging models are able to significantly improve quality of both\nsentence translation, in general, and alignment and translation of individual\nsource words with target words, in particular.', ""Existing research on response generation for chatbot focuses on \\textbf{First\nResponse Generation} which aims to teach the chatbot to say the first response\n(e.g. a sentence) appropriate to the conversation context (e.g. the user's\nquery). In this paper, we introduce a new task \\textbf{Second Response\nGeneration}, termed as Improv chat, which aims to teach the chatbot to say the\nsecond response after saying the first response with respect the conversation\ncontext, so as to lighten the burden on the user to keep the conversation\ngoing. Specifically, we propose a general learning based framework and develop\na retrieval based system which can generate the second responses with the\nusers' query and the chatbot's first response as input. We present the approach\nto building the conversation corpus for Improv chat from public forums and\nsocial networks, as well as the neural networks based models for response\nmatching and ranking. We include the preliminary experiments and results in\nthis paper. This work could be further advanced with better deep matching\nmodels for retrieval base systems or generative models for generation based\nsystems as well as extensive evaluations in real-life applications."", ""Searching for an available, reliable, official, and understandable\ninformation is not a trivial task due to scattered information across the\ninternet, and the availability lack of governmental communication channels\ncommunicating with African dialects and languages. In this paper, we introduce\nan Artificial Intelligence Powered chatbot for crisis communication that would\nbe omnichannel, multilingual and multi dialectal. We present our work on\nmodified StarSpace embedding tailored for African dialects for the\nquestion-answering task along with the architecture of the proposed chatbot\nsystem and a description of the different layers. English, French, Arabic,\nTunisian, Igbo,Yor\\`ub\\'a, and Hausa are used as languages and dialects.\nQuantitative and qualitative evaluation results are obtained for our real\ndeployed Covid-19 chatbot. Results show that users are satisfied and the\nconversation with the chatbot is meeting customer needs."", 'Despite being virtually ubiquitous, sequence-to-sequence models are\nchallenged by their lack of diversity and inability to be externally\ncontrolled. In this paper, we speculate that a fundamental shortcoming of\nsequence generation models is that the decoding is done strictly from\nleft-to-right, meaning that outputs values generated earlier have a profound\neffect on those generated later. To address this issue, we propose a novel\nmiddle-out decoder architecture that begins from an initial middle-word and\nsimultaneously expands the sequence in both directions. To facilitate\ninformation flow and maintain consistent decoding, we introduce a dual\nself-attention mechanism that allows us to model complex dependencies between\nthe outputs. We illustrate the performance of our model on the task of video\ncaptioning, as well as a synthetic sequence de-noising task. Our middle-out\ndecoder achieves significant improvements on de-noising and competitive\nperformance in the task of video captioning, while quantifiably improving the\ncaption diversity. Furthermore, we perform a qualitative analysis that\ndemonstrates our ability to effectively control the generation process of our\ndecoder.', ""Through advances in their conversational abilities, chatbots have started to\nrequest and process an increasing variety of sensitive personal information.\nThe accurate disclosure of sensitive information is essential where it is used\nto provide advice and support to users in the healthcare and finance sectors.\nIn this study, we explore users' concerns regarding factors associated with the\nuse of sensitive data by chatbot providers. We surveyed a representative sample\nof 491 British citizens. Our results show that the user concerns focus on\ndeleting personal information and concerns about their data's inappropriate\nuse. We also identified that individuals were concerned about losing control\nover their data after a conversation with conversational agents. We found no\neffect from a user's gender or education but did find an effect from the user's\nage, with those over 45 being more concerned than those under 45. We also\nconsidered the factors that engender trust in a chatbot. Our respondents'\nprimary focus was on the chatbot's technical elements, with factors such as the\nresponse quality being identified as the most critical factor. We again found\nno effect from the user's gender or education level; however, when we\nconsidered some social factors (e.g. avatars or perceived 'friendliness'), we\nfound those under 45 years old rated these as more important than those over\n45. The paper concludes with a discussion of these results within the context\nof designing inclusive, digital systems that support a wide range of users."", ""In this paper, we explore the problem of developing personalized chatbots. A\npersonalized chatbot is designed as a digital chatting assistant for a user.\nThe key characteristic of a personalized chatbot is that it should have a\nconsistent personality with the corresponding user. It can talk the same way as\nthe user when it is delegated to respond to others' messages. We present a\nretrieval-based personalized chatbot model, namely IMPChat, to learn an\nimplicit user profile from the user's dialogue history. We argue that the\nimplicit user profile is superior to the explicit user profile regarding\naccessibility and flexibility. IMPChat aims to learn an implicit user profile\nthrough modeling user's personalized language style and personalized\npreferences separately. To learn a user's personalized language style, we\nelaborately build language models from shallow to deep using the user's\nhistorical responses; To model a user's personalized preferences, we explore\nthe conditional relations underneath each post-response pair of the user. The\npersonalized preferences are dynamic and context-aware: we assign higher\nweights to those historical pairs that are topically related to the current\nquery when aggregating the personalized preferences. We match each response\ncandidate with the personalized language style and personalized preference,\nrespectively, and fuse the two matching signals to determine the final ranking\nscore. Comprehensive experiments on two large datasets show that our method\noutperforms all baseline models."", 'Errors in medication intake among elderly people are very common. One of the\nmain causes for this is their loss of ability to retain information. The high\namount of medicine intake required by the advanced age is another limiting\nfactor. Thence, the design of an interactive aid system, preferably using\nnatural language, to help the older population with medication is in demand. A\nchatbot based on a subtractive cluster algorithm, included in unsupervised\nlearned models, is the chosen solution since the processing of natural\nlanguages is a necessary step in view to construct a chatbot able to answer\nquestions that older people may pose upon themselves concerning a particular\ndrug. In this work, the subtractive mountain clustering algorithm has been\nadapted to the problem of natural languages processing. This algorithm version\nallows for the association of a set of words into clusters. After finding the\ncentre of every cluster -- the most relevant word, all the others are\naggregated according to a defined metric adapted to the language processing\nrealm. All the relevant stored information is processed, as well as the\nquestions, by the algorithm. The correct processing of the text enables the\nchatbot to produce answers that relate to the posed queries. To validate the\nmethod, we use the package insert of a drug as the available information and\nformulate associated questions.', 'This work presents an empirical approach to quantifying the loss of lexical\nrichness in Machine Translation (MT) systems compared to Human Translation\n(HT). Our experiments show how current MT systems indeed fail to render the\nlexical diversity of human generated or translated text. The inability of MT\nsystems to generate diverse outputs and its tendency to exacerbate already\nfrequent patterns while ignoring less frequent ones, might be the underlying\ncause for, among others, the currently heavily debated issues related to gender\nbiased output. Can we indeed, aside from biased data, talk about an algorithm\nthat exacerbates seen biases?', ""Most chatbot literature that focuses on improving the fluency and coherence\nof a chatbot, is dedicated to making chatbots more human-like. However, very\nlittle work delves into what really separates humans from chatbots -- humans\nintrinsically understand the effect their responses have on the interlocutor\nand often respond with an intention such as proposing an optimistic view to\nmake the interlocutor feel better. This paper proposes an innovative framework\nto train chatbots to possess human-like intentions. Our framework includes a\nguiding chatbot and an interlocutor model that plays the role of humans. The\nguiding chatbot is assigned an intention and learns to induce the interlocutor\nto reply with responses matching the intention, for example, long responses,\njoyful responses, responses with specific words, etc. We examined our framework\nusing three experimental setups and evaluated the guiding chatbot with four\ndifferent metrics to demonstrate flexibility and performance advantages.\nAdditionally, we performed trials with human interlocutors to substantiate the\nguiding chatbot's effectiveness in influencing the responses of humans to a\ncertain extent. Code will be made available to the public."", 'As the wide adoption of intelligent chatbot in human daily life, user demands\nfor such systems evolve from basic task-solving conversations to more casual\nand friend-like communication. To meet the user needs and build emotional bond\nwith users, it is essential for social chatbots to incorporate more human-like\nand advanced linguistic features. In this paper, we investigate the usage of a\ncommonly used rhetorical device by human -- metaphor for social chatbot. Our\nwork first designs a metaphor generation framework, which generates topic-aware\nand novel figurative sentences. By embedding the framework into a chatbot\nsystem, we then enables the chatbot to communicate with users using figurative\nlanguage. Human annotators validate the novelty and properness of the generated\nmetaphors. More importantly, we evaluate the effects of employing metaphors in\nhuman-chatbot conversations. Experiments indicate that our system effectively\narouses user interests in communicating with our chatbot, resulting in\nsignificantly longer human-chatbot conversations.', ""Chatbots are envisioned to dramatically change the future of Software\nEngineering, allowing practitioners to chat and inquire about their software\nprojects and interact with different services using natural language. At the\nheart of every chatbot is a Natural Language Understanding (NLU) component that\nenables the chatbot to understand natural language input. Recently, many NLU\nplatforms were provided to serve as an off-the-shelf NLU component for\nchatbots, however, selecting the best NLU for Software Engineering chatbots\nremains an open challenge.\n  Therefore, in this paper, we evaluate four of the most commonly used NLUs,\nnamely IBM Watson, Google Dialogflow, Rasa, and Microsoft LUIS to shed light on\nwhich NLU should be used in Software Engineering based chatbots. Specifically,\nwe examine the NLUs' performance in classifying intents, confidence scores\nstability, and extracting entities. To evaluate the NLUs, we use two datasets\nthat reflect two common tasks performed by Software Engineering practitioners,\n1) the task of chatting with the chatbot to ask questions about software\nrepositories 2) the task of asking development questions on Q&A forums (e.g.,\nStack Overflow). According to our findings, IBM Watson is the best performing\nNLU when considering the three aspects (intents classification, confidence\nscores, and entity extraction). However, the results from each individual\naspect show that, in intents classification, IBM Watson performs the best with\nan F1-measure > 84%, but in confidence scores, Rasa comes on top with a median\nconfidence score higher than 0.91. Our results also show that all NLUs, except\nfor Dialogflow, generally provide trustable confidence scores. For entity\nextraction, Microsoft LUIS and IBM Watson outperform other NLUs in the two SE\ntasks. Our results provide guidance to software engineering practitioners when\ndeciding which NLU to use in their chatbots."", 'The goal of text generation models is to fit the underlying real probability\ndistribution of text. For performance evaluation, quality and diversity metrics\nare usually applied. However, it is still not clear to what extend can the\nquality-diversity evaluation reflect the distribution-fitting goal. In this\npaper, we try to reveal such relation in a theoretical approach. We prove that\nunder certain conditions, a linear combination of quality and diversity\nconstitutes a divergence metric between the generated distribution and the real\ndistribution. We also show that the commonly used BLEU/Self-BLEU metric pair\nfails to match any divergence metric, thus propose CR/NRR as a substitute for\nquality/diversity metric pair.', 'In this paper we present the results of our experiments in training and\ndeploying a self-supervised retrieval-based chatbot trained with contrastive\nlearning for assisting customer support agents. In contrast to most existing\nresearch papers in this area where the focus is on solving just one component\nof a deployable chatbot, we present an end-to-end set of solutions to take the\nreader from an unlabelled chatlogs to a deployed chatbot. This set of solutions\nincludes creating a self-supervised dataset and a weakly labelled dataset from\nchatlogs, as well as a systematic approach to selecting a fixed list of canned\nresponses. We present a hierarchical-based RNN architecture for the response\nselection model, chosen for its ability to cache intermediate utterance\nembeddings, which helped to meet deployment inference speed requirements. We\ncompare the performance of this architecture across 3 different learning\nobjectives: self-supervised contrastive learning, binary classification, and\nmulti-class classification. We find that using a self-supervised contrastive\nlearning model outperforms training the binary and multi-class classification\nmodels on a weakly labelled dataset. Our results validate that the\nself-supervised contrastive learning approach can be effectively used for a\nreal-world chatbot scenario.', ""We present ParaBank, a large-scale English paraphrase dataset that surpasses\nprior work in both quantity and quality. Following the approach of ParaNMT, we\ntrain a Czech-English neural machine translation (NMT) system to generate novel\nparaphrases of English reference sentences. By adding lexical constraints to\nthe NMT decoding procedure, however, we are able to produce multiple\nhigh-quality sentential paraphrases per source sentence, yielding an English\nparaphrase resource with more than 4 billion generated tokens and exhibiting\ngreater lexical diversity. Using human judgments, we also demonstrate that\nParaBank's paraphrases improve over ParaNMT on both semantic similarity and\nfluency. Finally, we use ParaBank to train a monolingual NMT model with the\nsame support for lexically-constrained decoding for sentence rewriting tasks."", ""The majority of conversations a dialogue agent sees over its lifetime occur\nafter it has already been trained and deployed, leaving a vast store of\npotential training signal untapped. In this work, we propose the self-feeding\nchatbot, a dialogue agent with the ability to extract new training examples\nfrom the conversations it participates in. As our agent engages in\nconversation, it also estimates user satisfaction in its responses. When the\nconversation appears to be going well, the user's responses become new training\nexamples to imitate. When the agent believes it has made a mistake, it asks for\nfeedback; learning to predict the feedback that will be given improves the\nchatbot's dialogue abilities further. On the PersonaChat chit-chat dataset with\nover 131k training examples, we find that learning from dialogue with a\nself-feeding chatbot significantly improves performance, regardless of the\namount of traditional supervision."", 'In this paper, we focus on the personalized response generation for\nconversational systems. Based on the sequence to sequence learning, especially\nthe encoder-decoder framework, we propose a two-phase approach, namely\ninitialization then adaptation, to model the responding style of human and then\ngenerate personalized responses. For evaluation, we propose a novel human aided\nmethod to evaluate the performance of the personalized response generation\nmodels by online real-time conversation and offline human judgement. Moreover,\nthe lexical divergence of the responses generated by the 5 personalized models\nindicates that the proposed two-phase approach achieves good results on\nmodeling the responding style of human and generating personalized responses\nfor the conversational systems.', 'In this paper, we present a real-world conversational AI system to search for\nand book hotels through text messaging. Our architecture consists of a\nframe-based dialogue management system, which calls machine learning models for\nintent classification, named entity recognition, and information retrieval\nsubtasks. Our chatbot has been deployed on a commercial scale, handling tens of\nthousands of hotel searches every day. We describe the various opportunities\nand challenges of developing a chatbot in the travel industry.', ""Interview chatbots engage users in a text-based conversation to draw out\ntheir views and opinions. It is, however, challenging to build effective\ninterview chatbots that can handle user free-text responses to open-ended\nquestions and deliver engaging user experience. As the first step, we are\ninvestigating the feasibility and effectiveness of using publicly available,\npractical AI technologies to build effective interview chatbots. To demonstrate\nfeasibility, we built a prototype scoped to enable interview chatbots with a\nsubset of active listening skills - the abilities to comprehend a user's input\nand respond properly. To evaluate the effectiveness of our prototype, we\ncompared the performance of interview chatbots with or without active listening\nskills on four common interview topics in a live evaluation with 206 users. Our\nwork presents practical design implications for building effective interview\nchatbots, hybrid chatbot platforms, and empathetic chatbots beyond interview\ntasks."", 'We present TruthBot, an all-in-one multilingual conversational chatbot\ndesigned for seeking truth (trustworthy and verified information) on specific\ntopics. It helps users to obtain information specific to certain topics,\nfact-check information, and get recent news. The chatbot learns the intent of a\nquery by training a deep neural network from the data of the previous intents\nand responds appropriately when it classifies the intent in one of the classes\nabove. Each class is implemented as a separate module that uses either its own\ncurated knowledge-base or searches the web to obtain the correct information.\nThe topic of the chatbot is currently set to COVID-19. However, the bot can be\neasily customized to any topic-specific responses. Our experimental results\nshow that each module performs significantly better than its closest\ncompetitor, which is verified both quantitatively and through several\nuser-based surveys in multiple languages. TruthBot has been deployed in June\n2020 and is currently running.', 'Creating open-domain chatbots requires large amounts of conversational data\nand related benchmark tasks to evaluate them. Standardized evaluation tasks are\ncrucial for creating automatic evaluation metrics for model development;\notherwise, comparing the models would require resource-expensive human\nevaluation. While chatbot challenges have recently managed to provide a\nplethora of such resources for English, resources in other languages are not\nyet available. In this work, we provide a starting point for Finnish\nopen-domain chatbot research. We describe our collection efforts to create the\nFinnish chat conversation corpus FinChat, which is made available publicly.\nFinChat includes unscripted conversations on seven topics from people of\ndifferent ages. Using this corpus, we also construct a retrieval-based\nevaluation task for Finnish chatbot development. We observe that off-the-shelf\nchatbot models trained on conversational corpora do not perform better than\nchance at choosing the right answer based on automatic metrics, while humans\ncan do the same task almost perfectly. Similarly, in a human evaluation,\nresponses to questions from the evaluation set generated by the chatbots are\npredominantly marked as incoherent. Thus, FinChat provides a challenging\nevaluation set, meant to encourage chatbot development in Finnish.', 'Conversational systems typically focus on functional tasks such as scheduling\nappointments or creating todo lists. Instead we design and evaluate SlugBot\n(SB), one of 8 semifinalists in the 2018 AlexaPrize, whose goal is to support\ncasual open-domain social inter-action. This novel application requires both\nbroad topic coverage and engaging interactive skills. We developed a new\ntechnical approach to meet this demanding situation by crowd-sourcing novel\ncontent and introducing playful conversational strategies based on storytelling\nand games. We collected over 10,000 conversations during August 2018 as part of\nthe Alexa Prize competition. We also conducted an in-lab follow-up qualitative\nevaluation. Over-all users found SB moderately engaging; conversations averaged\n3.6 minutes and involved 26 user turns. However, users reacted very differently\nto different conversation subtypes. Storytelling and games were evaluated\npositively; these were seen as entertaining with predictable interactive\nstructure. They also led users to impute personality and intelligence to SB. In\ncontrast, search and general Chit-Chat induced coverage problems; here users\nfound it hard to infer what topics SB could understand, with these\nconversations seen as being too system-driven. Theoretical and design\nimplications suggest a move away from conversational systems that simply\nprovide factual information. Future systems should be designed to have their\nown opinions with personal stories to share, and SB provides an example of how\nwe might achieve this.', 'Data scarcity is a long-standing and crucial challenge that hinders quick\ndevelopment of task-oriented dialogue systems across multiple domains:\ntask-oriented dialogue models are expected to learn grammar, syntax, dialogue\nreasoning, decision making, and language generation from absurdly small amounts\nof task-specific data. In this paper, we demonstrate that recent progress in\nlanguage modeling pre-training and transfer learning shows promise to overcome\nthis problem. We propose a task-oriented dialogue model that operates solely on\ntext input: it effectively bypasses explicit policy and language generation\nmodules. Building on top of the TransferTransfo framework (Wolf et al., 2019)\nand generative model pre-training (Radford et al., 2019), we validate the\napproach on complex multi-domain task-oriented dialogues from the MultiWOZ\ndataset. Our automatic and human evaluations show that the proposed model is on\npar with a strong task-specific neural baseline. In the long run, our approach\nholds promise to mitigate the data scarcity problem, and to support the\nconstruction of more engaging and more eloquent task-oriented conversational\nagents.', 'An on-device DNN-HMM speech recognition system efficiently works with a\nlimited vocabulary in the presence of a variety of predictable noise. In such a\ncase, vocabulary and environment adaptation is highly effective. In this paper,\nwe propose a novel method of end-to-end (E2E) adaptation, which adjusts not\nonly an acoustic model (AM) but also a weighted finite-state transducer (WFST).\nWe convert a pretrained WFST to a trainable neural network and adapt the system\nto target environments/vocabulary by E2E joint training with an AM. We\nreplicate Viterbi decoding with forward--backward neural network computation,\nwhich is similar to recurrent neural networks (RNNs). By pooling output score\nsequences, a vocabulary posterior for each utterance is obtained and used for\ndiscriminative loss computation. Experiments using 2--10 hours of\nEnglish/Japanese adaptation datasets indicate that the fine-tuning of only\nWFSTs and that of only AMs are both comparable to a state-of-the-art adaptation\nmethod, and E2E joint training of the two components achieves the best\nrecognition performance. We also adapt each language system to the other\nlanguage using the adaptation data, and the results show that the proposed\nmethod also works well for language adaptations.', 'We investigate the effect of different model architectures, training\nobjectives, hyperparameter settings and decoding procedures on the diversity of\nautomatically generated image captions. Our results show that 1) simple\ndecoding by naive sampling, coupled with low temperature is a competitive and\nfast method to produce diverse and accurate caption sets; 2) training with\nCIDEr-based reward using Reinforcement learning harms the diversity properties\nof the resulting generator, which cannot be mitigated by manipulating decoding\nparameters. In addition, we propose a new metric AllSPICE for evaluating both\naccuracy and diversity of a set of captions by a single value.', 'This paper introduces a document grounded dataset for text conversations. We\ndefine ""Document Grounded Conversations"" as conversations that are about the\ncontents of a specified document. In this dataset the specified documents were\nWikipedia articles about popular movies. The dataset contains 4112\nconversations with an average of 21.43 turns per conversation. This positions\nthis dataset to not only provide a relevant chat history while generating\nresponses but also provide a source of information that the models could use.\nWe describe two neural architectures that provide benchmark performance on the\ntask of generating the next response. We also evaluate our models for\nengagement and fluency, and find that the information from the document helps\nin generating more engaging and fluent responses.', 'Customer service chatbots are conversational systems designed to provide\ninformation to customers about products/services offered by different\ncompanies. Particularly, intent recognition is one of the core components in\nthe natural language understating capabilities of a chatbot system. Among the\ndifferent intents that a chatbot is trained to recognize, there is a set of\nthem that is universal to any customer service chatbot. Universal intents may\ninclude salutation, switch the conversation to a human agent, farewells, among\nothers. A system to recognize those universal intents will be very helpful to\noptimize the training process of specific customer service chatbots. We propose\nthe development of a universal intent recognition system, which is trained to\nrecognize a selected group of 11 intents that are common in 28 different\nchatbots. The proposed system is trained considering state-of-the-art\nword-embedding models such as word2vec and BERT, and deep classifiers based on\nconvolutional and recurrent neural networks. The proposed model is able to\ndiscriminate between those universal intents with a balanced accuracy up to\n80.4\\%. In addition, the proposed system is equally accurate to recognize\nintents expressed both in short and long text requests. At the same time,\nmisclassification errors often occurs between intents with very similar\nsemantic fields such as farewells and positive comments. The proposed system\nwill be very helpful to optimize the training process of a customer service\nchatbot because some of the intents will be already available and detected by\nour system. At the same time, the proposed approach will be a suitable base\nmodel to train more specific chatbots by applying transfer learning strategies.', ""Smart systems for Universities powered by Artificial Intelligence have been\nmassively developed to help humans in various tasks. The chatbot concept is not\nsomething new in today society which is developing with recent technology.\nCollege students or candidates of college students often need actual\ninformation like asking for something to customer service, especially during\nthis pandemic, when it is difficult to have an immediate face-to-face meeting.\nChatbots are functionally helping in several things such as curriculum\ninformation, admission for new students, schedule info for any lecture courses,\nstudents grade information, and some adding features for Muslim worships\nschedule, also weather forecast information. This Chatbot is developed by Deep\nLearning models, which was adopted by an artificial intelligence model that\nreplicates human intelligence with some specific training schemes. This kind of\nDeep Learning is based on RNN which has some specific memory savings scheme for\nthe Deep Learning Model, specifically this chatbot using LSTM which already\nintegrates by RASA framework. LSTM is also known as Long Short Term Memory\nwhich efficiently saves some required memory but will remove some memory that\nis not needed. This Chatbot uses the FB platform because of the FB users have\nalready reached up to 60.8% of its entire population in Indonesia. Here's the\nchatbot only focuses on case studies at campus of the Magister Informatics FTI\nUniversity of Islamic Indonesia. This research is a first stage development\nwithin fairly sufficient simulate data."", ""The lack of time-efficient and reliable evaluation methods hamper the\ndevelopment of conversational dialogue systems (chatbots). Evaluations\nrequiring humans to converse with chatbots are time and cost-intensive, put\nhigh cognitive demands on the human judges, and yield low-quality results. In\nthis work, we introduce \\emph{Spot The Bot}, a cost-efficient and robust\nevaluation framework that replaces human-bot conversations with conversations\nbetween bots. Human judges then only annotate for each entity in a conversation\nwhether they think it is human or not (assuming there are humans participants\nin these conversations). These annotations then allow us to rank chatbots\nregarding their ability to mimic the conversational behavior of humans. Since\nwe expect that all bots are eventually recognized as such, we incorporate a\nmetric that measures which chatbot can uphold human-like behavior the longest,\ni.e., \\emph{Survival Analysis}. This metric has the ability to correlate a\nbot's performance to certain of its characteristics (e.g., \\ fluency or\nsensibleness), yielding interpretable results. The comparably low cost of our\nframework allows for frequent evaluations of chatbots during their evaluation\ncycle. We empirically validate our claims by applying \\emph{Spot The Bot} to\nthree domains, evaluating several state-of-the-art chatbots, and drawing\ncomparisons to related work. The framework is released as a ready-to-use tool."", ""In this work we explore a deep learning-based dialogue system that generates\nsarcastic and humorous responses from a conversation design perspective. We\ntrained a seq2seq model on a carefully curated dataset of 3000\nquestion-answering pairs, the core of our mean, grumpy, sarcastic chatbot. We\nshow that end-to-end systems learn patterns very quickly from small datasets\nand thus, are able to transfer simple linguistic structures representing\nabstract concepts to unseen settings. We also deploy our LSTM-based\nencoder-decoder model in the browser, where users can directly interact with\nthe chatbot. Human raters evaluated linguistic quality, creativity and\nhuman-like traits, revealing the system's strengths, limitations and potential\nfor future research."", 'As major progress is made in open-ended text generation, measuring how close\nmachine-generated text is to human language remains a critical open problem. We\nintroduce MAUVE, a comparison measure for open-ended text generation, which\ndirectly compares the learnt distribution from a text generation model to the\ndistribution of human-written text using divergence frontiers. MAUVE scales up\nto modern text generation models by computing information divergences in a\nquantized embedding space. Through an extensive empirical study on three\nopen-ended generation tasks, we find that MAUVE identifies known properties of\ngenerated text, scales naturally with model size, and correlates with human\njudgments, with fewer restrictions than existing distributional evaluation\nmetrics.', ""We investigate neural techniques for end-to-end computational argumentation\nmining (AM). We frame AM both as a token-based dependency parsing and as a\ntoken-based sequence tagging problem, including a multi-task learning setup.\nContrary to models that operate on the argument component level, we find that\nframing AM as dependency parsing leads to subpar performance results. In\ncontrast, less complex (local) tagging models based on BiLSTMs perform robustly\nacross classification scenarios, being able to catch long-range dependencies\ninherent to the AM problem. Moreover, we find that jointly learning 'natural'\nsubtasks, in a multi-task learning setup, improves performance."", 'We propose a new method to detect when users express the intent to leave a\nservice, also known as churn. While previous work focuses solely on social\nmedia, we show that this intent can be detected in chatbot conversations. As\ncompanies increasingly rely on chatbots they need an overview of potentially\nchurny users. To this end, we crowdsource and publish a dataset of churn intent\nexpressions in chatbot interactions in German and English. We show that\nclassifiers trained on social media data can detect the same intent in the\ncontext of chatbots.\n  We introduce a classification architecture that outperforms existing work on\nchurn intent detection in social media. Moreover, we show that, using bilingual\nword embeddings, a system trained on combined English and German data\noutperforms monolingual approaches. As the only existing dataset is in English,\nwe crowdsource and publish a novel dataset of German tweets. We thus underline\nthe universal aspect of the problem, as examples of churn intent in English\nhelp us identify churn in German tweets and chatbot conversations.', ""For words, rank-frequency distributions have long been heralded for adherence\nto a potentially-universal phenomenon known as Zipf's law. The hypothetical\nform of this empirical phenomenon was refined by Ben\\^{i}ot Mandelbrot to that\nwhich is presently referred to as the Zipf-Mandelbrot law. Parallel to this,\nHerbet Simon proposed a selection model potentially explaining Zipf's law.\nHowever, a significant dispute between Simon and Mandelbrot, notable empirical\nexceptions, and the lack of a strong empirical connection between Simon's model\nand the Zipf-Mandelbrot law have left the questions of universality and\nmechanistic generation open. We offer a resolution to these issues by\nexhibiting how the dark matter of word segmentation, i.e., space, punctuation,\netc., connect the Zipf-Mandelbrot law to Simon's mechanistic process. This\nexplains Mandelbrot's refinement as no more than a fudge factor, accommodating\nthe effects of the exclusion of the rank-frequency dark matter. Thus,\nintegrating these non-word objects resolves a more-generalized rank-frequency\nlaw. Since this relies upon the integration of space, etc., we find support for\nthe hypothesis that $all$ are generated by common processes, indicating from a\nphysical perspective that space is a word, too."", 'Multilingual speakers tend to alternate between languages within a\nconversation, a phenomenon referred to as ""code-switching"" (CS). CS is a\ncomplex phenomenon that not only encompasses linguistic challenges, but also\ncontains a great deal of complexity in terms of its dynamic behaviour across\nspeakers. This dynamic behaviour has been studied by sociologists and\npsychologists, identifying factors affecting CS. In this paper, we provide an\nempirical user study on Arabic-English CS, where we show the correlation\nbetween users\' CS frequency and character traits. We use machine learning (ML)\nto validate the findings, informing and confirming existing theories. The\npredictive models were able to predict users\' CS frequency with an accuracy\nhigher than 55%, where travel experiences and personality traits played the\nbiggest role in the modeling process.', 'Question: I have five fingers but I am not alive. What am I? Answer: a glove.\nAnswering such a riddle-style question is a challenging cognitive process, in\nthat it requires complex commonsense reasoning abilities, an understanding of\nfigurative language, and counterfactual reasoning skills, which are all\nimportant abilities for advanced natural language understanding (NLU). However,\nthere are currently no dedicated datasets aiming to test these abilities.\nHerein, we present RiddleSense, a new multiple-choice question answering task,\nwhich comes with the first large dataset (5.7k examples) for answering\nriddle-style commonsense questions. We systematically evaluate a wide range of\nmodels over the challenge, and point out that there is a large gap between the\nbest-supervised model and human performance -- suggesting intriguing future\nresearch in the direction of higher-order commonsense reasoning and linguistic\ncreativity towards building advanced NLU systems.', 'In this paper we present an approach for estimating the quality of machine\ntranslation system. There are various methods for estimating the quality of\noutput sentences, but in this paper we focus on Na\\""ive Bayes classifier to\nbuild model using features which are extracted from the input sentences. These\nfeatures are used for finding the likelihood of each of the sentences of the\ntraining data which are then further used for determining the scores of the\ntest data. On the basis of these scores we determine the class labels of the\ntest data.', 'This paper proposes a novel pipeline for automatic grammar augmentation that\nprovides a significant improvement in the voice command recognition accuracy\nfor systems with small footprint acoustic model (AM). The improvement is\nachieved by augmenting the user-defined voice command set, also called grammar\nset, with alternate grammar expressions. For a given grammar set, a set of\npotential grammar expressions (candidate set) for augmentation is constructed\nfrom an AM-specific statistical pronunciation dictionary that captures the\nconsistent patterns and errors in the decoding of AM induced by variations in\npronunciation, pitch, tempo, accent, ambiguous spellings, and noise conditions.\nUsing this candidate set, greedy optimization based and cross-entropy-method\n(CEM) based algorithms are considered to search for an augmented grammar set\nwith improved recognition accuracy utilizing a command-specific dataset. Our\nexperiments show that the proposed pipeline along with algorithms considered in\nthis paper significantly reduce the mis-detection and mis-classification rate\nwithout increasing the false-alarm rate. Experiments also demonstrate the\nconsistent superior performance of CEM method over greedy-based algorithms.', 'Emotion stimulus extraction is a fine-grained subtask of emotion analysis\nthat focuses on identifying the description of the cause behind an emotion\nexpression from a text passage (e.g., in the sentence ""I am happy that I passed\nmy exam"" the phrase ""passed my exam"" corresponds to the stimulus.). Previous\nwork mainly focused on Mandarin and English, with no resources or models for\nGerman. We fill this research gap by developing a corpus of 2006 German news\nheadlines annotated with emotions and 811 instances with annotations of\nstimulus phrases. Given that such corpus creation efforts are time-consuming\nand expensive, we additionally work on an approach for projecting the existing\nEnglish GoodNewsEveryone (GNE) corpus to a machine-translated German version.\nWe compare the performance of a conditional random field (CRF) model (trained\nmonolingually on German and cross-lingually via projection) with a multilingual\nXLM-RoBERTa (XLM-R) model. Our results show that training with the German\ncorpus achieves higher F1 scores than projection. Experiments with XLM-R\noutperform their respective CRF counterparts.', 'Argumentation mining (AM) requires the identification of complex discourse\nstructures and has lately been applied with success monolingually. In this\nwork, we show that the existing resources are, however, not adequate for\nassessing cross-lingual AM, due to their heterogeneity or lack of complexity.\nWe therefore create suitable parallel corpora by (human and machine)\ntranslating a popular AM dataset consisting of persuasive student essays into\nGerman, French, Spanish, and Chinese. We then compare (i) annotation projection\nand (ii) bilingual word embeddings based direct transfer strategies for\ncross-lingual AM, finding that the former performs considerably better and\nalmost eliminates the loss from cross-lingual transfer. Moreover, we find that\nannotation projection works equally well when using either costly human or\ncheap machine translations. Our code and data are available at\n\\url{http://github.com/UKPLab/coling2018-xling_argument_mining}.', 'We investigate whether and where multi-task learning (MTL) can improve\nperformance on NLP problems related to argumentation mining (AM), in particular\nargument component identification. Our results show that MTL performs\nparticularly well (and better than single-task learning) when little training\ndata is available for the main task, a common scenario in AM. Our findings\nchallenge previous assumptions that conceptualizations across AM datasets are\ndivergent and that MTL is difficult for semantic or higher-level tasks.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : In natural language processing tasks, it is essential to accurately represent the meanings of words to achieve optimal performance. However, this task is challenging due to the potential for multiple interpretations of a single word. As a solution, I propose a method that initially learns to identify the smallest unambiguous phrase component within sentences and paragraphs. Following this, I aim to accurately model their latent vectorized representations, taking into account their semantic and structural relationships with neighboring semantic components. My current approach involves capturing the semantic relationship using a pretraining method akin to word2vec. To encapsulate the structural syntactic relationships, I propose modeling each semantic component as a node, with the entire paragraph represented as a sparse graph structure. Subsequently, specialized graph-based neural network algorithms can be applied to obtain latent structural representations. Ultimately, my proposed embedding model should assign a more precise embedding to each text input, potentially enhancing performance in numerous downstream tasks.","['Existing neural semantic parsers mainly utilize a sequence encoder, i.e., a\nsequential LSTM, to extract word order features while neglecting other valuable\nsyntactic information such as dependency graph or constituent trees. In this\npaper, we first propose to use the \\textit{syntactic graph} to represent three\ntypes of syntactic information, i.e., word order, dependency and constituency\nfeatures. We further employ a graph-to-sequence model to encode the syntactic\ngraph and decode a logical form. Experimental results on benchmark datasets\nshow that our model is comparable to the state-of-the-art on Jobs640, ATIS and\nGeo880. Experimental results on adversarial examples demonstrate the robustness\nof the model is also improved by encoding more syntactic information.', 'Long text brings a big challenge to semantic matching due to their\ncomplicated semantic and syntactic structures. To tackle the challenge, we\nconsider using prior knowledge to help identify useful information and filter\nout noise to matching in long text. To this end, we propose a knowledge\nenhanced hybrid neural network (KEHNN). The model fuses prior knowledge into\nword representations by knowledge gates and establishes three matching channels\nwith words, sequential structures of sentences given by Gated Recurrent Units\n(GRU), and knowledge enhanced representations. The three channels are processed\nby a convolutional neural network to generate high level features for matching,\nand the features are synthesized as a matching score by a multilayer\nperceptron. The model extends the existing methods by conducting matching on\nwords, local structures of sentences, and global context of sentences.\nEvaluation results from extensive experiments on public data sets for question\nanswering and conversation show that KEHNN can significantly outperform\nthe-state-of-the-art matching models and particularly improve the performance\non pairs with long text.', 'Most unsupervised NLP models represent each word with a single point or\nsingle region in semantic space, while the existing multi-sense word embeddings\ncannot represent longer word sequences like phrases or sentences. We propose a\nnovel embedding method for a text sequence (a phrase or a sentence) where each\nsequence is represented by a distinct set of multi-mode codebook embeddings to\ncapture different semantic facets of its meaning. The codebook embeddings can\nbe viewed as the cluster centers which summarize the distribution of possibly\nco-occurring words in a pre-trained word embedding space. We introduce an\nend-to-end trainable neural model that directly predicts the set of cluster\ncenters from the input text sequence during test time. Our experiments show\nthat the per-sentence codebook embeddings significantly improve the\nperformances in unsupervised sentence similarity and extractive summarization\nbenchmarks. In phrase similarity experiments, we discover that the multi-facet\nembeddings provide an interpretable semantic representation but do not\noutperform the single-facet baseline.', ""Variational auto-encoders (VAEs) are widely used in natural language\ngeneration due to the regularization of the latent space. However, generating\nsentences from the continuous latent space does not explicitly model the\nsyntactic information. In this paper, we propose to generate sentences from\ndisentangled syntactic and semantic spaces. Our proposed method explicitly\nmodels syntactic information in the VAE's latent space by using the linearized\ntree sequence, leading to better performance of language generation.\nAdditionally, the advantage of sampling in the disentangled syntactic and\nsemantic latent spaces enables us to perform novel applications, such as the\nunsupervised paraphrase generation and syntax-transfer generation. Experimental\nresults show that our proposed model achieves similar or better performance in\nvarious tasks, compared with state-of-the-art related work."", 'Syntactic structure of sentences in a document substantially informs about\nits authorial writing style. Sentence representation learning has been widely\nexplored in recent years and it has been shown that it improves the\ngeneralization of different downstream tasks across many domains. Even though\nutilizing probing methods in several studies suggests that these learned\ncontextual representations implicitly encode some amount of syntax, explicit\nsyntactic information further improves the performance of deep neural models in\nthe domain of authorship attribution. These observations have motivated us to\ninvestigate the explicit representation learning of syntactic structure of\nsentences. In this paper, we propose a self-supervised framework for learning\nstructural representations of sentences. The self-supervised network contains\ntwo components; a lexical sub-network and a syntactic sub-network which take\nthe sequence of words and their corresponding structural labels as the input,\nrespectively. Due to the n-to-1 mapping of words to their structural labels,\neach word will be embedded into a vector representation which mainly carries\nstructural information. We evaluate the learned structural representations of\nsentences using different probing tasks, and subsequently utilize them in the\nauthorship attribution task. Our experimental results indicate that the\nstructural embeddings significantly improve the classification tasks when\nconcatenated with the existing pre-trained word embeddings.', 'Vector averaging remains one of the most popular sentence embedding methods\nin spite of its obvious disregard for syntactic structure. While more complex\nsequential or convolutional networks potentially yield superior classification\nperformance, the improvements in classification accuracy are typically mediocre\ncompared to the simple vector averaging. As an efficient alternative, we\npropose the use of discrete cosine transform (DCT) to compress word sequences\nin an order-preserving manner. The lower order DCT coefficients represent the\noverall feature patterns in sentences, which results in suitable embeddings for\ntasks that could benefit from syntactic features. Our results in semantic\nprobing tasks demonstrate that DCT embeddings indeed preserve more syntactic\ninformation compared with vector averaging. With practically equivalent\ncomplexity, the model yields better overall performance in downstream\nclassification tasks that correlate with syntactic features, which illustrates\nthe capacity of DCT to preserve word order information.', ""Capturing the meaning of sentences has long been a challenging task. Current\nmodels tend to apply linear combinations of word features to conduct semantic\ncomposition for bigger-granularity units e.g. phrases, sentences, and\ndocuments. However, the semantic linearity does not always hold in human\nlanguage. For instance, the meaning of the phrase `ivory tower' can not be\ndeduced by linearly combining the meanings of `ivory' and `tower'. To address\nthis issue, we propose a new framework that models different levels of semantic\nunits (e.g. sememe, word, sentence, and semantic abstraction) on a single\n\\textit{Semantic Hilbert Space}, which naturally admits a non-linear semantic\ncomposition by means of a complex-valued vector word representation. An\nend-to-end neural network~\\footnote{https://github.com/wabyking/qnn} is\nproposed to implement the framework in the text classification task, and\nevaluation results on six benchmarking text classification datasets demonstrate\nthe effectiveness, robustness and self-explanation power of the proposed model.\nFurthermore, intuitive case studies are conducted to help end users to\nunderstand how the framework works."", 'Most conventional sentence similarity methods only focus on similar parts of\ntwo input sentences, and simply ignore the dissimilar parts, which usually give\nus some clues and semantic meanings about the sentences. In this work, we\npropose a model to take into account both the similarities and dissimilarities\nby decomposing and composing lexical semantics over sentences. The model\nrepresents each word as a vector, and calculates a semantic matching vector for\neach word based on all words in the other sentence. Then, each word vector is\ndecomposed into a similar component and a dissimilar component based on the\nsemantic matching vector. After this, a two-channel CNN model is employed to\ncapture features by composing the similar and dissimilar components. Finally, a\nsimilarity score is estimated over the composed feature vectors. Experimental\nresults show that our model gets the state-of-the-art performance on the answer\nsentence selection task, and achieves a comparable result on the paraphrase\nidentification task.', 'Systematically discovering semantic relationships in text is an important and\nextensively studied area in Natural Language Processing, with various tasks\nsuch as entailment, semantic similarity, etc. Decomposability of sentence-level\nscores via subsequence alignments has been proposed as a way to make models\nmore interpretable. We study the problem of aligning components of sentences\nleading to an interpretable model for semantic textual similarity. In this\npaper, we introduce a novel pointer network based model with a sentinel gating\nfunction to align constituent chunks, which are represented using BERT. We\nimprove this base model with a loss function to equally penalize misalignments\nin both sentences, ensuring the alignments are bidirectional. Finally, to guide\nthe network with structured external knowledge, we introduce first-order logic\nconstraints based on ConceptNet and syntactic knowledge. The model achieves an\nF1 score of 97.73 and 96.32 on the benchmark SemEval datasets for the chunk\nalignment task, showing large improvements over the existing solutions. Source\ncode is available at\nhttps://github.com/manishb89/interpretable_sentence_similarity', 'Methods for learning word representations using large text corpora have\nreceived much attention lately due to their impressive performance in numerous\nnatural language processing (NLP) tasks such as, semantic similarity\nmeasurement, and word analogy detection. Despite their success, these\ndata-driven word representation learning methods do not consider the rich\nsemantic relational structure between words in a co-occurring context. On the\nother hand, already much manual effort has gone into the construction of\nsemantic lexicons such as the WordNet that represent the meanings of words by\ndefining the various relationships that exist among the words in a language. We\nconsider the question, can we improve the word representations learnt using a\ncorpora by integrating the knowledge from semantic lexicons?. For this purpose,\nwe propose a joint word representation learning method that simultaneously\npredicts the co-occurrences of two words in a sentence subject to the\nrelational constrains given by the semantic lexicon. We use relations that\nexist between words in the lexicon to regularize the word representations\nlearnt from the corpus. Our proposed method statistically significantly\noutperforms previously proposed methods for incorporating semantic lexicons\ninto word representations on several benchmark datasets for semantic similarity\nand word analogy.', 'The mathematical representation of semantics is a key issue for Natural\nLanguage Processing (NLP). A lot of research has been devoted to finding ways\nof representing the semantics of individual words in vector spaces.\nDistributional approaches --- meaning distributed representations that exploit\nco-occurrence statistics of large corpora --- have proved popular and\nsuccessful across a number of tasks. However, natural language usually comes in\nstructures beyond the word level, with meaning arising not only from the\nindividual words but also the structure they are contained in at the phrasal or\nsentential level. Modelling the compositional process by which the meaning of\nan utterance arises from the meaning of its parts is an equally fundamental\ntask of NLP.\n  This dissertation explores methods for learning distributed semantic\nrepresentations and models for composing these into representations for larger\nlinguistic units. Our underlying hypothesis is that neural models are a\nsuitable vehicle for learning semantically rich representations and that such\nrepresentations in turn are suitable vehicles for solving important tasks in\nnatural language processing. The contribution of this thesis is a thorough\nevaluation of our hypothesis, as part of which we introduce several new\napproaches to representation learning and compositional semantics, as well as\nmultiple state-of-the-art models which apply distributed semantic\nrepresentations to various tasks in NLP.', 'The bag-of-words (BOW) model is the common approach for classifying\ndocuments, where words are used as feature for training a classifier. This\ngenerally involves a huge number of features. Some techniques, such as Latent\nSemantic Analysis (LSA) or Latent Dirichlet Allocation (LDA), have been\ndesigned to summarize documents in a lower dimension with the least semantic\ninformation loss. Some semantic information is nevertheless always lost, since\nonly words are considered. Instead, we aim at using information coming from\nn-grams to overcome this limitation, while remaining in a low-dimension space.\nMany approaches, such as the Skip-gram model, provide good word vector\nrepresentations very quickly. We propose to average these representations to\nobtain representations of n-grams. All n-grams are thus embedded in a same\nsemantic space. A K-means clustering can then group them into semantic\nconcepts. The number of features is therefore dramatically reduced and\ndocuments can be represented as bag of semantic concepts. We show that this\nmodel outperforms LSA and LDA on a sentiment classification task, and yields\nsimilar results than a traditional BOW-model with far less features.', 'This paper proposes a neural semantic parsing approach -- Sequence-to-Action,\nwhich models semantic parsing as an end-to-end semantic graph generation\nprocess. Our method simultaneously leverages the advantages from two recent\npromising directions of semantic parsing. Firstly, our model uses a semantic\ngraph to represent the meaning of a sentence, which has a tight-coupling with\nknowledge bases. Secondly, by leveraging the powerful representation learning\nand prediction ability of neural network models, we propose a RNN model which\ncan effectively map sentences to action sequences for semantic graph\ngeneration. Experiments show that our method achieves state-of-the-art\nperformance on OVERNIGHT dataset and gets competitive performance on GEO and\nATIS datasets.', 'Attributes of words and relations between two words are central to numerous\ntasks in Artificial Intelligence such as knowledge representation, similarity\nmeasurement, and analogy detection. Often when two words share one or more\nattributes in common, they are connected by some semantic relations. On the\nother hand, if there are numerous semantic relations between two words, we can\nexpect some of the attributes of one of the words to be inherited by the other.\nMotivated by this close connection between attributes and relations, given a\nrelational graph in which words are inter- connected via numerous semantic\nrelations, we propose a method to learn a latent representation for the\nindividual words. The proposed method considers not only the co-occurrences of\nwords as done by existing approaches for word representation learning, but also\nthe semantic relations in which two words co-occur. To evaluate the accuracy of\nthe word representations learnt using the proposed method, we use the learnt\nword representations to solve semantic word analogy problems. Our experimental\nresults show that it is possible to learn better word representations by using\nsemantic semantics between words.', 'In this paper we propose a new language model called AGENT, which stands for\nAdversarial Generation and Encoding of Nested Texts. AGENT is designed for\nencoding, generating and refining documents that consist of a long and coherent\ntext, such as an entire book, provided they are hierarchically annotated\n(nested). i.e. divided into sentences, paragraphs and chapters. The core idea\nof our system is learning vector representations for each level of the text\nhierarchy (sentences, paragraphs, etc...), and train each such representation\nto perform 3 tasks: The task of reconstructing the sequence of vectors from a\nlower level that was used to create the representation, and generalized\nversions of the Masked Language Modeling (MLM) and ""Next Sentence Prediction""\ntasks from BERT Devlin et al. [2018]. Additionally we present a new adversarial\nmodel for long text generation and suggest a way to improve the coherence of\nthe generated text by traversing its vector representation tree.', 'Unsupervised text embeddings extraction is crucial for text understanding in\nmachine learning. Word2Vec and its variants have received substantial success\nin mapping words with similar syntactic or semantic meaning to vectors close to\neach other. However, extracting context-aware word-sequence embedding remains a\nchallenging task. Training over large corpus is difficult as labels are\ndifficult to get. More importantly, it is challenging for pre-trained models to\nobtain word-sequence embeddings that are universally good for all downstream\ntasks or for any new datasets. We propose a two-phased ConvDic+DeconvDec\nframework to solve the problem by combining a word-sequence dictionary learning\nmodel with a word-sequence embedding decode model. We propose a convolutional\ntensor decomposition mechanism to learn good word-sequence phrase dictionary in\nthe learning phase. It is proved to be more accurate and much more efficient\nthan the popular alternating minimization method. In the decode phase, we\nintroduce a deconvolution framework that is immune to the problem of varying\nsentence lengths. The word-sequence embeddings we extracted using\nConvDic+DeconvDec are universally good for a few downstream tasks we test on.\nThe framework requires neither pre-training nor prior/outside information.', 'We design a new technique for the distributional semantic modeling with a\nneural network-based approach to learn distributed term representations (or\nterm embeddings) - term vector space models as a result, inspired by the recent\nontology-related approach (using different types of contextual knowledge such\nas syntactic knowledge, terminological knowledge, semantic knowledge, etc.) to\nthe identification of terms (term extraction) and relations between them\n(relation extraction) called semantic pre-processing technology - SPT. Our\nmethod relies on automatic term extraction from the natural language texts and\nsubsequent formation of the problem-oriented or application-oriented (also\ndeeply annotated) text corpora where the fundamental entity is the term\n(includes non-compositional and compositional terms). This gives us an\nopportunity to changeover from distributed word representations (or word\nembeddings) to distributed term representations (or term embeddings). This\ntransition will allow to generate more accurate semantic maps of different\nsubject domains (also, of relations between input terms - it is useful to\nexplore clusters and oppositions, or to test your hypotheses about them). The\nsemantic map can be represented as a graph using Vec2graph - a Python library\nfor visualizing word embeddings (term embeddings in our case) as dynamic and\ninteractive graphs. The Vec2graph library coupled with term embeddings will not\nonly improve accuracy in solving standard NLP tasks, but also update the\nconventional concept of automated ontology development. The main practical\nresult of our work is the development kit (set of toolkits represented as web\nservice APIs and web application), which provides all necessary routines for\nthe basic linguistic pre-processing and the semantic pre-processing of the\nnatural language texts in Ukrainian for future training of term vector space\nmodels.', 'Neural word representations have proven useful in Natural Language Processing\n(NLP) tasks due to their ability to efficiently model complex semantic and\nsyntactic word relationships. However, most techniques model only one\nrepresentation per word, despite the fact that a single word can have multiple\nmeanings or ""senses"". Some techniques model words by using multiple vectors\nthat are clustered based on context. However, recent neural approaches rarely\nfocus on the application to a consuming NLP algorithm. Furthermore, the\ntraining process of recent word-sense models is expensive relative to\nsingle-sense embedding processes. This paper presents a novel approach which\naddresses these concerns by modeling multiple embeddings for each word based on\nsupervised disambiguation, which provides a fast and accurate way for a\nconsuming NLP model to select a sense-disambiguated embedding. We demonstrate\nthat these embeddings can disambiguate both contrastive senses such as nominal\nand verbal senses as well as nuanced senses such as sarcasm. We further\nevaluate Part-of-Speech disambiguated embeddings on neural dependency parsing,\nyielding a greater than 8% average error reduction in unlabeled attachment\nscores across 6 languages.', 'Many machine learning algorithms require the input to be represented as a\nfixed-length feature vector. When it comes to texts, one of the most common\nfixed-length features is bag-of-words. Despite their popularity, bag-of-words\nfeatures have two major weaknesses: they lose the ordering of the words and\nthey also ignore semantics of the words. For example, ""powerful,"" ""strong"" and\n""Paris"" are equally distant. In this paper, we propose Paragraph Vector, an\nunsupervised algorithm that learns fixed-length feature representations from\nvariable-length pieces of texts, such as sentences, paragraphs, and documents.\nOur algorithm represents each document by a dense vector which is trained to\npredict words in the document. Its construction gives our algorithm the\npotential to overcome the weaknesses of bag-of-words models. Empirical results\nshow that Paragraph Vectors outperform bag-of-words models as well as other\ntechniques for text representations. Finally, we achieve new state-of-the-art\nresults on several text classification and sentiment analysis tasks.', 'Fact checking is a challenging task because verifying the truthfulness of a\nclaim requires reasoning about multiple retrievable evidence. In this work, we\npresent a method suitable for reasoning about the semantic-level structure of\nevidence. Unlike most previous works, which typically represent evidence\nsentences with either string concatenation or fusing the features of isolated\nevidence sentences, our approach operates on rich semantic structures of\nevidence obtained by semantic role labeling. We propose two mechanisms to\nexploit the structure of evidence while leveraging the advances of pre-trained\nmodels like BERT, GPT or XLNet. Specifically, using XLNet as the backbone, we\nfirst utilize the graph structure to re-define the relative distances of words,\nwith the intuition that semantically related words should have short distances.\nThen, we adopt graph convolutional network and graph attention network to\npropagate and aggregate information from neighboring nodes on the graph. We\nevaluate our system on FEVER, a benchmark dataset for fact checking, and find\nthat rich structural information is helpful and both our graph-based mechanisms\nimprove the accuracy. Our model is the state-of-the-art system in terms of both\nofficial evaluation metrics, namely claim verification accuracy and FEVER\nscore.', 'Capturing the compositional process which maps the meaning of words to that\nof documents is a central challenge for researchers in Natural Language\nProcessing and Information Retrieval. We introduce a model that is able to\nrepresent the meaning of documents by embedding them in a low dimensional\nvector space, while preserving distinctions of word and sentence order crucial\nfor capturing nuanced semantics. Our model is based on an extended Dynamic\nConvolution Neural Network, which learns convolution filters at both the\nsentence and document level, hierarchically learning to capture and compose low\nlevel lexical features into high level semantic concepts. We demonstrate the\neffectiveness of this model on a range of document modelling tasks, achieving\nstrong results with no feature engineering and with a more compact model.\nInspired by recent advances in visualising deep convolution networks for\ncomputer vision, we present a novel visualisation technique for our document\nnetworks which not only provides insight into their learning process, but also\ncan be interpreted to produce a compelling automatic summarisation system for\ntexts.', 'As the first step in automated natural language processing, representing\nwords and sentences is of central importance and has attracted significant\nresearch attention. Different approaches, from the early one-hot and\nbag-of-words representation to more recent distributional dense and sparse\nrepresentations, were proposed. Despite the successful results that have been\nachieved, such vectors tend to consist of uninterpretable components and face\nnontrivial challenge in both memory and computational requirement in practical\napplications. In this paper, we designed a novel representation model that\nprojects dense word vectors into a higher dimensional space and favors a highly\nsparse and binary representation of word vectors with potentially interpretable\ncomponents, while trying to maintain pairwise inner products between original\nvectors as much as possible. Computationally, our model is relaxed as a\nsymmetric non-negative matrix factorization problem which admits a fast yet\neffective solution. In a series of empirical evaluations, the proposed model\nexhibited consistent improvement and high potential in practical applications.', 'Vector representation of sentences is important for many text processing\ntasks that involve clustering, classifying, or ranking sentences. Recently,\ndistributed representation of sentences learned by neural models from unlabeled\ndata has been shown to outperform the traditional bag-of-words representation.\nHowever, most of these learning methods consider only the content of a sentence\nand disregard the relations among sentences in a discourse by and large.\n  In this paper, we propose a series of novel models for learning latent\nrepresentations of sentences (Sen2Vec) that consider the content of a sentence\nas well as inter-sentence relations. We first represent the inter-sentence\nrelations with a language network and then use the network to induce contextual\ninformation into the content-based Sen2Vec models. Two different approaches are\nintroduced to exploit the information in the network. Our first approach\nretrofits (already trained) Sen2Vec vectors with respect to the network in two\ndifferent ways: (1) using the adjacency relations of a node, and (2) using a\nstochastic sampling method which is more flexible in sampling neighbors of a\nnode. The second approach uses a regularizer to encode the information in the\nnetwork into the existing Sen2Vec model. Experimental results show that our\nproposed models outperform existing methods in three fundamental information\nsystem tasks demonstrating the effectiveness of our approach. The models\nleverage the computational power of multi-core CPUs to achieve fine-grained\ncomputational efficiency. We make our code publicly available upon acceptance.', 'Learning a high-dimensional dense representation for vocabulary terms, also\nknown as a word embedding, has recently attracted much attention in natural\nlanguage processing and information retrieval tasks. The embedding vectors are\ntypically learned based on term proximity in a large corpus. This means that\nthe objective in well-known word embedding algorithms, e.g., word2vec, is to\naccurately predict adjacent word(s) for a given word or context. However, this\nobjective is not necessarily equivalent to the goal of many information\nretrieval (IR) tasks. The primary objective in various IR tasks is to capture\nrelevance instead of term proximity, syntactic, or even semantic similarity.\nThis is the motivation for developing unsupervised relevance-based word\nembedding models that learn word representations based on query-document\nrelevance information. In this paper, we propose two learning models with\ndifferent objective functions; one learns a relevance distribution over the\nvocabulary set for each query, and the other classifies each term as belonging\nto the relevant or non-relevant class for each query. To train our models, we\nused over six million unique queries and the top ranked documents retrieved in\nresponse to each query, which are assumed to be relevant to the query. We\nextrinsically evaluate our learned word representation models using two IR\ntasks: query expansion and query classification. Both query expansion\nexperiments on four TREC collections and query classification experiments on\nthe KDD Cup 2005 dataset suggest that the relevance-based word embedding models\nsignificantly outperform state-of-the-art proximity-based embedding models,\nsuch as word2vec and GloVe.', 'Sentence embeddings encode natural language sentences as low-dimensional\ndense vectors. A great deal of effort has been put into using sentence\nembeddings to improve several important natural language processing tasks.\nRelation extraction is such an NLP task that aims at identifying structured\nrelations defined in a knowledge base from unstructured text. A promising and\nmore efficient approach would be to embed both the text and structured\nknowledge in low-dimensional spaces and discover semantic alignments or\nmappings between them. Although a number of techniques have been proposed in\nthe literature for embedding both sentences and knowledge graphs, little is\nknown about the structural and semantic properties of these embedding spaces in\nterms of relation extraction. In this paper, we investigate the aforementioned\nproperties by evaluating the extent to which sentences carrying similar senses\nare embedded in close proximity sub-spaces, and if we can exploit that\nstructure to align sentences to a knowledge graph. We propose a set of\nexperiments using a widely-used large-scale data set for relation extraction\nand focusing on a set of key sentence embedding methods. We additionally\nprovide the code for reproducing these experiments at\nhttps://github.com/akalino/semantic-structural-sentences. These embedding\nmethods cover a wide variety of techniques ranging from simple word embedding\ncombination to transformer-based BERT-style model. Our experimental results\nshow that different embedding spaces have different degrees of strength for the\nstructural and semantic properties. These results provide useful information\nfor developing embedding-based relation extraction methods.', 'Semantic similarity measures are an important part in Natural Language\nProcessing tasks. However Semantic similarity measures built for general use do\nnot perform well within specific domains. Therefore in this study we introduce\na domain specific semantic similarity measure that was created by the\nsynergistic union of word2vec, a word embedding method that is used for\nsemantic similarity calculation and lexicon based (lexical) semantic similarity\nmethods. We prove that this proposed methodology out performs word embedding\nmethods trained on generic corpus and methods trained on domain specific corpus\nbut do not use lexical semantic similarity methods to augment the results.\nFurther, we prove that text lemmatization can improve the performance of word\nembedding methods.', 'Techniques in which words are represented as vectors have proved useful in\nmany applications in computational linguistics, however there is currently no\ngeneral semantic formalism for representing meaning in terms of vectors. We\npresent a framework for natural language semantics in which words, phrases and\nsentences are all represented as vectors, based on a theoretical analysis which\nassumes that meaning is determined by context.\n  In the theoretical analysis, we define a corpus model as a mathematical\nabstraction of a text corpus. The meaning of a string of words is assumed to be\na vector representing the contexts it occurs in in the corpus model. Based on\nthis assumption, we can show that the vector representations of words can be\nconsidered as elements of an algebra over a field. We note that in applications\nof vector spaces to representing meanings of words there is an underlying\nlattice structure; we interpret the partial ordering of the lattice as\ndescribing entailment between meanings. We also define the context-theoretic\nprobability of a string, and, based on this and the lattice structure, a degree\nof entailment between strings.\n  Together these properties form guidelines as to how to construct semantic\nrepresentations within the framework. A context theory is an implementation of\nthe framework; in an implementation strings are represented as vectors with the\nproperties deduced from the theoretical analysis.\n  We show how to incorporate logical semantics into context theories; this\nenables us to represent statistical information about uncertainty by taking\nweighted sums of individual representations. We also use the framework to\nanalyse approaches to the task of recognising textual entailment, to\nontological representations of meaning and to representing syntactic structure.\nFor the latter, we give new algebraic descriptions of link grammar.', 'Distributional semantic models provide vector representations for words by\ngathering co-occurrence frequencies from corpora of text. Compositional\ndistributional models extend these from words to phrases and sentences. In\ncategorical compositional distributional semantics, phrase and sentence\nrepresentations are functions of their grammatical structure and\nrepresentations of the words therein. In this setting, grammatical structures\nare formalised by morphisms of a compact closed category and meanings of words\nare formalised by objects of the same category. These can be instantiated in\nthe form of vectors or density matrices. This paper concerns the applications\nof this model to phrase and sentence level entailment. We argue that\nentropy-based distances of vectors and density matrices provide a good\ncandidate to measure word-level entailment, show the advantage of density\nmatrices over vectors for word level entailments, and prove that these\ndistances extend compositionally from words to phrases and sentences. We\nexemplify our theoretical constructions on real data and a toy entailment\ndataset and provide preliminary experimental evidence.', 'Distributional models that learn rich semantic word representations are a\nsuccess story of recent NLP research. However, developing models that learn\nuseful representations of phrases and sentences has proved far harder. We\npropose using the definitions found in everyday dictionaries as a means of\nbridging this gap between lexical and phrasal semantics. Neural language\nembedding models can be effectively trained to map dictionary definitions\n(phrases) to (lexical) representations of the words defined by those\ndefinitions. We present two applications of these architectures: ""reverse\ndictionaries"" that return the name of a concept given a definition or\ndescription and general-knowledge crossword question answerers. On both tasks,\nneural language embedding models trained on definitions from a handful of\nfreely-available lexical resources perform as well or better than existing\ncommercial systems that rely on significant task-specific engineering. The\nresults highlight the effectiveness of both neural embedding architectures and\ndefinition-based training for developing models that understand phrases and\nsentences.', 'For natural language processing systems, two kinds of evidence support the\nuse of text representations from neural language models ""pretrained"" on large\nunannotated corpora: performance on application-inspired benchmarks (Peters et\nal., 2018, inter alia), and the emergence of syntactic abstractions in those\nrepresentations (Tenney et al., 2019, inter alia). On the other hand, the lack\nof grounded supervision calls into question how well these representations can\never capture meaning (Bender and Koller, 2020). We apply novel probes to recent\nlanguage models -- specifically focusing on predicate-argument structure as\noperationalized by semantic dependencies (Ivanova et al., 2012) -- and find\nthat, unlike syntax, semantics is not brought to the surface by today\'s\npretrained models. We then use convolutional graph encoders to explicitly\nincorporate semantic parses into task-specific finetuning, yielding benefits to\nnatural language understanding (NLU) tasks in the GLUE benchmark. This approach\ndemonstrates the potential for general-purpose (rather than task-specific)\nlinguistic supervision, above and beyond conventional pretraining and\nfinetuning. Several diagnostics help to localize the benefits of our approach.', 'We propose a generative model for a sentence that uses two latent variables,\nwith one intended to represent the syntax of the sentence and the other to\nrepresent its semantics. We show we can achieve better disentanglement between\nsemantic and syntactic representations by training with multiple losses,\nincluding losses that exploit aligned paraphrastic sentences and word-order\ninformation. We also investigate the effect of moving from bag-of-words to\nrecurrent neural network modules. We evaluate our models as well as several\npopular pretrained embeddings on standard semantic similarity tasks and novel\nsyntactic similarity tasks. Empirically, we find that the model with the best\nperforming syntactic and semantic representations also gives rise to the most\ndisentangled representations.', 'Predicting patient mortality is an important and challenging problem in the\nhealthcare domain, especially for intensive care unit (ICU) patients.\nElectronic health notes serve as a rich source for learning patient\nrepresentations, that can facilitate effective risk assessment. However, a\nlarge portion of clinical notes are unstructured and also contain domain\nspecific terminologies, from which we need to extract structured information.\nIn this paper, we introduce an embedding framework to learn\nsemantically-plausible distributed representations of clinical notes that\nexploits the semantic correspondence between the unstructured texts and their\ncorresponding structured knowledge, known as semantic frame, in a hierarchical\nfashion. Our approach integrates text modeling and semantic correspondence\nlearning into a single model that comprises 1) an unstructured embedding module\nthat makes use of self-similarity matrix representations in order to inject\nstructural regularities of different segments inherent in clinical texts to\npromote local coherence, 2) a structured embedding module to embed the semantic\nframes (e.g., UMLS semantic types) with deep ConvNet and 3) a hierarchical\nsemantic correspondence module that embeds by enhancing the interactions\nbetween text-semantic frame embedding pairs at multiple levels (i.e., words,\nsentence, note). Evaluations on multiple embedding benchmarks on post discharge\nintensive care patient mortality prediction tasks demonstrate its effectiveness\ncompared to approaches that do not exploit the semantic interactions between\nstructured and unstructured information present in clinical notes.', 'In natural languages, words are used in association to construct sentences.\nIt is not words in isolation, but the appropriate combination of hierarchical\nstructures that conveys the meaning of the whole sentence. Neural networks can\ncapture expressive language features; however, insights into the link between\nwords and sentences are difficult to acquire automatically. In this work, we\ndesign a deep neural network architecture that explicitly wires lower and\nhigher linguistic components; we then evaluate its ability to perform the same\ntask at different hierarchical levels. Settling on broad text classification\ntasks, we show that our model, MHAL, learns to simultaneously solve them at\ndifferent levels of granularity by fluidly transferring knowledge between\nhierarchies. Using a multi-head attention mechanism to tie the representations\nbetween single words and full sentences, MHAL systematically outperforms\nequivalent models that are not incentivized towards developing compositional\nrepresentations. Moreover, we demonstrate that, with the proposed architecture,\nthe sentence information flows naturally to individual words, allowing the\nmodel to behave like a sequence labeller (which is a lower, word-level task)\neven without any word supervision, in a zero-shot fashion.', ""Semantic sentence embedding models encode natural language sentences into\nvectors, such that closeness in embedding space indicates closeness in the\nsemantics between the sentences. Bilingual data offers a useful signal for\nlearning such embeddings: properties shared by both sentences in a translation\npair are likely semantic, while divergent properties are likely stylistic or\nlanguage-specific. We propose a deep latent variable model that attempts to\nperform source separation on parallel sentences, isolating what they have in\ncommon in a latent semantic vector, and explaining what is left over with\nlanguage-specific latent vectors. Our proposed approach differs from past work\non semantic sentence encoding in two ways. First, by using a variational\nprobabilistic framework, we introduce priors that encourage source separation,\nand can use our model's posterior to predict sentence embeddings for\nmonolingual data at test time. Second, we use high-capacity transformers as\nboth data generating distributions and inference networks -- contrasting with\nmost past work on sentence embeddings. In experiments, our approach\nsubstantially outperforms the state-of-the-art on a standard suite of\nunsupervised semantic similarity evaluations. Further, we demonstrate that our\napproach yields the largest gains on more difficult subsets of these\nevaluations where simple word overlap is not a good indicator of similarity."", ""An important component of achieving language understanding is mastering the\ncomposition of sentence meaning, but an immediate challenge to solving this\nproblem is the opacity of sentence vector representations produced by current\nneural sentence composition models. We present a method to address this\nchallenge, developing tasks that directly target compositional meaning\ninformation in sentence vector representations with a high degree of precision\nand control. To enable the creation of these controlled tasks, we introduce a\nspecialized sentence generation system that produces large, annotated sentence\nsets meeting specified syntactic, semantic and lexical constraints. We describe\nthe details of the method and generation system, and then present results of\nexperiments applying our method to probe for compositional information in\nembeddings from a number of existing sentence composition models. We find that\nthe method is able to extract useful information about the differing capacities\nof these models, and we discuss the implications of our results with respect to\nthese systems' capturing of sentence information. We make available for public\nuse the datasets used for these experiments, as well as the generation system.""]","['Natural language generation of coherent long texts like paragraphs or longer\ndocuments is a challenging problem for recurrent networks models. In this\npaper, we explore an important step toward this generation task: training an\nLSTM (Long-short term memory) auto-encoder to preserve and reconstruct\nmulti-sentence paragraphs. We introduce an LSTM model that hierarchically\nbuilds an embedding for a paragraph from embeddings for sentences and words,\nthen decodes this embedding to reconstruct the original paragraph. We evaluate\nthe reconstructed paragraph using standard metrics like ROUGE and Entity Grid,\nshowing that neural models are able to encode texts in a way that preserve\nsyntactic, semantic, and discourse coherence. While only a first step toward\ngenerating coherent text units from neural models, our work has the potential\nto significantly impact natural language generation and\nsummarization\\footnote{Code for the three models described in this paper can be\nfound at www.stanford.edu/~jiweil/ .', 'Dense word embeddings, which encode semantic meanings of words to low\ndimensional vector spaces have become very popular in natural language\nprocessing (NLP) research due to their state-of-the-art performances in many\nNLP tasks. Word embeddings are substantially successful in capturing semantic\nrelations among words, so a meaningful semantic structure must be present in\nthe respective vector spaces. However, in many cases, this semantic structure\nis broadly and heterogeneously distributed across the embedding dimensions,\nwhich makes interpretation a big challenge. In this study, we propose a\nstatistical method to uncover the latent semantic structure in the dense word\nembeddings. To perform our analysis we introduce a new dataset (SEMCAT) that\ncontains more than 6500 words semantically grouped under 110 categories. We\nfurther propose a method to quantify the interpretability of the word\nembeddings; the proposed method is a practical alternative to the classical\nword intrusion test that requires human intervention.', 'Word2vec (Mikolov et al., 2013) has proven to be successful in natural\nlanguage processing by capturing the semantic relationships between different\nwords. Built on top of single-word embeddings, paragraph vectors (Le and\nMikolov, 2014) find fixed-length representations for pieces of text with\narbitrary lengths, such as documents, paragraphs, and sentences. In this work,\nwe propose a novel interpretation for neural-network-based paragraph vectors by\ndeveloping an unsupervised generative model whose maximum likelihood solution\ncorresponds to traditional paragraph vectors. This probabilistic formulation\nallows us to go beyond point estimates of parameters and to perform Bayesian\nposterior inference. We find that the entropy of paragraph vectors decreases\nwith the length of documents, and that information about posterior uncertainty\nimproves performance in supervised learning tasks such as sentiment analysis\nand paraphrase detection.', 'Recent progress on image captioning has made it possible to generate novel\nsentences describing images in natural language, but compressing an image into\na single sentence can describe visual content in only coarse detail. While one\nnew captioning approach, dense captioning, can potentially describe images in\nfiner levels of detail by captioning many regions within an image, it in turn\nis unable to produce a coherent story for an image. In this paper we overcome\nthese limitations by generating entire paragraphs for describing images, which\ncan tell detailed, unified stories. We develop a model that decomposes both\nimages and paragraphs into their constituent parts, detecting semantic regions\nin images and using a hierarchical recurrent neural network to reason about\nlanguage. Linguistic analysis confirms the complexity of the paragraph\ngeneration task, and thorough experiments on a new dataset of image and\nparagraph pairs demonstrate the effectiveness of our approach.', 'Sentence embedding is an important research topic in natural language\nprocessing. It is essential to generate a good embedding vector that fully\nreflects the semantic meaning of a sentence in order to achieve an enhanced\nperformance for various natural language processing tasks, such as machine\ntranslation and document classification. Thus far, various sentence embedding\nmodels have been proposed, and their feasibility has been demonstrated through\ngood performances on tasks following embedding, such as sentiment analysis and\nsentence classification. However, because the performances of sentence\nclassification and sentiment analysis can be enhanced by using a simple\nsentence representation method, it is not sufficient to claim that these models\nfully reflect the meanings of sentences based on good performances for such\ntasks. In this paper, inspired by human language recognition, we propose the\nfollowing concept of semantic coherence, which should be satisfied for a good\nsentence embedding method: similar sentences should be located close to each\nother in the embedding space. Then, we propose the Paraphrase-Thought\n(P-thought) model to pursue semantic coherence as much as possible.\nExperimental results on two paraphrase identification datasets (MS COCO and STS\nbenchmark) show that the P-thought models outperform the benchmarked sentence\nembedding methods.', ""Word vector representations are central to deep learning natural language\nprocessing models. Many forms of these vectors, known as embeddings, exist,\nincluding word2vec and GloVe. Embeddings are trained on large corpora and learn\nthe word's usage in context, capturing the semantic relationship between words.\nHowever, the semantics from such training are at the level of distinct words\n(known as word types), and can be ambiguous when, for example, a word type can\nbe either a noun or a verb. In question answering, parts-of-speech and named\nentity types are important, but encoding these attributes in neural models\nexpands the size of the input. This research employs holographic compression of\npre-trained embeddings, to represent a token, its part-of-speech, and named\nentity type, in the same dimension as representing only the token. The\nimplementation, in a modified question answering recurrent deep learning\nnetwork, shows that semantic relationships are preserved, and yields strong\nperformance."", ""Sentence representation at the semantic level is a challenging task for\nNatural Language Processing and Artificial Intelligence. Despite the advances\nin word embeddings (i.e. word vector representations), capturing sentence\nmeaning is an open question due to complexities of semantic interactions among\nwords. In this paper, we present an embedding method, which is aimed at\nlearning unsupervised sentence representations from unlabeled text. We propose\nan unsupervised method that models a sentence as a weighted series of word\nembeddings. The weights of the word embeddings are fitted by using Shannon's\nword entropies provided by the Term Frequency--Inverse Document Frequency\n(TF--IDF) transform. The hyperparameters of the model can be selected according\nto the properties of data (e.g. sentence length and textual gender).\nHyperparameter selection involves word embedding methods and dimensionalities,\nas well as weighting schemata. Our method offers advantages over existing\nmethods: identifiable modules, short-term training, online inference of\n(unseen) sentence representations, as well as independence from domain,\nexternal knowledge and language resources. Results showed that our model\noutperformed the state of the art in well-known Semantic Textual Similarity\n(STS) benchmarks. Moreover, our model reached state-of-the-art performance when\ncompared to supervised and knowledge-based STS systems."", 'Word embeddings are rich word representations, which in combination with deep\nneural networks, lead to large performance gains for many NLP tasks. However,\nword embeddings are represented by dense, real-valued vectors and they are\ntherefore not directly interpretable. Thus, computational operations based on\nthem are also not well understood. In this paper, we present an approach for\nanalyzing structures in the semantic vector space to get a better understanding\nof the underlying semantic encoding principles. We present a framework for\ndecomposing word embeddings into smaller meaningful units which we call\nsub-vectors. The framework opens up a wide range of possibilities analyzing\nphenomena in vector space semantics, as well as solving concrete NLP problems:\nWe introduce the category completion task and show that a sub-vector based\napproach is superior to supervised techniques; We present a sub-vector based\nmethod for solving the word analogy task, which substantially outperforms\ndifferent variants of the traditional vector-offset method.', 'This paper introduces a sentence to vector encoding framework suitable for\nadvanced natural language processing. Our latent representation is shown to\nencode sentences with common semantic information with similar vector\nrepresentations. The vector representation is extracted from an encoder-decoder\nmodel which is trained on sentence paraphrase pairs. We demonstrate the\napplication of the sentence representations for two different tasks -- sentence\nparaphrasing and paragraph summarization, making it attractive for commonly\nused recurrent frameworks that process text. Experimental results help gain\ninsight how vector representations are suitable for advanced language\nembedding.', 'In this paper, we present Hierarchical Graph Network (HGN) for multi-hop\nquestion answering. To aggregate clues from scattered texts across multiple\nparagraphs, a hierarchical graph is created by constructing nodes on different\nlevels of granularity (questions, paragraphs, sentences, entities), the\nrepresentations of which are initialized with pre-trained contextual encoders.\nGiven this hierarchical graph, the initial node representations are updated\nthrough graph propagation, and multi-hop reasoning is performed via traversing\nthrough the graph edges for each subsequent sub-task (e.g., paragraph\nselection, supporting facts extraction, answer prediction). By weaving\nheterogeneous nodes into an integral unified graph, this hierarchical\ndifferentiation of node granularity enables HGN to support different question\nanswering sub-tasks simultaneously. Experiments on the HotpotQA benchmark\ndemonstrate that the proposed model achieves new state of the art,\noutperforming existing multi-hop QA approaches.', 'Recently, there has been a lot of effort to represent words in continuous\nvector spaces. Those representations have been shown to capture both semantic\nand syntactic information about words. However, distributed representations of\nphrases remain a challenge. We introduce a novel model that jointly learns word\nvector representations and their summation. Word representations are learnt\nusing the word co-occurrence statistical information. To embed sequences of\nwords (i.e. phrases) with different sizes into a common semantic space, we\npropose to average word vector representations. In contrast with previous\nmethods which reported a posteriori some compositionality aspects by simple\nsummation, we simultaneously train words to sum, while keeping the maximum\ninformation from the original vectors. We evaluate the quality of the word\nrepresentations on several classical word evaluation tasks, and we introduce a\nnovel task to evaluate the quality of the phrase representations. While our\ndistributed representations compete with other methods of learning word\nrepresentations on word evaluations, we show that they give better performance\non the phrase evaluation. Such representations of phrases could be interesting\nfor many tasks in natural language processing.', 'Grounding referring expressions in images aims to locate the object instance\nin an image described by a referring expression. It involves a joint\nunderstanding of natural language and image content, and is essential for a\nrange of visual tasks related to human-computer interaction. As a\nlanguage-to-vision matching task, the core of this problem is to not only\nextract all the necessary information (i.e., objects and the relationships\namong them) in both the image and referring expression, but also make full use\nof context information to align cross-modal semantic concepts in the extracted\ninformation. Unfortunately, existing work on grounding referring expressions\nfails to accurately extract multi-order relationships from the referring\nexpression and associate them with the objects and their related contexts in\nthe image. In this paper, we propose a Cross-Modal Relationship Extractor\n(CMRE) to adaptively highlight objects and relationships (spatial and semantic\nrelations) related to the given expression with a cross-modal attention\nmechanism, and represent the extracted information as a language-guided visual\nrelation graph. In addition, we propose a Gated Graph Convolutional Network\n(GGCN) to compute multimodal semantic contexts by fusing information from\ndifferent modes and propagating multimodal information in the structured\nrelation graph. Experimental results on three common benchmark datasets show\nthat our Cross-Modal Relationship Inference Network, which consists of CMRE and\nGGCN, significantly surpasses all existing state-of-the-art methods. Code is\navailable at https://github.com/sibeiyang/sgmn/tree/master/lib/cmrin_models', 'Recently proposed Skip-gram model is a powerful method for learning\nhigh-dimensional word representations that capture rich semantic relationships\nbetween words. However, Skip-gram as well as most prior work on learning word\nrepresentations does not take into account word ambiguity and maintain only\nsingle representation per word. Although a number of Skip-gram modifications\nwere proposed to overcome this limitation and learn multi-prototype word\nrepresentations, they either require a known number of word meanings or learn\nthem using greedy heuristic approaches. In this paper we propose the Adaptive\nSkip-gram model which is a nonparametric Bayesian extension of Skip-gram\ncapable to automatically learn the required number of representations for all\nwords at desired semantic resolution. We derive efficient online variational\nlearning algorithm for the model and empirically demonstrate its efficiency on\nword-sense induction task.', 'Over the last few years, machine learning over graph structures has\nmanifested a significant enhancement in text mining applications such as event\ndetection, opinion mining, and news recommendation. One of the primary\nchallenges in this regard is structuring a graph that encodes and encompasses\nthe features of textual data for the effective machine learning algorithm.\nBesides, exploration and exploiting of semantic relations is regarded as a\nprincipal step in text mining applications. However, most of the traditional\ntext mining methods perform somewhat poor in terms of employing such relations.\nIn this paper, we propose a sentence-level graph-based text representation\nwhich includes stop words to consider semantic and term relations. Then, we\nemploy a representation learning approach on the combined graphs of sentences\nto extract the latent and continuous features of the documents. Eventually, the\nlearned features of the documents are fed into a deep neural network for the\nsentiment classification task. The experimental results demonstrate that the\nproposed method substantially outperforms the related sentiment analysis\napproaches based on several benchmark datasets. Furthermore, our method can be\ngeneralized on different datasets without any dependency on pre-trained word\nembeddings.', 'Natural language processing often involves computations with semantic or\nsyntactic graphs to facilitate sophisticated reasoning based on structural\nrelationships. While convolution kernels provide a powerful tool for comparing\ngraph structure based on node (word) level relationships, they are difficult to\ncustomize and can be computationally expensive. We propose a generalization of\nconvolution kernels, with a nonstationary model, for better expressibility of\nnatural languages in supervised settings. For a scalable learning of the\nparameters introduced with our model, we propose a novel algorithm that\nleverages stochastic sampling on k-nearest neighbor graphs, along with\napproximations based on locality-sensitive hashing. We demonstrate the\nadvantages of our approach on a challenging real-world (structured inference)\nproblem of automatically extracting biological models from the text of\nscientific papers.', 'As an efficient model for knowledge organization, the knowledge graph has\nbeen widely adopted in several fields, e.g., biomedicine, sociology, and\neducation. And there is a steady trend of learning embedding representations of\nknowledge graphs to facilitate knowledge graph construction and downstream\ntasks. In general, knowledge graph embedding techniques aim to learn vectorized\nrepresentations which preserve the structural information of the graph. And\nconventional embedding learning models rely on structural relationships among\nentities and relations. However, in educational knowledge graphs, structural\nrelationships are not the focus. Instead, rich literals of the graphs are more\nvaluable. In this paper, we focus on this problem and propose a novel model for\nembedding learning of educational knowledge graphs. Our model considers both\nstructural and literal information and jointly learns embedding\nrepresentations. Three experimental graphs were constructed based on an\neducational knowledge graph which has been applied in real-world teaching. We\nconducted two experiments on the three graphs and other common benchmark\ngraphs. The experimental results proved the effectiveness of our model and its\nsuperiority over other baselines when processing educational knowledge graphs.', 'Natural language processing models have attracted much interest in the deep\nlearning community. This branch of study is composed of some applications such\nas machine translation, sentiment analysis, named entity recognition, question\nand answer, and others. Word embeddings are continuous word representations,\nthey are an essential module for those applications and are generally used as\ninput word representation to the deep learning models. Word2Vec and GloVe are\ntwo popular methods to learn word embeddings. They achieve good word\nrepresentations, however, they learn representations with limited information\nbecause they ignore the morphological information of the words and consider\nonly one representation vector for each word. This approach implies that\nWord2Vec and GloVe are unaware of the word inner structure. To mitigate this\nproblem, the FastText model represents each word as a bag of characters\nn-grams. Hence, each n-gram has a continuous vector representation, and the\nfinal word representation is the sum of its characters n-grams vectors.\nNevertheless, the use of all n-grams character of a word is a poor approach\nsince some n-grams have no semantic relation with their words and increase the\namount of potentially useless information. This approach also increases the\ntraining phase time. In this work, we propose a new method for training word\nembeddings, and its goal is to replace the FastText bag of character n-grams\nfor a bag of word morphemes through the morphological analysis of the word.\nThus, words with similar context and morphemes are represented by vectors close\nto each other. To evaluate our new approach, we performed intrinsic evaluations\nconsidering 15 different tasks, and the results show a competitive performance\ncompared to FastText.', 'Distributed representations of meaning are a natural way to encode covariance\nrelationships between words and phrases in NLP. By overcoming data sparsity\nproblems, as well as providing information about semantic relatedness which is\nnot available in discrete representations, distributed representations have\nproven useful in many NLP tasks. Recent work has shown how compositional\nsemantic representations can successfully be applied to a number of monolingual\napplications such as sentiment analysis. At the same time, there has been some\ninitial success in work on learning shared word-level representations across\nlanguages. We combine these two approaches by proposing a method for learning\ndistributed representations in a multilingual setup. Our model learns to assign\nsimilar embeddings to aligned sentences and dissimilar ones to sentence which\nare not aligned while not requiring word alignments. We show that our\nrepresentations are semantically informative and apply them to a cross-lingual\ndocument classification task where we outperform the previous state of the art.\nFurther, by employing parallel corpora of multiple language pairs we find that\nour model learns representations that capture semantic relationships across\nlanguages for which no parallel data was used.', 'Due to the increasing amount of data on the internet, finding a\nhighly-informative, low-dimensional representation for text is one of the main\nchallenges for efficient natural language processing tasks including text\nclassification. This representation should capture the semantic information of\nthe text while retaining their relevance level for document classification.\nThis approach maps the documents with similar topics to a similar space in\nvector space representation. To obtain representation for large text, we\npropose the utilization of deep Siamese neural networks. To embed document\nrelevance in topics in the distributed representation, we use a Siamese neural\nnetwork to jointly learn document representations. Our Siamese network consists\nof two sub-network of multi-layer perceptron. We examine our representation for\nthe text categorization task on BBC news dataset. The results show that the\nproposed representations outperform the conventional and state-of-the-art\nrepresentations in the text classification task on this dataset.', 'The field of natural language understanding has experienced exponential\nprogress in the last few years, with impressive results in several tasks. This\nsuccess has motivated researchers to study the underlying knowledge encoded by\nthese models. Despite this, attempts to understand their semantic capabilities\nhave not been successful, often leading to non-conclusive, or contradictory\nconclusions among different works. Via a probing classifier, we extract the\nunderlying knowledge graph of nine of the most influential language models of\nthe last years, including word embeddings, text generators, and context\nencoders. This probe is based on concept relatedness, grounded on WordNet. Our\nresults reveal that all the models encode this knowledge, but suffer from\nseveral inaccuracies. Furthermore, we show that the different architectures and\ntraining strategies lead to different model biases. We conduct a systematic\nevaluation to discover specific factors that explain why some concepts are\nchallenging. We hope our insights will motivate the development of models that\ncapture concepts more precisely.', 'Current approaches to learning vector representations of text that are\ncompatible between different languages usually require some amount of parallel\ntext, aligned at word, sentence or at least document level. We hypothesize\nhowever, that different natural languages share enough semantic structure that\nit should be possible, in principle, to learn compatible vector representations\njust by analyzing the monolingual distribution of words.\n  In order to evaluate this hypothesis, we propose a scheme to map word vectors\ntrained on a source language to vectors semantically compatible with word\nvectors trained on a target language using an adversarial autoencoder.\n  We present preliminary qualitative results and discuss possible future\ndevelopments of this technique, such as applications to cross-lingual sentence\nrepresentations.', 'The recently introduced continuous Skip-gram model is an efficient method for\nlearning high-quality distributed vector representations that capture a large\nnumber of precise syntactic and semantic word relationships. In this paper we\npresent several extensions that improve both the quality of the vectors and the\ntraining speed. By subsampling of the frequent words we obtain significant\nspeedup and also learn more regular word representations. We also describe a\nsimple alternative to the hierarchical softmax called negative sampling. An\ninherent limitation of word representations is their indifference to word order\nand their inability to represent idiomatic phrases. For example, the meanings\nof ""Canada"" and ""Air"" cannot be easily combined to obtain ""Air Canada"".\nMotivated by this example, we present a simple method for finding phrases in\ntext, and show that learning good vector representations for millions of\nphrases is possible.', 'Semantic neologisms (SN) are defined as words that acquire a new word meaning\nwhile maintaining their form. Given the nature of this kind of neologisms, the\ntask of identifying these new word meanings is currently performed manually by\nspecialists at observatories of neology. To detect SN in a semi-automatic way,\nwe developed a system that implements a combination of the following\nstrategies: topic modeling, keyword extraction, and word sense disambiguation.\nThe role of topic modeling is to detect the themes that are treated in the\ninput text. Themes within a text give clues about the particular meaning of the\nwords that are used, for example: viral has one meaning in the context of\ncomputer science (CS) and another when talking about health. To extract\nkeywords, we used TextRank with POS tag filtering. With this method, we can\nobtain relevant words that are already part of the Spanish lexicon. We use a\ndeep learning model to determine if a given keyword could have a new meaning.\nEmbeddings that are different from all the known meanings (or topics) indicate\nthat a word might be a valid SN candidate. In this study, we examine the\nfollowing word embedding models: Word2Vec, Sense2Vec, and FastText. The models\nwere trained with equivalent parameters using Wikipedia in Spanish as corpora.\nThen we used a list of words and their concordances (obtained from our database\nof neologisms) to show the different embeddings that each model yields.\nFinally, we present a comparison of these outcomes with the concordances of\neach word to show how we can determine if a word could be a valid candidate for\nSN.', 'The words of a language reflect the structure of the human mind, allowing us\nto transmit thoughts between individuals. However, language can represent only\na subset of our rich and detailed cognitive architecture. Here, we ask what\nkinds of common knowledge (semantic memory) are captured by word meanings\n(lexical semantics). We examine a prominent computational model that represents\nwords as vectors in a multidimensional space, such that proximity between\nword-vectors approximates semantic relatedness. Because related words appear in\nsimilar contexts, such spaces - called ""word embeddings"" - can be learned from\npatterns of lexical co-occurrences in natural language. Despite their\npopularity, a fundamental concern about word embeddings is that they appear to\nbe semantically ""rigid"": inter-word proximity captures only overall similarity,\nyet human judgments about object similarities are highly context-dependent and\ninvolve multiple, distinct semantic features. For example, dolphins and\nalligators appear similar in size, but differ in intelligence and\naggressiveness. Could such context-dependent relationships be recovered from\nword embeddings? To address this issue, we introduce a powerful, domain-general\nsolution: ""semantic projection"" of word-vectors onto lines that represent\nvarious object features, like size (the line extending from the word ""small"" to\n""big""), intelligence (from ""dumb"" to ""smart""), or danger (from ""safe"" to\n""dangerous""). This method, which is intuitively analogous to placing objects\n""on a mental scale"" between two extremes, recovers human judgments across a\nrange of object categories and properties. We thus show that word embeddings\ninherit a wealth of common knowledge from word co-occurrence statistics and can\nbe flexibly manipulated to express context-dependent meanings.', 'Methods for learning word sense embeddings represent a single word with\nmultiple sense-specific vectors. These methods should not only produce\ninterpretable sense embeddings, but should also learn how to select which sense\nto use in a given context. We propose an unsupervised model that learns sense\nembeddings using a modified Gumbel softmax function, which allows for\ndifferentiable discrete sense selection. Our model produces sense embeddings\nthat are competitive (and sometimes state of the art) on multiple similarity\nbased downstream evaluations. However, performance on these downstream\nevaluations tasks does not correlate with interpretability of sense embeddings,\nas we discover through an interpretability comparison with competing\nmulti-sense embeddings. While many previous approaches perform well on\ndownstream evaluations, they do not produce interpretable embeddings and learn\nduplicated sense groups; our method achieves the best of both worlds.', 'Word-vector representations associate a high dimensional real-vector to every\nword from a corpus. Recently, neural-network based methods have been proposed\nfor learning this representation from large corpora. This type of\nword-to-vector embedding is able to keep, in the learned vector space, some of\nthe syntactic and semantic relationships present in the original word corpus.\nThis, in turn, serves to address different types of language classification\ntasks by doing algebraic operations defined on the vectors. The general\npractice is to assume that the semantic relationships between the words can be\ninferred by the application of a-priori specified algebraic operations. Our\ngeneral goal in this paper is to show that it is possible to learn methods for\nword composition in semantic spaces. Instead of expressing the compositional\nmethod as an algebraic operation, we will encode it as a program, which can be\nlinear, nonlinear, or involve more intricate expressions. More remarkably, this\nprogram will be evolved from a set of initial random programs by means of\ngenetic programming (GP). We show that our method is able to reproduce the same\nbehavior as human-designed algebraic operators. Using a word analogy task as\nbenchmark, we also show that GP-generated programs are able to obtain accuracy\nvalues above those produced by the commonly used human-designed rule for\nalgebraic manipulation of word vectors. Finally, we show the robustness of our\napproach by executing the evolved programs on the word2vec GoogleNews vectors,\nlearned over 3 billion running words, and assessing their accuracy in the same\nword analogy task.', 'Retrieving information from correlative paragraphs or documents to answer\nopen-domain multi-hop questions is very challenging. To deal with this\nchallenge, most of the existing works consider paragraphs as nodes in a graph\nand propose graph-based methods to retrieve them. However, in this paper, we\npoint out the intrinsic defect of such methods. Instead, we propose a new\narchitecture that models paragraphs as sequential data and considers multi-hop\ninformation retrieval as a kind of sequence labeling task. Specifically, we\ndesign a rewritable external memory to model the dependency among paragraphs.\nMoreover, a threshold gate mechanism is proposed to eliminate the distraction\nof noise paragraphs. We evaluate our method on both full wiki and distractor\nsubtask of HotpotQA, a public textual multi-hop QA dataset requiring multi-hop\ninformation retrieval. Experiments show that our method achieves significant\nimprovement over the published state-of-the-art method in retrieval and\ndownstream QA task performance.', 'This paper presents a new model for word sense disambiguation formulated in\nterms of evolutionary game theory, where each word to be disambiguated is\nrepresented as a node on a graph whose edges represent word relations and\nsenses are represented as classes. The words simultaneously update their class\nmembership preferences according to the senses that neighboring words are\nlikely to choose. We use distributional information to weigh the influence that\neach word has on the decisions of the others and semantic similarity\ninformation to measure the strength of compatibility among the choices. With\nthis information we can formulate the word sense disambiguation problem as a\nconstraint satisfaction problem and solve it using tools derived from game\ntheory, maintaining the textual coherence. The model is based on two ideas:\nsimilar words should be assigned to similar classes and the meaning of a word\ndoes not depend on all the words in a text but just on some of them. The paper\nprovides an in-depth motivation of the idea of modeling the word sense\ndisambiguation problem in terms of game theory, which is illustrated by an\nexample. The conclusion presents an extensive analysis on the combination of\nsimilarity measures to use in the framework and a comparison with\nstate-of-the-art systems. The results show that our model outperforms\nstate-of-the-art algorithms and can be applied to different tasks and in\ndifferent scenarios.', 'In the context of natural language processing, representation learning has\nemerged as a newly active research subject because of its excellent performance\nin many applications. Learning representations of words is a pioneering study\nin this school of research. However, paragraph (or sentence and document)\nembedding learning is more suitable/reasonable for some tasks, such as\nsentiment classification and document summarization. Nevertheless, as far as we\nare aware, there is relatively less work focusing on the development of\nunsupervised paragraph embedding methods. Classic paragraph embedding methods\ninfer the representation of a given paragraph by considering all of the words\noccurring in the paragraph. Consequently, those stop or function words that\noccur frequently may mislead the embedding learning process to produce a misty\nparagraph representation. Motivated by these observations, our major\ncontributions in this paper are twofold. First, we propose a novel unsupervised\nparagraph embedding method, named the essence vector (EV) model, which aims at\nnot only distilling the most representative information from a paragraph but\nalso excluding the general background information to produce a more informative\nlow-dimensional vector representation for the paragraph. Second, in view of the\nincreasing importance of spoken content processing, an extension of the EV\nmodel, named the denoising essence vector (D-EV) model, is proposed. The D-EV\nmodel not only inherits the advantages of the EV model but also can infer a\nmore robust representation for a given spoken paragraph against imperfect\nspeech recognition.', 'Neural word representations are at the core of many state-of-the-art natural\nlanguage processing models. A widely used approach is to pre-train, store and\nlook up word or character embedding matrices. While useful, such\nrepresentations occupy huge memory making it hard to deploy on-device and often\ndo not generalize to unknown words due to vocabulary pruning.\n  In this paper, we propose a skip-gram based architecture coupled with\nLocality-Sensitive Hashing (LSH) projections to learn efficient dynamically\ncomputable representations. Our model does not need to store lookup tables as\nrepresentations are computed on-the-fly and require low memory footprint. The\nrepresentations can be trained in an unsupervised fashion and can be easily\ntransferred to other NLP tasks. For qualitative evaluation, we analyze the\nnearest neighbors of the word representations and discover semantically similar\nwords even with misspellings. For quantitative evaluation, we plug our\ntransferable projections into a simple LSTM and run it on multiple NLP tasks\nand show how our transferable projections achieve better performance compared\nto prior work.', ""Understanding the semantic relationships between terms is a fundamental task\nin natural language processing applications. While structured resources that\ncan express those relationships in a formal way, such as ontologies, are still\nscarce, a large number of linguistic resources gathering dictionary definitions\nis becoming available, but understanding the semantic structure of natural\nlanguage definitions is fundamental to make them useful in semantic\ninterpretation tasks. Based on an analysis of a subset of WordNet's glosses, we\npropose a set of semantic roles that compose the semantic structure of a\ndictionary definition, and show how they are related to the definition's\nsyntactic configuration, identifying patterns that can be used in the\ndevelopment of information extraction frameworks and semantic models."", 'Common language models typically predict the next word given the context. In\nthis work, we propose a method that improves language modeling by learning to\nalign the given context and the following phrase. The model does not require\nany linguistic annotation of phrase segmentation. Instead, we define syntactic\nheights and phrase segmentation rules, enabling the model to automatically\ninduce phrases, recognize their task-specific heads, and generate phrase\nembeddings in an unsupervised learning manner. Our method can easily be applied\nto language models with different network architectures since an independent\nmodule is used for phrase induction and context-phrase alignment, and no change\nis required in the underlying language modeling network. Experiments have shown\nthat our model outperformed several strong baseline models on different data\nsets. We achieved a new state-of-the-art performance of 17.4 perplexity on the\nWikitext-103 dataset. Additionally, visualizing the outputs of the phrase\ninduction module showed that our model is able to learn approximate\nphrase-level structural knowledge without any annotation.', 'Natural language processing (NLP) tasks tend to suffer from a paucity of\nsuitably annotated training data, hence the recent success of transfer learning\nacross a wide variety of them. The typical recipe involves: (i) training a\ndeep, possibly bidirectional, neural network with an objective related to\nlanguage modeling, for which training data is plentiful; and (ii) using the\ntrained network to derive contextual representations that are far richer than\nstandard linear word embeddings such as word2vec, and thus result in important\ngains. In this work, we wonder whether the opposite perspective is also true:\ncan contextual representations trained for different NLP tasks improve language\nmodeling itself? Since language models (LMs) are predominantly locally\noptimized, other NLP tasks may help them make better predictions based on the\nentire semantic fabric of a document. We test the performance of several types\nof pre-trained embeddings in neural LMs, and we investigate whether it is\npossible to make the LM more aware of global semantic information through\nembeddings pre-trained with a domain classification model. Initial experiments\nsuggest that as long as the proper objective criterion is used during training,\npre-trained embeddings are likely to be beneficial for neural language\nmodeling.', 'A neural language model trained on a text corpus can be used to induce\ndistributed representations of words, such that similar words end up with\nsimilar representations. If the corpus is multilingual, the same model can be\nused to learn distributed representations of languages, such that similar\nlanguages end up with similar representations. We show that this holds even\nwhen the multilingual corpus has been translated into English, by picking up\nthe faint signal left by the source languages. However, just like it is a\nthorny problem to separate semantic from syntactic similarity in word\nrepresentations, it is not obvious what type of similarity is captured by\nlanguage representations. We investigate correlations and causal relationships\nbetween language representations learned from translations on one hand, and\ngenetic, geographical, and several levels of structural similarity between\nlanguages on the other. Of these, structural similarity is found to correlate\nmost strongly with language representation similarity, while genetic\nrelationships---a convenient benchmark used for evaluation in previous\nwork---appears to be a confounding factor. Apart from implications about\ntranslation effects, we see this more generally as a case where NLP and\nlinguistic typology can interact and benefit one another.', 'Textbook Question Answering (TQA) is a task that one should answer a\ndiagram/non-diagram question given a large multi-modal context consisting of\nabundant essays and diagrams. We argue that the explainability of this task\nshould place students as a key aspect to be considered. To address this issue,\nwe devise a novel architecture towards span-level eXplanations of the TQA\n(XTQA) based on our proposed coarse-to-fine grained algorithm, which can\nprovide not only the answers but also the span-level evidences to choose them\nfor students. This algorithm first coarsely chooses top $M$ paragraphs relevant\nto questions using the TF-IDF method, and then chooses top $K$ evidence spans\nfinely from all candidate spans within these paragraphs by computing the\ninformation gain of each span to questions. Experimental results shows that\nXTQA significantly improves the state-of-the-art performance compared with\nbaselines. The source code is available at\nhttps://github.com/keep-smile-001/opentqa', 'The growing complexity of legal cases has lead to an increasing interest in\nlegal information retrieval systems that can effectively satisfy user-specific\ninformation needs. However, such downstream systems typically require documents\nto be properly formatted and segmented, which is often done with relatively\nsimple pre-processing steps, disregarding topical coherence of segments.\nSystems generally rely on representations of individual sentences or\nparagraphs, which may lack crucial context, or document-level representations,\nwhich are too long for meaningful search results. To address this issue, we\npropose a segmentation system that can predict topical coherence of sequential\ntext segments spanning several paragraphs, effectively segmenting a document\nand providing a more balanced representation for downstream applications. We\nbuild our model on top of popular transformer networks and formulate structural\ntext segmentation as topical change detection, by performing a series of\nindependent classifications that allow for efficient fine-tuning on\ntask-specific data. We crawl a novel dataset consisting of roughly $74,000$\nonline Terms-of-Service documents, including hierarchical topic annotations,\nwhich we use for training. Results show that our proposed system significantly\noutperforms baselines, and adapts well to structural peculiarities of legal\ndocuments. We release both data and trained models to the research community\nfor future work.https://github.com/dennlinger/TopicalChange', 'This paper is concerned with the task of multi-hop open-domain Question\nAnswering (QA). This task is particularly challenging since it requires the\nsimultaneous performance of textual reasoning and efficient searching. We\npresent a method for retrieving multiple supporting paragraphs, nested amidst a\nlarge knowledge base, which contain the necessary evidence to answer a given\nquestion. Our method iteratively retrieves supporting paragraphs by forming a\njoint vector representation of both a question and a paragraph. The retrieval\nis performed by considering contextualized sentence-level representations of\nthe paragraphs in the knowledge source. Our method achieves state-of-the-art\nperformance over two well-known datasets, SQuAD-Open and HotpotQA, which serve\nas our single- and multi-hop open-domain QA benchmarks, respectively.', 'Analysis of word embedding properties to inform their use in downstream NLP\ntasks has largely been studied by assessing nearest neighbors. However,\ngeometric properties of the continuous feature space contribute directly to the\nuse of embedding features in downstream models, and are largely unexplored. We\nconsider four properties of word embedding geometry, namely: position relative\nto the origin, distribution of features in the vector space, global pairwise\ndistances, and local pairwise distances. We define a sequence of\ntransformations to generate new embeddings that expose subsets of these\nproperties to downstream models and evaluate change in task performance to\nunderstand the contribution of each property to NLP models. We transform\npublicly available pretrained embeddings from three popular toolkits (word2vec,\nGloVe, and FastText) and evaluate on a variety of intrinsic tasks, which model\nlinguistic information in the vector space, and extrinsic tasks, which use\nvectors as input to machine learning models. We find that intrinsic evaluations\nare highly sensitive to absolute position, while extrinsic tasks rely primarily\non local similarity. Our findings suggest that future embedding models and\npost-processing techniques should focus primarily on similarity to nearby\npoints in vector space.', 'Semantic vector embedding techniques have proven useful in learning semantic\nrepresentations of data across multiple domains. A key application enabled by\nsuch techniques is the ability to measure semantic similarity between given\ndata samples and find data most similar to a given sample. State-of-the-art\nembedding approaches assume all data is available on a single site. However, in\nmany business settings, data is distributed across multiple edge locations and\ncannot be aggregated due to a variety of constraints. Hence, the applicability\nof state-of-the-art embedding approaches is limited to freely shared datasets,\nleaving out applications with sensitive or mission-critical data. This paper\naddresses this gap by proposing novel unsupervised algorithms called\n\\emph{SEEC} for learning and applying semantic vector embedding in a variety of\ndistributed settings. Specifically, for scenarios where multiple edge locations\ncan engage in joint learning, we adapt the recently proposed federated learning\ntechniques for semantic vector embedding. Where joint learning is not possible,\nwe propose novel semantic vector translation algorithms to enable semantic\nquery across multiple edge locations, each with its own semantic vector-space.\nExperimental results on natural language as well as graph datasets show that\nthis may be a promising new direction.', 'Though there are some works on improving distributed word representations\nusing lexicons, the improper overfitting of the words that have multiple\nmeanings is a remaining issue deteriorating the learning when lexicons are\nused, which needs to be solved. An alternative method is to allocate a vector\nper sense instead of a vector per word. However, the word representations\nestimated in the former way are not as easy to use as the latter one. Our\nprevious work uses a probabilistic method to alleviate the overfitting, but it\nis not robust with a small corpus. In this paper, we propose a new neural\nnetwork to estimate distributed word representations using a lexicon and a\ncorpus. We add a lexicon layer in the continuous bag-of-words model and a\nthreshold node after the output of the lexicon layer. The threshold rejects the\nunreliable outputs of the lexicon layer that are less likely to be the same\nwith their inputs. In this way, it alleviates the overfitting of the polysemous\nwords. The proposed neural network can be trained using negative sampling,\nwhich maximizing the log probabilities of target words given the context words,\nby distinguishing the target words from random noises. We compare the proposed\nneural network with the continuous bag-of-words model, the other works\nimproving it, and the previous works estimating distributed word\nrepresentations using both a lexicon and a corpus. The experimental results\nshow that the proposed neural network is more efficient and balanced for both\nsemantic tasks and syntactic tasks than the previous works, and robust to the\nsize of the corpus.', 'Words in natural language follow a Zipfian distribution whereby some words\nare frequent but most are rare. Learning representations for words in the ""long\ntail"" of this distribution requires enormous amounts of data. Representations\nof rare words trained directly on end tasks are usually poor, requiring us to\npre-train embeddings on external data, or treat all rare words as\nout-of-vocabulary words with a unique representation. We provide a method for\npredicting embeddings of rare words on the fly from small amounts of auxiliary\ndata with a network trained end-to-end for the downstream task. We show that\nthis improves results against baselines where embeddings are trained on the end\ntask for reading comprehension, recognizing textual entailment and language\nmodeling.', 'Loading models pre-trained on the large-scale corpus in the general domain\nand fine-tuning them on specific downstream tasks is gradually becoming a\nparadigm in Natural Language Processing. Previous investigations prove that\nintroducing a further pre-training phase between pre-training and fine-tuning\nphases to adapt the model on the domain-specific unlabeled data can bring\npositive effects. However, most of these further pre-training works just keep\nrunning the conventional pre-training task, e.g., masked language model, which\ncan be regarded as the domain adaptation to bridge the data distribution gap.\nAfter observing diverse downstream tasks, we suggest that different tasks may\nalso need a further pre-training phase with appropriate training tasks to\nbridge the task formulation gap. To investigate this, we carry out a study for\nimproving multiple task-oriented dialogue downstream tasks through designing\nvarious tasks at the further pre-training phase. The experiment shows that\ndifferent downstream tasks prefer different further pre-training tasks, which\nhave intrinsic correlation and most further pre-training tasks significantly\nimprove certain target tasks rather than all. Our investigation indicates that\nit is of great importance and effectiveness to design appropriate further\npre-training tasks modeling specific information that benefit downstream tasks.\nBesides, we present multiple constructive empirical conclusions for enhancing\ntask-oriented dialogues.', 'We propose an end-to-end neural model for zero-shot abstractive text\nsummarization of paragraphs, and introduce a benchmark task, ROCSumm, based on\nROCStories, a subset for which we collected human summaries. In this task,\nfive-sentence stories (paragraphs) are summarized with one sentence, using\nhuman summaries only for evaluation. We show results for extractive and human\nbaselines to demonstrate a large abstractive gap in performance. Our model,\nSummAE, consists of a denoising auto-encoder that embeds sentences and\nparagraphs in a common space, from which either can be decoded. Summaries for\nparagraphs are generated by decoding a sentence from the paragraph\nrepresentations. We find that traditional sequence-to-sequence auto-encoders\nfail to produce good summaries and describe how specific architectural choices\nand pre-training techniques can significantly improve performance,\noutperforming extractive baselines. The data, training, evaluation code, and\nbest model weights are open-sourced.', 'We describe an algorithm for automatic classification of idiomatic and\nliteral expressions. Our starting point is that words in a given text segment,\nsuch as a paragraph, that are highranking representatives of a common topic of\ndiscussion are less likely to be a part of an idiomatic expression. Our\nadditional hypothesis is that contexts in which idioms occur, typically, are\nmore affective and therefore, we incorporate a simple analysis of the intensity\nof the emotions expressed by the contexts. We investigate the bag of words\ntopic representation of one to three paragraphs containing an expression that\nshould be classified as idiomatic or literal (a target phrase). We extract\ntopics from paragraphs containing idioms and from paragraphs containing\nliterals using an unsupervised clustering method, Latent Dirichlet Allocation\n(LDA) (Blei et al., 2003). Since idiomatic expressions exhibit the property of\nnon-compositionality, we assume that they usually present different semantics\nthan the words used in the local topic. We treat idioms as semantic outliers,\nand the identification of a semantic shift as outlier detection. Thus, this\ntopic representation allows us to differentiate idioms from literals using\nlocal semantic contexts. Our results are encouraging.', 'Language models (LMs) pretrained on a large text corpus and fine-tuned on a\ndownstream text corpus and fine-tuned on a downstream task becomes a de facto\ntraining strategy for several natural language processing (NLP) tasks.\nRecently, an adaptive pretraining method retraining the pretrained language\nmodel with task-relevant data has shown significant performance improvements.\nHowever, current adaptive pretraining methods suffer from underfitting on the\ntask distribution owing to a relatively small amount of data to re-pretrain the\nLM. To completely use the concept of adaptive pretraining, we propose a\nback-translated task-adaptive pretraining (BT-TAPT) method that increases the\namount of task-specific data for LM re-pretraining by augmenting the task data\nusing back-translation to generalize the LM to the target task domain. The\nexperimental results show that the proposed BT-TAPT yields improved\nclassification accuracy on both low- and high-resource data and better\nrobustness to noise than the conventional adaptive pretraining method.', 'Deep learning currently dominates the benchmarks for various NLP tasks and,\nat the basis of such systems, words are frequently represented as embeddings\n--vectors in a low dimensional space-- learned from large text corpora and\nvarious algorithms have been proposed to learn both word and concept\nembeddings. One of the claimed benefits of such embeddings is that they capture\nknowledge about semantic relations. Such embeddings are most often evaluated\nthrough tasks such as predicting human-rated similarity and analogy which only\ntest a few, often ill-defined, relations. In this paper, we propose a method\nfor (i) reliably generating word and concept pair datasets for a wide number of\nrelations by using a knowledge graph and (ii) evaluating to what extent\npre-trained embeddings capture those relations. We evaluate the approach\nagainst a proprietary and a public knowledge graph and analyze the results,\nshowing which lexico-semantic relational knowledge is captured by current\nembedding learning approaches.', 'We introduce semantic form mid-tuning, an approach for transferring semantic\nknowledge from semantic meaning representations into transformer-based language\nencoders. In mid-tuning, we learn to align the text of general sentences -- not\ntied to any particular inference task -- and structured semantic\nrepresentations of those sentences. Our approach does not require gold\nannotated semantic representations. Instead, it makes use of automatically\ngenerated semantic representations, such as from off-the-shelf PropBank and\nFrameNet semantic parsers. We show that this alignment can be learned\nimplicitly via classification or directly via triplet loss. Our method yields\nlanguage encoders that demonstrate improved predictive performance across\ninference, reading comprehension, textual similarity, and other semantic tasks\ndrawn from the GLUE, SuperGLUE, and SentEval benchmarks. We evaluate our\napproach on three popular baseline models, where our experimental results and\nanalysis concludes that current pre-trained language models can further benefit\nfrom structured semantic frames with the proposed mid-tuning method, as they\ninject additional task-agnostic knowledge to the encoder, improving the\ngenerated embeddings as well as the linguistic properties of the given model,\nas evident from improvements on a popular sentence embedding toolkit and a\nvariety of probing tasks.', 'The trends of open science have enabled several open scholarly datasets which\ninclude millions of papers and authors. Managing, exploring, and utilizing such\nlarge and complicated datasets effectively are challenging. In recent years,\nthe knowledge graph has emerged as a universal data format for representing\nknowledge about heterogeneous entities and their relationships. The knowledge\ngraph can be modeled by knowledge graph embedding methods, which represent\nentities and relations as embedding vectors in semantic space, then model the\ninteractions between these embedding vectors. However, the semantic structures\nin the knowledge graph embedding space are not well-studied, thus knowledge\ngraph embedding methods are usually only used for knowledge graph completion\nbut not data representation and analysis. In this paper, we propose to analyze\nthese semantic structures based on the well-studied word embedding space and\nuse them to support data exploration. We also define the semantic queries,\nwhich are algebraic operations between the embedding vectors in the knowledge\ngraph embedding space, to solve queries such as similarity and analogy between\nthe entities on the original datasets. We then design a general framework for\ndata exploration by semantic queries and discuss the solution to some\ntraditional scholarly data exploration tasks. We also propose some new\ninteresting tasks that can be solved based on the uncanny semantic structures\nof the embedding space.', 'Neural network pretraining is gaining attention due to its outstanding\nperformance in natural language processing applications. However, pretraining\nusually leverages predefined task sequences to learn general linguistic clues.\nThe lack of mechanisms in choosing proper tasks during pretraining makes the\nlearning and knowledge encoding inefficient. We thus propose using\nmeta-learning to select tasks that provide the most informative learning\nsignals in each episode of pretraining. With the proposed method, we aim to\nachieve better efficiency in computation and memory usage for the pretraining\nprocess and resulting networks while maintaining the performance. In this\npreliminary work, we discuss the algorithm of the method and its two variants,\ndownstream-aware and downstream-agnostic pretraining. Our experiment plan is\nalso summarized, while empirical results will be shared in our future works.', ""Word embeddings typically represent different meanings of a word in a single\nconflated vector. Empirical analysis of embeddings of ambiguous words is\ncurrently limited by the small size of manually annotated resources and by the\nfact that word senses are treated as unrelated individual concepts. We present\na large dataset based on manual Wikipedia annotations and word senses, where\nword senses from different words are related by semantic classes. This is the\nbasis for novel diagnostic tests for an embedding's content: we probe word\nembeddings for semantic classes and analyze the embedding space by classifying\nembeddings into semantic classes. Our main findings are: (i) Information about\na sense is generally represented well in a single-vector embedding - if the\nsense is frequent. (ii) A classifier can accurately predict whether a word is\nsingle-sense or multi-sense, based only on its embedding. (iii) Although rare\nsenses are not well represented in single-vector embeddings, this does not have\nnegative impact on an NLP application whose performance depends on frequent\nsenses."", 'Semantic measures are widely used today to estimate the strength of the\nsemantic relationship between elements of various types: units of language\n(e.g., words, sentences, documents), concepts or even instances semantically\ncharacterized (e.g., diseases, genes, geographical locations). Semantic\nmeasures play an important role to compare such elements according to semantic\nproxies: texts and knowledge representations, which support their meaning or\ndescribe their nature. Semantic measures are therefore essential for designing\nintelligent agents which will for example take advantage of semantic analysis\nto mimic human ability to compare abstract or concrete objects. This paper\nproposes a comprehensive survey of the broad notion of semantic measure for the\ncomparison of units of language, concepts or instances based on semantic proxy\nanalyses. Semantic measures generalize the well-known notions of semantic\nsimilarity, semantic relatedness and semantic distance, which have been\nextensively studied by various communities over the last decades (e.g.,\nCognitive Sciences, Linguistics, and Artificial Intelligence to mention a few).', 'The manifold hypothesis suggests that word vectors live on a submanifold\nwithin their ambient vector space. We argue that we should, more accurately,\nexpect them to live on a pinched manifold: a singular quotient of a manifold\nobtained by identifying some of its points. The identified, singular points\ncorrespond to polysemous words, i.e. words with multiple meanings. Our point of\nview suggests that monosemous and polysemous words can be distinguished based\non the topology of their neighbourhoods. We present two kinds of empirical\nevidence to support this point of view: (1) We introduce a topological measure\nof polysemy based on persistent homology that correlates well with the actual\nnumber of meanings of a word. (2) We propose a simple, topologically motivated\nsolution to the SemEval-2010 task on Word Sense Induction & Disambiguation that\nproduces competitive results.', 'Word embeddings, i.e., low-dimensional vector representations such as GloVe\nand SGNS, encode word ""meaning"" in the sense that distances between words\'\nvectors correspond to their semantic proximity. This enables transfer learning\nof semantics for a variety of natural language processing tasks.\n  Word embeddings are typically trained on large public corpora such as\nWikipedia or Twitter. We demonstrate that an attacker who can modify the corpus\non which the embedding is trained can control the ""meaning"" of new and existing\nwords by changing their locations in the embedding space. We develop an\nexplicit expression over corpus features that serves as a proxy for distance\nbetween words and establish a causative relationship between its values and\nembedding distances. We then show how to use this relationship for two\nadversarial objectives: (1) make a word a top-ranked neighbor of another word,\nand (2) move a word from one semantic cluster to another.\n  An attack on the embedding can affect diverse downstream tasks, demonstrating\nfor the first time the power of data poisoning in transfer learning scenarios.\nWe use this attack to manipulate query expansion in information retrieval\nsystems such as resume search, make certain names more or less visible to named\nentity recognition models, and cause new words to be translated to a particular\ntarget word regardless of the language. Finally, we show how the attacker can\ngenerate linguistically likely corpus modifications, thus fooling defenses that\nattempt to filter implausible sentences from the corpus using a language model.', 'We have developed a method for extracting the coherence features from a\nparagraph by matching similar words in its sentences. We conducted an\nexperiment with a parallel German corpus containing 2000 human-created and 2000\nmachine-translated paragraphs. The result showed that our method achieved the\nbest performance (accuracy = 72.3%, equal error rate = 29.8%) when it is\ncompared with previous methods on various computer-generated text including\ntranslation and paper generation (best accuracy = 67.9%, equal error rate =\n32.0%). Experiments on Dutch, another rich resource language, and a low\nresource one (Japanese) attained similar performances. It demonstrated the\nefficiency of the coherence features at distinguishing computer-translated from\nhuman-created paragraphs on diverse languages.', ""The offset method for solving word analogies has become a standard evaluation\ntool for vector-space semantic models: it is considered desirable for a space\nto represent semantic relations as consistent vector offsets. We show that the\nmethod's reliance on cosine similarity conflates offset consistency with\nlargely irrelevant neighborhood structure, and propose simple baselines that\nshould be used to improve the utility of the method in vector space evaluation."", 'We present a method for coarse-grained cross-lingual alignment of comparable\ntexts: segments consisting of contiguous paragraphs that discuss the same theme\n(e.g. history, economy) are aligned based on induced multilingual topics. The\nmethod combines three ideas: a two-level LDA model that filters out words that\ndo not convey themes, an HMM that models the ordering of themes in the\ncollection of documents, and language-independent concept annotations to serve\nas a cross-language bridge and to strengthen the connection between paragraphs\nin the same segment through concept relations. The method is evaluated on\nEnglish and French data previously used for monolingual alignment. The results\nshow state-of-the-art performance in both monolingual and cross-lingual\nsettings.', 'In this paper we show that if we want to obtain human evidence about\nconventionalization of some phrases, we should ask native speakers about\nassociations they have to a given phrase and its component words. We have shown\nthat if component words of a phrase have each other as frequent associations,\nthen this phrase can be considered as conventionalized. Another type of\nconventionalized phrases can be revealed using two factors: low entropy of\nphrase associations and low intersection of component word and phrase\nassociations. The association experiments were performed for the Russian\nlanguage.', 'Biased language commonly occurs around topics which are of controversial\nnature, thus, stirring disagreement between the different involved parties of a\ndiscussion. This is due to the fact that for language and its use,\nspecifically, the understanding and use of phrases, the stances are cohesive\nwithin the particular groups. However, such cohesiveness does not hold across\ngroups.\n  In collaborative environments or environments where impartial language is\ndesired (e.g. Wikipedia, news media), statements and the language therein\nshould represent equally the involved parties and be neutrally phrased. Biased\nlanguage is introduced through the presence of inflammatory words or phrases,\nor statements that may be incorrect or one-sided, thus violating such\nconsensus.\n  In this work, we focus on the specific case of phrasing bias, which may be\nintroduced through specific inflammatory words or phrases in a statement. For\nthis purpose, we propose an approach that relies on a recurrent neural networks\nin order to capture the inter-dependencies between words in a phrase that\nintroduced bias.\n  We perform a thorough experimental evaluation, where we show the advantages\nof a neural based approach over competitors that rely on word lexicons and\nother hand-crafted features in detecting biased language. We are able to\ndistinguish biased statements with a precision of P=0.92, thus significantly\noutperforming baseline models with an improvement of over 30%. Finally, we\nrelease the largest corpus of statements annotated for biased language.', 'Inspired by the success of BERT, several multimodal representation learning\napproaches have been proposed that jointly represent image and text. These\napproaches achieve superior performance by capturing high-level semantic\ninformation from large-scale multimodal pretraining. In particular, LXMERT and\nUNITER adopt visual region feature regression and label classification as\npretext tasks. However, they tend to suffer from the problems of noisy labels\nand sparse semantic annotations, based on the visual features having been\npretrained on a crowdsourced dataset with limited and inconsistent semantic\nlabeling. To overcome these issues, we propose unbiased Dense Contrastive\nVisual-Linguistic Pretraining (DCVLP), which replaces the region regression and\nclassification with cross-modality region contrastive learning that requires no\nannotations. Two data augmentation strategies (Mask Perturbation and\nIntra-/Inter-Adversarial Perturbation) are developed to improve the quality of\nnegative samples used in contrastive learning. Overall, DCVLP allows\ncross-modality dense region contrastive learning in a self-supervised setting\nindependent of any object annotations. We compare our method against prior\nvisual-linguistic pretraining frameworks to validate the superiority of dense\ncontrastive learning on multimodal representation learning.', 'Automatic question generation is an important problem in natural language\nprocessing. In this paper we propose a novel adaptive copying recurrent neural\nnetwork model to tackle the problem of question generation from sentences and\nparagraphs. The proposed model adds a copying mechanism component onto a\nbidirectional LSTM architecture to generate more suitable questions adaptively\nfrom the input data. Our experimental results show the proposed model can\noutperform the state-of-the-art question generation methods in terms of BLEU\nand ROUGE evaluation scores.', 'Measuring what linguistic information is encoded in neural models of language\nhas become popular in NLP. Researchers approach this enterprise by training\n""probes"" - supervised models designed to extract linguistic structure from\nanother model\'s output. One such probe is the structural probe (Hewitt and\nManning, 2019), designed to quantify the extent to which syntactic information\nis encoded in contextualised word representations. The structural probe has a\nnovel design, unattested in the parsing literature, the precise benefit of\nwhich is not immediately obvious. To explore whether syntactic probes would do\nbetter to make use of existing techniques, we compare the structural probe to a\nmore traditional parser with an identical lightweight parameterisation. The\nparser outperforms structural probe on UUAS in seven of nine analysed\nlanguages, often by a substantial amount (e.g. by 11.1 points in English).\nUnder a second less common metric, however, there is the opposite trend - the\nstructural probe outperforms the parser. This begs the question: which metric\nshould we prefer?', 'Automatic question generation from paragraphs is an important and challenging\nproblem, particularly due to the long context from paragraphs. In this paper,\nwe propose and study two hierarchical models for the task of question\ngeneration from paragraphs. Specifically, we propose (a) a novel hierarchical\nBiLSTM model with selective attention and (b) a novel hierarchical Transformer\narchitecture, both of which learn hierarchical representations of paragraphs.\nWe model a paragraph in terms of its constituent sentences, and a sentence in\nterms of its constituent words. While the introduction of the attention\nmechanism benefits the hierarchical BiLSTM model, the hierarchical Transformer,\nwith its inherent attention and positional encoding mechanisms also performs\nbetter than flat transformer model. We conducted empirical evaluation on the\nwidely used SQuAD and MS MARCO datasets using standard metrics. The results\ndemonstrate the overall effectiveness of the hierarchical models over their\nflat counterparts. Qualitatively, our hierarchical models are able to generate\nfluent and relevant questions', 'The following paper presents a method of comparing two sets of vectors. The\nmethod can be applied in all tasks, where it is necessary to measure the\ncloseness of two objects presented as sets of vectors. It may be applicable\nwhen we compare the meanings of two sentences as part of the problem of\nparaphrasing. This is the problem of measuring semantic similarity of two\nsentences (group of words). The existing methods are not sensible for the word\norder or syntactic connections in the considered sentences. The method appears\nto be advantageous because it neither presents a group of words as one scalar\nvalue, nor does it try to show the closeness through an aggregation vector,\nwhich is mean for the set of vectors. Instead of that we measure the cosine of\nthe angle as the mean for the first group vectors projections (the context) on\none side and each vector of the second group on the other side. The similarity\nof two sentences defined by these means does not lose any semantic\ncharacteristics and takes account of the words traits. The method was verified\non the comparison of sentence pairs in Russian.', ""We apply entropy agglomeration (EA), a recently introduced algorithm, to\ncluster the words of a literary text. EA is a greedy agglomerative procedure\nthat minimizes projection entropy (PE), a function that can quantify the\nsegmentedness of an element set. To apply it, the text is reduced to a feature\nallocation, a combinatorial object to represent the word occurences in the\ntext's paragraphs. The experiment results demonstrate that EA, despite its\nreduction and simplicity, is useful in capturing significant relationships\namong the words in the text. This procedure was implemented in Python and\npublished as a free software: REBUS."", 'Multi-hop reading comprehension (RC) questions are challenging because they\nrequire reading and reasoning over multiple paragraphs. We argue that it can be\ndifficult to construct large multi-hop RC datasets. For example, even highly\ncompositional questions can be answered with a single hop if they target\nspecific entity types, or the facts needed to answer them are redundant. Our\nanalysis is centered on HotpotQA, where we show that single-hop reasoning can\nsolve much more of the dataset than previously thought. We introduce a\nsingle-hop BERT-based RC model that achieves 67 F1---comparable to\nstate-of-the-art multi-hop models. We also design an evaluation setting where\nhumans are not shown all of the necessary paragraphs for the intended multi-hop\nreasoning but can still answer over 80% of questions. Together with detailed\nerror analysis, these results suggest there should be an increasing focus on\nthe role of evidence in multi-hop reasoning and possibly even a shift towards\ninformation retrieval style evaluations with large and diverse evidence\ncollections.', 'Text-to-speech systems are typically evaluated on single sentences. When\nlong-form content, such as data consisting of full paragraphs or dialogues is\nconsidered, evaluating sentences in isolation is not always appropriate as the\ncontext in which the sentences are synthesized is missing. In this paper, we\ninvestigate three different ways of evaluating the naturalness of long-form\ntext-to-speech synthesis. We compare the results obtained from evaluating\nsentences in isolation, evaluating whole paragraphs of speech, and presenting a\nselection of speech or text as context and evaluating the subsequent speech. We\nfind that, even though these three evaluations are based upon the same\nmaterial, the outcomes differ per setting, and moreover that these outcomes do\nnot necessarily correlate with each other. We show that our findings are\nconsistent between a single speaker setting of read paragraphs and a\ntwo-speaker dialogue scenario. We conclude that to evaluate the quality of\nlong-form speech, the traditional way of evaluating sentences in isolation does\nnot suffice, and that multiple evaluations are required.', 'We consider the problem of adapting neural paragraph-level question answering\nmodels to the case where entire documents are given as input. Our proposed\nsolution trains models to produce well calibrated confidence scores for their\nresults on individual paragraphs. We sample multiple paragraphs from the\ndocuments during training, and use a shared-normalization training objective\nthat encourages the model to produce globally correct output. We combine this\nmethod with a state-of-the-art pipeline for training models on document QA\ndata. Experiments demonstrate strong performance on several document QA\ndatasets. Overall, we are able to achieve a score of 71.3 F1 on the web portion\nof TriviaQA, a large improvement from the 56.7 F1 of the previous best system.', 'Generating syntactically and semantically valid and relevant questions from\nparagraphs is useful with many applications. Manual generation is a\nlabour-intensive task, as it requires the reading, parsing and understanding of\nlong passages of text. A number of question generation models based on\nsequence-to-sequence techniques have recently been proposed. Most of them\ngenerate questions from sentences only, and none of them is publicly available\nas an easy-to-use service. In this paper, we demonstrate ParaQG, a Web-based\nsystem for generating questions from sentences and paragraphs. ParaQG\nincorporates a number of novel functionalities to make the question generation\nprocess user-friendly. It provides an interactive interface for a user to\nselect answers with visual insights on generation of questions. It also employs\nvarious faceted views to group similar questions as well as filtering\ntechniques to eliminate unanswerable questions', 'One key consequence of the information revolution is a significant increase\nand a contamination of our information supply. The practice of fact checking\nwon\'t suffice to eliminate the biases in text data we observe, as the degree of\nfactuality alone does not determine whether biases exist in the spectrum of\nopinions visible to us. To better understand controversial issues, one needs to\nview them from a diverse yet comprehensive set of perspectives. For example,\nthere are many ways to respond to a claim such as ""animals should have lawful\nrights"", and these responses form a spectrum of perspectives, each with a\nstance relative to this claim and, ideally, with evidence supporting it.\nInherently, this is a natural language understanding task, and we propose to\naddress it as such. Specifically, we propose the task of substantiated\nperspective discovery where, given a claim, a system is expected to discover a\ndiverse set of well-corroborated perspectives that take a stance with respect\nto the claim. Each perspective should be substantiated by evidence paragraphs\nwhich summarize pertinent results and facts. We construct PERSPECTRUM, a\ndataset of claims, perspectives and evidence, making use of online debate\nwebsites to create the initial data collection, and augmenting it using search\nengines in order to expand and diversify our dataset. We use crowd-sourcing to\nfilter out noise and ensure high-quality data. Our dataset contains 1k claims,\naccompanied with pools of 10k and 8k perspective sentences and evidence\nparagraphs, respectively. We provide a thorough analysis of the dataset to\nhighlight key underlying language understanding challenges, and show that human\nbaselines across multiple subtasks far outperform ma-chine baselines built upon\nstate-of-the-art NLP techniques. This poses a challenge and opportunity for the\nNLP community to address.', 'Children learn word meanings by tapping into the commonalities across\ndifferent situations in which words are used and overcome the high level of\nuncertainty involved in early word learning experiences. We propose a modeling\nframework to investigate the role of mutual exclusivity bias - asserting\none-to-one mappings between words and their meanings - in reducing uncertainty\nin word learning. In a set of computational studies, we show that to\nsuccessfully learn word meanings in the face of uncertainty, a learner needs to\nuse two types of competition: words competing for association to a referent\nwhen learning from an observation and referents competing for a word when the\nword is used. Our work highlights the importance of an algorithmic-level\nanalysis to shed light on the utility of different mechanisms that can\nimplement the same computational-level theory.', 'This paper introduces a new framework for open-domain question answering in\nwhich the retriever and the reader iteratively interact with each other. The\nframework is agnostic to the architecture of the machine reading model, only\nrequiring access to the token-level hidden representations of the reader. The\nretriever uses fast nearest neighbor search to scale to corpora containing\nmillions of paragraphs. A gated recurrent unit updates the query at each step\nconditioned on the state of the reader and the reformulated query is used to\nre-rank the paragraphs by the retriever. We conduct analysis and show that\niterative interaction helps in retrieving informative paragraphs from the\ncorpus. Finally, we show that our multi-step-reasoning framework brings\nconsistent improvement when applied to two widely used reader architectures\nDrQA and BiDAF on various large open-domain datasets --- TriviaQA-unfiltered,\nQuasarT, SearchQA, and SQuAD-Open.', 'We study the time taken by a language learner to correctly identify the\nmeaning of all words in a lexicon under conditions where many plausible\nmeanings can be inferred whenever a word is uttered. We show that the most\nbasic form of cross-situational learning - whereby information from multiple\nepisodes is combined to eliminate incorrect meanings - can perform badly when\nwords are learned independently and meanings are drawn from a nonuniform\ndistribution. If learners further assume that no two words share a common\nmeaning, we find a phase transition between a maximally-efficient learning\nregime, where the learning time is reduced to the shortest it can possibly be,\nand a partially-efficient regime where incorrect candidate meanings for words\npersist at late times. We obtain exact results for the word-learning process\nthrough an equivalence to a statistical mechanical problem of enumerating loops\nin the space of word-meaning mappings.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : In my reinforcement learning (RL) research, I encountered a significant challenge: the inability of a policy learned in one environment to adapt or generalize to another. I am seeking an efficient method to transfer policies between environments without the need to train a new policy from scratch each time the environment changes. Currently, I am considering the use of natural language to describe the entire dynamics of the environment or to outline the anticipated changes in the environment. To implement this, I require a language model capable of encoding these natural language descriptions. The resulting encoded latent representations should aid in facilitating policy adaptation to different environments. It appears crucial to align the text's meaning with the environment's dynamics, including transition probability and reward functions. Ultimately, I aim to assess the effectiveness of this approach in terms of transfer and multi-task scenarios across various environments and compare it with existing models. I am interested in any literature related to the use of textual descriptions in RL and any existing datasets suitable for evaluating my proposed methodology.","['Obtaining policies that can generalise to new environments in reinforcement\nlearning is challenging. In this work, we demonstrate that language\nunderstanding via a reading policy learner is a promising vehicle for\ngeneralisation to new environments. We propose a grounded policy learning\nproblem, Read to Fight Monsters (RTFM), in which the agent must jointly reason\nover a language goal, relevant dynamics described in a document, and\nenvironment observations. We procedurally generate environment dynamics and\ncorresponding language descriptions of the dynamics, such that agents must read\nto understand new environment dynamics instead of memorising any particular\ninformation. In addition, we propose txt2$\\pi$, a model that captures three-way\ninteractions between the goal, document, and observations. On RTFM, txt2$\\pi$\ngeneralises to new environments with dynamics not seen during training via\nreading. Furthermore, our model outperforms baselines such as FiLM and\nlanguage-conditioned CNNs on RTFM. Through curriculum learning, txt2$\\pi$\nproduces policies that excel on complex RTFM tasks requiring several reasoning\nand coreference steps.', 'Despite recent breakthroughs in reinforcement learning (RL) and imitation\nlearning (IL), existing algorithms fail to generalize beyond the training\nenvironments. In reality, humans can adapt to new tasks quickly by leveraging\nprior knowledge about the world such as language descriptions. To facilitate\nthe research on language-guided agents with domain adaption, we propose a novel\nzero-shot compositional policy learning task, where the environments are\ncharacterized as a composition of different attributes. Since there are no\npublic environments supporting this study, we introduce a new research platform\nBabyAI++ in which the dynamics of environments are disentangled from visual\nappearance. At each episode, BabyAI++ provides varied vision-dynamics\ncombinations along with corresponding descriptive texts. To evaluate the\nadaption capability of learned agents, a set of vision-dynamics pairings are\nheld-out for testing on BabyAI++. Unsurprisingly, we find that current\nlanguage-guided RL/IL techniques overfit to the training environments and\nsuffer from a huge performance drop when facing unseen combinations. In\nresponse, we propose a multi-modal fusion method with an attention mechanism to\nperform visual language-grounding. Extensive experiments show strong evidence\nthat language grounding is able to improve the generalization of agents across\nenvironments with varied dynamics.', ""Deep Reinforcement Learning (RL) is mainly studied in a setting where the\ntraining and the testing environments are similar. But in many practical\napplications, these environments may differ. For instance, in control systems,\nthe robot(s) on which a policy is learned might differ from the robot(s) on\nwhich a policy will run. It can be caused by different internal factors (e.g.,\ncalibration issues, system attrition, defective modules) or also by external\nchanges (e.g., weather conditions). There is a need to develop RL methods that\ngeneralize well to variations of the training conditions. In this article, we\nconsider the simplest yet hard to tackle generalization setting where the test\nenvironment is unknown at train time, forcing the agent to adapt to the\nsystem's new dynamics. This online adaptation process can be computationally\nexpensive (e.g., fine-tuning) and cannot rely on meta-RL techniques since there\nis just a single train environment. To do so, we propose an approach where we\nlearn a subspace of policies within the parameter space. This subspace contains\nan infinite number of policies that are trained to solve the training\nenvironment while having different parameter values. As a consequence, two\npolicies in that subspace process information differently and exhibit different\nbehaviors when facing variations of the train environment. Our experiments\ncarried out over a large variety of benchmarks compare our approach with\nbaselines, including diversity-based methods. In comparison, our approach is\nsimple to tune, does not need any extra component (e.g., discriminator) and\nlearns policies able to gather a high reward on unseen environments."", ""Reinforcement Learning has shown success in a number of complex virtual\nenvironments. However, many challenges still exist towards solving problems\nwith natural language as a core component. Interactive Fiction Games (or Text\nGames) are one such problem type that offer a set of partially observable\nenvironments where natural language is required as part of the reinforcement\nlearning solutions.\n  Therefore, this survey's aim is to assist in the development of new Text Game\nproblem settings and solutions for Reinforcement Learning informed by natural\nlanguage. Specifically, this survey summarises: 1) the challenges introduced in\nText Game Reinforcement Learning problems, 2) the generation tools for\nevaluating Text Games and the subsequent environments generated and, 3) the\nagent architectures currently applied are compared to provide a systematic\nreview of benchmark methodologies and opportunities for future researchers."", 'Deep reinforcement learning (RL) has made it possible to solve complex\nrobotics problems using neural networks as function approximators. However, the\npolicies trained on stationary environments suffer in terms of generalization\nwhen transferred from one environment to another. In this work, we use Robust\nMarkov Decision Processes (RMDP) to train the drone control policy, which\ncombines ideas from Robust Control and RL. It opts for pessimistic optimization\nto handle potential gaps between policy transfer from one environment to\nanother. The trained control policy is tested on the task of quadcopter\npositional control. RL agents were trained in a MuJoCo simulator. During\ntesting, different environment parameters (unseen during the training) were\nused to validate the robustness of the trained policy for transfer from one\nenvironment to another. The robust policy outperformed the standard agents in\nthese environments, suggesting that the added robustness increases generality\nand can adapt to non-stationary environments.\n  Codes: https://github.com/adipandas/gym_multirotor', 'The ability to transfer a policy from one environment to another is a\npromising avenue for efficient robot learning in realistic settings where task\nsupervision is not available. This can allow us to take advantage of\nenvironments well suited for training, such as simulators or laboratories, to\nlearn a policy for a real robot in a home or office. To succeed, such policy\ntransfer must overcome both the visual domain gap (e.g. different illumination\nor background) and the dynamics domain gap (e.g. different robot calibration or\nmodelling error) between source and target environments. However, prior policy\ntransfer approaches either cannot handle a large domain gap or can only address\none type of domain gap at a time. In this paper, we propose a novel policy\ntransfer method with iterative ""environment grounding"", IDAPT, that alternates\nbetween (1) directly minimizing both visual and dynamics domain gaps by\ngrounding the source environment in the target environment domains, and (2)\ntraining a policy on the grounded source environment. This iterative training\nprogressively aligns the domains between the two environments and adapts the\npolicy to the target environment. Once trained, the policy can be directly\nexecuted on the target environment. The empirical results on locomotion and\nrobotic manipulation tasks demonstrate that our approach can effectively\ntransfer a policy across visual and dynamics domain gaps with minimal\nsupervision and interaction with the target environment. Videos and code are\navailable at https://clvrai.com/idapt .', 'Deep reinforcement learning could be used to learn dexterous robotic policies\nbut it is challenging to transfer them to new robots with vastly different\nhardware properties. It is also prohibitively expensive to learn a new policy\nfrom scratch for each robot hardware due to the high sample complexity of\nmodern state-of-the-art algorithms. We propose a novel approach called\n\\textit{Hardware Conditioned Policies} where we train a universal policy\nconditioned on a vector representation of robot hardware. We considered robots\nin simulation with varied dynamics, kinematic structure, kinematic lengths and\ndegrees-of-freedom. First, we use the kinematic structure directly as the\nhardware encoding and show great zero-shot transfer to completely novel robots\nnot seen during training. For robots with lower zero-shot success rate, we also\ndemonstrate that fine-tuning the policy network is significantly more\nsample-efficient than training a model from scratch. In tasks where knowing the\nagent dynamics is important for success, we learn an embedding for robot\nhardware and show that policies conditioned on the encoding of hardware tend to\ngeneralize and transfer well. The code and videos are available on the project\nwebpage: https://sites.google.com/view/robot-transfer-hcp.', 'Transfer and adaptation to new unknown environmental dynamics is a key\nchallenge for reinforcement learning (RL). An even greater challenge is\nperforming near-optimally in a single attempt at test time, possibly without\naccess to dense rewards, which is not addressed by current methods that require\nmultiple experience rollouts for adaptation. To achieve single episode transfer\nin a family of environments with related dynamics, we propose a general\nalgorithm that optimizes a probe and an inference model to rapidly estimate\nunderlying latent variables of test dynamics, which are then immediately used\nas input to a universal control policy. This modular approach enables\nintegration of state-of-the-art algorithms for variational inference or RL.\nMoreover, our approach does not require access to rewards at test time,\nallowing it to perform in settings where existing adaptive approaches cannot.\nIn diverse experimental domains with a single episode test constraint, our\nmethod significantly outperforms existing adaptive approaches and shows\nfavorable performance against baselines for robust transfer.', 'In this work we present a technique to use natural language to help\nreinforcement learning generalize to unseen environments. This technique uses\nneural machine translation, specifically the use of encoder-decoder networks,\nto learn associations between natural language behavior descriptions and\nstate-action information. We then use this learned model to guide agent\nexploration using a modified version of policy shaping to make it more\neffective at learning in unseen environments. We evaluate this technique using\nthe popular arcade game, Frogger, under ideal and non-ideal conditions. This\nevaluation shows that our modified policy shaping algorithm improves over a\nQ-learning agent as well as a baseline version of policy shaping.', 'The ability to adapt to changes in environmental contingencies is an\nimportant challenge in reinforcement learning. Indeed, transferring previously\nacquired knowledge to environments with unseen structural properties can\ngreatly enhance the flexibility and efficiency by which novel optimal policies\nmay be constructed. In this work, we study the problem of transfer learning\nunder changes in the environment dynamics. In this study, we apply causal\nreasoning in the offline reinforcement learning setting to transfer a learned\npolicy to new environments. Specifically, we use the Decision Transformer (DT)\narchitecture to distill a new policy on the new environment. The DT is trained\non data collected by performing policy rollouts on factual and counterfactual\nsimulations from the source environment. We show that this mechanism can\nbootstrap a successful policy on the target environment while retaining most of\nthe reward.', 'The ability to act in multiple environments and transfer previous knowledge\nto new situations can be considered a critical aspect of any intelligent agent.\nTowards this goal, we define a novel method of multitask and transfer learning\nthat enables an autonomous agent to learn how to behave in multiple tasks\nsimultaneously, and then generalize its knowledge to new domains. This method,\ntermed ""Actor-Mimic"", exploits the use of deep reinforcement learning and model\ncompression techniques to train a single policy network that learns how to act\nin a set of distinct tasks by using the guidance of several expert teachers. We\nthen show that the representations learnt by the deep policy network are\ncapable of generalizing to new tasks with no prior expert guidance, speeding up\nlearning in novel environments. Although our method can in general be applied\nto a wide range of problems, we use Atari games as a testing environment to\ndemonstrate these methods.', 'We investigate the use of natural language to drive the generalization of\ncontrol policies and introduce the new multi-task environment Messenger with\nfree-form text manuals describing the environment dynamics. Unlike previous\nwork, Messenger does not assume prior knowledge connecting text and state\nobservations $-$ the control policy must simultaneously ground the game manual\nto entity symbols and dynamics in the environment. We develop a new model, EMMA\n(Entity Mapper with Multi-modal Attention) which uses an entity-conditioned\nattention module that allows for selective focus over relevant descriptions in\nthe manual for each entity in the environment. EMMA is end-to-end\ndifferentiable and learns a latent grounding of entities and dynamics from text\nto observations using only environment rewards. EMMA achieves successful\nzero-shot generalization to unseen games with new dynamics, obtaining a 40%\nhigher win rate compared to multiple baselines. However, win rate on the\nhardest stage of Messenger remains low (10%), demonstrating the need for\nadditional work in this direction.', 'Recent breakthroughs in Go play and strategic games have witnessed the great\npotential of reinforcement learning in intelligently scheduling in uncertain\nenvironment, but some bottlenecks are also encountered when we generalize this\nparadigm to universal complex tasks. Among them, the low efficiency of data\nutilization in model-free reinforcement algorithms is of great concern. In\ncontrast, the model-based reinforcement learning algorithms can reveal\nunderlying dynamics in learning environments and seldom suffer the data\nutilization problem. To address the problem, a model-based reinforcement\nlearning algorithm with attention mechanism embedded is proposed as an\nextension of World Models in this paper. We learn the environment model through\nMixture Density Network Recurrent Network(MDN-RNN) for agents to interact, with\ncombinations of variational auto-encoder(VAE) and attention incorporated in\nstate value estimates during the process of learning policy. In this way, agent\ncan learn optimal policies through less interactions with actual environment,\nand final experiments demonstrate the effectiveness of our model in control\nproblem.', 'The ability to quickly solve a wide range of real-world tasks requires a\ncommonsense understanding of the world. Yet, how to best extract such knowledge\nfrom natural language corpora and integrate it with reinforcement learning (RL)\nagents remains an open challenge. This is partly due to the lack of lightweight\nsimulation environments that sufficiently reflect the semantics of the real\nworld and provide knowledge sources grounded with respect to observations in an\nRL environment. To better enable research on agents making use of commonsense\nknowledge, we propose WordCraft, an RL environment based on Little Alchemy 2.\nThis lightweight environment is fast to run and built upon entities and\nrelations inspired by real-world semantics. We evaluate several representation\nlearning methods on this new benchmark and propose a new method for integrating\nknowledge graphs with an RL agent.', 'Efficient and robust policy transfer remains a key challenge for\nreinforcement learning to become viable for real-wold robotics. Policy transfer\nthrough warm initialization, imitation, or interacting over a large set of\nagents with randomized instances, have been commonly applied to solve a variety\nof Reinforcement Learning tasks. However, this seems far from how skill\ntransfer happens in the biological world: Humans and animals are able to\nquickly adapt the learned behaviors between similar tasks and learn new skills\nwhen presented with new situations. Here we seek to answer the question: Will\nlearning to combine adaptation and exploration lead to a more efficient\ntransfer of policies between domains? We introduce a principled mechanism that\ncan ""Adapt-to-Learn"", that is adapt the source policy to learn to solve a\ntarget task with significant transition differences and uncertainties. We show\nthat the presented method learns to seamlessly combine learning from adaptation\nand exploration and leads to a robust policy transfer algorithm with\nsignificantly reduced sample complexity in transferring skills between related\ntasks.', 'Machine Learning models become increasingly proficient in complex tasks.\nHowever, even for experts in the field, it can be difficult to understand what\nthe model learned. This hampers trust and acceptance, and it obstructs the\npossibility to correct the model. There is therefore a need for transparency of\nmachine learning models. The development of transparent classification models\nhas received much attention, but there are few developments for achieving\ntransparent Reinforcement Learning (RL) models. In this study we propose a\nmethod that enables a RL agent to explain its behavior in terms of the expected\nconsequences of state transitions and outcomes. First, we define a translation\nof states and actions to a description that is easier to understand for human\nusers. Second, we developed a procedure that enables the agent to obtain the\nconsequences of a single action, as well as its entire policy. The method\ncalculates contrasts between the consequences of a policy derived from a user\nquery, and of the learned policy of the agent. Third, a format for generating\nexplanations was constructed. A pilot survey study was conducted to explore\npreferences of users for different explanation properties. Results indicate\nthat human users tend to favor explanations about policy rather than about\nsingle actions.']","['We present VisualHints, a novel environment for multimodal reinforcement\nlearning (RL) involving text-based interactions along with visual hints\n(obtained from the environment). Real-life problems often demand that agents\ninteract with the environment using both natural language information and\nvisual perception towards solving a goal. However, most traditional RL\nenvironments either solve pure vision-based tasks like Atari games or\nvideo-based robotic manipulation; or entirely use natural language as a mode of\ninteraction, like Text-based games and dialog systems. In this work, we aim to\nbridge this gap and unify these two approaches in a single environment for\nmultimodal RL. We introduce an extension of the TextWorld cooking environment\nwith the addition of visual clues interspersed throughout the environment. The\ngoal is to force an RL agent to use both text and visual features to predict\nnatural language action commands for solving the final task of cooking a meal.\nWe enable variations and difficulties in our environment to emulate various\ninteractive real-world scenarios. We present a baseline multimodal agent for\nsolving such problems using CNN-based feature extraction from visual hints and\nLSTMs for textual feature extraction. We believe that our proposed\nvisual-lingual environment will facilitate novel problem settings for the RL\ncommunity.', ""Efficiently adapting to new environments and changes in dynamics is critical\nfor agents to successfully operate in the real world. Reinforcement learning\n(RL) based approaches typically rely on external reward feedback for\nadaptation. However, in many scenarios this reward signal might not be readily\navailable for the target task, or the difference between the environments can\nbe implicit and only observable from the dynamics. To this end, we introduce a\nmethod that allows for self-adaptation of learned policies: No-Reward Meta\nLearning (NoRML). NoRML extends Model Agnostic Meta Learning (MAML) for RL and\nuses observable dynamics of the environment instead of an explicit reward\nfunction in MAML's finetune step. Our method has a more expressive update step\nthan MAML, while maintaining MAML's gradient based foundation. Additionally, in\norder to allow more targeted exploration, we implement an extension to MAML\nthat effectively disconnects the meta-policy parameters from the fine-tuned\npolicies' parameters. We first study our method on a number of synthetic\ncontrol problems and then validate our method on common benchmark environments,\nshowing that NoRML outperforms MAML when the dynamics change between tasks."", ""Being able to transfer existing skills to new situations is a key capability\nwhen training robots to operate in unpredictable real-world environments. A\nsuccessful transfer algorithm should not only minimize the number of samples\nthat the robot needs to collect in the new environment, but also prevent the\nrobot from damaging itself or the surrounding environment during the transfer\nprocess. In this work, we introduce a policy transfer algorithm for adapting\nrobot motor skills to novel scenarios while minimizing serious failures. Our\nalgorithm trains two control policies in the training environment: a task\npolicy that is optimized to complete the task of interest, and a protective\npolicy that is dedicated to keep the robot from unsafe events (e.g. falling to\nthe ground). To decide which policy to use during execution, we learn a safety\nestimator model in the training environment that estimates a continuous safety\nlevel of the robot. When used with a set of thresholds, the safety estimator\nbecomes a classifier for switching between the protective policy and the task\npolicy. We evaluate our approach on four simulated robot locomotion problems\nand a 2D navigation problem and show that our method can achieve successful\ntransfer to notably different environments while taking the robot's safety into\nconsideration."", ""A key challenge in reinforcement learning (RL) is environment generalization:\na policy trained to solve a task in one environment often fails to solve the\nsame task in a slightly different test environment. A common approach to\nimprove inter-environment transfer is to learn policies that are invariant to\nthe distribution of testing environments. However, we argue that instead of\nbeing invariant, the policy should identify the specific nuances of an\nenvironment and exploit them to achieve better performance. In this work, we\npropose the 'Environment-Probing' Interaction (EPI) policy, a policy that\nprobes a new environment to extract an implicit understanding of that\nenvironment's behavior. Once this environment-specific information is obtained,\nit is used as an additional input to a task-specific policy that can now\nperform environment-conditioned actions to solve a task. To learn these\nEPI-policies, we present a reward function based on transition predictability.\nSpecifically, a higher reward is given if the trajectory generated by the\nEPI-policy can be used to better predict transitions. We experimentally show\nthat EPI-conditioned task-specific policies significantly outperform commonly\nused policy generalization methods on novel testing environments."", 'One practical challenge in reinforcement learning (RL) is how to make quick\nadaptations when faced with new environments. In this paper, we propose a\nprincipled framework for adaptive RL, called \\textit{AdaRL}, that adapts\nreliably and efficiently to changes across domains with a few samples from the\ntarget domain, even in partially observable environments. Specifically, we\nleverage a parsimonious graphical representation that characterizes structural\nrelationships over variables in the RL system. Such graphical representations\nprovide a compact way to encode what and where the changes across domains are,\nand furthermore inform us with a minimal set of changes that one has to\nconsider for the purpose of policy adaptation. We show that by explicitly\nleveraging this compact representation to encode changes, we can efficiently\nadapt the policy to the target domain, in which only a few samples are needed\nand further policy optimization is avoided. We illustrate the efficacy of AdaRL\nthrough a series of experiments that vary factors in the observation,\ntransition, and reward functions for Cartpole and Atari games.', ""A wide range of reinforcement learning (RL) problems - including robustness,\ntransfer learning, unsupervised RL, and emergent complexity - require\nspecifying a distribution of tasks or environments in which a policy will be\ntrained. However, creating a useful distribution of environments is error\nprone, and takes a significant amount of developer time and effort. We propose\nUnsupervised Environment Design (UED) as an alternative paradigm, where\ndevelopers provide environments with unknown parameters, and these parameters\nare used to automatically produce a distribution over valid, solvable\nenvironments. Existing approaches to automatically generating environments\nsuffer from common failure modes: domain randomization cannot generate\nstructure or adapt the difficulty of the environment to the agent's learning\nprogress, and minimax adversarial training leads to worst-case environments\nthat are often unsolvable. To generate structured, solvable environments for\nour protagonist agent, we introduce a second, antagonist agent that is allied\nwith the environment-generating adversary. The adversary is motivated to\ngenerate environments which maximize regret, defined as the difference between\nthe protagonist and antagonist agent's return. We call our technique\nProtagonist Antagonist Induced Regret Environment Design (PAIRED). Our\nexperiments demonstrate that PAIRED produces a natural curriculum of\nincreasingly complex environments, and PAIRED agents achieve higher zero-shot\ntransfer performance when tested in highly novel environments."", 'To be successful in real-world tasks, Reinforcement Learning (RL) needs to\nexploit the compositional, relational, and hierarchical structure of the world,\nand learn to transfer it to the task at hand. Recent advances in representation\nlearning for language make it possible to build models that acquire world\nknowledge from text corpora and integrate this knowledge into downstream\ndecision making problems. We thus argue that the time is right to investigate a\ntight integration of natural language understanding into RL in particular. We\nsurvey the state of the field, including work on instruction following, text\ngames, and learning from textual domain knowledge. Finally, we call for the\ndevelopment of new environments as well as further investigation into the\npotential uses of recent Natural Language Processing (NLP) techniques for such\ntasks.', ""A grand goal in AI is to build a robot that can accurately navigate based on\nnatural language instructions, which requires the agent to perceive the scene,\nunderstand and ground language, and act in the real-world environment. One key\nchallenge here is to learn to navigate in new environments that are unseen\nduring training. Most of the existing approaches perform dramatically worse in\nunseen environments as compared to seen ones. In this paper, we present a\ngeneralizable navigational agent. Our agent is trained in two stages. The first\nstage is training via mixed imitation and reinforcement learning, combining the\nbenefits from both off-policy and on-policy optimization. The second stage is\nfine-tuning via newly-introduced 'unseen' triplets (environment, path,\ninstruction). To generate these unseen triplets, we propose a simple but\neffective 'environmental dropout' method to mimic unseen environments, which\novercomes the problem of limited seen environment variability. Next, we apply\nsemi-supervised learning (via back-translation) on these dropped-out\nenvironments to generate new paths and instructions. Empirically, we show that\nour agent is substantially better at generalizability when fine-tuned with\nthese triplets, outperforming the state-of-art approaches by a large margin on\nthe private unseen test set of the Room-to-Room task, and achieving the top\nrank on the leaderboard."", 'Despite ample motivation from costly exploration and limited trajectory data,\nrapidly adapting to new environments with few-shot reinforcement learning (RL)\ncan remain a challenging task, especially with respect to personalized\nsettings. Here, we consider the problem of recommending optimal policies to a\nset of multiple entities each with potentially different characteristics, such\nthat individual entities may parameterize distinct environments with unique\ntransition dynamics. Inspired by existing literature in meta-learning, we\nextend previous work by focusing on the notion that certain environments are\nmore similar to each other than others in personalized settings, and propose a\nmodel-free meta-learning algorithm that prioritizes past experiences by\nrelevance during gradient-based adaptation. Our algorithm involves\ncharacterizing past policy divergence through methods in inverse reinforcement\nlearning, and we illustrate how such metrics are able to effectively\ndistinguish past policy parameters by the environment they were deployed in,\nleading to more effective fast adaptation during test time. To study\npersonalization more effectively we introduce a navigation testbed to\nspecifically incorporate environment diversity across training episodes, and\ndemonstrate that our approach outperforms meta-learning alternatives with\nrespect to few-shot reinforcement learning in personalized settings.', 'Reinforcement learning would enjoy better success on real-world problems if\ndomain knowledge could be imparted to the algorithm by the modelers. Most\nproblems have both hidden state and unknown dynamics. Partially observable\nMarkov decision processes (POMDPs) allow for the modeling of both.\nUnfortunately, they do not provide a natural framework in which to specify\nknowledge about the domain dynamics. The designer must either admit to knowing\nnothing about the dynamics or completely specify the dynamics (thereby turning\nit into a planning problem). We propose a new framework called a partially\nknown Markov decision process (PKMDP) which allows the designer to specify\nknown dynamics while still leaving portions of the environment s dynamics\nunknown.The model represents NOT ONLY the environment dynamics but also the\nagents knowledge of the dynamics. We present a reinforcement learning algorithm\nfor this model based on importance sampling. The algorithm incorporates\nplanning based on the known dynamics and learning about the unknown dynamics.\nOur results clearly demonstrate the ability to add domain knowledge and the\nresulting benefits for learning.', 'Standard RL algorithms assume fixed environment dynamics and require a\nsignificant amount of interaction to adapt to new environments. We introduce\nPolicy-Dynamics Value Functions (PD-VF), a novel approach for rapidly adapting\nto dynamics different from those previously seen in training. PD-VF explicitly\nestimates the cumulative reward in a space of policies and environments. An\nensemble of conventional RL policies is used to gather experience on training\nenvironments, from which embeddings of both policies and environments can be\nlearned. Then, a value function conditioned on both embeddings is trained. At\ntest time, a few actions are sufficient to infer the environment embedding,\nenabling a policy to be selected by maximizing the learned value function\n(which requires no additional environment interaction). We show that our method\ncan rapidly adapt to new dynamics on a set of MuJoCo domains. Code available at\nhttps://github.com/rraileanu/policy-dynamics-value-functions.', 'A central capability of a long-lived reinforcement learning (RL) agent is to\nincrementally adapt its behavior as its environment changes, and to\nincrementally build upon previous experiences to facilitate future learning in\nreal-world scenarios. In this paper, we propose LifeLong Incremental\nReinforcement Learning (LLIRL), a new incremental algorithm for efficient\nlifelong adaptation to dynamic environments. We develop and maintain a library\nthat contains an infinite mixture of parameterized environment models, which is\nequivalent to clustering environment parameters in a latent space. The prior\ndistribution over the mixture is formulated as a Chinese restaurant process\n(CRP), which incrementally instantiates new environment models without any\nexternal information to signal environmental changes in advance. During\nlifelong learning, we employ the expectation maximization (EM) algorithm with\nonline Bayesian inference to update the mixture in a fully incremental manner.\nIn EM, the E-step involves estimating the posterior expectation of\nenvironment-to-cluster assignments, while the M-step updates the environment\nparameters for future learning. This method allows for all environment models\nto be adapted as necessary, with new models instantiated for environmental\nchanges and old models retrieved when previously seen environments are\nencountered again. Experiments demonstrate that LLIRL outperforms relevant\nexisting methods, and enables effective incremental adaptation to various\ndynamic environments for lifelong learning.', ""Reinforcement learning (RL), particularly in sparse reward settings, often\nrequires prohibitively large numbers of interactions with the environment,\nthereby limiting its applicability to complex problems. To address this,\nseveral prior approaches have used natural language to guide the agent's\nexploration. However, these approaches typically operate on structured\nrepresentations of the environment, and/or assume some structure in the natural\nlanguage commands. In this work, we propose a model that directly maps pixels\nto rewards, given a free-form natural language description of the task, which\ncan then be used for policy learning. Our experiments on the Meta-World robot\nmanipulation domain show that language-based rewards significantly improves the\nsample efficiency of policy learning, both in sparse and dense reward settings."", 'In this paper we target the problem of transferring policies across multiple\nenvironments with different dynamics parameters and motor noise variations, by\nintroducing a framework that decouples the processes of policy learning and\nsystem identification. Efficiently transferring learned policies to an unknown\nenvironment with changes in dynamics configurations in the presence of motor\nnoise is very important for operating robots in the real world, and our work is\na novel attempt in that direction. We introduce MANGA: Method Agnostic\nNeural-policy Generalization and Adaptation, that trains dynamics conditioned\npolicies and efficiently learns to estimate the dynamics parameters of the\nenvironment given off-policy state-transition rollouts in the environment. Our\nscheme is agnostic to the type of training method used - both reinforcement\nlearning (RL) and imitation learning (IL) strategies can be used. We\ndemonstrate the effectiveness of our approach by experimenting with four\ndifferent MuJoCo agents and comparing against previously proposed transfer\nbaselines.', 'Reinforcement Learning methods are capable of solving complex problems, but\nresulting policies might perform poorly in environments that are even slightly\ndifferent. In robotics especially, training and deployment conditions often\nvary and data collection is expensive, making retraining undesirable.\nSimulation training allows for feasible training times, but on the other hand\nsuffers from a reality-gap when applied in real-world settings. This raises the\nneed of efficient adaptation of policies acting in new environments. We\nconsider this as a problem of transferring knowledge within a family of similar\nMarkov decision processes.\n  For this purpose we assume that Q-functions are generated by some\nlow-dimensional latent variable. Given such a Q-function, we can find a master\npolicy that can adapt given different values of this latent variable. Our\nmethod learns both the generative mapping and an approximate posterior of the\nlatent variables, enabling identification of policies for new tasks by\nsearching only in the latent space, rather than the space of all policies. The\nlow-dimensional space, and master policy found by our method enables policies\nto quickly adapt to new environments. We demonstrate the method on both a\npendulum swing-up task in simulation, and for simulation-to-real transfer on a\npushing task.', 'This paper presents a comprehensive literature review on applications of deep\nreinforcement learning in communications and networking. Modern networks, e.g.,\nInternet of Things (IoT) and Unmanned Aerial Vehicle (UAV) networks, become\nmore decentralized and autonomous. In such networks, network entities need to\nmake decisions locally to maximize the network performance under uncertainty of\nnetwork environment. Reinforcement learning has been efficiently used to enable\nthe network entities to obtain the optimal policy including, e.g., decisions or\nactions, given their states when the state and action spaces are small.\nHowever, in complex and large-scale networks, the state and action spaces are\nusually large, and the reinforcement learning may not be able to find the\noptimal policy in reasonable time. Therefore, deep reinforcement learning, a\ncombination of reinforcement learning with deep learning, has been developed to\novercome the shortcomings. In this survey, we first give a tutorial of deep\nreinforcement learning from fundamental concepts to advanced models. Then, we\nreview deep reinforcement learning approaches proposed to address emerging\nissues in communications and networking. The issues include dynamic network\naccess, data rate control, wireless caching, data offloading, network security,\nand connectivity preservation which are all important to next generation\nnetworks such as 5G and beyond. Furthermore, we present applications of deep\nreinforcement learning for traffic routing, resource sharing, and data\ncollection. Finally, we highlight important challenges, open issues, and future\nresearch directions of applying deep reinforcement learning.', 'Reinforcement learning (RL) is a technique to learn the control policy for an\nagent that interacts with a stochastic environment. In any given state, the\nagent takes some action, and the environment determines the probability\ndistribution over the next state as well as gives the agent some reward. Most\nRL algorithms typically assume that the environment satisfies Markov\nassumptions (i.e. the probability distribution over the next state depends only\non the current state). In this paper, we propose a model-based RL technique for\na system that has non-Markovian dynamics. Such environments are common in many\nreal-world applications such as in human physiology, biological systems,\nmaterial science, and population dynamics. Model-based RL (MBRL) techniques\ntypically try to simultaneously learn a model of the environment from the data,\nas well as try to identify an optimal policy for the learned model. We propose\na technique where the non-Markovianity of the system is modeled through a\nfractional dynamical system. We show that we can quantify the difference in the\nperformance of an MBRL algorithm that uses bounded horizon model predictive\ncontrol from the optimal policy. Finally, we demonstrate our proposed framework\non a pharmacokinetic model of human blood glucose dynamics and show that our\nfractional models can capture distant correlations on real-world datasets.', 'Recently, text world games have been proposed to enable artificial agents to\nunderstand and reason about real-world scenarios. These text-based games are\nchallenging for artificial agents, as it requires an understanding of and\ninteraction using natural language in a partially observable environment.\nAgents observe the environment via textual descriptions designed to be\nchallenging enough for even human players. Past approaches have not paid enough\nattention to the language understanding capability of the proposed agents.\nTypically, these approaches train from scratch, an agent that learns both\ntextual representations and the gameplay online during training using a\ntemporal loss function. Given the sample-inefficiency of RL approaches, it is\ninefficient to learn rich enough textual representations to be able to\nunderstand and reason using the textual observation in such a complicated game\nenvironment setting. In this paper, we improve the semantic understanding of\nthe agent by proposing a simple RL with LM framework where we use\ntransformer-based language models with Deep RL models. We perform a detailed\nstudy of our framework to demonstrate how our model outperforms all existing\nagents on the popular game, Zork1, to achieve a score of 44.7, which is 1.6\nhigher than the state-of-the-art model. Overall, our proposed approach\noutperforms 4 games out of the 14 text-based games, while performing comparable\nto the state-of-the-art models on the remaining games.', 'Complex, multi-task problems have proven to be difficult to solve efficiently\nin a sparse-reward reinforcement learning setting. In order to be sample\nefficient, multi-task learning requires reuse and sharing of low-level\npolicies. To facilitate the automatic decomposition of hierarchical tasks, we\npropose the use of step-by-step human demonstrations in the form of natural\nlanguage instructions and action trajectories. We introduce a dataset of such\ndemonstrations in a crafting-based grid world. Our model consists of a\nhigh-level language generator and low-level policy, conditioned on language. We\nfind that human demonstrations help solve the most complex tasks. We also find\nthat incorporating natural language allows the model to generalize to unseen\ntasks in a zero-shot setting and to learn quickly from a few demonstrations.\nGeneralization is not only reflected in the actions of the agent, but also in\nthe generated natural language instructions in unseen tasks. Our approach also\ngives our trained agent interpretable behaviors because it is able to generate\na sequence of high-level descriptions of its actions.', 'Reinforcement learning (RL) algorithms typically start tabula rasa, without\nany prior knowledge of the environment, and without any prior skills. This\nhowever often leads to low sample efficiency, requiring a large amount of\ninteraction with the environment. This is especially true in a lifelong\nlearning setting, in which the agent needs to continually extend its\ncapabilities. In this paper, we examine how a pre-trained task-independent\nlanguage model can make a goal-conditional RL agent more sample efficient. We\ndo this by facilitating transfer learning between different related tasks. We\nexperimentally demonstrate our approach on a set of object navigation tasks.', 'This paper introduces a novel architecture for reinforcement learning with\ndeep neural networks designed to handle state and action spaces characterized\nby natural language, as found in text-based games. Termed a deep reinforcement\nrelevance network (DRRN), the architecture represents action and state spaces\nwith separate embedding vectors, which are combined with an interaction\nfunction to approximate the Q-function in reinforcement learning. We evaluate\nthe DRRN on two popular text games, showing superior performance over other\ndeep Q-learning architectures. Experiments with paraphrased action descriptions\nshow that the model is extracting meaning rather than simply memorizing strings\nof text.', 'Recent work has described neural-network-based agents that are trained with\nreinforcement learning (RL) to execute language-like commands in simulated\nworlds, as a step towards an intelligent agent or robot that can be instructed\nby human users. However, the optimisation of multi-goal motor policies via deep\nRL from scratch requires many episodes of experience. Consequently,\ninstruction-following with deep RL typically involves language generated from\ntemplates (by an environment simulator), which does not reflect the varied or\nambiguous expressions of real users. Here, we propose a conceptually simple\nmethod for training instruction-following agents with deep RL that are robust\nto natural human instructions. By applying our method with a state-of-the-art\npre-trained text-based language model (BERT), on tasks requiring agents to\nidentify and position everyday objects relative to other objects in a\nnaturalistic 3D simulated room, we demonstrate substantially-above-chance\nzero-shot transfer from synthetic template commands to natural instructions\ngiven by humans. Our approach is a general recipe for training any deep\nRL-based system to interface with human users, and bridges the gap between two\nresearch directions of notable recent success: agent-centric motor behavior and\ntext-based representation learning.', ""Reinforcement learning (RL) has been used in a range of simulated real-world\ntasks, e.g., sensor coordination, traffic light control, and on-demand mobility\nservices. However, real world deployments are rare, as RL struggles with\ndynamic nature of real world environments, requiring time for learning a task\nand adapting to changes in the environment. Transfer Learning (TL) can help\nlower these adaptation times. In particular, there is a significant potential\nof applying TL in multi-agent RL systems, where multiple agents can share\nknowledge with each other, as well as with new agents that join the system. To\nobtain the most from inter-agent transfer, transfer roles (i.e., determining\nwhich agents act as sources and which as targets), as well as relevant transfer\ncontent parameters (e.g., transfer size) should be selected dynamically in each\nparticular situation. As a first step towards fully dynamic transfers, in this\npaper we investigate the impact of TL transfer parameters with fixed source and\ntarget roles. Specifically, we label every agent-environment interaction with\nagent's epistemic confidence, and we filter the shared examples using varying\nthreshold levels and sample sizes. We investigate impact of these parameters in\ntwo scenarios, a standard predator-prey RL benchmark and a simulation of a\nride-sharing system with 200 vehicle agents and 10,000 ride-requests."", 'The capability of a reinforcement learning (RL) agent heavily depends on the\ndiversity of the learning scenarios generated by the environment. Generation of\ndiverse realistic scenarios is challenging for real-time strategy (RTS)\nenvironments. The RTS environments are characterized by intelligent\nentities/non-RL agents cooperating and competing with the RL agents with large\nstate and action spaces over a long period of time, resulting in an infinite\nspace of feasible, but not necessarily realistic, scenarios involving complex\ninteraction among different RL and non-RL agents. Yet, most of the existing\nsimulators rely on randomly generating the environments based on predefined\nsettings/layouts and offer limited flexibility and control over the environment\ndynamics for researchers to generate diverse, realistic scenarios as per their\ndemand. To address this issue, for the first time, we formally introduce the\nbenefits of adopting an existing formal scenario specification language,\nSCENIC, to assist researchers to model and generate diverse scenarios in an RTS\nenvironment in a flexible, systematic, and programmatic manner. To showcase the\nbenefits, we interfaced SCENIC to an existing RTS environment Google Research\nFootball(GRF) simulator and introduced a benchmark consisting of 32 realistic\nscenarios, encoded in SCENIC, to train RL agents and testing their\ngeneralization capabilities. We also show how researchers/RL practitioners can\nincorporate their domain knowledge to expedite the training process by\nintuitively modeling stochastic programmatic policies with SCENIC.', 'Robots that are trained to perform a task in a fixed environment often fail\nwhen facing unexpected changes to the environment due to a lack of exploration.\nWe propose a principled way to adapt the policy for better exploration in\nchanging sparse-reward environments. Unlike previous works which explicitly\nmodel environmental changes, we analyze the relationship between the value\nfunction and the optimal exploration for a Gaussian-parameterized policy and\nshow that our theory leads to an effective strategy for adjusting the variance\nof the policy, enabling fast adapt to changes in a variety of sparse-reward\nenvironments.', 'Deep reinforcement learning has over the past few years shown great potential\nin learning near-optimal control in complex simulated environments with little\nvisible information. Rainbow (Q-Learning) and PPO (Policy Optimisation) have\nshown outstanding performance in a variety of tasks, including Atari 2600,\nMuJoCo, and Roboschool test suite. While these algorithms are fundamentally\ndifferent, both suffer from high variance, low sample efficiency, and\nhyperparameter sensitivity that in practice, make these algorithms a no-go for\ncritical operations in the industry.\n  On the other hand, model-based reinforcement learning focuses on learning the\ntransition dynamics between states in an environment. If these environment\ndynamics are adequately learned, a model-based approach is perhaps the most\nsample efficient method for learning agents to act in an environment optimally.\nThe traits of model-based reinforcement are ideal for real-world environments\nwhere sampling is slow and for mission-critical operations. In the warehouse\nindustry, there is an increasing motivation to minimise time and to maximise\nproduction. Currently, autonomous agents act suboptimally using handcrafted\npolicies for significant portions of the state-space.\n  In this paper, we present The Dreaming Variational Autoencoder v2 (DVAE-2), a\nmodel-based reinforcement learning algorithm that increases sample efficiency,\nhence enable algorithms with low sample efficiency function better in\nreal-world environments. We introduce Deep Warehouse, a simulated environment\nfor industry-near testing of autonomous agents in grid-based warehouses.\nFinally, we illustrate that DVAE-2 improves the sample efficiency for the Deep\nWarehouse compared to model-free methods.', 'Many real-world systems problems require reasoning about the long term\nconsequences of actions taken to configure and manage the system. These\nproblems with delayed and often sequentially aggregated reward, are often\ninherently reinforcement learning problems and present the opportunity to\nleverage the recent substantial advances in deep reinforcement learning.\nHowever, in some cases, it is not clear why deep reinforcement learning is a\ngood fit for the problem. Sometimes, it does not perform better than the\nstate-of-the-art solutions. And in other cases, random search or greedy\nalgorithms could outperform deep reinforcement learning. In this paper, we\nreview, discuss, and evaluate the recent trends of using deep reinforcement\nlearning in system optimization. We propose a set of essential metrics to guide\nfuture works in evaluating the efficacy of using deep reinforcement learning in\nsystem optimization. Our evaluation includes challenges, the types of problems,\ntheir formulation in the deep reinforcement learning setting, embedding, the\nmodel used, efficiency, and robustness. We conclude with a discussion on open\nchallenges and potential directions for pushing further the integration of\nreinforcement learning in system optimization.', 'Autonomous Vehicles (AVs) are required to operate safely and efficiently in\ndynamic environments. For this, the AVs equipped with Joint\nRadar-Communications (JRC) functions can enhance the driving safety by\nutilizing both radar detection and data communication functions. However,\noptimizing the performance of the AV system with two different functions under\nuncertainty and dynamic of surrounding environments is very challenging. In\nthis work, we first propose an intelligent optimization framework based on the\nMarkov Decision Process (MDP) to help the AV make optimal decisions in\nselecting JRC operation functions under the dynamic and uncertainty of the\nsurrounding environment. We then develop an effective learning algorithm\nleveraging recent advances of deep reinforcement learning techniques to find\nthe optimal policy for the AV without requiring any prior information about\nsurrounding environment. Furthermore, to make our proposed framework more\nscalable, we develop a Transfer Learning (TL) mechanism that enables the AV to\nleverage valuable experiences for accelerating the training process when it\nmoves to a new environment. Extensive simulations show that the proposed\ntransferable deep reinforcement learning framework reduces the obstacle miss\ndetection probability by the AV up to 67% compared to other conventional deep\nreinforcement learning approaches.', 'Explicit engineering of reward functions for given environments has been a\nmajor hindrance to reinforcement learning methods. While Inverse Reinforcement\nLearning (IRL) is a solution to recover reward functions from demonstrations\nonly, these learned rewards are generally heavily \\textit{entangled} with the\ndynamics of the environment and therefore not portable or \\emph{robust} to\nchanging environments. Modern adversarial methods have yielded some success in\nreducing reward entanglement in the IRL setting. In this work, we leverage one\nsuch method, Adversarial Inverse Reinforcement Learning (AIRL), to propose an\nalgorithm that learns hierarchical disentangled rewards with a policy over\noptions. We show that this method has the ability to learn \\emph{generalizable}\npolicies and reward functions in complex transfer learning tasks, while\nyielding results in continuous control benchmarks that are comparable to those\nof the state-of-the-art methods.', 'In order for robots to perform mission-critical tasks, it is essential that\nthey are able to quickly adapt to changes in their environment as well as to\ninjuries and or other bodily changes. Deep reinforcement learning has been\nshown to be successful in training robot control policies for operation in\ncomplex environments. However, existing methods typically employ only a single\npolicy. This can limit the adaptability since a large environmental\nmodification might require a completely different behavior compared to the\nlearning environment. To solve this problem, we propose Map-based Multi-Policy\nReinforcement Learning (MMPRL), which aims to search and store multiple\npolicies that encode different behavioral features while maximizing the\nexpected reward in advance of the environment change. Thanks to these policies,\nwhich are stored into a multi-dimensional discrete map according to its\nbehavioral feature, adaptation can be performed within reasonable time without\nretraining the robot. An appropriate pre-trained policy from the map can be\nrecalled using Bayesian optimization. Our experiments show that MMPRL enables\nrobots to quickly adapt to large changes without requiring any prior knowledge\non the type of injuries that could occur. A highlight of the learned behaviors\ncan be found here: https://youtu.be/QwInbilXNOE .', ""One of the central challenges faced by a reinforcement learning (RL) agent is\nto effectively learn a (near-)optimal policy in environments with large state\nspaces having sparse and noisy feedback signals. In real-world applications, an\nexpert with additional domain knowledge can help in speeding up the learning\nprocess via \\emph{shaping the environment}, i.e., making the environment more\nlearner-friendly. A popular paradigm in literature is \\emph{potential-based\nreward shaping}, where the environment's reward function is augmented with\nadditional local rewards using a potential function. However, the applicability\nof potential-based reward shaping is limited in settings where (i) the state\nspace is very large, and it is challenging to compute an appropriate potential\nfunction, (ii) the feedback signals are noisy, and even with shaped rewards the\nagent could be trapped in local optima, and (iii) changing the rewards alone is\nnot sufficient, and effective shaping requires changing the dynamics. We\naddress these limitations of potential-based shaping methods and propose a\nnovel framework of \\emph{environment shaping using state abstraction}. Our key\nidea is to compress the environment's large state space with noisy signals to\nan abstracted space, and to use this abstraction in creating smoother and more\neffective feedback signals for the agent. We study the theoretical\nunderpinnings of our abstraction-based environment shaping, and show that the\nagent's policy learnt in the shaped environment preserves near-optimal behavior\nin the original environment."", 'In this paper, we present a new approach to Transfer Learning (TL) in\nReinforcement Learning (RL) for cross-domain tasks. Many of the available\ntechniques approach the transfer architecture as a method of speeding up the\ntarget task learning. We propose to adapt and reuse the mapped source task\noptimal-policy directly in related domains. We show the optimal policy from a\nrelated source task can be near optimal in target domain provided an adaptive\npolicy accounts for the model error between target and source. The main benefit\nof this policy augmentation is generalizing policies across multiple related\ndomains without having to re-learn the new tasks. Our results show that this\narchitecture leads to better sample efficiency in the transfer, reducing sample\ncomplexity of target task learning to target apprentice learning.', 'Existing work in language grounding typically study single environments. How\ndo we build unified models that apply across multiple environments? We propose\nthe multi-environment Symbolic Interactive Language Grounding benchmark (SILG),\nwhich unifies a collection of diverse grounded language learning environments\nunder a common interface. SILG consists of grid-world environments that require\ngeneralization to new dynamics, entities, and partially observed worlds (RTFM,\nMessenger, NetHack), as well as symbolic counterparts of visual worlds that\nrequire interpreting rich natural language with respect to complex scenes\n(ALFWorld, Touchdown). Together, these environments provide diverse grounding\nchallenges in richness of observation space, action space, language\nspecification, and plan complexity. In addition, we propose the first shared\nmodel architecture for RL on these environments, and evaluate recent advances\nsuch as egocentric local convolution, recurrent state-tracking, entity-centric\nattention, and pretrained LM using SILG. Our shared architecture achieves\ncomparable performance to environment-specific architectures. Moreover, we find\nthat many recent modelling advances do not result in significant gains on\nenvironments other than the one they were designed for. This highlights the\nneed for a multi-environment benchmark. Finally, the best models significantly\nunderperform humans on SILG, which suggests ample room for future work. We hope\nSILG enables the community to quickly identify new methodologies for language\ngrounding that generalize to a diverse set of environments and their associated\nchallenges.', 'We consider problems of making sequences of decisions to accomplish tasks,\ninteracting via the medium of language. These problems are often tackled with\nreinforcement learning approaches. We find that these models do not generalize\nwell when applied to novel task domains. However, the large amount of\ncomputation necessary to adequately train and explore the search space of\nsequential decision making, under a reinforcement learning paradigm, precludes\nthe inclusion of large contextualized language models, which might otherwise\nenable the desired generalization ability. We introduce a teacher-student\nimitation learning methodology and a means of converting a reinforcement\nlearning model into a natural language understanding model. Together, these\nmethodologies enable the introduction of contextualized language models into\nthe sequential decision making problem space. We show that models can learn\nfaster and generalize more, leveraging both the imitation learning and the\nreformulation. Our models exceed teacher performance on various held-out\ndecision problems, by up to 7% on in-domain problems and 24% on out-of-domain\nproblems.', 'Recent advances in reinforcement learning have shown its potential to tackle\ncomplex real-life tasks. However, as the dimensionality of the task increases,\nreinforcement learning methods tend to struggle. To overcome this, we explore\nmethods for representing the semantic information embedded in the state. While\nprevious methods focused on information in its raw form (e.g., raw visual\ninput), we propose to represent the state using natural language. Language can\nrepresent complex scenarios and concepts, making it a favorable candidate for\nrepresentation. Empirical evidence, within the domain of ViZDoom, suggests that\nnatural language based agents are more robust, converge faster and perform\nbetter than vision based agents, showing the benefit of using natural language\nrepresentations for reinforcement learning.', 'Reinforcement learning provides a powerful and general framework for decision\nmaking and control, but its application in practice is often hindered by the\nneed for extensive feature and reward engineering. Deep reinforcement learning\nmethods can remove the need for explicit engineering of policy or value\nfeatures, but still require a manually specified reward function. Inverse\nreinforcement learning holds the promise of automatic reward acquisition, but\nhas proven exceptionally difficult to apply to large, high-dimensional problems\nwith unknown dynamics. In this work, we propose adverserial inverse\nreinforcement learning (AIRL), a practical and scalable inverse reinforcement\nlearning algorithm based on an adversarial reward learning formulation. We\ndemonstrate that AIRL is able to recover reward functions that are robust to\nchanges in dynamics, enabling us to learn policies even under significant\nvariation in the environment seen during training. Our experiments show that\nAIRL greatly outperforms prior methods in these transfer settings.', 'The objective of a reinforcement learning agent is to behave so as to\nmaximise the sum of a suitable scalar function of state: the reward. These\nrewards are typically given and immutable. In this paper, we instead consider\nthe proposition that the reward function itself can be a good locus of learned\nknowledge. To investigate this, we propose a scalable meta-gradient framework\nfor learning useful intrinsic reward functions across multiple lifetimes of\nexperience. Through several proof-of-concept experiments, we show that it is\nfeasible to learn and capture knowledge about long-term exploration and\nexploitation into a reward function. Furthermore, we show that unlike policy\ntransfer methods that capture ""how"" the agent should behave, the learned reward\nfunctions can generalise to other kinds of agents and to changes in the\ndynamics of the environment by capturing ""what"" the agent should strive to do.', 'In this work we apply deep reinforcement learning to the problems of\nnavigating a three-dimensional environment and inferring the locations of human\nspeaker audio sources within, in the case where the only available information\nis the raw sound from the environment, as a simulated human listener placed in\nthe environment would hear it. For this purpose we create two virtual\nenvironments using the Unity game engine, one presenting an audio-based\nnavigation problem and one presenting an audio source localization problem. We\nalso create an autonomous agent based on PPO online reinforcement learning\nalgorithm and attempt to train it to solve these environments. Our experiments\nshow that our agent achieves adequate performance and generalization ability in\nboth environments, measured by quantitative metrics, even when a limited amount\nof training data are available or the environment parameters shift in ways not\nencountered during training. We also show that a degree of agent knowledge\ntransfer is possible between the environments.', ""To further improve the learning efficiency and performance of reinforcement\nlearning (RL), in this paper we propose a novel uncertainty-aware model-based\nRL (UA-MBRL) framework, and then implement and validate it in autonomous\ndriving under various task scenarios. First, an action-conditioned ensemble\nmodel with the ability of uncertainty assessment is established as the virtual\nenvironment model. Then, a novel uncertainty-aware model-based RL framework is\ndeveloped based on the adaptive truncation approach, providing virtual\ninteractions between the agent and environment model, and improving RL's\ntraining efficiency and performance. The developed algorithms are then\nimplemented in end-to-end autonomous vehicle control tasks, validated and\ncompared with state-of-the-art methods under various driving scenarios. The\nvalidation results suggest that the proposed UA-MBRL method surpasses the\nexisting model-based and model-free RL approaches, in terms of learning\nefficiency and achieved performance. The results also demonstrate the good\nability of the proposed method with respect to the adaptiveness and robustness,\nunder various autonomous driving scenarios."", ""In this paper, we consider the source of Deep Reinforcement Learning (DRL)'s\nsample complexity, asking how much derives from the requirement of learning\nuseful representations of environment states and how much is due to the sample\ncomplexity of learning a policy. While for DRL agents, the distinction between\nrepresentation and policy may not be clear, we seek new insight through a set\nof transfer learning experiments. In each experiment, we retain some fraction\nof layers trained on either the same game or a related game, comparing the\nbenefits of transfer learning to learning a policy from scratch. Interestingly,\nwe find that benefits due to transfer are highly variable in general and\nnon-symmetric across pairs of tasks. Our experiments suggest that perhaps\ntransfer from simpler environments can boost performance on more complex\ndownstream tasks and that the requirements of learning a useful representation\ncan range from negligible to the majority of the sample complexity, based on\nthe environment. Furthermore, we find that fine-tuning generally outperforms\ntraining with the transferred layers frozen, confirming an insight first noted\nin the classification setting."", 'Reinforcement learning from large-scale offline datasets provides us with the\nability to learn policies without potentially unsafe or impractical\nexploration. Significant progress has been made in the past few years in\ndealing with the challenge of correcting for differing behavior between the\ndata collection and learned policies. However, little attention has been paid\nto potentially changing dynamics when transferring a policy to the online\nsetting, where performance can be up to 90% reduced for existing methods. In\nthis paper we address this problem with Augmented World Models (AugWM). We\naugment a learned dynamics model with simple transformations that seek to\ncapture potential changes in physical properties of the robot, leading to more\nrobust policies. We not only train our policy in this new setting, but also\nprovide it with the sampled augmentation as a context, allowing it to adapt to\nchanges in the environment. At test time we learn the context in a\nself-supervised fashion by approximating the augmentation which corresponds to\nthe new environment. We rigorously evaluate our approach on over 100 different\nchanged dynamics settings, and show that this simple approach can significantly\nimprove the zero-shot generalization of a recent state-of-the-art baseline,\noften achieving successful policies where the baseline fails.', 'Sample-efficient generalisation of reinforcement learning approaches have\nalways been a challenge, especially, for complex scenes with many components.\nIn this work, we introduce Plug and Play Markov Decision Processes, an\nobject-based representation that allows zero-shot integration of new objects\nfrom known object classes. This is achieved by representing the global\ntransition dynamics as a union of local transition functions, each with respect\nto one active object in the scene. Transition dynamics from an object class can\nbe pre-learnt and thus would be ready to use in a new environment. Each active\nobject is also endowed with its reward function. Since there is no central\nreward function, addition or removal of objects can be handled efficiently by\nonly updating the reward functions of objects involved. A new transfer learning\nmechanism is also proposed to adapt reward function in such cases. Experiments\nshow that our representation can achieve sample-efficiency in a variety of\nset-ups.', 'In recent years some researchers have explored the use of reinforcement\nlearning (RL) algorithms as key components in the solution of various natural\nlanguage processing tasks. For instance, some of these algorithms leveraging\ndeep neural learning have found their way into conversational systems. This\npaper reviews the state of the art of RL methods for their possible use for\ndifferent problems of natural language processing, focusing primarily on\nconversational systems, mainly due to their growing relevance. We provide\ndetailed descriptions of the problems as well as discussions of why RL is\nwell-suited to solve them. Also, we analyze the advantages and limitations of\nthese methods. Finally, we elaborate on promising research directions in\nnatural language processing that might benefit from reinforcement learning.', 'A suitable state representation is a fundamental part of the learning process\nin Reinforcement Learning. In various tasks, the state can either be described\nby natural language or be natural language itself. This survey outlines the\nstrategies used in the literature to build natural language state\nrepresentations. We appeal for more linguistically interpretable and grounded\nrepresentations, careful justification of design decisions and evaluation of\nthe effectiveness of different approaches.', 'Humans generally use natural language to communicate task requirements to\neach other. Ideally, natural language should also be usable for communicating\ngoals to autonomous machines (e.g., robots) to minimize friction in task\nspecification. However, understanding and mapping natural language goals to\nsequences of states and actions is challenging. Specifically, existing work\nalong these lines has encountered difficulty in generalizing learned policies\nto new natural language goals and environments. In this paper, we propose a\nnovel adversarial inverse reinforcement learning algorithm to learn a\nlanguage-conditioned policy and reward function. To improve generalization of\nthe learned policy and reward function, we use a variational goal generator to\nrelabel trajectories and sample diverse goals during training. Our algorithm\noutperforms multiple baselines by a large margin on a vision-based natural\nlanguage instruction following dataset (Room-2-Room), demonstrating a promising\nadvance in enabling the use of natural language instructions in specifying\nagent goals.', 'Recent work has shown that deep reinforcement-learning agents can learn to\nfollow language-like instructions from infrequent environment rewards. However,\nthis places on environment designers the onus of designing language-conditional\nreward functions which may not be easily or tractably implemented as the\ncomplexity of the environment and the language scales. To overcome this\nlimitation, we present a framework within which instruction-conditional RL\nagents are trained using rewards obtained not from the environment, but from\nreward models which are jointly trained from expert examples. As reward models\nimprove, they learn to accurately reward agents for completing tasks for\nenvironment configurations---and for instructions---not present amongst the\nexpert data. This framework effectively separates the representation of what\ninstructions require from how they can be executed. In a simple grid world, it\nenables an agent to learn a range of commands requiring interaction with blocks\nand understanding of spatial relations and underspecified abstract\narrangements. We further show the method allows our agent to adapt to changes\nin the environment without requiring new expert examples.', 'The ability to transfer in reinforcement learning is key towards building an\nagent of general artificial intelligence. In this paper, we consider the\nproblem of learning to simultaneously transfer across both environments (ENV)\nand tasks (TASK), probably more importantly, by learning from only sparse (ENV,\nTASK) pairs out of all the possible combinations. We propose a novel\ncompositional neural network architecture which depicts a meta rule for\ncomposing policies from the environment and task embeddings. Notably, one of\nthe main challenges is to learn the embeddings jointly with the meta rule. We\nfurther propose new training methods to disentangle the embeddings, making them\nboth distinctive signatures of the environments and tasks and effective\nbuilding blocks for composing the policies. Experiments on GridWorld and Thor,\nof which the agent takes as input an egocentric view, show that our approach\ngives rise to high success rates on all the (ENV, TASK) pairs after learning\nfrom only 40% of them.', 'Reinforcement learning (RL) can automate a wide variety of robotic skills,\nbut learning each new skill requires considerable real-world data collection\nand manual representation engineering to design policy classes or features.\nUsing deep reinforcement learning to train general purpose neural network\npolicies alleviates some of the burden of manual representation engineering by\nusing expressive policy classes, but exacerbates the challenge of data\ncollection, since such methods tend to be less efficient than RL with\nlow-dimensional, hand-designed representations. Transfer learning can mitigate\nthis problem by enabling us to transfer information from one skill to another\nand even from one robot to another. We show that neural network policies can be\ndecomposed into ""task-specific"" and ""robot-specific"" modules, where the\ntask-specific modules are shared across robots, and the robot-specific modules\nare shared across all tasks on that robot. This allows for sharing task\ninformation, such as perception, between robots and sharing robot information,\nsuch as dynamics and kinematics, between tasks. We exploit this decomposition\nto train mix-and-match modules that can solve new robot-task combinations that\nwere not seen during training. Using a novel neural network architecture, we\ndemonstrate the effectiveness of our transfer method for enabling zero-shot\ngeneralization with a variety of robots and tasks in simulation for both visual\nand non-visual tasks.', 'Learning policies for complex tasks that require multiple different skills is\na major challenge in reinforcement learning (RL). It is also a requirement for\nits deployment in real-world scenarios. This paper proposes a novel framework\nfor efficient multi-task reinforcement learning. Our framework trains agents to\nemploy hierarchical policies that decide when to use a previously learned\npolicy and when to learn a new skill. This enables agents to continually\nacquire new skills during different stages of training. Each learned task\ncorresponds to a human language description. Because agents can only access\npreviously learned skills through these descriptions, the agent can always\nprovide a human-interpretable description of its choices. In order to help the\nagent learn the complex temporal dependencies necessary for the hierarchical\npolicy, we provide it with a stochastic temporal grammar that modulates when to\nrely on previously learned skills and when to execute new skills. We validate\nour approach on Minecraft games designed to explicitly test the ability to\nreuse previously learned skills while simultaneously learning new skills.', 'In most real world scenarios, a policy trained by reinforcement learning in\none environment needs to be deployed in another, potentially quite different\nenvironment. However, generalization across different environments is known to\nbe hard. A natural solution would be to keep training after deployment in the\nnew environment, but this cannot be done if the new environment offers no\nreward signal. Our work explores the use of self-supervision to allow the\npolicy to continue training after deployment without using any rewards. While\nprevious methods explicitly anticipate changes in the new environment, we\nassume no prior knowledge of those changes yet still obtain significant\nimprovements. Empirical evaluations are performed on diverse simulation\nenvironments from DeepMind Control suite and ViZDoom, as well as real robotic\nmanipulation tasks in continuously changing environments, taking observations\nfrom an uncalibrated camera. Our method improves generalization in 31 out of 36\nenvironments across various tasks and outperforms domain randomization on a\nmajority of environments.', 'Robots need to be able to work in multiple different environments. Even when\nperforming similar tasks, different behaviour should be deployed to best fit\nthe current environment. In this paper, We propose a new approach to\nnavigation, where it is treated as a multi-task learning problem. This enables\nthe robot to learn to behave differently in visual navigation tasks for\ndifferent environments while also learning shared expertise across\nenvironments. We evaluated our approach in both simulated environments as well\nas real-world data. Our method allows our system to converge with a 26%\nreduction in training time, while also increasing accuracy.', 'Traditional model-based RL relies on hand-specified or learned models of\ntransition dynamics of the environment. These methods are sample efficient and\nfacilitate learning in the real world but fail to generalize to subtle\nvariations in the underlying dynamics, e.g., due to differences in mass,\nfriction, or actuators across robotic agents or across time. We propose using\nvariational inference to learn an explicit latent representation of unknown\nenvironment properties that accelerates learning and facilitates generalization\non novel environments at test time. We use Online Bayesian Inference of these\nlearned latents to rapidly adapt online to changes in environments without\nretaining large replay buffers of recent data. Combined with a neural network\nensemble that models dynamics and captures uncertainty over dynamics, our\napproach demonstrates positive transfer during training and online adaptation\non the continuous control task HalfCheetah.', 'In typical reinforcement learning (RL), the environment is assumed given and\nthe goal of the learning is to identify an optimal policy for the agent taking\nactions through its interactions with the environment. In this paper, we extend\nthis setting by considering the environment is not given, but controllable and\nlearnable through its interaction with the agent at the same time. This\nextension is motivated by environment design scenarios in the real-world,\nincluding game design, shopping space design and traffic signal design.\nTheoretically, we find a dual Markov decision process (MDP) w.r.t. the\nenvironment to that w.r.t. the agent, and derive a policy gradient solution to\noptimizing the parametrized environment. Furthermore, discontinuous\nenvironments are addressed by a proposed general generative framework. Our\nexperiments on a Maze game design task show the effectiveness of the proposed\nalgorithms in generating diverse and challenging Mazes against various agent\nsettings.', 'Reinforcement learning is a learning paradigm for solving sequential\ndecision-making problems. Recent years have witnessed remarkable progress in\nreinforcement learning upon the fast development of deep neural networks. Along\nwith the promising prospects of reinforcement learning in numerous domains such\nas robotics and game-playing, transfer learning has arisen to tackle various\nchallenges faced by reinforcement learning, by transferring knowledge from\nexternal expertise to facilitate the efficiency and effectiveness of the\nlearning process. In this survey, we systematically investigate the recent\nprogress of transfer learning approaches in the context of deep reinforcement\nlearning. Specifically, we provide a framework for categorizing the\nstate-of-the-art transfer learning approaches, under which we analyze their\ngoals, methodologies, compatible reinforcement learning backbones, and\npractical applications. We also draw connections between transfer learning and\nother relevant topics from the reinforcement learning perspective and explore\ntheir potential challenges that await future research progress.', 'Agents trained with deep reinforcement learning algorithms are capable of\nperforming highly complex tasks including locomotion in continuous\nenvironments. We investigate transferring the learning acquired in one task to\na set of previously unseen tasks. Generalization and overfitting in deep\nreinforcement learning are not commonly addressed in current transfer learning\nresearch. Conducting a comparative analysis without an intermediate\nregularization step results in underperforming benchmarks and inaccurate\nalgorithm comparisons due to rudimentary assessments. In this study, we propose\nregularization techniques in deep reinforcement learning for continuous control\nthrough the application of sample elimination, early stopping and maximum\nentropy regularized adversarial learning. First, the importance of the\ninclusion of training iteration number to the hyperparameters in deep transfer\nreinforcement learning will be discussed. Because source task performance is\nnot indicative of the generalization capacity of the algorithm, we start by\nacknowledging the training iteration number as a hyperparameter. In line with\nthis, we introduce an additional step of resorting to earlier snapshots of\npolicy parameters to prevent overfitting to the source task. Then, to generate\nrobust policies, we discard the samples that lead to overfitting via a method\nwe call strict clipping. Furthermore, we increase the generalization capacity\nin widely used transfer learning benchmarks by using maximum entropy\nregularization, different critic methods, and curriculum learning in an\nadversarial setup. Subsequently, we propose maximum entropy adversarial\nreinforcement learning to increase the domain randomization. Finally, we\nevaluate the robustness of these methods on simulated robots in target\nenvironments where the morphology of the robot, gravity, and tangential\nfriction coefficient of the environment are altered.', 'Over its lifetime, a reinforcement learning agent is often tasked with\ndifferent tasks. How to efficiently adapt a previously learned control policy\nfrom one task to another, remains an open research question. In this paper, we\ninvestigate how instructions formulated in natural language can enable faster\nand more effective task adaptation. This can serve as the basis for developing\nlanguage instructed skills, which can be used in a lifelong learning setting.\nOur method is capable of assessing, given a set of developed base control\npolicies, which policy will adapt best to a new unseen task.', 'While reinforcement learning (RL) has been successful in natural language\nprocessing (NLP) domains such as dialogue generation and text-based games, it\ntypically faces the problem of sparse rewards that leads to slow or no\nconvergence. Traditional methods that use text descriptions to extract only a\nstate representation ignore the feedback inherently present in them. In\ntext-based games, for example, descriptions like ""Good Job! You ate the food}""\nindicate progress, and descriptions like ""You entered a new room"" indicate\nexploration. Positive and negative cues like these can be converted to rewards\nthrough sentiment analysis. This technique converts the sparse reward problem\ninto a dense one, which is easier to solve. Furthermore, this can enable\nreinforcement learning without rewards, in which the agent learns entirely from\nthese intrinsic sentiment rewards. This framework is similar to intrinsic\nmotivation, where the environment does not necessarily provide the rewards, but\nthe agent analyzes and realizes them by itself. We find that providing dense\nrewards in text-based games using sentiment analysis improves performance under\nsome conditions.', 'Agents trained by reinforcement learning (RL) often fail to generalize beyond\nthe environment they were trained in, even when presented with new scenarios\nthat seem similar to the training environment. We study the query complexity\nrequired to train RL agents that generalize to multiple environments.\nIntuitively, tractable generalization is only possible when the environments\nare similar or close in some sense. To capture this, we introduce Weak\nProximity, a natural structural condition that requires the environments to\nhave highly similar transition and reward functions and share a policy\nproviding optimal value. Despite such shared structure, we prove that tractable\ngeneralization is impossible in the worst case. This holds even when each\nindividual environment can be efficiently solved to obtain an optimal linear\npolicy, and when the agent possesses a generative model. Our lower bound\napplies to the more complex task of representation learning for the purpose of\nefficient generalization to multiple environments. On the positive side, we\nintroduce Strong Proximity, a strengthened condition which we prove is\nsufficient for efficient generalization.', 'Reinforcement learning is an active research area with a vast number of\napplications in robotics, and the RoboCup competition is an interesting\nenvironment for studying and evaluating reinforcement learning methods. A known\ndifficulty in applying reinforcement learning to robotics is the high number of\nexperience samples required, being the use of simulated environments for\ntraining the agents followed by transfer learning to real-world (sim-to-real) a\nviable path. This article introduces an open-source simulator for the IEEE Very\nSmall Size Soccer and the Small Size League optimized for reinforcement\nlearning experiments. We also propose a framework for creating OpenAI Gym\nenvironments with a set of benchmarks tasks for evaluating single-agent and\nmulti-agent robot soccer skills. We then demonstrate the learning capabilities\nof two state-of-the-art reinforcement learning methods as well as their\nlimitations in certain scenarios introduced in this framework. We believe this\nwill make it easier for more teams to compete in these categories using\nend-to-end reinforcement learning approaches and further develop this research\narea.', 'We propose a novel approach to address one aspect of the non-stationarity\nproblem in multi-agent reinforcement learning (RL), where the other agents may\nalter their policies due to environment changes during execution. This violates\nthe Markov assumption that governs most single-agent RL methods and is one of\nthe key challenges in multi-agent RL. To tackle this, we propose to train\nmultiple policies for each agent and postpone the selection of the best policy\nat execution time. Specifically, we model the environment non-stationarity with\na finite set of scenarios and train policies fitting each scenario. In addition\nto multiple policies, each agent also learns a policy predictor to determine\nwhich policy is the best with its local information. By doing so, each agent is\nable to adapt its policy when the environment changes and consequentially the\nother agents alter their policies during execution. We empirically evaluated\nour method on a variety of common benchmark problems proposed for multi-agent\ndeep RL in the literature. Our experimental results show that the agents\ntrained by our algorithm have better adaptiveness in changing environments and\noutperform the state-of-the-art methods in all the tested environments.', 'Deep reinforcement learning augments the reinforcement learning framework and\nutilizes the powerful representation of deep neural networks. Recent works have\ndemonstrated the remarkable successes of deep reinforcement learning in various\ndomains including finance, medicine, healthcare, video games, robotics, and\ncomputer vision. In this work, we provide a detailed review of recent and\nstate-of-the-art research advances of deep reinforcement learning in computer\nvision. We start with comprehending the theories of deep learning,\nreinforcement learning, and deep reinforcement learning. We then propose a\ncategorization of deep reinforcement learning methodologies and discuss their\nadvantages and limitations. In particular, we divide deep reinforcement\nlearning into seven main categories according to their applications in computer\nvision, i.e. (i)landmark localization (ii) object detection; (iii) object\ntracking; (iv) registration on both 2D image and 3D image volumetric data (v)\nimage segmentation; (vi) videos analysis; and (vii) other applications. Each of\nthese categories is further analyzed with reinforcement learning techniques,\nnetwork design, and performance. Moreover, we provide a comprehensive analysis\nof the existing publicly available datasets and examine source code\navailability. Finally, we present some open issues and discuss future research\ndirections on deep reinforcement learning in computer vision', 'This paper provides a roadmap that explores the question of how to imbue\nlearning agents with the ability to understand and generate contextually\nrelevant natural language in service of achieving a goal. We hypothesize that\ntwo key components in creating such agents are interactivity and environment\ngrounding, shown to be vital parts of language learning in humans, and posit\nthat interactive narratives should be the environments of choice for such\ntraining these agents. These games are simulations in which an agent interacts\nwith the world through natural language -- ""perceiving"", ""acting upon"", and\n""talking to"" the world using textual descriptions, commands, and dialogue --\nand as such exist at the intersection of natural language processing,\nstorytelling, and sequential decision making. We discuss the unique challenges\na text games\' puzzle-like structure combined with natural language\nstate-and-action spaces provides: knowledge representation, commonsense\nreasoning, and exploration. Beyond the challenges described so far, progress in\nthe realm of interactive narratives can be applied in adjacent problem domains.\nThese applications provide interesting challenges of their own as well as\nextensions to those discussed so far. We describe three of them in detail: (1)\nevaluating AI system\'s commonsense understanding by automatically creating\ninteractive narratives; (2) adapting abstract text-based policies to include\nother modalities such as vision; and (3) enabling multi-agent and human-AI\ncollaboration in shared, situated worlds.', 'Transfer in reinforcement learning refers to the notion that generalization\nshould occur not only within a task but also across tasks. We propose a\ntransfer framework for the scenario where the reward function changes between\ntasks but the environment\'s dynamics remain the same. Our approach rests on two\nkey ideas: ""successor features"", a value function representation that decouples\nthe dynamics of the environment from the rewards, and ""generalized policy\nimprovement"", a generalization of dynamic programming\'s policy improvement\noperation that considers a set of policies rather than a single one. Put\ntogether, the two ideas lead to an approach that integrates seamlessly within\nthe reinforcement learning framework and allows the free exchange of\ninformation across tasks. The proposed method also provides performance\nguarantees for the transferred policy even before any learning has taken place.\nWe derive two theorems that set our approach in firm theoretical ground and\npresent experiments that show that it successfully promotes transfer in\npractice, significantly outperforming alternative methods in a sequence of\nnavigation tasks and in the control of a simulated robotic arm.', ""We seek to align agent behavior with a user's objectives in a reinforcement\nlearning setting with unknown dynamics, an unknown reward function, and unknown\nunsafe states. The user knows the rewards and unsafe states, but querying the\nuser is expensive. To address this challenge, we propose an algorithm that\nsafely and interactively learns a model of the user's reward function. We start\nwith a generative model of initial states and a forward dynamics model trained\non off-policy data. Our method uses these models to synthesize hypothetical\nbehaviors, asks the user to label the behaviors with rewards, and trains a\nneural network to predict the rewards. The key idea is to actively synthesize\nthe hypothetical behaviors from scratch by maximizing tractable proxies for the\nvalue of information, without interacting with the environment. We call this\nmethod reward query synthesis via trajectory optimization (ReQueST). We\nevaluate ReQueST with simulated users on a state-based 2D navigation task and\nthe image-based Car Racing video game. The results show that ReQueST\nsignificantly outperforms prior methods in learning reward models that transfer\nto new environments with different initial state distributions. Moreover,\nReQueST safely trains the reward model to detect unsafe states, and corrects\nreward hacking before deploying the agent."", 'In this paper, we consider the task of learning control policies for\ntext-based games. In these games, all interactions in the virtual world are\nthrough text and the underlying state is not observed. The resulting language\nbarrier makes such environments challenging for automatic game players. We\nemploy a deep reinforcement learning framework to jointly learn state\nrepresentations and action policies using game rewards as feedback. This\nframework enables us to map text descriptions into vector representations that\ncapture the semantics of the game states. We evaluate our approach on two game\nworlds, comparing against baselines using bag-of-words and bag-of-bigrams for\nstate representations. Our algorithm outperforms the baselines on both worlds\ndemonstrating the importance of learning expressive representations.', 'How can we train a dialog model to produce better conversations by learning\nfrom human feedback, without the risk of humans teaching it harmful chat\nbehaviors? We start by hosting models online, and gather human feedback from\nreal-time, open-ended conversations, which we then use to train and improve the\nmodels using offline reinforcement learning (RL). We identify implicit\nconversational cues including language similarity, elicitation of laughter,\nsentiment, and more, which indicate positive human feedback, and embed these in\nmultiple reward functions. A well-known challenge is that learning an RL policy\nin an offline setting usually fails due to the lack of ability to explore and\nthe tendency to make over-optimistic estimates of future reward. These problems\nbecome even harder when using RL for language models, which can easily have a\n20,000 action vocabulary and many possible reward functions. We solve the\nchallenge by developing a novel class of offline RL algorithms. These\nalgorithms use KL-control to penalize divergence from a pre-trained prior\nlanguage model, and use a new strategy to make the algorithm pessimistic,\ninstead of optimistic, in the face of uncertainty. We test the resulting dialog\nmodel with ratings from 80 users in an open-domain setting and find it achieves\nsignificant improvements over existing deep offline RL approaches. The novel\noffline RL method is viable for improving any existing generative dialog model\nusing a static dataset of human feedback.', 'This paper gives a detailed review of reinforcement learning in combinatorial\noptimization, introduces the history of combinatorial optimization starting in\nthe 1960s, and compares it with the reinforcement learning algorithms in recent\nyears. We explicitly look at a famous combinatorial problem known as the\nTraveling Salesperson Problem (TSP). We compare the approach of the modern\nreinforcement learning algorithms on TSP with an approach published in 1970.\nThen, we discuss the similarities between these algorithms and how the approach\nof reinforcement learning changes due to the evolution of machine learning\ntechniques and computing power. We also mention the deep learning approach on\nthe TSP, which is named Deep Reinforcement Learning. We argue that deep\nlearning is a generic approach that can be integrated with traditional\nreinforcement learning algorithms and optimize the outcomes of the TSP.', 'We consider the challenge of policy simplification and verification in the\ncontext of policies learned through reinforcement learning (RL) in continuous\nenvironments. In well-behaved settings, RL algorithms have convergence\nguarantees in the limit. While these guarantees are valuable, they are\ninsufficient for safety-critical applications. Furthermore, they are lost when\napplying advanced techniques such as deep-RL. To recover guarantees when\napplying advanced RL algorithms to more complex environments with (i)\nreachability, (ii) safety-constrained reachability, or (iii) discounted-reward\nobjectives, we build upon the DeepMDP framework introduced by Gelada et al. to\nderive new bisimulation bounds between the unknown environment and a learned\ndiscrete latent model of it. Our bisimulation bounds enable the application of\nformal methods for Markov decision processes. Finally, we show how one can use\na policy obtained via state-of-the-art RL to efficiently train a variational\nautoencoder that yields a discrete latent model with provably approximately\ncorrect bisimulation guarantees. Additionally, we obtain a distilled version of\nthe policy for the latent model.', 'To achieve general artificial intelligence, reinforcement learning (RL)\nagents should learn not only to optimize returns for one specific task but also\nto constantly build more complex skills and scaffold their knowledge about the\nworld, without forgetting what has already been learned. In this paper, we\ndiscuss the desired characteristics of environments that can support the\ntraining and evaluation of lifelong reinforcement learning agents, review\nexisting environments from this perspective, and propose recommendations for\ndevising suitable environments in the future.', 'The goal of reinforcement learning (RL) is to let an agent learn an optimal\ncontrol policy in an unknown environment so that future expected rewards are\nmaximized. The model-free RL approach directly learns the policy based on data\nsamples. Although using many samples tends to improve the accuracy of policy\nlearning, collecting a large number of samples is often expensive in practice.\nOn the other hand, the model-based RL approach first estimates the transition\nmodel of the environment and then learns the policy based on the estimated\ntransition model. Thus, if the transition model is accurately learned from a\nsmall amount of data, the model-based approach can perform better than the\nmodel-free approach. In this paper, we propose a novel model-based RL method by\ncombining a recently proposed model-free policy search method called policy\ngradients with parameter-based exploration and the state-of-the-art transition\nmodel estimator called least-squares conditional density estimation. Through\nexperiments, we demonstrate the practical usefulness of the proposed method.', ""It is a long-standing challenge to enable an intelligent agent to learn in\none environment and generalize to an unseen environment without further data\ncollection and finetuning. In this paper, we consider a zero shot\ngeneralization problem setup that complies with biological intelligent agents'\nlearning and generalization processes. The agent is first presented with\nprevious experiences in the training environment, along with task description\nin the form of trajectory-level sparse rewards. Later when it is placed in the\nnew testing environment, it is asked to perform the task without any\ninteraction with the testing environment. We find this setting natural for\nbiological creatures and at the same time, challenging for previous methods.\nBehavior cloning, state-of-art RL along with other zero-shot learning methods\nperform poorly on this benchmark. Given a set of experiences in the training\nenvironment, our method learns a neural function that decomposes the sparse\nreward into particular regions in a contingency-aware observation as a per step\nreward. Based on such decomposed rewards, we further learn a dynamics model and\nuse Model Predictive Control (MPC) to obtain a policy. Since the rewards are\ndecomposed to finer-granularity observations, they are naturally generalizable\nto new environments that are composed of similar basic elements. We demonstrate\nour method on a wide range of environments, including a classic video game --\nSuper Mario Bros, as well as a robotic continuous control task. Please refer to\nthe project page for more visualized results."", 'Human intelligence has the remarkable ability to adapt to new tasks and\nenvironments quickly. Starting from a very young age, humans acquire new skills\nand learn how to solve new tasks either by imitating the behavior of others or\nby following provided natural language instructions. To facilitate research in\nthis direction, we propose IGLU: Interactive Grounded Language Understanding in\na Collaborative Environment. The primary goal of the competition is to approach\nthe problem of how to build interactive agents that learn to solve a task while\nprovided with grounded natural language instructions in a collaborative\nenvironment. Understanding the complexity of the challenge, we split it into\nsub-tasks to make it feasible for participants.\n  This research challenge is naturally related, but not limited, to two fields\nof study that are highly relevant to the NeurIPS community: Natural Language\nUnderstanding and Generation (NLU/G) and Reinforcement Learning (RL).\nTherefore, the suggested challenge can bring two communities together to\napproach one of the important challenges in AI. Another important aspect of the\nchallenge is the dedication to perform a human-in-the-loop evaluation as a\nfinal evaluation for the agents developed by contestants.', 'Language creates a compact representation of the world and allows the\ndescription of unlimited situations and objectives through compositionality.\nWhile these characterizations may foster instructing, conditioning or\nstructuring interactive agent behavior, it remains an open-problem to correctly\nrelate language understanding and reinforcement learning in even simple\ninstruction following scenarios. This joint learning problem is alleviated\nthrough expert demonstrations, auxiliary losses, or neural inductive biases. In\nthis paper, we propose an orthogonal approach called Hindsight Generation for\nExperience Replay (HIGhER) that extends the Hindsight Experience Replay (HER)\napproach to the language-conditioned policy setting. Whenever the agent does\nnot fulfill its instruction, HIGhER learns to output a new directive that\nmatches the agent trajectory, and it relabels the episode with a positive\nreward. To do so, HIGhER learns to map a state into an instruction by using\npast successful trajectories, which removes the need to have external expert\ninterventions to relabel episodes as in vanilla HER. We show the efficiency of\nour approach in the BabyAI environment, and demonstrate how it complements\nother instruction following methods.', 'Natural language is perhaps the most flexible and intuitive way for humans to\ncommunicate tasks to a robot. Prior work in imitation learning typically\nrequires each task be specified with a task id or goal image -- something that\nis often impractical in open-world environments. On the other hand, previous\napproaches in instruction following allow agent behavior to be guided by\nlanguage, but typically assume structure in the observations, actuators, or\nlanguage that limit their applicability to complex settings like robotics. In\nthis work, we present a method for incorporating free-form natural language\nconditioning into imitation learning. Our approach learns perception from\npixels, natural language understanding, and multitask continuous control\nend-to-end as a single neural network. Unlike prior work in imitation learning,\nour method is able to incorporate unlabeled and unstructured demonstration data\n(i.e. no task or language labels). We show this dramatically improves language\nconditioned performance, while reducing the cost of language annotation to less\nthan 1% of total data. At test time, a single language conditioned visuomotor\npolicy trained with our method can perform a wide variety of robotic\nmanipulation skills in a 3D environment, specified only with natural language\ndescriptions of each task (e.g. ""open the drawer...now pick up the block...now\npress the green button...""). To scale up the number of instructions an agent\ncan follow, we propose combining text conditioned policies with large\npretrained neural language models. We find this allows a policy to be robust to\nmany out-of-distribution synonym instructions, without requiring new\ndemonstrations. See videos of a human typing live text commands to our agent at\nlanguage-play.github.io', ""This paper studies the problem of autonomous exploration under localization\nuncertainty for a mobile robot with 3D range sensing. We present a framework\nfor self-learning a high-performance exploration policy in a single simulation\nenvironment, and transferring it to other environments, which may be physical\nor virtual. Recent work in transfer learning achieves encouraging performance\nby domain adaptation and domain randomization to expose an agent to scenarios\nthat fill the inherent gaps in sim2sim and sim2real approaches. However, it is\ninefficient to train an agent in environments with randomized conditions to\nlearn the important features of its current state. An agent can use domain\nknowledge provided by human experts to learn efficiently. We propose a novel\napproach that uses graph neural networks in conjunction with deep reinforcement\nlearning, enabling decision-making over graphs containing relevant exploration\ninformation provided by human experts to predict a robot's optimal sensing\naction in belief space. The policy, which is trained only in a single\nsimulation environment, offers a real-time, scalable, and transferable\ndecision-making strategy, resulting in zero-shot transfer to other simulation\nenvironments and even real-world environments."", 'Exploration is essential for reinforcement learning (RL). To face the\nchallenges of exploration, we consider a reward-free RL framework that\ncompletely separates exploration from exploitation and brings new challenges\nfor exploration algorithms. In the exploration phase, the agent learns an\nexploratory policy by interacting with a reward-free environment and collects a\ndataset of transitions by executing the policy. In the planning phase, the\nagent computes a good policy for any reward function based on the dataset\nwithout further interacting with the environment. This framework is suitable\nfor the meta RL setting where there are many reward functions of interest. In\nthe exploration phase, we propose to maximize the Renyi entropy over the\nstate-action space and justify this objective theoretically. The success of\nusing Renyi entropy as the objective results from its encouragement to explore\nthe hard-to-reach state-actions. We further deduce a policy gradient\nformulation for this objective and design a practical exploration algorithm\nthat can deal with complex environments. In the planning phase, we solve for\ngood policies given arbitrary reward functions using a batch RL algorithm.\nEmpirically, we show that our exploration algorithm is effective and sample\nefficient, and results in superior policies for arbitrary reward functions in\nthe planning phase.', ""The named concepts and compositional operators present in natural language\nprovide a rich source of information about the kinds of abstractions humans use\nto navigate the world. Can this linguistic background knowledge improve the\ngenerality and efficiency of learned classifiers and control policies? This\npaper aims to show that using the space of natural language strings as a\nparameter space is an effective way to capture natural task structure. In a\npretraining phase, we learn a language interpretation model that transforms\ninputs (e.g. images) into outputs (e.g. labels) given natural language\ndescriptions. To learn a new concept (e.g. a classifier), we search directly in\nthe space of descriptions to minimize the interpreter's loss on training\nexamples. Crucially, our models do not require language data to learn these\nconcepts: language is used only in pretraining to impose structure on\nsubsequent learning. Results on image classification, text editing, and\nreinforcement learning show that, in all settings, models with a linguistic\nparameterization outperform those without."", 'Deep reinforcement learning is revolutionizing the artificial intelligence\nfield. Currently, it serves as a good starting point for constructing\nintelligent autonomous systems which offer a better knowledge of the visual\nworld. It is possible to scale deep reinforcement learning with the use of deep\nlearning and do amazing tasks such as use of pixels in playing video games. In\nthis paper, key concepts of deep reinforcement learning including reward\nfunction, differences between reinforcement learning and supervised learning\nand models for implementation of reinforcement are discussed. Key challenges\nrelated to the implementation of reinforcement learning in conversational AI\ndomain are identified as well as discussed in detail. Various conversational\nmodels which are based on deep reinforcement learning (as well as deep\nlearning) are also discussed. In summary, this paper discusses key aspects of\ndeep reinforcement learning which are crucial for designing an efficient\nconversational AI.', 'Reinforcement learning has seen great advancements in the past five years.\nThe successful introduction of deep learning in place of more traditional\nmethods allowed reinforcement learning to scale to very complex domains\nachieving super-human performance in environments like the game of Go or\nnumerous video games. Despite great successes in multiple domains, these new\nmethods suffer from their own issues that make them often inapplicable to the\nreal world problems. Extreme lack of data efficiency, together with huge\nvariance and difficulty in enforcing safety constraints, is one of the three\nmost prominent issues in the field. Usually, millions of data points sampled\nfrom the environment are necessary for these algorithms to converge to\nacceptable policies.\n  This thesis proposes novel Generative Adversarial Imaginative Reinforcement\nLearning algorithm. It takes advantage of the recent introduction of highly\neffective generative adversarial models, and Markov property that underpins\nreinforcement learning setting, to model dynamics of the real environment\nwithin the internal imagination module. Rollouts from the imagination are then\nused to artificially simulate the real environment in a standard reinforcement\nlearning process to avoid, often expensive and dangerous, trial and error in\nthe real environment. Experimental results show that the proposed algorithm\nmore economically utilises experience from the real environment than the\ncurrent state-of-the-art Rainbow DQN algorithm, and thus makes an important\nstep towards sample efficient deep reinforcement learning.', 'Teaching an agent to navigate in an unseen 3D environment is a challenging\ntask, even in the event of simulated environments. To generalize to unseen\nenvironments, an agent needs to be robust to low-level variations (e.g. color,\ntexture, object changes), and also high-level variations (e.g. layout changes\nof the environment). To improve overall generalization, all types of variations\nin the environment have to be taken under consideration via different level of\ndata augmentation steps. To this end, we propose House3D, a rich, extensible\nand efficient environment that contains 45,622 human-designed 3D scenes of\nvisually realistic houses, ranging from single-room studios to multi-storied\nhouses, equipped with a diverse set of fully labeled 3D objects, textures and\nscene layouts, based on the SUNCG dataset (Song et.al.). The diversity in\nHouse3D opens the door towards scene-level augmentation, while the label-rich\nnature of House3D enables us to inject pixel- & task-level augmentations such\nas domain randomization (Toubin et. al.) and multi-task training. Using a\nsubset of houses in House3D, we show that reinforcement learning agents trained\nwith an enhancement of different levels of augmentations perform much better in\nunseen environments than our baselines with raw RGB input by over 8% in terms\nof navigation success rate. House3D is publicly available at\nhttp://github.com/facebookresearch/House3D.', ""Imitation learning and instruction-following are two common approaches to\ncommunicate a user's intent to a learning agent. However, as the complexity of\ntasks grows, it could be beneficial to use both demonstrations and language to\ncommunicate with an agent. In this work, we propose a novel setting where an\nagent is given both a demonstration and a description, and must combine\ninformation from both the modalities. Specifically, given a demonstration for a\ntask (the source task), and a natural language description of the differences\nbetween the demonstrated task and a related but different task (the target\ntask), our goal is to train an agent to complete the target task in a zero-shot\nsetting, that is, without any demonstrations for the target task. To this end,\nwe introduce Language-Aided Reward and Value Adaptation (LARVA) which, given a\nsource demonstration and a linguistic description of how the target task\ndiffers, learns to output a reward / value function that accurately describes\nthe target task. Our experiments show that on a diverse set of adaptations, our\napproach is able to complete more than 95% of target tasks when using\ntemplate-based descriptions, and more than 70% when using free-form natural\nlanguage."", 'Reward learning enables the application of reinforcement learning (RL) to\ntasks where reward is defined by human judgment, building a model of reward by\nasking humans questions. Most work on reward learning has used simulated\nenvironments, but complex information about values is often expressed in\nnatural language, and we believe reward learning for language is a key to\nmaking RL practical and safe for real-world tasks. In this paper, we build on\nadvances in generative pretraining of language models to apply reward learning\nto four natural language tasks: continuing text with positive sentiment or\nphysically descriptive language, and summarization tasks on the TL;DR and\nCNN/Daily Mail datasets. For stylistic continuation we achieve good results\nwith only 5,000 comparisons evaluated by humans. For summarization, models\ntrained with 60,000 comparisons copy whole sentences from the input but skip\nirrelevant preamble; this leads to reasonable ROUGE scores and very good\nperformance according to our human labelers, but may be exploiting the fact\nthat labelers rely on simple heuristics.', 'Several recent works have been dedicated to unsupervised reinforcement\nlearning in a single environment, in which a policy is first pre-trained with\nunsupervised interactions, and then fine-tuned towards the optimal policy for\nseveral downstream supervised tasks defined over the same environment. Along\nthis line, we address the problem of unsupervised reinforcement learning in a\nclass of multiple environments, in which the policy is pre-trained with\ninteractions from the whole class, and then fine-tuned for several tasks in any\nenvironment of the class. Notably, the problem is inherently multi-objective as\nwe can trade off the pre-training objective between environments in many ways.\nIn this work, we foster an exploration strategy that is sensitive to the most\nadverse cases within the class. Hence, we cast the exploration problem as the\nmaximization of the mean of a critical percentile of the state visitation\nentropy induced by the exploration strategy over the class of environments.\nThen, we present a policy gradient algorithm, $\\alpha$MEPOL, to optimize the\nintroduced objective through mediated interactions with the class. Finally, we\nempirically demonstrate the ability of the algorithm in learning to explore\nchallenging classes of continuous environments and we show that reinforcement\nlearning greatly benefits from the pre-trained exploration strategy w.r.t.\nlearning from scratch.', 'Reinforcement learning is one of the core components in designing an\nartificial intelligent system emphasizing real-time response. Reinforcement\nlearning influences the system to take actions within an arbitrary environment\neither having previous knowledge about the environment model or not. In this\npaper, we present a comprehensive study on Reinforcement Learning focusing on\nvarious dimensions including challenges, the recent development of different\nstate-of-the-art techniques, and future directions. The fundamental objective\nof this paper is to provide a framework for the presentation of available\nmethods of reinforcement learning that is informative enough and simple to\nfollow for the new researchers and academics in this domain considering the\nlatest concerns. First, we illustrated the core techniques of reinforcement\nlearning in an easily understandable and comparable way. Finally, we analyzed\nand depicted the recent developments in reinforcement learning approaches. My\nanalysis pointed out that most of the models focused on tuning policy values\nrather than tuning other things in a particular state of reasoning.', ""Due to its nature of learning from dynamic interactions and planning for\nlong-run performance, reinforcement learning (RL) recently has received much\nattention in interactive recommender systems (IRSs). IRSs usually face the\nlarge discrete action space problem, which makes most of the existing RL-based\nrecommendation methods inefficient. Moreover, data sparsity is another\nchallenging problem that most IRSs are confronted with. While the textual\ninformation like reviews and descriptions is less sensitive to sparsity,\nexisting RL-based recommendation methods either neglect or are not suitable for\nincorporating textual information. To address these two problems, in this\npaper, we propose a Text-based Deep Deterministic Policy Gradient framework\n(TDDPG-Rec) for IRSs. Specifically, we leverage textual information to map\nitems and users into a feature space, which greatly alleviates the sparsity\nproblem. Moreover, we design an effective method to construct an action\ncandidate set. By the policy vector dynamically learned from TDDPG-Rec that\nexpresses the user's preference, we can select actions from the candidate set\neffectively. Through experiments on three public datasets, we demonstrate that\nTDDPG-Rec achieves state-of-the-art performance over several baselines in a\ntime-efficient manner."", 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', 'Computer simulation provides an automatic and safe way for training robotic\ncontrol policies to achieve complex tasks such as locomotion. However, a policy\ntrained in simulation usually does not transfer directly to the real hardware\ndue to the differences between the two environments. Transfer learning using\ndomain randomization is a promising approach, but it usually assumes that the\ntarget environment is close to the distribution of the training environments,\nthus relying heavily on accurate system identification. In this paper, we\npresent a different approach that leverages domain randomization for\ntransferring control policies to unknown environments. The key idea that,\ninstead of learning a single policy in the simulation, we simultaneously learn\na family of policies that exhibit different behaviors. When tested in the\ntarget environment, we directly search for the best policy in the family based\non the task performance, without the need to identify the dynamic parameters.\nWe evaluate our method on five simulated robotic control problems with\ndifferent discrepancies in the training and testing environment and demonstrate\nthat our method can overcome larger modeling errors compared to training a\nrobust policy or an adaptive policy.', ""We consider the problem of imitation learning from a finite set of expert\ntrajectories, without access to reinforcement signals. The classical approach\nof extracting the expert's reward function via inverse reinforcement learning,\nfollowed by reinforcement learning is indirect and may be computationally\nexpensive. Recent generative adversarial methods based on matching the policy\ndistribution between the expert and the agent could be unstable during\ntraining. We propose a new framework for imitation learning by estimating the\nsupport of the expert policy to compute a fixed reward function, which allows\nus to re-frame imitation learning within the standard reinforcement learning\nsetting. We demonstrate the efficacy of our reward function on both discrete\nand continuous domains, achieving comparable or better performance than the\nstate of the art under different reinforcement learning algorithms."", 'The speed and accuracy with which robots are able to interpret natural\nlanguage is fundamental to realizing effective human-robot interaction. A great\ndeal of attention has been paid to developing models and approximate inference\nalgorithms that improve the efficiency of language understanding. However,\nexisting methods still attempt to reason over a representation of the\nenvironment that is flat and unnecessarily detailed, which limits scalability.\nAn open problem is then to develop methods capable of producing the most\ncompact environment model sufficient for accurate and efficient natural\nlanguage understanding. We propose a model that leverages environment-related\ninformation encoded within instructions to identify the subset of observations\nand perceptual classifiers necessary to perceive a succinct,\ninstruction-specific environment representation. The framework uses three\nprobabilistic graphical models trained from a corpus of annotated instructions\nto infer salient scene semantics, perceptual classifiers, and grounded symbols.\nExperimental results on two robots operating in different environments\ndemonstrate that by exploiting the content and the structure of the\ninstructions, our method learns compact environment representations that\nsignificantly improve the efficiency of natural language symbol grounding.', 'Reinforcement learning (RL) can enable task-oriented dialogue systems to\nsteer the conversation towards successful task completion. In an end-to-end\nsetting, a response can be constructed in a word-level sequential decision\nmaking process with the entire system vocabulary as action space. Policies\ntrained in such a fashion do not require expert-defined action spaces, but they\nhave to deal with large action spaces and long trajectories, making RL\nimpractical. Using the latent space of a variational model as action space\nalleviates this problem. However, current approaches use an uninformed prior\nfor training and optimize the latent distribution solely on the context. It is\ntherefore unclear whether the latent representation truly encodes the\ncharacteristics of different actions. In this paper, we explore three ways of\nleveraging an auxiliary task to shape the latent variable distribution: via\npre-training, to obtain an informed prior, and via multitask learning. We\nchoose response auto-encoding as the auxiliary task, as this captures the\ngenerative factors of dialogue responses while requiring low computational cost\nand neither additional data nor labels. Our approach yields a more\naction-characterized latent representations which support end-to-end dialogue\npolicy optimization and achieves state-of-the-art success rates. These results\nwarrant a more wide-spread use of RL in end-to-end dialogue models.', 'Reinforcement Learning (RL) is a powerful framework to address the\ndiscrepancy between loss functions used during training and the final\nevaluation metrics to be used at test time. When applied to neural Machine\nTranslation (MT), it minimises the mismatch between the cross-entropy loss and\nnon-differentiable evaluation metrics like BLEU. However, the suitability of\nthese metrics as reward function at training time is questionable: they tend to\nbe sparse and biased towards the specific words used in the reference texts. We\npropose to address this problem by making models less reliant on such metrics\nin two ways: (a) with an entropy-regularised RL method that does not only\nmaximise a reward function but also explore the action space to avoid peaky\ndistributions; (b) with a novel RL method that explores a dynamic unsupervised\nreward function to balance between exploration and exploitation. We base our\nproposals on the Soft Actor-Critic (SAC) framework, adapting the off-policy\nmaximum entropy model for language generation applications such as MT. We\ndemonstrate that SAC with BLEU reward tends to overfit less to the training\ndata and performs better on out-of-domain data. We also show that our dynamic\nunsupervised reward can lead to better translation of ambiguous words.', 'Most Knowledge Graphs (KGs) contain textual descriptions of entities in\nvarious natural languages. These descriptions of entities provide valuable\ninformation that may not be explicitly represented in the structured part of\nthe KG. Based on this fact, some link prediction methods which make use of the\ninformation presented in the textual descriptions of entities have been\nproposed to learn representations of (monolingual) KGs. However, these methods\nuse entity descriptions in only one language and ignore the fact that\ndescriptions given in different languages may provide complementary information\nand thereby also additional semantics. In this position paper, the problem of\neffectively leveraging multilingual entity descriptions for the purpose of link\nprediction in KGs will be discussed along with potential solutions to the\nproblem.', 'This work presents an exploration and imitation-learning-based agent capable\nof state-of-the-art performance in playing text-based computer games.\nText-based computer games describe their world to the player through natural\nlanguage and expect the player to interact with the game using text. These\ngames are of interest as they can be seen as a testbed for language\nunderstanding, problem-solving, and language generation by artificial agents.\nMoreover, they provide a learning environment in which these skills can be\nacquired through interactions with an environment rather than using fixed\ncorpora. One aspect that makes these games particularly challenging for\nlearning agents is the combinatorially large action space. Existing methods for\nsolving text-based games are limited to games that are either very simple or\nhave an action space restricted to a predetermined set of admissible actions.\nIn this work, we propose to use the exploration approach of Go-Explore for\nsolving text-based games. More specifically, in an initial exploration phase,\nwe first extract trajectories with high rewards, after which we train a policy\nto solve the game by imitating these trajectories. Our experiments show that\nthis approach outperforms existing solutions in solving text-based games, and\nit is more sample efficient in terms of the number of interactions with the\nenvironment. Moreover, we show that the learned policy can generalize better\nthan existing solutions to unseen games without using any restriction on the\naction space.', 'Reinforcement learning has recently experienced increased prominence in the\nmachine learning community. There are many approaches to solving reinforcement\nlearning problems with new techniques developed constantly. When solving\nproblems using reinforcement learning, there are various difficult challenges\nto overcome. To ensure progress in the field, benchmarks are important for\ntesting new algorithms and comparing with other approaches. The reproducibility\nof results for fair comparison is therefore vital in ensuring that improvements\nare accurately judged. This paper provides an overview of different\ncontributions to reinforcement learning benchmarking and discusses how they can\nassist researchers to address the challenges facing reinforcement learning. The\ncontributions discussed are the most used and recent in the literature. The\npaper discusses the contributions in terms of implementation, tasks and\nprovided algorithm implementations with benchmarks. The survey aims to bring\nattention to the wide range of reinforcement learning benchmarking tasks\navailable and to encourage research to take place in a standardised manner.\nAdditionally, this survey acts as an overview for researchers not familiar with\nthe different tasks that can be used to develop and test new reinforcement\nlearning algorithms.', 'Deep reinforcement learning is poised to revolutionise the field of AI and\nrepresents a step towards building autonomous systems with a higher level\nunderstanding of the visual world. Currently, deep learning is enabling\nreinforcement learning to scale to problems that were previously intractable,\nsuch as learning to play video games directly from pixels. Deep reinforcement\nlearning algorithms are also applied to robotics, allowing control policies for\nrobots to be learned directly from camera inputs in the real world. In this\nsurvey, we begin with an introduction to the general field of reinforcement\nlearning, then progress to the main streams of value-based and policy-based\nmethods. Our survey will cover central algorithms in deep reinforcement\nlearning, including the deep $Q$-network, trust region policy optimisation, and\nasynchronous advantage actor-critic. In parallel, we highlight the unique\nadvantages of deep neural networks, focusing on visual understanding via\nreinforcement learning. To conclude, we describe several current areas of\nresearch within the field.', 'Text Worlds are virtual environments for embodied agents that, unlike 2D or\n3D environments, are rendered exclusively using textual descriptions. These\nenvironments offer an alternative to higher-fidelity 3D environments due to\ntheir low barrier to entry, providing the ability to study semantics,\ncompositional inference, and other high-level tasks with rich high-level action\nspaces while controlling for perceptual input. This systematic survey outlines\nrecent developments in tooling, environments, and agent modeling for Text\nWorlds, while examining recent trends in knowledge graphs, common sense\nreasoning, transfer learning of Text World performance to higher-fidelity\nenvironments, as well as near-term development targets that, once achieved,\nmake Text Worlds an attractive general research paradigm for natural language\nprocessing.', ""Several approaches have recently been proposed for learning decentralized\ndeep multiagent policies that coordinate via a differentiable communication\nchannel. While these policies are effective for many tasks, interpretation of\ntheir induced communication strategies has remained a challenge. Here we\npropose to interpret agents' messages by translating them. Unlike in typical\nmachine translation problems, we have no parallel data to learn from. Instead\nwe develop a translation model based on the insight that agent messages and\nnatural language strings mean the same thing if they induce the same belief\nabout the world in a listener. We present theoretical guarantees and empirical\nevidence that our approach preserves both the semantics and pragmatics of\nmessages by ensuring that players communicating through a translation layer do\nnot suffer a substantial loss in reward relative to players with a common\nlanguage."", 'Recent work has shown that recurrent neural networks can be trained to\nseparate individual speakers in a sound mixture with high fidelity. Here we\nexplore convolutional neural network models as an alternative and show that\nthey achieve state-of-the-art results with an order of magnitude fewer\nparameters. We also characterize and compare the robustness and ability of\nthese different approaches to generalize under three different test conditions:\nlonger time sequences, the addition of intermittent noise, and different\ndatasets not seen during training. For the last condition, we create a new\ndataset, RealTalkLibri, to test source separation in real-world environments.\nWe show that the acoustics of the environment have significant impact on the\nstructure of the waveform and the overall performance of neural network models,\nwith the convolutional model showing superior ability to generalize to new\nenvironments. The code for our study is available at\nhttps://github.com/ShariqM/source_separation.', ""Human infants are able to acquire natural language seemingly easily at an\nearly age. Their language learning seems to occur simultaneously with learning\nother cognitive functions as well as with playful interactions with the\nenvironment and caregivers. From a neuroscientific perspective, natural\nlanguage is embodied, grounded in most, if not all, sensory and sensorimotor\nmodalities, and acquired by means of crossmodal integration. However,\ncharacterising the underlying mechanisms in the brain is difficult and\nexplaining the grounding of language in crossmodal perception and action\nremains challenging. In this paper, we present a neurocognitive model for\nlanguage grounding which reflects bio-inspired mechanisms such as an implicit\nadaptation of timescales as well as end-to-end multimodal abstraction. It\naddresses developmental robotic interaction and extends its learning\ncapabilities using larger-scale knowledge-based data. In our scenario, we\nutilise the humanoid robot NICO in obtaining the EMIL data collection, in which\nthe cognitive robot interacts with objects in a children's playground\nenvironment while receiving linguistic labels from a caregiver. The model\nanalysis shows that crossmodally integrated representations are sufficient for\nacquiring language merely from sensory input through interaction with objects\nin an environment. The representations self-organise hierarchically and embed\ntemporal and spatial information through composition and decomposition. This\nmodel can also provide the basis for further crossmodal integration of\nperceptually grounded cognitive representations."", 'We address the problem of solving mixed random linear equations. We have\nunlabeled observations coming from multiple linear regressions, and each\nobservation corresponds to exactly one of the regression models. The goal is to\nlearn the linear regressors from the observations. Classically, Alternating\nMinimization (AM) (which is a variant of Expectation Maximization (EM)) is used\nto solve this problem. AM iteratively alternates between the estimation of\nlabels and solving the regression problems with the estimated labels.\nEmpirically, it is observed that, for a large variety of non-convex problems\nincluding mixed linear regression, AM converges at a much faster rate compared\nto gradient based algorithms. However, the existing theory suggests similar\nrate of convergence for AM and gradient based methods, failing to capture this\nempirical behavior. In this paper, we close this gap between theory and\npractice for the special case of a mixture of $2$ linear regressions. We show\nthat, provided initialized properly, AM enjoys a \\emph{super-linear} rate of\nconvergence in certain parameter regimes. To the best of our knowledge, this is\nthe first work that theoretically establishes such rate for AM. Hence, if we\nwant to recover the unknown regressors upto an error (in $\\ell_2$ norm) of\n$\\epsilon$, AM only takes $\\mathcal{O}(\\log \\log (1/\\epsilon))$ iterations.\nFurthermore, we compare AM with a gradient based heuristic algorithm\nempirically and show that AM dominates in iteration complexity as well as\nwall-clock time.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently engaged in a project that involves compressing High Dynamic Range (HDR) images to be compatible with older JPEG formats. I have found that the existing JPEG XT standard, which is designed to support high dynamic range imaging, does not fully accommodate HDR images. Consequently, I am in need of a model that can convert HDR images into a JPEG XT version with minimal loss of information. I am interested in learning how to incorporate image compression methods into this task. My objective is to develop a universal model that can effectively handle this task without the need to find the optimal setting for each individual image. I am seeking guidance on how to design and train such a model, including the selection of appropriate training data and suitable performance metrics. I anticipate that a neural network-based compression algorithm could potentially outperform traditional algorithms. Therefore, I would appreciate any insights into the compression methods based on CNN, autoencoders, or GANs.","['Detection of contrast adjustments in the presence of JPEG postprocessing is\nknown to be a challenging task. JPEG post processing is often applied\ninnocently, as JPEG is the most common image format, or it may correspond to a\nlaundering attack, when it is purposely applied to erase the traces of\nmanipulation. In this paper, we propose a CNN-based detector for generic\ncontrast adjustment, which is robust to JPEG compression. The proposed system\nrelies on a patch-based Convolutional Neural Network (CNN), trained to\ndistinguish pristine images from contrast adjusted images, for some selected\nadjustment operators of different nature. Robustness to JPEG compression is\nachieved by training the CNN with JPEG examples, compressed over a range of\nQuality Factors (QFs). Experimental results show that the detector works very\nwell and scales well with respect to the adjustment type, yielding very good\nperformance under a large variety of unseen tonal adjustments.', 'Convolutional neural networks (CNNs) have achieved astonishing advances over\nthe past decade, defining state-of-the-art in several computer vision tasks.\nCNNs are capable of learning robust representations of the data directly from\nthe RGB pixels. However, most image data are usually available in compressed\nformat, from which the JPEG is the most widely used due to transmission and\nstorage purposes demanding a preliminary decoding process that have a high\ncomputational load and memory usage. For this reason, deep learning methods\ncapable of leaning directly from the compressed domain have been gaining\nattention in recent years. These methods adapt typical CNNs to work on the\ncompressed domain, but the common architectural modifications lead to an\nincrease in computational complexity and the number of parameters. In this\npaper, we investigate the usage of CNNs that are designed to work directly with\nthe DCT coefficients available in JPEG compressed images, proposing a\nhandcrafted and data-driven techniques for reducing the computational\ncomplexity and the number of parameters for these models in order to keep their\ncomputational cost similar to their RGB baselines. We make initial ablation\nstudies on a subset of ImageNet in order to analyse the impact of different\nfrequency ranges, image resolution, JPEG quality and classification task\ndifficulty on the performance of the models. Then, we evaluate the models on\nthe complete ImageNet dataset. Our results indicate that DCT models are capable\nof obtaining good performance, and that it is possible to reduce the\ncomputational complexity and the number of parameters from these models while\nretaining a similar classification accuracy through the use of our proposed\ntechniques.', ""With the growth of computer vision based applications and services, an\nexplosive amount of images have been uploaded to cloud servers which host such\ncomputer vision algorithms, usually in the form of deep learning models. JPEG\nhas been used as the {\\em de facto} compression and encapsulation method before\none uploads the images, due to its wide adaptation. However, standard JPEG\nconfiguration does not always perform well for compressing images that are to\nbe processed by a deep learning model, e.g., the standard quality level of JPEG\nleads to 50\\% of size overhead (compared with the best quality level selection)\non ImageNet under the same inference accuracy in popular computer vision models\nincluding InceptionNet, ResNet, etc. Knowing this, designing a better JPEG\nconfiguration for online computer vision services is still extremely\nchallenging: 1) Cloud-based computer vision models are usually a black box to\nend-users; thus it is difficult to design JPEG configuration without knowing\ntheir model structures. 2) JPEG configuration has to change when different\nusers use it. In this paper, we propose a reinforcement learning based JPEG\nconfiguration framework. In particular, we design an agent that adaptively\nchooses the compression level according to the input image's features and\nbackend deep learning models. Then we train the agent in a reinforcement\nlearning way to adapt it for different deep learning cloud services that act as\nthe {\\em interactive training environment} and feeding a reward with\ncomprehensive consideration of accuracy and data size. In our real-world\nevaluation on Amazon Rekognition, Face++ and Baidu Vision, our approach can\nreduce the size of images by 1/2 -- 1/3 while the overall classification\naccuracy only decreases slightly."", 'We propose an efficient two-layer near-lossless coding method using an\nextended histogram packing technique with backward compatibility to the legacy\nJPEG standard. The JPEG XT, which is the international standard to compress HDR\nimages, adopts a two-layer coding method for backward compatibility to the\nlegacy JPEG standard. However, there are two problems with this two-layer\ncoding method. One is that it does not exhibit better near-lossless performance\nthan other methods for HDR image compression with single-layer structure. The\nother problem is that the determining the appropriate values of the coding\nparameters may be required for each input image to achieve good compression\nperformance of near-lossless compression with the two-layer coding method of\nthe JPEG XT. To solve these problems, we focus on a histogram-packing technique\nthat takes into account the histogram sparseness of HDR images. We used\nzero-skip quantization, which is an extension of the histogram-packing\ntechnique proposed for lossless coding, for implementing the proposed\nnear-lossless coding method. The experimental results indicate that the\nproposed method exhibits not only a better near-lossless compression\nperformance than that of the two-layer coding method of the JPEG XT, but also\nthere are no issue regarding the combination of parameter values without losing\nbackward compatibility to the JPEG standard.', 'An efficient two-layer coding method using the histogram packing technique\nwith the backward compatibility to the legacy JPEG is proposed in this paper.\nThe JPEG XT, which is the international standard to compress HDR images, adopts\ntwo-layer coding scheme for backward compatibility to the legacy JPEG. However,\nthis two-layer coding structure does not give better lossless performance than\nthe other existing methods for HDR image compression with single-layer\nstructure. Moreover, the lossless compression of the JPEG XT has a problem on\ndetermination of the coding parameters; The lossless performance is affected by\nthe input images and/or the parameter values. That is, finding appropriate\ncombination of the values is necessary to achieve good lossless performance. It\nis firstly pointed out that the histogram packing technique considering the\nhistogram sparseness of HDR images is able to improve the performance of\nlossless compression. Then, a novel two-layer coding with the histogram packing\ntechnique and an additional lossless encoder is proposed. The experimental\nresults demonstrate that not only the proposed method has a better lossless\ncompression performance than that of the JPEG XT, but also there is no need to\ndetermine image-dependent parameter values for good compression performance\nwithout losing the backward compatibility to the well known legacy JPEG\nstandard.', 'Deep learning for computer vision depends on lossy image compression: it\nreduces the storage required for training and test data and lowers transfer\ncosts in deployment. Mainstream datasets and imaging pipelines all rely on\nstandard JPEG compression. In JPEG, the degree of quantization of frequency\ncoefficients controls the lossiness: an 8 by 8 quantization table (Q-table)\ndecides both the quality of the encoded image and the compression ratio. While\na long history of work has sought better Q-tables, existing work either seeks\nto minimize image distortion or to optimize for models of the human visual\nsystem. This work asks whether JPEG Q-tables exist that are ""better"" for\nspecific vision networks and can offer better quality--size trade-offs than\nones designed for human perception or minimal distortion. We reconstruct an\nImageNet test set with higher resolution to explore the effect of JPEG\ncompression under novel Q-tables. We attempt several approaches to tune a\nQ-table for a vision task. We find that a simple sorted random sampling method\ncan exceed the performance of the standard JPEG Q-table. We also use\nhyper-parameter tuning techniques including bounded random search, Bayesian\noptimization, and composite heuristic optimization methods. The new Q-tables we\nobtained can improve the compression rate by 10% to 200% when the accuracy is\nfixed, or improve accuracy up to $2\\%$ at the same compression rate.', ""Lossy compression brings artifacts into the compressed image and degrades the\nvisual quality. In recent years, many compression artifacts removal methods\nbased on convolutional neural network (CNN) have been developed with great\nsuccess. However, these methods usually train a model based on one specific\nvalue or a small range of quality factors. Obviously, if the test image's\nquality factor does not match to the assumed value range, then degraded\nperformance will be resulted. With this motivation and further consideration of\npractical usage, a highly robust compression artifacts removal network is\nproposed in this paper. Our proposed network is a single model approach that\ncan be trained for handling a wide range of quality factors while consistently\ndelivering superior or comparable image artifacts removal performance. To\ndemonstrate, we focus on the JPEG compression with quality factors, ranging\nfrom 1 to 60. Note that a turnkey success of our proposed network lies in the\nnovel utilization of the quantization tables as part of the training data.\nFurthermore, it has two branches in parallel---i.e., the restoration branch and\nthe global branch. The former effectively removes the local artifacts, such as\nringing artifacts removal. On the other hand, the latter extracts the global\nfeatures of the entire image that provides highly instrumental image quality\nimprovement, especially effective on dealing with the global artifacts, such as\nblocking, color shifting. Extensive experimental results performed on color and\ngrayscale images have clearly demonstrated the effectiveness and efficacy of\nour proposed single-model approach on the removal of compression artifacts from\nthe decoded image."", 'The popularity of Convolutional Neural Network (CNN) in the field of Image\nProcessing and Computer Vision has motivated researchers and industrialist\nexperts across the globe to solve different challenges with high accuracy. The\nsimplest way to train a CNN classifier is to directly feed the original RGB\npixels images into the network. However, if we intend to classify images\ndirectly with its compressed data, the same approach may not work better, like\nin case of JPEG compressed images. This research paper investigates the issues\nof modifying the input representation of the JPEG compressed data, and then\nfeeding into the CNN. The architecture is termed as DCT-CompCNN. This novel\napproach has shown that CNNs can also be trained with JPEG compressed DCT\ncoefficients, and subsequently can produce a better performance in comparison\nwith the conventional CNN approach. The efficiency of the modified input\nrepresentation is tested with the existing ResNet-50 architecture and the\nproposed DCT-CompCNN architecture on a public image classification datasets\nlike Dog Vs Cat and CIFAR-10 datasets, reporting a better performance', 'We propose a novel JPEG XT image compression with hue compensation for\ntwo-layer HDR coding. LDR images produced from JPEG XT bitstreams have some\ndistortion in hue due to tone mapping operations. In order to suppress the\ncolor distortion, we apply a novel hue compensation method based on the\nmaximally saturated colors. Moreover, the bitstreams generated by using the\nproposed method are fully compatible with the JPEG XT standard. In an\nexperiment, the proposed method is demonstrated not only to produce images with\nsmall hue degradation but also to maintain well-mapped luminance, in terms of\nthree kinds of criterion: TMQI, hue value in CIEDE2000, and the maximally\nsaturated color on the constant-hue plane.', 'We propose a new approach to the problem of optimizing autoencoders for lossy\nimage compression. New media formats, changing hardware technology, as well as\ndiverse requirements and content types create a need for compression algorithms\nwhich are more flexible than existing codecs. Autoencoders have the potential\nto address this need, but are difficult to optimize directly due to the\ninherent non-differentiabilty of the compression loss. We here show that\nminimal changes to the loss are sufficient to train deep autoencoders\ncompetitive with JPEG 2000 and outperforming recently proposed approaches based\non RNNs. Our network is furthermore computationally efficient thanks to a\nsub-pixel architecture, which makes it suitable for high-resolution images.\nThis is in contrast to previous work on autoencoders for compression using\ncoarser approximations, shallower architectures, computationally expensive\nmethods, or focusing on small images.', 'Most image data available are often stored in a compressed format, from which\nJPEG is the most widespread. To feed this data on a convolutional neural\nnetwork (CNN), a preliminary decoding process is required to obtain RGB pixels,\ndemanding a high computational load and memory usage. For this reason, the\ndesign of CNNs for processing JPEG compressed data has gained attention in\nrecent years. In most existing works, typical CNN architectures are adapted to\nfacilitate the learning with the DCT coefficients rather than RGB pixels.\nAlthough they are effective, their architectural changes either raise the\ncomputational costs or neglect relevant information from DCT inputs. In this\npaper, we examine different ways of speeding up CNNs designed for DCT inputs,\nexploiting learning strategies to reduce the computational complexity by taking\nfull advantage of DCT inputs. Our experiments were conducted on the ImageNet\ndataset. Results show that learning how to combine all DCT inputs in a\ndata-driven fashion is better than discarding them by hand, and its combination\nwith a reduction of layers has proven to be effective for reducing the\ncomputational costs while retaining accuracy.', ""Display technology must offer high dynamic range (HDR) contrast-based depth\ninduction and 3D personalization simultaneously. Efficient algorithms to\ncompress HDR stereo data is critical. Direct capturing of HDR content is\ncomplicated due to the high expense and scarcity of HDR cameras. The HDR 3D\nimages could be generated in low-cost by fusing low-dynamic-range (LDR) images\nacquired using a stereo camera with various exposure settings. In this paper,\nan efficient scheme for coding multi-exposure stereo images is proposed based\non a tensor low-rank approximation scheme. The multi-exposure fusion can be\nrealized to generate HDR stereo output at the decoder for increased realism and\nexaggerated binocular 3D depth cues.\n  For exploiting spatial redundancy in LDR stereo images, the stack of\nmulti-exposure stereo images is decomposed into a set of projection matrices\nand a core tensor following an alternating least squares Tucker decomposition\nmodel. The compact, low-rank representation of the scene, thus, generated is\nfurther processed by 3D extension of High Efficiency Video Coding standard. The\nencoding with 3D-HEVC enhance the proposed scheme efficiency by exploiting\nintra-frame, inter-view and the inter-component redundancies in low-rank\napproximated representation. We consider constant luminance property of IPT and\nY'CbCr color space to precisely approximate intensity prediction and\nperceptually minimize the encoding distortion. Besides, the proposed scheme\ngives flexibility to adjust the bitrate of tensor latent components by changing\nthe rank of core tensor and its quantization. Extensive experiments on natural\nscenes demonstrate that the proposed scheme outperforms state-of-the-art\nJPEG-XT and 3D-HEVC range coding standards."", ""Deep convolutional neural networks (DCNNs) have aided high dynamic range\n(HDR) imaging recently and have received a lot of attention. The quality of\nDCNN-generated HDR images has overperformed the traditional counterparts.\nHowever, DCNNs are prone to be computationally intensive and power-hungry, and\nhence cannot be implemented on various embedded computing platforms with\nlimited power and hardware resources. Embedded systems have a huge market, and\nutilizing DCNNs' powerful functionality into them will further reduce human\nintervention. To address the challenge, we propose LightFuse, a lightweight\nCNN-based algorithm for extreme dual-exposure image fusion, which achieves\nbetter functionality than a conventional DCNN and can be deployed in embedded\nsystems. Two sub-networks are utilized: a GlobalNet (G) and a DetailNet (D).\nThe goal of G is to learn the global illumination information on the spatial\ndimension, whereas D aims to enhance local details on the channel dimension.\nBoth G and D are based solely on depthwise convolution (D_Conv) and pointwise\nconvolution (P_Conv) to reduce required parameters and computations.\nExperimental results show that this proposed technique could generate HDR\nimages in extremely exposed regions with sufficient details to be legible. Our\nmodel outperforms other state-of-the-art approaches in peak signal-to-noise\nratio (PSNR) score by 0.9 to 8.7 while achieving 16.7 to 306.2 times parameter\nreduction."", 'Deep Neural Networks trained as image auto-encoders have recently emerged as\na promising direction for advancing the state-of-the-art in image compression.\nThe key challenge in learning such networks is twofold: To deal with\nquantization, and to control the trade-off between reconstruction error\n(distortion) and entropy (rate) of the latent image representation. In this\npaper, we focus on the latter challenge and propose a new technique to navigate\nthe rate-distortion trade-off for an image compression auto-encoder. The main\nidea is to directly model the entropy of the latent representation by using a\ncontext model: A 3D-CNN which learns a conditional probability model of the\nlatent distribution of the auto-encoder. During training, the auto-encoder\nmakes use of the context model to estimate the entropy of its representation,\nand the context model is concurrently updated to learn the dependencies between\nthe symbols in the latent representation. Our experiments show that this\napproach, when measured in MS-SSIM, yields a state-of-the-art image compression\nsystem based on a simple convolutional auto-encoder.', 'In this paper we propose a deep residual autoencoder exploiting\nResidual-in-Residual Dense Blocks (RRDB) to remove artifacts in JPEG compressed\nimages that is independent from the Quality Factor (QF) used. The proposed\napproach leverages both the learning capacity of deep residual networks and\nprior knowledge of the JPEG compression pipeline. The proposed model operates\nin the YCbCr color space and performs JPEG artifact restoration in two phases\nusing two different autoencoders: the first one restores the luma channel\nexploiting 2D convolutions; the second one, using the restored luma channel as\na guide, restores the chroma channels explotining 3D convolutions. Extensive\nexperimental results on three widely used benchmark datasets (i.e. LIVE1,\nBDS500, and CLASSIC-5) show that our model is able to outperform the state of\nthe art with respect to all the evaluation metrics considered (i.e. PSNR,\nPSNR-B, and SSIM). This results is remarkable since the approaches in the state\nof the art use a different set of weights for each compression quality, while\nthe proposed model uses the same weights for all of them, making it applicable\nto images in the wild where the QF used for compression is unkwnown.\nFurthermore, the proposed model shows a greater robustness than\nstate-of-the-art methods when applied to compression qualities not seen during\ntraining.', ""In the advent of a digital health revolution, vast amounts of clinical data\nare being generated, stored and processed on a daily basis. This has made the\nstorage and retrieval of large volumes of health-care data, especially,\nhigh-resolution medical images, particularly challenging. Effective image\ncompression for medical images thus plays a vital role in today's healthcare\ninformation system, particularly in teleradiology. In this work, an X-ray image\ncompression method based on a Convolutional Recurrent Neural Networks RNN-Conv\nis presented. The proposed architecture can provide variable compression rates\nduring deployment while it requires each network to be trained only once for a\nspecific dimension of X-ray images. The model uses a multi-level pooling scheme\nthat learns contextualized features for effective compression. We perform our\nimage compression experiments on the National Institute of Health (NIH)\nChestX-ray8 dataset and compare the performance of the proposed architecture\nwith a state-of-the-art RNN based technique and JPEG 2000. The experimental\nresults depict improved compression performance achieved by the proposed method\nin terms of Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio\n(PSNR) metrics. To the best of our knowledge, this is the first reported\nevaluation on using a deep convolutional RNN for medical image compression."", 'This paper proposes a new end-to-end trainable model for lossy image\ncompression, which includes several novel components. The method incorporates\n1) an adequate perceptual similarity metric; 2) saliency in the images; 3) a\nhierarchical auto-regressive model. This paper demonstrates that the popularly\nused evaluations metrics such as MS-SSIM and PSNR are inadequate for judging\nthe performance of image compression techniques as they do not align with the\nhuman perception of similarity. Alternatively, a new metric is proposed, which\nis learned on perceptual similarity data specific to image compression. The\nproposed compression model incorporates the salient regions and optimizes on\nthe proposed perceptual similarity metric. The model not only generates images\nwhich are visually better but also gives superior performance for subsequent\ncomputer vision tasks such as object detection and segmentation when compared\nto existing engineered or learned compression techniques.', 'We describe an image compression method, consisting of a nonlinear analysis\ntransformation, a uniform quantizer, and a nonlinear synthesis transformation.\nThe transforms are constructed in three successive stages of convolutional\nlinear filters and nonlinear activation functions. Unlike most convolutional\nneural networks, the joint nonlinearity is chosen to implement a form of local\ngain control, inspired by those used to model biological neurons. Using a\nvariant of stochastic gradient descent, we jointly optimize the entire model\nfor rate-distortion performance over a database of training images, introducing\na continuous proxy for the discontinuous loss function arising from the\nquantizer. Under certain conditions, the relaxed loss function may be\ninterpreted as the log likelihood of a generative model, as implemented by a\nvariational autoencoder. Unlike these models, however, the compression model\nmust operate at any given point along the rate-distortion curve, as specified\nby a trade-off parameter. Across an independent set of test images, we find\nthat the optimized method generally exhibits better rate-distortion performance\nthan the standard JPEG and JPEG 2000 compression methods. More importantly, we\nobserve a dramatic improvement in visual quality for all images at all bit\nrates, which is supported by objective quality estimates using MS-SSIM.', 'A large fraction of Internet traffic is now driven by requests from mobile\ndevices with relatively small screens and often stringent bandwidth\nrequirements. Due to these factors, it has become the norm for modern\ngraphics-heavy websites to transmit low-resolution, low-bytecount image\npreviews (thumbnails) as part of the initial page load process to improve\napparent page responsiveness. Increasing thumbnail compression beyond the\ncapabilities of existing codecs is therefore a current research focus, as any\nbyte savings will significantly enhance the experience of mobile device users.\nToward this end, we propose a general framework for variable-rate image\ncompression and a novel architecture based on convolutional and deconvolutional\nLSTM recurrent networks. Our models address the main issues that have prevented\nautoencoder neural networks from competing with existing image compression\nalgorithms: (1) our networks only need to be trained once (not per-image),\nregardless of input image dimensions and the desired compression rate; (2) our\nnetworks are progressive, meaning that the more bits are sent, the more\naccurate the image reconstruction; and (3) the proposed architecture is at\nleast as efficient as a standard purpose-trained autoencoder for a given number\nof bits. On a large-scale benchmark of 32$\\times$32 thumbnails, our LSTM-based\napproaches provide better visual quality than (headerless) JPEG, JPEG2000 and\nWebP, with a storage size that is reduced by 10% or more.', 'The field of neural image compression has witnessed exciting progress as\nrecently proposed architectures already surpass the established transform\ncoding based approaches. While, so far, research has mainly focused on\narchitecture and model improvements, in this work we explore content adaptive\noptimization. To this end, we introduce an iterative procedure which adapts the\nlatent representation to the specific content we wish to compress while keeping\nthe parameters of the network and the predictive model fixed. Our experiments\nshow that this allows for an overall increase in rate-distortion performance,\nindependently of the specific architecture used. Furthermore, we also evaluate\nthis strategy in the context of adapting a pretrained network to other content\nthat is different in visual appearance or resolution. Here, our experiments\nshow that our adaptation strategy can largely close the gap as compared to\nmodels specifically trained for the given content while having the benefit that\nno additional data in the form of model parameter updates has to be\ntransmitted.']","['In recent years we have witnessed an increasing interest in applying Deep\nNeural Networks (DNNs) to improve the rate-distortion performance in image\ncompression. However, the existing approaches either train a post-processing\nDNN on the decoder side, or propose learning for image compression in an\nend-to-end manner. This way, the trained DNNs are required in the decoder,\nleading to the incompatibility to the standard image decoders (e.g., JPEG) in\npersonal computers and mobiles. Therefore, we propose learning to improve the\nencoding performance with the standard decoder. In this paper, We work on JPEG\nas an example. Specifically, a frequency-domain pre-editing method is proposed\nto optimize the distribution of DCT coefficients, aiming at facilitating the\nJPEG compression. Moreover, we propose learning the JPEG quantization table\njointly with the pre-editing network. Most importantly, we do not modify the\nJPEG decoder and therefore our approach is applicable when viewing images with\nthe widely used standard JPEG decoder. The experiments validate that our\napproach successfully improves the rate-distortion performance of JPEG in terms\nof various quality metrics, such as PSNR, MS-SSIM and LPIPS. Visually, this\ntranslates to better overall color retention especially when strong compression\nis applied. The codes are available at\nhttps://github.com/YannickStruempler/LearnedJPEG.', 'High Dynamic Range (HDR) images are the ones that contain a greater range of\nluminosity as compared to the standard images. HDR images have a higher detail\nand clarity of structure, objects, and color, which the standard images lack.\nHDR images are useful in capturing scenes that pose high brightness, darker\nareas, and shadows, etc. An HDR image comprises multiple narrow-range-exposure\nimages combined into one high-quality image. As these HDR images cannot be\ndisplayed on standard display devices, the real challenge comes while\nconverting these HDR images to Low dynamic range (LDR) images. The conversion\nof HDR image to LDR image is performed using Tone-mapped operators (TMOs). This\nconversion results in the loss of much valuable information in structure,\ncolor, naturalness, and exposures. The loss of information in the LDR image may\nnot directly be visible to the human eye. To calculate how good an LDR image is\nafter conversion, various metrics have been proposed previously. Some are not\nnoise resilient, some work on separate color channels (Red, Green, and Blue one\nby one), and some lack capacity to identify the structure. To deal with this\nproblem, we propose a metric in this paper called the Tone Mapping Quality\nIndex (TMQI-3), which evaluates the quality of the LDR image based on its\nobjective score. TMQI-3 is noise resilient, takes account of structure and\nnaturalness, and works on all three color channels combined into one luminosity\ncomponent. This eliminates the need to use multiple metrics at the same time.\nWe compute results for several HDR and LDR images from the literature and show\nthat our quality index metric performs better than the baseline models.', 'Recent studies have shown that neural network (NN) based image classifiers\nare highly vulnerable to adversarial examples, which poses a threat to\nsecurity-sensitive image recognition task. Prior work has shown that JPEG\ncompression can combat the drop in classification accuracy on adversarial\nexamples to some extent. But, as the compression ratio increases, traditional\nJPEG compression is insufficient to defend those attacks but can cause an\nabrupt accuracy decline to the benign images. In this paper, with the aim of\nfully filtering the adversarial perturbations, we firstly make modifications to\ntraditional JPEG compression algorithm which becomes more favorable for NN.\nSpecifically, based on an analysis of the frequency coefficient, we design a\nNN-favored quantization table for compression. Considering compression as a\ndata augmentation strategy, we then combine our model-agnostic preprocess with\nnoisy training. We fine-tune the pre-trained model by training with images\nencoded at different compression levels, thus generating multiple classifiers.\nFinally, since lower (higher) compression ratio can remove both perturbations\nand original features slightly (aggressively), we use these trained multiple\nmodels for model ensemble. The majority vote of the ensemble of models is\nadopted as final predictions. Experiments results show our method can improve\ndefense efficiency while maintaining original accuracy.', 'As one of most fascinating machine learning techniques, deep neural network\n(DNN) has demonstrated excellent performance in various intelligent tasks such\nas image classification. DNN achieves such performance, to a large extent, by\nperforming expensive training over huge volumes of training data. To reduce the\ndata storage and transfer overhead in smart resource-limited Internet-of-Thing\n(IoT) systems, effective data compression is a ""must-have"" feature before\ntransferring real-time produced dataset for training or classification. While\nthere have been many well-known image compression approaches (such as JPEG), we\nfor the first time find that a human-visual based image compression approach\nsuch as JPEG compression is not an optimized solution for DNN systems,\nespecially with high compression ratios. To this end, we develop an image\ncompression framework tailored for DNN applications, named ""DeepN-JPEG"", to\nembrace the nature of deep cascaded information process mechanism of DNN\narchitecture. Extensive experiments, based on ""ImageNet"" dataset with various\nstate-of-the-art DNNs, show that ""DeepN-JPEG"" can achieve ~3.5x higher\ncompression rate over the popular JPEG solution while maintaining the same\naccuracy level for image recognition, demonstrating its great potential of\nstorage and power efficiency in DNN-based smart IoT system design.', 'A two-layer lossless image coding method compatible with JPEG XS is proposed.\nJPEG XS is a new international standard for still image coding that has the\ncharacteristics of very low latency and very low complexity. However, it does\nnot support lossless coding, although it can achieve visual lossless coding.\nThe proposed method has a two-layer structure similar to JPEG XT, which\nconsists of JPEG XS coding and a lossless coding method. As a result, it\nenables us to losslessly restore original images, while maintaining\ncompatibility with JPEG XS.', 'The prime goal of digital imaging techniques is to reproduce the realistic\nappearance of a scene. Low Dynamic Range (LDR) cameras are incapable of\nrepresenting the wide dynamic range of the real-world scene. The captured\nimages turn out to be either too dark (underexposed) or too bright\n(overexposed). Specifically, saturation in overexposed regions makes the task\nof reconstructing a High Dynamic Range (HDR) image from single LDR image\nchallenging. In this paper, we propose a deep learning based approach to\nrecover details in the saturated areas while reconstructing the HDR image. We\nformulate this problem as an image-to-image (I2I) translation task. To this\nend, we present a novel conditional GAN (cGAN) based framework trained in an\nend-to-end fashion over the HDR-REAL and HDR-SYNTH datasets. Our framework uses\nan overexposed mask obtained from a pre-trained segmentation model to\nfacilitate the hallucination task of adding details in the saturated regions.\nWe demonstrate the effectiveness of the proposed method by performing an\nextensive quantitative and qualitative comparison with several state-of-the-art\nsingle-image HDR reconstruction techniques.', 'Recent advances in generalized image understanding have seen a surge in the\nuse of deep convolutional neural networks (CNN) across a broad range of\nimage-based detection, classification and prediction tasks. Whilst the reported\nperformance of these approaches is impressive, this study investigates the\nhitherto unapproached question of the impact of commonplace image and video\ncompression techniques on the performance of such deep learning architectures.\nFocusing on the JPEG and H.264 (MPEG-4 AVC) as a representative proxy for\ncontemporary lossy image/video compression techniques that are in common use\nwithin network-connected image/video devices and infrastructure, we examine the\nimpact on performance across five discrete tasks: human pose estimation,\nsemantic segmentation, object detection, action recognition, and monocular\ndepth estimation. As such, within this study we include a variety of network\narchitectures and domains spanning end-to-end convolution, encoder-decoder,\nregion-based CNN (R-CNN), dual-stream, and generative adversarial networks\n(GAN). Our results show a non-linear and non-uniform relationship between\nnetwork performance and the level of lossy compression applied. Notably,\nperformance decreases significantly below a JPEG quality (quantization) level\nof 15% and a H.264 Constant Rate Factor (CRF) of 40. However, retraining said\narchitectures on pre-compressed imagery conversely recovers network performance\nby up to 78.4% in some cases. Furthermore, there is a correlation between\narchitectures employing an encoder-decoder pipeline and those that demonstrate\nresilience to lossy image compression. The characteristics of the relationship\nbetween input compression to output task performance can be used to inform\ndesign decisions within future image/video devices and infrastructure.', 'This paper presents a set of full-resolution lossy image compression methods\nbased on neural networks. Each of the architectures we describe can provide\nvariable compression rates during deployment without requiring retraining of\nthe network: each network need only be trained once. All of our architectures\nconsist of a recurrent neural network (RNN)-based encoder and decoder, a\nbinarizer, and a neural network for entropy coding. We compare RNN types (LSTM,\nassociative LSTM) and introduce a new hybrid of GRU and ResNet. We also study\n""one-shot"" versus additive reconstruction architectures and introduce a new\nscaled-additive framework. We compare to previous work, showing improvements of\n4.3%-8.8% AUC (area under the rate-distortion curve), depending on the\nperceptual metric used. As far as we know, this is the first neural network\narchitecture that is able to outperform JPEG at image compression across most\nbitrates on the rate-distortion curve on the Kodak dataset images, with and\nwithout the aid of entropy coding.', 'Mapping a single exposure low dynamic range (LDR) image into a high dynamic\nrange (HDR) is considered among the most strenuous image to image translation\ntasks due to exposure-related missing information. This study tackles the\nchallenges of single-shot LDR to HDR mapping by proposing a novel two-stage\ndeep network. Notably, our proposed method aims to reconstruct an HDR image\nwithout knowing hardware information, including camera response function (CRF)\nand exposure settings. Therefore, we aim to perform image enhancement task like\ndenoising, exposure correction, etc., in the first stage. Additionally, the\nsecond stage of our deep network learns tone mapping and bit-expansion from a\nconvex set of data samples. The qualitative and quantitative comparisons\ndemonstrate that the proposed method can outperform the existing LDR to HDR\nworks with a marginal difference. Apart from that, we collected an LDR image\ndataset incorporating different camera systems. The evaluation with our\ncollected real-world LDR images illustrates that the proposed method can\nreconstruct plausible HDR images without presenting any visual artefacts. Code\navailable: https://github. com/sharif-apu/twostageHDR_NTIRE21.', 'High dynamic range (HDR) imaging has recently drawn much attention in\nmultimedia community. In this paper, we proposed a HDR image forensics method\nbased on convolutional neural network (CNN).To our best knowledge, this is the\nfirst time to apply deep learning method on HDR image forensics. The proposed\nalgorithm uses CNN to distinguish HDR images generated by multiple low dynamic\nrange (LDR) images from that expanded by single LDR image using inverse tone\nmapping (iTM). To do this, we learn the change of statistical characteristics\nextracted by the proposed CNN architectures and classify two kinds of HDR\nimages. Comparision results with some traditional statistical characteristics\nshows efficiency of the proposed method in HDR image source identification.', 'Increasing popularity of high-dynamic-range (HDR) image and video content\nbrings the need for metrics that could predict the severity of image\nimpairments as seen on displays of different brightness levels and dynamic\nrange. Such metrics should be trained and validated on a sufficiently large\nsubjective image quality dataset to ensure robust performance. As the existing\nHDR quality datasets are limited in size, we created a Unified Photometric\nImage Quality dataset (UPIQ) with over 4,000 images by realigning and merging\nexisting HDR and standard-dynamic-range (SDR) datasets. The realigned quality\nscores share the same unified quality scale across all datasets. Such\nrealignment was achieved by collecting additional cross-dataset quality\ncomparisons and re-scaling data with a psychometric scaling method. Images in\nthe proposed dataset are represented in absolute photometric and colorimetric\nunits, corresponding to light emitted from a display. We use the new dataset to\nretrain existing HDR metrics and show that the dataset is sufficiently large\nfor training deep architectures. We show the utility of the dataset on\nbrightness aware image compression.', 'High dynamic range (HDR) imaging provides the capability of handling real\nworld lighting as opposed to the traditional low dynamic range (LDR) which\nstruggles to accurately represent images with higher dynamic range. However,\nmost imaging content is still available only in LDR. This paper presents a\nmethod for generating HDR content from LDR content based on deep Convolutional\nNeural Networks (CNNs) termed ExpandNet. ExpandNet accepts LDR images as input\nand generates images with an expanded range in an end-to-end fashion. The model\nattempts to reconstruct missing information that was lost from the original\nsignal due to quantization, clipping, tone mapping or gamma correction. The\nadded information is reconstructed from learned features, as the network is\ntrained in a supervised fashion using a dataset of HDR images. The approach is\nfully automatic and data driven; it does not require any heuristics or human\nexpertise. ExpandNet uses a multiscale architecture which avoids the use of\nupsampling layers to improve image quality. The method performs well compared\nto expansion/inverse tone mapping operators quantitatively on multiple metrics,\neven for badly exposed inputs.', 'Bottleneck autoencoders have been actively researched as a solution to image\ncompression tasks. However, we observed that bottleneck autoencoders produce\nsubjectively low quality reconstructed images. In this work, we explore the\nability of sparse coding to improve reconstructed image quality for the same\ndegree of compression. We observe that sparse image compression produces\nvisually superior reconstructed images and yields higher values of pixel-wise\nmeasures of reconstruction quality (PSNR and SSIM) compared to bottleneck\nautoencoders. % In addition, we find that using alternative metrics that\ncorrelate better with human perception, such as feature perceptual loss and the\nclassification accuracy, sparse image compression scores up to 18.06\\% and\n2.7\\% higher, respectively, compared to bottleneck autoencoders. Although\ncomputationally much more intensive, we find that sparse coding is otherwise\nsuperior to bottleneck autoencoders for the same degree of compression.', 'In this paper, we propose a novel deep neural network model that reconstructs\na high dynamic range (HDR) image from a single low dynamic range (LDR) image.\nThe proposed model is based on a convolutional neural network composed of\ndilated convolutional layers, and infers LDR images with various exposures and\nillumination from a single LDR image of the same scene. Then, the final HDR\nimage can be formed by merging these inference results. It is relatively easy\nfor the proposed method to find the mapping between the LDR and an HDR with a\ndifferent bit depth because of the chaining structure inferring the\nrelationship between the LDR images with brighter (or darker) exposures from a\ngiven LDR image. The method not only extends the range, but also has the\nadvantage of restoring the light information of the actual physical world. For\nthe HDR images obtained by the proposed method, the HDR-VDP2 Q score, which is\nthe most popular evaluation metric for HDR images, was 56.36 for a display with\na 1920$\\times$1200 resolution, which is an improvement of 6 compared with the\nscores of conventional algorithms. In addition, when comparing the peak\nsignal-to-noise ratio values for tone mapped HDR images generated by the\nproposed and conventional algorithms, the average value obtained by the\nproposed algorithm is 30.86 dB, which is 10 dB higher than those obtained by\nthe conventional algorithms.', 'Inpainting based image compression approaches, especially linear and\nnon-linear diffusion models, are an active research topic for lossy image\ncompression. The major challenge in these compression models is to find a small\nset of descriptive supporting points, which allow for an accurate\nreconstruction of the original image. It turns out in practice that this is a\nchallenging problem even for the simplest Laplacian interpolation model. In\nthis paper, we revisit the Laplacian interpolation compression model and\nintroduce two fast algorithms, namely successive preconditioning primal dual\nalgorithm and the recently proposed iPiano algorithm, to solve this problem\nefficiently. Furthermore, we extend the Laplacian interpolation based\ncompression model to a more general form, which is based on principles from\nbi-level optimization. We investigate two different variants of the Laplacian\nmodel, namely biharmonic interpolation and smoothed Total Variation\nregularization. Our numerical results show that significant improvements can be\nobtained from the biharmonic interpolation model, and it can recover an image\nwith very high quality from only 5% pixels.', 'Camera sensors can only capture a limited range of luminance simultaneously,\nand in order to create high dynamic range (HDR) images a set of different\nexposures are typically combined. In this paper we address the problem of\npredicting information that have been lost in saturated image areas, in order\nto enable HDR reconstruction from a single exposure. We show that this problem\nis well-suited for deep learning algorithms, and propose a deep convolutional\nneural network (CNN) that is specifically designed taking into account the\nchallenges in predicting HDR values. To train the CNN we gather a large dataset\nof HDR images, which we augment by simulating sensor saturation for a range of\ncameras. To further boost robustness, we pre-train the CNN on a simulated HDR\ndataset created from a subset of the MIT Places database. We demonstrate that\nour approach can reconstruct high-resolution visually convincing HDR results in\na wide range of situations, and that it generalizes well to reconstruction of\nimages captured with arbitrary and low-end cameras that use unknown camera\nresponse functions and post-processing. Furthermore, we compare to existing\nmethods for HDR expansion, and show high quality results also for image based\nlighting. Finally, we evaluate the results in a subjective experiment performed\non an HDR display. This shows that the reconstructed HDR images are visually\nconvincing, with large improvements as compared to existing methods.', 'Since its invention, Generative adversarial networks (GANs) have shown\noutstanding results in many applications. Generative Adversarial Networks are\npowerful yet, resource-hungry deep-learning models. Their main difference from\nordinary deep learning models is the nature of their output. For example, GAN\noutput can be a whole image versus other models detecting objects or\nclassifying images. Thus, the architecture and numeric precision of the network\naffect the quality and speed of the solution. Hence, accelerating GANs is\npivotal. Accelerating GANs can be classified into three main tracks: (1) Memory\ncompression, (2) Computation optimization, and (3) Data-flow optimization.\nBecause data transfer is the main source of energy usage, memory compression\nleads to the most savings. Thus, in this paper, we survey memory compression\ntechniques for CNN-Based GANs. Additionally, the paper summarizes opportunities\nand challenges in GANs acceleration and suggests open research problems to be\nfurther investigated.', 'Deep learning based image compression has recently witnessed exciting\nprogress and in some cases even managed to surpass transform coding based\napproaches that have been established and refined over many decades. However,\nstate-of-the-art solutions for deep image compression typically employ\nautoencoders which map the input to a lower dimensional latent space and thus\nirreversibly discard information already before quantization. Due to that, they\ninherently limit the range of quality levels that can be covered. In contrast,\ntraditional approaches in image compression allow for a larger range of quality\nlevels. Interestingly, they employ an invertible transformation before\nperforming the quantization step which explicitly discards information.\nInspired by this, we propose a deep image compression method that is able to go\nfrom low bit-rates to near lossless quality by leveraging normalizing flows to\nlearn a bijective mapping from the image space to a latent representation. In\naddition to this, we demonstrate further advantages unique to our solution,\nsuch as the ability to maintain constant quality results through re-encoding,\neven when performed multiple times. To the best of our knowledge, this is the\nfirst work to explore the opportunities for leveraging normalizing flows for\nlossy image compression.', 'JPEG is one of the most commonly used standards among lossy image compression\nmethods. However, JPEG compression inevitably introduces various kinds of\nartifacts, especially at high compression rates, which could greatly affect the\nQuality of Experience (QoE). Recently, convolutional neural network (CNN) based\nmethods have shown excellent performance for removing the JPEG artifacts. Lots\nof efforts have been made to deepen the CNNs and extract deeper features, while\nrelatively few works pay attention to the receptive field of the network. In\nthis paper, we illustrate that the quality of output images can be\nsignificantly improved by enlarging the receptive fields in many cases. One\nstep further, we propose a Dual-domain Multi-scale CNN (DMCNN) to take full\nadvantage of redundancies on both the pixel and DCT domains. Experiments show\nthat DMCNN sets a new state-of-the-art for the task of JPEG artifact removal.', 'The JPEG image compression algorithm is the most popular method of image\ncompression because of its ability for large compression ratios. However, to\nachieve such high compression, information is lost. For aggressive quantization\nsettings, this leads to a noticeable reduction in image quality. Artifact\ncorrection has been studied in the context of deep neural networks for some\ntime, but the current state-of-the-art methods require a different model to be\ntrained for each quality setting, greatly limiting their practical application.\nWe solve this problem by creating a novel architecture which is parameterized\nby the JPEG files quantization matrix. This allows our single model to achieve\nstate-of-the-art performance over models trained for specific quality settings.', 'Synthesizing high dynamic range (HDR) images from multiple low-dynamic range\n(LDR) exposures in dynamic scenes is challenging. There are two major problems\ncaused by the large motions of foreground objects. One is the severe\nmisalignment among the LDR images. The other is the missing content due to the\nover-/under-saturated regions caused by the moving objects, which may not be\neasily compensated for by the multiple LDR exposures. Thus, it requires the HDR\ngeneration model to be able to properly fuse the LDR images and restore the\nmissing details without introducing artifacts. To address these two problems,\nwe propose in this paper a novel GAN-based model, HDR-GAN, for synthesizing HDR\nimages from multi-exposed LDR images. To our best knowledge, this work is the\nfirst GAN-based approach for fusing multi-exposed LDR images for HDR\nreconstruction. By incorporating adversarial learning, our method is able to\nproduce faithful information in the regions with missing content. In addition,\nwe also propose a novel generator network, with a reference-based residual\nmerging block for aligning large object motions in the feature domain, and a\ndeep HDR supervision scheme for eliminating artifacts of the reconstructed HDR\nimages. Experimental results demonstrate that our model achieves\nstate-of-the-art reconstruction performance over the prior HDR methods on\ndiverse scenes.', 'This work evaluates six state-of-the-art deep neural network (DNN)\narchitectures applied to the problem of enhancing camera-captured document\nimages. The results from each network were evaluated both qualitatively and\nquantitatively using Image Quality Assessment (IQA) metrics, and also compared\nwith an existing approach based on traditional computer vision techniques. The\nbest performing architectures generally produced good enhancement compared to\nthe existing algorithm, showing that it is possible to use DNNs for document\nimage enhancement. Furthermore, the best performing architectures could work as\na baseline for future investigations on document enhancement using deep\nlearning techniques. The main contributions of this paper are: a baseline of\ndeep learning techniques that can be further improved to provide better\nresults, and a evaluation methodology using IQA metrics for quantitatively\ncomparing the produced images from the neural networks to a ground truth.', 'Document image enhancement and binarization methods are often used to improve\nthe accuracy and efficiency of document image analysis tasks such as text\nrecognition. Traditional non-machine-learning methods are constructed on\nlow-level features in an unsupervised manner but have difficulty with\nbinarization on documents with severely degraded backgrounds. Convolutional\nneural network-based methods focus only on grayscale images and on local\ntextual features. In this paper, we propose a two-stage color document image\nenhancement and binarization method using generative adversarial neural\nnetworks. In the first stage, four color-independent adversarial networks are\ntrained to extract color foreground information from an input image for\ndocument image enhancement. In the second stage, two independent adversarial\nnetworks with global and local features are trained for image binarization of\ndocuments of variable size. For the adversarial neural networks, we formulate\nloss functions between a discriminator and generators having an encoder-decoder\nstructure. Experimental results show that the proposed method achieves better\nperformance than many classical and state-of-the-art algorithms over the\nDocument Image Binarization Contest (DIBCO) datasets, the LRDE Document\nBinarization Dataset (LRDE DBD), and our shipping label image dataset. We plan\nto release the shipping label dataset as well as our implementation code at\ngithub.com/opensuh/DocumentBinarization/.', 'Image compression helps in storing the transmitted data in proficient way by\ndecreasing its redundancy. This technique helps in transferring more digital or\nmultimedia data over internet as it increases the storage space. It is\nimportant to maintain the image quality even if it is compressed to certain\nextent. Depend upon this the image compression is classified into two\ncategories : lossy and lossless image compression. There are many lossy digital\nimage compression techniques exists. Among this Incremental Self Organizing Map\nis a familiar one. The good pictures quality can be retrieved if image\ndenoising technique is used for compression and also provides better\ncompression ratio. Image denoising is an important pre-processing step for many\nimage analysis and computer vision system. It refers to the task of recovering\na good estimate of the true image from a degraded observation without altering\nand changing useful structure in the image such as discontinuities and edges.\nMany approaches have been proposed to remove the noise effectively while\npreserving the original image details and features as much as possible. This\npaper proposes a technique for image compression using Incremental Self\nOrganizing Map (ISOM) with Discret Wavelet Transform (DWT) by applying\nfiltering techniques which play a crucial role in enhancing the quality of a\nreconstructed image. The experimental result shows that the proposed technique\nobtained better compression ratio value.', 'Document Image Analysis, like any Digital Image Analysis requires\nidentification and extraction of proper features, which are generally extracted\nfrom uncompressed images, though in reality images are made available in\ncompressed form for the reasons such as transmission and storage efficiency.\nHowever, this implies that the compressed image should be decompressed, which\nindents additional computing resources. This limitation induces the motivation\nto research in extracting features directly from the compressed image. In this\nresearch, we propose to extract essential features such as projection profile,\nrun-histogram and entropy for text document analysis directly from run-length\ncompressed text-documents. The experimentation illustrates that features are\nextracted directly from the compressed image without going through the stage of\ndecompression, because of which the computing time is reduced. The feature\nvalues so extracted are exactly identical to those extracted from uncompressed\nimages.', 'The inherent noise in the observed (e.g., scanned) binary document image\ndegrades the image quality and harms the compression ratio through breaking the\npattern repentance and adding entropy to the document images. In this paper, we\ndesign a cost function in Bayesian framework with dictionary learning.\nMinimizing our cost function produces a restored image which has better quality\nthan that of the observed noisy image, and a dictionary for representing and\nencoding the image. After the restoration, we use this dictionary (from the\nsame cost function) to encode the restored image following the\nsymbol-dictionary framework by JBIG2 standard with the lossless mode.\nExperimental results with a variety of document images demonstrate that our\nmethod improves the image quality compared with the observed image, and\nsimultaneously improves the compression ratio. For the test images with\nsynthetic noise, our method reduces the number of flipped pixels by 48.2% and\nimproves the compression ratio by 36.36% as compared with the best encoding\nmethods. For the test images with real noise, our method visually improves the\nimage quality, and outperforms the cutting-edge method by 28.27% in terms of\nthe compression ratio.', 'For lossy image compression, we develop a neural-based system which learns a\nnonlinear estimator for decoding from quantized representations. The system\nlinks two recurrent networks that \\help"" each other reconstruct same target\nimage patches using complementary portions of spatial context that communicate\nvia gradient signals. This dual agent system builds upon prior work that\nproposed the iterative refinement algorithm for recurrent neural network\n(RNN)based decoding which improved image reconstruction compared to standard\ndecoding techniques. Our approach, which works with any encoder, neural or\nnon-neural, This system progressively reduces image patch reconstruction error\nover a fixed number of steps. Experiment with variants of RNN memory cells,\nwith and without future information, find that our model consistently creates\nlower distortion images of higher perceptual quality compared to other\napproaches. Specifically, on the Kodak Lossless True Color Image Suite, we\nobserve as much as a 1:64 decibel (dB) gain over JPEG, a 1:46 dB gain over JPEG\n2000, a 1:34 dB gain over the GOOG neural baseline, 0:36 over E2E (a modern\ncompetitive neural compression model), and 0:37 over a single iterative neural\ndecoder.', 'Convolutional Neural Networks (CNNs) are state-of-the-art models for document\nimage classification tasks. However, many of these approaches rely on\nparameters and architectures designed for classifying natural images, which\ndiffer from document images. We question whether this is appropriate and\nconduct a large empirical study to find what aspects of CNNs most affect\nperformance on document images. Among other results, we exceed the\nstate-of-the-art on the RVL-CDIP dataset by using shear transform data\naugmentation and an architecture designed for a larger input image.\nAdditionally, we analyze the learned features and find evidence that CNNs\ntrained on RVL-CDIP learn region-specific layout features.', 'High Dynamic Range (HDR) content is becoming ubiquitous due to the rapid\ndevelopment of capture technologies. Nevertheless, the dynamic range of common\ndisplay devices is still limited, therefore tone mapping (TM) remains a key\nchallenge for image visualization. Recent work has demonstrated that neural\nnetworks can achieve remarkable performance in this task when compared to\ntraditional methods, however, the quality of the results of these\nlearning-based methods is limited by the training data. Most existing works use\nas training set a curated selection of best-performing results from existing\ntraditional tone mapping operators (often guided by a quality metric),\ntherefore, the quality of newly generated results is fundamentally limited by\nthe performance of such operators. This quality might be even further limited\nby the pool of HDR content that is used for training. In this work we propose a\nlearning-based self-supervised tone mapping operator that is trained at test\ntime specifically for each HDR image and does not need any data labeling. The\nkey novelty of our approach is a carefully designed loss function built upon\nfundamental knowledge on contrast perception that allows for directly comparing\nthe content in the HDR and tone mapped images. We achieve this goal by\nreformulating classic VGG feature maps into feature contrast maps that\nnormalize local feature differences by their average magnitude in a local\nneighborhood, allowing our loss to account for contrast masking effects. We\nperform extensive ablation studies and exploration of parameters and\ndemonstrate that our solution outperforms existing approaches with a single set\nof fixed parameters, as confirmed by both objective and subjective metrics.', 'Spiking neural networks (SNNs) offer an inherent ability to process\nspatial-temporal data, or in other words, realworld sensory data, but suffer\nfrom the difficulty of training high accuracy models. A major thread of\nresearch on SNNs is on converting a pre-trained convolutional neural network\n(CNN) to an SNN of the same structure. State-of-the-art conversion methods are\napproaching the accuracy limit, i.e., the near-zero accuracy loss of SNN\nagainst the original CNN. However, we note that this is made possible only when\nsignificantly more energy is consumed to process an input. In this paper, we\nargue that this trend of ""energy for accuracy"" is not necessary -- a little\nenergy can go a long way to achieve the near-zero accuracy loss. Specifically,\nwe propose a novel CNN-to-SNN conversion method that is able to use a\nreasonably short spike train (e.g., 256 timesteps for CIFAR10 images) to\nachieve the near-zero accuracy loss. The new conversion method, named as\nexplicit current control (ECC), contains three techniques (current\nnormalisation, thresholding for residual elimination, and consistency\nmaintenance for batch-normalisation), in order to explicitly control the\ncurrents flowing through the SNN when processing inputs. We implement ECC into\na tool nicknamed SpKeras, which can conveniently import Keras CNN models and\nconvert them into SNNs. We conduct an extensive set of experiments with the\ntool -- working with VGG16 and various datasets such as CIFAR10 and CIFAR100 --\nand compare with state-of-the-art conversion methods. Results show that ECC is\na promising method that can optimise over energy consumption and accuracy loss\nsimultaneously.', 'High dynamic range (HDR) imaging is a technique that allows an extensive\ndynamic range of exposures, which is important in image processing, computer\ngraphics, and computer vision. In recent years, there has been a significant\nadvancement in HDR imaging using deep learning (DL). This study conducts a\ncomprehensive and insightful survey and analysis of recent developments in deep\nHDR imaging methodologies. We hierarchically and structurally group existing\ndeep HDR imaging methods into five categories based on (1) number/domain of\ninput exposures, (2) number of learning tasks, (3) novel sensor data, (4) novel\nlearning strategies, and (5) applications. Importantly, we provide a\nconstructive discussion on each category regarding its potential and\nchallenges. Moreover, we review some crucial aspects of deep HDR imaging, such\nas datasets and evaluation metrics. Finally, we highlight some open problems\nand point out future research directions.', 'In this paper, we propose a deep snapshot high dynamic range (HDR) imaging\nframework that can effectively reconstruct an HDR image from the RAW data\ncaptured using a multi-exposure color filter array (ME-CFA), which consists of\na mosaic pattern of RGB filters with different exposure levels. To effectively\nlearn the HDR image reconstruction network, we introduce the idea of luminance\nnormalization that simultaneously enables effective loss computation and input\ndata normalization by considering relative local contrasts in the\n""normalized-by-luminance"" HDR domain. This idea makes it possible to equally\nhandle the errors in both bright and dark areas regardless of absolute\nluminance levels, which significantly improves the visual image quality in a\ntone-mapped domain. Experimental results using two public HDR image datasets\ndemonstrate that our framework outperforms other snapshot methods and produces\nhigh-quality HDR images with fewer visual artifacts.', 'Segmentation of additive manufacturing (AM) defects in X-ray Computed\nTomography (XCT) images is challenging, due to the poor contrast, small sizes\nand variation in appearance of defects. Automatic segmentation can, however,\nprovide quality control for additive manufacturing. Over recent years,\nthree-dimensional convolutional neural networks (3D CNNs) have performed well\nin the volumetric segmentation of medical images. In this work, we leverage\ntechniques from the medical imaging domain and propose training a 3D U-Net\nmodel to automatically segment defects in XCT images of AM samples. This work\nnot only contributes to the use of machine learning for AM defect detection but\nalso demonstrates for the first time 3D volumetric segmentation in AM. We train\nand test with three variants of the 3D U-Net on an AM dataset, achieving a mean\nintersection of union (IOU) value of 88.4%.', 'Recently, experiments have been reported where researchers were able to\nperform high dynamic range (HDR) tomography in a heuristic fashion, by fusing\nmultiple tomographic projections. This approach to HDR tomography has been\ninspired by HDR photography and inherits the same disadvantages. Taking a\ncomputational imaging approach to the HDR tomography problem, we here suggest a\nnew model based on the Modulo Radon Transform (MRT), which we rigorously\nintroduce and analyze. By harnessing a joint design between hardware and\nalgorithms, we present a single-shot HDR tomography approach, which to our\nknowledge, is the only approach that is backed by mathematical guarantees.\n  On the hardware front, instead of recording the Radon Transform projections\nthat my potentially saturate, we propose to measure modulo values of the same.\nThis ensures that the HDR measurements are folded into a lower dynamic range.\nOn the algorithmic front, our recovery algorithms reconstruct the HDR images\nfrom folded measurements. Beyond mathematical aspects such as injectivity and\ninversion of the MRT for different scenarios including band-limited and\napproximately compactly supported images, we also provide a first\nproof-of-concept demonstration. To do so, we implement MRT by experimentally\nfolding tomographic measurements available as an open source data set using our\ncustom designed modulo hardware. Our reconstruction clearly shows the\nadvantages of our approach for experimental data. In this way, our MRT based\nsolution paves a path for HDR acquisition in a number of related imaging\nproblems.', 'Image correction aims to adjust an input image into a visually pleasing one.\nExisting approaches are proposed mainly from the perspective of image pixel\nmanipulation. They are not effective to recover the details in the under/over\nexposed regions. In this paper, we revisit the image formation procedure and\nnotice that the missing details in these regions exist in the corresponding\nhigh dynamic range (HDR) data. These details are well perceived by the human\neyes but diminished in the low dynamic range (LDR) domain because of the tone\nmapping process. Therefore, we formulate the image correction task as an HDR\ntransformation process and propose a novel approach called Deep Reciprocating\nHDR Transformation (DRHT). Given an input LDR image, we first reconstruct the\nmissing details in the HDR domain. We then perform tone mapping on the\npredicted HDR data to generate the output LDR image with the recovered details.\nTo this end, we propose a united framework consisting of two CNNs for HDR\nreconstruction and tone mapping. They are integrated end-to-end for joint\ntraining and prediction. Experiments on the standard benchmarks demonstrate\nthat the proposed method performs favorably against state-of-the-art image\ncorrection methods.', 'Tone mapping is a commonly used technique that maps the set of colors in\nhigh-dynamic-range (HDR) images to another set of colors in low-dynamic-range\n(LDR) images, to fit the need for print-outs, LCD monitors and projectors.\nUnfortunately, during the compression of dynamic range, the overall contrast\nand local details generally cannot be preserved simultaneously. Recently, with\nthe increased use of stereoscopic devices, the notion of binocular tone mapping\nhas been proposed in the existing research study. However, the existing\nresearch lacks the binocular perception study and is unable to generate the\noptimal binocular pair that presents the most visual content. In this paper, we\npropose a novel perception-based binocular tone mapping method, that can\ngenerate an optimal binocular image pair (generating left and right images\nsimultaneously) from an HDR image that presents the most visual content by\ndesigning a binocular perception metric. Our method outperforms the existing\nmethod in terms of both visual and time performance.', ""Stack-based high dynamic range (HDR) imaging is a technique for achieving a\nlarger dynamic range in an image by combining several low dynamic range images\nacquired at different exposures. Minimizing the set of images to combine, while\nensuring that the resulting HDR image fully captures the scene's irradiance, is\nimportant to avoid long image acquisition and post-processing times. The\nproblem of selecting the set of images has received much attention. However,\nexisting methods either are not fully automatic, can be slow, or can fail to\nfully capture more challenging scenes. In this paper, we propose a fully\nautomatic method for selecting the set of exposures to acquire that is both\nfast and more accurate. We show on an extensive set of benchmark scenes that\nour proposed method leads to improved HDR images as measured against ground\ntruth using the mean squared error, a pixel-based metric, and a visible\ndifference predictor and a quality score, both perception-based metrics."", 'High dynamic range (HDR) photography is becoming increasingly popular and\navailable by DSLR and mobile-phone cameras. While deep neural networks (DNN)\nhave greatly impacted other domains of image manipulation, their use for HDR\ntone-mapping is limited due to the lack of a definite notion of ground-truth\nsolution, which is needed for producing training data.\n  In this paper we describe a new tone-mapping approach guided by the distinct\ngoal of producing low dynamic range (LDR) renditions that best reproduce the\nvisual characteristics of native LDR images. This goal enables the use of an\nunpaired adversarial training based on unrelated sets of HDR and LDR images,\nboth of which are widely available and easy to acquire.\n  In order to achieve an effective training under this minimal requirements, we\nintroduce the following new steps and components: (i) a range-normalizing\npre-process which estimates and applies a different level of curve-based\ncompression, (ii) a loss that preserves the input content while allowing the\nnetwork to achieve its goal, and (iii) the use of a more concise discriminator\nnetwork, designed to promote the reproduction of low-level attributes native\nLDR possess.\n  Evaluation of the resulting network demonstrates its ability to produce\nphoto-realistic artifact-free tone-mapped images, and state-of-the-art\nperformance on different image fidelity indices and visual distances.', 'We proposed a lossless two-layer HDR coding method using a histogram packing\ntechnique. The proposed method was demonstrated to outperform the normative\nJPEG XT encoder, under the use of the default tone-mapping operator. However,\nthe performance under various tone-mapping operators has not been discussed. In\nthis paper, we aim to compare the performance of the proposed method with that\nof the JPEG XT encoder under the use of various tone-mapping operators to\nclearly show the characteristic difference between them.', 'With the development of human communications the usage of Visual\nCommunications has also increased. The advancement of image compression methods\nis one of the main reasons for the enhancement. This paper first presents main\nmodes of image compression methods such as JPEG and JPEG2000 without\nmathematical details. Also, the paper describes gradient Haar wavelet\ntransforms in order to construct a preliminary image compression algorithm.\nThen, a new image compression method is proposed based on the preliminary image\ncompression algorithm that can improve standards of image compression. The new\nmethod is compared with original modes of JPEG and JPEG2000 (based on Haar\nwavelet) by image quality measures such as MAE, PSNAR, and SSIM. The image\nquality and statistical results confirm that can boost image compression\nstandards. It is suggested that the new method is used in a part or all of an\nimage compression standard.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'This paper study the reconstruction of High Dynamic Range (HDR) video from\nsnapshot-coded LDR video. Constructing an HDR video requires restoring the HDR\nvalues for each frame and maintaining the consistency between successive\nframes. HDR image acquisition from single image capture, also known as snapshot\nHDR imaging, can be achieved in several ways. For example, the reconfigurable\nsnapshot HDR camera is realized by introducing an optical element into the\noptical stack of the camera; by placing a coded mask at a small standoff\ndistance in front of the sensor. High-quality HDR image can be recovered from\nthe captured coded image using deep learning methods. This study utilizes\n3D-CNNs to perform a joint demosaicking, denoising, and HDR video\nreconstruction from coded LDR video. We enforce more temporally consistent HDR\nvideo reconstruction by introducing a temporal loss function that considers the\nshort-term and long-term consistency. The obtained results are promising and\ncould lead to affordable HDR video capture using conventional cameras.', 'Since the low quality of document images will greatly undermine the chances\nof success in automatic text recognition and analysis, it is necessary to\nassess the quality of document images uploaded in online business process, so\nas to reject those images of low quality. In this paper, we attempt to achieve\ndocument image quality assessment and our contributions are twofold. Firstly,\nsince document image quality assessment is more interested in text, we propose\na text line based framework to estimate document image quality, which is\ncomposed of three stages: text line detection, text line quality prediction,\nand overall quality assessment. Text line detection aims to find potential text\nlines with a detector. In the text line quality prediction stage, the quality\nscore is computed for each text line with a CNN-based prediction model. The\noverall quality of document images is finally assessed with the ensemble of all\ntext line quality. Secondly, to train the prediction model, a large-scale\ndataset, comprising 52,094 text line images, is synthesized with diverse\nattributes. For each text line image, a quality label is computed with a\npiece-wise function. To demonstrate the effectiveness of the proposed\nframework, comprehensive experiments are evaluated on two popular document\nimage quality assessment benchmarks. Our framework significantly outperforms\nthe state-of-the-art methods by large margins on the large and complicated\ndataset.', 'Sparse decomposition has been extensively used for different applications\nincluding signal compression and denoising and document analysis. In this\npaper, sparse decomposition is used for image segmentation. The proposed\nalgorithm separates the background and foreground using a sparse-smooth\ndecomposition technique such that the smooth and sparse components correspond\nto the background and foreground respectively. This algorithm is tested on\nseveral test images from HEVC test sequences and is shown to have superior\nperformance over other methods, such as the hierarchical k-means clustering in\nDjVu. This segmentation algorithm can also be used for text extraction, video\ncompression and medical image segmentation.', 'The standard JPEG format is almost the optimum format in image compression.\nThe compression ratio in JPEG sometimes reaches 30:1. The compression ratio of\nJPEG could be increased by embedding the Five Modulus Method (FMM) into the\nJPEG algorithm. The novel algorithm gives twice the time as the standard JPEG\nalgorithm or more. The novel algorithm was called FJPEG (Five-JPEG). The\nquality of the reconstructed image after compression is approximately\napproaches the JPEG. Standard test images have been used to support and\nimplement the suggested idea in this paper and the error metrics have been\ncomputed and compared with JPEG.', ""Recently, high dynamic range (HDR) image reconstruction based on the multiple\nexposure stack from a given single exposure utilizes a deep learning framework\nto generate high-quality HDR images. These conventional networks focus on the\nexposure transfer task to reconstruct the multi-exposure stack. Therefore, they\noften fail to fuse the multi-exposure stack into a perceptually pleasant HDR\nimage as the inversion artifacts occur. We tackle the problem in stack\nreconstruction-based methods by proposing a novel framework with a fully\ndifferentiable high dynamic range imaging (HDRI) process. By explicitly using\nthe loss, which compares the network's output with the ground truth HDR image,\nour framework enables a neural network that generates the multiple exposure\nstack for HDRI to train stably. In other words, our differentiable HDR\nsynthesis layer helps the deep neural network to train to create multi-exposure\nstacks while reflecting the precise correlations between multi-exposure images\nin the HDRI process. In addition, our network uses the image decomposition and\nthe recursive process to facilitate the exposure transfer task and to\nadaptively respond to recursion frequency. The experimental results show that\nthe proposed network outperforms the state-of-the-art quantitative and\nqualitative results in terms of both the exposure transfer tasks and the whole\nHDRI process."", 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Digital camera and mobile document image acquisition are new trends arising\nin the world of Optical Character Recognition and text detection. In some\ncases, such process integrates many distortions and produces poorly scanned\ntext or text-photo images and natural images, leading to an unreliable OCR\ndigitization. In this paper, we present a novel nonparametric and unsupervised\nmethod to compensate for undesirable document image distortions aiming to\noptimally improve OCR accuracy. Our approach relies on a very efficient stack\nof document image enhancing techniques to recover deformation of the entire\ndocument image. First, we propose a local brightness and contrast adjustment\nmethod to effectively handle lighting variations and the irregular distribution\nof image illumination. Second, we use an optimized greyscale conversion\nalgorithm to transform our document image to greyscale level. Third, we sharpen\nthe useful information in the resulting greyscale image using Un-sharp Masking\nmethod. Finally, an optimal global binarization approach is used to prepare the\nfinal document image to OCR recognition. The proposed approach can\nsignificantly improve text detection rate and optical character recognition\naccuracy. To demonstrate the efficiency of our approach, an exhaustive\nexperimentation on a standard dataset is presented.', 'Estimating the primary quantization matrix of double JPEG compressed images\nis a problem of relevant importance in image forensics since it allows to infer\nimportant information about the past history of an image. In addition, the\ninconsistencies of the primary quantization matrices across different image\nregions can be used to localize splicing in double JPEG tampered images.\nTraditional model-based approaches work under specific assumptions on the\nrelationship between the first and second compression qualities and on the\nalignment of the JPEG grid. Recently, a deep learning-based estimator capable\nto work under a wide variety of conditions has been proposed, that outperforms\ntailored existing methods in most of the cases. The method is based on a\nConvolutional Neural Network (CNN) that is trained to solve the estimation as a\nstandard regression problem. By exploiting the integer nature of the\nquantization coefficients, in this paper, we propose a deep learning technique\nthat performs the estimation by resorting to a simil-classification\narchitecture. The CNN is trained with a loss function that takes into account\nboth the accuracy and the Mean Square Error (MSE) of the estimation. Results\nconfirm the superior performance of the proposed technique, compared to the\nstate-of-the art methods based on statistical analysis and, in particular, deep\nlearning regression. Moreover, the capability of the method to work under\ngeneral operative conditions, regarding the alignment of the second compression\ngrid with the one of first compression and the combinations of the JPEG\nqualities of former and second compression, is very relevant in practical\napplications, where these information are unknown a priori.', 'Machine learning is penetrating various domains virtually, thereby\nproliferating excellent results. It has also found an outlet in digital\nforensics, wherein it is becoming the prime driver of computational efficiency.\nA prominent feature that exhibits the effectiveness of ML algorithms is feature\nextraction that can be instrumental in the applications for digital forensics.\nConvolutional Neural Networks are further used to identify parts of the file.\nTo this end, we observed that the literature does not include sufficient\ninformation about the identification of the algorithms used to compress file\nfragments. With this research, we attempt to address this gap as compression\nalgorithms are beneficial in generating higher entropy comparatively as they\nmake the data more compact. We used a base dataset, compressed every file with\nvarious algorithms, and designed a model based on that. The used model was\naccurately able to identify files compressed using compress, lzip and bzip2.', 'Document image quality assessment (DIQA) is an important and challenging\nproblem in real applications. In order to predict the quality scores of\ndocument images, this paper proposes a novel no-reference DIQA method based on\ncharacter gradient, where the OCR accuracy is used as a ground-truth quality\nmetric. Character gradient is computed on character patches detected with the\nmaximally stable extremal regions (MSER) based method. Character patches are\nessentially significant to character recognition and therefore suitable for use\nin estimating document image quality. Experiments on a benchmark dataset show\nthat the proposed method outperforms the state-of-the-art methods in estimating\nthe quality score of document images.', 'Recent improvements to Generative Adversarial Networks (GANs) have made it\npossible to generate realistic images in high resolution based on natural\nlanguage descriptions such as image captions. Furthermore, conditional GANs\nallow us to control the image generation process through labels or even natural\nlanguage descriptions. However, fine-grained control of the image layout, i.e.\nwhere in the image specific objects should be located, is still difficult to\nachieve. This is especially true for images that should contain multiple\ndistinct objects at different spatial locations. We introduce a new approach\nwhich allows us to control the location of arbitrarily many objects within an\nimage by adding an object pathway to both the generator and the discriminator.\nOur approach does not need a detailed semantic layout but only bounding boxes\nand the respective labels of the desired objects are needed. The object pathway\nfocuses solely on the individual objects and is iteratively applied at the\nlocations specified by the bounding boxes. The global pathway focuses on the\nimage background and the general image layout. We perform experiments on the\nMulti-MNIST, CLEVR, and the more complex MS-COCO data set. Our experiments show\nthat through the use of the object pathway we can control object locations\nwithin images and can model complex scenes with multiple objects at various\nlocations. We further show that the object pathway focuses on the individual\nobjects and learns features relevant for these, while the global pathway\nfocuses on global image characteristics and the image background.', 'Despite significant progress on current state-of-the-art image generation\nmodels, synthesis of document images containing multiple and complex object\nlayouts is a challenging task. This paper presents a novel approach, called\nDocSynth, to automatically synthesize document images based on a given layout.\nIn this work, given a spatial layout (bounding boxes with object categories) as\na reference by the user, our proposed DocSynth model learns to generate a set\nof realistic document images consistent with the defined layout. Also, this\nframework has been adapted to this work as a superior baseline model for\ncreating synthetic document image datasets for augmenting real data during\ntraining for document layout analysis tasks. Different sets of learning\nobjectives have been also used to improve the model performance.\nQuantitatively, we also compare the generated results of our model with real\ndata using standard evaluation metrics. The results highlight that our model\ncan successfully generate realistic and diverse document images with multiple\nobjects. We also present a comprehensive qualitative analysis summary of the\ndifferent scopes of synthetic image generation tasks. Lastly, to our knowledge\nthis is the first work of its kind.', 'Documents often exhibit various forms of degradation, which make it hard to\nbe read and substantially deteriorate the performance of an OCR system. In this\npaper, we propose an effective end-to-end framework named Document Enhancement\nGenerative Adversarial Networks (DE-GAN) that uses the conditional GANs (cGANs)\nto restore severely degraded document images. To the best of our knowledge,\nthis practice has not been studied within the context of generative adversarial\ndeep networks. We demonstrate that, in different tasks (document clean up,\nbinarization, deblurring and watermark removal), DE-GAN can produce an enhanced\nversion of the degraded document with a high quality. In addition, our approach\nprovides consistent improvements compared to state-of-the-art methods over the\nwidely used DIBCO 2013, DIBCO 2017 and H-DIBCO 2018 datasets, proving its\nability to restore a degraded document image to its ideal condition. The\nobtained results on a wide variety of degradation reveal the flexibility of the\nproposed model to be exploited in other document enhancement problems.', 'This work analyses the impact of self-supervised pre-training on document\nimages in the context of document image classification. While previous\napproaches explore the effect of self-supervision on natural images, we show\nthat patch-based pre-training performs poorly on document images because of\ntheir different structural properties and poor intra-sample semantic\ninformation. We propose two context-aware alternatives to improve performance\non the Tobacco-3482 image classification task. We also propose a novel method\nfor self-supervision, which makes use of the inherent multi-modality of\ndocuments (image and text), which performs better than other popular\nself-supervised methods, including supervised ImageNet pre-training, on\ndocument image classification scenarios with a limited amount of data.', 'Computation of document image quality metrics often depends upon the\navailability of a ground truth image corresponding to the document. This limits\nthe applicability of quality metrics in applications such as hyperparameter\noptimization of image processing algorithms that operate on-the-fly on unseen\ndocuments. This work proposes the use of surrogate models to learn the behavior\nof a given document quality metric on existing datasets where ground truth\nimages are available. The trained surrogate model can later be used to predict\nthe metric value on previously unseen document images without requiring access\nto ground truth images. The surrogate model is empirically evaluated on the\nDocument Image Binarization Competition (DIBCO) and the Handwritten Document\nImage Binarization Competition (H-DIBCO) datasets.', 'Recognition of document images have important applications in restoring old\nand classical texts. The problem involves quality improvement before passing it\nto a properly trained OCR to get accurate recognition of the text. The image\nenhancement and quality improvement constitute important steps as subsequent\nrecognition depends upon the quality of the input image. There are scenarios\nwhen high resolution images are not available and our experiments show that the\nOCR accuracy reduces significantly with decrease in the spatial resolution of\ndocument images. Thus the only option is to improve the resolution of such\ndocument images. The goal is to construct a high resolution image, given a\nsingle low resolution binary image, which constitutes the problem of single\nimage super-resolution. Most of the previous work in super-resolution deal with\nnatural images which have more information-content than the document images.\nHere, we use Convolution Neural Network to learn the mapping between low and\nthe corresponding high resolution images. We experiment with different number\nof layers, parameter settings and non-linear functions to build a fast\nend-to-end framework for document image super-resolution. Our proposed model\nshows a very good PSNR improvement of about 4 dB on 75 dpi Tamil images,\nresulting in a 3 % improvement of word level accuracy by the OCR. It takes less\ntime than the recent sparse based natural image super-resolution technique,\nmaking it useful for real-time document recognition applications.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'Line separators are used to segregate text-lines from one another in document\nimage analysis. Finding the separator points at every line terminal in a\ndocument image would enable text-line segmentation. In particular, identifying\nthe separators in handwritten text could be a thrilling exercise. Obviously it\nwould be challenging to perform this in the compressed version of a document\nimage and that is the proposed objective in this research. Such an effort would\nprevent the computational burden of decompressing a document for text-line\nsegmentation. Since document images are generally compressed using run length\nencoding (RLE) technique as per the CCITT standards, the first column in the\nRLE will be a white column. The value (depth) in the white column is very low\nwhen a particular line is a text line and the depth could be larger at the\npoint of text line separation. A longer consecutive sequence of such larger\ndepth should indicate the gap between the text lines, which provides the\nseparator region. In case of over separation and under separation issues,\ncorrective actions such as deletion and insertion are suggested respectively.\nAn extensive experimentation is conducted on the compressed images of the\nbenchmark datasets of ICDAR13 and Alireza et al [17] to demonstrate the\nefficacy.', 'Document image binarization is often a challenging task due to various forms\nof degradation. Although there exist several binarization techniques in\nliterature, the binarized image is typically sensitive to control parameter\nsettings of the employed technique. This paper presents an automatic document\nimage binarization algorithm to segment the text from heavily degraded document\nimages. The proposed technique uses a two band-pass filtering approach for\nbackground noise removal, and Bayesian optimization for automatic\nhyperparameter selection for optimal results. The effectiveness of the proposed\nbinarization technique is empirically demonstrated on the Document Image\nBinarization Competition (DIBCO) and the Handwritten Document Image\nBinarization Competition (H-DIBCO) datasets.', 'While the importance of automatic image analysis is continuously increasing,\nrecent meta-research revealed major flaws with respect to algorithm validation.\nPerformance metrics are particularly key for meaningful, objective, and\ntransparent performance assessment and validation of the used automatic\nalgorithms, but relatively little attention has been given to the practical\npitfalls when using specific metrics for a given image analysis task. These are\ntypically related to (1) the disregard of inherent metric properties, such as\nthe behaviour in the presence of class imbalance or small target structures,\n(2) the disregard of inherent data set properties, such as the non-independence\nof the test cases, and (3) the disregard of the actual biomedical domain\ninterest that the metrics should reflect. This living dynamically document has\nthe purpose to illustrate important limitations of performance metrics commonly\napplied in the field of image analysis. In this context, it focuses on\nbiomedical image analysis problems that can be phrased as image-level\nclassification, semantic segmentation, instance segmentation, or object\ndetection task. The current version is based on a Delphi process on metrics\nconducted by an international consortium of image analysis experts from more\nthan 60 institutions worldwide.', 'The digital camera captured document images may often be warped and distorted\ndue to different camera angles or document surfaces. A robust technique is\nneeded to solve this kind of distortion. The research on dewarping of the\ndocument suffers due to the limited availability of benchmark public dataset.\nIn recent times, deep learning based approaches are used to solve the problems\naccurately. To train most of the deep neural networks a large number of\ndocument images is required and generating such a large volume of document\nimages manually is difficult. In this paper, we propose a technique to generate\na synthetic warped image from a flat-bedded scanned document image. It is done\nby calculating warping factors for each pixel position using two warping\nposition parameters (WPP) and eight warping control parameters (WCP). These\nparameters can be specified as needed depending upon the desired warping. The\nresults are compared with similar real captured images both qualitative and\nquantitative way.', 'Single-image high dynamic range (SI-HDR) reconstruction has recently emerged\nas a problem well-suited for deep learning methods. Each successive technique\ndemonstrates an improvement over existing methods by reporting higher image\nquality scores. This paper, however, highlights that such improvements in\nobjective metrics do not necessarily translate to visually superior images. The\nfirst problem is the use of disparate evaluation conditions in terms of data\nand metric parameters, calling for a standardized protocol to make it possible\nto compare between papers. The second problem, which forms the main focus of\nthis paper, is the inherent difficulty in evaluating SI-HDR reconstructions\nsince certain aspects of the reconstruction problem dominate objective\ndifferences, thereby introducing a bias. Here, we reproduce a typical\nevaluation using existing as well as simulated SI-HDR methods to demonstrate\nhow different aspects of the problem affect objective quality metrics.\nSurprisingly, we found that methods that do not even reconstruct HDR\ninformation can compete with state-of-the-art deep learning methods. We show\nhow such results are not representative of the perceived quality and that\nSI-HDR reconstruction needs better evaluation protocols.', 'We introduce the Brno Mobile OCR Dataset (B-MOD) for document Optical\nCharacter Recognition from low-quality images captured by handheld mobile\ndevices. While OCR of high-quality scanned documents is a mature field where\nmany commercial tools are available, and large datasets of text in the wild\nexist, no existing datasets can be used to develop and test document OCR\nmethods robust to non-uniform lighting, image blur, strong noise, built-in\ndenoising, sharpening, compression and other artifacts present in many\nphotographs from mobile devices.\n  This dataset contains 2 113 unique pages from random scientific papers, which\nwere photographed by multiple people using 23 different mobile devices. The\nresulting 19 728 photographs of various visual quality are accompanied by\nprecise positions and text annotations of 500k text lines. We further provide\nan evaluation methodology, including an evaluation server and a testset with\nnon-public annotations.\n  We provide a state-of-the-art text recognition baseline build on\nconvolutional and recurrent neural networks trained with Connectionist Temporal\nClassification loss. This baseline achieves 2 %, 22 % and 73 % word error rates\non easy, medium and hard parts of the dataset, respectively, confirming that\nthe dataset is challenging.\n  The presented dataset will enable future development and evaluation of\ndocument analysis for low-quality images. It is primarily intended for\nline-level text recognition, and can be further used for line localization,\nlayout analysis, image restoration and text binarization.', ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", ""This paper presents a novel iterative deep learning framework and apply it\nfor document enhancement and binarization. Unlike the traditional methods which\npredict the binary label of each pixel on the input image, we train the neural\nnetwork to learn the degradations in document images and produce the uniform\nimages of the degraded input images, which allows the network to refine the\noutput iteratively. Two different iterative methods have been studied in this\npaper: recurrent refinement (RR) which uses the same trained neural network in\neach iteration for document enhancement and stacked refinement (SR) which uses\na stack of different neural networks for iterative output refinement. Given the\nlearned uniform and enhanced image, the binarization map can be easy to obtain\nby a global or local threshold. The experimental results on several public\nbenchmark data sets show that our proposed methods provide a new clean version\nof the degraded image which is suitable for visualization and promising results\nof binarization using the global Otsu's threshold based on the enhanced images\nlearned iteratively by the neural network."", 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.', 'We propose an automatic method to infer high dynamic range illumination from\na single, limited field-of-view, low dynamic range photograph of an indoor\nscene. In contrast to previous work that relies on specialized image capture,\nuser input, and/or simple scene models, we train an end-to-end deep neural\nnetwork that directly regresses a limited field-of-view photo to HDR\nillumination, without strong assumptions on scene geometry, material\nproperties, or lighting. We show that this can be accomplished in a three step\nprocess: 1) we train a robust lighting classifier to automatically annotate the\nlocation of light sources in a large dataset of LDR environment maps, 2) we use\nthese annotations to train a deep neural network that predicts the location of\nlights in a scene from a single limited field-of-view photo, and 3) we\nfine-tune this network using a small dataset of HDR environment maps to predict\nlight intensities. This allows us to automatically recover high-quality HDR\nillumination estimates that significantly outperform previous state-of-the-art\nmethods. Consequently, using our illumination estimates for applications like\n3D object insertion, we can achieve results that are photo-realistic, which is\nvalidated via a perceptual user study.', 'In this work, we adhere to explore a Multi-Tasking learning (MTL) based\nnetwork to perform document attribute classification such as the font type,\nfont size, font emphasis and scanning resolution classification of a document\nimage. To accomplish these tasks, we operate on either segmented word level or\non uniformed size patches randomly cropped out of the document. Furthermore, a\nhybrid convolution neural network (CNN) architecture ""MTL+MI"", which is based\non the combination of MTL and Multi-Instance (MI) of patch and word is used to\naccomplish joint learning for the classification of the same document\nattributes. The contribution of this paper are three fold: firstly, based on\nsegmented word images and patches, we present a MTL based network for the\nclassification of a full document image. Secondly, we propose a MTL and MI\n(using segmented words and patches) based combined CNN architecture (""MTL+MI"")\nfor the classification of same document attributes. Thirdly, based on the\nmulti-tasking classifications of the words and/or patches, we propose an\nintelligent voting system which is based on the posterior probabilities of each\nwords and/or patches to perform the classification of document\'s attributes of\ncomplete document image.', 'The demand for on-device document recognition systems increases in\nconjunction with the emergence of more strict privacy and security\nrequirements. In such systems, there is no data transfer from the end device to\na third-party information processing servers. The response time is vital to the\nuser experience of on-device document recognition. Combined with the\nunavailability of discrete GPUs, powerful CPUs, or a large RAM capacity on\nconsumer-grade end devices such as smartphones, the time limitations put\nsignificant constraints on the computational complexity of the applied\nalgorithms for on-device execution.\n  In this work, we consider document location in an image without prior\nknowledge of the document content or its internal structure. In accordance with\nthe published works, at least 5 systems offer solutions for on-device document\nlocation. All these systems use a location method which can be considered\nHough-based. The precision of such systems seems to be lower than that of the\nstate-of-the-art solutions which were not designed to account for the limited\ncomputational resources.\n  We propose an advanced Hough-based method. In contrast with other approaches,\nit accounts for the geometric invariants of the central projection model and\ncombines both edge and color features for document boundary detection. The\nproposed method allowed for the second best result for SmartDoc dataset in\nterms of precision, surpassed by U-net like neural network. When evaluated on a\nmore challenging MIDV-500 dataset, the proposed algorithm guaranteed the best\nprecision compared to published methods. Our method retained the applicability\nto on-device computations.', 'We present a framework for learning an efficient holistic representation for\nhandwritten word images. The proposed method uses a deep convolutional neural\nnetwork with traditional classification loss. The major strengths of our work\nlie in: (i) the efficient usage of synthetic data to pre-train a deep network,\n(ii) an adapted version of the ResNet-34 architecture with the region of\ninterest pooling (referred to as HWNet v2) which learns discriminative features\nfor variable sized word images, and (iii) a realistic augmentation of training\ndata with multiple scales and distortions which mimics the natural process of\nhandwriting. We further investigate the process of transfer learning to reduce\nthe domain gap between synthetic and real domain, and also analyze the\ninvariances learned at different layers of the network using visualization\ntechniques proposed in the literature.\n  Our representation leads to a state-of-the-art word spotting performance on\nstandard handwritten datasets and historical manuscripts in different languages\nwith minimal representation size. On the challenging IAM dataset, our method is\nfirst to report an mAP of around 0.90 for word spotting with a representation\nsize of just 32 dimensions. Furthermore, we also present results on printed\ndocument datasets in English and Indic scripts which validates the generic\nnature of the proposed framework for learning word image representation.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', 'Graphical elements: particularly tables and figures contain a visual summary\nof the most valuable information contained in a document. Therefore,\nlocalization of such graphical objects in the document images is the initial\nstep to understand the content of such graphical objects or document images. In\nthis paper, we present a novel end-to-end trainable deep learning based\nframework to localize graphical objects in the document images called as\nGraphical Object Detection (GOD). Our framework is data-driven and does not\nrequire any heuristics or meta-data to locate graphical objects in the document\nimages. The GOD explores the concept of transfer learning and domain adaptation\nto handle scarcity of labeled training images for graphical object detection\ntask in the document images. Performance analysis carried out on the various\npublic benchmark data sets: ICDAR-2013, ICDAR-POD2017,and UNLV shows that our\nmodel yields promising results as compared to state-of-the-art techniques.', 'Identity documents recognition is an important sub-field of document\nanalysis, which deals with tasks of robust document detection, type\nidentification, text fields recognition, as well as identity fraud prevention\nand document authenticity validation given photos, scans, or video frames of an\nidentity document capture. Significant amount of research has been published on\nthis topic in recent years, however a chief difficulty for such research is\nscarcity of datasets, due to the subject matter being protected by security\nrequirements. A few datasets of identity documents which are available lack\ndiversity of document types, capturing conditions, or variability of document\nfield values. In addition, the published datasets were typically designed only\nfor a subset of document recognition problems, not for a complex identity\ndocument analysis. In this paper, we present a dataset MIDV-2020 which consists\nof 1000 video clips, 2000 scanned images, and 1000 photos of 1000 unique mock\nidentity documents, each with unique text field values and unique artificially\ngenerated faces, with rich annotation. For the presented benchmark dataset\nbaselines are provided for such tasks as document location and identification,\ntext fields recognition, and face detection. With 72409 annotated images in\ntotal, to the date of publication the proposed dataset is the largest publicly\navailable identity documents dataset with variable artificially generated data,\nand we believe that it will prove invaluable for advancement of the field of\ndocument analysis and recognition. The dataset is available for download at\nftp://smartengines.com/midv-2020 and http://l3i-share.univ-lr.fr .', 'Today all kind of information is getting digitized and along with all this\ndigitization, the huge archive of various kinds of documents is being digitized\ntoo. We know that, Optical Character Recognition is the method through which,\nnewspapers and other paper documents convert into digital resources. But, it is\na fact that this method works on texts only. As a result, if we try to process\nany document which contains non-textual zones, then we will get garbage texts\nas output. That is why; in order to digitize documents properly they should be\nprepossessed carefully. And while preprocessing, segmenting document in\ndifferent regions according to the category properly is most important. But,\nthe Optical Character Recognition processes available for Bangla language have\nno such algorithm that can categorize a newspaper/book page fully. So we worked\nto decompose a document into its several parts like headlines, sub headlines,\ncolumns, images etc. And if the input is skewed and rotated, then the input was\nalso deskewed and de-rotated. To decompose any Bangla document we found out the\nedges of the input image. Then we find out the horizontal and vertical area of\nevery pixel where it lies in. Later on the input image was cut according to\nthese areas. Then we pick each and every sub image and found out their\nheight-width ratio, line height. Then according to these values the sub images\nwere categorized. To deskew the image we found out the skew angle and de skewed\nthe image according to this angle. To de-rotate the image we used the line\nheight, matra line, pixel ratio of matra line.', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', 'This paper presents results of Document Visual Question Answering Challenge\norganized as part of ""Text and Documents in the Deep Learning Era"" workshop, in\nCVPR 2020. The challenge introduces a new problem - Visual Question Answering\non document images. The challenge comprised two tasks. The first task concerns\nwith asking questions on a single document image. On the other hand, the second\ntask is set as a retrieval task where the question is posed over a collection\nof images. For the task 1 a new dataset is introduced comprising 50,000\nquestions-answer(s) pairs defined over 12,767 document images. For task 2\nanother dataset has been created comprising 20 questions over 14,362 document\nimages which share the same document template.', 'We introduce a novel type of text representation that preserves the 2D layout\nof a document. This is achieved by encoding each document page as a\ntwo-dimensional grid of characters. Based on this representation, we present a\ngeneric document understanding pipeline for structured documents. This pipeline\nmakes use of a fully convolutional encoder-decoder network that predicts a\nsegmentation mask and bounding boxes. We demonstrate its capabilities on an\ninformation extraction task from invoices and show that it significantly\noutperforms approaches based on sequential text or document images.', 'We introduce FontCode, an information embedding technique for text documents.\nProvided a text document with specific fonts, our method embeds user-specified\ninformation in the text by perturbing the glyphs of text characters while\npreserving the text content. We devise an algorithm to chooses unobtrusive yet\nmachine-recognizable glyph perturbations, leveraging a recently developed\ngenerative model that alters the glyphs of each character continuously on a\nfont manifold. We then introduce an algorithm that embeds a user-provided\nmessage in the text document and produces an encoded document whose appearance\nis minimally perturbed from the original document. We also present a glyph\nrecognition method that recovers the embedded information from an encoded\ndocument stored as a vector graphic or pixel image, or even on a printed paper.\nIn addition, we introduce a new error-correction coding scheme that rectifies a\ncertain number of recognition errors. Lastly, we demonstrate that our technique\nenables a wide array of applications, using it as a text document metadata\nholder, an unobtrusive optical barcode, a cryptographic message embedding\nscheme, and a text document signature.', 'When digitizing a document into an image, it is common to include a\nsurrounding border region to visually indicate that the entire document is\npresent in the image. However, this border should be removed prior to automated\nprocessing. In this work, we present a deep learning based system, PageNet,\nwhich identifies the main page region in an image in order to segment content\nfrom both textual and non-textual border noise. In PageNet, a Fully\nConvolutional Network obtains a pixel-wise segmentation which is post-processed\ninto the output quadrilateral region. We evaluate PageNet on 4 collections of\nhistorical handwritten documents and obtain over 94% mean intersection over\nunion on all datasets and approach human performance on 2 of these collections.\nAdditionally, we show that PageNet can segment documents that are overlayed on\ntop of other documents.', 'Document segmentation is a method of rending the document into distinct\nregions. A document is an assortment of information and a standard mode of\nconveying information to others. Pursuance of data from documents involves ton\nof human effort, time intense and might severely prohibit the usage of data\nsystems. So, automatic information pursuance from the document has become a big\nissue. It is been shown that document segmentation will facilitate to beat such\nproblems. This paper proposes a new approach to segment and classify the\ndocument regions as text, image, drawings and table. Document image is divided\ninto blocks using Run length smearing rule and features are extracted from\nevery blocks. Discipulus tool has been used to construct the Genetic\nprogramming based classifier model and located 97.5% classification accuracy.', 'This report explores the latest advances in the field of digital document\nrecognition. With the focus on printed document imagery, we discuss the major\ndevelopments in optical character recognition (OCR) and document image\nenhancement/restoration in application to Latin and non-Latin scripts. In\naddition, we review and discuss the available technologies for hand-written\ndocument recognition. In this report, we also provide some company-accumulated\nbenchmark results on available OCR engines.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently working on a project that involves recognizing text in images, utilizing the attention-based encoder-decoder framework. However, I am encountering difficulties when the system misinterprets words with extra or missing characters in the text, which complicates the training process and reduces accuracy. I am in search of a model capable of assigning probabilities to words within the images, specifically regarding their likelihood of containing missing or superfluous characters. Additionally, I require techniques that can be integrated into the existing framework. Ideally, these techniques would enable the framework to identify potential spelling errors and attempt to restore the original, correct word. I anticipate that the implementation of such a model and techniques will enhance the performance of my text recognition framework. To test this, I am also seeking datasets that contain spelling errors. I would appreciate any guidance on how to effectively train the model with these datasets and how to evaluate the models’ performance in a reliable way.","['Scanned documents in electronic health records (EHR) have been a challenge\nfor decades, and are expected to stay in the foreseeable future. Current\napproaches for processing often include image preprocessing, optical character\nrecognition (OCR), and text mining. However, there is limited work that\nevaluates the choice of image preprocessing methods, the selection of NLP\nmodels, and the role of document layout. The impact of each element remains\nunknown. We evaluated this method on a use case of two key indicators for sleep\napnea, Apnea hypopnea index (AHI) and oxygen saturation (SaO2) values, from\nscanned sleep study reports. Our data that included 955 manually annotated\nreports was secondarily utilized from a previous study in the University of\nTexas Medical Branch. We performed image preprocessing: gray-scaling followed\nby 1 iteration of dilating and erode, and 20% contrast increasing. The OCR was\nimplemented with the Tesseract OCR engine. A total of seven Bag-of-Words models\n(Logistic Regression, Ridge Regression, Lasso Regression, Support Vector\nMachine, k-Nearest Neighbor, Na\\""ive Bayes, and Random Forest) and three deep\nlearning-based models (BiLSTM, BERT, and Clinical BERT) were evaluated. We also\nevaluated the combinations of image preprocessing methods (gray-scaling, dilate\n& erode, increased contrast by 20%, increased contrast by 60%), and two deep\nlearning architectures (with and without structured input that provides\ndocument layout information). Our proposed method using Clinical BERT reached\nan AUROC of 0.9743 and document accuracy of 94.76% for AHI, and an AUROC of\n0.9523, and document accuracy of 91.61% for SaO2. We demonstrated the proper\nuse of image preprocessing and document layout could be beneficial to scanned\ndocument processing.']","['State-of-the-art text spotting systems typically aim to detect isolated words\nor word-by-word text in images of natural scenes and ignore the semantic\ncoherence within a region of text. However, when interpreted together,\nseemingly isolated words may be easier to recognize. On this basis, we propose\na novel ""semantic-based text recognition"" (STR) deep learning model that reads\ntext in images with the help of understanding context. STR consists of several\nmodules. We introduce the Text Grouping and Arranging (TGA) algorithm to\nconnect and order isolated text regions. A text-recognition network interprets\nisolated words. Benefiting from semantic information, a sequenceto-sequence\nnetwork model efficiently corrects inaccurate and uncertain phrases produced\nearlier in the STR pipeline. We present experiments on two new distinct\ndatasets that contain scanned catalog images of interior designs and\nphotographs of protesters with hand-written signs, respectively. Our results\nshow that our STR model outperforms a baseline method that uses\nstate-of-the-art single-wordrecognition techniques on both datasets. STR yields\na high accuracy rate of 90% on the catalog images and 71% on the more difficult\nprotest images, suggesting its generality in recognizing text.', 'Recent models for cross-modal retrieval have benefited from an increasingly\nrich understanding of visual scenes, afforded by scene graphs and object\ninteractions to mention a few. This has resulted in an improved matching\nbetween the visual representation of an image and the textual representation of\nits caption. Yet, current visual representations overlook a key aspect: the\ntext appearing in images, which may contain crucial information for retrieval.\nIn this paper, we first propose a new dataset that allows exploration of\ncross-modal retrieval where images contain scene-text instances. Then, armed\nwith this dataset, we describe several approaches which leverage scene text,\nincluding a better scene-text aware cross-modal retrieval method which uses\nspecialized representations for text from the captions and text from the visual\nscene, and reconcile them in a common embedding space. Extensive experiments\nconfirm that cross-modal retrieval approaches benefit from scene text and\nhighlight interesting research questions worth exploring further. Dataset and\ncode are available at http://europe.naverlabs.com/stacmr', 'Despite the ubiquity of mobile and wearable text messaging applications, the\nproblem of keyboard text decoding is not tackled sufficiently in the light of\nthe enormous success of the deep learning Recurrent Neural Network (RNN) and\nConvolutional Neural Networks (CNN) for natural language understanding. In\nparticular, considering that the keyboard decoders should operate on devices\nwith memory and processor resource constraints, makes it challenging to deploy\nindustrial scale deep neural network (DNN) models. This paper proposes a\nsequence-to-sequence neural attention network system for automatic text\ncorrection and completion. Given an erroneous sequence, our model encodes\ncharacter level hidden representations and then decodes the revised sequence\nthus enabling auto-correction and completion. We achieve this by a combination\nof character level CNN and gated recurrent unit (GRU) encoder along with and a\nword level gated recurrent unit (GRU) attention decoder. Unlike traditional\nlanguage models that learn from billions of words, our corpus size is only 12\nmillion words; an order of magnitude smaller. The memory footprint of our\nlearnt model for inference and prediction is also an order of magnitude smaller\nthan the conventional language model based text decoders. We report baseline\nperformance for neural keyboard decoders in such limited domain. Our models\nachieve a word level accuracy of $90\\%$ and a character error rate CER of\n$2.4\\%$ over the Twitter typo dataset. We present a novel dataset of noisy to\ncorrected mappings by inducing the noise distribution from the Twitter data\nover the OpenSubtitles 2009 dataset; on which our model predicts with a word\nlevel accuracy of $98\\%$ and sequence accuracy of $68.9\\%$. In our user study,\nour model achieved an average CER of $2.6\\%$ with the state-of-the-art\nnon-neural touch-screen keyboard decoder at CER of $1.6\\%$.', ""This paper presents a novel training method for end-to-end scene text\nrecognition. End-to-end scene text recognition offers high recognition\naccuracy, especially when using the encoder-decoder model based on Transformer.\nTo train a highly accurate end-to-end model, we need to prepare a large\nimage-to-text paired dataset for the target language. However, it is difficult\nto collect this data, especially for resource-poor languages. To overcome this\ndifficulty, our proposed method utilizes well-prepared large datasets in\nresource-rich languages such as English, to train the resource-poor\nencoder-decoder model. Our key idea is to build a model in which the encoder\nreflects knowledge of multiple languages while the decoder specializes in\nknowledge of just the resource-poor language. To this end, the proposed method\npre-trains the encoder by using a multilingual dataset that combines the\nresource-poor language's dataset and the resource-rich language's dataset to\nlearn language-invariant knowledge for scene text recognition. The proposed\nmethod also pre-trains the decoder by using the resource-poor language's\ndataset to make the decoder better suited to the resource-poor language.\nExperiments on Japanese scene text recognition using a small, publicly\navailable dataset demonstrate the effectiveness of the proposed method."", 'Recent advancements in the area of Computer Vision with state-of-art Neural\nNetworks has given a boost to Optical Character Recognition (OCR) accuracies.\nHowever, extracting characters/text alone is often insufficient for relevant\ninformation extraction as documents also have a visual structure that is not\ncaptured by OCR. Extracting information from tables, charts, footnotes, boxes,\nheadings and retrieving the corresponding structured representation for the\ndocument remains a challenge and finds application in a large number of\nreal-world use cases. In this paper, we propose a novel enterprise based\nend-to-end framework called DeepReader which facilitates information extraction\nfrom document images via identification of visual entities and populating a\nmeta relational model across different entities in the document image. The\nmodel schema allows for an easy to understand abstraction of the entities\ndetected by the deep vision models and the relationships between them.\nDeepReader has a suite of state-of-the-art vision algorithms which are applied\nto recognize handwritten and printed text, eliminate noisy effects, identify\nthe type of documents and detect visual entities like tables, lines and boxes.\nDeep Reader maps the extracted entities into a rich relational schema so as to\ncapture all the relevant relationships between entities (words, textboxes,\nlines etc) detected in the document. Relevant information and fields can then\nbe extracted from the document by writing SQL queries on top of the\nrelationship tables. A natural language based interface is added on top of the\nrelationship schema so that a non-technical user, specifying the queries in\nnatural language, can fetch the information with minimal effort. In this paper,\nwe also demonstrate many different capabilities of Deep Reader and report\nresults on a real-world use case.', 'Scene text recognition is a hot research topic in computer vision. Recently,\nmany recognition methods based on the encoder-decoder framework have been\nproposed, and they can handle scene texts of perspective distortion and curve\nshape. Nevertheless, they still face lots of challenges like image blur, uneven\nillumination, and incomplete characters. We argue that most encoder-decoder\nmethods are based on local visual features without explicit global semantic\ninformation. In this work, we propose a semantics enhanced encoder-decoder\nframework to robustly recognize low-quality scene texts. The semantic\ninformation is used both in the encoder module for supervision and in the\ndecoder module for initializing. In particular, the state-of-the art ASTER\nmethod is integrated into the proposed framework as an exemplar. Extensive\nexperiments demonstrate that the proposed framework is more robust for\nlow-quality text images, and achieves state-of-the-art results on several\nbenchmark datasets.', 'In this work we present a state-of-the-art approach for unconstrained natural\nscene text recognition. We propose a cascade approach that incorporates a\nconvolutional neural network (CNN) architecture followed by a long short term\nmemory model (LSTM). The CNN learns visual features for the characters and uses\nthem with a softmax layer to detect sequence of characters. While the CNN gives\nvery good recognition results, it does not model relation between characters,\nhence gives rise to false positive and false negative cases (confusing\ncharacters due to visual similarities like ""g"" and ""9"", or confusing background\npatches with characters; either removing existing characters or adding\nnon-existing ones) To alleviate these problems we leverage recent developments\nin LSTM architectures to encode contextual information. We show that the LSTM\ncan dramatically reduce such errors and achieve state-of-the-art accuracy in\nthe task of unconstrained natural scene text recognition. Moreover we manually\nremove all occurrences of the words that exist in the test set from our\ntraining set to test whether our approach will generalize to unseen data. We\nuse the ICDAR 13 test set for evaluation and compare the results with the state\nof the art approaches [11, 18]. We finally present an application of the work\nin the domain of for traffic monitoring.', 'Segmentation of additive manufacturing (AM) defects in X-ray Computed\nTomography (XCT) images is challenging, due to the poor contrast, small sizes\nand variation in appearance of defects. Automatic segmentation can, however,\nprovide quality control for additive manufacturing. Over recent years,\nthree-dimensional convolutional neural networks (3D CNNs) have performed well\nin the volumetric segmentation of medical images. In this work, we leverage\ntechniques from the medical imaging domain and propose training a 3D U-Net\nmodel to automatically segment defects in XCT images of AM samples. This work\nnot only contributes to the use of machine learning for AM defect detection but\nalso demonstrates for the first time 3D volumetric segmentation in AM. We train\nand test with three variants of the 3D U-Net on an AM dataset, achieving a mean\nintersection of union (IOU) value of 88.4%.', 'Recent studies on machine reading comprehension have focused on text-level\nunderstanding but have not yet reached the level of human understanding of the\nvisual layout and content of real-world documents. In this study, we introduce\na new visual machine reading comprehension dataset, named VisualMRC, wherein\ngiven a question and a document image, a machine reads and comprehends texts in\nthe image to answer the question in natural language. Compared with existing\nvisual question answering (VQA) datasets that contain texts in images,\nVisualMRC focuses more on developing natural language understanding and\ngeneration abilities. It contains 30,000+ pairs of a question and an\nabstractive answer for 10,000+ document images sourced from multiple domains of\nwebpages. We also introduce a new model that extends existing\nsequence-to-sequence models, pre-trained with large-scale text corpora, to take\ninto account the visual layout and content of documents. Experiments with\nVisualMRC show that this model outperformed the base sequence-to-sequence\nmodels and a state-of-the-art VQA model. However, its performance is still\nbelow that of humans on most automatic evaluation metrics. The dataset will\nfacilitate research aimed at connecting vision and language understanding.', 'Recognizing irregular text in natural scene images is challenging due to the\nlarge variance in text appearance, such as curvature, orientation and\ndistortion. Most existing approaches rely heavily on sophisticated model\ndesigns and/or extra fine-grained annotations, which, to some extent, increase\nthe difficulty in algorithm implementation and data collection. In this work,\nwe propose an easy-to-implement strong baseline for irregular scene text\nrecognition, using off-the-shelf neural network components and only word-level\nannotations. It is composed of a $31$-layer ResNet, an LSTM-based\nencoder-decoder framework and a 2-dimensional attention module. Despite its\nsimplicity, the proposed method is robust and achieves state-of-the-art\nperformance on both regular and irregular scene text recognition benchmarks.\nCode is available at: https://tinyurl.com/ShowAttendRead', 'Text normalization is a ubiquitous process that appears as the first step of\nmany Natural Language Processing problems. However, previous Deep Learning\napproaches have suffered from so-called silly errors, which are undetectable on\nunsupervised frameworks, making those models unsuitable for deployment. In this\nwork, we make use of an attention-based encoder-decoder architecture that\novercomes these undetectable errors by using a fine-grained character-level\napproach rather than a word-level one. Furthermore, our new general-purpose\nencoder based on causal convolutions, called Causal Feature Extractor (CFE), is\nintroduced and compared to other common encoders. The experimental results show\nthe feasibility of this encoder, which leverages the attention mechanisms the\nmost and obtains better results in terms of accuracy, number of parameters and\nconvergence time. While our method results in a slightly worse initial accuracy\n(92.74%), errors can be automatically detected and, thus, more readily solved,\nobtaining a more robust model for deployment. Furthermore, there is still\nplenty of room for future improvements that will push even further these\nadvantages.', 'Encoder-decoder models have become an effective approach for sequence\nlearning tasks like machine translation, image captioning and speech\nrecognition, but have yet to show competitive results for handwritten text\nrecognition. To this end, we propose an attention-based sequence-to-sequence\nmodel. It combines a convolutional neural network as a generic feature\nextractor with a recurrent neural network to encode both the visual\ninformation, as well as the temporal context between characters in the input\nimage, and uses a separate recurrent neural network to decode the actual\ncharacter sequence. We make experimental comparisons between various attention\nmechanisms and positional encodings, in order to find an appropriate alignment\nbetween the input and output sequence. The model can be trained end-to-end and\nthe optional integration of a hybrid loss allows the encoder to retain an\ninterpretable and usable output, if desired. We achieve competitive results on\nthe IAM and ICFHR2016 READ data sets compared to the state-of-the-art without\nthe use of a language model, and we significantly improve over any recent\nsequence-to-sequence approaches.', 'The pursuit of high performance on public benchmarks has been the driving\nforce for research in scene text recognition, and notable progress has been\nachieved. However, a close investigation reveals a startling fact that the\nstate-of-the-art methods perform well on images with words within vocabulary\nbut generalize poorly to images with words outside vocabulary. We call this\nphenomenon ""vocabulary reliance"". In this paper, we establish an analytical\nframework to conduct an in-depth study on the problem of vocabulary reliance in\nscene text recognition. Key findings include: (1) Vocabulary reliance is\nubiquitous, i.e., all existing algorithms more or less exhibit such\ncharacteristic; (2) Attention-based decoders prove weak in generalizing to\nwords outside vocabulary and segmentation-based decoders perform well in\nutilizing visual features; (3) Context modeling is highly coupled with the\nprediction layers. These findings provide new insights and can benefit future\nresearch in scene text recognition. Furthermore, we propose a simple yet\neffective mutual learning strategy to allow models of two families\n(attention-based and segmentation-based) to learn collaboratively. This remedy\nalleviates the problem of vocabulary reliance and improves the overall scene\ntext recognition performance.', 'In the Gastric Histopathology Image Classification (GHIC) tasks, which are\nusually weakly supervised learning missions, there is inevitably redundant\ninformation in the images. Therefore, designing networks that can focus on\neffective distinguishing features has become a popular research topic. In this\npaper, to accomplish the tasks of GHIC superiorly and to assist pathologists in\nclinical diagnosis, an intelligent Hierarchical Conditional Random Field based\nAttention Mechanism (HCRF-AM) model is proposed. The HCRF-AM model consists of\nan Attention Mechanism (AM) module and an Image Classification (IC) module. In\nthe AM module, an HCRF model is built to extract attention regions. In the IC\nmodule, a Convolutional Neural Network (CNN) model is trained with the\nattention regions selected and then an algorithm called Classification\nProbability-based Ensemble Learning is applied to obtain the image-level\nresults from patch-level output of the CNN. In the experiment, a classification\nspecificity of 96.67% is achieved on a gastric histopathology dataset with 700\nimages. Our HCRF-AM model demonstrates high classification performance and\nshows its effectiveness and future potential in the GHIC field.', 'Key information extraction from document images is of paramount importance in\noffice automation. Conventional template matching based approaches fail to\ngeneralize well to document images of unseen templates, and are not robust\nagainst text recognition errors. In this paper, we propose an end-to-end\nSpatial Dual-Modality Graph Reasoning method (SDMG-R) to extract key\ninformation from unstructured document images. We model document images as\ndual-modality graphs, nodes of which encode both the visual and textual\nfeatures of detected text regions, and edges of which represent the spatial\nrelations between neighboring text regions. The key information extraction is\nsolved by iteratively propagating messages along graph edges and reasoning the\ncategories of graph nodes. In order to roundly evaluate our proposed method as\nwell as boost the future research, we release a new dataset named WildReceipt,\nwhich is collected and annotated tailored for the evaluation of key information\nextraction from document images of unseen templates in the wild. It contains 25\nkey information categories, a total of about 69000 text boxes, and is about 2\ntimes larger than the existing public datasets. Extensive experiments validate\nthat all information including visual features, textual features and spatial\nrelations can benefit key information extraction. It has been shown that SDMG-R\ncan effectively extract key information from document images of unseen\ntemplates, and obtain new state-of-the-art results on the recent popular\nbenchmark SROIE and our WildReceipt. Our code and dataset will be publicly\nreleased.', 'Text recognition has attracted considerable research interests because of its\nvarious applications. The cutting-edge text recognition methods are based on\nattention mechanisms. However, most of attention methods usually suffer from\nserious alignment problem due to its recurrency alignment operation, where the\nalignment relies on historical decoding results. To remedy this issue, we\npropose a decoupled attention network (DAN), which decouples the alignment\noperation from using historical decoding results. DAN is an effective, flexible\nand robust end-to-end text recognizer, which consists of three components: 1) a\nfeature encoder that extracts visual features from the input image; 2) a\nconvolutional alignment module that performs the alignment operation based on\nvisual features from the encoder; and 3) a decoupled text decoder that makes\nfinal prediction by jointly using the feature map and attention maps.\nExperimental results show that DAN achieves state-of-the-art performance on\nmultiple text recognition tasks, including offline handwritten text recognition\nand regular/irregular scene text recognition.', ""Spelling error correction is one of topics which have a long history in\nnatural language processing. Although previous studies have achieved remarkable\nresults, challenges still exist. In the Vietnamese language, a state-of-the-art\nmethod for the task infers a syllable's context from its adjacent syllables.\nThe method's accuracy can be unsatisfactory, however, because the model may\nlose the context if two (or more) spelling mistakes stand near each other. In\nthis paper, we propose a novel method to correct Vietnamese spelling errors. We\ntackle the problems of mistyped errors and misspelled errors by using a deep\nlearning model. The embedding layer, in particular, is powered by the byte pair\nencoding technique. The sequence to sequence model based on the Transformer\narchitecture makes our approach different from the previous works on the same\nproblem. In the experiment, we train the model with a large synthetic dataset,\nwhich is randomly introduced spelling errors. We test the performance of the\nproposed method using a realistic dataset. This dataset contains 11,202\nhuman-made misspellings in 9,341 different Vietnamese sentences. The\nexperimental results show that our method achieves encouraging performance with\n86.8% errors detected and 81.5% errors corrected, which improves the\nstate-of-the-art approach 5.6% and 2.2%, respectively."", 'In this paper, we study the problem of text line recognition. Unlike most\napproaches targeting specific domains such as scene-text or handwritten\ndocuments, we investigate the general problem of developing a universal\narchitecture that can extract text from any image, regardless of source or\ninput modality. We consider two decoder families (Connectionist Temporal\nClassification and Transformer) and three encoder modules (Bidirectional LSTMs,\nSelf-Attention, and GRCLs), and conduct extensive experiments to compare their\naccuracy and performance on widely used public datasets of scene and\nhandwritten text. We find that a combination that so far has received little\nattention in the literature, namely a Self-Attention encoder coupled with the\nCTC decoder, when compounded with an external language model and trained on\nboth public and internal data, outperforms all the others in accuracy and\ncomputational complexity. Unlike the more common Transformer-based models, this\narchitecture can handle inputs of arbitrary length, a requirement for universal\nline recognition. Using an internal dataset collected from multiple sources, we\nalso expose the limitations of current public datasets in evaluating the\naccuracy of line recognizers, as the relatively narrow image width and sequence\nlength distributions do not allow to observe the quality degradation of the\nTransformer approach when applied to the transcription of long lines.', 'Texts appearing in daily scenes that can be recognized by OCR (Optical\nCharacter Recognition) tools contain significant information, such as street\nname, product brand and prices. Two tasks -- text-based visual question\nanswering and text-based image captioning, with a text extension from existing\nvision-language applications, are catching on rapidly. To address these\nproblems, many sophisticated multi-modality encoding frameworks (such as\nheterogeneous graph structure) are being used. In this paper, we argue that a\nsimple attention mechanism can do the same or even better job without any bells\nand whistles. Under this mechanism, we simply split OCR token features into\nseparate visual- and linguistic-attention branches, and send them to a popular\nTransformer decoder to generate answers or captions. Surprisingly, we find this\nsimple baseline model is rather strong -- it consistently outperforms\nstate-of-the-art (SOTA) models on two popular benchmarks, TextVQA and all three\ntasks of ST-VQA, although these SOTA models use far more complex encoding\nmechanisms. Transferring it to text-based image captioning, we also surpass the\nTextCaps Challenge 2020 winner. We wish this work to set the new baseline for\nthis two OCR text related applications and to inspire new thinking of\nmulti-modality encoder design. Code is available at\nhttps://github.com/ZephyrZhuQi/ssbaseline', 'Structured information extraction from document images usually consists of\nthree steps: text detection, text recognition, and text field labeling. While\ntext detection and text recognition have been heavily studied and improved a\nlot in literature, text field labeling is less explored and still faces many\nchallenges. Existing learning based methods for text labeling task usually\nrequire a large amount of labeled examples to train a specific model for each\ntype of document. However, collecting large amounts of document images and\nlabeling them is difficult and sometimes impossible due to privacy issues.\nDeploying separate models for each type of document also consumes a lot of\nresources. Facing these challenges, we explore one-shot learning for the text\nfield labeling task. Existing one-shot learning methods for the task are mostly\nrule-based and have difficulty in labeling fields in crowded regions with few\nlandmarks and fields consisting of multiple separate text regions. To alleviate\nthese problems, we proposed a novel deep end-to-end trainable approach for\none-shot text field labeling, which makes use of attention mechanism to\ntransfer the layout information between document images. We further applied\nconditional random field on the transferred layout information for the\nrefinement of field labeling. We collected and annotated a real-world one-shot\nfield labeling dataset with a large variety of document types and conducted\nextensive experiments to examine the effectiveness of the proposed model. To\nstimulate research in this direction, the collected dataset and the one-shot\nmodel will be released1.', ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'Attention-based scene text recognizers have gained huge success, which\nleverages a more compact intermediate representation to learn 1d- or 2d-\nattention by a RNN-based encoder-decoder architecture. However, such methods\nsuffer from attention-drift problem because high similarity among encoded\nfeatures leads to attention confusion under the RNN-based local attention\nmechanism. Moreover, RNN-based methods have low efficiency due to poor\nparallelization. To overcome these problems, we propose the MASTER, a\nself-attention based scene text recognizer that (1) not only encodes the\ninput-output attention but also learns self-attention which encodes\nfeature-feature and target-target relationships inside the encoder and decoder\nand (2) learns a more powerful and robust intermediate representation to\nspatial distortion, and (3) owns a great training efficiency because of high\ntraining parallelization and a high-speed inference because of an efficient\nmemory-cache mechanism. Extensive experiments on various benchmarks demonstrate\nthe superior performance of our MASTER on both regular and irregular scene\ntext. Pytorch code can be found at https://github.com/wenwenyu/MASTER-pytorch,\nand Tensorflow code can be found at https://github.com/jiangxiluning/MASTER-TF.', 'Reading irregular scene text of arbitrary shape in natural images is still a\nchallenging problem, despite the progress made recently. Many existing\napproaches incorporate sophisticated network structures to handle various\nshapes, use extra annotations for stronger supervision, or employ hard-to-train\nrecurrent neural networks for sequence modeling. In this work, we propose a\nsimple yet strong approach for scene text recognition. With no need to convert\ninput images to sequence representations, we directly connect two-dimensional\nCNN features to an attention-based sequence decoder which guided by holistic\nrepresentation. The holistic representation can guide the attention-based\ndecoder focus on more accurate area. As no recurrent module is adopted, our\nmodel can be trained in parallel. It achieves 1.5x to 9.4x acceleration to\nbackward pass and 1.3x to 7.9x acceleration to forward pass, compared with the\nRNN counterparts. The proposed model is trained with only word-level\nannotations. With this simple design, our method achieves state-of-the-art or\ncompetitive recognition performance on the evaluated regular and irregular\nscene text benchmark datasets.', 'Texts from scene images typically consist of several characters and exhibit a\ncharacteristic sequence structure. Existing methods capture the structure with\nthe sequence-to-sequence models by an encoder to have the visual\nrepresentations and then a decoder to translate the features into the label\nsequence. In this paper, we study text recognition framework by considering the\nlong-term temporal dependencies in the encoder stage. We demonstrate that the\nproposed Temporal Convolutional Encoder with increased sequential extents\nimproves the accuracy of text recognition. We also study the impact of\ndifferent attention modules in convolutional blocks for learning accurate text\nrepresentations. We conduct comparisons on seven datasets and the experiments\ndemonstrate the effectiveness of our proposed approach.', 'Scene text recognition has attracted a great many researches due to its\nimportance to various applications. Existing methods mainly adopt recurrence or\nconvolution based networks. Though have obtained good performance, these\nmethods still suffer from two limitations: slow training speed due to the\ninternal recurrence of RNNs, and high complexity due to stacked convolutional\nlayers for long-term feature extraction. This paper, for the first time,\nproposes a no-recurrence sequence-to-sequence text recognizer, named NRTR, that\ndispenses with recurrences and convolutions entirely. NRTR follows the\nencoder-decoder paradigm, where the encoder uses stacked self-attention to\nextract image features, and the decoder applies stacked self-attention to\nrecognize texts based on encoder output. NRTR relies solely on self-attention\nmechanism thus could be trained with more parallelization and less complexity.\nConsidering scene image has large variation in text and background, we further\ndesign a modality-transform block to effectively transform 2D input images to\n1D sequences, combined with the encoder to extract more discriminative\nfeatures. NRTR achieves state-of-the-art or highly competitive performance on\nboth regular and irregular benchmarks, while requires only a small fraction of\ntraining time compared to the best model from the literature (at least 8 times\nfaster).', 'Scene text recognition has been a hot research topic in computer vision due\nto its various applications. The state of the art is the attention-based\nencoder-decoder framework that learns the mapping between input images and\noutput sequences in a purely data-driven way. However, we observe that existing\nattention-based methods perform poorly on complicated and/or low-quality\nimages. One major reason is that existing methods cannot get accurate\nalignments between feature areas and targets for such images. We call this\nphenomenon ""attention drift"". To tackle this problem, in this paper we propose\nthe FAN (the abbreviation of Focusing Attention Network) method that employs a\nfocusing attention mechanism to automatically draw back the drifted attention.\nFAN consists of two major components: an attention network (AN) that is\nresponsible for recognizing character targets as in the existing methods, and a\nfocusing network (FN) that is responsible for adjusting attention by evaluating\nwhether AN pays attention properly on the target areas in the images.\nFurthermore, different from the existing methods, we adopt a ResNet-based\nnetwork to enrich deep representations of scene text images. Extensive\nexperiments on various benchmarks, including the IIIT5k, SVT and ICDAR\ndatasets, show that the FAN method substantially outperforms the existing\nmethods.', 'Scene Text Recognition (STR), the task of recognizing text against complex\nimage backgrounds, is an active area of research. Current state-of-the-art\n(SOTA) methods still struggle to recognize text written in arbitrary shapes. In\nthis paper, we introduce a novel architecture for STR, named Selective Context\nATtentional Text Recognizer (SCATTER). SCATTER utilizes a stacked block\narchitecture with intermediate supervision during training, that paves the way\nto successfully train a deep BiLSTM encoder, thus improving the encoding of\ncontextual dependencies. Decoding is done using a two-step 1D attention\nmechanism. The first attention step re-weights visual features from a CNN\nbackbone together with contextual features computed by a BiLSTM layer. The\nsecond attention step, similar to previous papers, treats the features as a\nsequence and attends to the intra-sequence relationships. Experiments show that\nthe proposed approach surpasses SOTA performance on irregular text recognition\nbenchmarks by 3.7\\% on average.', 'The paper proposes a new text recognition network for scene-text images. Many\nstate-of-the-art methods employ the attention mechanism either in the text\nencoder or decoder for the text alignment. Although the encoder-based attention\nyields promising results, these schemes inherit noticeable limitations. They\nperform the feature extraction (FE) and visual attention (VA) sequentially,\nwhich bounds the attention mechanism to rely only on the FE final single-scale\noutput. Moreover, the utilization of the attention process is limited by only\napplying it directly to the single scale feature-maps. To address these issues,\nwe propose a new multi-scale and encoder-based attention network for text\nrecognition that performs the multi-scale FE and VA in parallel. The\nmulti-scale channels also undergo regular fusion with each other to develop the\ncoordinated knowledge together. Quantitative evaluation and robustness analysis\non the standard benchmarks demonstrate that the proposed network outperforms\nthe state-of-the-art in most cases.', 'The attention-based encoder-decoder framework has recently achieved\nimpressive results for scene text recognition, and many variants have emerged\nwith improvements in recognition quality. However, it performs poorly on\ncontextless texts (e.g., random character sequences) which is unacceptable in\nmost of real application scenarios. In this paper, we first deeply investigate\nthe decoding process of the decoder. We empirically find that a representative\ncharacter-level sequence decoder utilizes not only context information but also\npositional information. Contextual information, which the existing approaches\nheavily rely on, causes the problem of attention drift. To suppress such\nside-effect, we propose a novel position enhancement branch, and dynamically\nfuse its outputs with those of the decoder attention module for scene text\nrecognition. Specifically, it contains a position aware module to enable the\nencoder to output feature vectors encoding their own spatial positions, and an\nattention module to estimate glimpses using the positional clue (i.e., the\ncurrent decoding time step) only. The dynamic fusion is conducted for more\nrobust feature via an element-wise gate mechanism. Theoretically, our proposed\nmethod, dubbed \\emph{RobustScanner}, decodes individual characters with dynamic\nratio between context and positional clues, and utilizes more positional ones\nwhen the decoding sequences with scarce context, and thus is robust and\npractical. Empirically, it has achieved new state-of-the-art results on popular\nregular and irregular text recognition benchmarks while without much\nperformance drop on contextless benchmarks, validating its robustness in both\ncontextual and contextless application scenarios.', 'Scene text recognition has attracted particular research interest because it\nis a very challenging problem and has various applications. The most\ncutting-edge methods are attentional encoder-decoder frameworks that learn the\nalignment between the input image and output sequences. In particular, the\ndecoder recurrently outputs predictions, using the prediction of the previous\nstep as a guidance for every time step. In this study, we point out that the\ninappropriate use of previous predictions in existing attention mechanisms\nrestricts the recognition performance and brings instability. To handle this\nproblem, we propose a novel module, namely adaptive embedding gate(AEG). The\nproposed AEG focuses on introducing high-order character language models to\nattention mechanism by controlling the information transmission between\nadjacent characters. AEG is a flexible module and can be easily integrated into\nthe state-of-the-art attentional methods. We evaluate its effectiveness as well\nas robustness on a number of standard benchmarks, including the IIIT$5$K, SVT,\nSVT-P, CUTE$80$, and ICDAR datasets. Experimental results demonstrate that AEG\ncan significantly boost recognition performance and bring better robustness.', 'In this paper, we propose an end-to-end trainable framework for restoring\nhistorical documents content that follows the correct reading order. In this\nframework, two branches named character branch and layout branch are added\nbehind the feature extraction network. The character branch localizes\nindividual characters in a document image and recognizes them simultaneously.\nThen we adopt a post-processing method to group them into text lines. The\nlayout branch based on fully convolutional network outputs a binary mask. We\nthen use Hough transform for line detection on the binary mask and combine\ncharacter results with the layout information to restore document content.\nThese two branches can be trained in parallel and are easy to train.\nFurthermore, we propose a re-score mechanism to minimize recognition error.\nExperiment results on the extended Chinese historical document MTHv2 dataset\ndemonstrate the effectiveness of the proposed framework.', 'In this work, we adhere to explore a Multi-Tasking learning (MTL) based\nnetwork to perform document attribute classification such as the font type,\nfont size, font emphasis and scanning resolution classification of a document\nimage. To accomplish these tasks, we operate on either segmented word level or\non uniformed size patches randomly cropped out of the document. Furthermore, a\nhybrid convolution neural network (CNN) architecture ""MTL+MI"", which is based\non the combination of MTL and Multi-Instance (MI) of patch and word is used to\naccomplish joint learning for the classification of the same document\nattributes. The contribution of this paper are three fold: firstly, based on\nsegmented word images and patches, we present a MTL based network for the\nclassification of a full document image. Secondly, we propose a MTL and MI\n(using segmented words and patches) based combined CNN architecture (""MTL+MI"")\nfor the classification of same document attributes. Thirdly, based on the\nmulti-tasking classifications of the words and/or patches, we propose an\nintelligent voting system which is based on the posterior probabilities of each\nwords and/or patches to perform the classification of document\'s attributes of\ncomplete document image.', 'Recognizing scene text is a challenging problem, even more so than the\nrecognition of scanned documents. This problem has gained significant attention\nfrom the computer vision community in recent years, and several methods based\non energy minimization frameworks and deep learning approaches have been\nproposed. In this work, we focus on the energy minimization framework and\npropose a model that exploits both bottom-up and top-down cues for recognizing\ncropped words extracted from street images. The bottom-up cues are derived from\nindividual character detections from an image. We build a conditional random\nfield model on these detections to jointly model the strength of the detections\nand the interactions between them. These interactions are top-down cues\nobtained from a lexicon-based prior, i.e., language statistics. The optimal\nword represented by the text image is obtained by minimizing the energy\nfunction corresponding to the random field model. We evaluate our proposed\nalgorithm extensively on a number of cropped scene text benchmark datasets,\nnamely Street View Text, ICDAR 2003, 2011 and 2013 datasets, and IIIT 5K-word,\nand show better performance than comparable methods. We perform a rigorous\nanalysis of all the steps in our approach and analyze the results. We also show\nthat state-of-the-art convolutional neural network features can be integrated\nin our framework to further improve the recognition performance.', ""Given image labels as the only supervisory signal, we focus on harvesting, or\nmining, thoracic disease localizations from chest X-ray images. Harvesting such\nlocalizations from existing datasets allows for the creation of improved data\nsources for computer-aided diagnosis and retrospective analyses. We train a\nconvolutional neural network (CNN) for image classification and propose an\nattention mining (AM) strategy to improve the model's sensitivity or saliency\nto disease patterns. The intuition of AM is that once the most salient disease\narea is blocked or hidden from the CNN model, it will pay attention to\nalternative image regions, while still attempting to make correct predictions.\nHowever, the model requires to be properly constrained during AM, otherwise, it\nmay overfit to uncorrelated image parts and forget the valuable knowledge that\nit has learned from the original image classification task. To alleviate such\nside effects, we then design a knowledge preservation (KP) loss, which\nminimizes the discrepancy between responses for X-ray images from the original\nand the updated networks. Furthermore, we modify the CNN model to include\nmulti-scale aggregation (MSA), improving its localization ability on\nsmall-scale disease findings, e.g., lung nodules. We experimentally validate\nour method on the publicly-available ChestX-ray14 dataset, outperforming a\nclass activation map (CAM)-based approach, and demonstrating the value of our\nnovel framework for mining disease locations."", 'From the literature, it is demonstrated that performing text-line\nsegmentation directly in the run-length compressed handwritten document images\nsignificantly reduces the computational time and memory space. In this paper,\nwe investigate the issues of word and character segmentation directly on the\nrun-length compressed document images. Primarily, the spreads of the characters\nare intelligently extracted from the foreground runs of the compressed data and\nsubsequently connected components are established. The spacing between the\nconnected components would be larger between the adjacent words when compared\nto that of intra-words. With this knowledge, a threshold is empirically chosen\nfor inter-word separation. Every connected component within a word is further\nanalysed for character segmentation. Here, min-cut graph concept is used for\nseparating the touching characters. Over-segmentation and under-segmentation\nissues are addressed by insertion and deletion operations respectively. The\napproach has been developed particularly for compressed handwritten English\ndocument images. However, the model has been tested on non-English document\nimages.', 'The paper approaches the task of handwritten text recognition (HTR) with\nattentional encoder-decoder networks trained on sequences of characters, rather\nthan words. We experiment on lines of text from popular handwriting datasets\nand compare different activation functions for the attention mechanism used for\naligning image pixels and target characters. We find that softmax attention\nfocuses heavily on individual characters, while sigmoid attention focuses on\nmultiple characters at each step of the decoding. When the sequence alignment\nis one-to-one, softmax attention is able to learn a more precise alignment at\neach step of the decoding, whereas the alignment generated by sigmoid attention\nis much less precise. When a linear function is used to obtain attention\nweights, the model predicts a character by looking at the entire sequence of\ncharacters and performs poorly because it lacks a precise alignment between the\nsource and target. Future research may explore HTR in natural scene images,\nsince the model is capable of transcribing handwritten text without the need\nfor producing segmentations or bounding boxes of text in images.', 'The recently proposed Multilinear Compressive Learning (MCL) framework\ncombines Multilinear Compressive Sensing and Machine Learning into an\nend-to-end system that takes into account the multidimensional structure of the\nsignals when designing the sensing and feature synthesis components. The key\nidea behind MCL is the assumption of the existence of a tensor subspace which\ncan capture the essential features from the signal for the downstream learning\ntask. Thus, the ability to find such a discriminative tensor subspace and\noptimize the system to project the signals onto that data manifold plays an\nimportant role in Multilinear Compressive Learning. In this paper, we propose a\nnovel solution to address both of the aforementioned requirements, i.e., How to\nfind those tensor subspaces in which the signals of interest are highly\nseparable? and How to optimize the sensing and feature synthesis components to\ntransform the original signals to the data manifold found in the first\nquestion? In our proposal, the discovery of a high-quality data manifold is\nconducted by training a nonlinear compressive learning system on the inference\ntask. Its knowledge of the data manifold of interest is then progressively\ntransferred to the MCL components via multi-stage supervised training with the\nsupervisory information encoding how the compressed measurements, the\nsynthesized features, and the predictions should be like. The proposed\nknowledge transfer algorithm also comes with a semi-supervised adaption that\nenables compressive learning models to utilize unlabeled data effectively.\nExtensive experiments demonstrate that the proposed knowledge transfer method\ncan effectively train MCL models to compressively sense and synthesize better\nfeatures for the learning tasks with improved performances, especially when the\ncomplexity of the learning task increases.', ""Attention-based encoder-decoder framework is widely used in the scene text\nrecognition task. However, for the current state-of-the-art(SOTA) methods,\nthere is room for improvement in terms of the efficient usage of local visual\nand global context information of the input text image, as well as the robust\ncorrelation between the scene processing module(encoder) and the text\nprocessing module(decoder). In this paper, we propose a Representation and\nCorrelation Enhanced Encoder-Decoder Framework(RCEED) to address these\ndeficiencies and break performance bottleneck. In the encoder module, local\nvisual feature, global context feature, and position information are aligned\nand fused to generate a small-size comprehensive feature map. In the decoder\nmodule, two methods are utilized to enhance the correlation between scene and\ntext feature space. 1) The decoder initialization is guided by the holistic\nfeature and global glimpse vector exported from the encoder. 2) The feature\nenriched glimpse vector produced by the Multi-Head General Attention is used to\nassist the RNN iteration and the character prediction at each time step.\nMeanwhile, we also design a Layernorm-Dropout LSTM cell to improve model's\ngeneralization towards changeable texts. Extensive experiments on the\nbenchmarks demonstrate the advantageous performance of RCEED in scene text\nrecognition tasks, especially the irregular ones."", 'Here we address a challenging problem: recognizing multiple text sequences\nfrom an image by pure end-to-end learning. It is twofold: 1) Multiple text\nsequences recognition. Each image may contain multiple text sequences of\ndifferent content, location and orientation, and we try to recognize all the\ntext sequences contained in the image. 2) Pure end-to-end (PEE) learning.We\nsolve the problem in a pure end-to-end learning way where each training image\nis labeled by only text transcripts of all contained sequences, without any\ngeometric annotations. Most existing works recognize multiple text sequences\nfrom an image in a non-end-to-end (NEE) or quasi-end-to-end (QEE) way, in which\neach image is trained with both text transcripts and text locations.Only\nrecently, a PEE method was proposed to recognize text sequences from an image\nwhere the text sequence was split to several lines in the image. However, it\ncannot be directly applied to recognizing multiple text sequences from an\nimage. So in this paper, we propose a pure end-to-end learning method to\nrecognize multiple text sequences from an image. Our method directly learns\nmultiple sequences of probability distribution conditioned on each input image,\nand outputs multiple text transcripts with a well-designed decoding strategy.To\nevaluate the proposed method, we constructed several datasets mainly based on\nan existing public dataset andtwo real application scenarios. Experimental\nresults show that the proposed method can effectively recognize multiple text\nsequences from images, and outperforms CTC-based and attention-based baseline\nmethods.', 'In this paper, we address the problem of having characters with different\nscales in scene text recognition. We propose a novel scale aware feature\nencoder (SAFE) that is designed specifically for encoding characters with\ndifferent scales. SAFE is composed of a multi-scale convolutional encoder and a\nscale attention network. The multi-scale convolutional encoder targets at\nextracting character features under multiple scales, and the scale attention\nnetwork is responsible for selecting features from the most relevant scale(s).\nSAFE has two main advantages over the traditional single-CNN encoder used in\ncurrent state-of-the-art text recognizers. First, it explicitly tackles the\nscale problem by extracting scale-invariant features from the characters. This\nallows the recognizer to put more effort in handling other challenges in scene\ntext recognition, like those caused by view distortion and poor image quality.\nSecond, it can transfer the learning of feature encoding across different\ncharacter scales. This is particularly important when the training set has a\nvery unbalanced distribution of character scales, as training with such a\ndataset will make the encoder biased towards extracting features from the\npredominant scale. To evaluate the effectiveness of SAFE, we design a simple\ntext recognizer named scale-spatial attention network (S-SAN) that employs SAFE\nas its feature encoder, and carry out experiments on six public benchmarks.\nExperimental results demonstrate that S-SAN can achieve state-of-the-art (or,\nin some cases, extremely competitive) performance without any post-processing.', 'Scene text recognition is a challenging task due to diverse variations of\ntext instances in natural scene images. Conventional methods based on\nCNN-RNN-CTC or encoder-decoder with attention mechanism may not fully\ninvestigate stable and efficient feature representations for multi-oriented\nscene texts. In this paper, we propose a primitive representation learning\nmethod that aims to exploit intrinsic representations of scene text images. We\nmodel elements in feature maps as the nodes of an undirected graph. A pooling\naggregator and a weighted aggregator are proposed to learn primitive\nrepresentations, which are transformed into high-level visual text\nrepresentations by graph convolutional networks. A Primitive REpresentation\nlearning Network (PREN) is constructed to use the visual text representations\nfor parallel decoding. Furthermore, by integrating visual text representations\ninto an encoder-decoder model with the 2D attention mechanism, we propose a\nframework called PREN2D to alleviate the misalignment problem in\nattention-based methods. Experimental results on both English and Chinese scene\ntext recognition tasks demonstrate that PREN keeps a balance between accuracy\nand efficiency, while PREN2D achieves state-of-the-art performance.', 'Inspired by speech recognition, recent state-of-the-art algorithms mostly\nconsider scene text recognition as a sequence prediction problem. Though\nachieving excellent performance, these methods usually neglect an important\nfact that text in images are actually distributed in two-dimensional space. It\nis a nature quite different from that of speech, which is essentially a\none-dimensional signal. In principle, directly compressing features of text\ninto a one-dimensional form may lose useful information and introduce extra\nnoise. In this paper, we approach scene text recognition from a two-dimensional\nperspective. A simple yet effective model, called Character Attention Fully\nConvolutional Network (CA-FCN), is devised for recognizing the text of\narbitrary shapes. Scene text recognition is realized with a semantic\nsegmentation network, where an attention mechanism for characters is adopted.\nCombined with a word formation module, CA-FCN can simultaneously recognize the\nscript and predict the position of each character. Experiments demonstrate that\nthe proposed algorithm outperforms previous methods on both regular and\nirregular text datasets. Moreover, it is proven to be more robust to imprecise\nlocalizations in the text detection phase, which are very common in practice.', 'This paper investigates how to correct Chinese text errors with types of\nmistaken, missing and redundant characters, which is common for Chinese native\nspeakers. Most existing models based on detect-correct framework can correct\nmistaken characters errors, but they cannot deal with missing or redundant\ncharacters. The reason is that lengths of sentences before and after correction\nare not the same, leading to the inconsistence between model inputs and\noutputs. Although the Seq2Seq-based or sequence tagging methods provide\nsolutions to the problem and achieved relatively good results on English\ncontext, but they do not perform well in Chinese context according to our\nexperimental results. In our work, we propose a novel detect-correct framework\nwhich is alignment-agnostic, meaning that it can handle both text aligned and\nnon-aligned occasions, and it can also serve as a cold start model when there\nare no annotated data provided. Experimental results on three datasets\ndemonstrate that our method is effective and achieves the best performance\namong existing published models.', 'Offline handwritten text recognition from images is an important problem for\nenterprises attempting to digitize large volumes of handmarked scanned\ndocuments/reports. Deep recurrent models such as Multi-dimensional LSTMs have\nbeen shown to yield superior performance over traditional Hidden Markov Model\nbased approaches that suffer from the Markov assumption and therefore lack the\nrepresentational power of RNNs. In this paper we introduce a novel approach\nthat combines a deep convolutional network with a recurrent Encoder-Decoder\nnetwork to map an image to a sequence of characters corresponding to the text\npresent in the image. The entire model is trained end-to-end using Focal Loss,\nan improvement over the standard Cross-Entropy loss that addresses the class\nimbalance problem, inherent to text recognition. To enhance the decoding\ncapacity of the model, Beam Search algorithm is employed which searches for the\nbest sequence out of a set of hypotheses based on a joint distribution of\nindividual characters. Our model takes as input a downsampled version of the\noriginal image thereby making it both computationally and memory efficient. The\nexperimental results were benchmarked against two publicly available datasets,\nIAM and RIMES. We surpass the state-of-the-art word level accuracy on the\nevaluation set of both datasets by 3.5% & 1.1%, respectively.', 'Natural language correction has the potential to help language learners\nimprove their writing skills. While approaches with separate classifiers for\ndifferent error types have high precision, they do not flexibly handle errors\nsuch as redundancy or non-idiomatic phrasing. On the other hand, word and\nphrase-based machine translation methods are not designed to cope with\northographic errors, and have recently been outpaced by neural models.\nMotivated by these issues, we present a neural network-based approach to\nlanguage correction. The core component of our method is an encoder-decoder\nrecurrent neural network with an attention mechanism. By operating at the\ncharacter level, the network avoids the problem of out-of-vocabulary words. We\nillustrate the flexibility of our approach on dataset of noisy, user-generated\ntext collected from an English learner forum. When combined with a language\nmodel, our method achieves a state-of-the-art $F_{0.5}$-score on the CoNLL 2014\nShared Task. We further demonstrate that training the network on additional\ndata with synthesized errors can improve performance.', 'Generating textual descriptions for images has been an attractive problem for\nthe computer vision and natural language processing researchers in recent\nyears. Dozens of models based on deep learning have been proposed to solve this\nproblem. The existing approaches are based on neural encoder-decoder structures\nequipped with the attention mechanism. These methods strive to train decoders\nto minimize the log likelihood of the next word in a sentence given the\nprevious ones, which results in the sparsity of the output space. In this work,\nwe propose a new approach to train decoders to regress the word embedding of\nthe next word with respect to the previous ones instead of minimizing the log\nlikelihood. The proposed method is able to learn and extract long-term\ninformation and can generate longer fine-grained captions without introducing\nany external memory cell. Furthermore, decoders trained by the proposed\ntechnique can take the importance of the generated words into consideration\nwhile generating captions. In addition, a novel semantic attention mechanism is\nproposed that guides attention points through the image, taking the meaning of\nthe previously generated word into account. We evaluate the proposed approach\nwith the MS-COCO dataset. The proposed model outperformed the state of the art\nmodels especially in generating longer captions. It achieved a CIDEr score\nequal to 125.0 and a BLEU-4 score equal to 50.5, while the best scores of the\nstate of the art models are 117.1 and 48.0, respectively.', 'Attention-based encoder-decoder model has achieved impressive results for\nboth automatic speech recognition (ASR) and text-to-speech (TTS) tasks. This\napproach takes advantage of the memorization capacity of neural networks to\nlearn the mapping from the input sequence to the output sequence from scratch,\nwithout the assumption of prior knowledge such as the alignments. However, this\nmodel is prone to overfitting, especially when the amount of training data is\nlimited. Inspired by SpecAugment and BERT, in this paper, we propose a semantic\nmask based regularization for training such kind of end-to-end (E2E) model. The\nidea is to mask the input features corresponding to a particular output token,\ne.g., a word or a word-piece, in order to encourage the model to fill the token\nbased on the contextual information. While this approach is applicable to the\nencoder-decoder framework with any type of neural network architecture, we\nstudy the transformer-based model for ASR in this work. We perform experiments\non Librispeech 960h and TedLium2 data sets, and achieve the state-of-the-art\nperformance on the test set in the scope of E2E models.', ""In this technical report I present my method for automatic synthetic dataset\ngeneration for object detection and demonstrate it on the video game League of\nLegends. This report furthermore serves as a handbook on how to automatically\ngenerate datasets and as an introduction on the dataset generation part of the\nLeagueAI framework. The LeagueAI framework is a software framework that\nprovides detailed information about the game League of Legends based on the\nsame input a human player would have, namely vision. The framework allows\nresearchers and enthusiasts to develop their own intelligent agents or to\nextract detailed information about the state of the game. A big problem of\nmachine vision applications usually is the laborious work of gathering large\namounts of hand labeled data. Thus, a crucial part of the vision pipeline of\nthe LeagueAI framework, the dataset generation, is presented in this report.\nThe method involves extracting image raw data from the game's 3D models and\ncombining them with the game background to create game-like synthetic images\nand to generate the corresponding labels automatically. In an experiment I\ncompared a model trained on synthetic data to a model trained on hand labeled\ndata and a model trained on a combined dataset. The model trained on the\nsynthetic data showed higher detection precision on more classes and more\nreliable tracking performance of the player character. The model trained on the\ncombined dataset did not perform better because of the different formats of the\nolder hand labeled dataset and the synthetic data."", 'Detection and recognition of text in natural images are two main problems in\nthe field of computer vision that have a wide variety of applications in\nanalysis of sports videos, autonomous driving, industrial automation, to name a\nfew. They face common challenging problems that are factors in how text is\nrepresented and affected by several environmental conditions. The current\nstate-of-the-art scene text detection and/or recognition methods have exploited\nthe witnessed advancement in deep learning architectures and reported a\nsuperior accuracy on benchmark datasets when tackling multi-resolution and\nmulti-oriented text. However, there are still several remaining challenges\naffecting text in the wild images that cause existing methods to underperform\ndue to there models are not able to generalize to unseen data and the\ninsufficient labeled data. Thus, unlike previous surveys in this field, the\nobjectives of this survey are as follows: first, offering the reader not only a\nreview on the recent advancement in scene text detection and recognition, but\nalso presenting the results of conducting extensive experiments using a unified\nevaluation framework that assesses pre-trained models of the selected methods\non challenging cases, and applies the same evaluation criteria on these\ntechniques. Second, identifying several existing challenges for detecting or\nrecognizing text in the wild images, namely, in-plane-rotation, multi-oriented\nand multi-resolution text, perspective distortion, illumination reflection,\npartial occlusion, complex fonts, and special characters. Finally, the paper\nalso presents insight into the potential research directions in this field to\naddress some of the mentioned challenges that are still encountering scene text\ndetection and recognition techniques.', 'In the big data era, the impetus to digitize the vast reservoirs of data\ntrapped in unstructured scanned documents such as invoices, bank documents and\ncourier receipts has gained fresh momentum. The scanning process often results\nin the introduction of artifacts such as background noise, blur due to camera\nmotion, watermarkings, coffee stains, or faded text. These artifacts pose many\nreadability challenges to current text recognition algorithms and significantly\ndegrade their performance. Existing learning based denoising techniques require\na dataset comprising of noisy documents paired with cleaned versions. In such\nscenarios, a model can be trained to generate clean documents from noisy\nversions. However, very often in the real world such a paired dataset is not\navailable, and all we have for training our denoising model are unpaired sets\nof noisy and clean images. This paper explores the use of GANs to generate\ndenoised versions of the noisy documents. In particular, where paired\ninformation is available, we formulate the problem as an image-to-image\ntranslation task i.e, translating a document from noisy domain ( i.e.,\nbackground noise, blurred, faded, watermarked ) to a target clean document\nusing Generative Adversarial Networks (GAN). However, in the absence of paired\nimages for training, we employed CycleGAN which is known to learn a mapping\nbetween the distributions of the noisy images to the denoised images using\nunpaired data to achieve image-to-image translation for cleaning the noisy\ndocuments. We compare the performance of CycleGAN for document cleaning tasks\nusing unpaired images with a Conditional GAN trained on paired data from the\nsame dataset. Experiments were performed on a public document dataset on which\ndifferent types of noise were artificially induced, results demonstrate that\nCycleGAN learns a more robust mapping from the space of noisy to clean\ndocuments.', 'Since the low quality of document images will greatly undermine the chances\nof success in automatic text recognition and analysis, it is necessary to\nassess the quality of document images uploaded in online business process, so\nas to reject those images of low quality. In this paper, we attempt to achieve\ndocument image quality assessment and our contributions are twofold. Firstly,\nsince document image quality assessment is more interested in text, we propose\na text line based framework to estimate document image quality, which is\ncomposed of three stages: text line detection, text line quality prediction,\nand overall quality assessment. Text line detection aims to find potential text\nlines with a detector. In the text line quality prediction stage, the quality\nscore is computed for each text line with a CNN-based prediction model. The\noverall quality of document images is finally assessed with the ensemble of all\ntext line quality. Secondly, to train the prediction model, a large-scale\ndataset, comprising 52,094 text line images, is synthesized with diverse\nattributes. For each text line image, a quality label is computed with a\npiece-wise function. To demonstrate the effectiveness of the proposed\nframework, comprehensive experiments are evaluated on two popular document\nimage quality assessment benchmarks. Our framework significantly outperforms\nthe state-of-the-art methods by large margins on the large and complicated\ndataset.', 'The demand for on-device document recognition systems increases in\nconjunction with the emergence of more strict privacy and security\nrequirements. In such systems, there is no data transfer from the end device to\na third-party information processing servers. The response time is vital to the\nuser experience of on-device document recognition. Combined with the\nunavailability of discrete GPUs, powerful CPUs, or a large RAM capacity on\nconsumer-grade end devices such as smartphones, the time limitations put\nsignificant constraints on the computational complexity of the applied\nalgorithms for on-device execution.\n  In this work, we consider document location in an image without prior\nknowledge of the document content or its internal structure. In accordance with\nthe published works, at least 5 systems offer solutions for on-device document\nlocation. All these systems use a location method which can be considered\nHough-based. The precision of such systems seems to be lower than that of the\nstate-of-the-art solutions which were not designed to account for the limited\ncomputational resources.\n  We propose an advanced Hough-based method. In contrast with other approaches,\nit accounts for the geometric invariants of the central projection model and\ncombines both edge and color features for document boundary detection. The\nproposed method allowed for the second best result for SmartDoc dataset in\nterms of precision, surpassed by U-net like neural network. When evaluated on a\nmore challenging MIDV-500 dataset, the proposed algorithm guaranteed the best\nprecision compared to published methods. Our method retained the applicability\nto on-device computations.', 'In the Bag-of-Words (BoW) model based image retrieval task, the precision of\nvisual matching plays a critical role in improving retrieval performance.\nConventionally, local cues of a keypoint are employed. However, such strategy\ndoes not consider the contextual evidences of a keypoint, a problem which would\nlead to the prevalence of false matches. To address this problem, this paper\ndefines ""true match"" as a pair of keypoints which are similar on three levels,\ni.e., local, regional, and global. Then, a principled probabilistic framework\nis established, which is capable of implicitly integrating discriminative cues\nfrom all these feature levels.\n  Specifically, the Convolutional Neural Network (CNN) is employed to extract\nfeatures from regional and global patches, leading to the so-called ""Deep\nEmbedding"" framework. CNN has been shown to produce excellent performance on a\ndozen computer vision tasks such as image classification and detection, but few\nworks have been done on BoW based image retrieval. In this paper, firstly we\nshow that proper pre-processing techniques are necessary for effective usage of\nCNN feature. Then, in the attempt to fit it into our model, a novel indexing\nstructure called ""Deep Indexing"" is introduced, which dramatically reduces\nmemory usage.\n  Extensive experiments on three benchmark datasets demonstrate that, the\nproposed Deep Embedding method greatly promotes the retrieval accuracy when CNN\nfeature is integrated. We show that our method is efficient in terms of both\nmemory and time cost, and compares favorably with the state-of-the-art methods.', 'Synthesizing high-quality images from text descriptions is a challenging\nproblem in computer vision and has many practical applications. Samples\ngenerated by existing text-to-image approaches can roughly reflect the meaning\nof the given descriptions, but they fail to contain necessary details and vivid\nobject parts. In this paper, we propose Stacked Generative Adversarial Networks\n(StackGAN) to generate 256x256 photo-realistic images conditioned on text\ndescriptions. We decompose the hard problem into more manageable sub-problems\nthrough a sketch-refinement process. The Stage-I GAN sketches the primitive\nshape and colors of the object based on the given text description, yielding\nStage-I low-resolution images. The Stage-II GAN takes Stage-I results and text\ndescriptions as inputs, and generates high-resolution images with\nphoto-realistic details. It is able to rectify defects in Stage-I results and\nadd compelling details with the refinement process. To improve the diversity of\nthe synthesized images and stabilize the training of the conditional-GAN, we\nintroduce a novel Conditioning Augmentation technique that encourages\nsmoothness in the latent conditioning manifold. Extensive experiments and\ncomparisons with state-of-the-arts on benchmark datasets demonstrate that the\nproposed method achieves significant improvements on generating photo-realistic\nimages conditioned on text descriptions.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', ""Recently, artificial intelligence technologies and algorithms have become a\nmajor focus for advancements in treatment planning for radiation therapy. As\nthese are starting to become incorporated into the clinical workflow, a major\nconcern from clinicians is not whether the model is accurate, but whether the\nmodel can express to a human operator when it does not know if its answer is\ncorrect. We propose to use Monte Carlo dropout (MCDO) and the bootstrap\naggregation (bagging) technique on deep learning models to produce uncertainty\nestimations for radiation therapy dose prediction. We show that both models are\ncapable of generating a reasonable uncertainty map, and, with our proposed\nscaling technique, creating interpretable uncertainties and bounds on the\nprediction and any relevant metrics. Performance-wise, bagging provides\nstatistically significant reduced loss value and errors in most of the metrics\ninvestigated in this study. The addition of bagging was able to further reduce\nerrors by another 0.34% for Dmean and 0.19% for Dmax, on average, when compared\nto the baseline framework. Overall, the bagging framework provided\nsignificantly lower MAE of 2.62, as opposed to the baseline framework's MAE of\n2.87. The usefulness of bagging, from solely a performance standpoint, does\nhighly depend on the problem and the acceptable predictive error, and its high\nupfront computational cost during training should be factored in to deciding\nwhether it is advantageous to use it. In terms of deployment with uncertainty\nestimations turned on, both frameworks offer the same performance time of about\n12 seconds. As an ensemble-based metaheuristic, bagging can be used with\nexisting machine learning architectures to improve stability and performance,\nand MCDO can be applied to any deep learning models that have dropout as part\nof their architecture."", 'Despite recent advances in the field of supervised deep learning for text\nline segmentation, unsupervised deep learning solutions are beginning to gain\npopularity. In this paper, we present an unsupervised deep learning method that\nembeds document image patches to a compact Euclidean space where distances\ncorrespond to a coarse text line pattern similarity. Once this space has been\nproduced, text line segmentation can be easily implemented using standard\ntechniques with the embedded feature vectors. To train the model, we extract\nrandom pairs of document image patches with the assumption that neighbour\npatches contain a similar coarse trend of text lines, whereas if one of them is\nrotated, they contain different coarse trends of text lines. Doing well on this\ntask requires the model to learn to recognize the text lines and their salient\nparts. The benefit of our approach is zero manual labelling effort. We evaluate\nthe method qualitatively and quantitatively on several variants of text line\nsegmentation datasets to demonstrate its effectivity.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'Annotating words in a historical document image archive for word image\nrecognition purpose demands time and skilled human resource (like historians,\npaleographers). In a real-life scenario, obtaining sample images for all\npossible words is also not feasible. However, Zero-shot learning methods could\naptly be used to recognize unseen/out-of-lexicon words in such historical\ndocument images. Based on previous state-of-the-art method for zero-shot word\nrecognition Pho(SC)Net, we propose a hybrid model based on the CTC framework\n(Pho(SC)-CTC) that takes advantage of the rich features learned by Pho(SC)Net\nfollowed by a connectionist temporal classification (CTC) framework to perform\nthe final classification. Encouraging results were obtained on two publicly\navailable historical document datasets and one synthetic handwritten dataset,\nwhich justifies the efficacy of Pho(SC)-CTC and Pho(SC)Net.', 'Recognizing text in natural images is a challenging task with many unsolved\nproblems. Different from those in documents, words in natural images often\npossess irregular shapes, which are caused by perspective distortion, curved\ncharacter placement, etc. We propose RARE (Robust text recognizer with\nAutomatic REctification), a recognition model that is robust to irregular text.\nRARE is a specially-designed deep neural network, which consists of a Spatial\nTransformer Network (STN) and a Sequence Recognition Network (SRN). In testing,\nan image is firstly rectified via a predicted Thin-Plate-Spline (TPS)\ntransformation, into a more ""readable"" image for the following SRN, which\nrecognizes text through a sequence recognition approach. We show that the model\nis able to recognize several types of irregular text, including perspective\ntext and curved text. RARE is end-to-end trainable, requiring only images and\nassociated text labels, making it convenient to train and deploy the model in\npractical systems. State-of-the-art or highly-competitive performance achieved\non several benchmarks well demonstrates the effectiveness of the proposed\nmodel.', 'Visual text recognition is undoubtedly one of the most extensively researched\ntopics in computer vision. Great progress have been made to date, with the\nlatest models starting to focus on the more practical ""in-the-wild"" setting.\nHowever, a salient problem still hinders practical deployment -- prior arts\nmostly struggle with recognising unseen (or rarely seen) character sequences.\nIn this paper, we put forward a novel framework to specifically tackle this\n""unseen"" problem. Our framework is iterative in nature, in that it utilises\npredicted knowledge of character sequences from a previous iteration, to\naugment the main network in improving the next prediction. Key to our success\nis a unique cross-modal variational autoencoder to act as a feedback module,\nwhich is trained with the presence of textual error distribution data. This\nmodule importantly translate a discrete predicted character space, to a\ncontinuous affine transformation parameter space used to condition the visual\nfeature map at next iteration. Experiments on common datasets have shown\ncompetitive performance over state-of-the-arts under the conventional setting.\nMost importantly, under the new disjoint setup where train-test labels are\nmutually exclusive, ours offers the best performance thus showcasing the\ncapability of generalising onto unseen words.', 'Recently, there has been a growing interest in end-to-end speech recognition\nthat directly transcribes speech to text without any predefined alignments. In\nthis paper, we explore the use of attention-based encoder-decoder model for\nMandarin speech recognition on a voice search task. Previous attempts have\nshown that applying attention-based encoder-decoder to Mandarin speech\nrecognition was quite difficult due to the logographic orthography of Mandarin,\nthe large vocabulary and the conditional dependency of the attention model. In\nthis paper, we use character embedding to deal with the large vocabulary.\nSeveral tricks are used for effective model training, including L2\nregularization, Gaussian weight noise and frame skipping. We compare two\nattention mechanisms and use attention smoothing to cover long context in the\nattention model. Taken together, these tricks allow us to finally achieve a\ncharacter error rate (CER) of 3.58% and a sentence error rate (SER) of 7.43% on\nthe MiTV voice search dataset. While together with a trigram language model,\nCER and SER reach 2.81% and 5.77%, respectively.', 'Recently, there has been an increasing interest in end-to-end speech\nrecognition that directly transcribes speech to text without any predefined\nalignments. One approach is the attention-based encoder-decoder framework that\nlearns a mapping between variable-length input and output sequences in one step\nusing a purely data-driven method. The attention model has often been shown to\nimprove the performance over another end-to-end approach, the Connectionist\nTemporal Classification (CTC), mainly because it explicitly uses the history of\nthe target character without any conditional independence assumptions. However,\nwe observed that the performance of the attention has shown poor results in\nnoisy condition and is hard to learn in the initial training stage with long\ninput sequences. This is because the attention model is too flexible to predict\nproper alignments in such cases due to the lack of left-to-right constraints as\nused in CTC. This paper presents a novel method for end-to-end speech\nrecognition to improve robustness and achieve fast convergence by using a joint\nCTC-attention model within the multi-task learning framework, thereby\nmitigating the alignment issue. An experiment on the WSJ and CHiME-4 tasks\ndemonstrates its advantages over both the CTC and attention-based\nencoder-decoder baselines, showing 5.4-14.6% relative improvements in Character\nError Rate (CER).', 'Many tasks are related to determining if a particular text string exists in\nan image. In this work, we propose a new framework that learns this task in an\nend-to-end way. The framework takes an image and a text string as input and\nthen outputs the probability of the text string being present in the image.\nThis is the first end-to-end framework that learns such relationships between\ntext and images in scene text area. The framework does not require explicit\nscene text detection or recognition and thus no bounding box annotations are\nneeded for it. It is also the first work in scene text area that tackles suh a\nweakly labeled problem. Based on this framework, we developed a model called\nGuided Attention. Our designed model achieves much better results than several\nstate-of-the-art scene text reading based solutions for a challenging Street\nView Business Matching task. The task tries to find correct business names for\nstorefront images and the dataset we collected for it is substantially larger,\nand more challenging than existing scene text dataset. This new real-world task\nprovides a new perspective for studying scene text related problems. We also\ndemonstrate the uniqueness of our task via a comparison between our problem and\na typical Visual Question Answering problem.', 'Arabic text recognition is a challenging task because of the cursive nature\nof Arabic writing system, its joint writing scheme, the large number of\nligatures and many other challenges. Deep Learning DL models achieved\nsignificant progress in numerous domains including computer vision and sequence\nmodelling. This paper presents a model that can recognize Arabic text that was\nprinted using multiple font types including fonts that mimic Arabic handwritten\nscripts. The proposed model employs a hybrid DL network that can recognize\nArabic printed text without the need for character segmentation. The model was\ntested on a custom dataset comprised of over two million word samples that were\ngenerated using 18 different Arabic font types. The objective of the testing\nprocess was to assess the model capability in recognizing a diverse set of\nArabic fonts representing a varied cursive styles. The model achieved good\nresults in recognizing characters and words and it also achieved promising\nresults in recognizing characters when it was tested on unseen data. The\nprepared model, the custom datasets and the toolkit for generating similar\ndatasets are made publicly available, these tools can be used to prepare models\nfor recognizing other font types as well as to further extend and enhance the\nperformance of the proposed model.', '""If I provide you a face image of mine (without telling you the actual age\nwhen I took the picture) and a large amount of face images that I crawled\n(containing labeled faces of different ages but not necessarily paired), can\nyou show me what I would look like when I am 80 or what I was like when I was\n5?"" The answer is probably a ""No."" Most existing face aging works attempt to\nlearn the transformation between age groups and thus would require the paired\nsamples as well as the labeled query image. In this paper, we look at the\nproblem from a generative modeling perspective such that no paired samples is\nrequired. In addition, given an unlabeled image, the generative model can\ndirectly produce the image with desired age attribute. We propose a conditional\nadversarial autoencoder (CAAE) that learns a face manifold, traversing on which\nsmooth age progression and regression can be realized simultaneously. In CAAE,\nthe face is first mapped to a latent vector through a convolutional encoder,\nand then the vector is projected to the face manifold conditional on age\nthrough a deconvolutional generator. The latent vector preserves personalized\nface features (i.e., personality) and the age condition controls progression\nvs. regression. Two adversarial networks are imposed on the encoder and\ngenerator, respectively, forcing to generate more photo-realistic faces.\nExperimental results demonstrate the appealing performance and flexibility of\nthe proposed framework by comparing with the state-of-the-art and ground truth.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'We introduce a novel type of text representation that preserves the 2D layout\nof a document. This is achieved by encoding each document page as a\ntwo-dimensional grid of characters. Based on this representation, we present a\ngeneric document understanding pipeline for structured documents. This pipeline\nmakes use of a fully convolutional encoder-decoder network that predicts a\nsegmentation mask and bounding boxes. We demonstrate its capabilities on an\ninformation extraction task from invoices and show that it significantly\noutperforms approaches based on sequential text or document images.', 'Document Layout Analysis is a fundamental step in Handwritten Text Processing\nsystems, from the extraction of the text lines to the type of zone it belongs\nto. We present a system based on artificial neural networks which is able to\ndetermine not only the baselines of text lines present in the document, but\nalso performs geometric and logic layout analysis of the document. Experiments\nin three different datasets demonstrate the potential of the method and show\ncompetitive results with respect to state-of-the-art methods.', 'Image denoising is always a challenging task in the field of computer vision\nand image processing. In this paper, we have proposed an encoder-decoder model\nwith direct attention, which is capable of denoising and reconstruct highly\ncorrupted images. Our model consists of an encoder and a decoder, where the\nencoder is a convolutional neural network and decoder is a multilayer Long\nShort-Term memory network. In the proposed model, the encoder reads an image\nand catches the abstraction of that image in a vector, where decoder takes that\nvector as well as the corrupted image to reconstruct a clean image. We have\ntrained our model on MNIST handwritten digit database after making lower half\nof every image as black as well as adding noise top of that. After a massive\ndestruction of the images where it is hard for a human to understand the\ncontent of those images, our model can retrieve that image with minimal error.\nOur proposed model has been compared with convolutional encoder-decoder, where\nour model has performed better at generating missing part of the images than\nconvolutional autoencoder.', 'Handwritten document image binarization is challenging due to high\nvariability in the written content and complex background attributes such as\npage style, paper quality, stains, shadow gradients, and non-uniform\nillumination. While the traditional thresholding methods do not effectively\ngeneralize on such challenging real-world scenarios, deep learning-based\nmethods have performed relatively well when provided with sufficient training\ndata. However, the existing datasets are limited in size and diversity. This\nwork proposes LS-HDIB - a large-scale handwritten document image binarization\ndataset containing over a million document images that span numerous real-world\nscenarios. Additionally, we introduce a novel technique that uses a combination\nof adaptive thresholding and seamless cloning methods to create the dataset\nwith accurate ground truths. Through an extensive quantitative and qualitative\nevaluation over eight different deep learning based models, we demonstrate the\nenhancement in the performance of these models when trained on the LS-HDIB\ndataset and tested on unseen images.', 'Identity documents recognition is an important sub-field of document\nanalysis, which deals with tasks of robust document detection, type\nidentification, text fields recognition, as well as identity fraud prevention\nand document authenticity validation given photos, scans, or video frames of an\nidentity document capture. Significant amount of research has been published on\nthis topic in recent years, however a chief difficulty for such research is\nscarcity of datasets, due to the subject matter being protected by security\nrequirements. A few datasets of identity documents which are available lack\ndiversity of document types, capturing conditions, or variability of document\nfield values. In addition, the published datasets were typically designed only\nfor a subset of document recognition problems, not for a complex identity\ndocument analysis. In this paper, we present a dataset MIDV-2020 which consists\nof 1000 video clips, 2000 scanned images, and 1000 photos of 1000 unique mock\nidentity documents, each with unique text field values and unique artificially\ngenerated faces, with rich annotation. For the presented benchmark dataset\nbaselines are provided for such tasks as document location and identification,\ntext fields recognition, and face detection. With 72409 annotated images in\ntotal, to the date of publication the proposed dataset is the largest publicly\navailable identity documents dataset with variable artificially generated data,\nand we believe that it will prove invaluable for advancement of the field of\ndocument analysis and recognition. The dataset is available for download at\nftp://smartengines.com/midv-2020 and http://l3i-share.univ-lr.fr .', 'Despite significant progress on current state-of-the-art image generation\nmodels, synthesis of document images containing multiple and complex object\nlayouts is a challenging task. This paper presents a novel approach, called\nDocSynth, to automatically synthesize document images based on a given layout.\nIn this work, given a spatial layout (bounding boxes with object categories) as\na reference by the user, our proposed DocSynth model learns to generate a set\nof realistic document images consistent with the defined layout. Also, this\nframework has been adapted to this work as a superior baseline model for\ncreating synthetic document image datasets for augmenting real data during\ntraining for document layout analysis tasks. Different sets of learning\nobjectives have been also used to improve the model performance.\nQuantitatively, we also compare the generated results of our model with real\ndata using standard evaluation metrics. The results highlight that our model\ncan successfully generate realistic and diverse document images with multiple\nobjects. We also present a comprehensive qualitative analysis summary of the\ndifferent scopes of synthetic image generation tasks. Lastly, to our knowledge\nthis is the first work of its kind.', ""Class labels have been empirically shown useful in improving the sample\nquality of generative adversarial nets (GANs). In this paper, we mathematically\nstudy the properties of the current variants of GANs that make use of class\nlabel information. With class aware gradient and cross-entropy decomposition,\nwe reveal how class labels and associated losses influence GAN's training.\nBased on that, we propose Activation Maximization Generative Adversarial\nNetworks (AM-GAN) as an advanced solution. Comprehensive experiments have been\nconducted to validate our analysis and evaluate the effectiveness of our\nsolution, where AM-GAN outperforms other strong baselines and achieves\nstate-of-the-art Inception Score (8.91) on CIFAR-10. In addition, we\ndemonstrate that, with the Inception ImageNet classifier, Inception Score\nmainly tracks the diversity of the generator, and there is, however, no\nreliable evidence that it can reflect the true sample quality. We thus propose\na new metric, called AM Score, to provide a more accurate estimation of the\nsample quality. Our proposed model also outperforms the baseline methods in the\nnew metric."", 'We study how to train a student deep neural network for visual recognition by\ndistilling knowledge from a blackbox teacher model in a data-efficient manner.\nProgress on this problem can significantly reduce the dependence on large-scale\ndatasets for learning high-performing visual recognition models. There are two\nmajor challenges. One is that the number of queries into the teacher model\nshould be minimized to save computational and/or financial costs. The other is\nthat the number of images used for the knowledge distillation should be small;\notherwise, it violates our expectation of reducing the dependence on\nlarge-scale datasets. To tackle these challenges, we propose an approach that\nblends mixup and active learning. The former effectively augments the few\nunlabeled images by a big pool of synthetic images sampled from the convex hull\nof the original images, and the latter actively chooses from the pool hard\nexamples for the student neural network and query their labels from the teacher\nmodel. We validate our approach with extensive experiments.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'This work analyses the impact of self-supervised pre-training on document\nimages in the context of document image classification. While previous\napproaches explore the effect of self-supervision on natural images, we show\nthat patch-based pre-training performs poorly on document images because of\ntheir different structural properties and poor intra-sample semantic\ninformation. We propose two context-aware alternatives to improve performance\non the Tobacco-3482 image classification task. We also propose a novel method\nfor self-supervision, which makes use of the inherent multi-modality of\ndocuments (image and text), which performs better than other popular\nself-supervised methods, including supervised ImageNet pre-training, on\ndocument image classification scenarios with a limited amount of data.', 'Generative Adversarial Networks have been crucial in the developments made in\nunsupervised learning in recent times. Exemplars of image synthesis from text\nor other images, these networks have shown remarkable improvements over\nconventional methods in terms of performance. Trained on the adversarial\ntraining philosophy, these networks aim to estimate the potential distribution\nfrom the real data and then use this as input to generate the synthetic data.\nBased on this fundamental principle, several frameworks can be generated that\nare paragon implementations in several real-life applications such as art\nsynthesis, generation of high resolution outputs and synthesis of images from\nhuman drawn sketches, to name a few. While theoretically GANs present better\nresults and prove to be an improvement over conventional methods in many\nfactors, the implementation of these frameworks for dedicated applications\nremains a challenge. This study explores and presents a taxonomy of these\nframeworks and their use in various image to image synthesis and text to image\nsynthesis applications. The basic GANs, as well as a variety of different niche\nframeworks, are critically analyzed. The advantages of GANs for image\ngeneration over conventional methods as well their disadvantages amongst other\nframeworks are presented. The future applications of GANs in industries such as\nhealthcare, art and entertainment are also discussed.', 'Computation of document image quality metrics often depends upon the\navailability of a ground truth image corresponding to the document. This limits\nthe applicability of quality metrics in applications such as hyperparameter\noptimization of image processing algorithms that operate on-the-fly on unseen\ndocuments. This work proposes the use of surrogate models to learn the behavior\nof a given document quality metric on existing datasets where ground truth\nimages are available. The trained surrogate model can later be used to predict\nthe metric value on previously unseen document images without requiring access\nto ground truth images. The surrogate model is empirically evaluated on the\nDocument Image Binarization Competition (DIBCO) and the Handwritten Document\nImage Binarization Competition (H-DIBCO) datasets.', 'Scene text recognition has been a hot topic in computer vision. Recent\nmethods adopt the attention mechanism for sequence prediction which achieve\nconvincing results. However, we argue that the existing attention mechanism\nfaces the problem of attention diffusion, in which the model may not focus on a\ncertain character area. In this paper, we propose Gaussian Constrained\nAttention Network to deal with this problem. It is a 2D attention-based method\nintegrated with a novel Gaussian Constrained Refinement Module, which predicts\nan additional Gaussian mask to refine the attention weights. Different from\nadopting an additional supervision on the attention weights simply, our\nproposed method introduces an explicit refinement. In this way, the attention\nweights will be more concentrated and the attention-based recognition network\nachieves better performance. The proposed Gaussian Constrained Refinement\nModule is flexible and can be applied to existing attention-based methods\ndirectly. The experiments on several benchmark datasets demonstrate the\neffectiveness of our proposed method. Our code has been available at\nhttps://github.com/Pay20Y/GCAN.', 'Automated recognition of texts in scenes has been a research challenge for\nyears, largely due to the arbitrary variation of text appearances in\nperspective distortion, text line curvature, text styles and different types of\nimaging artifacts. The recent deep networks are capable of learning robust\nrepresentations with respect to imaging artifacts and text style changes, but\nstill face various problems while dealing with scene texts with perspective and\ncurvature distortions. This paper presents an end-to-end trainable scene text\nrecognition system (ESIR) that iteratively removes perspective distortion and\ntext line curvature as driven by better scene text recognition performance. An\ninnovative rectification network is developed which employs a novel\nline-fitting transformation to estimate the pose of text lines in scenes. In\naddition, an iterative rectification pipeline is developed where scene text\ndistortions are corrected iteratively towards a fronto-parallel view. The ESIR\nis also robust to parameter initialization and the training needs only scene\ntext images and word-level annotations as required by most scene text\nrecognition systems. Extensive experiments over a number of public datasets\nshow that the proposed ESIR is capable of rectifying scene text distortions\naccurately, achieving superior recognition performance for both normal scene\ntext images and those suffering from perspective and curvature distortions.', 'Recent progress has been made on developing a unified framework for joint\ntext detection and recognition in natural images, but existing joint models\nwere mostly built on two-stage framework by involving ROI pooling, which can\ndegrade the performance on recognition task. In this work, we propose\nconvolutional character networks, referred as CharNet, which is an one-stage\nmodel that can process two tasks simultaneously in one pass. CharNet directly\noutputs bounding boxes of words and characters, with corresponding character\nlabels. We utilize character as basic element, allowing us to overcome the main\ndifficulty of existing approaches that attempted to optimize text detection\njointly with a RNN-based recognition branch. In addition, we develop an\niterative character detection approach able to transform the ability of\ncharacter detection learned from synthetic data to real-world images. These\ntechnical improvements result in a simple, compact, yet powerful one-stage\nmodel that works reliably on multi-orientation and curved text. We evaluate\nCharNet on three standard benchmarks, where it consistently outperforms the\nstate-of-the-art approaches [25, 24] by a large margin, e.g., with improvements\nof 65.33%->71.08% (with generic lexicon) on ICDAR 2015, and 54.0%->69.23% on\nTotal-Text, on end-to-end text recognition. Code is available at:\nhttps://github.com/MalongTech/research-charnet.', 'State-of-the-art approaches for image captioning require supervised training\ndata consisting of captions with paired image data. These methods are typically\nunable to use unsupervised data such as textual data with no corresponding\nimages, which is a much more abundant commodity. We here propose a novel way of\nusing such textual data by artificially generating missing visual information.\nWe evaluate this learning approach on a newly designed model that detects\nvisual concepts present in an image and feed them to a reviewer-decoder\narchitecture with an attention mechanism. Unlike previous approaches that\nencode visual concepts using word embeddings, we instead suggest using regional\nimage features which capture more intrinsic information. The main benefit of\nthis architecture is that it synthesizes meaningful thought vectors that\ncapture salient image properties and then applies a soft attentive decoder to\ndecode the thought vectors and generate image captions. We evaluate our model\non both Microsoft COCO and Flickr30K datasets and demonstrate that this model\ncombined with our semi-supervised learning method can largely improve\nperformance and help the model to generate more accurate and diverse captions.', 'Spelling normalization for low resource languages is a challenging task\nbecause the patterns are hard to predict and large corpora are usually required\nto collect enough examples. This work shows a comparison of a neural model and\ncharacter language models with varying amounts on target language data. Our\nusage scenario is interactive correction with nearly zero amounts of training\nexamples, improving models as more data is collected, for example within a chat\napp. Such models are designed to be incrementally improved as feedback is given\nfrom users. In this work, we design a knowledge-base and prediction model\nembedded system for spelling correction in low-resource languages. Experimental\nresults on multiple languages show that the model could become effective with a\nsmall amount of data. We perform experiments on both natural and synthetic\ndata, as well as on data from two endangered languages (Ainu and Griko). Last,\nwe built a prototype system that was used for a small case study on Hinglish,\nwhich further demonstrated the suitability of our approach in real world\nscenarios.', 'This work addresses the problem of Question Answering (QA) on handwritten\ndocument collections. Unlike typical QA and Visual Question Answering (VQA)\nformulations where the answer is a short text, we aim to locate a document\nsnippet where the answer lies. The proposed approach works without recognizing\nthe text in the documents. We argue that the recognition-free approach is\nsuitable for handwritten documents and historical collections where robust text\nrecognition is often difficult. At the same time, for human users, document\nimage snippets containing answers act as a valid alternative to textual\nanswers. The proposed approach uses an off-the-shelf deep embedding network\nwhich can project both textual words and word images into a common sub-space.\nThis embedding bridges the textual and visual domains and helps us retrieve\ndocument snippets that potentially answer a question. We evaluate results of\nthe proposed approach on two new datasets: (i) HW-SQuAD: a synthetic,\nhandwritten document image counterpart of SQuAD1.0 dataset and (ii) BenthamQA:\na smaller set of QA pairs defined on documents from the popular Bentham\nmanuscripts collection. We also present a thorough analysis of the proposed\nrecognition-free approach compared to a recognition-based approach which uses\ntext recognized from the images using an OCR. Datasets presented in this work\nare available to download at docvqa.org', 'Cloud computing offers the potential to help scientists to process massive\nnumber of computing resources often required in machine learning application\nsuch as computer vision problems. This proposal would like to show that which\nbenefits can be obtained from cloud in order to help medical image analysis\nusers (including scientists, clinicians, and research institutes). As security\nand privacy of algorithms are important for most of algorithms inventors, these\nalgorithms can be hidden in a cloud to allow the users to use the algorithms as\na package without any access to see/change their inside. In another word, in\nthe user part, users send their images to the cloud and configure the algorithm\nvia an interface. In the cloud part, the algorithms are applied to this image\nand the results are returned back to the user. My proposal has two parts: (1)\ninvestigate the potential of cloud computing for computer vision problems and\n(2) study the components of a proposed cloud-based framework for medical image\nanalysis application and develop them (depending on the length of the\ninternship). The investigation part will involve a study on several aspects of\nthe problem including security, usability (for medical end users of the\nservice), appropriate programming abstractions for vision problems, scalability\nand resource requirements. In the second part of this proposal I am going to\nthoroughly study of the proposed framework components and their relations and\ndevelop them. The proposed cloud-based framework includes an integrated\nenvironment to enable scientists and clinicians to access to the previous and\ncurrent medical image analysis algorithms using a handful user interface\nwithout any access to the algorithm codes and procedures.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'A lot of research has been devoted to identity documents analysis and\nrecognition on mobile devices. However, no publicly available datasets designed\nfor this particular problem currently exist. There are a few datasets which are\nuseful for associated subtasks but in order to facilitate a more comprehensive\nscientific and technical approach to identity document recognition more\nspecialized datasets are required. In this paper we present a Mobile Identity\nDocument Video dataset (MIDV-500) consisting of 500 video clips for 50\ndifferent identity document types with ground truth which allows to perform\nresearch in a wide scope of document analysis problems. The paper presents\ncharacteristics of the dataset and evaluation results for existing methods of\nface detection, text line recognition, and document fields data extraction.\nSince an important feature of identity documents is their sensitiveness as they\ncontain personal data, all source document images used in MIDV-500 are either\nin public domain or distributed under public copyright licenses.\n  The main goal of this paper is to present a dataset. However, in addition and\nas a baseline, we present evaluation results for existing methods for face\ndetection, text line recognition, and document data extraction, using the\npresented dataset.\n  (The dataset is available for download at ftp://smartengines.com/midv-500/.)', 'Graphical elements: particularly tables and figures contain a visual summary\nof the most valuable information contained in a document. Therefore,\nlocalization of such graphical objects in the document images is the initial\nstep to understand the content of such graphical objects or document images. In\nthis paper, we present a novel end-to-end trainable deep learning based\nframework to localize graphical objects in the document images called as\nGraphical Object Detection (GOD). Our framework is data-driven and does not\nrequire any heuristics or meta-data to locate graphical objects in the document\nimages. The GOD explores the concept of transfer learning and domain adaptation\nto handle scarcity of labeled training images for graphical object detection\ntask in the document images. Performance analysis carried out on the various\npublic benchmark data sets: ICDAR-2013, ICDAR-POD2017,and UNLV shows that our\nmodel yields promising results as compared to state-of-the-art techniques.', 'When digitizing a document into an image, it is common to include a\nsurrounding border region to visually indicate that the entire document is\npresent in the image. However, this border should be removed prior to automated\nprocessing. In this work, we present a deep learning based system, PageNet,\nwhich identifies the main page region in an image in order to segment content\nfrom both textual and non-textual border noise. In PageNet, a Fully\nConvolutional Network obtains a pixel-wise segmentation which is post-processed\ninto the output quadrilateral region. We evaluate PageNet on 4 collections of\nhistorical handwritten documents and obtain over 94% mean intersection over\nunion on all datasets and approach human performance on 2 of these collections.\nAdditionally, we show that PageNet can segment documents that are overlayed on\ntop of other documents.', 'There is a large collection of Handwritten English paper documents of\nHistorical and Scientific importance. But paper documents are not recognized\ndirectly by computer. Hence the closest way of indexing these documents is by\nstoring their document digital image. Hence a large database of document images\ncan replace the paper documents. But the document and data corresponding to\neach image cannot be directly recognized by the computer.\n  This paper applies the technique of word spotting using Modified Character\nShape Code to Handwritten English document images for quick and efficient query\nsearch of words on a database of document images. It is different from other\nWord Spotting techniques as it implements two level of selection for word\nsegments to match search query. First based on word size and then based on\ncharacter shape code of query. It makes the process faster and more efficient\nand reduces the need of multiple pre-processing.', 'This paper describes the short-term competition on the Components\nSegmentation Task of Document Photos that was prepared in the context of the\n16th International Conference on Document Analysis and Recognition (ICDAR\n2021). This competition aims to bring together researchers working in the field\nof identification document image processing and provides them a suitable\nbenchmark to compare their techniques on the component segmentation task of\ndocument images. Three challenge tasks were proposed entailing different\nsegmentation assignments to be performed on a provided dataset. The collected\ndata are from several types of Brazilian ID documents, whose personal\ninformation was conveniently replaced. There were 16 participants whose results\nobtained for some or all the three tasks show different rates for the adopted\nmetrics, like Dice Similarity Coefficient ranging from 0.06 to 0.99. Different\nDeep Learning models were applied by the entrants with diverse strategies to\nachieve the best results in each of the tasks. Obtained results show that the\ncurrently applied methods for solving one of the proposed tasks (document\nboundary detection) are already well established. However, for the other two\nchallenge tasks (text zone and handwritten sign detection) research and\ndevelopment of more robust approaches are still required to achieve acceptable\nresults.', 'Documents often exhibit various forms of degradation, which make it hard to\nbe read and substantially deteriorate the performance of an OCR system. In this\npaper, we propose an effective end-to-end framework named Document Enhancement\nGenerative Adversarial Networks (DE-GAN) that uses the conditional GANs (cGANs)\nto restore severely degraded document images. To the best of our knowledge,\nthis practice has not been studied within the context of generative adversarial\ndeep networks. We demonstrate that, in different tasks (document clean up,\nbinarization, deblurring and watermark removal), DE-GAN can produce an enhanced\nversion of the degraded document with a high quality. In addition, our approach\nprovides consistent improvements compared to state-of-the-art methods over the\nwidely used DIBCO 2013, DIBCO 2017 and H-DIBCO 2018 datasets, proving its\nability to restore a degraded document image to its ideal condition. The\nobtained results on a wide variety of degradation reveal the flexibility of the\nproposed model to be exploited in other document enhancement problems.', 'The antinuclear antibody detection with human epithelial cells is a popular\napproach for autoimmune diseases diagnosis. The manual evaluation demands time,\neffort and capital, and automation in screening can greatly aid the physicians\nin these respects. In this work, we employ simple, efficient and visually more\ninterpretable, class-specific features which defined based on the visual\ncharacteristics of each class. We believe that defining features with a good\nvisual interpretation, is indeed important in a scenario, where such an\napproach is used in an interactive CAD system for pathologists. Considering\nthat problem consists of few classes, and our rather simplistic feature\ndefinitions, frameworks can be structured as hierarchies of various binary\nclassifiers. These variants include frameworks which are earlier explored and\nsome which are not explored for this task. We perform various experiments which\ninclude traditional texture features and demonstrate the effectiveness of\nclass-specific features in various frameworks. We make insightful comparisons\nbetween different types of classification frameworks given their silent aspects\nand pros and cons over each other. We also demonstrate an experiment with only\nintermediates samples for testing. The proposed work yields encouraging results\nwith respect to the state-of-the-art and highlights the role of class-specific\nfeatures in different classification frameworks.', 'Video text spotting is still an important research topic due to its various\nreal-applications. Previous approaches usually fall into the four-staged\npipeline: text detection in individual images, framewisely recognizing\nlocalized text regions, tracking text streams and generating final results with\ncomplicated post-processing skills, which might suffer from the huge\ncomputational cost as well as the interferences of low-quality text. In this\npaper, we propose a fast and robust video text spotting framework by only\nrecognizing the localized text one-time instead of frame-wisely recognition.\nSpecifically, we first obtain text regions in videos with a well-designed\nspatial-temporal detector. Then we concentrate on developing a novel text\nrecommender for selecting the highest-quality text from text streams and only\nrecognizing the selected ones. Here, the recommender assembles text tracking,\nquality scoring and recognition into an end-to-end trainable module, which not\nonly avoids the interferences from low-quality text but also dramatically\nspeeds up the video text spotting process. In addition, we collect a larger\nscale video text dataset (LSVTD) for promoting the video text spotting\ncommunity, which contains 100 text videos from 22 different real-life\nscenarios. Extensive experiments on two public benchmarks show that our method\ngreatly speeds up the recognition process averagely by 71 times compared with\nthe frame-wise manner, and also achieves the remarkable state-of-the-art.', 'Machine learning techniques have excelled in the automatic semantic analysis\nof images, reaching human-level performances on challenging benchmarks. Yet,\nthe semantic analysis of videos remains challenging due to the significantly\nhigher dimensionality of the input data, respectively, the significantly higher\nneed for annotated training examples. By studying the automatic recognition of\nGerman sign language videos, we demonstrate that on the relatively scarce\ntraining data of 2.800 videos, modern deep learning architectures for video\nanalysis (such as ResNeXt) along with transfer learning on large gesture\nrecognition tasks, can achieve about 75% character accuracy. Considering that\nthis leaves us with a probability of under 25% that a 5 letter word is spelled\ncorrectly, spell-correction systems are crucial for producing readable outputs.\nThe contribution of this paper is to propose a convolutional neural network for\nspell-correction that expects the softmax outputs of the character recognition\nnetwork (instead of a misspelled word) as an input. We demonstrate that purely\nlearning on softmax inputs in combination with scarce training data yields\noverfitting as the network learns the inputs by heart. In contrast, training\nthe network on several variants of the logits of the classification output i.e.\nscaling by a constant factor, adding of random noise, mixing of softmax and\nhardmax inputs or purely training on hardmax inputs, leads to better\ngeneralization while benefitting from the significant information hidden in\nthese outputs (that have 98% top-5 accuracy), yielding a readable text despite\nthe comparably low character accuracy.', 'Automatic document content processing is affected by artifacts caused by the\nshape of the paper, non-uniform and diverse color of lighting conditions.\nFully-supervised methods on real data are impossible due to the large amount of\ndata needed. Hence, the current state of the art deep learning models are\ntrained on fully or partially synthetic images. However, document shadow or\nshading removal results still suffer because: (a) prior methods rely on\nuniformity of local color statistics, which limit their application on\nreal-scenarios with complex document shapes and textures and; (b) synthetic or\nhybrid datasets with non-realistic, simulated lighting conditions are used to\ntrain the models. In this paper we tackle these problems with our two main\ncontributions. First, a physically constrained learning-based method that\ndirectly estimates document reflectance based on intrinsic image formation\nwhich generalizes to challenging illumination conditions. Second, a new dataset\nthat clearly improves previous synthetic ones, by adding a large range of\nrealistic shading and diverse multi-illuminant conditions, uniquely customized\nto deal with documents in-the-wild. The proposed architecture works in a\nself-supervised manner where only the synthetic texture is used as a weak\ntraining signal (obviating the need for very costly ground truth with\ndisentangled versions of shading and reflectance). The proposed approach leads\nto a significant generalization of document reflectance estimation in real\nscenes with challenging illumination. We extensively evaluate on the real\nbenchmark datasets available for intrinsic image decomposition and document\nshadow removal tasks. Our reflectance estimation scheme, when used as a\npre-processing step of an OCR pipeline, shows a 26% improvement of character\nerror rate (CER), thus, proving the practical applicability.', 'Document image binarization is the initial step and a crucial in many\ndocument analysis and recognition scheme. In fact, it is still a relevant\nresearch subject and a fundamental challenge due to its importance and\ninfluence. This paper provides an original multi-phases system that hybridizes\nvarious efficient image thresholding methods in order to get the best\nbinarization output. First, to improve contrast in particularly defective\nimages, the application of CLAHE algorithm is suggested and justified. We then\nuse a cooperative technique to segment image into two separated classes. At the\nend, a special transformation is applied for the purpose of removing scattered\nnoise and of correcting characters forms. Experimentations demonstrate the\nprecision and the robustness of our framework applied on historical degraded\ndocuments images within three benchmarks compared to other noted methods.', 'The digital camera captured document images may often be warped and distorted\ndue to different camera angles or document surfaces. A robust technique is\nneeded to solve this kind of distortion. The research on dewarping of the\ndocument suffers due to the limited availability of benchmark public dataset.\nIn recent times, deep learning based approaches are used to solve the problems\naccurately. To train most of the deep neural networks a large number of\ndocument images is required and generating such a large volume of document\nimages manually is difficult. In this paper, we propose a technique to generate\na synthetic warped image from a flat-bedded scanned document image. It is done\nby calculating warping factors for each pixel position using two warping\nposition parameters (WPP) and eight warping control parameters (WCP). These\nparameters can be specified as needed depending upon the desired warping. The\nresults are compared with similar real captured images both qualitative and\nquantitative way.', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.', 'We present a novel data generation tool for document processing. The tool\nfocuses on providing a maximal level of visual information in a normal type\ndocument, ranging from character position to paragraph-level position. It also\nenables working with a large dataset on low-resource languages as well as\nproviding a mean of processing thorough full-level information of the\ndocumented text. The data generation tools come with a dataset of 320000\nVietnamese synthetic document images and an instruction to generate a dataset\nof similar size in other languages. The repository can be found at:\nhttps://github.com/tson1997/SDL-Document-Image-Generation', 'Textual information in a captured scene plays an important role in scene\ninterpretation and decision making. Though there exist methods that can\nsuccessfully detect and interpret complex text regions present in a scene, to\nthe best of our knowledge, there is no significant prior work that aims to\nmodify the textual information in an image. The ability to edit text directly\non images has several advantages including error correction, text restoration\nand image reusability. In this paper, we propose a method to modify text in an\nimage at character-level. We approach the problem in two stages. At first, the\nunobserved character (target) is generated from an observed character (source)\nbeing modified. We propose two different neural network architectures - (a)\nFANnet to achieve structural consistency with source font and (b) Colornet to\npreserve source color. Next, we replace the source character with the generated\ncharacter maintaining both geometric and visual consistency with neighboring\ncharacters. Our method works as a unified platform for modifying text in\nimages. We present the effectiveness of our method on COCO-Text and ICDAR\ndatasets both qualitatively and quantitatively.', 'This paper presents results of Document Visual Question Answering Challenge\norganized as part of ""Text and Documents in the Deep Learning Era"" workshop, in\nCVPR 2020. The challenge introduces a new problem - Visual Question Answering\non document images. The challenge comprised two tasks. The first task concerns\nwith asking questions on a single document image. On the other hand, the second\ntask is set as a retrieval task where the question is posed over a collection\nof images. For the task 1 a new dataset is introduced comprising 50,000\nquestions-answer(s) pairs defined over 12,767 document images. For task 2\nanother dataset has been created comprising 20 questions over 14,362 document\nimages which share the same document template.', 'Segmentation of a text-document into lines, words and characters, which is\nconsidered to be the crucial pre-processing stage in Optical Character\nRecognition (OCR) is traditionally carried out on uncompressed documents,\nalthough most of the documents in real life are available in compressed form,\nfor the reasons such as transmission and storage efficiency. However, this\nimplies that the compressed image should be decompressed, which indents\nadditional computing resources. This limitation has motivated us to take up\nresearch in document image analysis using compressed documents. In this paper,\nwe think in a new way to carry out segmentation at line, word and character\nlevel in run-length compressed printed-text-documents. We extract the\nhorizontal projection profile curve from the compressed file and using the\nlocal minima points perform line segmentation. However, tracing vertical\ninformation which leads to tracking words-characters in a run-length compressed\nfile is not very straight forward. Therefore, we propose a novel technique for\ncarrying out simultaneous word and character segmentation by popping out column\nruns from each row in an intelligent sequence. The proposed algorithms have\nbeen validated with 1101 text-lines, 1409 words and 7582 characters from a\ndata-set of 35 noise and skew free compressed documents of Bengali, Kannada and\nEnglish Scripts.', 'Numerous activities in our daily life, including transactions, access to\nservices and transportation, require us to verify who we are by showing our ID\ndocuments containing face images, e.g. passports and driver licenses. An\nautomatic system for matching ID document photos to live face images in real\ntime with high accuracy would speedup the verification process and remove the\nburden on human operators. In this paper, by employing the transfer learning\ntechnique, we propose a new method, DocFace, to train a domain-specific network\nfor ID document photo matching without a large dataset. Compared with the\nbaseline of applying existing methods for general face recognition to this\nproblem, our method achieves considerable improvement. A cross validation on an\nID-Selfie dataset shows that DocFace improves the TAR from 61.14% to 92.77% at\nFAR=0.1%. Experimental results also indicate that given more training data, a\nviable system for automatic ID document photo matching can be developed and\ndeployed.', 'Multimodal prediction results are essential for trajectory prediction task as\nthere is no single correct answer for the future. Previous frameworks can be\ndivided into three categories: regression, generation and classification\nframeworks. However, these frameworks have weaknesses in different aspects so\nthat they cannot model the multimodal prediction task comprehensively. In this\npaper, we present a novel insight along with a brand-new prediction framework\nby formulating multimodal prediction into three steps: modality clustering,\nclassification and synthesis, and address the shortcomings of earlier\nframeworks. Exhaustive experiments on popular benchmarks have demonstrated that\nour proposed method surpasses state-of-the-art works even without introducing\nsocial and map information. Specifically, we achieve 19.2% and 20.8%\nimprovement on ADE and FDE respectively on ETH/UCY dataset. Our code will be\nmade publicly availabe.', 'In Convolutional Neural Network (CNN) based image processing, most studies\npropose networks that are optimized to single-level (or single-objective);\nthus, they underperform on other levels and must be retrained for delivery of\noptimal performance. Using multiple models to cover multiple levels involves\nvery high computational costs. To solve these problems, recent approaches train\nnetworks on two different levels and propose their own interpolation methods to\nenable arbitrary intermediate levels. However, many of them fail to generalize\nor have certain side effects in practical usage. In this paper, we define these\nframeworks as network tuning and interpolation and propose a novel module for\ncontinuous-level learning, called Filter Transition Network (FTN). This module\nis a structurally smoother module than existing ones. Therefore, the frameworks\nwith FTN generalize well across various tasks and networks and cause fewer\nundesirable side effects. For stable learning of FTN, we additionally propose a\nmethod to initialize non-linear neural network layers with identity mappings.\nExtensive results for various image processing tasks indicate that the\nperformance of FTN is comparable in multiple continuous levels, and is\nsignificantly smoother and lighter than that of other frameworks.', 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', 'Everyone knows that thousand of words are represented by a single image. As a\nresult image search has become a very popular mechanism for the Web searchers.\nImage search means, the search results are produced by the search engine should\nbe a set of images along with their Web page Unified Resource Locator. Now Web\nsearcher can perform two types of image search, they are Text to Image and\nImage to Image search. In Text to Image search, search query should be a text.\nBased on the input text data system will generate a set of images along with\ntheir Web page URL as an output. On the other hand, in Image to Image search,\nsearch query should be an image and based on this image system will generate a\nset of images along with their Web page URL as an output. According to the\ncurrent scenarios, Text to Image search mechanism always not returns perfect\nresult. It matches the text data and then displays the corresponding images as\nan output, which is not always perfect. To resolve this problem, Web\nresearchers have introduced the Image to Image search mechanism. In this paper,\nwe have also proposed an alternate approach of Image to Image search mechanism\nusing Histogram.', 'Document segmentation is a method of rending the document into distinct\nregions. A document is an assortment of information and a standard mode of\nconveying information to others. Pursuance of data from documents involves ton\nof human effort, time intense and might severely prohibit the usage of data\nsystems. So, automatic information pursuance from the document has become a big\nissue. It is been shown that document segmentation will facilitate to beat such\nproblems. This paper proposes a new approach to segment and classify the\ndocument regions as text, image, drawings and table. Document image is divided\ninto blocks using Run length smearing rule and features are extracted from\nevery blocks. Discipulus tool has been used to construct the Genetic\nprogramming based classifier model and located 97.5% classification accuracy.', 'The image processing community has witnessed remarkable advances in enhancing\nand restoring images. Nevertheless, restoring the visual quality of underwater\nimages remains a great challenge. End-to-end frameworks might fail to enhance\nthe visual quality of underwater images since in several scenarios it is not\nfeasible to provide the ground truth of the scene radiance. In this work, we\npropose a CNN-based approach that does not require ground truth data since it\nuses a set of image quality metrics to guide the restoration learning process.\nThe experiments showed that our method improved the visual quality of\nunderwater images preserving their edges and also performed well considering\nthe UCIQE metric.', 'Digital camera and mobile document image acquisition are new trends arising\nin the world of Optical Character Recognition and text detection. In some\ncases, such process integrates many distortions and produces poorly scanned\ntext or text-photo images and natural images, leading to an unreliable OCR\ndigitization. In this paper, we present a novel nonparametric and unsupervised\nmethod to compensate for undesirable document image distortions aiming to\noptimally improve OCR accuracy. Our approach relies on a very efficient stack\nof document image enhancing techniques to recover deformation of the entire\ndocument image. First, we propose a local brightness and contrast adjustment\nmethod to effectively handle lighting variations and the irregular distribution\nof image illumination. Second, we use an optimized greyscale conversion\nalgorithm to transform our document image to greyscale level. Third, we sharpen\nthe useful information in the resulting greyscale image using Un-sharp Masking\nmethod. Finally, an optimal global binarization approach is used to prepare the\nfinal document image to OCR recognition. The proposed approach can\nsignificantly improve text detection rate and optical character recognition\naccuracy. To demonstrate the efficiency of our approach, an exhaustive\nexperimentation on a standard dataset is presented.', 'Line separators are used to segregate text-lines from one another in document\nimage analysis. Finding the separator points at every line terminal in a\ndocument image would enable text-line segmentation. In particular, identifying\nthe separators in handwritten text could be a thrilling exercise. Obviously it\nwould be challenging to perform this in the compressed version of a document\nimage and that is the proposed objective in this research. Such an effort would\nprevent the computational burden of decompressing a document for text-line\nsegmentation. Since document images are generally compressed using run length\nencoding (RLE) technique as per the CCITT standards, the first column in the\nRLE will be a white column. The value (depth) in the white column is very low\nwhen a particular line is a text line and the depth could be larger at the\npoint of text line separation. A longer consecutive sequence of such larger\ndepth should indicate the gap between the text lines, which provides the\nseparator region. In case of over separation and under separation issues,\ncorrective actions such as deletion and insertion are suggested respectively.\nAn extensive experimentation is conducted on the compressed images of the\nbenchmark datasets of ICDAR13 and Alireza et al [17] to demonstrate the\nefficacy.', 'In this work\\footnote {This work was supported in part by the National\nScience Foundation under grant IIS-1212948.}, we present a method to represent\na video with a sequence of words, and learn the temporal sequencing of such\nwords as the key information for predicting and recognizing human actions. We\nleverage core concepts from the Natural Language Processing (NLP) literature\nused in sentence classification to solve the problems of action prediction and\naction recognition. Each frame is converted into a word that is represented as\na vector using the Bag of Visual Words (BoW) encoding method. The words are\nthen combined into a sentence to represent the video, as a sentence. The\nsequence of words in different actions are learned with a simple but effective\nTemporal Convolutional Neural Network (T-CNN) that captures the temporal\nsequencing of information in a video sentence. We demonstrate that a key\ncharacteristic of the proposed method is its low-latency, i.e. its ability to\npredict an action accurately with a partial sequence (sentence). Experiments on\ntwo datasets, \\textit{UCF101} and \\textit{HMDB51} show that the method on\naverage reaches 95\\% of its accuracy within half the video frames. Results,\nalso demonstrate that our method achieves compatible state-of-the-art\nperformance in action recognition (i.e. at the completion of the sentence) in\naddition to action prediction.', 'We introduce FontCode, an information embedding technique for text documents.\nProvided a text document with specific fonts, our method embeds user-specified\ninformation in the text by perturbing the glyphs of text characters while\npreserving the text content. We devise an algorithm to chooses unobtrusive yet\nmachine-recognizable glyph perturbations, leveraging a recently developed\ngenerative model that alters the glyphs of each character continuously on a\nfont manifold. We then introduce an algorithm that embeds a user-provided\nmessage in the text document and produces an encoded document whose appearance\nis minimally perturbed from the original document. We also present a glyph\nrecognition method that recovers the embedded information from an encoded\ndocument stored as a vector graphic or pixel image, or even on a printed paper.\nIn addition, we introduce a new error-correction coding scheme that rectifies a\ncertain number of recognition errors. Lastly, we demonstrate that our technique\nenables a wide array of applications, using it as a text document metadata\nholder, an unobtrusive optical barcode, a cryptographic message embedding\nscheme, and a text document signature.', 'Imagery texts are usually organized as a hierarchy of several visual\nelements, i.e. characters, words, text lines and text blocks. Among these\nelements, character is the most basic one for various languages such as\nWestern, Chinese, Japanese, mathematical expression and etc. It is natural and\nconvenient to construct a common text detection engine based on character\ndetectors. However, training character detectors requires a vast of location\nannotated characters, which are expensive to obtain. Actually, the existing\nreal text datasets are mostly annotated in word or line level. To remedy this\ndilemma, we propose a weakly supervised framework that can utilize word\nannotations, either in tight quadrangles or the more loose bounding boxes, for\ncharacter detector training. When applied in scene text detection, we are thus\nable to train a robust character detector by exploiting word annotations in the\nrich large-scale real scene text datasets, e.g. ICDAR15 and COCO-text. The\ncharacter detector acts as a key role in the pipeline of our text detection\nengine. It achieves the state-of-the-art performance on several challenging\nscene text detection benchmarks. We also demonstrate the flexibility of our\npipeline by various scenarios, including deformed text detection and math\nexpression recognition.', 'We address the problem of predicting similarity between a pair of handwritten\ndocument images written by different individuals. This has applications related\nto matching and mining in image collections containing handwritten content. A\nsimilarity score is computed by detecting patterns of text re-usages between\ndocument images irrespective of the minor variations in word morphology, word\nordering, layout and paraphrasing of the content. Our method does not depend on\nan accurate segmentation of words and lines. We formulate the document matching\nproblem as a structured comparison of the word distributions across two\ndocument images. To match two word images, we propose a convolutional neural\nnetwork (CNN) based feature descriptor. Performance of this representation\nsurpasses the state-of-the-art on handwritten word spotting. Finally, we\ndemonstrate the applicability of our method on a practical problem of matching\nhandwritten assignments.', 'Recognition of document images have important applications in restoring old\nand classical texts. The problem involves quality improvement before passing it\nto a properly trained OCR to get accurate recognition of the text. The image\nenhancement and quality improvement constitute important steps as subsequent\nrecognition depends upon the quality of the input image. There are scenarios\nwhen high resolution images are not available and our experiments show that the\nOCR accuracy reduces significantly with decrease in the spatial resolution of\ndocument images. Thus the only option is to improve the resolution of such\ndocument images. The goal is to construct a high resolution image, given a\nsingle low resolution binary image, which constitutes the problem of single\nimage super-resolution. Most of the previous work in super-resolution deal with\nnatural images which have more information-content than the document images.\nHere, we use Convolution Neural Network to learn the mapping between low and\nthe corresponding high resolution images. We experiment with different number\nof layers, parameter settings and non-linear functions to build a fast\nend-to-end framework for document image super-resolution. Our proposed model\nshows a very good PSNR improvement of about 4 dB on 75 dpi Tamil images,\nresulting in a 3 % improvement of word level accuracy by the OCR. It takes less\ntime than the recent sparse based natural image super-resolution technique,\nmaking it useful for real-time document recognition applications.', 'Smartphones have enabled effortless capturing and sharing of documents in\ndigital form. The documents, however, often undergo various types of\ndegradation due to aging, stains, or shortcoming of capturing environment such\nas shadow, non-uniform lighting, etc., which reduces the comprehensibility of\nthe document images. In this work, we consider the problem of document image\ncleanup on embedded applications such as smartphone apps, which usually have\nmemory, energy, and latency limitations due to the device and/or for best human\nuser experience. We propose a light-weight encoder decoder based convolutional\nneural network architecture for removing the noisy elements from document\nimages. To compensate for generalization performance with a low network\ncapacity, we incorporate the perceptual loss for knowledge transfer from\npre-trained deep CNN network in our loss function. In terms of the number of\nparameters and product-sum operations, our models are 65-1030 and 3-27 times,\nrespectively, smaller than existing state-of-the-art document enhancement\nmodels. Overall, the proposed models offer a favorable resource versus accuracy\ntrade-off and we empirically illustrate the efficacy of our approach on several\nreal-world benchmark datasets.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently involved in a project that aims to improve the safety and efficiency of self-driving cars by predicting the paths of surrounding objects in urban driving situations. Current methods, such as multimodal regression and occupancy maps, have shown limitations in accurately predicting an object’s path. My plan is to treat trajectory prediction as a ranking task, sorting through the set of all potential paths. I need algorithms that can generate a diverse set of potential paths, accommodating a wide range of situations while concurrently discarding paths that are physically unfeasible. To enhance this method, I plan to generate these path sets based on the vehicle's current situation. For example, when the vehicle is on a highway, the model should not generate paths for objects on the roadside. I am also interested in integrating real-time traffic data and weather conditions into the model to increase its accuracy. Lastly, I aim to validate my novel approach using real-world autonomous vehicle data and compare its performance with existing trajectory prediction models.","['Autonomous explorative robots frequently encounter scenarios where multiple\nfuture trajectories can be pursued. Often these are cases with multiple paths\naround an obstacle or trajectory options towards various frontiers. Humans in\nsuch situations can inherently perceive and reason about the surrounding\nenvironment to identify several possibilities of either manoeuvring around the\nobstacles or moving towards various frontiers. In this work, we propose a 2\nstage Convolutional Neural Network architecture which mimics such an ability to\nmap the perceived surroundings to multiple trajectories that a robot can choose\nto traverse. The first stage is a Trajectory Proposal Network which suggests\ndiverse regions in the environment which can be occupied in the future. The\nsecond stage is a Trajectory Sampling network which provides a finegrained\ntrajectory over the regions proposed by Trajectory Proposal Network. We\nevaluate our framework in diverse and complicated real life settings. For the\noutdoor case, we use the KITTI dataset and our own outdoor driving dataset. In\nthe indoor setting, we use an autonomous drone to navigate various scenarios\nand also a ground robot which can explore the environment using the\ntrajectories proposed by our framework. Our experiments suggest that the\nframework is able to develop a semantic understanding of the obstacles, open\nregions and identify diverse trajectories that a robot can traverse. Our\ncomparisons portray the performance gain of the proposed architecture over a\ndiverse set of methods against which it is compared.', ""To safely and efficiently navigate in complex urban traffic, autonomous\nvehicles must make responsible predictions in relation to surrounding\ntraffic-agents (vehicles, bicycles, pedestrians, etc.). A challenging and\ncritical task is to explore the movement patterns of different traffic-agents\nand predict their future trajectories accurately to help the autonomous vehicle\nmake reasonable navigation decision. To solve this problem, we propose a long\nshort-term memory-based (LSTM-based) realtime traffic prediction algorithm,\nTrafficPredict. Our approach uses an instance layer to learn instances'\nmovements and interactions and has a category layer to learn the similarities\nof instances belonging to the same type to refine the prediction. In order to\nevaluate its performance, we collected trajectory datasets in a large city\nconsisting of varying conditions and traffic densities. The dataset includes\nmany challenging scenarios where vehicles, bicycles, and pedestrians move among\none another. We evaluate the performance of TrafficPredict on our new dataset\nand highlight its higher accuracy for trajectory prediction by comparing with\nprior prediction methods."", 'To navigate safely in urban environments, an autonomous vehicle (ego vehicle)\nmust understand and anticipate its surroundings, in particular the behavior and\nintents of other road users (neighbors). Most of the times, multiple decision\nchoices are acceptable for all road users (e.g., turn right or left, or\ndifferent ways of avoiding an obstacle), leading to a highly uncertain and\nmulti-modal decision space. We focus here on predicting multiple feasible\nfuture trajectories for both ego vehicle and neighbors through a probabilistic\nframework. We rely on a conditional imitation learning algorithm, conditioned\nby a navigation command for the ego vehicle (e.g., ""turn right""). Our model\nprocesses ego vehicle front-facing camera images and bird-eye view grid,\ncomputed from Lidar point clouds, with detections of past and present objects,\nin order to generate multiple trajectories for both ego vehicle and its\nneighbors. Our approach is computationally efficient and relies only on\non-board sensors. We evaluate our method offline on the publicly available\ndataset nuScenes, achieving state-of-the-art performance, investigate the\nimpact of our architecture choices on online simulated experiments and show\npreliminary insights for real vehicle control', 'Mimicking human ability to forecast future positions or interpret complex\ninteractions in urban scenarios, such as streets, shopping malls or squares, is\nessential to develop socially compliant robots or self-driving cars. Autonomous\nsystems may gain advantage on anticipating human motion to avoid collisions or\nto naturally behave alongside people. To foresee plausible trajectories, we\nconstruct an LSTM (long short-term memory)-based model considering three\nfundamental factors: people interactions, past observations in terms of\npreviously crossed areas and semantics of surrounding space. Our model\nencompasses several pooling mechanisms to join the above elements defining\nmultiple tensors, namely social, navigation and semantic tensors. The network\nis tested in unstructured environments where complex paths emerge according to\nboth internal (intentions) and external (other people, not accessible areas)\nmotivations. As demonstrated, modeling paths unaware of social interactions or\ncontext information, is insufficient to correctly predict future positions.\nExperimental results corroborate the effectiveness of the proposed framework in\ncomparison to LSTM-based models for human path prediction.', ""Predicting a vehicle's trajectory is an essential ability for autonomous\nvehicles navigating through complex urban traffic scenes. Bird's-eye-view\nroadmap information provides valuable information for making trajectory\npredictions, and while state-of-the-art models extract this information via\nimage convolution, auxiliary loss functions can augment patterns inferred from\ndeep learning by further encoding common knowledge of social and legal driving\nbehaviors. Since human driving behavior is inherently multimodal, models which\nallow for multimodal output tend to outperform single-prediction models on\nstandard metrics. We propose a loss function which enhances such models by\nenforcing expected driving rules on all predicted modes. Our contribution to\ntrajectory prediction is twofold; we propose a new metric which addresses\nfailure cases of the off-road rate metric by penalizing trajectories that\noppose the ascribed heading (flow direction) of a driving lane, and we show\nthis metric to be differentiable and therefore suitable as an auxiliary loss\nfunction. We then use this auxiliary loss to extend the the standard multiple\ntrajectory prediction (MTP) and MultiPath models, achieving improved results on\nthe nuScenes prediction benchmark by predicting trajectories which better\nconform to the lane-following rules of the road."", ""Annually, a large number of injuries and deaths around the world are related\nto motor vehicle accidents. This value has recently been reduced to some\nextent, via the use of driver-assistance systems. Developing driver-assistance\nsystems (i.e., automated driving systems) can play a crucial role in reducing\nthis number. Estimating and predicting surrounding vehicles' movement is\nessential for an automated vehicle and advanced safety systems. Moreover,\npredicting the trajectory is influenced by numerous factors, such as drivers'\nbehavior during accidents, history of the vehicle's movement and the\nsurrounding vehicles, and their position on the traffic scene. The vehicle must\nmove over a safe path in traffic and react to other drivers' unpredictable\nbehaviors in the shortest time. Herein, to predict automated vehicles' path, a\nmodel with low computational complexity is proposed, which is trained by images\ntaken from the road's aerial image. Our method is based on an encoder-decoder\nmodel that utilizes a social tensor to model the effect of the surrounding\nvehicles' movement on the target vehicle. The proposed model can predict the\nvehicle's future path in any freeway only by viewing the images related to the\nhistory of the target vehicle's movement and its neighbors. Deep learning was\nused as a tool for extracting the features of these images. Using the HighD\ndatabase, an image dataset of the road's aerial image was created, and the\nmodel's performance was evaluated on this new database. We achieved the RMSE of\n1.91 for the next 5 seconds and found that the proposed method had less error\nthan the best path-prediction methods in previous studies."", 'Predicting the future trajectory of surrounding vehicles in a randomly\nvarying traffic level is one of the most challenging problems in developing an\nautonomous vehicle. Since there is no pre-defined number of interacting\nvehicles participate in, the prediction network has to be scalable with respect\nto the vehicle number in order to guarantee the consistency in terms of both\naccuracy and computational load. In this paper, the first fully scalable\ntrajectory prediction network, SCALE-Net, is proposed that can ensure both\nhigher prediction performance and consistent computational load regardless of\nthe number of surrounding vehicles. The SCALE-Net employs the Edge-enhance\nGraph Convolutional Neural Network (EGCN) for the inter-vehicular interaction\nembedding network. Since the proposed EGCN is inherently scalable with respect\nto the graph node (an agent in this study), the model can be operated\nindependently from the total number of vehicles considered. We evaluated the\nscalability of the SCALE-Net on the publically available NGSIM datasets by\ncomparing variations on computation time and prediction accuracy per single\ndriving scene with respect to the varying vehicle number. The experimental test\nshows that both computation time and prediction performance of the SCALE-Net\nconsistently outperform those of previous models regardless of the level of\ntraffic complexities.', ""Urban environments manifest a high level of complexity, and therefore it is\nof vital importance for safety systems embedded within autonomous vehicles\n(AVs) to be able to accurately predict the short-term future motion of nearby\nagents. This problem can be further understood as generating a sequence of\nfuture coordinates for a given agent based on its past motion data e.g.\nposition, velocity, acceleration etc, and whilst current approaches demonstrate\nplausible results they have a propensity to neglect a scene's physical\nconstrains. In this paper we propose the model based on a combination of the\nCNN and LSTM encoder-decoder architecture that learns to extract a relevant\nroad features from semantic maps as well as general motion of agents and uses\nthis learned representation to predict their short-term future trajectories. We\ntrain and validate the model on the publicly available dataset that provides\ndata from urban areas, allowing us to examine it in challenging and uncertain\nscenarios. We show that our model is not only capable of anticipating future\nmotion whilst taking into consideration road boundaries, but can also\neffectively and precisely predict trajectories for a longer time horizon than\ninitially trained for."", ""Understanding the intentions of drivers at intersections is a critical\ncomponent for autonomous vehicles. Urban intersections that do not have traffic\nsignals are a common epicentre of highly variable vehicle movement and\ninteractions. We present a method for predicting driver intent at urban\nintersections through multi-modal trajectory prediction with uncertainty. Our\nmethod is based on recurrent neural networks combined with a mixture density\nnetwork output layer. To consolidate the multi-modal nature of the output\nprobability distribution, we introduce a clustering algorithm that extracts the\nset of possible paths that exist in the prediction output, and ranks them\naccording to likelihood. To verify the method's performance and\ngeneralizability, we present a real-world dataset that consists of over 23,000\nvehicles traversing five different intersections, collected using a vehicle\nmounted Lidar based tracking system. An array of metrics is used to demonstrate\nthe performance of the model against several baselines."", 'Autonomous vehicles should be able to predict the future states of its\nenvironment and respond appropriately. Specifically, predicting the behavior of\nsurrounding human drivers is vital for such platforms to share the same road\nwith humans. Behavior of each of the surrounding vehicles is governed by the\nmotion of its neighbor vehicles. This paper focuses on predicting the behavior\nof the surrounding vehicles of an autonomous vehicle on highways. We are\nmotivated by improving the prediction accuracy when a surrounding vehicle\nperforms lane change and highway merging maneuvers. We propose a novel pooling\nstrategy to capture the inter-dependencies between the neighbor vehicles.\nDepending solely on Euclidean trajectory representation, the existing pooling\nstrategies do not model the context information of the maneuvers intended by a\nsurrounding vehicle. In contrast, our pooling mechanism employs polar\ntrajectory representation, vehicles orientation and radial velocity. This\nresults in an implicitly maneuver-aware pooling operation. We incorporated the\nproposed pooling mechanism into a generative encoder-decoder model, and\nevaluated our method on the public NGSIM dataset. The results of maneuver-based\ntrajectory predictions demonstrate the effectiveness of the proposed method\ncompared with the state-of-the-art approaches. Our ""Pooling Toolbox"" code is\navailable at https://github.com/m-hasan-n/pooling.', 'The trajectory prediction is a critical and challenging problem in the design\nof an autonomous driving system. Many AI-oriented companies, such as Google\nWaymo, Uber and DiDi, are investigating more accurate vehicle trajectory\nprediction algorithms. However, the prediction performance is governed by lots\nof entangled factors, such as the stochastic behaviors of surrounding vehicles,\nhistorical information of self-trajectory, and relative positions of neighbors,\netc. In this paper, we propose a novel graph-based information sharing network\n(GISNet) that allows the information sharing between the target vehicle and its\nsurrounding vehicles. Meanwhile, the model encodes the historical trajectory\ninformation of all the vehicles in the scene. Experiments are carried out on\nthe public NGSIM US-101 and I-80 Dataset and the prediction performance is\nmeasured by the Root Mean Square Error (RMSE). The quantitative and qualitative\nexperimental results show that our model significantly improves the trajectory\nprediction accuracy, by up to 50.00%, compared to existing models.', 'Reliable prediction of surround vehicle motion is a critical requirement for\npath planning for autonomous vehicles. In this paper we propose a unified\nframework for surround vehicle maneuver classification and motion prediction\nthat exploits multiple cues, namely, the estimated motion of vehicles, an\nunderstanding of typical motion patterns of freeway traffic and inter-vehicle\ninteraction. We report our results in terms of maneuver classification accuracy\nand mean and median absolute error of predicted trajectories against the ground\ntruth for real traffic data collected using vehicle mounted sensors on\nfreeways. An ablative analysis is performed to analyze the relative importance\nof each cue for trajectory prediction. Additionally, an analysis of execution\ntime for the components of the framework is presented. Finally, we present\nmultiple case studies analyzing the outputs of our model for complex traffic\nscenarios', ""Predicting future trajectories of surrounding obstacles is a crucial task for\nautonomous driving cars to achieve a high degree of road safety. There are\nseveral challenges in trajectory prediction in real-world traffic scenarios,\nincluding obeying traffic rules, dealing with social interactions, handling\ntraffic of multi-class movement, and predicting multi-modal trajectories with\nprobability. Inspired by people's natural habit of navigating traffic with\nattention to their goals and surroundings, this paper presents a unique dynamic\ngraph attention network to solve all those challenges. The network is designed\nto model the dynamic social interactions among agents and conform to traffic\nrules with a semantic map. By extending the anchor-based method to multiple\ntypes of agents, the proposed method can predict multi-modal trajectories with\nprobabilities for multi-class movements using a single model. We validate our\napproach on the proprietary autonomous driving dataset for the logistic\ndelivery scenario and two publicly available datasets. The results show that\nour method outperforms state-of-the-art techniques and demonstrates the\npotential for trajectory prediction in real-world traffic."", 'The ability to predict multiple possible future positions of the ego-vehicle\ngiven the surrounding context while also estimating their probabilities is key\nto safe autonomous driving. Most of the current state-of-the-art Deep Learning\napproaches are trained on trajectory data to achieve this task. However\ntrajectory data captured by sensor systems is highly imbalanced, since by far\nmost of the trajectories follow straight lines with an approximately constant\nvelocity. This poses a huge challenge for the task of predicting future\npositions, which is inherently a regression problem. Current state-of-the-art\napproaches alleviate this problem only by major preprocessing of the training\ndata, e.g. resampling, clustering into anchors etc. In this paper we propose an\napproach which reformulates the prediction problem as a classification task,\nallowing for powerful tools, e.g. focal loss, to combat the imbalance. To this\nend we design a generative probabilistic model consisting of a deep neural\nnetwork with a Mixture of Gaussian head. A smart choice of the latent variable\nallows for the reformulation of the log-likelihood function as a combination of\na classification problem and a much simplified regression problem. The output\nof our model is an estimate of the probability density function of future\npositions, hence allowing for prediction of multiple possible positions while\nalso estimating their probabilities. The proposed approach can easily\nincorporate context information and does not require any preprocessing of the\ndata.', ""For autonomous agents to successfully operate in real world, the ability to\nanticipate future motions of surrounding entities in the scene can greatly\nenhance their safety levels since potentially dangerous situations could be\navoided in advance. While impressive results have been shown on predicting each\nagent's behavior independently, we argue that it is not valid to consider road\nentities individually since transitions of vehicle states are highly coupled.\nMoreover, as the predicted horizon becomes longer, modeling prediction\nuncertainties and multi-modal distributions over future sequences will turn\ninto a more challenging task. In this paper, we address this challenge by\npresenting a multi-modal probabilistic prediction approach. The proposed method\nis based on a generative model and is capable of jointly predicting sequential\nmotions of each pair of interacting agents. Most importantly, our model is\ninterpretable, which can explain the underneath logic as well as obtain more\nreliability to use in real applications. A complicate real-world roundabout\nscenario is utilized to implement and examine the proposed method."", 'Trajectory and intention prediction of traffic participants is an important\ntask in automated driving and crucial for safe interaction with the\nenvironment. In this paper, we present a new approach to vehicle trajectory\nprediction based on automatically generated maps containing statistical\ninformation about the behavior of traffic participants in a given area. These\nmaps are generated based on trajectory observations using image processing and\nmap matching techniques and contain all typical vehicle movements and\nprobabilities in the considered area. Our prediction approach matches an\nobserved trajectory to a behavior contained in the map and uses this\ninformation to generate a prediction. We evaluated our approach on a dataset\ncontaining over 14000 trajectories and found that it produces significantly\nmore precise mid-term predictions compared to motion model-based prediction\napproaches.', 'As autonomous vehicles (AVs) need to interact with other road users, it is of\nimportance to comprehensively understand the dynamic traffic environment,\nespecially the future possible trajectories of surrounding vehicles. This paper\npresents an algorithm for long-horizon trajectory prediction of surrounding\nvehicles using a dual long short term memory (LSTM) network, which is capable\nof effectively improving prediction accuracy in strongly interactive driving\nenvironments. In contrast to traditional approaches which require trajectory\nmatching and manual feature selection, this method can automatically learn\nhigh-level spatial-temporal features of driver behaviors from naturalistic\ndriving data through sequence learning. By employing two blocks of LSTMs, the\nproposed method feeds the sequential trajectory to the first LSTM for driver\nintention recognition as an intermediate indicator, which is immediately\nfollowed by a second LSTM for future trajectory prediction. Test results from\nreal-world highway driving data show that the proposed method can, in\ncomparison to state-of-art methods, output more accurate and reasonable\nestimate of different future trajectories over 5s time horizon with root mean\nsquare error (RMSE) for longitudinal and lateral prediction less than 5.77m and\n0.49m, respectively.', 'In a given scenario, simultaneously and accurately predicting every possible\ninteraction of traffic participants is an important capability for autonomous\nvehicles. The majority of current researches focused on the prediction of an\nsingle entity without incorporating the environment information. Although some\napproaches aimed to predict multiple vehicles, they either predicted each\nvehicle independently with no considerations on possible interaction with\nsurrounding entities or generated discretized joint motions which cannot be\ndirectly used in decision making and motion planning for autonomous vehicle. In\nthis paper, we present a probabilistic framework that is able to jointly\npredict continuous motions for multiple interacting road participants under any\ndriving scenarios and is capable of forecasting the duration of each\ninteraction, which can enhance the prediction performance and efficiency. The\nproposed traffic scene prediction framework contains two hierarchical modules:\nthe upper module and the lower module. The upper module forecasts the intention\nof the predicted vehicle, while the lower module predicts motions for\ninteracting scene entities. An exemplar real-world scenario is used to\nimplement and examine the proposed framework.', ""The environmental perception of an autonomous vehicle is limited by its\nphysical sensor ranges and algorithmic performance, as well as by occlusions\nthat degrade its understanding of an ongoing traffic situation. This not only\nposes a significant threat to safety and limits driving speeds, but it can also\nlead to inconvenient maneuvers. Intelligent Infrastructure Systems can help to\nalleviate these problems. An Intelligent Infrastructure System can fill in the\ngaps in a vehicle's perception and extend its field of view by providing\nadditional detailed information about its surroundings, in the form of a\ndigital model of the current traffic situation, i.e. a digital twin. However,\ndetailed descriptions of such systems and working prototypes demonstrating\ntheir feasibility are scarce. In this paper, we propose a hardware and software\narchitecture that enables such a reliable Intelligent Infrastructure System to\nbe built. We have implemented this system in the real world and demonstrate its\nability to create an accurate digital twin of an extended highway stretch, thus\nenhancing an autonomous vehicle's perception beyond the limits of its on-board\nsensors. Furthermore, we evaluate the accuracy and reliability of the digital\ntwin by using aerial images and earth observation methods for generating ground\ntruth data."", 'Accurate prediction of pedestrian and bicyclist paths is integral to the\ndevelopment of reliable autonomous vehicles in dense urban environments. The\ninteractions between vehicle and pedestrian or bicyclist have a significant\nimpact on the trajectories of traffic participants e.g. stopping or turning to\navoid collisions. Although recent datasets and trajectory prediction approaches\nhave fostered the development of autonomous vehicles yet the amount of\nvehicle-pedestrian (bicyclist) interactions modeled are sparse. In this work,\nwe propose Euro-PVI, a dataset of pedestrian and bicyclist trajectories. In\nparticular, our dataset caters more diverse and complex interactions in dense\nurban scenarios compared to the existing datasets. To address the challenges in\npredicting future trajectories with dense interactions, we develop a joint\ninference model that learns an expressive multi-modal shared latent space\nacross agents in the urban scene. This enables our Joint-$\\beta$-cVAE approach\nto better model the distribution of future trajectories. We achieve state of\nthe art results on the nuScenes and Euro-PVI datasets demonstrating the\nimportance of capturing interactions between ego-vehicle and pedestrians\n(bicyclists) for accurate predictions.', 'This paper addresses the problem of path prediction for multiple interacting\nagents in a scene, which is a crucial step for many autonomous platforms such\nas self-driving cars and social robots. We present \\textit{SoPhie}; an\ninterpretable framework based on Generative Adversarial Network (GAN), which\nleverages two sources of information, the path history of all the agents in a\nscene, and the scene context information, using images of the scene. To predict\na future path for an agent, both physical and social information must be\nleveraged. Previous work has not been successful to jointly model physical and\nsocial interactions. Our approach blends a social attention mechanism with a\nphysical attention that helps the model to learn where to look in a large scene\nand extract the most salient parts of the image relevant to the path. Whereas,\nthe social attention component aggregates information across the different\nagent interactions and extracts the most important trajectory information from\nthe surrounding neighbors. SoPhie also takes advantage of GAN to generates more\nrealistic samples and to capture the uncertain nature of the future paths by\nmodeling its distribution. All these mechanisms enable our approach to predict\nsocially and physically plausible paths for the agents and to achieve\nstate-of-the-art performance on several different trajectory forecasting\nbenchmarks.', 'The ability to predict the future movements of other vehicles is a\nsubconscious and effortless skill for humans and key to safe autonomous\ndriving. Therefore, trajectory prediction for autonomous cars has gained a lot\nof attention in recent years. It is, however, still a hard task to achieve\nhuman-level performance. Interdependencies between vehicle behaviors and the\nmultimodal nature of future intentions in a dynamic and complex driving\nenvironment render trajectory prediction a challenging problem. In this work,\nwe propose a new, data-driven approach for predicting the motion of vehicles in\na road environment. The model allows for inferring future intentions from the\npast interaction among vehicles in highway driving scenarios. Using our\nneighborhood-based data representation, the proposed system jointly exploits\ncorrelations in the spatial and temporal domain using convolutional neural\nnetworks. Our system considers multiple possible maneuver intentions and their\ncorresponding motion and predicts the trajectory for five seconds into the\nfuture. We implemented our approach and evaluated it on two highway datasets\ntaken in different countries and are able to achieve a competitive prediction\nperformance.', ""Highway driving invariably combines high speeds with the need to interact\nclosely with other drivers. Prediction methods enable autonomous vehicles (AVs)\nto anticipate drivers' future trajectories and plan accordingly. Kinematic\nmethods for prediction have traditionally ignored the presence of other\ndrivers, or made predictions only for a limited set of scenarios. Data-driven\napproaches fill this gap by learning from large datasets to predict\ntrajectories in general scenarios. While they achieve high accuracy, they also\nlose the interpretability and tools for model validation enjoyed by kinematic\nmethods. This letter proposes a novel kinematic model to describe car-following\nand lane change behavior, and extends it to predict trajectories in general\nscenarios. Experiments on highway datasets under varied sensing conditions\ndemonstrate that the proposed method outperforms state-of-the-art methods."", ""Accurately forecasting the future movements of surrounding vehicles is\nessential for safe and efficient operations of autonomous driving cars. This\ntask is difficult because a vehicle's moving trajectory is greatly determined\nby its driver's intention, which is often hard to estimate. By leveraging\nattention mechanisms along with long short-term memory (LSTM) networks, this\nwork learns the relation between a driver's intention and the vehicle's\nchanging positions relative to road infrastructures, and uses it to guide the\nprediction. Different from other state-of-the-art solutions, our work treats\nthe on-road lanes as non-Euclidean structures, unfolds the vehicle's moving\nhistory to form a spatio-temporal graph, and uses methods from Graph Neural\nNetworks to solve the problem. Not only is our approach a pioneering attempt in\nusing non-Euclidean methods to process static environmental features around a\npredicted object, our model also outperforms other state-of-the-art models in\nseveral metrics. The practicability and interpretability analysis of the model\nshows great potential for large-scale deployment in various autonomous driving\nsystems in addition to our own."", 'Path planning in dynamic environments is essential to high-risk applications\nsuch as unmanned aerial vehicles, self-driving cars, and autonomous underwater\nvehicles. In this paper, we generate collision-free trajectories for a robot\nwithin any given environment with temporal and spatial uncertainties caused due\nto randomly moving obstacles. We use two Poisson distributions to model the\nmovements of obstacles across the generated trajectory of a robot in both space\nand time to determine the probability of collision with an obstacle. Measures\nare taken to avoid an obstacle by intelligently manipulating the speed of the\nrobot at space-time intervals where a larger number of obstacles intersect the\ntrajectory of the robot. Our method potentially reduces the use of\ncomputationally expensive collision detection libraries. Based on our\nexperiments, there has been a significant improvement over existing methods in\nterms of safety, accuracy, execution time and computational cost. Our results\nshow a high level of accuracy between the predicted and actual number of\ncollisions with moving obstacles.', 'An efficient path planner for autonomous car-like vehicles should handle the\nstrong kinematic constraints, particularly in confined spaces commonly\nencountered while maneuvering in city traffic, and should enable rapid\nplanning, as the city traffic scenarios are highly dynamic. State-of-the-art\nplanning algorithms handle such difficult cases at high computational cost,\noften yielding non-deterministic results. However, feasible local paths can be\nquickly generated leveraging the past planning experience gained in the same or\nsimilar environment. While learning through supervised training is problematic\nfor real traffic scenarios, we introduce in this paper a novel neural\nnetwork-based method for path planning, which employs a gradient-based\nself-supervised learning algorithm to predict feasible paths. This approach\nstrongly exploits the experience gained in the past and rapidly yields feasible\nmaneuver plans for car-like vehicles with limited steering-angle. The\neffectiveness of such an approach has been confirmed by computational\nexperiments.', 'We address multi-modal trajectory forecasting of agents in unknown scenes by\nformulating it as a planning problem. We present an approach consisting of\nthree models; a goal prediction model to identify potential goals of the agent,\nan inverse reinforcement learning model to plan optimal paths to each goal, and\na trajectory generator to obtain future trajectories along the planned paths.\nAnalysis of predictions on the Stanford drone dataset, shows generalizability\nof our approach to novel scenes.', 'In this paper, we present Goal-GAN, an interpretable and end-to-end trainable\nmodel for human trajectory prediction. Inspired by human navigation, we model\nthe task of trajectory prediction as an intuitive two-stage process: (i) goal\nestimation, which predicts the most likely target positions of the agent,\nfollowed by a (ii) routing module which estimates a set of plausible\ntrajectories that route towards the estimated goal. We leverage information\nabout the past trajectory and visual context of the scene to estimate a\nmulti-modal probability distribution over the possible goal positions, which is\nused to sample a potential goal during the inference. The routing is governed\nby a recurrent neural network that reacts to physical constraints in the nearby\nsurroundings and generates feasible paths that route towards the sampled goal.\nOur extensive experimental evaluation shows that our method establishes a new\nstate-of-the-art on several benchmarks while being able to generate a realistic\nand diverse set of trajectories that conform to physical constraints.', 'It is critical to predict the motion of surrounding vehicles for self-driving\nplanning, especially in a socially compliant and flexible way. However, future\nprediction is challenging due to the interaction and uncertainty in driving\nbehaviors. We propose planning-informed trajectory prediction (PiP) to tackle\nthe prediction problem in the multi-agent setting. Our approach is\ndifferentiated from the traditional manner of prediction, which is only based\non historical information and decoupled with planning. By informing the\nprediction process with the planning of ego vehicle, our method achieves the\nstate-of-the-art performance of multi-agent forecasting on highway datasets.\nMoreover, our approach enables a novel pipeline which couples the prediction\nand planning, by conditioning PiP on multiple candidate trajectories of the ego\nvehicle, which is highly beneficial for autonomous driving in interactive\nscenarios.', 'Vision-based navigation of autonomous vehicles primarily depends on the Deep\nNeural Network (DNN) based systems in which the controller obtains input from\nsensors/detectors, such as cameras and produces a vehicle control output, such\nas a steering wheel angle to navigate the vehicle safely in a roadway traffic\nenvironment. Typically, these DNN-based systems of the autonomous vehicle are\ntrained through supervised learning; however, recent studies show that a\ntrained DNN-based system can be compromised by perturbation or adversarial\ninputs. Similarly, this perturbation can be introduced into the DNN-based\nsystems of autonomous vehicle by unexpected roadway hazards, such as debris and\nroadblocks. In this study, we first introduce a roadway hazardous environment\n(both intentional and unintentional roadway hazards) that can compromise the\nDNN-based navigational system of an autonomous vehicle, and produces an\nincorrect steering wheel angle, which can cause crashes resulting in fatality\nand injury. Then, we develop a DNN-based autonomous vehicle driving system\nusing object detection and semantic segmentation to mitigate the adverse effect\nof this type of hazardous environment, which helps the autonomous vehicle to\nnavigate safely around such hazards. We find that our developed DNN-based\nautonomous vehicle driving system including hazardous object detection and\nsemantic segmentation improves the navigational ability of an autonomous\nvehicle to avoid a potential hazard by 21% compared to the traditional\nDNN-based autonomous vehicle driving system.', ""When driving, people make decisions based on current traffic as well as their\ndesired route. They have a mental map of known routes and are often able to\nnavigate without needing directions. Current self-driving models improve their\nperformances when using additional GPS information. Here we aim to push forward\nself-driving research and perform route planning even in the absence of GPS.\nOur system learns to predict in real-time vehicle's current location and future\ntrajectory, as a function of time, on a known map, given only the raw video\nstream and the intended destination. The GPS signal is available only at\ntraining time, with training data annotation being fully automatic. Different\nfrom other published models, we predict the vehicle's trajectory for up to\nseven seconds ahead, from which complete steering, speed and acceleration\ninformation can be derived for the entire time span. Trajectories capture\nnavigational information on multiple levels, from instant steering commands\nthat depend on present traffic and obstacles ahead, to longer-term navigation\ndecisions, towards a specific destination. We collect our dataset with a\nregular car and a smartphone that records video and GPS streams. The GPS data\nis used to derive ground-truth supervision labels and create an analytical\nrepresentation of the traversed map. In tests, our system outperforms published\nmethods on visual localization and steering and gives accurate navigation\nassistance between any two known locations.""]","['Path prediction is a fundamental task for estimating how pedestrians or\nvehicles are going to move in a scene. Because path prediction as a task of\ncomputer vision uses video as input, various information used for prediction,\nsuch as the environment surrounding the target and the internal state of the\ntarget, need to be estimated from the video in addition to predicting paths.\nMany prediction approaches that include understanding the environment and the\ninternal state have been proposed. In this survey, we systematically summarize\nmethods of path prediction that take video as input and and extract features\nfrom the video. Moreover, we introduce datasets used to evaluate path\nprediction methods quantitatively.', 'We present a new method for multi-modal, long-term vehicle trajectory\nprediction. Our approach relies on using lane centerlines captured in rich maps\nof the environment to generate a set of proposed goal paths for each vehicle.\nUsing these paths -- which are generated at run time and therefore dynamically\nadapt to the scene -- as spatial anchors, we predict a set of goal-based\ntrajectories along with a categorical distribution over the goals. This\napproach allows us to directly model the goal-directed behavior of traffic\nactors, which unlocks the potential for more accurate long-term prediction. Our\nexperimental results on both a large-scale internal driving dataset and on the\npublic nuScenes dataset show that our model outperforms state-of-the-art\napproaches for vehicle trajectory prediction over a 6-second horizon. We also\nempirically demonstrate that our model is better able to generalize to road\nscenes from a completely new city than existing methods.', 'Urban autonomous driving decision making is challenging due to complex road\ngeometry and multi-agent interactions. Current decision making methods are\nmostly manually designing the driving policy, which might result in sub-optimal\nsolutions and is expensive to develop, generalize and maintain at scale. On the\nother hand, with reinforcement learning (RL), a policy can be learned and\nimproved automatically without any manual designs. However, current RL methods\ngenerally do not work well on complex urban scenarios. In this paper, we\npropose a framework to enable model-free deep reinforcement learning in\nchallenging urban autonomous driving scenarios. We design a specific input\nrepresentation and use visual encoding to capture the low-dimensional latent\nstates. Several state-of-the-art model-free deep RL algorithms are implemented\ninto our framework, with several tricks to improve their performance. We\nevaluate our method in a challenging roundabout task with dense surrounding\nvehicles in a high-definition driving simulator. The result shows that our\nmethod can solve the task well and is significantly better than the baseline.', 'Trajectory prediction is a fundamental and challenging task for numerous\napplications, such as autonomous driving and intelligent robots. Currently,\nmost of existing work treat the pedestrian trajectory as a series of fixed\ntwo-dimensional coordinates. However, in real scenarios, the trajectory often\nexhibits randomness, and has its own probability distribution. Inspired by this\nobserved fact, also considering other movement characteristics of pedestrians,\nwe propose one simple and intuitive movement description, probability\ntrajectory, which maps the coordinate points of pedestrian trajectory into\ntwo-dimensional Gaussian distribution in images. Based on this unique\ndescription, we develop one novel trajectory prediction method, called social\nprobability. The method combines the new probability trajectory and powerful\nconvolution recurrent neural networks together. Both the input and output of\nour method are probability trajectories, which provide the recurrent neural\nnetwork with sufficient spatial and random information of moving pedestrians.\nAnd the social probability extracts spatio-temporal features directly on the\nnew movement description to generate robust and accurate predicted results. The\nexperiments on public benchmark datasets show the effectiveness of the proposed\nmethod.', 'Progress towards advanced systems for assisted and autonomous driving is\nleveraging recent advances in recognition and segmentation methods. Yet, we are\nstill facing challenges in bringing reliable driving to inner cities, as those\nare composed of highly dynamic scenes observed from a moving platform at\nconsiderable speeds. Anticipation becomes a key element in order to react\ntimely and prevent accidents. In this paper we argue that it is necessary to\npredict at least 1 second and we thus propose a new model that jointly predicts\nego motion and people trajectories over such large time horizons. We pay\nparticular attention to modeling the uncertainty of our estimates arising from\nthe non-deterministic nature of natural traffic scenes. Our experimental\nresults show that it is indeed possible to predict people trajectories at the\ndesired time horizons and that our uncertainty estimates are informative of the\nprediction error. We also show that both sequence modeling of trajectories as\nwell as our novel method of long term odometry prediction are essential for\nbest performance.', ""In this paper, we propose an efficient vehicle trajectory prediction\nframework based on recurrent neural network. Basically, the characteristic of\nthe vehicle's trajectory is different from that of regular moving objects since\nit is affected by various latent factors including road structure, traffic\nrules, and driver's intention. Previous state of the art approaches use\nsophisticated vehicle behavior model describing these factors and derive the\ncomplex trajectory prediction algorithm, which requires a system designer to\nconduct intensive model optimization for practical use. Our approach is\ndata-driven and simple to use in that it learns complex behavior of the\nvehicles from the massive amount of trajectory data through deep neural network\nmodel. The proposed trajectory prediction method employs the recurrent neural\nnetwork called long short-term memory (LSTM) to analyze the temporal behavior\nand predict the future coordinate of the surrounding vehicles. The proposed\nscheme feeds the sequence of vehicles' coordinates obtained from sensor\nmeasurements to the LSTM and produces the probabilistic information on the\nfuture location of the vehicles over occupancy grid map. The experiments\nconducted using the data collected from highway driving show that the proposed\nmethod can produce reasonably good estimate of future trajectory."", 'We present a weakly-supervised approach to segmenting proposed drivable paths\nin images with the goal of autonomous driving in complex urban environments.\nUsing recorded routes from a data collection vehicle, our proposed method\ngenerates vast quantities of labelled images containing proposed paths and\nobstacles without requiring manual annotation, which we then use to train a\ndeep semantic segmentation network. With the trained network we can segment\nproposed paths and obstacles at run-time using a vehicle equipped with only a\nmonocular camera without relying on explicit modelling of road or lane\nmarkings. We evaluate our method on the large-scale KITTI and Oxford RobotCar\ndatasets and demonstrate reliable path proposal and obstacle segmentation in a\nwide variety of environments under a range of lighting, weather and traffic\nconditions. We illustrate how the method can generalise to multiple path\nproposals at intersections and outline plans to incorporate the system into a\nframework for autonomous urban driving.', 'Predicting future behavior of other traffic participants is an essential task\nthat needs to be solved by automated vehicles and human drivers alike to\nachieve safe and situationaware driving. Modern approaches to vehicles\ntrajectory prediction typically rely on data-driven models like neural\nnetworks, in particular LSTMs (Long Short-Term Memorys), achieving promising\nresults. However, the question of optimal composition of the underlying\ntraining data has received less attention. In this paper, we expand on previous\nwork on vehicle trajectory prediction based on neural network models employing\ndistributed representations to encode automotive scenes in a semantic vector\nsubstrate. We analyze the influence of variations in the training data on the\nperformance of our prediction models. Thereby, we show that the models\nemploying our semantic vector representation outperform the numerical model\nwhen trained on an adequate data set and thereby, that the composition of\ntraining data in vehicle trajectory prediction is crucial for successful\ntraining. We conduct our analysis on challenging real-world driving data.', 'Motion Planning, as a fundamental technology of automatic navigation for the\nautonomous vehicle, is still an open challenging issue in the real-life traffic\nsituation and is mostly applied by the model-based approaches. However, due to\nthe complexity of the traffic situations and the uncertainty of the edge cases,\nit is hard to devise a general motion planning system for the autonomous\nvehicle. In this paper, we proposed a motion planning model based on deep\nlearning (named as spatiotemporal LSTM network), which is able to generate a\nreal-time reflection based on spatiotemporal information extraction. To be\nspecific, the model based on spatiotemporal LSTM network has three main\nstructure. Firstly, the Convolutional Long-short Term Memory (Conv-LSTM) is\nused to extract hidden features through sequential image data. Then, the 3D\nConvolutional Neural Network(3D-CNN) is applied to extract the spatiotemporal\ninformation from the multi-frame feature information. Finally, the fully\nconnected neural networks are used to construct a control model for autonomous\nvehicle steering angle. The experiments demonstrated that the proposed method\ncan generate a robust and accurate visual motion planning results for the\nautonomous vehicle.', 'Understanding and predicting the intention of pedestrians is essential to\nenable autonomous vehicles and mobile robots to navigate crowds. This problem\nbecomes increasingly complex when we consider the uncertainty and multimodality\nof pedestrian motion, as well as the implicit interactions between members of a\ncrowd, including any response to a vehicle. Our approach, Probabilistic Crowd\nGAN, extends recent work in trajectory prediction, combining Recurrent Neural\nNetworks (RNNs) with Mixture Density Networks (MDNs) to output probabilistic\nmultimodal predictions, from which likely modal paths are found and used for\nadversarial training. We also propose the use of Graph Vehicle-Pedestrian\nAttention Network (GVAT), which models social interactions and allows input of\na shared vehicle feature, showing that inclusion of this module leads to\nimproved trajectory prediction both with and without the presence of a vehicle.\nThrough evaluation on various datasets, we demonstrate improvements on the\nexisting state of the art methods for trajectory prediction and illustrate how\nthe true multimodal and uncertain nature of crowd interactions can be directly\nmodelled.', 'Trajectory prediction is an important task, especially in autonomous driving.\nThe ability to forecast the position of other moving agents can yield to an\neffective planning, ensuring safety for the autonomous vehicle as well for the\nobserved entities. In this work we propose a data driven approach based on\nMarkov Chains to generate synthetic trajectories, which are useful for training\na multiple future trajectory predictor. The advantages are twofold: on the one\nhand synthetic samples can be used to augment existing datasets and train more\neffective predictors; on the other hand, it allows to generate samples with\nmultiple ground truths, corresponding to diverse equally likely outcomes of the\nobserved trajectory. We define a trajectory prediction model and a loss that\nexplicitly address the multimodality of the problem and we show that combining\nsynthetic and real data leads to prediction improvements, obtaining state of\nthe art results.', ""The vision of automated driving is to increase both road safety and\nefficiency, while offering passengers a convenient travel experience. This\nrequires that autonomous systems correctly estimate the current traffic scene\nand its likely evolution. In highway scenarios early recognition of cut-in\nmaneuvers is essential for risk-aware maneuver planning. In this paper, a\nstatistical approach is proposed, which advantageously utilizes a set of\nprototypical lane change trajectories to realize both early maneuver detection\nand uncertainty-aware trajectory prediction for traffic participants.\nGeneration of prototype trajectories from real traffic data is accomplished by\nAgglomerative Hierarchical Clustering. During clustering, the alignment of the\ncluster prototypes to each other is optimized and the cohesion of the resulting\nprototype is limited when two clusters merge. In the prediction stage, the\nsimilarity of observed vehicle motion and typical lane change patterns in the\ndata base is evaluated to construct a set of significant features for maneuver\nclassification via Boosted Decision Trees. The future trajectory is predicted\ncombining typical lane change realizations in a mixture model. B-splines based\ntrajectory adaptations guarantee continuity during transition from actually\nobserved to predicted vehicle states. Quantitative evaluation results\ndemonstrate the proposed concept's improved performance for both maneuver and\ntrajectory prediction compared to a previously implemented reference approach."", 'As autonomous driving systems mature, motion forecasting has received\nincreasing attention as a critical requirement for planning. Of particular\nimportance are interactive situations such as merges, unprotected turns, etc.,\nwhere predicting individual object motion is not sufficient. Joint predictions\nof multiple objects are required for effective route planning. There has been a\ncritical need for high-quality motion data that is rich in both interactions\nand annotation to develop motion planning models. In this work, we introduce\nthe most diverse interactive motion dataset to our knowledge, and provide\nspecific labels for interacting objects suitable for developing joint\nprediction models. With over 100,000 scenes, each 20 seconds long at 10 Hz, our\nnew dataset contains more than 570 hours of unique data over 1750 km of\nroadways. It was collected by mining for interesting interactions between\nvehicles, pedestrians, and cyclists across six cities within the United States.\nWe use a high-accuracy 3D auto-labeling system to generate high quality 3D\nbounding boxes for each road agent, and provide corresponding high definition\n3D maps for each scene. Furthermore, we introduce a new set of metrics that\nprovides a comprehensive evaluation of both single agent and joint agent\ninteraction motion forecasting models. Finally, we provide strong baseline\nmodels for individual-agent prediction and joint-prediction. We hope that this\nnew large-scale interactive motion dataset will provide new opportunities for\nadvancing motion forecasting models.', ""Autonomous driving consists of a multitude of interacting modules, where each\nmodule must contend with errors from the others. Typically, the motion\nprediction module depends upon a robust tracking system to capture each agent's\npast movement. In this work, we systematically explore the importance of the\ntracking module for the motion prediction task and ultimately conclude that the\noverall motion prediction performance is highly sensitive to the tracking\nmodule's imperfections. We explicitly compare models that use tracking\ninformation to models that do not across multiple scenarios and conditions. We\nfind that the tracking information plays an essential role and improves motion\nprediction performance in noise-free conditions. However, in the presence of\ntracking noise, it can potentially affect the overall performance if not\nstudied thoroughly. We thus argue practitioners should be mindful of noise when\ndeveloping and testing motion/tracking modules, or that they should consider\ntracking free alternatives."", 'Accurate pedestrian trajectory prediction is of great importance for\ndownstream tasks such as autonomous driving and mobile robot navigation. Fully\ninvestigating the social interactions within the crowd is crucial for accurate\npedestrian trajectory prediction. However, most existing methods do not capture\ngroup level interactions well, focusing only on pairwise interactions and\nneglecting group-wise interactions. In this work, we propose a hierarchical\ngraph convolutional network, HGCN-GJS, for trajectory prediction which well\nleverages group level interactions within the crowd. Furthermore, we introduce\na novel joint sampling scheme for modeling the joint distribution of multiple\npedestrians in the future trajectories. Based on the group information, this\nscheme associates the trajectory of one person with the trajectory of other\npeople in the group, but maintains the independence of the trajectories of\noutsiders. We demonstrate the performance of our network on several trajectory\nprediction datasets, achieving state-of-the-art results on all datasets\nconsidered.', 'This paper presents a method for future localization: to predict a set of\nplausible trajectories of ego-motion given a depth image. We predict paths\navoiding obstacles, between objects, even paths turning around a corner into\nspace behind objects. As a byproduct of the predicted trajectories of\nego-motion, we discover in the image the empty space occluded by foreground\nobjects. We use no image based features such as semantic labeling/segmentation\nor object detection/recognition for this algorithm. Inspired by proxemics, we\nrepresent the space around a person using an EgoSpace map, akin to an\nillustrated tourist map, that measures a likelihood of occlusion at the\negocentric coordinate system. A future trajectory of ego-motion is modeled by a\nlinear combination of compact trajectory bases allowing us to constrain the\npredicted trajectory. We learn the relationship between the EgoSpace map and\ntrajectory from the EgoMotion dataset providing in-situ measurements of the\nfuture trajectory. A cost function that takes into account partial occlusion\ndue to foreground objects is minimized to predict a trajectory. This cost\nfunction generates a trajectory that passes through the occluded space, which\nallows us to discover the empty space behind the foreground objects. We\nquantitatively evaluate our method to show predictive validity and apply to\nvarious real world scenes including walking, shopping, and social interactions.', 'To plan a safe and efficient route, an autonomous vehicle should anticipate\nfuture motions of other agents around it. Motion prediction is an extremely\nchallenging task which recently gained significant attention of the research\ncommunity. In this work, we present a simple and yet strong baseline for\nuncertainty aware motion prediction based purely on transformer neural\nnetworks, which has shown its effectiveness in conditions of domain change.\nWhile being easy-to-implement, the proposed approach achieves competitive\nperformance and ranks 1$^{st}$ on the 2021 Shifts Vehicle Motion Prediction\nCompetition.', 'Deep learning has enabled impressive progress in the accuracy of semantic\nsegmentation. Yet, the ability to estimate uncertainty and detect failure is\nkey for safety-critical applications like autonomous driving. Existing\nuncertainty estimates have mostly been evaluated on simple tasks, and it is\nunclear whether these methods generalize to more complex scenarios. We present\nFishyscapes, the first public benchmark for uncertainty estimation in a\nreal-world task of semantic segmentation for urban driving. It evaluates\npixel-wise uncertainty estimates towards the detection of anomalous objects in\nfront of the vehicle. We~adapt state-of-the-art methods to recent semantic\nsegmentation models and compare approaches based on softmax confidence,\nBayesian learning, and embedding density. Our results show that anomaly\ndetection is far from solved even for ordinary situations, while our benchmark\nallows measuring advancements beyond the state-of-the-art.', ""Pedestrian trajectory prediction is a critical yet challenging task,\nespecially for crowded scenes. We suggest that introducing an attention\nmechanism to infer the importance of different neighbors is critical for\naccurate trajectory prediction in scenes with varying crowd size. In this work,\nwe propose a novel method, AVGCN, for trajectory prediction utilizing graph\nconvolutional networks (GCN) based on human attention (A denotes attention, V\ndenotes visual field constraints). First, we train an attention network that\nestimates the importance of neighboring pedestrians, using gaze data collected\nas subjects perform a bird's eye view crowd navigation task. Then, we\nincorporate the learned attention weights modulated by constraints on the\npedestrian's visual field into a trajectory prediction network that uses a GCN\nto aggregate information from neighbors efficiently. AVGCN also considers the\nstochastic nature of pedestrian trajectories by taking advantage of variational\ntrajectory prediction. Our approach achieves state-of-the-art performance on\nseveral trajectory prediction benchmarks, and the lowest average prediction\nerror over all considered benchmarks."", 'End-to-end approaches to autonomous driving have high sample complexity and\nare difficult to scale to realistic urban driving. Simulation can help\nend-to-end driving systems by providing a cheap, safe, and diverse training\nenvironment. Yet training driving policies in simulation brings up the problem\nof transferring such policies to the real world. We present an approach to\ntransferring driving policies from simulation to reality via modularity and\nabstraction. Our approach is inspired by classic driving systems and aims to\ncombine the benefits of modular architectures and end-to-end deep learning\napproaches. The key idea is to encapsulate the driving policy such that it is\nnot directly exposed to raw perceptual input or low-level vehicle dynamics. We\nevaluate the presented approach in simulated urban environments and in the real\nworld. In particular, we transfer a driving policy trained in simulation to a\n1/5-scale robotic truck that is deployed in a variety of conditions, with no\nfinetuning, on two continents. The supplementary video can be viewed at\nhttps://youtu.be/BrMDJqI6H5U', 'Autonomous vehicles rely heavily upon their perception subsystems to see the\nenvironment in which they operate. Unfortunately, the effect of variable\nweather conditions presents a significant challenge to object detection\nalgorithms, and thus it is imperative to test the vehicle extensively in all\nconditions which it may experience. However, development of robust autonomous\nvehicle subsystems requires repeatable, controlled testing - while real weather\nis unpredictable and cannot be scheduled. Real-world testing in adverse\nconditions is an expensive and time-consuming task, often requiring access to\nspecialist facilities. Simulation is commonly relied upon as a substitute, with\nincreasingly visually realistic representations of the real-world being\ndeveloped. In the context of the complete autonomous vehicle control pipeline,\nsubsystems downstream of perception need to be tested with accurate recreations\nof the perception system output, rather than focusing on subjective visual\nrealism of the input - whether in simulation or the real world. This study\ndevelops the untapped potential of a lightweight weather augmentation method in\nan autonomous racing vehicle - focusing not on visual accuracy, but rather the\neffect upon perception subsystem performance in real time. With minimal\nadjustment, the prototype developed in this study can replicate the effects of\nwater droplets on the camera lens, and fading light conditions. This approach\nintroduces a latency of less than 8 ms using compute hardware well suited to\nbeing carried in the vehicle - rendering it ideal for real-time implementation\nthat can be run during experiments in simulation, and augmented reality testing\nin the real world.', 'Vision-based prediction algorithms have a wide range of applications\nincluding autonomous driving, surveillance, human-robot interaction, weather\nprediction. The objective of this paper is to provide an overview of the field\nin the past five years with a particular focus on deep learning approaches. For\nthis purpose, we categorize these algorithms into video prediction, action\nprediction, trajectory prediction, body motion prediction, and other prediction\napplications. For each category, we highlight the common architectures,\ntraining methods and types of data used. In addition, we discuss the common\nevaluation metrics and datasets used for vision-based prediction tasks. A\ndatabase of all the information presented in this survey including,\ncross-referenced according to papers, datasets and metrics, can be found online\nat https://github.com/aras62/vision-based-prediction.', 'Motion prediction of vehicles is critical but challenging due to the\nuncertainties in complex environments and the limited visibility caused by\nocclusions and limited sensor ranges. In this paper, we study a new task,\nsafety-aware motion prediction with unseen vehicles for autonomous driving.\nUnlike the existing trajectory prediction task for seen vehicles, we aim at\npredicting an occupancy map that indicates the earliest time when each location\ncan be occupied by either seen and unseen vehicles. The ability to predict\nunseen vehicles is critical for safety in autonomous driving. To tackle this\nchallenging task, we propose a safety-aware deep learning model with three new\nloss functions to predict the earliest occupancy map. Experiments on the\nlarge-scale autonomous driving nuScenes dataset show that our proposed model\nsignificantly outperforms the state-of-the-art baselines on the safety-aware\nmotion prediction task. To the best of our knowledge, our approach is the first\none that can predict the existence of unseen vehicles in most cases. Project\npage at {\\url{https://github.com/xrenaa/Safety-Aware-Motion-Prediction}}.', 'This paper is focused on the task of searching for a specific vehicle that\nappeared in the surveillance networks. Existing methods usually assume the\nvehicle images are well cropped from the surveillance videos, then use visual\nattributes, like colors and types, or license plate numbers to match the target\nvehicle in the image set. However, a complete vehicle search system should\nconsider the problems of vehicle detection, representation, indexing, storage,\nmatching, and so on. Besides, attribute-based search cannot accurately find the\nsame vehicle due to intra-instance changes in different cameras and the\nextremely uncertain environment. Moreover, the license plates may be\nmisrecognized in surveillance scenes due to the low resolution and noise. In\nthis paper, a Progressive Vehicle Search System, named as PVSS, is designed to\nsolve the above problems. PVSS is constituted of three modules: the crawler,\nthe indexer, and the searcher. The vehicle crawler aims to detect and track\nvehicles in surveillance videos and transfer the captured vehicle images,\nmetadata and contextual information to the server or cloud. Then multi-grained\nattributes, such as the visual features and license plate fingerprints, are\nextracted and indexed by the vehicle indexer. At last, a query triplet with an\ninput vehicle image, the time range, and the spatial scope is taken as the\ninput by the vehicle searcher. The target vehicle will be searched in the\ndatabase by a progressive process. Extensive experiments on the public dataset\nfrom a real surveillance network validate the effectiveness of the PVSS.', ""Deciphering human behaviors to predict their future paths/trajectories and\nwhat they would do from videos is important in many applications. Motivated by\nthis idea, this paper studies predicting a pedestrian's future path jointly\nwith future activities. We propose an end-to-end, multi-task learning system\nutilizing rich visual features about human behavioral information and\ninteraction with their surroundings. To facilitate the training, the network is\nlearned with an auxiliary task of predicting future location in which the\nactivity will happen. Experimental results demonstrate our state-of-the-art\nperformance over two public benchmarks on future trajectory prediction.\nMoreover, our method is able to produce meaningful future activity prediction\nin addition to the path. The result provides the first empirical evidence that\njoint modeling of paths and activities benefits future path prediction."", ""Most of researchers use the vehicle re-identification based on\nclassification. This always requires an update with the new vehicle models in\nthe market. In this paper, two types of vehicle re-identification will be\npresented. First, the standard method, which needs an image from the search\nvehicle. VRIC and VehicleID data set are suitable for training this module. It\nwill be explained in detail how to improve the performance of this method using\na trained network, which is designed for the classification. The second method\ntakes as input a representative image of the search vehicle with similar\nmake/model, released year and colour. It is very useful when an image from the\nsearch vehicle is not available. It produces as output a shape and a colour\nfeatures. This could be used by the matching across a database to re-identify\nvehicles, which look similar to the search vehicle. To get a robust module for\nthe re-identification, a fine-grained classification has been trained, which\nits class consists of four elements: the make of a vehicle refers to the\nvehicle's manufacturer, e.g. Mercedes-Benz, the model of a vehicle refers to\ntype of model within that manufacturer's portfolio, e.g. C Class, the year\nrefers to the iteration of the model, which may receive progressive alterations\nand upgrades by its manufacturer and the perspective of the vehicle. Thus, all\nfour elements describe the vehicle at increasing degree of specificity. The aim\nof the vehicle shape classification is to classify the combination of these\nfour elements. The colour classification has been separately trained. The\nresults of vehicle re-identification will be shown. Using a developed tool, the\nre-identification of vehicles on video images and on controlled data set will\nbe demonstrated. This work was partially funded under the grant."", 'Varying weather conditions, including rainfall and snowfall, are generally\nregarded as a challenge for computer vision algorithms. One proposed solution\nto the challenges induced by rain and snowfall is to artificially remove the\nrain from images or video using rain removal algorithms. It is the promise of\nthese algorithms that the rain-removed image frames will improve the\nperformance of subsequent segmentation and tracking algorithms. However, rain\nremoval algorithms are typically evaluated on their ability to remove synthetic\nrain on a small subset of images. Currently, their behavior is unknown on\nreal-world videos when integrated with a typical computer vision pipeline. In\nthis paper, we review the existing rain removal algorithms and propose a new\ndataset that consists of 22 traffic surveillance sequences under a broad\nvariety of weather conditions that all include either rain or snowfall. We\npropose a new evaluation protocol that evaluates the rain removal algorithms on\ntheir ability to improve the performance of subsequent segmentation, instance\nsegmentation, and feature tracking algorithms under rain and snow. If\nsuccessful, the de-rained frames of a rain removal algorithm should improve\nsegmentation performance and increase the number of accurately tracked\nfeatures. The results show that a recent single-frame-based rain removal\nalgorithm increases the segmentation performance by 19.7% on our proposed\ndataset, but it eventually decreases the feature tracking performance and\nshowed mixed results with recent instance segmentation methods. However, the\nbest video-based rain removal algorithm improves the feature tracking accuracy\nby 7.72%.', ""We introduce and show preliminary results of a fast randomized method that\nfinds a set of K paths lying in distinct homotopy classes. We frame the path\nplanning task as a graph search problem, where the navigation graph is based on\na Voronoi diagram. The search is biased by a cost function derived from the\nsocial force model that is used to generate and select the paths. We compare\nour method to Yen's algorithm, and empirically show that our approach is faster\nto find a subset of homotopy classes. Furthermore our approach computes a set\nof more diverse paths with respect to the baseline while obtaining a negligible\nloss in path quality."", 'Despite significant progress on current state-of-the-art image generation\nmodels, synthesis of document images containing multiple and complex object\nlayouts is a challenging task. This paper presents a novel approach, called\nDocSynth, to automatically synthesize document images based on a given layout.\nIn this work, given a spatial layout (bounding boxes with object categories) as\na reference by the user, our proposed DocSynth model learns to generate a set\nof realistic document images consistent with the defined layout. Also, this\nframework has been adapted to this work as a superior baseline model for\ncreating synthetic document image datasets for augmenting real data during\ntraining for document layout analysis tasks. Different sets of learning\nobjectives have been also used to improve the model performance.\nQuantitatively, we also compare the generated results of our model with real\ndata using standard evaluation metrics. The results highlight that our model\ncan successfully generate realistic and diverse document images with multiple\nobjects. We also present a comprehensive qualitative analysis summary of the\ndifferent scopes of synthetic image generation tasks. Lastly, to our knowledge\nthis is the first work of its kind.', 'Autonomous driving system development is critically dependent on the ability\nto replay complex and diverse traffic scenarios in simulation. In such\nscenarios, the ability to accurately simulate the vehicle sensors such as\ncameras, lidar or radar is essential. However, current sensor simulators\nleverage gaming engines such as Unreal or Unity, requiring manual creation of\nenvironments, objects and material properties. Such approaches have limited\nscalability and fail to produce realistic approximations of camera, lidar, and\nradar data without significant additional work.\n  In this paper, we present a simple yet effective approach to generate\nrealistic scenario sensor data, based only on a limited amount of lidar and\ncamera data collected by an autonomous vehicle. Our approach uses\ntexture-mapped surfels to efficiently reconstruct the scene from an initial\nvehicle pass or set of passes, preserving rich information about object 3D\ngeometry and appearance, as well as the scene conditions. We then leverage a\nSurfelGAN network to reconstruct realistic camera images for novel positions\nand orientations of the self-driving vehicle and moving objects in the scene.\nWe demonstrate our approach on the Waymo Open Dataset and show that it can\nsynthesize realistic camera data for simulated scenarios. We also create a\nnovel dataset that contains cases in which two self-driving vehicles observe\nthe same scene at the same time. We use this dataset to provide additional\nevaluation and demonstrate the usefulness of our SurfelGAN model.', ""Occlusion is a long-standing problem that causes many modern tracking methods\nto be erroneous. In this paper, we address the occlusion problem by exploiting\nthe current and future possible locations of the target object from its past\ntrajectory. To achieve this, we introduce a learning-based tracking method that\ntakes into account background motion modeling and trajectory prediction. Our\ntrajectory prediction module predicts the target object's locations in the\ncurrent and future frames based on the object's past trajectory. Since, in the\ninput video, the target object's trajectory is not only affected by the object\nmotion but also the camera motion, our background motion module estimates the\ncamera motion. So that the object's trajectory can be made independent from it.\nTo dynamically switch between the appearance-based tracker and the trajectory\nprediction, we employ a network that can assess how good a tracking prediction\nis, and we use the assessment scores to choose between the appearance-based\ntracker's prediction and the trajectory-based prediction. Comprehensive\nevaluations show that the proposed method sets a new state-of-the-art\nperformance on commonly used tracking benchmarks."", ""Driver attention prediction is currently becoming the focus in safe driving\nresearch community, such as the DR(eye)VE project and newly emerged Berkeley\nDeepDrive Attention (BDD-A) database in critical situations. In safe driving,\nan essential task is to predict the incoming accidents as early as possible.\nBDD-A was aware of this problem and collected the driver attention in\nlaboratory because of the rarity of such scenes. Nevertheless, BDD-A focuses\nthe critical situations which do not encounter actual accidents, and just faces\nthe driver attention prediction task, without a close step for accident\nprediction. In contrast to this, we explore the view of drivers' eyes for\ncapturing multiple kinds of accidents, and construct a more diverse and larger\nvideo benchmark than ever before with the driver attention and the driving\naccident annotation simultaneously (named as DADA-2000), which has 2000 video\nclips owning about 658,476 frames on 54 kinds of accidents. These clips are\ncrowd-sourced and captured in various occasions (highway, urban, rural, and\ntunnel), weather (sunny, rainy and snowy) and light conditions (daytime and\nnighttime). For the driver attention representation, we collect the maps of\nfixations, saccade scan path and focusing time. The accidents are annotated by\ntheir categories, the accident window in clips and spatial locations of the\ncrash-objects. Based on the analysis, we obtain a quantitative and positive\nanswer for the question in this paper."", 'This document is a document that has written procedures and methods for\ncollecting objects and unstructured dynamic data on the road for the\ndevelopment of object recognition technology for self-driving cars, and\noutlines the methods of collecting data, annotation data, object classifier\ncriteria, and data processing methods. On-road object and unstructured dynamic\ndata were collected in various environments, such as weather, time and traffic\nconditions, and additional reception calls for police and safety personnel were\ncollected. Finally, 100,000 images of various objects existing on pedestrians\nand roads, 200,000 images of police and traffic safety personnel, 5,000 images\nof police and traffic safety personnel, and data sets consisting of 5,000 image\ndata were collected and built.', 'In this work we propose a novel interpretation of residual networks showing\nthat they can be seen as a collection of many paths of differing length.\nMoreover, residual networks seem to enable very deep networks by leveraging\nonly the short paths during training. To support this observation, we rewrite\nresidual networks as an explicit collection of paths. Unlike traditional\nmodels, paths through residual networks vary in length. Further, a lesion study\nreveals that these paths show ensemble-like behavior in the sense that they do\nnot strongly depend on each other. Finally, and most surprising, most paths are\nshorter than one might expect, and only the short paths are needed during\ntraining, as longer paths do not contribute any gradient. For example, most of\nthe gradient in a residual network with 110 layers comes from paths that are\nonly 10-34 layers deep. Our results reveal one of the key characteristics that\nseem to enable the training of very deep networks: Residual networks avoid the\nvanishing gradient problem by introducing short paths which can carry gradient\nthroughout the extent of very deep networks.', 'The inference of minimum spanning arborescences within a set of objects is a\ngeneral problem which translates into numerous application-specific\nunsupervised learning tasks. We introduce a unified and generic structure\ncalled edit arborescence that relies on edit paths between data in a\ncollection, as well as the Min Edit Arborescence Problem, which asks for an\nedit arborescence that minimizes the sum of costs of its inner edit paths.\nThrough the use of suitable cost functions, this generic framework allows to\nmodel a variety of problems. In particular, we show that by introducing\nencoding size preserving edit costs, it can be used as an efficient method for\ncompressing collections of labeled graphs. Experiments on various graph\ndatasets, with comparisons to standard compression tools, show the potential of\nour method.', 'We introduce a novel type of text representation that preserves the 2D layout\nof a document. This is achieved by encoding each document page as a\ntwo-dimensional grid of characters. Based on this representation, we present a\ngeneric document understanding pipeline for structured documents. This pipeline\nmakes use of a fully convolutional encoder-decoder network that predicts a\nsegmentation mask and bounding boxes. We demonstrate its capabilities on an\ninformation extraction task from invoices and show that it significantly\noutperforms approaches based on sequential text or document images.', 'In recent years, attention models have been extensively used for person and\nvehicle re-identification. Most re-identification methods are designed to focus\nattention on key-point locations. However, depending on the orientation, the\ncontribution of each key-point varies. In this paper, we present a novel\ndual-path adaptive attention model for vehicle re-identification (AAVER). The\nglobal appearance path captures macroscopic vehicle features while the\norientation conditioned part appearance path learns to capture localized\ndiscriminative features by focusing attention on the most informative\nkey-points. Through extensive experimentation, we show that the proposed AAVER\nmethod is able to accurately re-identify vehicles in unconstrained scenarios,\nyielding state of the art results on the challenging dataset VeRi-776. As a\nbyproduct, the proposed system is also able to accurately predict vehicle\nkey-points and shows an improvement of more than 7% over state of the art. The\ncode for key-point estimation model is available at\nhttps://github.com/Pirazh/Vehicle_Key_Point_Orientation_Estimation.', 'The digital camera captured document images may often be warped and distorted\ndue to different camera angles or document surfaces. A robust technique is\nneeded to solve this kind of distortion. The research on dewarping of the\ndocument suffers due to the limited availability of benchmark public dataset.\nIn recent times, deep learning based approaches are used to solve the problems\naccurately. To train most of the deep neural networks a large number of\ndocument images is required and generating such a large volume of document\nimages manually is difficult. In this paper, we propose a technique to generate\na synthetic warped image from a flat-bedded scanned document image. It is done\nby calculating warping factors for each pixel position using two warping\nposition parameters (WPP) and eight warping control parameters (WCP). These\nparameters can be specified as needed depending upon the desired warping. The\nresults are compared with similar real captured images both qualitative and\nquantitative way.', 'Methods to quantify the complexity of trajectory datasets are still a missing\npiece in benchmarking human trajectory prediction models. In order to gain a\nbetter understanding of the complexity of trajectory prediction tasks and\nfollowing the intuition, that more complex datasets contain more information,\nan approach for quantifying the amount of information contained in a dataset\nfrom a prototype-based dataset representation is proposed. The dataset\nrepresentation is obtained by first employing a non-trivial spatial sequence\nalignment, which enables a subsequent learning vector quantization (LVQ) stage.\nA large-scale complexity analysis is conducted on several human trajectory\nprediction benchmarking datasets, followed by a brief discussion on indications\nfor human trajectory prediction and benchmarking.', 'The demand for on-device document recognition systems increases in\nconjunction with the emergence of more strict privacy and security\nrequirements. In such systems, there is no data transfer from the end device to\na third-party information processing servers. The response time is vital to the\nuser experience of on-device document recognition. Combined with the\nunavailability of discrete GPUs, powerful CPUs, or a large RAM capacity on\nconsumer-grade end devices such as smartphones, the time limitations put\nsignificant constraints on the computational complexity of the applied\nalgorithms for on-device execution.\n  In this work, we consider document location in an image without prior\nknowledge of the document content or its internal structure. In accordance with\nthe published works, at least 5 systems offer solutions for on-device document\nlocation. All these systems use a location method which can be considered\nHough-based. The precision of such systems seems to be lower than that of the\nstate-of-the-art solutions which were not designed to account for the limited\ncomputational resources.\n  We propose an advanced Hough-based method. In contrast with other approaches,\nit accounts for the geometric invariants of the central projection model and\ncombines both edge and color features for document boundary detection. The\nproposed method allowed for the second best result for SmartDoc dataset in\nterms of precision, surpassed by U-net like neural network. When evaluated on a\nmore challenging MIDV-500 dataset, the proposed algorithm guaranteed the best\nprecision compared to published methods. Our method retained the applicability\nto on-device computations.', 'This work addresses the problem of discovering, in an unsupervised manner,\ninterpretable paths in the latent space of pretrained GANs, so as to provide an\nintuitive and easy way of controlling the underlying generative factors. In\ndoing so, it addresses some of the limitations of the state-of-the-art works,\nnamely, a) that they discover directions that are independent of the latent\ncode, i.e., paths that are linear, and b) that their evaluation relies either\non visual inspection or on laborious human labeling. More specifically, we\npropose to learn non-linear warpings on the latent space, each one parametrized\nby a set of RBF-based latent space warping functions, and where each warping\ngives rise to a family of non-linear paths via the gradient of the function.\nBuilding on the work of Voynov and Babenko, that discovers linear paths, we\noptimize the trainable parameters of the set of RBFs, so as that images that\nare generated by codes along different paths, are easily distinguishable by a\ndiscriminator network. This leads to easily distinguishable image\ntransformations, such as pose and facial expressions in facial images. We show\nthat linear paths can be derived as a special case of our method, and show\nexperimentally that non-linear paths in the latent space lead to steeper, more\ndisentangled and interpretable changes in the image space than in state-of-the\nart methods, both qualitatively and quantitatively. We make the code and the\npretrained models publicly available at:\nhttps://github.com/chi0tzp/WarpedGANSpace.', 'Weather Recognition plays an important role in our daily lives and many\ncomputer vision applications. However, recognizing the weather conditions from\na single image remains challenging and has not been studied thoroughly.\nGenerally, most previous works treat weather recognition as a single-label\nclassification task, namely, determining whether an image belongs to a specific\nweather class or not. This treatment is not always appropriate, since more than\none weather conditions may appear simultaneously in a single image. To address\nthis problem, we make the first attempt to view weather recognition as a\nmulti-label classification task, i.e., assigning an image more than one labels\naccording to the displayed weather conditions. Specifically, a CNN-RNN based\nmulti-label classification approach is proposed in this paper. The\nconvolutional neural network (CNN) is extended with a channel-wise attention\nmodel to extract the most correlated visual features. The Recurrent Neural\nNetwork (RNN) further processes the features and excavates the dependencies\namong weather classes. Finally, the weather labels are predicted step by step.\nBesides, we construct two datasets for the weather recognition task and explore\nthe relationships among different weather conditions. Experimental results\ndemonstrate the superiority and effectiveness of the proposed approach. The new\nconstructed datasets will be available at\nhttps://github.com/wzgwzg/Multi-Label-Weather-Recognition.', 'In urban theory, urban form is related to social and economic status. This\npaper explores to uncover zip-code level income through urban form by analyzing\nfigure-ground map, a simple, prevailing and precise representation of urban\nform in the field of urban study. Deep learning in computer vision enables such\nrepresentation maps to be studied at a large scale. We propose to train a DCNN\nmodel to identify and uncover the internal bridge between social class and\nurban form. Further, using hand-crafted informative visual features related\nwith urban form properties (building size, building density, etc.), we apply a\nrandom forest classifier to interpret how morphological properties are related\nwith social class.', 'Training a supernet matters for one-shot neural architecture search (NAS)\nmethods since it serves as a basic performance estimator for different\narchitectures (paths). Current methods mainly hold the assumption that a\nsupernet should give a reasonable ranking over all paths. They thus treat all\npaths equally, and spare much effort to train paths. However, it is harsh for a\nsingle supernet to evaluate accurately on such a huge-scale search space (e.g.,\n$7^{21}$). In this paper, instead of covering all paths, we ease the burden of\nsupernet by encouraging it to focus more on evaluation of those\npotentially-good ones, which are identified using a surrogate portion of\nvalidation data. Concretely, during training, we propose a multi-path sampling\nstrategy with rejection, and greedily filter the weak paths. The training\nefficiency is thus boosted since the training space has been greedily shrunk\nfrom all paths to those potentially-good ones. Moreover, we further adopt an\nexploration and exploitation policy by introducing an empirical candidate path\npool. Our proposed method GreedyNAS is easy-to-follow, and experimental results\non ImageNet dataset indicate that it can achieve better Top-1 accuracy under\nsame search space and FLOPs or latency level, but with only $\\sim$60\\% of\nsupernet training cost. By searching on a larger space, our GreedyNAS can also\nobtain new state-of-the-art architectures.', 'Nowadays autonomous technologies are a very heavily explored area and\nparticularly computer vision as the main component of vehicle perception. The\nquality of the whole vision system based on neural networks relies on the\ndataset it was trained on. It is extremely difficult to find traffic sign\ndatasets from most of the counties of the world. Meaning autonomous vehicle\nfrom the USA will not be able to drive though Lithuania recognizing all road\nsigns on the way. In this paper, we propose a solution on how to update model\nusing a small dataset from the country vehicle will be used in. It is important\nto mention that is not panacea, rather small upgrade which can boost autonomous\ncar development in countries with limited data access. We achieved about 10\npercent quality raise and expect even better results during future experiments.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'Since the low quality of document images will greatly undermine the chances\nof success in automatic text recognition and analysis, it is necessary to\nassess the quality of document images uploaded in online business process, so\nas to reject those images of low quality. In this paper, we attempt to achieve\ndocument image quality assessment and our contributions are twofold. Firstly,\nsince document image quality assessment is more interested in text, we propose\na text line based framework to estimate document image quality, which is\ncomposed of three stages: text line detection, text line quality prediction,\nand overall quality assessment. Text line detection aims to find potential text\nlines with a detector. In the text line quality prediction stage, the quality\nscore is computed for each text line with a CNN-based prediction model. The\noverall quality of document images is finally assessed with the ensemble of all\ntext line quality. Secondly, to train the prediction model, a large-scale\ndataset, comprising 52,094 text line images, is synthesized with diverse\nattributes. For each text line image, a quality label is computed with a\npiece-wise function. To demonstrate the effectiveness of the proposed\nframework, comprehensive experiments are evaluated on two popular document\nimage quality assessment benchmarks. Our framework significantly outperforms\nthe state-of-the-art methods by large margins on the large and complicated\ndataset.', 'Document Layout Analysis is a fundamental step in Handwritten Text Processing\nsystems, from the extraction of the text lines to the type of zone it belongs\nto. We present a system based on artificial neural networks which is able to\ndetermine not only the baselines of text lines present in the document, but\nalso performs geometric and logic layout analysis of the document. Experiments\nin three different datasets demonstrate the potential of the method and show\ncompetitive results with respect to state-of-the-art methods.', 'Training a good supernet in one-shot NAS methods is difficult since the\nsearch space is usually considerably huge (e.g., $13^{21}$). In order to\nenhance the supernet\'s evaluation ability, one greedy strategy is to sample\ngood paths, and let the supernet lean towards the good ones and ease its\nevaluation burden as a result. However, in practice the search can be still\nquite inefficient since the identification of good paths is not accurate enough\nand sampled paths still scatter around the whole search space. In this paper,\nwe leverage an explicit path filter to capture the characteristics of paths and\ndirectly filter those weak ones, so that the search can be thus implemented on\nthe shrunk space more greedily and efficiently. Concretely, based on the fact\nthat good paths are much less than the weak ones in the space, we argue that\nthe label of ""weak paths"" will be more confident and reliable than that of\n""good paths"" in multi-path sampling. In this way, we thus cast the training of\npath filter in the positive and unlabeled (PU) learning paradigm, and also\nencourage a \\textit{path embedding} as better path/operation representation to\nenhance the identification capacity of the learned filter. By dint of this\nembedding, we can further shrink the search space by aggregating similar\noperations with similar embeddings, and the search can be more efficient and\naccurate. Extensive experiments validate the effectiveness of the proposed\nmethod GreedyNASv2. For example, our obtained GreedyNASv2-L achieves $81.1\\%$\nTop-1 accuracy on ImageNet dataset, significantly outperforming the ResNet-50\nstrong baselines.', 'Automatic document content processing is affected by artifacts caused by the\nshape of the paper, non-uniform and diverse color of lighting conditions.\nFully-supervised methods on real data are impossible due to the large amount of\ndata needed. Hence, the current state of the art deep learning models are\ntrained on fully or partially synthetic images. However, document shadow or\nshading removal results still suffer because: (a) prior methods rely on\nuniformity of local color statistics, which limit their application on\nreal-scenarios with complex document shapes and textures and; (b) synthetic or\nhybrid datasets with non-realistic, simulated lighting conditions are used to\ntrain the models. In this paper we tackle these problems with our two main\ncontributions. First, a physically constrained learning-based method that\ndirectly estimates document reflectance based on intrinsic image formation\nwhich generalizes to challenging illumination conditions. Second, a new dataset\nthat clearly improves previous synthetic ones, by adding a large range of\nrealistic shading and diverse multi-illuminant conditions, uniquely customized\nto deal with documents in-the-wild. The proposed architecture works in a\nself-supervised manner where only the synthetic texture is used as a weak\ntraining signal (obviating the need for very costly ground truth with\ndisentangled versions of shading and reflectance). The proposed approach leads\nto a significant generalization of document reflectance estimation in real\nscenes with challenging illumination. We extensively evaluate on the real\nbenchmark datasets available for intrinsic image decomposition and document\nshadow removal tasks. Our reflectance estimation scheme, when used as a\npre-processing step of an OCR pipeline, shows a 26% improvement of character\nerror rate (CER), thus, proving the practical applicability.', 'In this paper, a new task is proposed, namely, weather translation, which\nrefers to transferring weather conditions of the image from one category to\nanother. It is important for photographic style transfer. Although lots of\napproaches have been proposed in traditional image translation tasks, few of\nthem can handle the multi-category weather translation task, since weather\nconditions have rich categories and highly complex semantic structures. To\naddress this problem, we develop a multi-domain weather translation approach\nbased on generative adversarial networks (GAN), denoted as Weather GAN, which\ncan achieve the transferring of weather conditions among sunny, cloudy, foggy,\nrainy and snowy. Specifically, the weather conditions in the image are\ndetermined by various weather-cues, such as cloud, blue sky, wet ground, etc.\nTherefore, it is essential for weather translation to focus the main attention\non weather-cues. To this end, the generator of Weather GAN is composed of an\ninitial translation module, an attention module and a weather-cue segmentation\nmodule. The initial translation module performs global translation during\ngeneration procedure. The weather-cue segmentation module identifies the\nstructure and exact distribution of weather-cues. The attention module learns\nto focus on the interesting areas of the image while keeping other areas\nunaltered. The final generated result is synthesized by these three parts. This\napproach suppresses the distortion and deformation caused by weather\ntranslation. our approach outperforms the state-of-the-arts has been shown by a\nlarge number of experiments and evaluations.', 'This paper mainly focuses on environment perception in snowy situations which\nforms the backbone of the autonomous driving technology. For the purpose,\nsemantic segmentation is employed to classify the objects while the vehicle is\ndriven autonomously. We train the Fully Convolutional Networks (FCN) on our own\ndataset and present the experimental results. Finally, the outcomes are\nanalyzed to give a conclusion. It can be concluded that the database still\nneeds to be optimized and a favorable algorithm should be proposed to get\nbetter results.', 'Graphical elements: particularly tables and figures contain a visual summary\nof the most valuable information contained in a document. Therefore,\nlocalization of such graphical objects in the document images is the initial\nstep to understand the content of such graphical objects or document images. In\nthis paper, we present a novel end-to-end trainable deep learning based\nframework to localize graphical objects in the document images called as\nGraphical Object Detection (GOD). Our framework is data-driven and does not\nrequire any heuristics or meta-data to locate graphical objects in the document\nimages. The GOD explores the concept of transfer learning and domain adaptation\nto handle scarcity of labeled training images for graphical object detection\ntask in the document images. Performance analysis carried out on the various\npublic benchmark data sets: ICDAR-2013, ICDAR-POD2017,and UNLV shows that our\nmodel yields promising results as compared to state-of-the-art techniques.', 'This study developed a traffic sign detection and recognition algorithm based\non the RetinaNet. Two main aspects were revised to improve the detection of\ntraffic signs: image cropping to address the issue of large image and small\ntraffic signs; and using more anchors with various scales to detect traffic\nsigns with different sizes and shapes. The proposed algorithm was trained and\ntested in a series of autonomous driving front-view images in a virtual\nsimulation environment. Results show that the algorithm performed extremely\nwell under good illumination and weather conditions. Its drawbacks are that it\nsometimes failed to detect object under bad weather conditions like snow and\nfailed to distinguish speed limits signs with different limit values.', 'Documents often exhibit various forms of degradation, which make it hard to\nbe read and substantially deteriorate the performance of an OCR system. In this\npaper, we propose an effective end-to-end framework named Document Enhancement\nGenerative Adversarial Networks (DE-GAN) that uses the conditional GANs (cGANs)\nto restore severely degraded document images. To the best of our knowledge,\nthis practice has not been studied within the context of generative adversarial\ndeep networks. We demonstrate that, in different tasks (document clean up,\nbinarization, deblurring and watermark removal), DE-GAN can produce an enhanced\nversion of the degraded document with a high quality. In addition, our approach\nprovides consistent improvements compared to state-of-the-art methods over the\nwidely used DIBCO 2013, DIBCO 2017 and H-DIBCO 2018 datasets, proving its\nability to restore a degraded document image to its ideal condition. The\nobtained results on a wide variety of degradation reveal the flexibility of the\nproposed model to be exploited in other document enhancement problems.', 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'This paper presents results of Document Visual Question Answering Challenge\norganized as part of ""Text and Documents in the Deep Learning Era"" workshop, in\nCVPR 2020. The challenge introduces a new problem - Visual Question Answering\non document images. The challenge comprised two tasks. The first task concerns\nwith asking questions on a single document image. On the other hand, the second\ntask is set as a retrieval task where the question is posed over a collection\nof images. For the task 1 a new dataset is introduced comprising 50,000\nquestions-answer(s) pairs defined over 12,767 document images. For task 2\nanother dataset has been created comprising 20 questions over 14,362 document\nimages which share the same document template.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'Weather conditions often disrupt the proper functioning of transportation\nsystems. Present systems either deploy an array of sensors or use an in-vehicle\ncamera to predict weather conditions. These solutions have resulted in\nincremental cost and limited scope. To ensure smooth operation of all\ntransportation services in all-weather conditions, a reliable detection system\nis necessary to classify weather in wild. The challenges involved in solving\nthis problem is that weather conditions are diverse in nature and there is an\nabsence of discriminate features among various weather conditions. The existing\nworks to solve this problem have been scene specific and have targeted\nclassification of two categories of weather. In this paper, we have created a\nnew open source dataset consisting of images depicting three classes of weather\ni.e rain, snow and fog called RFS Dataset. A novel algorithm has also been\nproposed which has used super pixel delimiting masks as a form of data\naugmentation, leading to reasonable results with respect to ten Convolutional\nNeural Network architectures.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'Document segmentation is a method of rending the document into distinct\nregions. A document is an assortment of information and a standard mode of\nconveying information to others. Pursuance of data from documents involves ton\nof human effort, time intense and might severely prohibit the usage of data\nsystems. So, automatic information pursuance from the document has become a big\nissue. It is been shown that document segmentation will facilitate to beat such\nproblems. This paper proposes a new approach to segment and classify the\ndocument regions as text, image, drawings and table. Document image is divided\ninto blocks using Run length smearing rule and features are extracted from\nevery blocks. Discipulus tool has been used to construct the Genetic\nprogramming based classifier model and located 97.5% classification accuracy.', 'Computation of document image quality metrics often depends upon the\navailability of a ground truth image corresponding to the document. This limits\nthe applicability of quality metrics in applications such as hyperparameter\noptimization of image processing algorithms that operate on-the-fly on unseen\ndocuments. This work proposes the use of surrogate models to learn the behavior\nof a given document quality metric on existing datasets where ground truth\nimages are available. The trained surrogate model can later be used to predict\nthe metric value on previously unseen document images without requiring access\nto ground truth images. The surrogate model is empirically evaluated on the\nDocument Image Binarization Competition (DIBCO) and the Handwritten Document\nImage Binarization Competition (H-DIBCO) datasets.', 'Identity of a vehicle is done through the vehicle license plate by traffic\npolice in general. Au- tomatic vehicle license plate recognition has several\napplications in intelligent traffic management systems. The security situation\nacross the globe and particularly in India demands a need to equip the traffic\npolice with a system that enables them to get instant details of a vehicle. The\nsystem should be easy to use, should be mobile, and work 24 x 7. In this paper,\nwe describe a mobile phone based, client-server architected, license plate\nrecognition system. While we use the state of the art image processing and\npattern recognition algorithms tuned for Indian conditions to automatically\nrecognize non-uniform license plates, the main contribution is in creating an\nend to end usable solution. The client application runs on a mobile device and\na server application, with access to vehicle information database, is hosted\ncentrally. The solution enables capture of license plate image captured by the\nphone camera and passes to the server; on the server the license plate number\nis recognized; the data associated with the number plate is then sent back to\nthe mobile device, instantaneously. We describe the end to end system\narchitecture in detail. A working prototype of the proposed system has been\nimplemented in the lab environment.', 'Identity documents recognition is an important sub-field of document\nanalysis, which deals with tasks of robust document detection, type\nidentification, text fields recognition, as well as identity fraud prevention\nand document authenticity validation given photos, scans, or video frames of an\nidentity document capture. Significant amount of research has been published on\nthis topic in recent years, however a chief difficulty for such research is\nscarcity of datasets, due to the subject matter being protected by security\nrequirements. A few datasets of identity documents which are available lack\ndiversity of document types, capturing conditions, or variability of document\nfield values. In addition, the published datasets were typically designed only\nfor a subset of document recognition problems, not for a complex identity\ndocument analysis. In this paper, we present a dataset MIDV-2020 which consists\nof 1000 video clips, 2000 scanned images, and 1000 photos of 1000 unique mock\nidentity documents, each with unique text field values and unique artificially\ngenerated faces, with rich annotation. For the presented benchmark dataset\nbaselines are provided for such tasks as document location and identification,\ntext fields recognition, and face detection. With 72409 annotated images in\ntotal, to the date of publication the proposed dataset is the largest publicly\navailable identity documents dataset with variable artificially generated data,\nand we believe that it will prove invaluable for advancement of the field of\ndocument analysis and recognition. The dataset is available for download at\nftp://smartengines.com/midv-2020 and http://l3i-share.univ-lr.fr .', 'This paper describes the short-term competition on the Components\nSegmentation Task of Document Photos that was prepared in the context of the\n16th International Conference on Document Analysis and Recognition (ICDAR\n2021). This competition aims to bring together researchers working in the field\nof identification document image processing and provides them a suitable\nbenchmark to compare their techniques on the component segmentation task of\ndocument images. Three challenge tasks were proposed entailing different\nsegmentation assignments to be performed on a provided dataset. The collected\ndata are from several types of Brazilian ID documents, whose personal\ninformation was conveniently replaced. There were 16 participants whose results\nobtained for some or all the three tasks show different rates for the adopted\nmetrics, like Dice Similarity Coefficient ranging from 0.06 to 0.99. Different\nDeep Learning models were applied by the entrants with diverse strategies to\nachieve the best results in each of the tasks. Obtained results show that the\ncurrently applied methods for solving one of the proposed tasks (document\nboundary detection) are already well established. However, for the other two\nchallenge tasks (text zone and handwritten sign detection) research and\ndevelopment of more robust approaches are still required to achieve acceptable\nresults.', 'This report explores the latest advances in the field of digital document\nrecognition. With the focus on printed document imagery, we discuss the major\ndevelopments in optical character recognition (OCR) and document image\nenhancement/restoration in application to Latin and non-Latin scripts. In\naddition, we review and discuss the available technologies for hand-written\ndocument recognition. In this report, we also provide some company-accumulated\nbenchmark results on available OCR engines.', 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : In our security company, we often deal with low-resolution images of faces captured from surveillance cameras, which makes it difficult to identify individuals accurately. We are looking for a solution that can generate high-resolution face images from these low-resolution inputs, especially when the input resolution is extremely low (e.g., only several hundred pixels) and the images are captured under uncontrolled settings with large pose and illumination variations. We hope the solution is based on a deep learning framework that can efficiently utilize global and local constraints defining a face. It would be beneficial if the framework had the following 2-step processes: one for holistic face reconstruction according to global constraints and another for enhancing face-specific details and enforcing local patch statistics. Additionally, we would like the deep network to be optimized using a loss function for super-resolution that combines reconstruction error with a learned face quality measured in an adversarial setting, resulting in robust visual results. We are interested in a solution that has been extensively tested in both controlled and uncontrolled setups and has demonstrated improvements over the current state of the art.","['To learn disentangled representations of facial images, we present a Dual\nEncoder-Decoder based Generative Adversarial Network (DED-GAN). In the proposed\nmethod, both the generator and discriminator are designed with deep\nencoder-decoder architectures as their backbones. To be more specific, the\nencoder-decoder structured generator is used to learn a pose disentangled face\nrepresentation, and the encoder-decoder structured discriminator is tasked to\nperform real/fake classification, face reconstruction, determining identity and\nestimating face pose. We further improve the proposed network architecture by\nminimising the additional pixel-wise loss defined by the Wasserstein distance\nat the output of the discriminator so that the adversarial framework can be\nbetter trained. Additionally, we consider face pose variation to be continuous,\nrather than discrete in existing literature, to inject richer pose information\ninto our model. The pose estimation task is formulated as a regression problem,\nwhich helps to disentangle identity information from pose variations. The\nproposed network is evaluated on the tasks of pose-invariant face recognition\n(PIFR) and face synthesis across poses. An extensive quantitative and\nqualitative evaluation carried out on several controlled and in-the-wild\nbenchmarking datasets demonstrates the superiority of the proposed DED-GAN\nmethod over the state-of-the-art approaches.', 'Recent learning-based approaches, in which models are trained by single-view\nimages have shown promising results for monocular 3D face reconstruction, but\nthey suffer from the ill-posed face pose and depth ambiguity issue. In contrast\nto previous works that only enforce 2D feature constraints, we propose a\nself-supervised training architecture by leveraging the multi-view geometry\nconsistency, which provides reliable constraints on face pose and depth\nestimation. We first propose an occlusion-aware view synthesis method to apply\nmulti-view geometry consistency to self-supervised learning. Then we design\nthree novel loss functions for multi-view consistency, including the pixel\nconsistency loss, the depth consistency loss, and the facial landmark-based\nepipolar loss. Our method is accurate and robust, especially under large\nvariations of expressions, poses, and illumination conditions. Comprehensive\nexperiments on the face alignment and 3D face reconstruction benchmarks have\ndemonstrated superiority over state-of-the-art methods. Our code and model are\nreleased in https://github.com/jiaxiangshang/MGCNet.', 'Robust face reconstruction from monocular image in general lighting\nconditions is challenging. Methods combining deep neural network encoders with\ndifferentiable rendering have opened up the path for very fast monocular\nreconstruction of geometry, lighting and reflectance. They can also be trained\nin self-supervised manner for increased robustness and better generalization.\nHowever, their differentiable rasterization based image formation models, as\nwell as underlying scene parameterization, limit them to Lambertian face\nreflectance and to poor shape details. More recently, ray tracing was\nintroduced for monocular face reconstruction within a classic\noptimization-based framework and enables state-of-the art results. However\noptimization-based approaches are inherently slow and lack robustness. In this\npaper, we build our work on the aforementioned approaches and propose a new\nmethod that greatly improves reconstruction quality and robustness in general\nscenes. We achieve this by combining a CNN encoder with a differentiable ray\ntracer, which enables us to base the reconstruction on much more advanced\npersonalized diffuse and specular albedos, a more sophisticated illumination\nmodel and a plausible representation of self-shadows. This enables to take a\nbig leap forward in reconstruction quality of shape, appearance and lighting\neven in scenes with difficult illumination. With consistent face attributes\nreconstruction, our method leads to practical applications such as relighting\nand self-shadows removal. Compared to state-of-the-art methods, our results\nshow improved accuracy and validity of the approach.', 'There are many factors affecting visual face recognition, such as low\nresolution images, aging, illumination and pose variance, etc. One of the most\nimportant problem is low resolution face images which can result in bad\nperformance on face recognition. Most of the general face recognition\nalgorithms usually assume a sufficient resolution for the face images. However,\nin practice many applications often do not have sufficient image resolutions.\nThe modern face hallucination models demonstrate reasonable performance to\nreconstruct high-resolution images from its corresponding low resolution\nimages. However, they do not consider identity level information during\nhallucination which directly affects results of the recognition of low\nresolution faces. To address this issue, we propose a Face Hallucination\nGenerative Adversarial Network (FH-GAN) which improves the quality of low\nresolution face images and accurately recognize those low quality images.\nConcretely, we make the following contributions: 1) we propose FH-GAN network,\nan end-to-end system, that improves both face hallucination and face\nrecognition simultaneously. The novelty of this proposed network depends on\nincorporating identity information in a GAN-based face hallucination algorithm\nvia combining a face recognition network for identity preserving. 2) We also\npropose a new face hallucination network, namely Dense Sparse Network (DSNet),\nwhich improves upon the state-of-art in face hallucination. 3) We demonstrate\nbenefits of training the face recognition and GAN-based DSNet jointly by\nreporting good result on face hallucination and recognition.', 'Despite generative adversarial networks (GANs) can hallucinate\nphoto-realistic high-resolution (HR) faces from low-resolution (LR) faces, they\ncannot guarantee preserving the identities of hallucinated HR faces, making the\nHR faces poorly recognizable. To address this problem, we propose a Siamese GAN\n(SiGAN) to reconstruct HR faces that visually resemble their corresponding\nidentities. On top of a Siamese network, the proposed SiGAN consists of a pair\nof two identical generators and one discriminator. We incorporate\nreconstruction error and identity label information in the loss function of\nSiGAN in a pairwise manner. By iteratively optimizing the loss functions of the\ngenerator pair and discriminator of SiGAN, we cannot only achieve\nphoto-realistic face reconstruction, but also ensures the reconstructed\ninformation is useful for identity recognition. Experimental results\ndemonstrate that SiGAN significantly outperforms existing face hallucination\nGANs in objective face verification performance, while achieving\nphoto-realistic reconstruction. Moreover, for input LR faces from unknown\nidentities who are not included in training, SiGAN can still do a good job.', 'We present a deep learning approach for high resolution face completion with\nmultiple controllable attributes (e.g., male and smiling) under arbitrary\nmasks. Face completion entails understanding both structural meaningfulness and\nappearance consistency locally and globally to fill in ""holes"" whose content do\nnot appear elsewhere in an input image. It is a challenging task with the\ndifficulty level increasing significantly with respect to high resolution, the\ncomplexity of ""holes"" and the controllable attributes of filled-in fragments.\nOur system addresses the challenges by learning a fully end-to-end framework\nthat trains generative adversarial networks (GANs) progressively from low\nresolution to high resolution with conditional vectors encoding controllable\nattributes.\n  We design novel network architectures to exploit information across multiple\nscales effectively and efficiently. We introduce new loss functions encouraging\nsharp completion. We show that our system can complete faces with large\nstructural and appearance variations using a single feed-forward pass of\ncomputation with mean inference time of 0.007 seconds for images at 1024 x 1024\nresolution. We also perform a pilot human study that shows our approach\noutperforms state-of-the-art face completion methods in terms of rank analysis.\nThe code will be released upon publication.', 'State-of-the-art face super-resolution methods employ deep convolutional\nneural networks to learn a mapping between low- and high- resolution facial\npatterns by exploring local appearance knowledge. However, most of these\nmethods do not well exploit facial structures and identity information, and\nstruggle to deal with facial images that exhibit large pose variations. In this\npaper, we propose a novel face super-resolution method that explicitly\nincorporates 3D facial priors which grasp the sharp facial structures. Our work\nis the first to explore 3D morphable knowledge based on the fusion of\nparametric descriptions of face attributes (e.g., identity, facial expression,\ntexture, illumination, and face pose). Furthermore, the priors can easily be\nincorporated into any network and are extremely efficient in improving the\nperformance and accelerating the convergence speed. Firstly, a 3D face\nrendering branch is set up to obtain 3D priors of salient facial structures and\nidentity knowledge. Secondly, the Spatial Attention Module is used to better\nexploit this hierarchical information (i.e., intensity similarity, 3D facial\nstructure, and identity content) for the super-resolution problem. Extensive\nexperiments demonstrate that the proposed 3D priors achieve superior face\nsuper-resolution results over the state-of-the-arts.', 'Over the past few decades, numerous attempts have been made to address the\nproblem of recovering a high-resolution (HR) facial image from its\ncorresponding low-resolution (LR) counterpart, a task commonly referred to as\nface hallucination. Despite the impressive performance achieved by\nposition-patch and deep learning-based methods, most of these techniques are\nstill unable to recover identity-specific features of faces. The former group\nof algorithms often produces blurry and oversmoothed outputs particularly in\nthe presence of higher levels of degradation, whereas the latter generates\nfaces which sometimes by no means resemble the individuals in the input images.\nIn this paper, a novel face super-resolution approach will be introduced, in\nwhich the hallucinated face is forced to lie in a subspace spanned by the\navailable training faces. Therefore, in contrast to the majority of existing\nface hallucination techniques and thanks to this face subspace prior, the\nreconstruction is performed in favor of recovering person-specific facial\nfeatures, rather than merely increasing image quantitative scores. Furthermore,\ninspired by recent advances in the area of 3D face reconstruction, an efficient\n3D dictionary alignment scheme is also presented, through which the algorithm\nbecomes capable of dealing with low-resolution faces taken in uncontrolled\nconditions. In extensive experiments carried out on several well-known face\ndatasets, the proposed algorithm shows remarkable performance by generating\ndetailed and close to ground truth results which outperform the\nstate-of-the-art face hallucination algorithms by significant margins both in\nquantitative and qualitative evaluations.', 'Traditional methods for image-based 3D face reconstruction and facial motion\nretargeting fit a 3D morphable model (3DMM) to the face, which has limited\nmodeling capacity and fail to generalize well to in-the-wild data. Use of\ndeformation transfer or multilinear tensor as a personalized 3DMM for\nblendshape interpolation does not address the fact that facial expressions\nresult in different local and global skin deformations in different persons.\nMoreover, existing methods learn a single albedo per user which is not enough\nto capture the expression-specific skin reflectance variations. We propose an\nend-to-end framework that jointly learns a personalized face model per user and\nper-frame facial motion parameters from a large corpus of in-the-wild videos of\nuser expressions. Specifically, we learn user-specific expression blendshapes\nand dynamic (expression-specific) albedo maps by predicting personalized\ncorrections on top of a 3DMM prior. We introduce novel constraints to ensure\nthat the corrected blendshapes retain their semantic meanings and the\nreconstructed geometry is disentangled from the albedo. Experimental results\nshow that our personalization accurately captures fine-grained facial dynamics\nin a wide range of conditions and efficiently decouples the learned face model\nfrom facial motion, resulting in more accurate face reconstruction and facial\nmotion retargeting compared to state-of-the-art methods.', 'Face frontalization refers to the process of synthesizing the frontal view of\na face from a given profile. Due to self-occlusion and appearance distortion in\nthe wild, it is extremely challenging to recover faithful results and preserve\ntexture details in a high-resolution. This paper proposes a High Fidelity Pose\nInvariant Model (HF-PIM) to produce photographic and identity-preserving\nresults. HF-PIM frontalizes the profiles through a novel texture warping\nprocedure and leverages a dense correspondence field to bind the 2D and 3D\nsurface spaces. We decompose the prerequisite of warping into dense\ncorrespondence field estimation and facial texture map recovering, which are\nboth well addressed by deep networks. Different from those reconstruction\nmethods relying on 3D data, we also propose Adversarial Residual Dictionary\nLearning (ARDL) to supervise facial texture map recovering with only monocular\nimages. Exhaustive experiments on both controlled and uncontrolled environments\ndemonstrate that the proposed method not only boosts the performance of\npose-invariant face recognition but also dramatically improves high-resolution\nfrontalization appearances.', 'Monocular 3D facial shape reconstruction from a single 2D facial image has\nbeen an active research area due to its wide applications. Inspired by the\nsuccess of deep neural networks (DNN), we propose a DNN-based approach for\nEnd-to-End 3D FAce Reconstruction (UH-E2FAR) from a single 2D image. Different\nfrom recent works that reconstruct and refine the 3D face in an iterative\nmanner using both an RGB image and an initial 3D facial shape rendering, our\nDNN model is end-to-end, and thus the complicated 3D rendering process can be\navoided. Moreover, we integrate in the DNN architecture two components, namely\na multi-task loss function and a fusion convolutional neural network (CNN) to\nimprove facial expression reconstruction. With the multi-task loss function, 3D\nface reconstruction is divided into neutral 3D facial shape reconstruction and\nexpressive 3D facial shape reconstruction. The neutral 3D facial shape is\nclass-specific. Therefore, higher layer features are useful. In comparison, the\nexpressive 3D facial shape favors lower or intermediate layer features. With\nthe fusion-CNN, features from different intermediate layers are fused and\ntransformed for predicting the 3D expressive facial shape. Through extensive\nexperiments, we demonstrate the superiority of our end-to-end framework in\nimproving the accuracy of 3D face reconstruction.', 'Face Super-Resolution (SR) is a subfield of the SR domain that specifically\ntargets the reconstruction of face images. The main challenge of face SR is to\nrestore essential facial features without distortion. We propose a novel face\nSR method that generates photo-realistic 8x super-resolved face images with\nfully retained facial details. To that end, we adopt a progressive training\nmethod, which allows stable training by splitting the network into successive\nsteps, each producing output with a progressively higher resolution. We also\npropose a novel facial attention loss and apply it at each step to focus on\nrestoring facial attributes in greater details by multiplying the pixel\ndifference and heatmap values. Lastly, we propose a compressed version of the\nstate-of-the-art face alignment network (FAN) for landmark heatmap extraction.\nWith the proposed FAN, we can extract the heatmaps suitable for face SR and\nalso reduce the overall training time. Experimental results verify that our\nmethod outperforms state-of-the-art methods in both qualitative and\nquantitative measurements, especially in perceptual quality.', 'State-of-the-art deep neural network models have reached near perfect face\nrecognition accuracy rates on controlled high-resolution face images. However,\ntheir performance is drastically degraded when they are tested with very\nlow-resolution face images. This is particularly critical in surveillance\nsystems, where a low-resolution probe image is to be matched with\nhigh-resolution gallery images. super-resolution techniques aim at producing\nhigh-resolution face images from low-resolution counterparts. While they are\ncapable of reconstructing images that are visually appealing, the\nidentity-related information is not preserved. Here, we propose an\nidentity-preserving end-to-end image-to-image translation deep neural network\nwhich is capable of super-resolving very low-resolution faces to their\nhigh-resolution counterparts while preserving identity-related information. We\nachieved this by training a very deep convolutional encoder-decoder network\nwith a symmetric contracting path between corresponding layers. This network\nwas trained with a combination of a reconstruction and an identity-preserving\nloss, on multi-scale low-resolution conditions. Extensive quantitative\nevaluations of our proposed model demonstrated that it outperforms competing\nsuper-resolution and low-resolution face recognition methods on natural and\nartificial low-resolution face data sets and even unseen identities.', 'Face hallucination is a domain-specific super-resolution problem with the\ngoal to generate high-resolution (HR) faces from low-resolution (LR) input\nimages. In contrast to existing methods that often learn a single\npatch-to-patch mapping from LR to HR images and are regardless of the\ncontextual interdependency between patches, we propose a novel Attention-aware\nFace Hallucination (Attention-FH) framework which resorts to deep reinforcement\nlearning for sequentially discovering attended patches and then performing the\nfacial part enhancement by fully exploiting the global interdependency of the\nimage. Specifically, in each time step, the recurrent policy network is\nproposed to dynamically specify a new attended region by incorporating what\nhappened in the past. The state (i.e., face hallucination result for the whole\nimage) can thus be exploited and updated by the local enhancement network on\nthe selected region. The Attention-FH approach jointly learns the recurrent\npolicy network and local enhancement network through maximizing the long-term\nreward that reflects the hallucination performance over the whole image.\nTherefore, our proposed Attention-FH is capable of adaptively personalizing an\noptimal searching path for each face image according to its own characteristic.\nExtensive experiments show our approach significantly surpasses the\nstate-of-the-arts on in-the-wild faces with large pose and illumination\nvariations.', 'We present a self-supervised learning approach to learning monocular 3D face\nreconstruction with a pose guidance network (PGN). First, we unveil the\nbottleneck of pose estimation in prior parametric 3D face learning methods, and\npropose to utilize 3D face landmarks for estimating pose parameters. With our\nspecially designed PGN, our model can learn from both faces with fully labeled\n3D landmarks and unlimited unlabeled in-the-wild face images. Our network is\nfurther augmented with a self-supervised learning scheme, which exploits face\ngeometry information embedded in multiple frames of the same person, to\nalleviate the ill-posed nature of regressing 3D face geometry from a single\nimage. These three insights yield a single approach that combines the\ncomplementary strengths of parametric model learning and data-driven learning\ntechniques. We conduct a rigorous evaluation on the challenging AFLW2000-3D,\nFlorence and FaceWarehouse datasets, and show that our method outperforms the\nstate-of-the-art for all metrics.', 'Certain facial parts are salient (unique) in appearance, which substantially\ncontribute to the holistic recognition of a subject. Occlusion of these salient\nparts deteriorates the performance of face recognition algorithms. In this\npaper, we propose a generative model to reconstruct the missing parts of the\nface which are under occlusion. The proposed generative model (SD-GAN)\nreconstructs a face preserving the illumination variation and identity of the\nface. A novel adversarial training algorithm has been designed for a bimodal\nmutually exclusive Generative Adversarial Network (GAN) model, for faster\nconvergence. A novel adversarial ""structural"" loss function is also proposed,\ncomprising of two components: a holistic and a local loss, characterized by\nSSIM and patch-wise MSE. Ablation studies on real and synthetically occluded\nface datasets reveal that our proposed technique outperforms the competing\nmethods by a considerable margin, even for boosting the performance of Face\nRecognition.', 'Face recognition performance has improved remarkably in the last decade. Much\nof this success can be attributed to the development of deep learning\ntechniques such as convolutional neural networks (CNNs). While CNNs have pushed\nthe state-of-the-art forward, their training process requires a large amount of\nclean and correctly labelled training data. If a CNN is intended to tolerate\nfacial pose, then we face an important question: should this training data be\ndiverse in its pose distribution, or should face images be normalized to a\nsingle pose in a pre-processing step? To address this question, we evaluate a\nnumber of popular facial landmarking and pose correction algorithms to\nunderstand their effect on facial recognition performance. Additionally, we\nintroduce a new, automatic, single-image frontalization scheme that exceeds the\nperformance of current algorithms. CNNs trained using sets of different\npre-processing methods are used to extract features from the Point and Shoot\nChallenge (PaSC) and CMU Multi-PIE datasets. We assert that the subsequent\nverification and recognition performance serves to quantify the effectiveness\nof each pose correction scheme.', 'Surveillance scenarios are prone to several problems since they usually\ninvolve low-resolution footage, and there is no control of how far the subjects\nmay be from the camera in the first place. This situation is suitable for the\napplication of upsampling (super-resolution) algorithms since they may be able\nto recover the discriminant properties of the subjects involved. While general\nsuper-resolution approaches were proposed to enhance image quality for\nhuman-level perception, biometrics super-resolution methods seek the best\n""computer perception"" version of the image since their focus is on improving\nautomatic recognition performance. Convolutional neural networks and deep\nlearning algorithms, in general, have been applied to computer vision tasks and\nare now state-of-the-art for several sub-domains, including image\nclassification, restoration, and super-resolution. However, no work has\nevaluated the effects that the latest proposed super-resolution methods may\nhave upon the accuracy and face verification performance in low-resolution\n""in-the-wild"" data. This project aimed at evaluating and adapting different\ndeep neural network architectures for the task of face super-resolution driven\nby face recognition performance in real-world low-resolution images. The\nexperimental results in a real-world surveillance and attendance datasets\nshowed that general super-resolution architectures might enhance face\nverification performance of deep neural networks trained on high-resolution\nfaces. Also, since neural networks are function approximators and can be\ntrained based on specific objective functions, the use of a customized loss\nfunction optimized for feature extraction showed promising results for\nrecovering discriminant features in low-resolution face images.', 'The reconstruction of dense 3D models of face geometry and appearance from a\nsingle image is highly challenging and ill-posed. To constrain the problem,\nmany approaches rely on strong priors, such as parametric face models learned\nfrom limited 3D scan data. However, prior models restrict generalization of the\ntrue diversity in facial geometry, skin reflectance and illumination. To\nalleviate this problem, we present the first approach that jointly learns 1) a\nregressor for face shape, expression, reflectance and illumination on the basis\nof 2) a concurrently learned parametric face model. Our multi-level face model\ncombines the advantage of 3D Morphable Models for regularization with the\nout-of-space generalization of a learned corrective space. We train end-to-end\non in-the-wild images without dense annotations by fusing a convolutional\nencoder with a differentiable expert-designed renderer and a self-supervised\ntraining loss, both defined at multiple detail levels. Our approach compares\nfavorably to the state-of-the-art in terms of reconstruction quality, better\ngeneralizes to real world faces, and runs at over 250 Hz.', 'General image super-resolution techniques have difficulties in recovering\ndetailed face structures when applying to low resolution face images. Recent\ndeep learning based methods tailored for face images have achieved improved\nperformance by jointly trained with additional task such as face parsing and\nlandmark prediction. However, multi-task learning requires extra manually\nlabeled data. Besides, most of the existing works can only generate relatively\nlow resolution face images (e.g., $128\\times128$), and their applications are\ntherefore limited. In this paper, we introduce a novel SPatial Attention\nResidual Network (SPARNet) built on our newly proposed Face Attention Units\n(FAUs) for face super-resolution. Specifically, we introduce a spatial\nattention mechanism to the vanilla residual blocks. This enables the\nconvolutional layers to adaptively bootstrap features related to the key face\nstructures and pay less attention to those less feature-rich regions. This\nmakes the training more effective and efficient as the key face structures only\naccount for a very small portion of the face image. Visualization of the\nattention maps shows that our spatial attention network can capture the key\nface structures well even for very low resolution faces (e.g., $16\\times16$).\nQuantitative comparisons on various kinds of metrics (including PSNR, SSIM,\nidentity similarity, and landmark detection) demonstrate the superiority of our\nmethod over current state-of-the-arts. We further extend SPARNet with\nmulti-scale discriminators, named as SPARNetHD, to produce high resolution\nresults (i.e., $512\\times512$). We show that SPARNetHD trained with synthetic\ndata cannot only produce high quality and high resolution outputs for\nsynthetically degraded face images, but also show good generalization ability\nto real world low quality face images.', 'Face hallucination is a domain-specific super-resolution problem that aims to\ngenerate a high-resolution (HR) face image from a low-resolution~(LR) input. In\ncontrast to the existing patch-wise super-resolution models that divide a face\nimage into regular patches and independently apply LR to HR mapping to each\npatch, we implement deep reinforcement learning and develop a novel\nattention-aware face hallucination (Attention-FH) framework, which recurrently\nlearns to attend a sequence of patches and performs facial part enhancement by\nfully exploiting the global interdependency of the image. Specifically, our\nproposed framework incorporates two components: a recurrent policy network for\ndynamically specifying a new attended region at each time step based on the\nstatus of the super-resolved image and the past attended region sequence, and a\nlocal enhancement network for selected patch hallucination and global state\nupdating. The Attention-FH model jointly learns the recurrent policy network\nand local enhancement network through maximizing a long-term reward that\nreflects the hallucination result with respect to the whole HR image. Extensive\nexperiments demonstrate that our Attention-FH significantly outperforms the\nstate-of-the-art methods on in-the-wild face images with large pose and\nillumination variations.', 'In recent decades, 3D morphable model (3DMM) has been commonly used in\nimage-based photorealistic 3D face reconstruction. However, face images are\noften corrupted by serious occlusion by non-face objects including eyeglasses,\nmasks, and hands. Such objects block the correct capture of landmarks and\nshading information. Therefore, the reconstructed 3D face model is hardly\nreusable. In this paper, a novel method is proposed to restore de-occluded face\nimages based on inverse use of 3DMM and generative adversarial network. We\nutilize the 3DMM prior to the proposed adversarial network and combine a global\nand local adversarial convolutional neural network to learn face de-occlusion\nmodel. The 3DMM serves not only as geometric prior but also proposes the face\nregion for the local discriminator. Experiment results confirm the\neffectiveness and robustness of the proposed algorithm in removing challenging\ntypes of occlusions with various head poses and illumination. Furthermore, the\nproposed method reconstructs the correct 3D face model with de-occluded\ntextures.', 'This paper studies the problem of blind face restoration from an\nunconstrained blurry, noisy, low-resolution, or compressed image (i.e.,\ndegraded observation). For better recovery of fine facial details, we modify\nthe problem setting by taking both the degraded observation and a high-quality\nguided image of the same identity as input to our guided face restoration\nnetwork (GFRNet). However, the degraded observation and guided image generally\nare different in pose, illumination and expression, thereby making plain CNNs\n(e.g., U-Net) fail to recover fine and identity-aware facial details. To tackle\nthis issue, our GFRNet model includes both a warping subnetwork (WarpNet) and a\nreconstruction subnetwork (RecNet). The WarpNet is introduced to predict flow\nfield for warping the guided image to correct pose and expression (i.e., warped\nguidance), while the RecNet takes the degraded observation and warped guidance\nas input to produce the restoration result. Due to that the ground-truth flow\nfield is unavailable, landmark loss together with total variation\nregularization are incorporated to guide the learning of WarpNet. Furthermore,\nto make the model applicable to blind restoration, our GFRNet is trained on the\nsynthetic data with versatile settings on blur kernel, noise level,\ndownsampling scale factor, and JPEG quality factor. Experiments show that our\nGFRNet not only performs favorably against the state-of-the-art image and face\nrestoration methods, but also generates visually photo-realistic results on\nreal degraded facial images.', 'Reconstructing the detailed geometric structure of a face from a given image\nis a key to many computer vision and graphics applications, such as motion\ncapture and reenactment. The reconstruction task is challenging as human faces\nvary extensively when considering expressions, poses, textures, and intrinsic\ngeometries. While many approaches tackle this complexity by using additional\ndata to reconstruct the face of a single subject, extracting facial surface\nfrom a single image remains a difficult problem. As a result, single-image\nbased methods can usually provide only a rough estimate of the facial geometry.\nIn contrast, we propose to leverage the power of convolutional neural networks\nto produce a highly detailed face reconstruction from a single image. For this\npurpose, we introduce an end-to-end CNN framework which derives the shape in a\ncoarse-to-fine fashion. The proposed architecture is composed of two main\nblocks, a network that recovers the coarse facial geometry (CoarseNet),\nfollowed by a CNN that refines the facial features of that geometry (FineNet).\nThe proposed networks are connected by a novel layer which renders a depth\nimage given a mesh in 3D. Unlike object recognition and detection problems,\nthere are no suitable datasets for training CNNs to perform face geometry\nreconstruction. Therefore, our training regime begins with a supervised phase,\nbased on synthetic images, followed by an unsupervised phase that uses only\nunconstrained facial images. The accuracy and robustness of the proposed model\nis demonstrated by both qualitative and quantitative evaluation tests.']","['Recently, a lot of attention has been focused on the incorporation of 3D data\ninto face analysis and its applications. Despite providing a more accurate\nrepresentation of the face, 3D facial images are more complex to acquire than\n2D pictures. As a consequence, great effort has been invested in developing\nsystems that reconstruct 3D faces from an uncalibrated 2D image. However, the\n3D-from-2D face reconstruction problem is ill-posed, thus prior knowledge is\nneeded to restrict the solutions space. In this work, we review 3D face\nreconstruction methods proposed in the last decade, focusing on those that only\nuse 2D pictures captured under uncontrolled conditions. We present a\nclassification of the proposed methods based on the technique used to add prior\nknowledge, considering three main strategies, namely, statistical model\nfitting, photometry, and deep learning, and reviewing each of them separately.\nIn addition, given the relevance of statistical 3D facial models as prior\nknowledge, we explain the construction procedure and provide a list of the most\npopular publicly available 3D facial models. After the exhaustive study of\n3D-from-2D face reconstruction approaches, we observe that the deep learning\nstrategy is rapidly growing since the last few years, becoming the standard\nchoice in replacement of the widespread statistical model fitting. Unlike the\nother two strategies, photometry-based methods have decreased in number due to\nthe need for strong underlying assumptions that limit the quality of their\nreconstructions compared to statistical model fitting and deep learning\nmethods. The review also identifies current challenges and suggests avenues for\nfuture research.', 'Face super-resolution methods usually aim at producing visually appealing\nresults rather than preserving distinctive features for further face\nidentification. In this work, we propose a deep learning method for face\nverification on very low-resolution face images that involves\nidentity-preserving face super-resolution. Our framework includes a\nsuper-resolution network and a feature extraction network. We train a VGG-based\ndeep face recognition network (Parkhi et al. 2015) to be used as feature\nextractor. Our super-resolution network is trained to minimize the feature\ndistance between the high resolution ground truth image and the super-resolved\nimage, where features are extracted using our pre-trained feature extraction\nnetwork. We carry out experiments on FRGC, Multi-PIE, LFW-a, and MegaFace\ndatasets to evaluate our method in controlled and uncontrolled settings. The\nresults show that the presented method outperforms conventional\nsuper-resolution methods in low-resolution face verification.', 'Prior knowledge of face shape and structure plays an important role in face\ninpainting. However, traditional face inpainting methods mainly focus on the\ngenerated image resolution of the missing portion without consideration of the\nspecial particularities of the human face explicitly and generally produce\ndiscordant facial parts. To solve this problem, we present a domain embedded\nmulti-model generative adversarial model for inpainting of face images with\nlarge cropped regions. We firstly represent only face regions using the latent\nvariable as the domain knowledge and combine it with the non-face parts\ntextures to generate high-quality face images with plausible contents. Two\nadversarial discriminators are finally used to judge whether the generated\ndistribution is close to the real distribution or not. It can not only\nsynthesize novel image structures but also explicitly utilize the embedded face\ndomain knowledge to generate better predictions with consistency on structures\nand appearance. Experiments on both CelebA and CelebA-HQ face datasets\ndemonstrate that our proposed approach achieved state-of-the-art performance\nand generates higher quality inpainting results than existing ones.', 'Given a monocular face image as input, 3D face geometry reconstruction aims\nto recover a corresponding 3D face mesh. Recently, both optimization-based and\nlearning-based face reconstruction methods have taken advantage of the emerging\ndifferentiable renderer and shown promising results. However, the\ndifferentiable renderer, mainly based on graphics rules, simplifies the\nrealistic mechanism of the illumination, reflection, \\etc, of the real world,\nthus cannot produce realistic images. This brings a lot of domain-shift noise\nto the optimization or training process. In this work, we introduce a novel\nGenerative Adversarial Renderer (GAR) and propose to tailor its inverted\nversion to the general fitting pipeline, to tackle the above problem.\nSpecifically, the carefully designed neural renderer takes a face normal map\nand a latent code representing other factors as inputs and renders a realistic\nface image. Since the GAR learns to model the complicated real-world image,\ninstead of relying on the simplified graphics rules, it is capable of producing\nrealistic images, which essentially inhibits the domain-shift noise in training\nand optimization. Equipped with the elaborated GAR, we further proposed a novel\napproach to predict 3D face parameters, in which we first obtain fine initial\nparameters via Renderer Inverting and then refine it with gradient-based\noptimizers. Extensive experiments have been conducted to demonstrate the\neffectiveness of the proposed generative adversarial renderer and the novel\noptimization-based face reconstruction framework. Our method achieves\nstate-of-the-art performances on multiple face reconstruction datasets.', 'In this paper, we propose an effective face completion algorithm using a deep\ngenerative model. Different from well-studied background completion, the face\ncompletion task is more challenging as it often requires to generate\nsemantically new pixels for the missing key components (e.g., eyes and mouths)\nthat contain large appearance variations. Unlike existing nonparametric\nalgorithms that search for patches to synthesize, our algorithm directly\ngenerates contents for missing regions based on a neural network. The model is\ntrained with a combination of a reconstruction loss, two adversarial losses and\na semantic parsing loss, which ensures pixel faithfulness and local-global\ncontents consistency. With extensive experimental results, we demonstrate\nqualitatively and quantitatively that our model is able to deal with a large\narea of missing pixels in arbitrary shapes and generate realistic face\ncompletion results.', 'This paper addresses 2 challenging tasks: improving the quality of low\nresolution facial images and accurately locating the facial landmarks on such\npoor resolution images. To this end, we make the following 5 contributions: (a)\nwe propose Super-FAN: the very first end-to-end system that addresses both\ntasks simultaneously, i.e. both improves face resolution and detects the facial\nlandmarks. The novelty or Super-FAN lies in incorporating structural\ninformation in a GAN-based super-resolution algorithm via integrating a\nsub-network for face alignment through heatmap regression and optimizing a\nnovel heatmap loss. (b) We illustrate the benefit of training the two networks\njointly by reporting good results not only on frontal images (as in prior work)\nbut on the whole spectrum of facial poses, and not only on synthetic low\nresolution images (as in prior work) but also on real-world images. (c) We\nimprove upon the state-of-the-art in face super-resolution by proposing a new\nresidual-based architecture. (d) Quantitatively, we show large improvement over\nthe state-of-the-art for both face super-resolution and alignment. (e)\nQualitatively, we show for the first time good results on real-world low\nresolution images.', 'Several factors contribute to the appearance of an object in a visual scene,\nincluding pose, illumination, and deformation, among others. Each factor\naccounts for a source of variability in the data, while the multiplicative\ninteractions of these factors emulate the entangled variability, giving rise to\nthe rich structure of visual object appearance. Disentangling such unobserved\nfactors from visual data is a challenging task, especially when the data have\nbeen captured in uncontrolled recording conditions (also referred to as\n""in-the-wild"") and label information is not available.\n  In this paper, we propose the first unsupervised deep learning method (with\npseudo-supervision) for disentangling multiple latent factors of variation in\nface images captured in-the-wild. To this end, we propose a deep latent\nvariable model, where the multiplicative interactions of multiple latent\nfactors of variation are explicitly modelled by means of multilinear (tensor)\nstructure. We demonstrate that the proposed approach indeed learns disentangled\nrepresentations of facial expressions and pose, which can be used in various\napplications, including face editing, as well as 3D face reconstruction and\nclassification of facial expression, identity and pose.', 'Deep neural networks (DNNs) trained on large-scale datasets have recently\nachieved impressive improvements in face recognition. But a persistent\nchallenge remains to develop methods capable of handling large pose variations\nthat are relatively underrepresented in training data. This paper presents a\nmethod for learning a feature representation that is invariant to pose, without\nrequiring extensive pose coverage in training data. We first propose to\ngenerate non-frontal views from a single frontal face, in order to increase the\ndiversity of training data while preserving accurate facial details that are\ncritical for identity discrimination. Our next contribution is to seek a rich\nembedding that encodes identity features, as well as non-identity ones such as\npose and landmark locations. Finally, we propose a new feature reconstruction\nmetric learning to explicitly disentangle identity and pose, by demanding\nalignment between the feature reconstructions through various combinations of\nidentity and pose features, which is obtained from two images of the same\nsubject. Experiments on both controlled and in-the-wild face datasets, such as\nMultiPIE, 300WLP and the profile view database CFP, show that our method\nconsistently outperforms the state-of-the-art, especially on images with large\nhead pose variations. Detail results and resource are referred to\nhttps://sites.google.com/site/xipengcshomepage/iccv2017', 'In recent years, with the advent of deep-learning, face recognition has\nachieved exceptional success. However, many of these deep face recognition\nmodels perform much better in handling frontal faces compared to profile faces.\nThe major reason for poor performance in handling of profile faces is that it\nis inherently difficult to learn pose-invariant deep representations that are\nuseful for profile face recognition. In this paper, we hypothesize that the\nprofile face domain possesses a latent connection with the frontal face domain\nin a latent feature subspace. We look to exploit this latent connection by\nprojecting the profile faces and frontal faces into a common latent subspace\nand perform verification or retrieval in the latent domain. We leverage a\ncoupled conditional generative adversarial network (cpGAN) structure to find\nthe hidden relationship between the profile and frontal images in a latent\ncommon embedding subspace. Specifically, the cpGAN framework consists of two\nconditional GAN-based sub-networks, one dedicated to the frontal domain and the\nother dedicated to the profile domain. Each sub-network tends to find a\nprojection that maximizes the pair-wise correlation between the two feature\ndomains in a common embedding feature subspace. The efficacy of our approach\ncompared with the state-of-the-art is demonstrated using the CFP, CMU\nMulti-PIE, IJB-A, and IJB-C datasets. Additionally, we have also implemented a\ncoupled convolutional neural network (cpCNN) and an adversarial discriminative\ndomain adaptation network (ADDA) for profile to frontal face recognition. We\nhave evaluated the performance of cpCNN and ADDA and compared it with the\nproposed cpGAN. Finally, we have also evaluated our cpGAN for reconstruction of\nfrontal faces from input profile faces contained in the VGGFace2 dataset.', 'Face Super-Resolution (SR) is a domain-specific super-resolution problem. The\nspecific facial prior knowledge could be leveraged for better super-resolving\nface images. We present a novel deep end-to-end trainable Face Super-Resolution\nNetwork (FSRNet), which makes full use of the geometry prior, i.e., facial\nlandmark heatmaps and parsing maps, to super-resolve very low-resolution (LR)\nface images without well-aligned requirement. Specifically, we first construct\na coarse SR network to recover a coarse high-resolution (HR) image. Then, the\ncoarse HR image is sent to two branches: a fine SR encoder and a prior\ninformation estimation network, which extracts the image features, and\nestimates landmark heatmaps/parsing maps respectively. Both image features and\nprior information are sent to a fine SR decoder to recover the HR image. To\nfurther generate realistic faces, we propose the Face Super-Resolution\nGenerative Adversarial Network (FSRGAN) to incorporate the adversarial loss\ninto FSRNet. Moreover, we introduce two related tasks, face alignment and\nparsing, as the new evaluation metrics for face SR, which address the\ninconsistency of classic metrics w.r.t. visual perception. Extensive benchmark\nexperiments show that FSRNet and FSRGAN significantly outperforms state of the\narts for very LR face SR, both quantitatively and qualitatively. Code will be\nmade available upon publication.', 'Existing face super-resolution (SR) methods mainly assume the input image to\nbe noise-free. Their performance degrades drastically when applied to\nreal-world scenarios where the input image is always contaminated by noise. In\nthis paper, we propose a Facial Attribute Capsules Network (FACN) to deal with\nthe problem of high-scale super-resolution of noisy face image. Capsule is a\ngroup of neurons whose activity vector models different properties of the same\nentity. Inspired by the concept of capsule, we propose an integrated\nrepresentation model of facial information, which named Facial Attribute\nCapsule (FAC). In the SR processing, we first generated a group of FACs from\nthe input LR face, and then reconstructed the HR face from this group of FACs.\nAiming to effectively improve the robustness of FAC to noise, we generate FAC\nin semantic, probabilistic and facial attributes manners by means of integrated\nlearning strategy. Each FAC can be divided into two sub-capsules: Semantic\nCapsule (SC) and Probabilistic Capsule (PC). Them describe an explicit facial\nattribute in detail from two aspects of semantic representation and probability\ndistribution. The group of FACs model an image as a combination of facial\nattribute information in the semantic space and probabilistic space by an\nattribute-disentangling way. The diverse FACs could better combine the face\nprior information to generate the face images with fine-grained semantic\nattributes. Extensive benchmark experiments show that our method achieves\nsuperior hallucination results and outperforms state-of-the-art for very low\nresolution (LR) noise face image super resolution.', 'In this paper we address the problem of hallucinating high-resolution facial\nimages from unaligned low-resolution inputs at high magnification factors. We\napproach the problem with convolutional neural networks (CNNs) and propose a\nnovel (deep) face hallucination model that incorporates identity priors into\nthe learning procedure. The model consists of two main parts: i) a cascaded\nsuper-resolution network that upscales the low-resolution images, and ii) an\nensemble of face recognition models that act as identity priors for the\nsuper-resolution network during training. Different from competing\nsuper-resolution approaches that typically rely on a single model for upscaling\n(even with large magnification factors), our network uses a cascade of multiple\nSR models that progressively upscale the low-resolution images using steps of\n$2\\times$. This characteristic allows us to apply supervision signals (target\nappearances) at different resolutions and incorporate identity constraints at\nmultiple-scales. Our model is able to upscale (very) low-resolution images\ncaptured in unconstrained conditions and produce visually convincing results.\nWe rigorously evaluate the proposed model on a large datasets of facial images\nand report superior performance compared to the state-of-the-art.', 'Facial image super-resolution (SR) is an important preprocessing for facial\nimage analysis, face recognition, and image-based 3D face reconstruction.\nRecent convolutional neural network (CNN) based method has shown excellent\nperformance by learning mapping relation using pairs of low-resolution (LR) and\nhigh-resolution (HR) facial images. However, since the HR facial image\nreconstruction using CNN is conventionally aimed to increase the PSNR and SSIM\nmetrics, the reconstructed HR image might not be realistic even with high\nscores. An adversarial framework is proposed in this study to reconstruct the\nHR facial image by simultaneously generating an HR image with and without blur.\nFirst, the spatial resolution of the LR facial image is increased by eight\ntimes using a five-layer CNN. Then, the encoder extracts the features of the\nup-scaled image. These features are finally sent to two branches (decoders) to\ngenerate an HR facial image with and without blur. In addition, local and\nglobal discriminators are combined to focus on the reconstruction of HR facial\nstructures. Experiment results show that the proposed algorithm generates a\nrealistic HR facial image. Furthermore, the proposed method can generate a\nvariety of different facial images.', 'Enhancing low resolution images via super-resolution or image synthesis for\ncross-resolution face recognition has been well studied. Several image\nprocessing and machine learning paradigms have been explored for addressing the\nsame. In this research, we propose Synthesis via Deep Sparse Representation\nalgorithm for synthesizing a high resolution face image from a low resolution\ninput image. The proposed algorithm learns multi-level sparse representation\nfor both high and low resolution gallery images, along with an identity aware\ndictionary and a transformation function between the two representations for\nface identification scenarios. With low resolution test data as input, the high\nresolution test image is synthesized using the identity aware dictionary and\ntransformation which is then used for face recognition. The performance of the\nproposed SDSR algorithm is evaluated on four databases, including one real\nworld dataset. Experimental results and comparison with existing seven\nalgorithms demonstrate the efficacy of the proposed algorithm in terms of both\nface identification and image quality measures.', 'We introduce InverseFaceNet, a deep convolutional inverse rendering framework\nfor faces that jointly estimates facial pose, shape, expression, reflectance\nand illumination from a single input image. By estimating all parameters from\njust a single image, advanced editing possibilities on a single face image,\nsuch as appearance editing and relighting, become feasible in real time. Most\nprevious learning-based face reconstruction approaches do not jointly recover\nall dimensions, or are severely limited in terms of visual quality. In\ncontrast, we propose to recover high-quality facial pose, shape, expression,\nreflectance and illumination using a deep neural network that is trained using\na large, synthetically created training corpus. Our approach builds on a novel\nloss function that measures model-space similarity directly in parameter space\nand significantly improves reconstruction accuracy. We further propose a\nself-supervised bootstrapping process in the network training loop, which\niteratively updates the synthetic training corpus to better reflect the\ndistribution of real-world imagery. We demonstrate that this strategy\noutperforms completely synthetically trained networks. Finally, we show\nhigh-quality reconstructions and compare our approach to several\nstate-of-the-art approaches.', 'The performance of a convolutional neural network (CNN) based face\nrecognition model largely relies on the richness of labelled training data.\nCollecting a training set with large variations of a face identity under\ndifferent poses and illumination changes, however, is very expensive, making\nthe diversity of within-class face images a critical issue in practice. In this\npaper, we propose a 3D model-assisted domain-transferred face augmentation\nnetwork (DotFAN) that can generate a series of variants of an input face based\non the knowledge distilled from existing rich face datasets collected from\nother domains. DotFAN is structurally a conditional CycleGAN but has two\nadditional subnetworks, namely face expert network (FEM) and face shape\nregressor (FSR), for latent code control. While FSR aims to extract face\nattributes, FEM is designed to capture a face identity. With their aid, DotFAN\ncan learn a disentangled face representation and effectively generate face\nimages of various facial attributes while preserving the identity of augmented\nfaces. Experiments show that DotFAN is beneficial for augmenting small face\ndatasets to improve their within-class diversity so that a better face\nrecognition model can be learned from the augmented dataset.', 'With the powerfulness of convolution neural networks (CNN), CNN based face\nreconstruction has recently shown promising performance in reconstructing\ndetailed face shape from 2D face images. The success of CNN-based methods\nrelies on a large number of labeled data. The state-of-the-art synthesizes such\ndata using a coarse morphable face model, which however has difficulty to\ngenerate detailed photo-realistic images of faces (with wrinkles). This paper\npresents a novel face data generation method. Specifically, we render a large\nnumber of photo-realistic face images with different attributes based on\ninverse rendering. Furthermore, we construct a fine-detailed face image dataset\nby transferring different scales of details from one image to another. We also\nconstruct a large number of video-type adjacent frame pairs by simulating the\ndistribution of real video data. With these nicely constructed datasets, we\npropose a coarse-to-fine learning framework consisting of three convolutional\nnetworks. The networks are trained for real-time detailed 3D face\nreconstruction from monocular video as well as from a single image. Extensive\nexperimental results demonstrate that our framework can produce high-quality\nreconstruction but with much less computation time compared to the\nstate-of-the-art. Moreover, our method is robust to pose, expression and\nlighting due to the diversity of data.', 'Given a really low-resolution input image of a face (say 16x16 or 8x8\npixels), the goal of this paper is to reconstruct a high-resolution version\nthereof. This, by itself, is an ill-posed problem, as the high-frequency\ninformation is missing in the low-resolution input and needs to be\nhallucinated, based on prior knowledge about the image content. Rather than\nrelying on a generic face prior, in this paper, we explore the use of a set of\nexemplars, i.e. other high-resolution images of the same person. These guide\nthe neural network as we condition the output on them. Multiple exemplars work\nbetter than a single one. To combine the information from multiple exemplars\neffectively, we introduce a pixel-wise weight generation module. Besides\nstandard face super-resolution, our method allows to perform subtle face\nediting simply by replacing the exemplars with another set with different\nfacial features. A user study is conducted and shows the super-resolved images\ncan hardly be distinguished from real images on the CelebA dataset. A\nqualitative comparison indicates our model outperforms methods proposed in the\nliterature on the CelebA and WebFace dataset.', 'It has been recently shown that neural networks can recover the geometric\nstructure of a face from a single given image. A common denominator of most\nexisting face geometry reconstruction methods is the restriction of the\nsolution space to some low-dimensional subspace. While such a model\nsignificantly simplifies the reconstruction problem, it is inherently limited\nin its expressiveness. As an alternative, we propose an Image-to-Image\ntranslation network that jointly maps the input image to a depth image and a\nfacial correspondence map. This explicit pixel-based mapping can then be\nutilized to provide high quality reconstructions of diverse faces under extreme\nexpressions, using a purely geometric refinement process. In the spirit of\nrecent approaches, the network is trained only with synthetic data, and is then\nevaluated on in-the-wild facial images. Both qualitative and quantitative\nanalyses demonstrate the accuracy and the robustness of our approach.', 'We present a novel deep learning approach to synthesize complete face images\nin the presence of large ocular region occlusions. This is motivated by recent\nsurge of VR/AR displays that hinder face-to-face communications. Different from\nthe state-of-the-art face inpainting methods that have no control over the\nsynthesized content and can only handle frontal face pose, our approach can\nfaithfully recover the missing content under various head poses while\npreserving the identity. At the core of our method is a novel generative\nnetwork with dedicated constraints to regularize the synthesis process. To\npreserve the identity, our network takes an arbitrary occlusion-free image of\nthe target identity to infer the missing content, and its high-level CNN\nfeatures as an identity prior to regularize the searching space of generator.\nSince the input reference image may have a different pose, a pose map and a\nnovel pose discriminator are further adopted to supervise the learning of\nimplicit pose transformations. Our method is capable of generating coherent\nfacial inpainting with consistent identity over videos with large variations of\nhead motions. Experiments on both synthesized and real data demonstrate that\nour method greatly outperforms the state-of-the-art methods in terms of both\nsynthesis quality and robustness.', 'Face deblurring aims to restore a clear face image from a blurred input image\nwith more explicit structure and facial details. However, most conventional\nimage and face deblurring methods focus on the whole generated image resolution\nwithout consideration of special face part texture and generally produce\nunsufficient details. Considering that faces and backgrounds have different\ndistribution information, in this study, we designed an effective face\ndeblurring network based on separable normalization and adaptive\ndenormalization (SNADNet). First, We fine-tuned the face parsing network to\nobtain an accurate face structure. Then, we divided the face parsing feature\ninto face foreground and background. Moreover, we constructed a new feature\nadaptive denormalization to regularize fafcial structures as a condition of the\nauxiliary to generate more harmonious and undistorted face structure. In\naddition, we proposed a texture extractor and multi-patch discriminator to\nenhance the generated facial texture information. Experimental results on both\nCelebA and CelebA-HQ datasets demonstrate that the proposed face deblurring\nnetwork restores face structure with more facial details and performs favorably\nagainst state-of-the-art methods in terms of structured similarity indexing\nmethod (SSIM), peak signal-to-noise ratio (PSNR), Frechet inception distance\n(FID) and L1, and qualitative comparisons.', 'Recently, it was shown that excellent results can be achieved in both face\nlandmark localization and pose-invariant face recognition. These breakthroughs\nare attributed to the efforts of the community to manually annotate facial\nimages in many different poses and to collect 3D faces data. In this paper, we\npropose a novel method for joint face landmark localization and frontal face\nreconstruction (pose correction) using a small set of frontal images only. By\nobserving that the frontal facial image is the one with the minimum rank from\nall different poses we formulate an appropriate model which is able to jointly\nrecover the facial landmarks as well as the frontalized version of the face. To\nthis end, a suitable optimization problem, involving the minimization of the\nnuclear norm and the matrix $\\ell_1$ norm, is solved. The proposed method is\nassessed in frontal face reconstruction (pose correction), face landmark\nlocalization, and pose-invariant face recognition and verification by\nconducting experiments on $6$ facial images databases. The experimental results\ndemonstrate the effectiveness of the proposed method.', 'Deep convolutional networks (CNNs) have achieved great success in face\ncompletion to generate plausible facial structures. These methods, however, are\nlimited in maintaining global consistency among face components and recovering\nfine facial details. On the other hand, reflectional symmetry is a prominent\nproperty of face image and benefits face recognition and consistency modeling,\nyet remaining uninvestigated in deep face completion. In this work, we leverage\ntwo kinds of symmetry-enforcing subnets to form a symmetry-consistent CNN model\n(i.e., SymmFCNet) for effective face completion. For missing pixels on only one\nof the half-faces, an illumination-reweighted warping subnet is developed to\nguide the warping and illumination reweighting of the other half-face. As for\nmissing pixels on both of half-faces, we present a generative reconstruction\nsubnet together with a perceptual symmetry loss to enforce symmetry consistency\nof recovered structures. The SymmFCNet is constructed by stacking generative\nreconstruction subnet upon illumination-reweighted warping subnet, and can be\nend-to-end learned from training set of unaligned face images. Experiments show\nthat SymmFCNet can generate high quality results on images with synthetic and\nreal occlusion, and performs favorably against state-of-the-arts.', 'Although face recognition systems have achieved impressive performance in\nrecent years, the low-resolution face recognition (LRFR) task remains\nchallenging, especially when the LR faces are captured under non-ideal\nconditions, as is common in surveillance-based applications. Faces captured in\nsuch conditions are often contaminated by blur, nonuniform lighting, and\nnonfrontal face pose. In this paper, we analyze face recognition techniques\nusing data captured under low-quality conditions in the wild. We provide a\ncomprehensive analysis of experimental results for two of the most important\napplications in real surveillance applications, and demonstrate practical\napproaches to handle both cases that show promising performance. The following\nthree contributions are made: {\\em (i)} we conduct experiments to evaluate\nsuper-resolution methods for low-resolution face recognition; {\\em (ii)} we\nstudy face re-identification on various public face datasets including real\nsurveillance and low-resolution subsets of large-scale datasets, present a\nbaseline result for several deep learning based approaches, and improve them by\nintroducing a GAN pre-training approach and fully convolutional architecture;\nand {\\em (iii)} we explore low-resolution face identification by employing a\nstate-of-the-art supervised discriminative learning approach. Evaluations are\nconducted on challenging portions of the SCFace and UCCSface datasets.', 'Face recognition has been of great importance in many applications as a\nbiometric for its throughput, convenience, and non-invasiveness. Recent\nadvancements in deep Convolutional Neural Network (CNN) architectures have\nboosted significantly the performance of face recognition based on\ntwo-dimensional (2D) facial texture images and outperformed the previous state\nof the art using conventional methods. However, the accuracy of 2D face\nrecognition is still challenged by the change of pose, illumination, make-up,\nand expression. On the other hand, the geometric information contained in\nthree-dimensional (3D) face data has the potential to overcome the fundamental\nlimitations of 2D face data.\n  We propose a multi-Channel deep 3D face network for face recognition based on\n3D face data. We compute the geometric information of a 3D face based on its\npiecewise-linear triangular mesh structure and then conformally flatten\ngeometric information along with the color from 3D to 2D plane to leverage the\nstate-of-the-art deep CNN architectures. We modify the input layer of the\nnetwork to take images with nine channels instead of three only such that more\ngeometric information can be explicitly fed to it. We pre-train the network\nusing images from the VGG-Face \\cite{Parkhi2015} and then fine-tune it with the\ngenerated multi-channel face images. The face recognition accuracy of the\nmulti-Channel deep 3D face network has achieved 98.6. The experimental results\nalso clearly show that the network performs much better when a 9-channel image\nis flattened to plane based on the conformal map compared with the orthographic\nprojection.', 'In video based face recognition, face images are typically captured over\nmultiple frames in uncontrolled conditions, where head pose, illumination,\nshadowing, motion blur and focus change over the sequence. Additionally,\ninaccuracies in face localisation can also introduce scale and alignment\nvariations. Using all face images, including images of poor quality, can\nactually degrade face recognition performance. While one solution it to use\nonly the ""best"" subset of images, current face selection techniques are\nincapable of simultaneously handling all of the abovementioned issues. We\npropose an efficient patch-based face image quality assessment algorithm which\nquantifies the similarity of a face image to a probabilistic face model,\nrepresenting an ""ideal"" face. Image characteristics that affect recognition are\ntaken into account, including variations in geometric alignment (shift,\nrotation and scale), sharpness, head pose and cast shadows. Experiments on\nFERET and PIE datasets show that the proposed algorithm is able to identify\nimages which are simultaneously the most frontal, aligned, sharp and well\nilluminated. Further experiments on a new video surveillance dataset (termed\nChokePoint) show that the proposed method provides better face subsets than\nexisting face selection techniques, leading to significant improvements in\nrecognition accuracy.', '3D face reconstruction is an important task in the field of computer vision.\nAlthough 3D face reconstruction has being developing rapidly in recent years,\nit is still a challenge for face reconstruction under large pose. That is\nbecause much of the information about a face in a large pose will be\nunknowable. In order to address this issue, this paper proposes a novel 3D face\nreconstruction algorithm (PIFR) based on 3D Morphable Model (3DMM). After input\na single face image, it generates a frontal image by normalizing the image.\nThen we set weighted sum of the 3D parameters of the two images. Our method\nsolves the problem of face reconstruction of a single image of a traditional\nmethod in a large pose, works on arbitrary Pose and Expressions, greatly\nimproves the accuracy of reconstruction. Experiments on the challenging AFW,\nLFPW and AFLW database show that our algorithm significantly improves the\naccuracy of 3D face reconstruction even under extreme poses .', 'Face images in the wild undergo large intra-personal variations, such as\nposes, illuminations, occlusions, and low resolutions, which cause great\nchallenges to face-related applications. This paper addresses this challenge by\nproposing a new deep learning framework that can recover the canonical view of\nface images. It dramatically reduces the intra-person variances, while\nmaintaining the inter-person discriminativeness. Unlike the existing face\nreconstruction methods that were either evaluated in controlled 2D environment\nor employed 3D information, our approach directly learns the transformation\nfrom the face images with a complex set of variations to their canonical views.\nAt the training stage, to avoid the costly process of labeling canonical-view\nimages from the training set by hand, we have devised a new measurement to\nautomatically select or synthesize a canonical-view image for each identity. As\nan application, this face recovery approach is used for face verification.\nFacial features are learned from the recovered canonical-view face images by\nusing a facial component-based convolutional neural network. Our approach\nachieves the state-of-the-art performance on the LFW dataset.', 'Face recognition approaches often rely on equal image resolution for\nverifying faces on two images. However, in practical applications, those image\nresolutions are usually not in the same range due to different image capture\nmechanisms or sources. In this work, we first analyze the impact of image\nresolutions on face verification performance with a state-of-the-art face\nrecognition model. For images synthetically reduced to $5\\,\\times\\,5$ px\nresolution, the verification performance drops from $99.23\\%$ increasingly down\nto almost $55\\%$. Especially for cross-resolution image pairs (one high- and\none low-resolution image), the verification accuracy decreases even further. We\ninvestigate this behavior more in-depth by looking at the feature distances for\nevery 2-image test pair. To tackle this problem, we propose the following two\nmethods: 1) Train a state-of-the-art face-recognition model straightforwardly\nwith $50\\%$ low-resolution images directly within each batch. 2) Train a\nsiamese-network structure and add a cosine distance feature loss between high-\nand low-resolution features. Both methods show an improvement for\ncross-resolution scenarios and can increase the accuracy at very low resolution\nto approximately $70\\%$. However, a disadvantage is that a specific model needs\nto be trained for every resolution pair. Thus, we extend the aforementioned\nmethods by training them with multiple image resolutions at once. The\nperformances for particular testing image resolutions are slightly worse, but\nthe advantage is that this model can be applied to arbitrary resolution images\nand achieves an overall better performance ($97.72\\%$ compared to $96.86\\%$).\nDue to the lack of a benchmark for arbitrary resolution images for the\ncross-resolution and equal-resolution task, we propose an evaluation protocol\nfor five well-known datasets, focusing on high, mid, and low-resolution images.', 'Convolutional Neural Networks have reached extremely high performances on the\nFace Recognition task. Largely used datasets, such as VGGFace2, focus on\ngender, pose and age variations trying to balance them to achieve better\nresults. However, the fact that images have different resolutions is not\nusually discussed and resize to 256 pixels before cropping is used. While\nspecific datasets for very low resolution faces have been proposed, less\nattention has been payed on the task of cross-resolution matching. Such\nscenarios are of particular interest for forensic and surveillance systems in\nwhich it usually happens that a low-resolution probe has to be matched with\nhigher-resolution galleries. While it is always possible to either increase the\nresolution of the probe image or to reduce the size of the gallery images, to\nthe best of our knowledge an extensive experimentation of cross-resolution\nmatching was missing in the recent deep learning based literature. In the\ncontext of low- and cross-resolution Face Recognition, the contributions of our\nwork are: i) we proposed a training method to fine-tune a state-of-the-art\nmodel in order to make it able to extract resolution-robust deep features; ii)\nwe tested our models on the benchmark datasets IJB-B/C considering images at\nboth full and low resolutions in order to show the effectiveness of the\nproposed training algorithm. To the best of our knowledge, this is the first\nwork testing extensively the performance of a FR model in a cross-resolution\nscenario; iii) we tested our models on the low resolution and low quality\ndatasets QMUL-SurvFace and TinyFace and showed their superior performances,\neven though we did not train our model on low-resolution faces only and our\nmain focus was cross-resolution; iv) we showed that our approach can be more\neffective with respect to preprocessing faces with super resolution techniques.', 'In this paper, we present an effective and efficient face deblurring\nalgorithm by exploiting semantic cues via deep convolutional neural networks\n(CNNs). As face images are highly structured and share several key semantic\ncomponents (e.g., eyes and mouths), the semantic information of a face provides\na strong prior for restoration. As such, we propose to incorporate global\nsemantic priors as input and impose local structure losses to regularize the\noutput within a multi-scale deep CNN. We train the network with perceptual and\nadversarial losses to generate photo-realistic results and develop an\nincremental training strategy to handle random blur kernels in the wild.\nQuantitative and qualitative evaluations demonstrate that the proposed face\ndeblurring algorithm restores sharp images with more facial details and\nperforms favorably against state-of-the-art methods in terms of restoration\nquality, face recognition and execution speed.', 'As a domain-specific super-resolution problem, facial image hallucination has\nenjoyed a series of breakthroughs thanks to the advances of deep convolutional\nneural networks. However, the direct migration of existing methods to video is\nstill difficult to achieve good performance due to its lack of alignment and\nconsistency modelling in temporal domain. Taking advantage of high inter-frame\ndependency in videos, we propose a self-enhanced convolutional network for\nfacial video hallucination. It is implemented by making full usage of preceding\nsuper-resolved frames and a temporal window of adjacent low-resolution frames.\nSpecifically, the algorithm first obtains the initial high-resolution inference\nof each frame by taking into consideration a sequence of consecutive\nlow-resolution inputs through temporal consistency modelling. It further\nrecurrently exploits the reconstructed results and intermediate features of a\nsequence of preceding frames to improve the initial super-resolution of the\ncurrent frame by modelling the coherence of structural facial features across\nframes. Quantitative and qualitative evaluations demonstrate the superiority of\nthe proposed algorithm against state-of-the-art methods. Moreover, our\nalgorithm also achieves excellent performance in the task of general video\nsuper-resolution in a single-shot setting.', 'Face super-resolution (FSR), also known as face hallucination, which is aimed\nat enhancing the resolution of low-resolution (LR) face images to generate\nhigh-resolution (HR) face images, is a domain-specific image super-resolution\nproblem. Recently, FSR has received considerable attention and witnessed\ndazzling advances with the development of deep learning techniques. To date,\nfew summaries of the studies on the deep learning-based FSR are available. In\nthis survey, we present a comprehensive review of deep learning-based FSR\nmethods in a systematic manner. First, we summarize the problem formulation of\nFSR and introduce popular assessment metrics and loss functions. Second, we\nelaborate on the facial characteristics and popular datasets used in FSR.\nThird, we roughly categorize existing methods according to the utilization of\nfacial characteristics. In each category, we start with a general description\nof design principles, then present an overview of representative approaches,\nand then discuss the pros and cons among them. Fourth, we evaluate the\nperformance of some state-of-the-art methods. Fifth, joint FSR and other tasks,\nand FSR-related applications are roughly introduced. Finally, we envision the\nprospects of further technological advancement in this field. A curated list of\npapers and resources to face super-resolution are available at\n\\url{https://github.com/junjun-jiang/Face-Hallucination-Benchmark}', 'Face analysis techniques have become a crucial component of human-machine\ninteraction in the fields of assistive and humanoid robotics. However, the\nvariations in head-pose that arise naturally in these environments are still a\ngreat challenge. In this paper, we present a real-time capable 3D face\nmodelling framework for 2D in-the-wild images that is applicable for robotics.\nThe fitting of the 3D Morphable Model is based exclusively on automatically\ndetected landmarks. After fitting, the face can be corrected in pose and\ntransformed back to a frontal 2D representation that is more suitable for face\nrecognition. We conduct face recognition experiments with non-frontal images\nfrom the MUCT database and uncontrolled, in the wild images from the PaSC\ndatabase, the most challenging face recognition database to date, showing an\nimproved performance. Finally, we present our SCITOS G5 robot system, which\nincorporates our framework as a means of image pre-processing for face\nanalysis.', 'We propose an Ensemble of Robust Constrained Local Models for alignment of\nfaces in the presence of significant occlusions and of any unknown pose and\nexpression. To account for partial occlusions we introduce, Robust Constrained\nLocal Models, that comprises of a deformable shape and local landmark\nappearance model and reasons over binary occlusion labels. Our occlusion\nreasoning proceeds by a hypothesize-and-test search over occlusion labels.\nHypotheses are generated by Constrained Local Model based shape fitting over\nrandomly sampled subsets of landmark detector responses and are evaluated by\nthe quality of face alignment. To span the entire range of facial pose and\nexpression variations we adopt an ensemble of independent Robust Constrained\nLocal Models to search over a discretized representation of pose and\nexpression. We perform extensive evaluation on a large number of face images,\nboth occluded and unoccluded. We find that our face alignment system trained\nentirely on facial images captured ""in-the-lab"" exhibits a high degree of\ngeneralization to facial images captured ""in-the-wild"". Our results are\naccurate and stable over a wide spectrum of occlusions, pose and expression\nvariations resulting in excellent performance on many real-world face datasets.', 'Recently, deep learning based 3D face reconstruction methods have shown\npromising results in both quality and efficiency.However, training deep neural\nnetworks typically requires a large volume of data, whereas face images with\nground-truth 3D face shapes are scarce. In this paper, we propose a novel deep\n3D face reconstruction approach that 1) leverages a robust, hybrid loss\nfunction for weakly-supervised learning which takes into account both low-level\nand perception-level information for supervision, and 2) performs multi-image\nface reconstruction by exploiting complementary information from different\nimages for shape aggregation. Our method is fast, accurate, and robust to\nocclusion and large pose. We provide comprehensive experiments on three\ndatasets, systematically comparing our method with fifteen recent methods and\ndemonstrating its state-of-the-art performance.', 'In this paper, we address the problem of face hallucination by proposing a\nnovel multi-scale generative adversarial network (GAN) architecture optimized\nfor face verification. First, we propose a multi-scale generator architecture\nfor face hallucination with a high up-scaling ratio factor, which has multiple\nintermediate outputs at different resolutions. The intermediate outputs have\nthe growing goal of synthesizing small to large images. Second, we incorporate\na face verifier with the original GAN discriminator and propose a novel\ndiscriminator which learns to discriminate different identities while\ndistinguishing fake generated HR face images from their ground truth images. In\nparticular, the learned generator cares for not only the visual quality of\nhallucinated face images but also preserving the discriminative features in the\nhallucination process. In addition, to capture perceptually relevant\ndifferences we employ a perceptual similarity loss, instead of similarity in\npixel space. We perform a quantitative and qualitative evaluation of our\nframework on the LFW and CelebA datasets. The experimental results show the\nadvantages of our proposed method against the state-of-the-art methods on the\n8x downsampled testing dataset.', 'We address the problem of restoring a high-resolution face image from a\nblurry low-resolution input. This problem is difficult as super-resolution and\ndeblurring need to be tackled simultaneously. Moreover, existing algorithms\ncannot handle face images well as low-resolution face images do not have much\ntexture which is especially critical for deblurring. In this paper, we propose\nan effective algorithm by utilizing the domain-specific knowledge of human\nfaces to recover high-quality faces. We first propose a facial component guided\ndeep Convolutional Neural Network (CNN) to restore a coarse face image, which\nis denoted as the base image where the facial component is automatically\ngenerated from the input face image. However, the CNN based method cannot\nhandle image details well. We further develop a novel exemplar-based detail\nenhancement algorithm via facial component matching. Extensive experiments show\nthat the proposed method outperforms the state-of-the-art algorithms both\nquantitatively and qualitatively.', 'Recent advances in image-based 3D human shape estimation have been driven by\nthe significant improvement in representation power afforded by deep neural\nnetworks. Although current approaches have demonstrated the potential in real\nworld settings, they still fail to produce reconstructions with the level of\ndetail often present in the input images. We argue that this limitation stems\nprimarily form two conflicting requirements; accurate predictions require large\ncontext, but precise predictions require high resolution. Due to memory\nlimitations in current hardware, previous approaches tend to take low\nresolution images as input to cover large spatial context, and produce less\nprecise (or low resolution) 3D estimates as a result. We address this\nlimitation by formulating a multi-level architecture that is end-to-end\ntrainable. A coarse level observes the whole image at lower resolution and\nfocuses on holistic reasoning. This provides context to an fine level which\nestimates highly detailed geometry by observing higher-resolution images. We\ndemonstrate that our approach significantly outperforms existing\nstate-of-the-art techniques on single image human shape reconstruction by fully\nleveraging 1k-resolution input images.', 'Face recognition in real life situations like low illumination condition is\nstill an open challenge in biometric security. It is well established that the\nstate-of-the-art methods in face recognition provide low accuracy in the case\nof poor illumination. In this work, we propose an algorithm for a more robust\nillumination invariant face recognition using a multi-modal approach. We\npropose a new dataset consisting of aligned faces of thermal and visual images\nof a hundred subjects. We then apply face detection on thermal images using the\nbiggest blob extraction method and apply them for fusing images of different\nmodalities for the purpose of face recognition. An algorithm is proposed to\nimplement fusion of thermal and visual images. We reason for why relying on\nonly one modality can give erroneous results. We use a lighter and faster CNN\nmodel called MobileNet for the purpose of face recognition with faster\ninferencing and to be able to be use it in real time biometric systems. We test\nour proposed method on our own created dataset to show that real-time face\nrecognition on fused images shows far better results than using visual or\nthermal images separately.', 'Over the last years, with the advent of Generative Adversarial Networks\n(GANs), many face analysis tasks have accomplished astounding performance, with\napplications including, but not limited to, face generation and 3D face\nreconstruction from a single ""in-the-wild"" image. Nevertheless, to the best of\nour knowledge, there is no method which can produce high-resolution\nphotorealistic 3D faces from ""in-the-wild"" images and this can be attributed to\nthe: (a) scarcity of available data for training, and (b) lack of robust\nmethodologies that can successfully be applied on very high-resolution data. In\nthis paper, we introduce AvatarMe, the first method that is able to reconstruct\nphotorealistic 3D faces from a single ""in-the-wild"" image with an increasing\nlevel of detail. To achieve this, we capture a large dataset of facial shape\nand reflectance and build on a state-of-the-art 3D texture and shape\nreconstruction method and successively refine its results, while generating the\nper-pixel diffuse and specular components that are required for realistic\nrendering. As we demonstrate in a series of qualitative and quantitative\nexperiments, AvatarMe outperforms the existing arts by a significant margin and\nreconstructs authentic, 4K by 6K-resolution 3D faces from a single\nlow-resolution image that, for the first time, bridges the uncanny valley.', 'This paper proposes an encoder-decoder network to disentangle shape features\nduring 3D face reconstruction from single 2D images, such that the tasks of\nreconstructing accurate 3D face shapes and learning discriminative shape\nfeatures for face recognition can be accomplished simultaneously. Unlike\nexisting 3D face reconstruction methods, our proposed method directly regresses\ndense 3D face shapes from single 2D images, and tackles identity and residual\n(i.e., non-identity) components in 3D face shapes explicitly and separately\nbased on a composite 3D face shape model with latent representations. We devise\na training process for the proposed network with a joint loss measuring both\nface identification error and 3D face shape reconstruction error. To construct\ntraining data we develop a method for fitting 3D morphable model (3DMM) to\nmultiple 2D images of a subject. Comprehensive experiments have been done on\nMICC, BU3DFE, LFW and YTF databases. The results show that our method expands\nthe capacity of 3DMM for capturing discriminative shape features and facial\ndetail, and thus outperforms existing methods both in 3D face reconstruction\naccuracy and in face recognition accuracy.', 'We propose a framework based on Generative Adversarial Networks to\ndisentangle the identity and attributes of faces, such that we can conveniently\nrecombine different identities and attributes for identity preserving face\nsynthesis in open domains. Previous identity preserving face synthesis\nprocesses are largely confined to synthesizing faces with known identities that\nare already in the training dataset. To synthesize a face with identity outside\nthe training dataset, our framework requires one input image of that subject to\nproduce an identity vector, and any other input face image to extract an\nattribute vector capturing, e.g., pose, emotion, illumination, and even the\nbackground. We then recombine the identity vector and the attribute vector to\nsynthesize a new face of the subject with the extracted attribute. Our proposed\nframework does not need to annotate the attributes of faces in any way. It is\ntrained with an asymmetric loss function to better preserve the identity and\nstabilize the training process. It can also effectively leverage large amounts\nof unlabeled training face images to further improve the fidelity of the\nsynthesized faces for subjects that are not presented in the labeled training\nface dataset. Our experiments demonstrate the efficacy of the proposed\nframework. We also present its usage in a much broader set of applications\nincluding face frontalization, face attribute morphing, and face adversarial\nexample detection.', 'Face verification and recognition problems have seen rapid progress in recent\nyears, however recognition from small size images remains a challenging task\nthat is inherently intertwined with the task of face super-resolution. Tackling\nthis problem using multiple frames is an attractive idea, yet requires solving\nthe alignment problem that is also challenging for low-resolution faces. Here\nwe present a holistic system for multi-frame recognition, alignment, and\nsuperresolution of faces. Our neural network architecture restores the central\nframe of each input sequence additionally taking into account a number of\nadjacent frames and making use of sub-pixel movements. We present our results\nusing the popular dataset for video face recognition (YouTube Faces). We show a\nnotable improvement of identification score compared to several baselines\nincluding the one based on single-image super-resolution.', 'This paper is on image and face super-resolution. The vast majority of prior\nwork for this problem focus on how to increase the resolution of low-resolution\nimages which are artificially generated by simple bilinear down-sampling (or in\na few cases by blurring followed by down-sampling).We show that such methods\nfail to produce good results when applied to real-world low-resolution, low\nquality images. To circumvent this problem, we propose a two-stage process\nwhich firstly trains a High-to-Low Generative Adversarial Network (GAN) to\nlearn how to degrade and downsample high-resolution images requiring, during\ntraining, only unpaired high and low-resolution images. Once this is achieved,\nthe output of this network is used to train a Low-to-High GAN for image\nsuper-resolution using this time paired low- and high-resolution images. Our\nmain result is that this network can be now used to efectively increase the\nquality of real-world low-resolution images. We have applied the proposed\npipeline for the problem of face super-resolution where we report large\nimprovement over baselines and prior work although the proposed method is\npotentially applicable to other object categories.', 'State-of-the-art deep face recognition approaches report near perfect\nperformance on popular benchmarks, e.g., Labeled Faces in the Wild. However,\ntheir performance deteriorates significantly when they are applied on low\nquality images, such as those acquired by surveillance cameras. A further\nchallenge for low resolution face recognition for surveillance applications is\nthe matching of recorded low resolution probe face images with high resolution\nreference images, which could be the case in watchlist scenarios. In this\npaper, we have addressed these problems and investigated the factors that would\ncontribute to the identification performance of the state-of-the-art deep face\nrecognition models when they are applied to low resolution face recognition\nunder mismatched conditions. We have observed that the following factors affect\nperformance in a positive way: appearance variety and resolution distribution\nof the training dataset, resolution matching between the gallery and probe\nimages, and the amount of information included in the probe images. By\nleveraging this information, we have utilized deep face models trained on\nMS-Celeb-1M and fine-tuned on VGGFace2 dataset and achieved state-of-the-art\naccuracies on the SCFace and ICB-RW benchmarks, even without using any training\ndata from the datasets of these benchmarks.', 'We propose a novel couple mappings method for low resolution face recognition\nusing deep convolutional neural networks (DCNNs). The proposed architecture\nconsists of two branches of DCNNs to map the high and low resolution face\nimages into a common space with nonlinear transformations. The branch\ncorresponding to transformation of high resolution images consists of 14 layers\nand the other branch which maps the low resolution face images to the common\nspace includes a 5-layer super-resolution network connected to a 14-layer\nnetwork. The distance between the features of corresponding high and low\nresolution images are backpropagated to train the networks. Our proposed method\nis evaluated on FERET data set and compared with state-of-the-art competing\nmethods. Our extensive experimental results show that the proposed method\nsignificantly improves the recognition performance especially for very low\nresolution probe face images (11.4% improvement in recognition accuracy).\nFurthermore, it can reconstruct a high resolution image from its corresponding\nlow resolution probe image which is comparable with state-of-the-art\nsuper-resolution methods in terms of visual quality.', 'In this paper, we present a deep learning based image feature extraction\nmethod designed specifically for face images. To train the feature extraction\nmodel, we construct a large scale photo-realistic face image dataset with\nground-truth correspondence between multi-view face images, which are\nsynthesized from real photographs via an inverse rendering procedure. The deep\nface feature (DFF) is trained using correspondence between face images rendered\nfrom different views. Using the trained DFF model, we can extract a feature\nvector for each pixel of a face image, which distinguishes different facial\nregions and is shown to be more effective than general-purpose feature\ndescriptors for face-related tasks such as matching and alignment. Based on the\nDFF, we develop a robust face alignment method, which iteratively updates\nlandmarks, pose and 3D shape. Extensive experiments demonstrate that our method\ncan achieve state-of-the-art results for face alignment under highly\nunconstrained face images.', 'Low-resolution face recognition (LRFR) has received increasing attention over\nthe past few years. Its applications lie widely in the real-world environment\nwhen high-resolution or high-quality images are hard to capture. One of the\nbiggest demands for LRFR technologies is video surveillance. As the the number\nof surveillance cameras in the city increases, the videos that captured will\nneed to be processed automatically. However, those videos or images are usually\ncaptured with large standoffs, arbitrary illumination condition, and diverse\nangles of view. Faces in these images are generally small in size. Several\nstudies addressed this problem employed techniques like super resolution,\ndeblurring, or learning a relationship between different resolution domains. In\nthis paper, we provide a comprehensive review of approaches to low-resolution\nface recognition in the past five years. First, a general problem definition is\ngiven. Later, systematically analysis of the works on this topic is presented\nby catogory. In addition to describing the methods, we also focus on datasets\nand experiment settings. We further address the related works on unconstrained\nlow-resolution face recognition and compare them with the result that use\nsynthetic low-resolution data. Finally, we summarized the general limitations\nand speculate a priorities for the future effort.', 'The area of face recognition is one of the most widely researched areas in\nthe domain of computer vision and biometric. This is because, the non-intrusive\nnature of face biometric makes it comparatively more suitable for application\nin area of surveillance at public places such as airports. The application of\nprimitive methods in face recognition could not give very satisfactory\nperformance. However, with the advent of machine and deep learning methods and\ntheir application in face recognition, several major breakthroughs were\nobtained. The use of 2D Convolution Neural networks(2D CNN) in face recognition\ncrossed the human face recognition accuracy and reached to 99%. Still, robust\nface recognition in the presence of real world conditions such as variation in\nresolution, illumination and pose is a major challenge for researchers in face\nrecognition. In this work, we used video as input to the 3D CNN architectures\nfor capturing both spatial and time domain information from the video for face\nrecognition in real world environment. For the purpose of experimentation, we\nhave developed our own video dataset called CVBL video dataset. The use of 3D\nCNN for face recognition in videos shows promising results with DenseNets\nperforming the best with an accuracy of 97% on CVBL dataset.', 'Majority of the face recognition algorithms use query faces captured from\nuncontrolled, in the wild, environment. Often caused by the cameras limited\ncapabilities, it is common for these captured facial images to be blurred or\nlow resolution. Super resolution algorithms are therefore crucial in improving\nthe resolution of such images especially when the image size is small requiring\nenlargement. This paper aims to demonstrate the effect of one of the\nstate-of-the-art algorithms in the field of image super resolution. To\ndemonstrate the functionality of the algorithm, various before and after 3D\nface alignment cases are provided using the images from the Labeled Faces in\nthe Wild (lfw). Resulting images are subject to testing on a closed set face\nrecognition protocol using unsupervised algorithms with high dimension\nextracted features. The inclusion of super resolution algorithm resulted in\nsignificant improved recognition rate over recently reported results obtained\nfrom unsupervised algorithms.', 'With the recent success of deep neural networks, remarkable progress has been\nachieved on face recognition. However, collecting large-scale real-world\ntraining data for face recognition has turned out to be challenging, especially\ndue to the label noise and privacy issues. Meanwhile, existing face recognition\ndatasets are usually collected from web images, lacking detailed annotations on\nattributes (e.g., pose and expression), so the influences of different\nattributes on face recognition have been poorly investigated. In this paper, we\naddress the above-mentioned issues in face recognition using synthetic face\nimages, i.e., SynFace. Specifically, we first explore the performance gap\nbetween recent state-of-the-art face recognition models trained with synthetic\nand real face images. We then analyze the underlying causes behind the\nperformance gap, e.g., the poor intra-class variations and the domain gap\nbetween synthetic and real face images. Inspired by this, we devise the SynFace\nwith identity mixup (IM) and domain mixup (DM) to mitigate the above\nperformance gap, demonstrating the great potentials of synthetic data for face\nrecognition. Furthermore, with the controllable face synthesis model, we can\neasily manage different factors of synthetic face generation, including pose,\nexpression, illumination, the number of identities, and samples per identity.\nTherefore, we also perform a systematically empirical analysis on synthetic\nface images to provide some insights on how to effectively utilize synthetic\ndata for face recognition.', 'The objective of this paper is a neural network model that controls the pose\nand expression of a given face, using another face or modality (e.g. audio).\nThis model can then be used for lightweight, sophisticated video and image\nediting.\n  We make the following three contributions. First, we introduce a network,\nX2Face, that can control a source face (specified by one or more frames) using\nanother face in a driving frame to produce a generated frame with the identity\nof the source frame but the pose and expression of the face in the driving\nframe. Second, we propose a method for training the network fully\nself-supervised using a large collection of video data. Third, we show that the\ngeneration process can be driven by other modalities, such as audio or pose\ncodes, without any further training of the network.\n  The generation results for driving a face with another face are compared to\nstate-of-the-art self-supervised/supervised methods. We show that our approach\nis more robust than other methods, as it makes fewer assumptions about the\ninput data. We also show examples of using our framework for video face\nediting.', 'As the deep learning makes big progresses in still-image face recognition,\nunconstrained video face recognition is still a challenging task due to low\nquality face images caused by pose, blur, occlusion, illumination etc. In this\npaper we propose a network for face recognition which gives an explicit and\nquantitative quality score at the same time when a feature vector is extracted.\nTo our knowledge this is the first network that implements these two functions\nin one network online. This network is very simple by adding a quality network\nbranch to the baseline network of face recognition. It does not require\ntraining datasets with annotated face quality labels. We evaluate this network\non both still-image face datasets and video face datasets and achieve the\nstate-of-the-art performance in many cases. This network enables a lot of\napplications where an explicit face quality scpre is used. We demonstrate three\napplications of the explicit face quality, one of which is a progressive\nfeature aggregation scheme in online video face recognition. We design an\nexperiment to prove the benefits of using the face quality in this application.\nCode will be available at \\url{https://github.com/deepcam-cn/facequality}.', 'Nowadays, due to the ubiquitous visual media there are vast amounts of\nalready available high-resolution (HR) face images. Therefore, for\nsuper-resolving a given very low-resolution (LR) face image of a person it is\nvery likely to find another HR face image of the same person which can be used\nto guide the process. In this paper, we propose a convolutional neural network\n(CNN)-based solution, namely GWAInet, which applies super-resolution (SR) by a\nfactor 8x on face images guided by another unconstrained HR face image of the\nsame person with possible differences in age, expression, pose or size. GWAInet\nis trained in an adversarial generative manner to produce the desired high\nquality perceptual image results. The utilization of the HR guiding image is\nrealized via the use of a warper subnetwork that aligns its contents to the\ninput image and the use of a feature fusion chain for the extracted features\nfrom the warped guiding image and the input image. In training, the identity\nloss further helps in preserving the identity related features by minimizing\nthe distance between the embedding vectors of SR and HR ground truth images.\nContrary to the current state-of-the-art in face super-resolution, our method\ndoes not require facial landmark points for its training, which helps its\nrobustness and allows it to produce fine details also for the surrounding face\nregion in a uniform manner. Our method GWAInet produces photo-realistic images\nin upscaling factor 8x and outperforms state-of-the-art in quantitative terms\nand perceptual quality.', 'Landmark detection algorithms trained on high resolution images perform\npoorly on datasets containing low resolution images. This deters the\nperformance of algorithms relying on quality landmarks, for example, face\nrecognition. To the best of our knowledge, there does not exist any dataset\nconsisting of low resolution face images along with their annotated landmarks,\nmaking supervised training infeasible. In this paper, we present a\nsemi-supervised approach to predict landmarks on low resolution images by\nlearning them from labeled high resolution images. The objective of this work\nis to show that predicting landmarks directly on low resolution images is more\neffective than the current practice of aligning images after rescaling or\nsuperresolution. In a two-step process, the proposed approach first learns to\ngenerate low resolution images by modeling the distribution of target low\nresolution images. In the second stage, the roles of generated images and real\nlow resolution images are switched and the model learns to predict landmarks\nfor real low resolution images from generated low resolution images. With\nextensive experimentation, we study the impact of each of the design choices\nand also show that prediction of landmarks directly on low resolution images\nimproves the performance of important tasks such as face recognition in low\nresolution images.', 'Face liveness detection is an essential prerequisite for face recognition\napplications. Previous face liveness detection methods usually train a binary\nclassifier to differentiate between a fake face and a real face before face\nrecognition. The client identity information is not utilized in previous face\nliveness detection methods. However, in practical face recognition\napplications, face spoofing attacks are always aimed at a specific client, and\nthe client identity information can provide useful clues for face liveness\ndetection. In this paper, we propose a face liveness detection method based on\nthe client identity using Siamese network. We detect face liveness after face\nrecognition instead of before face recognition, that is, we detect face\nliveness with the client identity information. We train a Siamese network with\nimage pairs. Each image pair consists of two real face images or one real and\none fake face images. The face images in each pair come from a same client.\nGiven a test face image, the face image is firstly recognized by face\nrecognition system, then the real face image of the identified client is\nretrieved to help the face liveness detection. Experiment results demonstrate\nthe effectiveness of our method.', 'With the growing importance of preventing the COVID-19 virus, face images\nobtained in most video surveillance scenarios are low resolution with mask\nsimultaneously. However, most of the previous face super-resolution solutions\ncan not handle both tasks in one model. In this work, we treat the mask\nocclusion as image noise and construct a joint and collaborative learning\nnetwork, called JDSR-GAN, for the masked face super-resolution task. Given a\nlow-quality face image with the mask as input, the role of the generator\ncomposed of a denoising module and super-resolution module is to acquire a\nhigh-quality high-resolution face image. The discriminator utilizes some\ncarefully designed loss functions to ensure the quality of the recovered face\nimages. Moreover, we incorporate the identity information and attention\nmechanism into our network for feasible correlated feature expression and\ninformative feature learning. By jointly performing denoising and face\nsuper-resolution, the two tasks can complement each other and attain promising\nperformance. Extensive qualitative and quantitative results show the\nsuperiority of our proposed JDSR-GAN over some comparable methods which perform\nthe previous two tasks separately.', 'Face quality assessment aims at estimating the utility of a face image for\nthe purpose of recognition. It is a key factor to achieve high face recognition\nperformances. Currently, the high performance of these face recognition systems\ncome with the cost of a strong bias against demographic and non-demographic\nsub-groups. Recent work has shown that face quality assessment algorithms\nshould adapt to the deployed face recognition system, in order to achieve\nhighly accurate and robust quality estimations. However, this could lead to a\nbias transfer towards the face quality assessment leading to discriminatory\neffects e.g. during enrolment. In this work, we present an in-depth analysis of\nthe correlation between bias in face recognition and face quality assessment.\nExperiments were conducted on two publicly available datasets captured under\ncontrolled and uncontrolled circumstances with two popular face embeddings. We\nevaluated four state-of-the-art solutions for face quality assessment towards\nbiases to pose, ethnicity, and age. The experiments showed that the face\nquality assessment solutions assign significantly lower quality values towards\nsubgroups affected by the recognition bias demonstrating that these approaches\nare biased as well. This raises ethical questions towards fairness and\ndiscrimination which future works have to address.', 'Learned 3D representations of human faces are useful for computer vision\nproblems such as 3D face tracking and reconstruction from images, as well as\ngraphics applications such as character generation and animation. Traditional\nmodels learn a latent representation of a face using linear subspaces or\nhigher-order tensor generalizations. Due to this linearity, they can not\ncapture extreme deformations and non-linear expressions. To address this, we\nintroduce a versatile model that learns a non-linear representation of a face\nusing spectral convolutions on a mesh surface. We introduce mesh sampling\noperations that enable a hierarchical mesh representation that captures\nnon-linear variations in shape and expression at multiple scales within the\nmodel. In a variational setting, our model samples diverse realistic 3D faces\nfrom a multivariate Gaussian distribution. Our training data consists of 20,466\nmeshes of extreme expressions captured over 12 different subjects. Despite\nlimited training data, our trained model outperforms state-of-the-art face\nmodels with 50% lower reconstruction error, while using 75% fewer parameters.\nWe also show that, replacing the expression space of an existing\nstate-of-the-art face model with our autoencoder, achieves a lower\nreconstruction error. Our data, model and code are available at\nhttp://github.com/anuragranj/coma', 'Fast and robust three-dimensional reconstruction of facial geometric\nstructure from a single image is a challenging task with numerous applications.\nHere, we introduce a learning-based approach for reconstructing a\nthree-dimensional face from a single image. Recent face recovery methods rely\non accurate localization of key characteristic points. In contrast, the\nproposed approach is based on a Convolutional-Neural-Network (CNN) which\nextracts the face geometry directly from its image. Although such deep\narchitectures outperform other models in complex computer vision problems,\ntraining them properly requires a large dataset of annotated examples. In the\ncase of three-dimensional faces, currently, there are no large volume data\nsets, while acquiring such big-data is a tedious task. As an alternative, we\npropose to generate random, yet nearly photo-realistic, facial images for which\nthe geometric form is known. The suggested model successfully recovers facial\nshapes from real images, even for faces with extreme expressions and under\nvarious lighting conditions.', 'Recent progress in deep learning-based models has improved photo-realistic\n(or perceptual) single-image super-resolution significantly. However, despite\ntheir powerful performance, many methods are difficult to apply to real-world\napplications because of the heavy computational requirements. To facilitate the\nuse of a deep model under such demands, we focus on keeping the network\nefficient while maintaining its performance. In detail, we design an\narchitecture that implements a cascading mechanism on a residual network to\nboost the performance with limited resources via multi-level feature fusion. In\naddition, our proposed model adopts group convolution and recursive schemes in\norder to achieve extreme efficiency. We further improve the perceptual quality\nof the output by employing the adversarial learning paradigm and a multi-scale\ndiscriminator approach. The performance of our method is investigated through\nextensive internal experiments and benchmarks using various datasets. Our\nresults show that our models outperform the recent methods with similar\ncomplexity, for both traditional pixel-based and perception-based tasks.', ""Deep Convolutional Neural Networks (CNNs) have been one of the most\ninfluential recent developments in computer vision, particularly for\ncategorization. There is an increasing demand for explainable AI as these\nsystems are deployed in the real world. However, understanding the information\nrepresented and processed in CNNs remains in most cases challenging. Within\nthis paper, we explore the use of new information theoretic techniques\ndeveloped in the field of neuroscience to enable novel understanding of how a\nCNN represents information. We trained a 10-layer ResNet architecture to\nidentify 2,000 face identities from 26M images generated using a rigorously\ncontrolled 3D face rendering model that produced variations of intrinsic (i.e.\nface morphology, gender, age, expression and ethnicity) and extrinsic factors\n(i.e. 3D pose, illumination, scale and 2D translation). With our methodology,\nwe demonstrate that unlike human's network overgeneralizes face identities even\nwith extreme changes of face shape, but it is more sensitive to changes of\ntexture. To understand the processing of information underlying these\ncounterintuitive properties, we visualize the features of shape and texture\nthat the network processes to identify faces. Then, we shed a light into the\ninner workings of the black box and reveal how hidden layers represent these\nfeatures and whether the representations are invariant to pose. We hope that\nour methodology will provide an additional valuable tool for interpretability\nof CNNs."", 'We present a versatile model, FaceAnime, for various video generation tasks\nfrom still images. Video generation from a single face image is an interesting\nproblem and usually tackled by utilizing Generative Adversarial Networks (GANs)\nto integrate information from the input face image and a sequence of sparse\nfacial landmarks. However, the generated face images usually suffer from\nquality loss, image distortion, identity change, and expression mismatching due\nto the weak representation capacity of the facial landmarks. In this paper, we\npropose to ""imagine"" a face video from a single face image according to the\nreconstructed 3D face dynamics, aiming to generate a realistic and\nidentity-preserving face video, with precisely predicted pose and facial\nexpression. The 3D dynamics reveal changes of the facial expression and motion,\nand can serve as a strong prior knowledge for guiding highly realistic face\nvideo generation. In particular, we explore face video prediction and exploit a\nwell-designed 3D dynamic prediction network to predict a 3D dynamic sequence\nfor a single face image. The 3D dynamics are then further rendered by the\nsparse texture mapping algorithm to recover structural details and sparse\ntextures for generating face frames. Our model is versatile for various AR/VR\nand entertainment applications, such as face video retargeting and face video\nprediction. Superior experimental results have well demonstrated its\neffectiveness in generating high-fidelity, identity-preserving, and visually\npleasant face video clips from a single source face image.', 'Face detection and tracking in low resolution image is not a trivial task due\nto the limitation in the appearance features for face characterization.\nMoreover, facial expression gives additional distortion on this small and noisy\nface. In this paper, we propose deformable symmetric Gabor wavelet network face\nmodel for face detection in low resolution image. Our model optimizes the\nrotation, translation, dilation, perspective and partial deformation amount of\nthe face model with symmetry constraints. Symmetry constraints help our model\nto be more robust to noise and distortion. Experimental results on our low\nresolution face image dataset and videos show promising face detection and\ntracking results under various challenging conditions.', 'Face recognition (FR) is one of the most extensively investigated problems in\ncomputer vision. Significant progress in FR has been made due to the recent\nintroduction of the larger scale FR challenges, particularly with constrained\nsocial media web images, e.g. high-resolution photos of celebrity faces taken\nby professional photo-journalists. However, the more challenging FR in\nunconstrained and low-resolution surveillance images remains largely\nunder-studied. To facilitate more studies on developing FR models that are\neffective and robust for low-resolution surveillance facial images, we\nintroduce a new Surveillance Face Recognition Challenge, which we call the\nQMUL-SurvFace benchmark. This new benchmark is the largest and more importantly\nthe only true surveillance FR benchmark to our best knowledge, where\nlow-resolution images are not synthesised by artificial down-sampling of native\nhigh-resolution images. This challenge contains 463,507 face images of 15,573\ndistinct identities captured in real-world uncooperative surveillance scenes\nover wide space and time. As a consequence, it presents an extremely\nchallenging FR benchmark. We benchmark the FR performance on this challenge\nusing five representative deep learning face recognition models, in comparison\nto existing benchmarks. We show that the current state of the arts are still\nfar from being satisfactory to tackle the under-investigated surveillance FR\nproblem in practical forensic scenarios. Face recognition is generally more\ndifficult in an open-set setting which is typical for surveillance scenarios,\nowing to a large number of non-target people (distractors) appearing open\nspaced scenes. This is evidently so that on the new Surveillance FR Challenge,\nthe top-performing CentreFace deep learning FR model on the MegaFace benchmark\ncan now only achieve 13.2% success rate (at Rank-20) at a 10% false alarm rate.', 'Most of the face hallucination methods are designed for complete inputs. They\nwill not work well if the inputs are very tiny or contaminated by large\nocclusion. Inspired by this fact, we propose an obscured face hallucination\nnetwork(OFHNet). The OFHNet consists of four parts: an inpainting network, an\nupsampling network, a discriminative network, and a fixed facial landmark\ndetection network. The inpainting network restores the low-resolution(LR)\nobscured face images. The following upsampling network is to upsample the\noutput of inpainting network. In order to ensure the generated\nhigh-resolution(HR) face images more photo-realistic, we utilize the\ndiscriminative network and the facial landmark detection network to better the\nresult of upsampling network. In addition, we present a semantic structure\nloss, which makes the generated HR face images more pleasing. Extensive\nexperiments show that our framework can restore the appealing HR face images\nfrom 1/4 missing area LR face images with a challenging scaling factor of 8x.', 'Visual kinship verification entails confirming whether or not two individuals\nin a given pair of images or videos share a hypothesized kin relation. As a\ngeneralized face verification task, visual kinship verification is particularly\ndifficult with low-quality found Internet data. Due to uncontrolled variations\nin background, pose, facial expression, blur, illumination and occlusion,\nstate-of-the-art methods fail to provide high level of recognition accuracy. As\nwith many other visual recognition tasks, kinship verification may benefit from\ncombining visual and audio signals. However, voice-based kinship verification\nhas received very little prior attention. We hypothesize that the human voice\ncontains kin-related cues that are complementary to visual cues. In this paper,\nwe address, for the first time, the use of audio-visual information from face\nand voice modalities to perform kinship verification. We first propose a new\nmulti-modal kinship dataset, called TALking KINship (TALKIN), that contains\nseveral pairs of Internet-quality video sequences. Using TALKIN, we then study\nthe utility of various kinship verification methods including traditional local\nfeature based methods, statistical methods and more recent deep learning\napproaches. Then, early and late fusion methods are evaluated on the TALKIN\ndataset for the study of kinship verification with both face and voice\nmodalities. Finally, we propose a deep Siamese fusion network with contrastive\nloss for multi-modal fusion of kinship relations. Extensive experiments on the\nTALKIN dataset indicate that by combining face and voice modalities, the\nproposed Siamese network can provide a significantly higher level of accuracy\ncompared to baseline uni-modal and multi-modal fusion techniques. Experimental\nresults also indicate that audio (vocal) information is complementary (to\nfacial information) and useful for kinship verification.', 'We propose using the Wasserstein loss for training in inverse problems. In\nparticular, we consider a learned primal-dual reconstruction scheme for\nill-posed inverse problems using the Wasserstein distance as loss function in\nthe learning. This is motivated by miss-alignments in training data, which when\nusing standard mean squared error loss could severely degrade reconstruction\nquality. We prove that training with the Wasserstein loss gives a\nreconstruction operator that correctly compensates for miss-alignments in\ncertain cases, whereas training with the mean squared error gives a smeared\nreconstruction. Moreover, we demonstrate these effects by training a\nreconstruction algorithm using both mean squared error and optimal transport\nloss for a problem in computerized tomography.', 'Face recognition has been widely studied due to its importance in different\napplications; however, most of the proposed methods fail when face images are\noccluded or captured under illumination and pose variations. Recently several\nlow-rank dictionary learning methods have been proposed and achieved promising\nresults for noisy observations. While these methods are mostly developed for\nsingle-modality scenarios, recent studies demonstrated the advantages of\nfeature fusion from multiple inputs. We propose a multi-modal structured\nlow-rank dictionary learning method for robust face recognition, using raw\npixels of face images and their illumination invariant representation. The\nproposed method learns robust and discriminative representations from\ncontaminated face images, even if there are few training samples with large\nintra-class variations. Extensive experiments on different datasets validate\nthe superior performance and robustness of our method to severe illumination\nvariations and occlusion.', '3D face reconstruction is a fundamental task that can facilitate numerous\napplications such as robust facial analysis and augmented reality. It is also a\nchallenging task due to the lack of high-quality datasets that can fuel current\ndeep learning-based methods. However, existing datasets are limited in\nquantity, realisticity and diversity. To circumvent these hurdles, we introduce\nPixel-Face, a large-scale, high-resolution and diverse 3D face dataset with\nmassive annotations. Specifically, Pixel-Face contains 855 subjects aging from\n18 to 80. Each subject has more than 20 samples with various expressions. Each\nsample is composed of high-resolution multi-view RGB images and 3D meshes with\nvarious expressions. Moreover, we collect precise landmarks annotation and 3D\nregistration result for each data. To demonstrate the advantages of Pixel-Face,\nwe re-parameterize the 3D Morphable Model (3DMM) into Pixel-3DM using the\ncollected data. We show that the obtained Pixel-3DM is better in modeling a\nwide range of face shapes and expressions. We also carefully benchmark existing\n3D face reconstruction methods on our dataset. Moreover, Pixel-Face serves as\nan effective training source. We observe that the performance of current face\nreconstruction models significantly improves both on existing benchmarks and\nPixel-Face after being fine-tuned using our newly collected data. Extensive\nexperiments demonstrate the effectiveness of Pixel-3DM and the usefulness of\nPixel-Face.', '3D human shape and pose estimation from monocular images has been an active\narea of research in computer vision, having a substantial impact on the\ndevelopment of new applications, from activity recognition to creating virtual\navatars. Existing deep learning methods for 3D human shape and pose estimation\nrely on relatively high-resolution input images; however, high-resolution\nvisual content is not always available in several practical scenarios such as\nvideo surveillance and sports broadcasting. Low-resolution images in real\nscenarios can vary in a wide range of sizes, and a model trained in one\nresolution does not typically degrade gracefully across resolutions. Two common\napproaches to solve the problem of low-resolution input are applying\nsuper-resolution techniques to the input images which may result in visual\nartifacts, or simply training one model for each resolution, which is\nimpractical in many realistic applications. To address the above issues, this\npaper proposes a novel algorithm called RSC-Net, which consists of a\nResolution-aware network, a Self-supervision loss, and a Contrastive learning\nscheme. The proposed network is able to learn the 3D body shape and pose across\ndifferent resolutions with a single model. The self-supervision loss encourages\nscale-consistency of the output, and the contrastive learning scheme enforces\nscale-consistency of the deep features. We show that both these new training\nlosses provide robustness when learning 3D shape and pose in a\nweakly-supervised manner. Extensive experiments demonstrate that the RSC-Net\ncan achieve consistently better results than the state-of-the-art methods for\nchallenging low-resolution images.', 'We present a simple method to reconstruct a high-resolution video from a\nface-video, where the identity of a person is obscured by pixelization. This\nconcealment method is popular because the viewer can still perceive a human\nface figure and the overall head motion. However, we show in our experiments\nthat a fairly good approximation of the original video can be reconstructed in\na way that compromises anonymity. Our system exploits the simultaneous\nsimilarity and small disparity between close-by video frames depicting a human\nface, and employs a spatial transformation component that learns the alignment\nbetween the pixelated frames. Each frame, supported by its aligned surrounding\nframes, is first encoded, then decoded to a higher resolution. Reconstruction\nand perceptual losses promote adherence to the ground-truth, and an adversarial\nloss assists in maintaining domain faithfulness. There is no need for explicit\ntemporal coherency loss as it is maintained implicitly by the alignment of\nneighboring frames and reconstruction. Although simple, our framework\nsynthesizes high-quality face reconstructions, demonstrating that given the\nstatistical prior of a human face, multiple aligned pixelated frames contain\nsufficient information to reconstruct a high-quality approximation of the\noriginal signal.', 'Face detection is to search all the possible regions for faces in images and\nlocate the faces if there are any. Many applications including face\nrecognition, facial expression recognition, face tracking and head-pose\nestimation assume that both the location and the size of faces are known in the\nimage. In recent decades, researchers have created many typical and efficient\nface detectors from the Viola-Jones face detector to current CNN-based ones.\nHowever, with the tremendous increase in images and videos with variations in\nface scale, appearance, expression, occlusion and pose, traditional face\ndetectors are challenged to detect various ""in the wild"" faces. The emergence\nof deep learning techniques brought remarkable breakthroughs to face detection\nalong with the price of a considerable increase in computation. This paper\nintroduces representative deep learning-based methods and presents a deep and\nthorough analysis in terms of accuracy and efficiency. We further compare and\ndiscuss the popular and challenging datasets and their evaluation metrics. A\ncomprehensive comparison of several successful deep learning-based face\ndetectors is conducted to uncover their efficiency using two metrics: FLOPs and\nlatency. The paper can guide to choose appropriate face detectors for different\napplications and also to develop more efficient and accurate detectors.', 'Face detection is a well-explored problem. Many challenges on face detectors\nlike extreme pose, illumination, low resolution and small scales are studied in\nthe previous work. However, previous proposed models are mostly trained and\ntested on good-quality images which are not always the case for practical\napplications like surveillance systems. In this paper, we first review the\ncurrent state-of-the-art face detectors and their performance on benchmark\ndataset FDDB, and compare the design protocols of the algorithms. Secondly, we\ninvestigate their performance degradation while testing on low-quality images\nwith different levels of blur, noise, and contrast. Our results demonstrate\nthat both hand-crafted and deep-learning based face detectors are not robust\nenough for low-quality images. It inspires researchers to produce more robust\ndesign for face detection in the wild.', '3D human pose and shape estimation from monocular images has been an active\nresearch area in computer vision. Existing deep learning methods for this task\nrely on high-resolution input, which however, is not always available in many\nscenarios such as video surveillance and sports broadcasting. Two common\napproaches to deal with low-resolution images are applying super-resolution\ntechniques to the input, which may result in unpleasant artifacts, or simply\ntraining one model for each resolution, which is impractical in many realistic\napplications.\n  To address the above issues, this paper proposes a novel algorithm called\nRSC-Net, which consists of a Resolution-aware network, a Self-supervision loss,\nand a Contrastive learning scheme. The proposed method is able to learn 3D body\npose and shape across different resolutions with one single model. The\nself-supervision loss enforces scale-consistency of the output, and the\ncontrastive learning scheme enforces scale-consistency of the deep features. We\nshow that both these new losses provide robustness when learning in a\nweakly-supervised manner. Moreover, we extend the RSC-Net to handle\nlow-resolution videos and apply it to reconstruct textured 3D pedestrians from\nlow-resolution input. Extensive experiments demonstrate that the RSC-Net can\nachieve consistently better results than the state-of-the-art methods for\nchallenging low-resolution images.', 'Although deep learning approaches have achieved performance surpassing humans\nfor still image-based face recognition, unconstrained video-based face\nrecognition is still a challenging task due to large volume of data to be\nprocessed and intra/inter-video variations on pose, illumination, occlusion,\nscene, blur, video quality, etc. In this work, we consider challenging\nscenarios for unconstrained video-based face recognition from multiple-shot\nvideos and surveillance videos with low-quality frames. To handle these\nproblems, we propose a robust and efficient system for unconstrained\nvideo-based face recognition, which is composed of modules for face/fiducial\ndetection, face association, and face recognition. First, we use multi-scale\nsingle-shot face detectors to efficiently localize faces in videos. The\ndetected faces are then grouped respectively through carefully designed face\nassociation methods, especially for multi-shot videos. Finally, the faces are\nrecognized by the proposed face matcher based on an unsupervised subspace\nlearning approach and a subspace-to-subspace similarity metric. Extensive\nexperiments on challenging video datasets, such as Multiple Biometric Grand\nChallenge (MBGC), Face and Ocular Challenge Series (FOCS), IARPA Janus\nSurveillance Video Benchmark (IJB-S) for low-quality surveillance videos and\nIARPA JANUS Benchmark B (IJB-B) for multiple-shot videos, demonstrate that the\nproposed system can accurately detect and associate faces from unconstrained\nvideos and effectively learn robust and discriminative features for\nrecognition.', 'Face recognition technology has been widely adopted in many mission-critical\nscenarios like means of human identification, controlled admission, and mobile\ndevice access, etc. Security surveillance is a typical scenario of face\nrecognition technology. Because the low-resolution feature of surveillance\nvideo and images makes it difficult for high-resolution face recognition\nalgorithms to extract effective feature information, Algorithms applied to\nhigh-resolution face recognition are difficult to migrate directly to\nlow-resolution situations. As face recognition in security surveillance becomes\nmore important in the era of dense urbanization, it is essential to develop\nalgorithms that are able to provide satisfactory performance in processing the\nvideo frames generated by low-resolution surveillance cameras. This paper study\non the Correlation Features-based Face Recognition (CoFFaR) method which using\nfor homogeneous low-resolution surveillance videos, the theory, experimental\ndetails, and experimental results are elaborated in detail. The experimental\nresults validate the effectiveness of the correlation features method that\nimproves the accuracy of homogeneous face recognition in surveillance security\nscenarios.', '3D face reconstruction is a challenging problem but also an important task in\nthe field of computer vision and graphics. Recently, many researchers put\nattention to the problem and a large number of articles have been published.\nSingle image reconstruction is one of the branches of 3D face reconstruction,\nwhich has a lot of applications in our life. This paper is a review of the\nrecent literature on 3D face reconstruction from a single image.', 'While existing face recognition systems based on local features are robust to\nissues such as misalignment, they can exhibit accuracy degradation when\ncomparing images of differing resolutions. This is common in surveillance\nenvironments where a gallery of high resolution mugshots is compared to low\nresolution CCTV probe images, or where the size of a given image is not a\nreliable indicator of the underlying resolution (eg. poor optics). To alleviate\nthis degradation, we propose a compensation framework which dynamically chooses\nthe most appropriate face recognition system for a given pair of image\nresolutions. This framework applies a novel resolution detection method which\ndoes not rely on the size of the input images, but instead exploits the\nsensitivity of local features to resolution using a probabilistic multi-region\nhistogram approach. Experiments on a resolution-modified version of the\n""Labeled Faces in the Wild"" dataset show that the proposed resolution detector\nfrontend obtains a 99% average accuracy in selecting the most appropriate face\nrecognition system, resulting in higher overall face discrimination accuracy\n(across several resolutions) compared to the individual baseline face\nrecognition systems.', 'This paper presents an easy and efficient face detection and face recognition\napproach using free software components from the internet. Face detection and\nface recognition problems have wide applications in home and office security.\nTherefore this work will helpful for those searching for a free face\noff-the-shelf face detection system. Using this system, faces can be detected\nin uncontrolled environments. In the detection phase, every individual face is\ndetected and in the recognition phase the detected faces are compared with the\nfaces in a given data set and recognized.', 'The widespread deployment of surveillance cameras for facial recognition\ngives rise to many privacy concerns. This study proposes a privacy-friendly\nalternative to large scale facial recognition. While there are multiple\ntechniques to preserve privacy, our work is based on the minimization principle\nwhich implies minimizing the amount of collected personal data. Instead of\nrunning facial recognition software on all video data, we propose to\nautomatically extract a high quality snapshot of each detected person without\nrevealing his or her identity. This snapshot is then encrypted and access is\nonly granted after legal authorization. We introduce a novel unsupervised face\nimage quality assessment method which is used to select the high quality\nsnapshots. For this, we train a variational autoencoder on high quality face\nimages from a publicly available dataset and use the reconstruction probability\nas a metric to estimate the quality of each face crop. We experimentally\nconfirm that the reconstruction probability can be used as biometric quality\npredictor. Unlike most previous studies, we do not rely on a manually defined\nface quality metric as everything is learned from data. Our face quality\nassessment method outperforms supervised, unsupervised and general image\nquality assessment methods on the task of improving face verification\nperformance by rejecting low quality images. The effectiveness of the whole\nsystem is validated qualitatively on still images and videos.', 'Video super-resolution aims at generating a high-resolution video from its\nlow-resolution counterpart. With the rapid rise of deep learning, many recently\nproposed video super-resolution methods use convolutional neural networks in\nconjunction with explicit motion compensation to capitalize on statistical\ndependencies within and across low-resolution frames. Two common issues of such\nmethods are noteworthy. Firstly, the quality of the final reconstructed HR\nvideo is often very sensitive to the accuracy of motion estimation. Secondly,\nthe warp grid needed for motion compensation, which is specified by the two\nflow maps delineating pixel displacements in horizontal and vertical\ndirections, tends to introduce additional errors and jeopardize the temporal\nconsistency across video frames. To address these issues, we propose a novel\ndynamic local filter network to perform implicit motion estimation and\ncompensation by employing, via locally connected layers, sample-specific and\nposition-specific dynamic local filters that are tailored to the target pixels.\nWe also propose a global refinement network based on ResBlock and autoencoder\nstructures to exploit non-local correlations and enhance the spatial\nconsistency of super-resolved frames. The experimental results demonstrate that\nthe proposed method outperforms the state-of-the-art, and validate its strength\nin terms of local transformation handling, temporal consistency as well as edge\nsharpness.', 'Cameras mounted on Micro Aerial Vehicles (MAVs) are increasingly used for\nrecreational photography. However, aerial photographs of public places often\ncontain faces of bystanders thus leading to a perceived or actual violation of\nprivacy. To address this issue, we propose to pseudo-randomly modify the\nappearance of face regions in the images using a privacy filter that prevents a\nhuman or a face recogniser from inferring the identities of people. The filter,\nwhich is applied only when the resolution is high enough for a face to be\nrecognisable, adaptively distorts the face appearance as a function of its\nresolution. Moreover, the proposed filter locally changes its parameters to\ndiscourage attacks that use parameter estimation. The filter exploits both\nglobal adaptiveness to reduce distortion and local hopping of the parameters to\nmake their estimation difficult for an attacker. In order to evaluate the\nefficiency of the proposed approach, we use a state-of-the-art face recognition\nalgorithm and synthetically generated face data with 3D geometric image\ntransformations that mimic faces captured from an MAV at different heights and\npitch angles. Experimental results show that the proposed filter protects\nprivacy while reducing distortion and exhibits resilience against attacks.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently engaged in a project that involves identifying and isolating concealed objects in videos, with a particular emphasis on tracking their movements. I am encountering difficulties in consistently outlining the edges of camouflaged objects, especially when certain parts of the video exhibit no movement. I require a system capable of aligning video frames based on the background while simultaneously tracking target objects and ensuring consistent tracking. The intended application of this model is in nature reserves, where it will be used to monitor the living conditions of endangered species. Therefore, I am seeking a model that can effectively handle various animal types and accurately label them during video storage. Ideally, the model should be trained on a comprehensive dataset encompassing a broad spectrum of animal species. Furthermore, I am interested in understanding how to apply supervised training models to unsupervised tasks, given that some of the animals may be too rare to be included in the training dataset.",[],"['Object tracking is one of the most important problems in computer vision. The\naim of video tracking is to extract the trajectories of a target or object of\ninterest, i.e. accurately locate a moving target in a video sequence and\ndiscriminate target from non-targets in the feature space of the sequence. So,\nfeature descriptors can have significant effects on such discrimination. In\nthis paper, we use the basic idea of many trackers which consists of three main\ncomponents of the reference model, i.e., object modeling, object detection and\nlocalization, and model updating. However, there are major improvements in our\nsystem. Our forth component, occlusion handling, utilizes the r-spatiogram to\ndetect the best target candidate. While spatiogram contains some moments upon\nthe coordinates of the pixels, r-spatiogram computes region-based compactness\non the distribution of the given feature in the image that captures richer\nfeatures to represent the objects. The proposed research develops an efficient\nand robust way to keep tracking the object throughout video sequences in the\npresence of significant appearance variations and severe occlusions. The\nproposed method is evaluated on the Princeton RGBD tracking dataset considering\nsequences with different challenges and the obtained results demonstrate the\neffectiveness of the proposed method.', 'Tracking is one of the most important but still difficult tasks in computer\nvision and pattern recognition. The main difficulties in the tracking field are\nappearance variation and occlusion. Most traditional tracking methods set the\nparameters or templates to track target objects in advance and should be\nmodified accordingly. Thus, we propose a new and robust tracking method using a\nFully Convolutional Network (FCN) to obtain an object probability map and\nDynamic Programming (DP) to seek the globally optimal path through all frames\nof video. Our proposed method solves the object appearance variation problem\nwith the use of a FCN and deals with occlusion by DP. We show that our method\nis effective in tracking various single objects through video frames.', 'For many years, multi-object tracking benchmarks have focused on a handful of\ncategories. Motivated primarily by surveillance and self-driving applications,\nthese datasets provide tracks for people, vehicles, and animals, ignoring the\nvast majority of objects in the world. By contrast, in the related field of\nobject detection, the introduction of large-scale, diverse datasets (e.g.,\nCOCO) have fostered significant progress in developing highly robust solutions.\nTo bridge this gap, we introduce a similarly diverse dataset for Tracking Any\nObject (TAO). It consists of 2,907 high resolution videos, captured in diverse\nenvironments, which are half a minute long on average. Importantly, we adopt a\nbottom-up approach for discovering a large vocabulary of 833 categories, an\norder of magnitude more than prior tracking benchmarks. To this end, we ask\nannotators to label objects that move at any point in the video, and give names\nto them post factum. Our vocabulary is both significantly larger and\nqualitatively different from existing tracking datasets. To ensure scalability\nof annotation, we employ a federated approach that focuses manual effort on\nlabeling tracks for those relevant objects in a video (e.g., those that move).\nWe perform an extensive evaluation of state-of-the-art trackers and make a\nnumber of important discoveries regarding large-vocabulary tracking in an\nopen-world. In particular, we show that existing single- and multi-object\ntrackers struggle when applied to this scenario in the wild, and that\ndetection-based, multi-object trackers are in fact competitive with\nuser-initialized ones. We hope that our dataset and analysis will boost further\nprogress in the tracking community.', 'Current state-of-the-art trackers only rely on a target appearance model in\norder to localize the object in each frame. Such approaches are however prone\nto fail in case of e.g. fast appearance changes or presence of distractor\nobjects, where a target appearance model alone is insufficient for robust\ntracking. Having the knowledge about the presence and locations of other\nobjects in the surrounding scene can be highly beneficial in such cases. This\nscene information can be propagated through the sequence and used to, for\ninstance, explicitly avoid distractor objects and eliminate target candidate\nregions.\n  In this work, we propose a novel tracking architecture which can utilize\nscene information for tracking. Our tracker represents such information as\ndense localized state vectors, which can encode, for example, if the local\nregion is target, background, or distractor. These state vectors are propagated\nthrough the sequence and combined with the appearance model output to localize\nthe target. Our network is learned to effectively utilize the scene information\nby directly maximizing tracking performance on video segments. The proposed\napproach sets a new state-of-the-art on 3 tracking benchmarks, achieving an AO\nscore of 63.6% on the recent GOT-10k dataset.', 'Robots navigating autonomously need to perceive and track the motion of\nobjects and other agents in its surroundings. This information enables planning\nand executing robust and safe trajectories. To facilitate these processes, the\nmotion should be perceived in 3D Cartesian space. However, most recent\nmulti-object tracking (MOT) research has focused on tracking people and moving\nobjects in 2D RGB video sequences. In this work we present JRMOT, a novel 3D\nMOT system that integrates information from RGB images and 3D point clouds to\nachieve real-time, state-of-the-art tracking performance. Our system is built\nwith recent neural networks for re-identification, 2D and 3D detection and\ntrack description, combined into a joint probabilistic data-association\nframework within a multi-modal recursive Kalman architecture. As part of our\nwork, we release the JRDB dataset, a novel large scale 2D+3D dataset and\nbenchmark, annotated with over 2 million boxes and 3500 time consistent 2D+3D\ntrajectories across 54 indoor and outdoor scenes. JRDB contains over 60 minutes\nof data including 360 degree cylindrical RGB video and 3D pointclouds in social\nsettings that we use to develop, train and evaluate JRMOT. The presented 3D MOT\nsystem demonstrates state-of-the-art performance against competing methods on\nthe popular 2D tracking KITTI benchmark and serves as first 3D tracking\nsolution for our benchmark. Real-robot tests on our social robot JackRabbot\nindicate that the system is capable of tracking multiple pedestrians fast and\nreliably. We provide the ROS code of our tracker at\nhttps://sites.google.com/view/jrmot.', ""Object tracking is an essential task in computer vision that has been studied\nsince the early days of the field. Being able to follow objects that undergo\ndifferent transformations in the video sequence, including changes in scale,\nillumination, shape and occlusions, makes the problem extremely difficult. One\nof the real challenges is to keep track of the changes in objects appearance\nand not drift towards the background clutter. Different from previous\napproaches, we obtain robustness against background with a tracker model that\nis composed of many different parts. They are classifiers that respond at\ndifferent scales and locations. The tracker system functions as a society of\nparts, each having its own role and level of credibility. Reliable classifiers\ndecide the tracker's next move, while newcomers are first monitored before\ngaining the necessary level of reliability to participate in the decision\nprocess. Some parts that loose their consistency are rejected, while others\nthat show consistency for a sufficiently long time are promoted to permanent\nroles. The tracker system, as a whole, could also go through different phases,\nfrom the usual, normal functioning to states of weak agreement and even crisis.\nThe tracker system has different governing rules in each state. What truly\ndistinguishes our work from others is not necessarily the strength of\nindividual tracking parts, but the way in which they work together and build a\nstrong and robust organization. We also propose an efficient way to learn\nsimultaneously many tracking parts, with a single closed-form formulation. We\nobtain a fast and robust tracker with state of the art performance on the\nchallenging OTB50 dataset."", ""Visual object tracking is an important function in many real-time video\nsurveillance applications, such as localization and spatio-temporal recognition\nof persons. In real-world applications, an object detector and tracker must\ninteract on a periodic basis to discover new objects, and thereby to initiate\ntracks. Periodic interactions with the detector can also allow the tracker to\nvalidate and/or update its object template with new bounding boxes. However,\nbounding boxes provided by a state-of-the-art detector are noisy, due to\nchanges in appearance, background and occlusion, which can cause the tracker to\ndrift. Moreover, CNN-based detectors can provide a high level of accuracy at\nthe expense of computational complexity, so interactions should be minimized\nfor real-time applications.\n  In this paper, a new approach is proposed to manage detector-tracker\ninteractions for trackers from the Siamese-FC family. By integrating a change\ndetection mechanism into a deep Siamese-FC tracker, its template can be adapted\nin response to changes in a target's appearance that lead to drifts during\ntracking. An abrupt change detection triggers an update of tracker template\nusing the bounding box produced by the detector, while in the case of a gradual\nchange, the detector is used to update an evolving set of templates for robust\nmatching.\n  Experiments were performed using state-of-the-art Siamese-FC trackers and the\nYOLOv3 detector on a subset of videos from the OTB-100 dataset that mimic video\nsurveillance scenarios. Results highlight the importance for reliable VOT of\nusing accurate detectors. They also indicate that our adaptive Siamese trackers\nare robust to noisy object detections, and can significantly improve the\nperformance of Siamese-FC tracking."", 'Traditionally, object tracking and segmentation are treated as two separate\nproblems and solved independently. However, in this paper, we argue that\ntracking and segmentation are actually closely related and solving one should\nhelp the other. On one hand, the object track, which is a set of bounding boxes\nwith one bounding box in every frame, would provide strong high-level guidance\nfor the target/background segmentation task. On the other hand, the object\nsegmentation would separate object from other objects and background, which\nwill be useful for determining track locations in every frame. We propose a\nnovel framework which combines online multiple target tracking and segmentation\nin a video. In our approach, the tracking and segmentation problems are coupled\nby Lagrange dual decomposition, which leads to more accurate segmentation\nresults and also \\emph{helps resolve typical difficulties in multiple target\ntracking, such as occlusion handling, ID-switch and track drifting}. To track\ntargets, an individual appearance model is learned for each target via\nstructured learning and network flow is employed to generate tracks from\ndensely sampled candidates. For segmentation, multi-label Conditional Random\nField (CRF) is applied to a superpixel based spatio-temporal graph in a segment\nof video to assign background or target labels to every superpixel. The\nexperiments on diverse sequences show that our method outperforms\nstate-of-the-art approaches for multiple target tracking as well as\nsegmentation.', 'We address Unsupervised Video Object Segmentation (UVOS), the task of\nautomatically generating accurate pixel masks for salient objects in a video\nsequence and of tracking these objects consistently through time, without any\ninput about which objects should be tracked. Towards solving this task, we\npresent UnOVOST (Unsupervised Offline Video Object Segmentation and Tracking)\nas a simple and generic algorithm which is able to track and segment a large\nvariety of objects. This algorithm builds up tracks in a number stages, first\ngrouping segments into short tracklets that are spatio-temporally consistent,\nbefore merging these tracklets into long-term consistent object tracks based on\ntheir visual similarity. In order to achieve this we introduce a novel\ntracklet-based Forest Path Cutting data association algorithm which builds up a\ndecision forest of track hypotheses before cutting this forest into paths that\nform long-term consistent object tracks. When evaluating our approach on the\nDAVIS 2017 Unsupervised dataset we obtain state-of-the-art performance with a\nmean J &F score of 67.9% on the val, 58% on the test-dev and 56.4% on the\ntest-challenge benchmarks, obtaining first place in the DAVIS 2019 Unsupervised\nVideo Object Segmentation Challenge. UnOVOST even performs competitively with\nmany semi-supervised video object segmentation algorithms even though it is not\ngiven any input as to which objects should be tracked and segmented.', ""Animals have been a common sighting on roads in India which leads to several\naccidents between them and vehicles every year. This makes it vital to develop\na support system for driverless vehicles that assists in preventing these forms\nof accidents. In this paper, we propose a neoteric framework for avoiding\nvehicle-to-animal collisions by developing an efficient approach for the\ndetection of animals on highways using deep learning and computer vision\ntechniques on dashcam video. Our approach leverages the Mask R-CNN model for\ndetecting and identifying various commonly found animals. Then, we perform lane\ndetection to deduce whether a detected animal is on the vehicle's lane or not\nand track its location and direction of movement using a centroid based object\ntracking algorithm. This approach ensures that the framework is effective at\ndetermining whether an animal is obstructing the path or not of an autonomous\nvehicle in addition to predicting its movement and giving feedback accordingly.\nThis system was tested under various lighting and weather conditions and was\nobserved to perform relatively well, which leads the way for prominent\ndriverless vehicle's support systems for avoiding vehicular collisions with\nanimals on Indian roads in real-time."", 'The problem of video object segmentation can become extremely challenging\nwhen multiple instances co-exist. While each instance may exhibit large scale\nand pose variations, the problem is compounded when instances occlude each\nother causing failures in tracking. In this study, we formulate a deep\nrecurrent network that is capable of segmenting and tracking objects in video\nsimultaneously by their temporal continuity, yet able to re-identify them when\nthey re-appear after a prolonged occlusion. We combine both temporal\npropagation and re-identification functionalities into a single framework that\ncan be trained end-to-end. In particular, we present a re-identification module\nwith template expansion to retrieve missing objects despite their large\nappearance changes. In addition, we contribute a new attention-based recurrent\nmask propagation approach that is robust to distractors not belonging to the\ntarget segment. Our approach achieves a new state-of-the-art global mean\n(Region Jaccard and Boundary F measure) of 68.2 on the challenging DAVIS 2017\nbenchmark (test-dev set), outperforming the winning solution which achieves a\nglobal mean of 66.1 on the same partition.', 'End-to-end production of object tracklets from high resolution video in\nreal-time and with high accuracy remains a challenging problem due to the cost\nof object detection on each frame. In this work we present Localization-based\nTracking (LBT), an extension to any tracker that follows the tracking by\ndetection or joint detection and tracking paradigms. Localization-based\nTracking focuses only on regions likely to contain objects to boost detection\nspeed and avoid matching errors. We evaluate LBT as an extension to two example\ntrackers (KIOU and SORT) on the UA-DETRAC and MOT20 datasets. LBT-extended\ntrackers outperform all other reported algorithms in terms of PR-MOTA, PR-MOTP,\nand mostly tracked objects on the UA-DETRAC benchmark, establishing a new\nstate-of-the art. relative to tracking by detection with KIOU, LBT-extended\nKIOU achieves a 25% higher frame-rate and is 1.1% more accurate in terms of\nPR-MOTA on the UA-DETRAC dataset. LBT-extended SORT achieves a 62% speedup and\na 3.2% increase in PR-MOTA on the UA-DETRAC dataset. On MOT20, LBT-extended\nKIOU has a 50% higher frame-rate than tracking by detection and is 0.4% more\naccurate in terms of MOTA. As of submission time, our LBT-extended KIOU tracker\nplaces 10th overall on the MOT20 benchmark.', 'This paper presents a method, called AOGTracker, for simultaneously tracking,\nlearning and parsing (TLP) of unknown objects in video sequences with a\nhierarchical and compositional And-Or graph (AOG) representation. %The AOG\ncaptures both structural and appearance variations of a target object in a\nprincipled way. The TLP method is formulated in the Bayesian framework with a\nspatial and a temporal dynamic programming (DP) algorithms inferring object\nbounding boxes on-the-fly. During online learning, the AOG is discriminatively\nlearned using latent SVM to account for appearance (e.g., lighting and partial\nocclusion) and structural (e.g., different poses and viewpoints) variations of\na tracked object, as well as distractors (e.g., similar objects) in background.\nThree key issues in online inference and learning are addressed: (i)\nmaintaining purity of positive and negative examples collected online, (ii)\ncontroling model complexity in latent structure learning, and (iii) identifying\ncritical moments to re-learn the structure of AOG based on its intrackability.\nThe intrackability measures uncertainty of an AOG based on its score maps in a\nframe. In experiments, our AOGTracker is tested on two popular tracking\nbenchmarks with the same parameter setting: the TB-100/50/CVPR2013 benchmarks,\nand the VOT benchmarks --- VOT 2013, 2014, 2015 and TIR2015 (thermal imagery\ntracking). In the former, our AOGTracker outperforms state-of-the-art tracking\nalgorithms including two trackers based on deep convolutional network. In the\nlatter, our AOGTracker outperforms all other trackers in VOT2013 and is\ncomparable to the state-of-the-art methods in VOT2014, 2015 and TIR2015.', 'In this paper, Object Detection and Tracking System (ODTS) in combination\nwith a well-known deep learning network, Faster Regional Convolution Neural\nNetwork (Faster R-CNN), for Object Detection and Conventional Object Tracking\nalgorithm will be introduced and applied for automatic detection and monitoring\nof unexpected events on CCTVs in tunnels, which are likely to (1) Wrong-Way\nDriving (WWD), (2) Stop, (3) Person out of vehicle in tunnel (4) Fire. ODTS\naccepts a video frame in time as an input to obtain Bounding Box (BBox) results\nby Object Detection and compares the BBoxs of the current and previous video\nframes to assign a unique ID number to each moving and detected object. This\nsystem makes it possible to track a moving object in time, which is not usual\nto be achieved in conventional object detection frameworks. A deep learning\nmodel in ODTS was trained with a dataset of event images in tunnels to Average\nPrecision (AP) values of 0.8479, 0.7161 and 0.9085 for target objects: Car,\nPerson, and Fire, respectively. Then, based on trained deep learning model, the\nODTS based Tunnel CCTV Accident Detection System was tested using four accident\nvideos which including each accident. As a result, the system can detect all\naccidents within 10 seconds. The more important point is that the detection\ncapacity of ODTS could be enhanced automatically without any changes in the\nprogram codes as the training dataset becomes rich.', 'In this paper, we address the basic problem of recognizing moving objects in\nvideo images using Visual Vocabulary model and Bag of Words and track our\nobject of interest in the subsequent video frames using species inspired PSO.\nInitially, the shadow free images are obtained by background modelling followed\nby foreground modeling to extract the blobs of our object of interest.\nSubsequently, we train a cubic SVM with human body datasets in accordance with\nour domain of interest for recognition and tracking. During training, using the\nprinciple of Bag of Words we extract necessary features of certain domains and\nobjects for classification. Subsequently, matching these feature sets with\nthose of the extracted object blobs that are obtained by subtracting the shadow\nfree background from the foreground, we detect successfully our object of\ninterest from the test domain. The performance of the classification by cubic\nSVM is satisfactorily represented by confusion matrix and ROC curve reflecting\nthe accuracy of each module. After classification, our object of interest is\ntracked in the test domain using species inspired PSO. By combining the\nadaptive learning tools with the efficient classification of description, we\nachieve optimum accuracy in recognition of the moving objects. We evaluate our\nalgorithm benchmark datasets: iLIDS, VIVID, Walking2, Woman. Comparative\nanalysis of our algorithm against the existing state-of-the-art trackers shows\nvery satisfactory and competitive results.', 'We propose an automatic system for organizing the content of a collection of\nunstructured videos of an articulated object class (e.g. tiger, horse). By\nexploiting the recurring motion patterns of the class across videos, our\nsystem: 1) identifies its characteristic behaviors; and 2) recovers\npixel-to-pixel alignments across different instances. Our system can be useful\nfor organizing video collections for indexing and retrieval. Moreover, it can\nbe a platform for learning the appearance or behaviors of object classes from\nInternet video. Traditional supervised techniques cannot exploit this wealth of\ndata directly, as they require a large amount of time-consuming manual\nannotations.\n  The behavior discovery stage generates temporal video intervals, each\nautomatically trimmed to one instance of the discovered behavior, clustered by\ntype. It relies on our novel motion representation for articulated motion based\non the displacement of ordered pairs of trajectories (PoTs). The alignment\nstage aligns hundreds of instances of the class to a great accuracy despite\nconsiderable appearance variations (e.g. an adult tiger and a cub). It uses a\nflexible Thin Plate Spline deformation model that can vary through time. We\ncarefully evaluate each step of our system on a new, fully annotated dataset.\nOn behavior discovery, we outperform the state-of-the-art Improved DTF\ndescriptor. On spatial alignment, we outperform the popular SIFT Flow\nalgorithm.', 'Having accurate, detailed, and up-to-date information about the location and\nbehavior of animals in the wild would revolutionize our ability to study and\nconserve ecosystems. We investigate the ability to automatically, accurately,\nand inexpensively collect such data, which could transform many fields of\nbiology, ecology, and zoology into ""big data"" sciences. Motion sensor ""camera\ntraps"" enable collecting wildlife pictures inexpensively, unobtrusively, and\nfrequently. However, extracting information from these pictures remains an\nexpensive, time-consuming, manual task. We demonstrate that such information\ncan be automatically extracted by deep learning, a cutting-edge type of\nartificial intelligence. We train deep convolutional neural networks to\nidentify, count, and describe the behaviors of 48 species in the\n3.2-million-image Snapshot Serengeti dataset. Our deep neural networks\nautomatically identify animals with over 93.8% accuracy, and we expect that\nnumber to improve rapidly in years to come. More importantly, if our system\nclassifies only images it is confident about, our system can automate animal\nidentification for 99.3% of the data while still performing at the same 96.6%\naccuracy as that of crowdsourced teams of human volunteers, saving more than\n8.4 years (at 40 hours per week) of human labeling effort (i.e. over 17,000\nhours) on this 3.2-million-image dataset. Those efficiency gains immediately\nhighlight the importance of using deep neural networks to automate data\nextraction from camera-trap images. Our results suggest that this technology\ncould enable the inexpensive, unobtrusive, high-volume, and even real-time\ncollection of a wealth of information about vast numbers of animals in the\nwild.', 'Predicting on-road abnormalities such as road accidents or traffic violations\nis a challenging task in traffic surveillance. If such predictions can be done\nin advance, many damages can be controlled. Here in our wok, we tried to\nformulate a solution for automated collision prediction in traffic surveillance\nvideos with computer vision and deep networks. It involves object detection,\ntracking, trajectory estimation, and collision prediction. We propose an\nend-to-end collision prediction system, named as COLLIDE-PRED, that\nintelligently integrates the information of past and future trajectories of\nmoving objects to predict collisions in videos. It is a pipeline that starts\nwith object detection, which is used for object tracking, and then trajectory\nprediction is performed which concludes by collision detection. The probable\nplace of collision, and the objects those may cause the collision, both can be\nidentified correctly with COLLIDE-PRED. The proposed method is experimentally\nvalidated with a number of different videos and proves to be effective in\nidentifying accident in advance.', 'In this paper, we propose a novel approach for exploiting structural\nrelations to track multiple objects that may undergo long-term occlusion and\nabrupt motion. We use a model-free approach that relies only on annotations\ngiven in the first frame of the video to track all the objects online, i.e.\nwithout knowledge from future frames. We initialize a probabilistic Attributed\nRelational Graph (ARG) from the first frame, which is incrementally updated\nalong the video. Instead of using the structural information only to evaluate\nthe scene, the proposed approach considers it to generate new tracking\nhypotheses. In this way, our method is capable of generating relevant object\ncandidates that are used to improve or recover the track of lost objects. The\nproposed method is evaluated on several videos of table tennis, volleyball, and\non the ACASVA dataset. The results show that our approach is very robust,\nflexible and able to outperform other state-of-the-art methods in sports videos\nthat present structural patterns.', 'The Discriminative Correlation Filter (CF) uses a circulant convolution\noperation to provide several training samples for the design of a classifier\nthat can distinguish the target from the background. The filter design may be\ninterfered by objects close to the target during the tracking process,\nresulting in tracking failure. This paper proposes an adaptive distraction\ncontext aware tracking algorithm to solve this problem. In the response map\nobtained for the previous frame by the CF algorithm, we adaptively find the\nimage blocks that are similar to the target and use them as negative samples.\nThis diminishes the influence of similar image blocks on the classifier in the\ntracking process and its accuracy is improved. The tracking results on video\nsequences show that the algorithm can cope with rapid changes such as occlusion\nand rotation, and can adaptively use the distractive objects around the target\nas negative samples to improve the accuracy of target tracking.', 'This paper presents the development of a real time tracking algorithm that\nruns on a 1.2 GHz PC/104 computer on-board a small UAV. The algorithm uses zero\nmean normalized cross correlation to detect and locate an object in the image.\nA kalman filter is used to make the tracking algorithm computationally\nefficient. Object position in an image frame is predicted using the motion\nmodel and a search window, centered at the predicted position is generated.\nObject position is updated with the measurement from object detection. The\ndetected position is sent to the motion controller to move the gimbal so that\nthe object stays at the center of the image frame. Detection and tracking is\nautonomously carried out on the payload computer and the system is able to work\nin two different methods. The first method starts detecting and tracking using\na stored image patch. The second method allows the operator on the ground to\nselect the interest object for the UAV to track. The system is capable of\nre-detecting an object, in the event of tracking failure. Performance of the\ntracking system was verified both in the lab and on the field by mounting the\npayload on a vehicle and simulating a flight. Tests show that the system can\ndetect and track a diverse set of objects in real time. Flight testing of the\nsystem will be conducted at the next available opportunity.', 'This paper proposes a novel approach to create an automated visual\nsurveillance system which is very efficient in detecting and tracking moving\nobjects in a video captured by moving camera without any apriori information\nabout the captured scene. Separating foreground from the background is\nchallenging job in videos captured by moving camera as both foreground and\nbackground information change in every consecutive frames of the image\nsequence; thus a pseudo-motion is perceptive in background. In the proposed\nalgorithm, the pseudo-motion in background is estimated and compensated using\nphase correlation of consecutive frames based on the principle of Fourier shift\ntheorem. Then a method is proposed to model an acting background from recent\nhistory of commonality of the current frame and the foreground is detected by\nthe differences between the background model and the current frame. Further\nexploiting the recent history of dissimilarities of the current frame, actual\nmoving objects are detected in the foreground. Next, a two-stepped\nmorphological operation is proposed to refine the object region for an optimum\nobject size. Each object is attributed by its centroid, dimension and three\nhighest peaks of its gray value histogram. Finally, each object is tracked\nusing Kalman filter based on its attributes. The major advantage of this\nalgorithm over most of the existing object detection and tracking algorithms is\nthat, it does not require initialization of object position in the first frame\nor training on sample data to perform. Performance of the algorithm is tested\non benchmark videos containing variable background and very satisfiable results\nis achieved. The performance of the algorithm is also comparable with some of\nthe state-of-the-art algorithms for object detection and tracking.', 'In this paper we propose a novel approach for detecting and tracking objects\nin videos with variable background i.e. videos captured by moving cameras\nwithout any additional sensor. In a video captured by a moving camera, both the\nbackground and foreground are changing in each frame of the image sequence. So\nfor these videos, modeling a single background with traditional background\nmodeling methods is infeasible and thus the detection of actual moving object\nin a variable background is a challenging task. To detect actual moving object\nin this work, spatio-temporal blobs have been generated in each frame by\nspatio-temporal analysis of the image sequence using a three-dimensional Gabor\nfilter. Then individual blobs, which are parts of one object are merged using\nMinimum Spanning Tree to form the moving object in the variable background. The\nheight, width and four-bin gray-value histogram of the object are calculated as\nits features and an object is tracked in each frame using these features to\ngenerate the trajectories of the object through the video sequence. In this\nwork, problem of data association during tracking is solved by Linear\nAssignment Problem and occlusion is handled by the application of kalman\nfilter. The major advantage of our method over most of the existing tracking\nalgorithms is that, the proposed method does not require initialization in the\nfirst frame or training on sample data to perform. Performance of the algorithm\nhas been tested on benchmark videos and very satisfactory result has been\nachieved. The performance of the algorithm is also comparable and superior with\nrespect to some benchmark algorithms.', 'We propose an unsupervised visual tracking method in this paper. Different\nfrom existing approaches using extensive annotated data for supervised\nlearning, our CNN model is trained on large-scale unlabeled videos in an\nunsupervised manner. Our motivation is that a robust tracker should be\neffective in both the forward and backward predictions (i.e., the tracker can\nforward localize the target object in successive frames and backtrace to its\ninitial position in the first frame). We build our framework on a Siamese\ncorrelation filter network, which is trained using unlabeled raw videos.\nMeanwhile, we propose a multiple-frame validation method and a cost-sensitive\nloss to facilitate unsupervised learning. Without bells and whistles, the\nproposed unsupervised tracker achieves the baseline accuracy of fully\nsupervised trackers, which require complete and accurate labels during\ntraining. Furthermore, unsupervised framework exhibits a potential in\nleveraging unlabeled or weakly labeled data to further improve the tracking\naccuracy.', ""The study of animal behavior increasingly relies on (semi-) automatic methods\nfor the extraction of relevant behavioral features from video or picture data.\nTo date, several specialized software products exist to detect and track\nanimals' positions in simple (laboratory) environments. Tracking animals in\ntheir natural environments, however, often requires substantial customization\nof the image processing algorithms to the problem-specific image\ncharacteristics. Here we introduce BioTracker, an open-source computer vision\nframework, that provides programmers with core functionalities that are\nessential parts of a tracking software, such as video I/O, graphics overlays\nand mouse and keyboard interfaces. BioTracker additionally provides a number of\ndifferent tracking algorithms suitable for a variety of image recording\nconditions. The main feature of BioTracker is however the straightforward\nimplementation of new problem-specific tracking modules and vision algorithms\nthat can build upon BioTracker's core functionalities. With this open-source\nframework the scientific community can accelerate their research and focus on\nthe development of new vision algorithms."", 'Wildlife monitoring is crucial to nature conservation and has been done by\nmanual observations from motion-triggered camera traps deployed in the field.\nWidespread adoption of such in-situ sensors has resulted in unprecedented data\nvolumes being collected over the last decade. A significant challenge exists to\nprocess and reliably identify what is in these images efficiently. Advances in\ncomputer vision are poised to provide effective solutions with custom AI models\nbuilt to automatically identify images of interest and label the species in\nthem. Here we outline the data unification effort for the Wildlife Insights\nplatform from various conservation partners, and the challenges involved. Then\nwe present an initial deep convolutional neural network model, trained on 2.9M\nimages across 465 fine-grained species, with a goal to reduce the load on human\nexperts to classify species in images manually. The long-term goal is to enable\nscientists to make conservation recommendations from near real-time analysis of\nspecies abundance and population health.', 'A robust visual tracking system requires an object appearance model that is\nable to handle occlusion, pose, and illumination variations in the video\nstream. This can be difficult to accomplish when the model is trained using\nonly a single image. In this paper, we first propose a tracking approach based\non affine subspaces (constructed from several images) which are able to\naccommodate the abovementioned variations. We use affine subspaces not only to\nrepresent the object, but also the candidate areas that the object may occupy.\nWe furthermore propose a novel approach to measure affine subspace-to-subspace\ndistance via the use of non-Euclidean geometry of Grassmann manifolds. The\ntracking problem is then considered as an inference task in a Markov Chain\nMonte Carlo framework via particle filtering. Quantitative evaluation on\nchallenging video sequences indicates that the proposed approach obtains\nconsiderably better performance than several recent state-of-the-art methods\nsuch as Tracking-Learning-Detection and MILtrack.', 'Multi-object tracking from RGB-D video sequences is a challenging problem due\nto the combination of changing viewpoints, motion, and occlusions over time. We\nobserve that having the complete geometry of objects aids in their tracking,\nand thus propose to jointly infer the complete geometry of objects as well as\ntrack them, for rigidly moving objects over time. Our key insight is that\ninferring the complete geometry of the objects significantly helps in tracking.\nBy hallucinating unseen regions of objects, we can obtain additional\ncorrespondences between the same instance, thus providing robust tracking even\nunder strong change of appearance. From a sequence of RGB-D frames, we detect\nobjects in each frame and learn to predict their complete object geometry as\nwell as a dense correspondence mapping into a canonical space. This allows us\nto derive 6DoF poses for the objects in each frame, along with their\ncorrespondence between frames, providing robust object tracking across the\nRGB-D sequence. Experiments on both synthetic and real-world RGB-D data\ndemonstrate that we achieve state-of-the-art performance on dynamic object\ntracking. Furthermore, we show that our object completion significantly helps\ntracking, providing an improvement of $6.5\\%$ in mean MOTA.', 'This paper presents a review of human activity recognition and behaviour\nunderstanding in video sequence. The key objective of this paper is to provide\na general review on the overall process of a surveillance system used in the\ncurrent trend. Visual surveillance system is directed on automatic\nidentification of events of interest, especially on tracking and classification\nof moving objects. The processing step of the video surveillance system\nincludes the following stages: Surrounding model, object representation, object\ntracking, activity recognition and behaviour understanding. It describes\ntechniques that use to define a general set of activities that are applicable\nto a wide range of scenes and environments in video sequence.', 'The ability to detect and track objects in the visual world is a crucial\nskill for any intelligent agent, as it is a necessary precursor to any\nobject-level reasoning process. Moreover, it is important that agents learn to\ntrack objects without supervision (i.e. without access to annotated training\nvideos) since this will allow agents to begin operating in new environments\nwith minimal human assistance. The task of learning to discover and track\nobjects in videos, which we call \\textit{unsupervised object tracking}, has\ngrown in prominence in recent years; however, most architectures that address\nit still struggle to deal with large scenes containing many objects. In the\ncurrent work, we propose an architecture that scales well to the large-scene,\nmany-object setting by employing spatially invariant computations (convolutions\nand spatial attention) and representations (a spatially local object\nspecification scheme). In a series of experiments, we demonstrate a number of\nattractive features of our architecture; most notably, that it outperforms\ncompeting methods at tracking objects in cluttered scenes with many objects,\nand that it can generalize well to videos that are larger and/or contain more\nobjects than videos encountered during training.', 'Tracking by detection, the dominant approach for online multi-object\ntracking, alternates between localization and association steps. As a result,\nit strongly depends on the quality of instantaneous observations, often failing\nwhen objects are not fully visible. In contrast, tracking in humans is\nunderlined by the notion of object permanence: once an object is recognized, we\nare aware of its physical existence and can approximately localize it even\nunder full occlusions. In this work, we introduce an end-to-end trainable\napproach for joint object detection and tracking that is capable of such\nreasoning. We build on top of the recent CenterTrack architecture, which takes\npairs of frames as input, and extend it to videos of arbitrary length. To this\nend, we augment the model with a spatio-temporal, recurrent memory module,\nallowing it to reason about object locations and identities in the current\nframe using all the previous history. It is, however, not obvious how to train\nsuch an approach. We study this question on a new, large-scale, synthetic\ndataset for multi-object tracking, which provides ground truth annotations for\ninvisible objects, and propose several approaches for supervising tracking\nbehind occlusions. Our model, trained jointly on synthetic and real data,\noutperforms the state of the art on KITTI and MOT17 datasets thanks to its\nrobustness to occlusions.', 'Understanding visual reality involves acquiring common-sense knowledge about\ncountless regularities in the visual world, e.g., how illumination alters the\nappearance of objects in a scene, and how motion changes their apparent spatial\nrelationship. These regularities are hard to label for training supervised\nmachine learning algorithms; consequently, algorithms need to learn these\nregularities from the real world in an unsupervised way. We present a novel\nnetwork meta-architecture that can learn world dynamics from raw, continuous\nvideo. The components of this network can be implemented using any algorithm\nthat possesses three key capabilities: prediction of a signal over time,\nreduction of signal dimensionality (compression), and the ability to use\nsupplementary contextual information to inform the prediction. The presented\narchitecture is highly-parallelized and scalable, and is implemented using\nlocalized connectivity, processing, and learning. We demonstrate an\nimplementation of this architecture where the components are built from\nmulti-layer perceptrons. We apply the implementation to create a system capable\nof stable and robust visual tracking of objects as seen by a moving camera.\nResults show performance on par with or exceeding state-of-the-art tracking\nalgorithms. The tracker can be trained in either fully supervised or\nunsupervised-then-briefly-supervised regimes. Success of the briefly-supervised\nregime suggests that the unsupervised portion of the model extracts useful\ninformation about visual reality. The results suggest a new class of AI\nalgorithms that uniquely combine prediction and scalability in a way that makes\nthem suitable for learning from and --- and eventually acting within --- the\nreal world.', 'The advancement of visual tracking has continuously been brought by deep\nlearning models. Typically, supervised learning is employed to train these\nmodels with expensive labeled data. In order to reduce the workload of manual\nannotations and learn to track arbitrary objects, we propose an unsupervised\nlearning method for visual tracking. The motivation of our unsupervised\nlearning is that a robust tracker should be effective in bidirectional\ntracking. Specifically, the tracker is able to forward localize a target object\nin successive frames and backtrace to its initial position in the first frame.\nBased on such a motivation, in the training process, we measure the consistency\nbetween forward and backward trajectories to learn a robust tracker from\nscratch merely using unlabeled videos. We build our framework on a Siamese\ncorrelation filter network, and propose a multi-frame validation scheme and a\ncost-sensitive loss to facilitate unsupervised learning. Without bells and\nwhistles, the proposed unsupervised tracker achieves the baseline accuracy as\nclassic fully supervised trackers while achieving a real-time speed.\nFurthermore, our unsupervised framework exhibits a potential in leveraging more\nunlabeled or weakly labeled data to further improve the tracking accuracy.', 'Recently, the Kernelized Correlation Filters tracker (KCF) achieved\ncompetitive performance and robustness in visual object tracking. On the other\nhand, visual trackers are not typically used in multiple object tracking. In\nthis paper, we investigate how a robust visual tracker like KCF can improve\nmultiple object tracking. Since KCF is a fast tracker, many can be used in\nparallel and still result in fast tracking. We build a multiple object tracking\nsystem based on KCF and background subtraction. Background subtraction is\napplied to extract moving objects and get their scale and size in combination\nwith KCF outputs, while KCF is used for data association and to handle\nfragmentation and occlusion problems. As a result, KCF and background\nsubtraction help each other to take tracking decision at every frame. Sometimes\nKCF outputs are the most trustworthy (e.g. during occlusion), while in some\nother case, it is the background subtraction outputs. To validate the\neffectiveness of our system, the algorithm is demonstrated on four urban video\nrecordings from a standard dataset. Results show that our method is competitive\nwith state-of-the-art trackers even if we use a much simpler data association\nstep.', 'Recent interest in self-supervised dense tracking has yielded rapid progress,\nbut performance still remains far from supervised methods. We propose a dense\ntracking model trained on videos without any annotations that surpasses\nprevious self-supervised methods on existing benchmarks by a significant margin\n(+15%), and achieves performance comparable to supervised methods. In this\npaper, we first reassess the traditional choices used for self-supervised\ntraining and reconstruction loss by conducting thorough experiments that\nfinally elucidate the optimal choices. Second, we further improve on existing\nmethods by augmenting our architecture with a crucial memory component. Third,\nwe benchmark on large-scale semi-supervised video object segmentation(aka.\ndense tracking), and propose a new metric: generalizability. Our first two\ncontributions yield a self-supervised network that for the first time is\ncompetitive with supervised methods on standard evaluation metrics of dense\ntracking. When measuring generalizability, we show self-supervised approaches\nare actually superior to the majority of supervised methods. We believe this\nnew generalizability metric can better capture the real-world use-cases for\ndense tracking, and will spur new interest in this research direction.', ""Visual animal biometrics is rapidly gaining popularity as it enables a\nnon-invasive and cost-effective approach for wildlife monitoring applications.\nWidespread usage of camera traps has led to large volumes of collected images,\nmaking manual processing of visual content hard to manage. In this work, we\ndevelop a framework for automatic detection and recognition of individuals in\ndifferent patterned species like tigers, zebras and jaguars. Most existing\nsystems primarily rely on manual input for localizing the animal, which does\nnot scale well to large datasets. In order to automate the detection process\nwhile retaining robustness to blur, partial occlusion, illumination and pose\nvariations, we use the recently proposed Faster-RCNN object detection framework\nto efficiently detect animals in images. We further extract features from\nAlexNet of the animal's flank and train a logistic regression (or Linear SVM)\nclassifier to recognize the individuals. We primarily test and evaluate our\nframework on a camera trap tiger image dataset that contains images that vary\nin overall image quality, animal pose, scale and lighting. We also evaluate our\nrecognition system on zebra and jaguar images to show generalization to other\npatterned species. Our framework gives perfect detection results in camera\ntrapped tiger images and a similar or better individual recognition performance\nwhen compared with state-of-the-art recognition techniques."", 'Anomaly detection in surveillance videos is currently a challenge because of\nthe diversity of possible events. We propose a deep convolutional neural\nnetwork (CNN) that addresses this problem by learning a correspondence between\ncommon object appearances (e.g. pedestrian, background, tree, etc.) and their\nassociated motions. Our model is designed as a combination of a reconstruction\nnetwork and an image translation model that share the same encoder. The former\nsub-network determines the most significant structures that appear in video\nframes and the latter one attempts to associate motion templates to such\nstructures. The training stage is performed using only videos of normal events\nand the model is then capable to estimate frame-level scores for an unknown\ninput. The experiments on 6 benchmark datasets demonstrate the competitive\nperformance of the proposed approach with respect to state-of-the-art methods.', 'Quantifying behavior is crucial for many applications in neuroscience.\nVideography provides easy methods for the observation and recording of animal\nbehavior in diverse settings, yet extracting particular aspects of a behavior\nfor further analysis can be highly time consuming. In motor control studies,\nhumans or other animals are often marked with reflective markers to assist with\ncomputer-based tracking, yet markers are intrusive (especially for smaller\nanimals), and the number and location of the markers must be determined a\npriori. Here, we present a highly efficient method for markerless tracking\nbased on transfer learning with deep neural networks that achieves excellent\nresults with minimal training data. We demonstrate the versatility of this\nframework by tracking various body parts in a broad collection of experimental\nsettings: mice odor trail-tracking, egg-laying behavior in drosophila, and\nmouse hand articulation in a skilled forelimb task. For example, during the\nskilled reaching behavior, individual joints can be automatically tracked (and\na confidence score is reported). Remarkably, even when a small number of frames\nare labeled ($\\approx 200$), the algorithm achieves excellent tracking\nperformance on test frames that is comparable to human accuracy.', ""Egocentric vision is an emerging field of computer vision that is\ncharacterized by the acquisition of images and video from the first person\nperspective. In this paper we address the challenge of egocentric human action\nrecognition by utilizing the presence and position of detected regions of\ninterest in the scene explicitly, without further use of visual features.\n  Initially, we recognize that human hands are essential in the execution of\nactions and focus on obtaining their movements as the principal cues that\ndefine actions. We employ object detection and region tracking techniques to\nlocate hands and capture their movements. Prior knowledge about egocentric\nviews facilitates hand identification between left and right. With regard to\ndetection and tracking, we contribute a pipeline that successfully operates on\nunseen egocentric videos to find the camera wearer's hands and associate them\nthrough time. Moreover, we emphasize on the value of scene information for\naction recognition. We acknowledge that the presence of objects is significant\nfor the execution of actions by humans and in general for the description of a\nscene. To acquire this information, we utilize object detection for specific\nclasses that are relevant to the actions we want to recognize.\n  Our experiments are targeted on videos of kitchen activities from the\nEpic-Kitchens dataset. We model action recognition as a sequence learning\nproblem of the detected spatial positions in the frames. Our results show that\nexplicit hand and object detections with no other visual information can be\nrelied upon to classify hand-related human actions. Testing against methods\nfully dependent on visual features, signals that for actions where hand motions\nare conceptually important, a region-of-interest-based description of a video\ncontains equally expressive information with comparable classification\nperformance."", 'Visual tracking addresses the problem of identifying and localizing an\nunknown target in a video given the target specified by a bounding box in the\nfirst frame. In this paper, we propose a dual network to better utilize\nfeatures among layers for visual tracking. It is observed that features in\nhigher layers encode semantic context while its counterparts in lower layers\nare sensitive to discriminative appearance. Thus we exploit the hierarchical\nfeatures in different layers of a deep model and design a dual structure to\nobtain better feature representation from various streams, which is rarely\ninvestigated in previous work. To highlight geometric contours of the target,\nwe integrate the hierarchical feature maps with an edge detector as the coarse\nprior maps to further embed local details around the target. To leverage the\nrobustness of our dual network, we train it with random patches measuring the\nsimilarities between the network activation and target appearance, which serves\nas a regularization to enforce the dual network to focus on target object. The\nproposed dual network is updated online in a unique manner based on the\nobservation that the target being tracked in consecutive frames should share\nmore similar feature representations than those in the surrounding background.\nIt is also found that for a target object, the prior maps can help further\nenhance performance by passing message into the output maps of the dual\nnetwork. Therefore, an independent component analysis with reference algorithm\n(ICA-R) is employed to extract target context using prior maps as guidance.\nOnline tracking is conducted by maximizing the posterior estimate on the final\nmaps with stochastic and periodic update. Quantitative and qualitative\nevaluations on two large-scale benchmark data sets show that the proposed\nalgorithm performs favourably against the state-of-the-arts.', 'In this paper, we present a new method for detecting road users in an urban\nenvironment which leads to an improvement in multiple object tracking. Our\nmethod takes as an input a foreground image and improves the object detection\nand segmentation. This new image can be used as an input to trackers that use\nforeground blobs from background subtraction. The first step is to create\nforeground images for all the frames in an urban video. Then, starting from the\noriginal blobs of the foreground image, we merge the blobs that are close to\none another and that have similar optical flow. The next step is extracting the\nedges of the different objects to detect multiple objects that might be very\nclose (and be merged in the same blob) and to adjust the size of the original\nblobs. At the same time, we use the optical flow to detect occlusion of objects\nthat are moving in opposite directions. Finally, we make a decision on which\ninformation we keep in order to construct a new foreground image with blobs\nthat can be used for tracking. The system is validated on four videos of an\nurban traffic dataset. Our method improves the recall and precision metrics for\nthe object detection task compared to the vanilla background subtraction method\nand improves the CLEAR MOT metrics in the tracking tasks for most videos.', 'In this paper, we propose and study a novel visual object tracking approach\nbased on convolutional networks and recurrent networks. The proposed approach\nis distinct from the existing approaches to visual object tracking, such as\nfiltering-based ones and tracking-by-detection ones, in the sense that the\ntracking system is explicitly trained off-line to track anonymous objects in a\nnoisy environment. The proposed visual tracking model is end-to-end trainable,\nminimizing any adversarial effect from mismatches in object representation and\nbetween the true underlying dynamics and learning dynamics. We empirically show\nthat the proposed tracking approach works well in various scenarios by\ngenerating artificial video sequences with varying conditions; the number of\nobjects, amount of noise and the match between the training shapes and test\nshapes.', 'In this paper, the task of video panoptic segmentation is studied and two\ndifferent methods to solve the task will be proposed. Video panoptic\nsegmentation (VPS) is a recently introduced computer vision task that requires\nclassifying and tracking every pixel in a given video. The nature of this task\nmakes the cost of annotating datasets for it prohibiting. To understand video\npanoptic segmentation, first, earlier introduced constituent tasks that focus\non semantics and tracking separately will be researched. Thereafter, two\ndata-driven approaches which do not require training on a tailored VPS dataset\nwill be selected to solve it. The first approach will show how a model for\nvideo panoptic segmentation can be built by heuristically fusing the outputs of\na pre-trained semantic segmentation model and a pre-trained multi-object\ntracking model. This can be desired if one wants to easily extend the\ncapabilities of either model. The second approach will counter some of the\nshortcomings of the first approach by building on top of a shared neural\nnetwork backbone with task-specific heads. This network is designed for\npanoptic segmentation and will be extended by a mask propagation module to link\ninstance masks across time, yielding the video panoptic segmentation format.', 'We propose an end-to-end learning framework for segmenting generic objects in\nvideos. Our method learns to combine appearance and motion information to\nproduce pixel level segmentation masks for all prominent objects in videos. We\nformulate this task as a structured prediction problem and design a two-stream\nfully convolutional neural network which fuses together motion and appearance\nin a unified framework. Since large-scale video datasets with pixel level\nsegmentations are problematic, we show how to bootstrap weakly annotated videos\ntogether with existing image recognition datasets for training. Through\nexperiments on three challenging video segmentation benchmarks, our method\nsubstantially improves the state-of-the-art for segmenting generic (unseen)\nobjects. Code and pre-trained models are available on the project website.', 'We present the first systematic study on concealed object detection (COD),\nwhich aims to identify objects that are ""perfectly"" embedded in their\nbackground. The high intrinsic similarities between the concealed objects and\ntheir background make COD far more challenging than traditional object\ndetection/segmentation. To better understand this task, we collect a\nlarge-scale dataset, called COD10K, which consists of 10,000 images covering\nconcealed objects in diverse real-world scenarios from 78 object categories.\nFurther, we provide rich annotations including object categories, object\nboundaries, challenging attributes, object-level labels, and instance-level\nannotations. Our COD10K is the largest COD dataset to date, with the richest\nannotations, which enables comprehensive concealed object understanding and can\neven be used to help progress several other vision tasks, such as detection,\nsegmentation, classification, etc. Motivated by how animals hunt in the wild,\nwe also design a simple but strong baseline for COD, termed the Search\nIdentification Network (SINet). Without any bells and whistles, SINet\noutperforms 12 cutting-edge baselines on all datasets tested, making them\nrobust, general architectures that could serve as catalysts for future research\nin COD. Finally, we provide some interesting findings and highlight several\npotential applications and future directions. To spark research in this new\nfield, our code, dataset, and online demo are available on our project page:\nhttp://mmcheng.net/cod.', 'The world is composed of objects, the ground, and the sky. Visual perception\nof objects requires solving two fundamental challenges: segmenting visual input\ninto discrete units, and tracking identities of these units despite appearance\nchanges due to object deformation, changing perspective, and dynamic occlusion.\nCurrent computer vision approaches to segmentation and tracking that approach\nhuman performance all require learning, raising the question: can objects be\nsegmented and tracked without learning? Here, we show that the mathematical\nstructure of light rays reflected from environment surfaces yields a natural\nrepresentation of persistent surfaces, and this surface representation provides\na solution to both the segmentation and tracking problems. We describe how to\ngenerate this surface representation from continuous visual input, and\ndemonstrate that our approach can segment and invariantly track objects in\ncluttered synthetic video despite severe appearance changes, without requiring\nlearning.', ""This paper introduces an automated vision system for animal detection in\ntrail-camera images taken from a field under the administration of the Texas\nParks and Wildlife Department. As traditional wildlife counting techniques are\nintrusive and labor intensive to conduct, trail-camera imaging is a\ncomparatively non-intrusive method for capturing wildlife activity. However,\ngiven the large volume of images produced from trail-cameras, manual analysis\nof the images remains time-consuming and inefficient. We implemented a\ntwo-stage deep convolutional neural network pipeline to find animal-containing\nimages in the first stage and then process these images to detect birds in the\nsecond stage. The animal classification system classifies animal images with\noverall 93% sensitivity and 96% specificity. The bird detection system achieves\nbetter than 93% sensitivity, 92% specificity, and 68% average\nIntersection-over-Union rate. The entire pipeline processes an image in less\nthan 0.5 seconds as opposed to an average 30 seconds for a human labeler. We\nalso addressed post-deployment issues related to data drift for the animal\nclassification system as image features vary with seasonal changes. This system\nutilizes an automatic retraining algorithm to detect data drift and update the\nsystem. We introduce a novel technique for detecting drifted images and\ntriggering the retraining procedure. Two statistical experiments are also\npresented to explain the prediction behavior of the animal classification\nsystem. These experiments investigate the cues that steers the system towards a\nparticular decision. Statistical hypothesis testing demonstrates that the\npresence of an animal in the input image significantly contributes to the\nsystem's decisions."", 'This paper addresses the problem of tracking moving objects of variable\nappearance in challenging scenes rich with features and texture. Reliable\ntracking is of pivotal importance in surveillance applications. It is made\nparticularly difficult by the nature of objects encountered in such scenes:\nthese too change in appearance and scale, and are often articulated (e.g.\nhumans). We propose a method which uses fast motion detection and segmentation\nas a constraint for both building appearance models and their robust\npropagation (matching) in time. The appearance model is based on sets of local\nappearances automatically clustered using spatio-kinetic similarity, and is\nupdated with each new appearance seen. This integration of all seen appearances\nof a tracked object makes it extremely resilient to errors caused by occlusion\nand the lack of permanence of due to low data quality, appearance change or\nbackground clutter. These theoretical strengths of our algorithm are\nempirically demonstrated on two hour long video footage of a busy city\nmarketplace.', 'From video streaming to security and surveillance applications, video data\nplay an important role in our daily living today. However, managing a large\namount of video data and retrieving the most useful information for the user\nremain a challenging task. In this paper, we propose a novel video\nclassification system that would benefit the scene understanding task. We\ndefine our classification problem as classifying background and foreground\nmotions using the same feature representation for outdoor scenes. This means\nthat the feature representation needs to be robust enough and adaptable to\ndifferent classification tasks. We propose a lightweight Loss Switching Fusion\nNetwork (LSFNet) for the fusion of spatiotemporal descriptors and a similarity\nsearch scheme with soft voting to boost the classification performance. The\nproposed system has a variety of potential applications such as content-based\nvideo clustering, video filtering, etc. Evaluation results on two private\nindustry datasets show that our system is robust in both classifying different\nbackground motions and detecting human motions from these background motions.', 'In this paper, an online adaptive model-free tracker is proposed to track\nsingle objects in video sequences to deal with real-world tracking challenges\nlike low-resolution, object deformation, occlusion and motion blur. The novelty\nlies in the construction of a strong appearance model that captures features\nfrom the initialized bounding box and then are assembled into anchor-point\nfeatures. These features memorize the global pattern of the object and have an\ninternal star graph-like structure. These features are unique and flexible and\nhelps tracking generic and deformable objects with no limitation on specific\nobjects. In addition, the relevance of each feature is evaluated online using\nshort-term consistency and long-term consistency. These parameters are adapted\nto retain consistent features that vote for the object location and that deal\nwith outliers for long-term tracking scenarios. Additionally, voting in a\nGaussian manner helps in tackling inherent noise of the tracking system and in\naccurate object localization. Furthermore, the proposed tracker uses pairwise\ndistance measure to cope with scale variations and combines pixel-level binary\nfeatures and global weighted color features for model update. Finally,\nexperimental results on a visual tracking benchmark dataset are presented to\ndemonstrate the effectiveness and competitiveness of the proposed tracker.', 'In this paper we propose a system capable of tracking multiple soccer players\nin different types of video quality. The main goal, in contrast to most\nstate-of-art soccer player tracking systems, is the ability of execute\neffectively tracking in videos of low-quality. We adapted a state-of-art\nMultiple Object Tracking to the task. In order to do that adaptation, we\ncreated a Detection and a Tracking Dataset for 3 different qualities of video.\nThe results of our system are conclusive of its high performance.', 'Tracking moving objects from a video sequence requires segmentation of these\nobjects from the background image. However, getting the actual background image\nautomatically without object detection and using only the video is difficult.\nIn this paper, we describe a novel algorithm that generates background from\nreal world images without foreground detection. The algorithm assumes that the\nbackground image is shown in the majority of the video. Given this simple\nassumption, the method described in this paper is able to accurately generate,\nwith high probability, the background image from a video using only a small\nnumber of binary operations.', 'Class-agnostic object tracking is particularly difficult in cluttered\nenvironments as target specific discriminative models cannot be learned a\npriori. Inspired by how the human visual cortex employs spatial attention and\nseparate ""where"" and ""what"" processing pathways to actively suppress irrelevant\nvisual features, this work develops a hierarchical attentive recurrent model\nfor single object tracking in videos. The first layer of attention discards the\nmajority of background by selecting a region containing the object of interest,\nwhile the subsequent layers tune in on visual features particular to the\ntracked object. This framework is fully differentiable and can be trained in a\npurely data driven fashion by gradient methods. To improve training\nconvergence, we augment the loss function with terms for a number of auxiliary\ntasks relevant for tracking. Evaluation of the proposed model is performed on\ntwo datasets: pedestrian tracking on the KTH activity recognition dataset and\nthe more difficult KITTI object tracking dataset.', ""Being heavily reliant on animals, it is our ethical obligation to improve\ntheir well-being by understanding their needs. Several studies show that animal\nneeds are often expressed through their faces. Though remarkable progress has\nbeen made towards the automatic understanding of human faces, this has\nregrettably not been the case with animal faces. There exists significant room\nand appropriate need to develop automatic systems capable of interpreting\nanimal faces. Among many transformative impacts, such a technology will foster\nbetter and cheaper animal healthcare, and further advance animal psychology\nunderstanding.\n  We believe the underlying research progress is mainly obstructed by the lack\nof an adequately annotated dataset of animal faces, covering a wide spectrum of\nanimal species. To this end, we introduce a large-scale, hierarchical annotated\ndataset of animal faces, featuring 21.9K faces from 334 diverse species and 21\nanimal orders across biological taxonomy. These faces are captured\n`in-the-wild' conditions and are consistently annotated with 9 landmarks on key\nfacial features. The proposed dataset is structured and scalable by design; its\ndevelopment underwent four systematic stages involving rigorous, manual\nannotation effort of over 6K man-hours. We benchmark it for face alignment\nusing the existing art under novel problem settings. Results showcase its\nchallenging nature, unique attributes and present definite prospects for novel,\nadaptive, and generalized face-oriented CV algorithms. We further benchmark the\ndataset for face detection and fine-grained recognition tasks, to demonstrate\nmulti-task applications and room for improvement. Experiments indicate that\nthis dataset will push the algorithmic advancements across many related CV\ntasks and encourage the development of novel systems for animal facial\nbehaviour monitoring. We will make the dataset publicly available."", 'Photographs of wild animals in their natural habitats can be recorded\nunobtrusively via cameras that are triggered by motion nearby. The installation\nof such camera traps is becoming increasingly common across the world. Although\nthis is a convenient source of invaluable data for biologists, ecologists and\nconservationists, the arduous task of poring through potentially millions of\npictures each season introduces prohibitive costs and frustrating delays. We\ndevelop automatic algorithms that are able to detect animals, identify the\nspecies of animals and to recognize individual animals for two species. we\npropose the first fully-automatic tool that can recognize specific individuals\nof leopard and tiger due to their characteristic body markings. We adopt a\nclass of supervised learning approach of machine learning where a Deep\nConvolutional Neural Network (DCNN) is trained using several instances of\nmanually-labelled images for each of the three classification tasks. We\ndemonstrate the effectiveness of our approach on a data set of camera-trap\nimages recorded in the jungles of Southern India.', 'We propose the first multi-frame video object detection framework trained to\ndetect great apes. It is applicable to challenging camera trap footage in\ncomplex jungle environments and extends a traditional feature pyramid\narchitecture by adding self-attention driven feature blending in both the\nspatial as well as the temporal domain. We demonstrate that this extension can\ndetect distinctive species appearance and motion signatures despite significant\npartial occlusion. We evaluate the framework using 500 camera trap videos of\ngreat apes from the Pan African Programme containing 180K frames, which we\nmanually annotated with accurate per-frame animal bounding boxes. These clips\ncontain significant partial occlusions, challenging lighting, dynamic\nbackgrounds, and natural camouflage effects. We show that our approach performs\nhighly robustly and significantly outperforms frame-based detectors. We also\nperform detailed ablation studies and validation on the full ILSVRC 2015 VID\ndata corpus to demonstrate wider applicability at adequate performance levels.\nWe conclude that the framework is ready to assist human camera trap inspection\nefforts. We publish code, weights, and ground truth annotations with this\npaper.', 'Animal pose estimation has recently come into the limelight due to its\napplication in biology, zoology, and aquaculture. Deep learning methods have\neffectively been applied to human pose estimation. However, the major\nbottleneck to the application of these methods to animal pose estimation is the\nunavailability of sufficient quantities of labeled data. Though there are ample\nquantities of unlabelled data publicly available, it is economically\nimpractical to label large quantities of data for each animal. In addition, due\nto the wide variety of body shapes in the animal kingdom, the transfer of\nknowledge across domains is ineffective. Given the fact that the human brain is\nable to recognize animal pose without requiring large amounts of labeled data,\nit is only reasonable that we exploit unsupervised learning to tackle the\nproblem of animal pose recognition from the available, unlabelled data. In this\npaper, we introduce a novel architecture that is able to recognize the pose of\nmultiple animals fromunlabelled data. We do this by (1) removing background\ninformation from each image and employing an edge detection algorithm on the\nbody of the animal, (2) Tracking motion of the edge pixels and performing\nagglomerative clustering to segment body parts, (3) employing contrastive\nlearning to discourage grouping of distant body parts together. Hence we are\nable to distinguish between body parts of the animal, based on their visual\nbehavior, instead of the underlying anatomy. Thus, we are able to achieve a\nmore effective classification of the data than their human-labeled\ncounterparts. We test our model on the TigDog and WLD (WildLife Documentary)\ndatasets, where we outperform state-of-the-art approaches by a significant\nmargin. We also study the performance of our model on other public data to\ndemonstrate the generalization ability of our model.', 'Satellite video cameras can provide continuous observation for a large-scale\narea, which is important for many remote sensing applications. However,\nachieving moving object detection and tracking in satellite videos remains\nchallenging due to the insufficient appearance information of objects and lack\nof high-quality datasets. In this paper, we first build a large-scale satellite\nvideo dataset with rich annotations for the task of moving object detection and\ntracking. This dataset is collected by the Jilin-1 satellite constellation and\ncomposed of 47 high-quality videos with 1,646,038 instances of interest for\nobject detection and 3,711 trajectories for object tracking. We then introduce\na motion modeling baseline to improve the detection rate and reduce false\nalarms based on accumulative multi-frame differencing and robust matrix\ncompletion. Finally, we establish the first public benchmark for moving object\ndetection and tracking in satellite videos, and extensively evaluate the\nperformance of several representative approaches on our dataset. Comprehensive\nexperimental analyses and insightful conclusions are also provided. The dataset\nis available at https://github.com/QingyongHu/VISO.', 'The problem of tracking multiple objects in a video sequence poses several\nchallenging tasks. For tracking-by-detection, these include object\nre-identification, motion prediction and dealing with occlusions. We present a\ntracker (without bells and whistles) that accomplishes tracking without\nspecifically targeting any of these tasks, in particular, we perform no\ntraining or optimization on tracking data. To this end, we exploit the bounding\nbox regression of an object detector to predict the position of an object in\nthe next frame, thereby converting a detector into a Tracktor. We demonstrate\nthe potential of Tracktor and provide a new state-of-the-art on three\nmulti-object tracking benchmarks by extending it with a straightforward\nre-identification and camera motion compensation. We then perform an analysis\non the performance and failure cases of several state-of-the-art tracking\nmethods in comparison to our Tracktor. Surprisingly, none of the dedicated\ntracking methods are considerably better in dealing with complex tracking\nscenarios, namely, small and occluded objects or missing detections. However,\nour approach tackles most of the easy tracking scenarios. Therefore, we\nmotivate our approach as a new tracking paradigm and point out promising future\nresearch directions. Overall, Tracktor yields superior tracking performance\nthan any current tracking method and our analysis exposes remaining and\nunsolved tracking challenges to inspire future research directions.', 'Target tracking in hyperspectral videos is a new research topic. In this\npaper, a novel method based on convolutional network and Kernelized Correlation\nFilter (KCF) framework is presented for tracking objects of interest in\nhyperspectral videos. We extract a set of normalized three-dimensional cubes\nfrom the target region as fixed convolution filters which contain spectral\ninformation surrounding a target. The feature maps generated by convolutional\noperations are combined to form a three-dimensional representation of an\nobject, thereby providing effective encoding of local spectral-spatial\ninformation. We show that a simple two-layer convolutional networks is\nsufficient to learn robust representations without the need of offline training\nwith a large dataset. In the tracking step, KCF is adopted to distinguish\ntargets from neighboring environment. Experimental results demonstrate that the\nproposed method performs well on sample hyperspectral videos, and outperforms\nseveral state-of-the-art methods tested on grayscale and color videos in the\nsame scene.', 'Understanding the abundance of a species is the first step towards\nunderstanding both its long-term sustainability and the impact that we may be\nhaving upon it. Ecologists use camera traps to remotely survey for the presence\nof specific animal species. Previous studies have shown that deep learning\nmodels can be trained to automatically detect and classify animals within\ncamera trap imagery with high levels of confidence. However, the ability to\ntrain these models is reliant upon having enough high-quality training data.\nWhat happens when the animal is rare or the data sets are non-existent? This\nresearch proposes an approach of using images of rare animals in captivity\n(focusing on the Scottish wildcat) to generate the training dataset. We explore\nthe challenges associated with generalising a model trained on captivity data\nwhen applied to data collected in the wild. The research is contextualised by\nthe needs of ecologists in planning/engineering. Following precedents from\nother research, this project establishes an ensemble for object detection,\nimage segmentation and image classification models which are then tested using\ndifferent image manipulation and class structuring techniques to encourage\nmodel generalisation. The research concludes, in the context of Scottish\nwildcat, that models trained on captivity imagery cannot be generalised to wild\ncamera trap imagery using existing techniques. However, final model\nperformances based on a two-class model Wildcat vs Not Wildcat achieved an\noverall accuracy score of 81.6% and Wildcat accuracy score of 54.8% on a test\nset in which only 1% of images contained a wildcat. This suggests using\ncaptivity images is feasible with further research. This is the first research\nwhich attempts to generate a training set based on captivity data and the first\nto explore the development of such models in the context of ecologists in\nplanning/engineering.', 'By assuming that the most frequently occuring color in a video or a region of\na video I propose a new algorithm for detecting foreground objects in a video.\nThe process of detecting the foreground objects is complicated because of the\nfact that there may be swaying trees, objects of the background being moved\naround or lighting changes in the video. To deal with such complexities many\nhave come up with solutions which heavily rely on expensive floating point\noperations. In this paper I used a data structure called Octree which is\nimplemented only using binary operations. Traditionally octrees were used for\ncolor quantization but here in this paper I used it as a data structure to\nstore the most frequently occuring colors in a video as well. For each of the\nstarting few video frames, I constructed a Octree using all the colors of that\nframe. Next I pruned all the trees by removing nodes below a certain height and\ngave the leaf nodes a color which is dependant on the topological path from\nthat node to its parent. Hence any two leaf nodes in two different octrees with\nthe same topological path from themselves to the root will represent the same\ncolor. Next I merged all these individual trees into a single tree retaining\nonly those nodes whose topological path to itself from the root is most common\namong all the trees. The colors represented by the leaf nodes in the resultant\ntree will be the most frequently occuring colors in the starting video frames\nof the video. Hence any color of an incomming frame that is not close to any of\nthe colors represented by the leaf node of the merged tree can be regarded as\nbelonging to a foreground object.\n  As an Octree is constructed using only binary operations, it is very fast\ncompared to other leading algorithms.', 'Motion models play a great role in visual tracking applications for\npredicting the possible locations of objects in the next frame. Unlike target\ntracking in radar or aerospace domain which considers only points, object\ntracking in computer vision involves sizes of objects. Constant velocity motion\nmodel is the most widely used motion model for visual tracking, however, there\nis no clear and understandable derivation involving sizes of objects specially\nfor new researchers joining this research field. In this document, we derive\nthe constant velocity motion model that incorporates sizes of objects that, we\nthink, can help the new researchers to adapt to it very quickly.', 'Object tracking is one of the most challenging task and has secured\nsignificant attention of computer vision researchers in the past two decades.\nRecent deep learning based trackers have shown good performance on various\ntracking challenges. A tracking method should track objects in sequential\nframes accurately in challenges such as deformation, low resolution, occlusion,\nscale and light variations. Most trackers achieve good performance on specific\nchallenges instead of all tracking problems, hence there is a lack of general\npurpose tracking algorithms that can perform well in all conditions. Moreover,\nperformance of tracking techniques has not been evaluated in noisy\nenvironments. Visual object tracking has real world applications and there is\ngood chance that noise may get added during image acquisition in surveillance\ncameras. We aim to study the robustness of two state of the art trackers in the\npresence of noise including Efficient Convolutional Operators (ECO) and\nCorrelation Filter Network (CFNet). Our study demonstrates that the performance\nof these trackers degrades as the noise level increases, which demonstrate the\nneed to design more robust tracking algorithms.', ""This paper investigates long-term face tracking of a specific person given\nhis/her face image in a single frame as a query in a video stream. Through\ntaking advantage of pre-trained deep learning models on big data, a novel\nsystem is developed for accurate video face tracking in the unconstrained\nenvironments depicting various people and objects moving in and out of the\nframe. In the proposed system, we present a detection-verification-tracking\nmethod (dubbed as 'DVT') which accomplishes the long-term face tracking task\nthrough the collaboration of face detection, face verification, and\n(short-term) face tracking. An offline trained detector based on cascaded\nconvolutional neural networks localizes all faces appeared in the frames, and\nan offline trained face verifier based on deep convolutional neural networks\nand similarity metric learning decides if any face or which face corresponds to\nthe queried person. An online trained tracker follows the face from frame to\nframe. When validated on a sitcom episode and a TV show, the DVT method\noutperforms tracking-learning-detection (TLD) and face-TLD in terms of recall\nand precision. The proposed system is also tested on many other types of videos\nand shows very promising results."", 'For multi-target tracking, target representation plays a crucial rule in\nperformance. State-of-the-art approaches rely on the deep learning-based visual\nrepresentation that gives an optimal performance at the cost of high\ncomputational complexity. In this paper, we come up with a simple yet effective\ntarget representation for human tracking. Our inspiration comes from the fact\nthat the human body goes through severe deformation and inter/intra occlusion\nover the passage of time. So, instead of tracking the whole body part, a\nrelative rigid organ tracking is selected for tracking the human over an\nextended period of time. Hence, we followed the tracking-by-detection paradigm\nand generated the target hypothesis of only the spatial locations of heads in\nevery frame. After the localization of head location, a Kalman filter with a\nconstant velocity motion model is instantiated for each target that follows the\ntemporal evolution of the targets in the scene. For associating the targets in\nthe consecutive frames, combinatorial optimization is used that associates the\ncorresponding targets in a greedy fashion. Qualitative results are evaluated on\nfour challenging video surveillance dataset and promising results has been\nachieved.', 'The primary aim of this project is to build a contextual Question-Answering\nmodel for videos. The current methodologies provide a robust model for image\nbased Question-Answering, but we are aim to generalize this approach to be\nvideos. We propose a graphical representation of video which is able to handle\nseveral types of queries across the whole video. For example, if a frame has an\nimage of a man and a cat sitting, it should be able to handle queries like,\nwhere is the cat sitting with respect to the man? or ,what is the man holding\nin his hand?. It should be able to answer queries relating to temporal\nrelationships also.', 'In the field of bioimaging, an important part of analyzing the motion of\nobjects is tracking. We propose a method that applies the Sinkhorn distance for\nsolving the optimal transport problem to track objects. The advantage of this\nmethod is that it can flexibly incorporate various assumptions in tracking as a\ncost matrix. First, we extend the Sinkhorn distance from two dimensions to\nthree dimensions. Using this three-dimensional distance, we compare the\nperformance of two types of tracking technique, namely tracking that associates\nobjects that are close to each other, which conventionally uses the\nnearest-neighbor method, and tracking that assumes that the object is moving at\nconstant velocity, using three types of simulation data. The results suggest\nthat when tracking objects moving at constant velocity, our method is superior\nto conventional nearest-neighbor tracking as long as the added noise is not\nexcessively large. We show that the Sinkhorn method can be applied effectively\nto object tracking. Our simulation data analysis suggests that when objects are\nmoving at constant velocity, our method, which sets acceleration as a cost,\noutperforms the traditional nearest-neighbor method in terms of tracking\nobjects. To apply the proposed method to real bioimaging data, it is necessary\nto set an appropriate cost indicator based on the movement features.', 'Camera trapping is increasingly used to monitor wildlife, but this technology\ntypically requires extensive data annotation. Recently, deep learning has\nsignificantly advanced automatic wildlife recognition. However, current methods\nare hampered by a dependence on large static data sets when wildlife data is\nintrinsically dynamic and involves long-tailed distributions. These two\ndrawbacks can be overcome through a hybrid combination of machine learning and\nhumans in the loop. Our proposed iterative human and automated identification\napproach is capable of learning from wildlife imagery data with a long-tailed\ndistribution. Additionally, it includes self-updating learning that facilitates\ncapturing the community dynamics of rapidly changing natural systems. Extensive\nexperiments show that our approach can achieve a ~90% accuracy employing only\n~20% of the human annotations of existing approaches. Our synergistic\ncollaboration of humans and machines transforms deep learning from a relatively\ninefficient post-annotation tool to a collaborative on-going annotation tool\nthat vastly relieves the burden of human annotation and enables efficient and\nconstant model updates.', 'Moving object detection and tracking have various applications, including\nsurveillance, anomaly detection, vehicle navigation, etc. The literature on\nobject detection and tracking is rich enough, and several essential survey\npapers exist. However, the research on camouflage object detection and tracking\nlimited due to the complexity of the problem. Existing work on this problem has\nbeen done based on either biological characteristics of the camouflaged objects\nor computer vision techniques. In this article, we review the existing\ncamouflaged object detection and tracking techniques using computer vision\nalgorithms from the theoretical point of view. This article also addresses\nseveral issues of interest as well as future research direction on this area.\nWe hope this review will help the reader to learn the recent advances in\ncamouflaged object detection and tracking.', 'Animals have evolved highly functional visual systems to understand motion,\nassisting perception even under complex environments. In this paper, we work\ntowards developing a computer vision system able to segment objects by\nexploiting motion cues, i.e. motion segmentation. We make the following\ncontributions: First, we introduce a simple variant of the Transformer to\nsegment optical flow frames into primary objects and the background. Second, we\ntrain the architecture in a self-supervised manner, i.e. without using any\nmanual annotations. Third, we analyze several critical components of our method\nand conduct thorough ablation studies to validate their necessity. Fourth, we\nevaluate the proposed architecture on public benchmarks (DAVIS2016, SegTrackv2,\nand FBMS59). Despite using only optical flow as input, our approach achieves\nsuperior or comparable results to previous state-of-the-art self-supervised\nmethods, while being an order of magnitude faster. We additionally evaluate on\na challenging camouflage dataset (MoCA), significantly outperforming the other\nself-supervised approaches, and comparing favourably to the top supervised\napproach, highlighting the importance of motion cues, and the potential bias\ntowards visual appearance in existing video segmentation models.', ""The robust detection of small targets against cluttered background is\nimportant for future artificial visual systems in searching and tracking\napplications. The insects' visual systems have demonstrated excellent ability\nto avoid predators, find prey or identify conspecifics - which always appear as\nsmall dim speckles in the visual field. Build a computational model of the\ninsects' visual pathways could provide effective solutions to detect small\nmoving targets. Although a few visual system models have been proposed, they\nonly make use of small-field visual features for motion detection and their\ndetection results often contain a number of false positives. To address this\nissue, we develop a new visual system model for small target motion detection\nagainst cluttered moving backgrounds. Compared to the existing models, the\nsmall-field and wide-field visual features are separately extracted by two\nmotion-sensitive neurons to detect small target motion and background motion.\nThese two types of motion information are further integrated to filter out\nfalse positives. Extensive experiments showed that the proposed model can\noutperform the existing models in terms of detection rates."", ""We present the first method to perform automatic 3D pose, shape and texture\ncapture of animals from images acquired in-the-wild. In particular, we focus on\nthe problem of capturing 3D information about Grevy's zebras from a collection\nof images. The Grevy's zebra is one of the most endangered species in Africa,\nwith only a few thousand individuals left. Capturing the shape and pose of\nthese animals can provide biologists and conservationists with information\nabout animal health and behavior. In contrast to research on human pose, shape\nand texture estimation, training data for endangered species is limited, the\nanimals are in complex natural scenes with occlusion, they are naturally\ncamouflaged, travel in herds, and look similar to each other. To overcome these\nchallenges, we integrate the recent SMAL animal model into a network-based\nregression pipeline, which we train end-to-end on synthetically generated\nimages with pose, shape, and background variation. Going beyond\nstate-of-the-art methods for human shape and pose estimation, our method learns\na shape space for zebras during training. Learning such a shape space from\nimages using only a photometric loss is novel, and the approach can be used to\nlearn shape in other settings with limited 3D supervision. Moreover, we couple\n3D pose and shape prediction with the task of texture synthesis, obtaining a\nfull texture map of the animal from a single image. We show that the predicted\ntexture map allows a novel per-instance unsupervised optimization over the\nnetwork features. This method, SMALST (SMAL with learned Shape and Texture)\ngoes beyond previous work, which assumed manual keypoints and/or segmentation,\nto regress directly from pixels to 3D animal shape, pose and texture. Code and\ndata are available at https://github.com/silviazuffi/smalst."", 'In this paper, we are interested in pose estimation of animals. Animals\nusually exhibit a wide range of variations on poses and there is no available\nanimal pose dataset for training and testing. To address this problem, we build\nan animal pose dataset to facilitate training and evaluation. Considering the\nheavy labor needed to label dataset and it is impossible to label data for all\nconcerned animal species, we, therefore, proposed a novel cross-domain\nadaptation method to transform the animal pose knowledge from labeled animal\nclasses to unlabeled animal classes. We use the modest animal pose dataset to\nadapt learned knowledge to multiple animals species. Moreover, humans also\nshare skeleton similarities with some animals (especially four-footed mammals).\nTherefore, the easily available human pose dataset, which is of a much larger\nscale than our labeled animal dataset, provides important prior knowledge to\nboost up the performance on animal pose estimation. Experiments show that our\nproposed method leverages these pieces of prior knowledge well and achieves\nconvincing results on animal pose estimation.', 'Self-driving vehicle vision systems must deal with an extremely broad and\nchallenging set of scenes. They can potentially exploit an enormous amount of\ntraining data collected from vehicles in the field, but the volumes are too\nlarge to train offline naively. Not all training instances are equally valuable\nthough, and importance sampling can be used to prioritize which training images\nto collect. This approach assumes that objects in images are labeled with high\naccuracy. To generate accurate labels in the field, we exploit the\nspatio-temporal coherence of vehicle video. We use a near-to-far labeling\nstrategy by first labeling large, close objects in the video, and tracking them\nback in time to induce labels on small distant presentations of those objects.\nIn this paper we demonstrate the feasibility of this approach in several steps.\nFirst, we note that an optimal subset (relative to all the objects encountered\nand labeled) of labeled objects in images can be obtained by importance\nsampling using gradients of the recognition network. Next we show that these\ngradients can be approximated with very low error using the loss function,\nwhich is already available when the CNN is running inference. Then, we\ngeneralize these results to objects in a larger scene using an object detection\nsystem. Finally, we describe a self-labeling scheme using object tracking.\nObjects are tracked back in time (near-to-far) and labels of near objects are\nused to check accuracy of those objects in the far field. We then evaluate the\naccuracy of models trained on importance sampled data vs models trained on\ncomplete data.', 'The demand for on-device document recognition systems increases in\nconjunction with the emergence of more strict privacy and security\nrequirements. In such systems, there is no data transfer from the end device to\na third-party information processing servers. The response time is vital to the\nuser experience of on-device document recognition. Combined with the\nunavailability of discrete GPUs, powerful CPUs, or a large RAM capacity on\nconsumer-grade end devices such as smartphones, the time limitations put\nsignificant constraints on the computational complexity of the applied\nalgorithms for on-device execution.\n  In this work, we consider document location in an image without prior\nknowledge of the document content or its internal structure. In accordance with\nthe published works, at least 5 systems offer solutions for on-device document\nlocation. All these systems use a location method which can be considered\nHough-based. The precision of such systems seems to be lower than that of the\nstate-of-the-art solutions which were not designed to account for the limited\ncomputational resources.\n  We propose an advanced Hough-based method. In contrast with other approaches,\nit accounts for the geometric invariants of the central projection model and\ncombines both edge and color features for document boundary detection. The\nproposed method allowed for the second best result for SmartDoc dataset in\nterms of precision, surpassed by U-net like neural network. When evaluated on a\nmore challenging MIDV-500 dataset, the proposed algorithm guaranteed the best\nprecision compared to published methods. Our method retained the applicability\nto on-device computations.', 'We introduce a novel type of text representation that preserves the 2D layout\nof a document. This is achieved by encoding each document page as a\ntwo-dimensional grid of characters. Based on this representation, we present a\ngeneric document understanding pipeline for structured documents. This pipeline\nmakes use of a fully convolutional encoder-decoder network that predicts a\nsegmentation mask and bounding boxes. We demonstrate its capabilities on an\ninformation extraction task from invoices and show that it significantly\noutperforms approaches based on sequential text or document images.', 'In this paper, we present supervision-by-registration, an unsupervised\napproach to improve the precision of facial landmark detectors on both images\nand video. Our key observation is that the detections of the same landmark in\nadjacent frames should be coherent with registration, i.e., optical flow.\nInterestingly, the coherency of optical flow is a source of supervision that\ndoes not require manual labeling, and can be leveraged during detector\ntraining. For example, we can enforce in the training loss function that a\ndetected landmark at frame$_{t-1}$ followed by optical flow tracking from\nframe$_{t-1}$ to frame$_t$ should coincide with the location of the detection\nat frame$_t$. Essentially, supervision-by-registration augments the training\nloss function with a registration loss, thus training the detector to have\noutput that is not only close to the annotations in labeled images, but also\nconsistent with registration on large amounts of unlabeled videos. End-to-end\ntraining with the registration loss is made possible by a differentiable\nLucas-Kanade operation, which computes optical flow registration in the forward\npass, and back-propagates gradients that encourage temporal coherency in the\ndetector. The output of our method is a more precise image-based facial\nlandmark detector, which can be applied to single images or video. With\nsupervision-by-registration, we demonstrate (1) improvements in facial landmark\ndetection on both images (300W, ALFW) and video (300VW, Youtube-Celebrities),\nand (2) significant reduction of jittering in video detections.', 'Despite significant progress on current state-of-the-art image generation\nmodels, synthesis of document images containing multiple and complex object\nlayouts is a challenging task. This paper presents a novel approach, called\nDocSynth, to automatically synthesize document images based on a given layout.\nIn this work, given a spatial layout (bounding boxes with object categories) as\na reference by the user, our proposed DocSynth model learns to generate a set\nof realistic document images consistent with the defined layout. Also, this\nframework has been adapted to this work as a superior baseline model for\ncreating synthetic document image datasets for augmenting real data during\ntraining for document layout analysis tasks. Different sets of learning\nobjectives have been also used to improve the model performance.\nQuantitatively, we also compare the generated results of our model with real\ndata using standard evaluation metrics. The results highlight that our model\ncan successfully generate realistic and diverse document images with multiple\nobjects. We also present a comprehensive qualitative analysis summary of the\ndifferent scopes of synthetic image generation tasks. Lastly, to our knowledge\nthis is the first work of its kind.', 'Segmentation of additive manufacturing (AM) defects in X-ray Computed\nTomography (XCT) images is challenging, due to the poor contrast, small sizes\nand variation in appearance of defects. Automatic segmentation can, however,\nprovide quality control for additive manufacturing. Over recent years,\nthree-dimensional convolutional neural networks (3D CNNs) have performed well\nin the volumetric segmentation of medical images. In this work, we leverage\ntechniques from the medical imaging domain and propose training a 3D U-Net\nmodel to automatically segment defects in XCT images of AM samples. This work\nnot only contributes to the use of machine learning for AM defect detection but\nalso demonstrates for the first time 3D volumetric segmentation in AM. We train\nand test with three variants of the 3D U-Net on an AM dataset, achieving a mean\nintersection of union (IOU) value of 88.4%.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'A lot of research has been devoted to identity documents analysis and\nrecognition on mobile devices. However, no publicly available datasets designed\nfor this particular problem currently exist. There are a few datasets which are\nuseful for associated subtasks but in order to facilitate a more comprehensive\nscientific and technical approach to identity document recognition more\nspecialized datasets are required. In this paper we present a Mobile Identity\nDocument Video dataset (MIDV-500) consisting of 500 video clips for 50\ndifferent identity document types with ground truth which allows to perform\nresearch in a wide scope of document analysis problems. The paper presents\ncharacteristics of the dataset and evaluation results for existing methods of\nface detection, text line recognition, and document fields data extraction.\nSince an important feature of identity documents is their sensitiveness as they\ncontain personal data, all source document images used in MIDV-500 are either\nin public domain or distributed under public copyright licenses.\n  The main goal of this paper is to present a dataset. However, in addition and\nas a baseline, we present evaluation results for existing methods for face\ndetection, text line recognition, and document data extraction, using the\npresented dataset.\n  (The dataset is available for download at ftp://smartengines.com/midv-500/.)', ""We study the problem of segmenting moving objects in unconstrained videos.\nGiven a video, the task is to segment all the objects that exhibit independent\nmotion in at least one frame. We formulate this as a learning problem and\ndesign our framework with three cues: (i) independent object motion between a\npair of frames, which complements object recognition, (ii) object appearance,\nwhich helps to correct errors in motion estimation, and (iii) temporal\nconsistency, which imposes additional constraints on the segmentation. The\nframework is a two-stream neural network with an explicit memory module. The\ntwo streams encode appearance and motion cues in a video sequence respectively,\nwhile the memory module captures the evolution of objects over time, exploiting\nthe temporal consistency. The motion stream is a convolutional neural network\ntrained on synthetic videos to segment independently moving objects in the\noptical flow field. The module to build a 'visual memory' in video, i.e., a\njoint representation of all the video frames, is realized with a convolutional\nrecurrent unit learned from a small number of training video sequences.\n  For every pixel in a frame of a test video, our approach assigns an object or\nbackground label based on the learned spatio-temporal features as well as the\n'visual memory' specific to the video. We evaluate our method extensively on\nthree benchmarks, DAVIS, Freiburg-Berkeley motion segmentation dataset and\nSegTrack. In addition, we provide an extensive ablation study to investigate\nboth the choice of the training data and the influence of each component in the\nproposed framework."", 'The digital camera captured document images may often be warped and distorted\ndue to different camera angles or document surfaces. A robust technique is\nneeded to solve this kind of distortion. The research on dewarping of the\ndocument suffers due to the limited availability of benchmark public dataset.\nIn recent times, deep learning based approaches are used to solve the problems\naccurately. To train most of the deep neural networks a large number of\ndocument images is required and generating such a large volume of document\nimages manually is difficult. In this paper, we propose a technique to generate\na synthetic warped image from a flat-bedded scanned document image. It is done\nby calculating warping factors for each pixel position using two warping\nposition parameters (WPP) and eight warping control parameters (WCP). These\nparameters can be specified as needed depending upon the desired warping. The\nresults are compared with similar real captured images both qualitative and\nquantitative way.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'Computation of document image quality metrics often depends upon the\navailability of a ground truth image corresponding to the document. This limits\nthe applicability of quality metrics in applications such as hyperparameter\noptimization of image processing algorithms that operate on-the-fly on unseen\ndocuments. This work proposes the use of surrogate models to learn the behavior\nof a given document quality metric on existing datasets where ground truth\nimages are available. The trained surrogate model can later be used to predict\nthe metric value on previously unseen document images without requiring access\nto ground truth images. The surrogate model is empirically evaluated on the\nDocument Image Binarization Competition (DIBCO) and the Handwritten Document\nImage Binarization Competition (H-DIBCO) datasets.', 'This paper presents results of Document Visual Question Answering Challenge\norganized as part of ""Text and Documents in the Deep Learning Era"" workshop, in\nCVPR 2020. The challenge introduces a new problem - Visual Question Answering\non document images. The challenge comprised two tasks. The first task concerns\nwith asking questions on a single document image. On the other hand, the second\ntask is set as a retrieval task where the question is posed over a collection\nof images. For the task 1 a new dataset is introduced comprising 50,000\nquestions-answer(s) pairs defined over 12,767 document images. For task 2\nanother dataset has been created comprising 20 questions over 14,362 document\nimages which share the same document template.', 'Convolutional neural networks perform well on object recognition because of a\nnumber of recent advances: rectified linear units (ReLUs), data augmentation,\ndropout, and large labelled datasets. Unsupervised data has been proposed as\nanother way to improve performance. Unfortunately, unsupervised pre-training is\nnot used by state-of-the-art methods leading to the following question: Is\nunsupervised pre-training still useful given recent advances? If so, when? We\nanswer this in three parts: we 1) develop an unsupervised method that\nincorporates ReLUs and recent unsupervised regularization techniques, 2)\nanalyze the benefits of unsupervised pre-training compared to data augmentation\nand dropout on CIFAR-10 while varying the ratio of unsupervised to supervised\nsamples, 3) verify our findings on STL-10. We discover unsupervised\npre-training, as expected, helps when the ratio of unsupervised to supervised\nsamples is high, and surprisingly, hurts when the ratio is low. We also use\nunsupervised pre-training with additional color augmentation to achieve near\nstate-of-the-art performance on STL-10.', 'Disentanglement is at the forefront of unsupervised learning, as disentangled\nrepresentations of data improve generalization, interpretability, and\nperformance in downstream tasks. Current unsupervised approaches remain\ninapplicable for real-world datasets since they are highly variable in their\nperformance and fail to reach levels of disentanglement of (semi-)supervised\napproaches. We introduce population-based training (PBT) for improving\nconsistency in training variational autoencoders (VAEs) and demonstrate the\nvalidity of this approach in a supervised setting (PBT-VAE). We then use\nUnsupervised Disentanglement Ranking (UDR) as an unsupervised heuristic to\nscore models in our PBT-VAE training and show how models trained this way tend\nto consistently disentangle only a subset of the generative factors. Building\non top of this observation we introduce the recursive rPU-VAE approach. We\ntrain the model until convergence, remove the learned factors from the dataset\nand reiterate. In doing so, we can label subsets of the dataset with the\nlearned factors and consecutively use these labels to train one model that\nfully disentangles the whole dataset. With this approach, we show striking\nimprovement in state-of-the-art unsupervised disentanglement performance and\nrobustness across multiple datasets and metrics.', 'Imagine trying to track one particular fruitfly in a swarm of hundreds.\nHigher biological visual systems have evolved to track moving objects by\nrelying on both appearance and motion features. We investigate if\nstate-of-the-art deep neural networks for visual tracking are capable of the\nsame. For this, we introduce PathTracker, a synthetic visual challenge that\nasks human observers and machines to track a target object in the midst of\nidentical-looking ""distractor"" objects. While humans effortlessly learn\nPathTracker and generalize to systematic variations in task design,\nstate-of-the-art deep networks struggle. To address this limitation, we\nidentify and model circuit mechanisms in biological brains that are implicated\nin tracking objects based on motion cues. When instantiated as a recurrent\nnetwork, our circuit model learns to solve PathTracker with a robust visual\nstrategy that rivals human performance and explains a significant proportion of\ntheir decision-making on the challenge. We also show that the success of this\ncircuit model extends to object tracking in natural videos. Adding it to a\ntransformer-based architecture for object tracking builds tolerance to visual\nnuisances that affect object appearance, resulting in a new state-of-the-art\nperformance on the large-scale TrackingNet object tracking challenge. Our work\nhighlights the importance of building artificial vision models that can help us\nbetter understand human vision and improve computer vision.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', ""Class labels have been empirically shown useful in improving the sample\nquality of generative adversarial nets (GANs). In this paper, we mathematically\nstudy the properties of the current variants of GANs that make use of class\nlabel information. With class aware gradient and cross-entropy decomposition,\nwe reveal how class labels and associated losses influence GAN's training.\nBased on that, we propose Activation Maximization Generative Adversarial\nNetworks (AM-GAN) as an advanced solution. Comprehensive experiments have been\nconducted to validate our analysis and evaluate the effectiveness of our\nsolution, where AM-GAN outperforms other strong baselines and achieves\nstate-of-the-art Inception Score (8.91) on CIFAR-10. In addition, we\ndemonstrate that, with the Inception ImageNet classifier, Inception Score\nmainly tracks the diversity of the generator, and there is, however, no\nreliable evidence that it can reflect the true sample quality. We thus propose\na new metric, called AM Score, to provide a more accurate estimation of the\nsample quality. Our proposed model also outperforms the baseline methods in the\nnew metric."", 'In the Gastric Histopathology Image Classification (GHIC) tasks, which are\nusually weakly supervised learning missions, there is inevitably redundant\ninformation in the images. Therefore, designing networks that can focus on\neffective distinguishing features has become a popular research topic. In this\npaper, to accomplish the tasks of GHIC superiorly and to assist pathologists in\nclinical diagnosis, an intelligent Hierarchical Conditional Random Field based\nAttention Mechanism (HCRF-AM) model is proposed. The HCRF-AM model consists of\nan Attention Mechanism (AM) module and an Image Classification (IC) module. In\nthe AM module, an HCRF model is built to extract attention regions. In the IC\nmodule, a Convolutional Neural Network (CNN) model is trained with the\nattention regions selected and then an algorithm called Classification\nProbability-based Ensemble Learning is applied to obtain the image-level\nresults from patch-level output of the CNN. In the experiment, a classification\nspecificity of 96.67% is achieved on a gastric histopathology dataset with 700\nimages. Our HCRF-AM model demonstrates high classification performance and\nshows its effectiveness and future potential in the GHIC field.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Document segmentation is a method of rending the document into distinct\nregions. A document is an assortment of information and a standard mode of\nconveying information to others. Pursuance of data from documents involves ton\nof human effort, time intense and might severely prohibit the usage of data\nsystems. So, automatic information pursuance from the document has become a big\nissue. It is been shown that document segmentation will facilitate to beat such\nproblems. This paper proposes a new approach to segment and classify the\ndocument regions as text, image, drawings and table. Document image is divided\ninto blocks using Run length smearing rule and features are extracted from\nevery blocks. Discipulus tool has been used to construct the Genetic\nprogramming based classifier model and located 97.5% classification accuracy.', 'What happens if one pushes a cup sitting on a table toward the edge of the\ntable? How about pushing a desk against a wall? In this paper, we study the\nproblem of understanding the movements of objects as a result of applying\nexternal forces to them. For a given force vector applied to a specific\nlocation in an image, our goal is to predict long-term sequential movements\ncaused by that force. Doing so entails reasoning about scene geometry, objects,\ntheir attributes, and the physical rules that govern the movements of objects.\nWe design a deep neural network model that learns long-term sequential\ndependencies of object movements while taking into account the geometry and\nappearance of the scene by combining Convolutional and Recurrent Neural\nNetworks. Training our model requires a large-scale dataset of object movements\ncaused by external forces. To build a dataset of forces in scenes, we\nreconstructed all images in SUN RGB-D dataset in a physics simulator to\nestimate the physical movements of objects caused by external forces applied to\nthem. Our Forces in Scenes (ForScene) dataset contains 10,335 images in which a\nvariety of external forces are applied to different types of objects resulting\nin more than 65,000 object movements represented in 3D. Our experimental\nevaluations show that the challenging task of predicting long-term movements of\nobjects as their reaction to external forces is possible from a single image.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'Video understanding is one of the most challenging topics in computer vision.\nIn this paper, a four-stage video understanding pipeline is presented to\nsimultaneously recognize all atomic actions and the single on-going activity in\na video. This pipeline uses objects and motions from the video and a\ngraph-based knowledge representation network as prior reference. Two deep\nnetworks are trained to identify objects and motions in each video sequence\nassociated with an action. Low Level image features are then used to identify\nobjects of interest in that video sequence. Confidence scores are assigned to\nobjects of interest based on their involvement in the action and to motion\nclasses based on results from a deep neural network that classifies the\non-going action in video into motion classes. Confidence scores are computed\nfor each candidate functional unit associated with an action using a knowledge\nrepresentation network, object confidences, and motion confidences. Each action\nis therefore associated with a functional unit and the sequence of actions is\nfurther evaluated to identify the single on-going activity in the video. The\nknowledge representation used in the pipeline is called the functional\nobject-oriented network which is a graph-based network useful for encoding\nknowledge about manipulation tasks. Experiments are performed on a dataset of\ncooking videos to test the proposed algorithm with action inference and\nactivity classification. Experiments show that using functional object oriented\nnetwork improves video understanding significantly.', 'Graphical elements: particularly tables and figures contain a visual summary\nof the most valuable information contained in a document. Therefore,\nlocalization of such graphical objects in the document images is the initial\nstep to understand the content of such graphical objects or document images. In\nthis paper, we present a novel end-to-end trainable deep learning based\nframework to localize graphical objects in the document images called as\nGraphical Object Detection (GOD). Our framework is data-driven and does not\nrequire any heuristics or meta-data to locate graphical objects in the document\nimages. The GOD explores the concept of transfer learning and domain adaptation\nto handle scarcity of labeled training images for graphical object detection\ntask in the document images. Performance analysis carried out on the various\npublic benchmark data sets: ICDAR-2013, ICDAR-POD2017,and UNLV shows that our\nmodel yields promising results as compared to state-of-the-art techniques.', ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'Identity documents recognition is an important sub-field of document\nanalysis, which deals with tasks of robust document detection, type\nidentification, text fields recognition, as well as identity fraud prevention\nand document authenticity validation given photos, scans, or video frames of an\nidentity document capture. Significant amount of research has been published on\nthis topic in recent years, however a chief difficulty for such research is\nscarcity of datasets, due to the subject matter being protected by security\nrequirements. A few datasets of identity documents which are available lack\ndiversity of document types, capturing conditions, or variability of document\nfield values. In addition, the published datasets were typically designed only\nfor a subset of document recognition problems, not for a complex identity\ndocument analysis. In this paper, we present a dataset MIDV-2020 which consists\nof 1000 video clips, 2000 scanned images, and 1000 photos of 1000 unique mock\nidentity documents, each with unique text field values and unique artificially\ngenerated faces, with rich annotation. For the presented benchmark dataset\nbaselines are provided for such tasks as document location and identification,\ntext fields recognition, and face detection. With 72409 annotated images in\ntotal, to the date of publication the proposed dataset is the largest publicly\navailable identity documents dataset with variable artificially generated data,\nand we believe that it will prove invaluable for advancement of the field of\ndocument analysis and recognition. The dataset is available for download at\nftp://smartengines.com/midv-2020 and http://l3i-share.univ-lr.fr .', 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', 'The unsupervised training of GANs and VAEs has enabled them to generate\nrealistic images mimicking real-world distributions and perform image-based\nunsupervised clustering or semi-supervised classification. Combining the power\nof these two generative models, we introduce Multi-Adversarial Variational\nautoEncoder Networks (MAVENs), a novel network architecture that incorporates\nan ensemble of discriminators in a VAE-GAN network, with simultaneous\nadversarial learning and variational inference. We apply MAVENs to the\ngeneration of synthetic images and propose a new distribution measure to\nquantify the quality of the generated images. Our experimental results using\ndatasets from the computer vision and medical imaging domains---Street View\nHouse Numbers, CIFAR-10, and Chest X-Ray datasets---demonstrate competitive\nperformance against state-of-the-art semi-supervised models both in image\ngeneration and classification tasks.', 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently engaged in a project aimed at identifying critical areas within medical images, such as X-rays, to aid in disease diagnosis. For instance, in lung X-rays, the region with a tumor is of greater significance, while other areas may not be as relevant. I am seeking a method to pinpoint these crucial areas in the images. Subsequently, we can enhance efficiency by capturing high-resolution images of these important areas. I am interested in models that can execute this concept by initially analyzing a low-resolution image, detecting significant features, and then focusing on more localized but detailed aspects. Additionally, I require techniques that can deliver the same performance level but with reduced computational power and storage, considering the vast quantity of medical images. I need a newer and larger dataset to train and test the model. I would also appreciate recommendations for other image analysis models that are more efficient yet still yield accurate results.","['To enable a deep learning-based system to be used in the medical domain as a\ncomputer-aided diagnosis system, it is essential to not only classify diseases\nbut also present the locations of the diseases. However, collecting\ninstance-level annotations for various thoracic diseases is expensive.\nTherefore, weakly supervised localization methods have been proposed that use\nonly image-level annotation. While the previous methods presented the disease\nlocation as the most discriminative part for classification, this causes a deep\nnetwork to localize wrong areas for indistinguishable X-ray images. To solve\nthis issue, we propose a spatial attention method using disease masks that\ndescribe the areas where diseases mainly occur. We then apply the spatial\nattention to find the precise disease area by highlighting the highest\nprobability of disease occurrence. Meanwhile, the various sizes, rotations and\nnoise in chest X-ray images make generating the disease masks challenging. To\nreduce the variation among images, we employ an alignment module to transform\nan input X-ray image into a generalized image. Through extensive experiments on\nthe NIH-Chest X-ray dataset with eight kinds of diseases, we show that the\nproposed method results in superior localization performances compared to\nstate-of-the-art methods.', 'COVID-19 is a virus with high transmission rate that demands rapid\nidentification of the infected patients to reduce the spread of the disease.\nThe current gold-standard test, Reverse-Transcription Polymerase Chain Reaction\n(RT-PCR), has a high rate of false negatives. Diagnosing from CT-scan images as\na more accurate alternative has the challenge of distinguishing COVID-19 from\nother pneumonia diseases. Artificial intelligence can help radiologists and\nphysicians to accelerate the process of diagnosis, increase its accuracy, and\nmeasure the severity of the disease. We designed a new interpretable deep\nneural network to distinguish healthy people, patients with COVID-19, and\npatients with other pneumonia diseases from axial lung CT-scan images. Our\nmodel also detects the infected areas and calculates the percentage of the\ninfected lung volume. We first preprocessed the images to eliminate the batch\neffects of different devices, and then adopted a weakly supervised method to\ntrain the model without having any tags for the infected parts. We trained and\nevaluated the model on a large dataset of 3359 samples from 6 different medical\ncenters. The model reached sensitivities of 97.75% and 98.15%, and\nspecificities of 87% and 81.03% in separating healthy people from the diseased\nand COVID-19 from other diseases, respectively. It also demonstrated similar\nperformance for 1435 samples from 6 different medical centers which proves its\ngeneralizability. The performance of the model on a large diverse dataset, its\ngeneralizability, and interpretability makes it suitable to be used as a\nreliable diagnostic system.', 'Medical images differ from natural images in significantly higher resolutions\nand smaller regions of interest. Because of these differences, neural network\narchitectures that work well for natural images might not be applicable to\nmedical image analysis. In this work, we extend the globally-aware multiple\ninstance classifier, a framework we proposed to address these unique properties\nof medical images. This model first uses a low-capacity, yet memory-efficient,\nnetwork on the whole image to identify the most informative regions. It then\napplies another higher-capacity network to collect details from chosen regions.\nFinally, it employs a fusion module that aggregates global and local\ninformation to make a final prediction. While existing methods often require\nlesion segmentation during training, our model is trained with only image-level\nlabels and can generate pixel-level saliency maps indicating possible malignant\nfindings. We apply the model to screening mammography interpretation:\npredicting the presence or absence of benign and malignant lesions. On the NYU\nBreast Cancer Screening Dataset, consisting of more than one million images,\nour model achieves an AUC of 0.93 in classifying breasts with malignant\nfindings, outperforming ResNet-34 and Faster R-CNN. Compared to ResNet-34, our\nmodel is 4.1x faster for inference while using 78.4% less GPU memory.\nFurthermore, we demonstrate, in a reader study, that our model surpasses\nradiologist-level AUC by a margin of 0.11. The proposed model is available\nonline: https://github.com/nyukat/GMIC.', 'Medical image analysis practitioners have embraced big data methodologies.\nThis has created a need for large annotated datasets. The source of big data is\ntypically large image collections and clinical reports recorded for these\nimages. In many cases, however, building algorithms aimed at segmentation and\ndetection of disease requires a training dataset with markings of the areas of\ninterest on the image that match with the described anomalies. This process of\nannotation is expensive and needs the involvement of clinicians. In this work\nwe propose two separate deep neural network architectures for automatic marking\nof a region of interest (ROI) on the image best representing a finding\nlocation, given a textual report or a set of keywords. One architecture\nconsists of LSTM and CNN components and is trained end to end with images,\nmatching text, and markings of ROIs for those images. The output layer\nestimates the coordinates of the vertices of a polygonal region. The second\narchitecture uses a network pre-trained on a large dataset of the same image\ntypes for learning feature representations of the findings of interest. We show\nthat for a variety of findings from chest X-ray images, both proposed\narchitectures learn to estimate the ROI, as validated by clinical annotations.\nThere is a clear advantage obtained from the architecture with pre-trained\nimaging network. The centroids of the ROIs marked by this network were on\naverage at a distance equivalent to 5.1% of the image width from the centroids\nof the ground truth ROIs.', 'The automatic detection of critical findings in chest X-rays (CXR), such as\npneumothorax, is important for assisting radiologists in their clinical\nworkflow like triaging time-sensitive cases and screening for incidental\nfindings. While deep learning (DL) models has become a promising predictive\ntechnology with near-human accuracy, they commonly suffer from a lack of\nexplainability, which is an important aspect for clinical deployment of DL\nmodels in the highly regulated healthcare industry. For example, localizing\ncritical findings in an image is useful for explaining the predictions of DL\nclassification algorithms. While there have been a host of joint classification\nand localization methods for computer vision, the state-of-the-art DL models\nrequire locally annotated training data in the form of pixel level labels or\nbounding box coordinates. In the medical domain, this requires an expensive\namount of manual annotation by medical experts for each critical finding. This\nrequirement becomes a major barrier for training models that can rapidly scale\nto various findings. In this work, we address these shortcomings with an\ninterpretable DL algorithm based on multi-instance learning that jointly\nclassifies and localizes critical findings in CXR without the need for local\nannotations. We show competitive classification results on three different\ncritical findings (pneumothorax, pneumonia, and pulmonary edema) from three\ndifferent CXR datasets.', 'Advances in deep learning for natural images have prompted a surge of\ninterest in applying similar techniques to medical images. The majority of the\ninitial attempts focused on replacing the input of a deep convolutional neural\nnetwork with a medical image, which does not take into consideration the\nfundamental differences between these two types of images. Specifically, fine\ndetails are necessary for detection in medical images, unlike in natural images\nwhere coarse structures matter most. This difference makes it inadequate to use\nthe existing network architectures developed for natural images, because they\nwork on heavily downscaled images to reduce the memory requirements. This hides\ndetails necessary to make accurate predictions. Additionally, a single exam in\nmedical imaging often comes with a set of views which must be fused in order to\nreach a correct conclusion. In our work, we propose to use a multi-view deep\nconvolutional neural network that handles a set of high-resolution medical\nimages. We evaluate it on large-scale mammography-based breast cancer screening\n(BI-RADS prediction) using 886,000 images. We focus on investigating the impact\nof the training set size and image size on the prediction accuracy. Our results\nhighlight that performance increases with the size of training set, and that\nthe best performance can only be achieved using the original resolution. In the\nreader study, performed on a random subset of the test set, we confirmed the\nefficacy of our model, which achieved performance comparable to a committee of\nradiologists when presented with the same data.', 'Radiologists usually observe anatomical regions of chest X-ray images as well\nas the overall image before making a decision. However, most existing deep\nlearning models only look at the entire X-ray image for classification, failing\nto utilize important anatomical information. In this paper, we propose a novel\nmulti-label chest X-ray classification model that accurately classifies the\nimage finding and also localizes the findings to their correct anatomical\nregions. Specifically, our model consists of two modules, the detection module\nand the anatomical dependency module. The latter utilizes graph convolutional\nnetworks, which enable our model to learn not only the label dependency but\nalso the relationship between the anatomical regions in the chest X-ray. We\nfurther utilize a method to efficiently create an adjacency matrix for the\nanatomical regions using the correlation of the label across the different\nregions. Detailed experiments and analysis of our results show the\neffectiveness of our method when compared to the current state-of-the-art\nmulti-label chest X-ray image classification methods while also providing\naccurate location information.', 'Deep learning (DL) has drawn tremendous attention in object localization and\nrecognition for both natural and medical images. U-Net segmentation models have\ndemonstrated superior performance compared to conventional handcrafted\nfeature-based methods. Medical image modality-specific DL models are better at\ntransferring domain knowledge to a relevant target task than those that are\npretrained on stock photography images. This helps improve model adaptation,\ngeneralization, and class-specific region of interest (ROI) localization. In\nthis study, we train chest X-ray (CXR) modality-specific U-Nets and other\nstate-of-the-art U-Net models for semantic segmentation of tuberculosis\n(TB)-consistent findings. Automated segmentation of such manifestations could\nhelp radiologists reduce errors and supplement decision-making while improving\npatient care and productivity. Our approach uses the publicly available TBX11K\nCXR dataset with weak TB annotations, typically provided as bounding boxes, to\ntrain a set of U-Net models. Next, we improve the results by augmenting the\ntraining data with weak localizations, post-processed into an ROI mask, from a\nDL classifier that is trained to classify CXRs as showing normal lungs or\nsuspected TB manifestations. Test data are individually derived from the TBX11K\nCXR training distribution and other cross-institutional collections including\nthe Shenzhen TB and Montgomery TB CXR datasets. We observe that our augmented\ntraining strategy helped the CXR modality-specific U-Net models achieve\nsuperior performance with test data derived from the TBX11K CXR training\ndistribution as well as from cross-institutional collections (p < 0.05).', 'The deployment of automated systems to diagnose diseases from medical images\nis challenged by the requirement to localise the diagnosed diseases to justify\nor explain the classification decision. This requirement is hard to fulfil\nbecause most of the training sets available to develop these systems only\ncontain global annotations, making the localisation of diseases a weakly\nsupervised approach. The main methods designed for weakly supervised disease\nclassification and localisation rely on saliency or attention maps that are not\nspecifically trained for localisation, or on region proposals that can not be\nrefined to produce accurate detections. In this paper, we introduce a new model\nthat combines region proposal and saliency detection to overcome both\nlimitations for weakly supervised disease classification and localisation.\nUsing the ChestX-ray14 data set, we show that our proposed model establishes\nthe new state-of-the-art for weakly-supervised disease diagnosis and\nlocalisation.', 'Building a highly accurate predictive model for classification and\nlocalization of abnormalities in chest X-rays usually requires a large number\nof manually annotated labels and pixel regions (bounding boxes) of\nabnormalities. However, it is expensive to acquire such annotations, especially\nthe bounding boxes. Recently, contrastive learning has shown strong promise in\nleveraging unlabeled natural images to produce highly generalizable and\ndiscriminative features. However, extending its power to the medical image\ndomain is under-explored and highly non-trivial, since medical images are much\nless amendable to data augmentations. In contrast, their prior knowledge, as\nwell as radiomic features, is often crucial. To bridge this gap, we propose an\nend-to-end semi-supervised knowledge-augmented contrastive learning framework,\nthat simultaneously performs disease classification and localization tasks. The\nkey knob of our framework is a unique positive sampling approach tailored for\nthe medical images, by seamlessly integrating radiomic features as a knowledge\naugmentation. Specifically, we first apply an image encoder to classify the\nchest X-rays and to generate the image features. We next leverage Grad-CAM to\nhighlight the crucial (abnormal) regions for chest X-rays (even when\nunannotated), from which we extract radiomic features. The radiomic features\nare then passed through another dedicated encoder to act as the positive sample\nfor the image features generated from the same chest X-ray. In this way, our\nframework constitutes a feedback loop for image and radiomic modality features\nto mutually reinforce each other. Their contrasting yields knowledge-augmented\nrepresentations that are both robust and interpretable. Extensive experiments\non the NIH Chest X-ray dataset demonstrate that our approach outperforms\nexisting baselines in both classification and localization tasks.', 'In the last few years, deep learning classifiers have shown promising results\nin image-based medical diagnosis. However, interpreting the outputs of these\nmodels remains a challenge. In cancer diagnosis, interpretability can be\nachieved by localizing the region of the input image responsible for the\noutput, i.e. the location of a lesion. Alternatively, segmentation or detection\nmodels can be trained with pixel-wise annotations indicating the locations of\nmalignant lesions. Unfortunately, acquiring such labels is labor-intensive and\nrequires medical expertise. To overcome this difficulty, weakly-supervised\nlocalization can be utilized. These methods allow neural network classifiers to\noutput saliency maps highlighting the regions of the input most relevant to the\nclassification task (e.g. malignant lesions in mammograms) using only\nimage-level labels (e.g. whether the patient has cancer or not) during\ntraining. When applied to high-resolution images, existing methods produce\nlow-resolution saliency maps. This is problematic in applications in which\nsuspicious lesions are small in relation to the image size. In this work, we\nintroduce a novel neural network architecture to perform weakly-supervised\nsegmentation of high-resolution images. The proposed model selects regions of\ninterest via coarse-level localization, and then performs fine-grained\nsegmentation of those regions. We apply this model to breast cancer diagnosis\nwith screening mammography, and validate it on a large clinically-realistic\ndataset. Measured by Dice similarity score, our approach outperforms existing\nmethods by a large margin in terms of localization performance of benign and\nmalignant lesions, relatively improving the performance by 39.6% and 20.0%,\nrespectively. Code and the weights of some of the models are available at\nhttps://github.com/nyukat/GLAM', 'Recently, chest X-ray report generation, which aims to automatically generate\ndescriptions of given chest X-ray images, has received growing research\ninterests. The key challenge of chest X-ray report generation is to accurately\ncapture and describe the abnormal regions. In most cases, the normal regions\ndominate the entire chest X-ray image, and the corresponding descriptions of\nthese normal regions dominate the final report. Due to such data bias,\nlearning-based models may fail to attend to abnormal regions. In this work, to\neffectively capture and describe abnormal regions, we propose the Contrastive\nAttention (CA) model. Instead of solely focusing on the current input image,\nthe CA model compares the current input image with normal images to distill the\ncontrastive information. The acquired contrastive information can better\nrepresent the visual features of abnormal regions. According to the experiments\non the public IU-X-ray and MIMIC-CXR datasets, incorporating our CA into\nseveral existing models can boost their performance across most metrics. In\naddition, according to the analysis, the CA model can help existing models\nbetter attend to the abnormal regions and provide more accurate descriptions\nwhich are crucial for an interpretable diagnosis. Specifically, we achieve the\nstate-of-the-art results on the two public datasets.', 'In recent years, besides the medical treatment methods in medical field,\nComputer Aided Diagnosis (CAD) systems which can facilitate the decision making\nphase of the physician and can detect the disease at an early stage have\nstarted to be used frequently. The diagnosis of Idiopathic Pulmonary Fibrosis\n(IPF) disease by using CAD systems is very important in that it can be followed\nby doctors and radiologists. It has become possible to diagnose and follow up\nthe disease with the help of CAD systems by the development of high resolution\ncomputed imaging scanners and increasing size of computation power. The purpose\nof this project is to design a tool that will help specialists diagnose and\nfollow up the IPF disease by identifying areas of honeycombing and ground glass\npatterns in High Resolution Computed Tomography (HRCT) lung images. Creating a\nprogram module that segments the lung pair and creating a self-learner deep\nlearning model from given Computed Tomography (CT) images for the specific\ndiseased regions thanks to doctors are the main purposes of this work. Through\nthe created model, program module will be able to find special regions in given\nnew CT images. In this study, the performance of lung segmentation was tested\nby the S{\\o}rensen-Dice coefficient method and the mean performance was\nmeasured as 90.7%, testing of the created model was performed with data not\nused in the training stage of the CNN network, and the average performance was\nmeasured as 87.8% for healthy regions, 73.3% for ground-glass areas and 69.1%\nfor honeycombing zones.', 'We consider the problem of abnormality localization for clinical\napplications. While deep learning has driven much recent progress in medical\nimaging, many clinical challenges are not fully addressed, limiting its broader\nusage. While recent methods report high diagnostic accuracies, physicians have\nconcerns trusting these algorithm results for diagnostic decision-making\npurposes because of a general lack of algorithm decision reasoning and\ninterpretability. One potential way to address this problem is to further train\nthese models to localize abnormalities in addition to just classifying them.\nHowever, doing this accurately will require a large amount of disease\nlocalization annotations by clinical experts, a task that is prohibitively\nexpensive to accomplish for most applications. In this work, we take a step\ntowards addressing these issues by means of a new attention-driven weakly\nsupervised algorithm comprising a hierarchical attention mining framework that\nunifies activation- and gradient-based visual attention in a holistic manner.\nOur key algorithmic innovations include the design of explicit ordinal\nattention constraints, enabling principled model training in a\nweakly-supervised fashion, while also facilitating the generation of\nvisual-attention-driven model explanations by means of localization cues. On\ntwo large-scale chest X-ray datasets (NIH ChestX-ray14 and CheXpert), we\ndemonstrate significant localization performance improvements over the current\nstate of the art while also achieving competitive classification performance.\nOur code is available on https://github.com/oyxhust/HAM.', ""Landmark Localization plays a very important role in processing medical\nimages as well as in disease identification. However, In medical field, it's a\nchallenging task because of the complexity of medical images and the high\nrequirement of accuracy for disease identification and treatment.There are two\ndominant ways to regress landmark coordination, one using the full\nconvolutional network to regress the heatmaps of landmarks , which is a complex\nway and heatmap post-process strategies are needed, and the other way is to\nregress the coordination using CNN + Full Connective Network directly, which is\nvery simple and faster training , but larger dataset and deeper model are\nneeded to achieve higher accuracy. Though with data augmentation and deeper\nnetwork it can reach a reasonable accuracy, but the accuracy still not reach\nthe requirement of medical field. In addition, a deeper networks also means\nlarger space consumption. To achieve a higher accuracy, we contrived a new\nlandmark regression method which combing heatmap regression and direct\ncoordinate regression base on probability methods and system control theory."", 'Telementoring surgeons as they perform surgery can be essential in the\ntreatment of patients when in situ expertise is not available. Nonetheless,\nexpert mentors are often unavailable to provide trainees with real-time medical\nguidance. When mentors are unavailable, a fallback autonomous mechanism should\nprovide medical practitioners with the required guidance. However,\nAI/autonomous mentoring in medicine has been limited by the availability of\ngeneralizable prediction models, and surgical procedures datasets to train\nthose models with. This work presents the initial steps towards the development\nof an intelligent artificial system for autonomous medical mentoring.\nSpecifically, we present the first Database for AI Surgical Instruction\n(DAISI). DAISI leverages on images and instructions to provide step-by-step\ndemonstrations of how to perform procedures from various medical disciplines.\nThe dataset was acquired from real surgical procedures and data from academic\ntextbooks. We used DAISI to train an encoder-decoder neural network capable of\npredicting medical instructions given a current view of the surgery.\nAfterwards, the instructions predicted by the network were evaluated using\ncumulative BLEU scores and input from expert physicians. According to the BLEU\nscores, the predicted and ground truth instructions were as high as 67%\nsimilar. Additionally, expert physicians subjectively assessed the algorithm\nusing Likert scale, and considered that the predicted descriptions were related\nto the images. This work provides a baseline for AI algorithms to assist in\nautonomous medical mentoring.', 'Deep learning models designed for visual classification tasks on natural\nimages have become prevalent in medical image analysis. However, medical images\ndiffer from typical natural images in many ways, such as significantly higher\nresolutions and smaller regions of interest. Moreover, both the global\nstructure and local details play important roles in medical image analysis\ntasks. To address these unique properties of medical images, we propose a\nneural network that is able to classify breast cancer lesions utilizing\ninformation from both a global saliency map and multiple local patches. The\nproposed model outperforms the ResNet-based baseline and achieves\nradiologist-level performance in the interpretation of screening mammography.\nAlthough our model is trained only with image-level labels, it is able to\ngenerate pixel-level saliency maps that provide localization of possible\nmalignant findings.', 'The identification and localization of diseases in medical images using deep\nlearning models have recently attracted significant interest. Existing methods\nonly consider training the networks with each image independently and most\nleverage an activation map for disease localization. In this paper, we propose\nan alternative approach that learns discriminative features among triplets of\nimages and cyclically trains on region features to verify whether attentive\nregions contain information indicative of a disease. Concretely, we adapt a\ndistance learning framework for multi-label disease classification to\ndifferentiate subtle disease features. Additionally, we feed back the features\nof the predicted class-specific regions to a separate classifier during\ntraining to better verify the localized diseases. Our model can achieve\nstate-of-the-art classification performance on the challenging ChestX-ray14\ndataset, and our ablation studies indicate that both distance learning and\nregion verification contribute to overall classification performance. Moreover,\nthe distance learning and region verification modules can capture essential\ninformation for better localization than baseline models without these modules.', 'The field of medical diagnostics contains a wealth of challenges which\nclosely resemble classical machine learning problems; practical constraints,\nhowever, complicate the translation of these endpoints naively into classical\narchitectures. Many tasks in radiology, for example, are largely problems of\nmulti-label classification wherein medical images are interpreted to indicate\nmultiple present or suspected pathologies. Clinical settings drive the\nnecessity for high accuracy simultaneously across a multitude of pathological\noutcomes and greatly limit the utility of tools which consider only a subset.\nThis issue is exacerbated by a general scarcity of training data and maximizes\nthe need to extract clinically relevant features from available samples --\nideally without the use of pre-trained models which may carry forward\nundesirable biases from tangentially related tasks. We present and evaluate a\npartial solution to these constraints in using LSTMs to leverage\ninterdependencies among target labels in predicting 14 pathologic patterns from\nchest x-rays and establish state of the art results on the largest publicly\navailable chest x-ray dataset from the NIH without pre-training. Furthermore,\nwe propose and discuss alternative evaluation metrics and their relevance in\nclinical practice.', 'One issue with computer based histopathology image analysis is that the size\nof the raw image is usually very large. Taking the raw image as input to the\ndeep learning model would be computationally expensive while resizing the raw\nimage to low resolution would incur information loss. In this paper, we present\na novel deep hybrid attention approach to breast cancer classification. It\nfirst adaptively selects a sequence of coarse regions from the raw image by a\nhard visual attention algorithm, and then for each such region it is able to\ninvestigate the abnormal parts based on a soft-attention mechanism. A recurrent\nnetwork is then built to make decisions to classify the image region and also\nto predict the location of the image region to be investigated at the next time\nstep. As the region selection process is non-differentiable, we optimize the\nwhole network through a reinforcement approach to learn an optimal policy to\nclassify the regions. Based on this novel Look, Investigate and Classify\napproach, we only need to process a fraction of the pixels in the raw image\nresulting in significant saving in computational resources without sacrificing\nperformances. Our approach is evaluated on a public breast cancer\nhistopathology database, where it demonstrates superior performance to the\nstate-of-the-art deep learning approaches, achieving around 96\\% classification\naccuracy while only 15% of raw pixels are used.', 'Lung cancer is the leading cause of cancer death worldwide. The critical\nreason for the deaths is delayed diagnosis and poor prognosis. With the\naccelerated development of deep learning techniques, it has been successfully\napplied extensively in many real-world applications, including health sectors\nsuch as medical image interpretation and disease diagnosis. By combining more\nmodalities that being engaged in the processing of information, multimodal\nlearning can extract better features and improve predictive ability. The\nconventional methods for lung cancer survival analysis normally utilize\nclinical data and only provide a statistical probability. To improve the\nsurvival prediction accuracy and help prognostic decision-making in clinical\npractice for medical experts, we for the first time propose a multimodal deep\nlearning method for non-small cell lung cancer (NSCLC) survival analysis, named\nDeepMMSA. This method leverages CT images in combination with clinical data,\nenabling the abundant information hold within medical images to be associate\nwith lung cancer survival information. We validate our method on the data of\n422 NSCLC patients from The Cancer Imaging Archive (TCIA). Experimental results\nsupport our hypothesis that there is an underlying relationship between\nprognostic information and radiomic images. Besides, quantitative results\nshowing that the established multimodal model can be applied to traditional\nmethod and has the potential to break bottleneck of existing methods and\nincrease the the percentage of concordant pairs(right predicted pairs) in\noverall population by 4%.', 'Generative Adversarial Networks (GAN) have many potential medical imaging\napplications, including data augmentation, domain adaptation, and model\nexplanation. Due to the limited memory of Graphical Processing Units (GPUs),\nmost current 3D GAN models are trained on low-resolution medical images, these\nmodels either cannot scale to high-resolution or are prone to patchy artifacts.\nIn this work, we propose a novel end-to-end GAN architecture that can generate\nhigh-resolution 3D images. We achieve this goal by using different\nconfigurations between training and inference. During training, we adopt a\nhierarchical structure that simultaneously generates a low-resolution version\nof the image and a randomly selected sub-volume of the high-resolution image.\nThe hierarchical design has two advantages: First, the memory demand for\ntraining on high-resolution images is amortized among sub-volumes. Furthermore,\nanchoring the high-resolution sub-volumes to a single low-resolution image\nensures anatomical consistency between sub-volumes. During inference, our model\ncan directly generate full high-resolution images. We also incorporate an\nencoder with a similar hierarchical structure into the model to extract\nfeatures from the images. Experiments on 3D thorax CT and brain MRI demonstrate\nthat our approach outperforms state of the art in image generation. We also\ndemonstrate clinical applications of the proposed model in data augmentation\nand clinical-relevant feature extraction.', 'In the Gastric Histopathology Image Classification (GHIC) tasks, which are\nusually weakly supervised learning missions, there is inevitably redundant\ninformation in the images. Therefore, designing networks that can focus on\neffective distinguishing features has become a popular research topic. In this\npaper, to accomplish the tasks of GHIC superiorly and to assist pathologists in\nclinical diagnosis, an intelligent Hierarchical Conditional Random Field based\nAttention Mechanism (HCRF-AM) model is proposed. The HCRF-AM model consists of\nan Attention Mechanism (AM) module and an Image Classification (IC) module. In\nthe AM module, an HCRF model is built to extract attention regions. In the IC\nmodule, a Convolutional Neural Network (CNN) model is trained with the\nattention regions selected and then an algorithm called Classification\nProbability-based Ensemble Learning is applied to obtain the image-level\nresults from patch-level output of the CNN. In the experiment, a classification\nspecificity of 96.67% is achieved on a gastric histopathology dataset with 700\nimages. Our HCRF-AM model demonstrates high classification performance and\nshows its effectiveness and future potential in the GHIC field.', ""Brain aging, and more specifically the difference between the chronological\nand the biological age of a person, may be a promising biomarker for\nidentifying neurodegenerative diseases. For this purpose accurate prediction is\nimportant but the localisation of the areas that play a significant role in the\nprediction is also crucial, in order to gain clinicians' trust and reassurance\nabout the performance of a prediction model. Most interpretability methods are\nfocused on classification tasks and cannot be directly transferred to\nregression tasks. In this study, we focus on the task of brain age regression\nfrom 3D brain Magnetic Resonance (MR) images using a Convolutional Neural\nNetwork, termed prediction model. We interpret its predictions by extracting\nimportance maps, which discover the parts of the brain that are the most\nimportant for brain age. In order to do so, we assume that voxels that are not\nuseful for the regression are resilient to noise addition. We implement a noise\nmodel which aims to add as much noise as possible to the input without harming\nthe performance of the prediction model. We average the importance maps of the\nsubjects and end up with a population-based importance map, which displays the\nregions of the brain that are influential for the task. We test our method on\n13,750 3D brain MR images from the UK Biobank, and our findings are consistent\nwith the existing neuropathology literature, highlighting that the hippocampus\nand the ventricles are the most relevant regions for brain aging."", 'Applying deep learning methods to mammography assessment has remained a\nchallenging topic. Dense noise with sparse expressions, mega-pixel raw data\nresolution, lack of diverse examples have all been factors affecting\nperformance. The lack of pixel-level ground truths have especially limited\nsegmentation methods in pushing beyond approximately bounding regions. We\npropose a classification approach grounded in high performance tissue\nassessment as an alternative to all-in-one localization and assessment models\nthat is also capable of pinpointing the causal pixels. First, the objective of\nthe mammography assessment task is formalized in the context of local tissue\nclassifiers. Then, the accuracy of a convolutional neural net is evaluated on\nclassifying patches of tissue with suspicious findings at varying scales, where\nhighest obtained AUC is above $0.9$. The local evaluations of one such expert\ntissue classifier is used to augment the results of a heatmap regression model\nand additionally recover the exact causal regions at high resolution as a\nsaliency image suitable for clinical settings.', ""Given image labels as the only supervisory signal, we focus on harvesting, or\nmining, thoracic disease localizations from chest X-ray images. Harvesting such\nlocalizations from existing datasets allows for the creation of improved data\nsources for computer-aided diagnosis and retrospective analyses. We train a\nconvolutional neural network (CNN) for image classification and propose an\nattention mining (AM) strategy to improve the model's sensitivity or saliency\nto disease patterns. The intuition of AM is that once the most salient disease\narea is blocked or hidden from the CNN model, it will pay attention to\nalternative image regions, while still attempting to make correct predictions.\nHowever, the model requires to be properly constrained during AM, otherwise, it\nmay overfit to uncorrelated image parts and forget the valuable knowledge that\nit has learned from the original image classification task. To alleviate such\nside effects, we then design a knowledge preservation (KP) loss, which\nminimizes the discrepancy between responses for X-ray images from the original\nand the updated networks. Furthermore, we modify the CNN model to include\nmulti-scale aggregation (MSA), improving its localization ability on\nsmall-scale disease findings, e.g., lung nodules. We experimentally validate\nour method on the publicly-available ChestX-ray14 dataset, outperforming a\nclass activation map (CAM)-based approach, and demonstrating the value of our\nnovel framework for mining disease locations."", 'Medical image datasets can have large number of images representing patients\nwith different health conditions and various disease severity. When dealing\nwith raw unlabeled image datasets, the large number of samples often makes it\nhard for experts and non-experts to understand the variety of images present in\na dataset. Supervised learning methods rely on labeled images which requires a\nconsiderable effort by medical experts to first understand the communities of\nimages present in the data and then labeling the images. Here, we propose an\nalgorithm to facilitate the automatic identification of communities in medical\nimage datasets. We further demonstrate that such analysis can be insightful in\na supervised setting when the images are already labeled. Such insights are\nuseful because, health and disease severity can be considered a continuous\nspectrum, and within each class, there usually are finer communities worthy of\ninvestigation, especially when they have similarities to communities in other\nclasses. In our approach, we use wavelet decomposition of images in tandem with\nspectral methods. We show that the eigenvalues of a graph Laplacian can reveal\nthe number of notable communities in an image dataset. Moreover, analyzing the\nsimilarities may be used to infer a spectrum representing the severity of the\ndisease. In our experiments, we use a dataset of images labeled with different\nconditions for COVID patients. We detect 25 communities in the dataset and then\nobserve that only 6 of those communities contain patients with pneumonia. We\nalso investigate the contents of a colorectal cancer histology dataset.', 'Convolutional neural networks (CNNs) show impressive performance for image\nclassification and detection, extending heavily to the medical image domain.\nNevertheless, medical experts are sceptical in these predictions as the\nnonlinear multilayer structure resulting in a classification outcome is not\ndirectly graspable. Recently, approaches have been shown which help the user to\nunderstand the discriminative regions within an image which are decisive for\nthe CNN to conclude to a certain class. Although these approaches could help to\nbuild trust in the CNNs predictions, they are only slightly shown to work with\nmedical image data which often poses a challenge as the decision for a class\nrelies on different lesion areas scattered around the entire image. Using the\nDiaretDB1 dataset, we show that on retina images different lesion areas\nfundamental for diabetic retinopathy are detected on an image level with high\naccuracy, comparable or exceeding supervised methods. On lesion level, we\nachieve few false positives with high sensitivity, though, the network is\nsolely trained on image-level labels which do not include information about\nexisting lesions. Classifying between diseased and healthy images, we achieve\nan AUC of 0.954 on the DiaretDB1.', 'With the increasing usage of radiograph images as a most common medical\nimaging system for diagnosis, treatment planning, and clinical studies, it is\nincreasingly becoming a vital factor to use machine learning-based systems to\nprovide reliable information for surgical pre-planning. Segmentation of pelvic\nbone in radiograph images is a critical preprocessing step for some\napplications such as automatic pose estimation and disease detection. However,\nthe encoder-decoder style network known as U-Net has demonstrated limited\nresults due to the challenging complexity of the pelvic shapes, especially in\nsevere patients. In this paper, we propose a novel multi-task segmentation\nmethod based on Mask R-CNN architecture. For training, the network weights were\ninitialized by large non-medical dataset and fine-tuned with radiograph images.\nFurthermore, in the training process, augmented data was generated to improve\nnetwork performance. Our experiments show that Mask R-CNN utilizing multi-task\nlearning, transfer learning, and data augmentation techniques achieve 0.96 DICE\ncoefficient, which significantly outperforms the U-Net. Notably, for a fair\ncomparison, the same transfer learning and data augmentation techniques have\nbeen used for U-net training.', 'Medical imaging is an essential tool in many areas of medical applications,\nused for both diagnosis and treatment. However, reading medical images and\nmaking diagnosis or treatment recommendations require specially trained medical\nspecialists. The current practice of reading medical images is labor-intensive,\ntime-consuming, costly, and error-prone. It would be more desirable to have a\ncomputer-aided system that can automatically make diagnosis and treatment\nrecommendations. Recent advances in deep learning enable us to rethink the ways\nof clinician diagnosis based on medical images. In this thesis, we will\nintroduce 1) mammograms for detecting breast cancers, the most frequently\ndiagnosed solid cancer for U.S. women, 2) lung CT images for detecting lung\ncancers, the most frequently diagnosed malignant cancer, and 3) head and neck\nCT images for automated delineation of organs at risk in radiotherapy. First,\nwe will show how to employ the adversarial concept to generate the hard\nexamples improving mammogram mass segmentation. Second, we will demonstrate how\nto use the weakly labeled data for the mammogram breast cancer diagnosis by\nefficiently design deep learning for multi-instance learning. Third, the thesis\nwill walk through DeepLung system which combines deep 3D ConvNets and GBM for\nautomated lung nodule detection and classification. Fourth, we will show how to\nuse weakly labeled data to improve existing lung nodule detection system by\nintegrating deep learning with a probabilistic graphic model. Lastly, we will\ndemonstrate the AnatomyNet which is thousands of times faster and more accurate\nthan previous methods on automated anatomy segmentation.', 'The recent progress of computing, machine learning, and especially deep\nlearning, for image recognition brings a meaningful effect for automatic\ndetection of various diseases from chest X-ray images (CXRs). Here efficiency\nof lung segmentation and bone shadow exclusion techniques is demonstrated for\nanalysis of 2D CXRs by deep learning approach to help radiologists identify\nsuspicious lesions and nodules in lung cancer patients. Training and validation\nwas performed on the original JSRT dataset (dataset #01), BSE-JSRT dataset,\ni.e. the same JSRT dataset, but without clavicle and rib shadows (dataset #02),\noriginal JSRT dataset after segmentation (dataset #03), and BSE-JSRT dataset\nafter segmentation (dataset #04). The results demonstrate the high efficiency\nand usefulness of the considered pre-processing techniques in the simplified\nconfiguration even. The pre-processed dataset without bones (dataset #02)\ndemonstrates the much better accuracy and loss results in comparison to the\nother pre-processed datasets after lung segmentation (datasets #02 and #03).', 'This study aims to automatically diagnose thoracic diseases depicted on the\nchest x-ray (CXR) images using deep convolutional neural networks. The existing\nmethods generally used the entire CXR images for training purposes, but this\nstrategy may suffer from two drawbacks. First, potential misalignment or the\nexistence of irrelevant objects in the entire CXR images may cause unnecessary\nnoise and thus limit the network performance. Second, the relatively low image\nresolution caused by the resizing operation, which is a common preprocessing\nprocedure for training neural networks, may lead to the loss of image details,\nmaking it difficult to detect pathologies with small lesion regions. To address\nthese issues, we present a novel method termed as segmentation-based deep\nfusion network (SDFN), which leverages the domain knowledge and the\nhigherresolution information of local lung regions. Specifically, the local\nlung regions were identified and cropped by the Lung Region Generator (LRG).\nTwo CNN-based classification models were then used as feature extractors to\nobtain the discriminative features of the entire CXR images and the cropped\nlung region images. Lastly, the obtained features were fused by the feature\nfusion module for disease classification. Evaluated by the NIH benchmark split\non the Chest X-ray 14 Dataset, our experimental result demonstrated that the\ndeveloped method achieved more accurate disease classification compared with\nthe available approaches via the receiver operating characteristic (ROC)\nanalyses. It was also found that the SDFN could localize the lesion regions\nmore precisely as compared to the traditional method.', 'Early detection of lung cancer is essential in reducing mortality. Recent\nstudies have demonstrated the clinical utility of low-dose computed tomography\n(CT) to detect lung cancer among individuals selected based on very limited\nclinical information. However, this strategy yields high false positive rates,\nwhich can lead to unnecessary and potentially harmful procedures. To address\nsuch challenges, we established a pipeline that co-learns from detailed\nclinical demographics and 3D CT images. Toward this end, we leveraged data from\nthe Consortium for Molecular and Cellular Characterization of Screen-Detected\nLesions (MCL), which focuses on early detection of lung cancer. A 3D\nattention-based deep convolutional neural net (DCNN) is proposed to identify\nlung cancer from the chest CT scan without prior anatomical location of the\nsuspicious nodule. To improve upon the non-invasive discrimination between\nbenign and malignant, we applied a random forest classifier to a dataset\nintegrating clinical information to imaging data. The results show that the AUC\nobtained from clinical demographics alone was 0.635 while the attention network\nalone reached an accuracy of 0.687. In contrast when applying our proposed\npipeline integrating clinical and imaging variables, we reached an AUC of 0.787\non the testing dataset. The proposed network both efficiently captures\nanatomical information for classification and also generates attention maps\nthat explain the features that drive performance.']","['Convolutional Neural Networks (CNNs) are state-of-the-art models for document\nimage classification tasks. However, many of these approaches rely on\nparameters and architectures designed for classifying natural images, which\ndiffer from document images. We question whether this is appropriate and\nconduct a large empirical study to find what aspects of CNNs most affect\nperformance on document images. Among other results, we exceed the\nstate-of-the-art on the RVL-CDIP dataset by using shear transform data\naugmentation and an architecture designed for a larger input image.\nAdditionally, we analyze the learned features and find evidence that CNNs\ntrained on RVL-CDIP learn region-specific layout features.', 'Recently deep learning methods, in particular, convolutional neural networks\n(CNNs), have led to a massive breakthrough in the range of computer vision.\nAlso, the large-scale annotated dataset is the essential key to a successful\ntraining procedure. However, it is a huge challenge to get such datasets in the\nmedical domain. Towards this, we present a data augmentation method for\ngenerating synthetic medical images using cycle-consistency Generative\nAdversarial Networks (GANs). We add semi-supervised attention modules to\ngenerate images with convincing details. We treat tumor images and normal\nimages as two domains. The proposed GANs-based model can generate a tumor image\nfrom a normal image, and in turn, it can also generate a normal image from a\ntumor image. Furthermore, we show that generated medical images can be used for\nimproving the performance of ResNet18 for medical image classification. Our\nmodel is applied to three limited datasets of tumor MRI images. We first\ngenerate MRI images on limited datasets, then we trained three popular\nclassification models to get the best model for tumor classification. Finally,\nwe train the classification model using real images with classic data\naugmentation methods and classification models using synthetic images. The\nclassification results between those trained models showed that the proposed\nSAG-GAN data augmentation method can boost Accuracy and AUC compare with\nclassic data augmentation methods. We believe the proposed data augmentation\nmethod can apply to other medical image domains, and improve the accuracy of\ncomputer-assisted diagnosis.', 'Segmentation of additive manufacturing (AM) defects in X-ray Computed\nTomography (XCT) images is challenging, due to the poor contrast, small sizes\nand variation in appearance of defects. Automatic segmentation can, however,\nprovide quality control for additive manufacturing. Over recent years,\nthree-dimensional convolutional neural networks (3D CNNs) have performed well\nin the volumetric segmentation of medical images. In this work, we leverage\ntechniques from the medical imaging domain and propose training a 3D U-Net\nmodel to automatically segment defects in XCT images of AM samples. This work\nnot only contributes to the use of machine learning for AM defect detection but\nalso demonstrates for the first time 3D volumetric segmentation in AM. We train\nand test with three variants of the 3D U-Net on an AM dataset, achieving a mean\nintersection of union (IOU) value of 88.4%.', 'The science of solving clinical problems by analyzing images generated in\nclinical practice is known as medical image analysis. The aim is to extract\ninformation in an effective and efficient manner for improved clinical\ndiagnosis. The recent advances in the field of biomedical engineering has made\nmedical image analysis one of the top research and development area. One of the\nreason for this advancement is the application of machine learning techniques\nfor the analysis of medical images. Deep learning is successfully used as a\ntool for machine learning, where a neural network is capable of automatically\nlearning features. This is in contrast to those methods where traditionally\nhand crafted features are used. The selection and calculation of these features\nis a challenging task. Among deep learning techniques, deep convolutional\nnetworks are actively used for the purpose of medical image analysis. This\ninclude application areas such as segmentation, abnormality detection, disease\nclassification, computer aided diagnosis and retrieval. In this study, a\ncomprehensive review of the current state-of-the-art in medical image analysis\nusing deep convolutional networks is presented. The challenges and potential of\nthese techniques are also highlighted.', 'Deep learning based medical image classifiers have shown remarkable prowess\nin various application areas like ophthalmology, dermatology, pathology, and\nradiology. However, the acceptance of these Computer-Aided Diagnosis (CAD)\nsystems in real clinical setups is severely limited primarily because their\ndecision-making process remains largely obscure. This work aims at elucidating\na deep learning based medical image classifier by verifying that the model\nlearns and utilizes similar disease-related concepts as described and employed\nby dermatologists. We used a well-trained and high performing neural network\ndeveloped by REasoning for COmplex Data (RECOD) Lab for classification of three\nskin tumours, i.e. Melanocytic Naevi, Melanoma and Seborrheic Keratosis and\nperformed a detailed analysis on its latent space. Two well established and\npublicly available skin disease datasets, PH2 and derm7pt, are used for\nexperimentation. Human understandable concepts are mapped to RECOD image\nclassification model with the help of Concept Activation Vectors (CAVs),\nintroducing a novel training and significance testing paradigm for CAVs. Our\nresults on an independent evaluation set clearly shows that the classifier\nlearns and encodes human understandable concepts in its latent representation.\nAdditionally, TCAV scores (Testing with CAVs) suggest that the neural network\nindeed makes use of disease-related concepts in the correct way when making\npredictions. We anticipate that this work can not only increase confidence of\nmedical practitioners on CAD but also serve as a stepping stone for further\ndevelopment of CAV-based neural network interpretation methods.', 'Chest X-ray is the most common medical imaging exam used to assess multiple\npathologies. Automated algorithms and tools have the potential to support the\nreading workflow, improve efficiency, and reduce reading errors. With the\navailability of large scale data sets, several methods have been proposed to\nclassify pathologies on chest X-ray images. However, most methods report\nperformance based on random image based splitting, ignoring the high\nprobability of the same patient appearing in both training and test set. In\naddition, most methods fail to explicitly incorporate the spatial information\nof abnormalities or utilize the high resolution images. We propose a novel\napproach based on location aware Dense Networks (DNetLoc), whereby we\nincorporate both high-resolution image data and spatial information for\nabnormality classification. We evaluate our method on the largest data set\nreported in the community, containing a total of 86,876 patients and 297,541\nchest X-ray images. We achieve (i) the best average AUC score for published\ntraining and test splits on the single benchmarking data set (ChestX-Ray14),\nand (ii) improved AUC scores when the pathology location information is\nexplicitly used. To foster future research we demonstrate the limitations of\nthe current benchmarking setup and provide new reference patient-wise splits\nfor the used data sets. This could support consistent and meaningful\nbenchmarking of future methods on the largest publicly available data sets.', ""The recent thrust on digital agriculture (DA) has renewed significant\nresearch interest in the automated delineation of agricultural fields. Most\nprior work addressing this problem have focused on detecting medium to large\nfields, while there is strong evidence that around 40\\% of the fields\nworld-wide and 70% of the fields in Asia and Africa are small. The lack of\nadequate labeled images for small fields, huge variations in their color,\ntexture, and shape, and faint boundary lines separating them make it difficult\nto develop an end-to-end learning model for detecting such fields. Hence, in\nthis paper, we present a multi-stage approach that uses a combination of\nmachine learning and image processing techniques. In the first stage, we\nleverage state-of-the-art edge detection algorithms such as holistically-nested\nedge detection (HED) to extract first-level contours and polygons. In the\nsecond stage, we propose image-processing techniques to identify polygons that\nare non-fields, over-segmentations, or noise and eliminate them. The next stage\ntackles under-segmentations using a combination of a novel ``cut-point'' based\ntechnique and localized second-level edge detection to obtain individual\nparcels. Since a few small, non-cropped but vegetated or constructed pockets\ncan be interspersed in areas that are predominantly croplands, in the final\nstage, we train a classifier for identifying each parcel from the previous\nstage as an agricultural field or not. In an evaluation using high-resolution\nimagery, we show that our approach has a high F-Score of 0.84 in areas with\nlarge fields and reasonable accuracy with an F-Score of 0.73 in areas with\nsmall fields, which is encouraging."", 'Document image classification remains a popular research area because it can\nbe commercialized in many enterprise applications across different industries.\nRecent advancements in large pre-trained computer vision and language models\nand graph neural networks has lent document image classification many tools.\nHowever using large pre-trained models usually requires substantial computing\nresources which could defeat the cost-saving advantages of automatic document\nimage classification. In the paper we propose an efficient document image\nclassification framework that uses graph convolution neural networks and\nincorporates textual, visual and layout information of the document. We have\nrigorously benchmarked our proposed algorithm against several state-of-art\nvision and language models on both publicly available dataset and a real-life\ninsurance document classification dataset. Empirical results on both publicly\navailable and real-world data show that our methods achieve near SOTA\nperformance yet require much less computing resources and time for model\ntraining and inference. This results in solutions than offer better cost\nadvantages, especially in scalable deployment for enterprise applications. The\nresults showed that our algorithm can achieve classification performance quite\nclose to SOTA. We also provide comprehensive comparisons of computing\nresources, model sizes, train and inference time between our proposed methods\nand baselines. In addition we delineate the cost per image using our method and\nother baselines.', 'Deep neural network models used for medical image segmentation are large\nbecause they are trained with high-resolution three-dimensional (3D) images.\nGraphics processing units (GPUs) are widely used to accelerate the trainings.\nHowever, the memory on a GPU is not large enough to train the models. A popular\napproach to tackling this problem is patch-based method, which divides a large\nimage into small patches and trains the models with these small patches.\nHowever, this method would degrade the segmentation quality if a target object\nspans multiple patches. In this paper, we propose a novel approach for 3D\nmedical image segmentation that utilizes the data-swapping, which swaps out\nintermediate data from GPU memory to CPU memory to enlarge the effective GPU\nmemory size, for training high-resolution 3D medical images without patching.\nWe carefully tuned parameters in the data-swapping method to obtain the best\ntraining performance for 3D U-Net, a widely used deep neural network model for\nmedical image segmentation. We applied our tuning to train 3D U-Net with\nfull-size images of 192 x 192 x 192 voxels in brain tumor dataset. As a result,\ncommunication overhead, which is the most important issue, was reduced by\n17.1%. Compared with the patch-based method for patches of 128 x 128 x 128\nvoxels, our training for full-size images achieved improvement on the mean Dice\nscore by 4.48% and 5.32 % for detecting whole tumor sub-region and tumor core\nsub-region, respectively. The total training time was reduced from 164 hours to\n47 hours, resulting in 3.53 times of acceleration.', 'We present in this paper a new approach for the automatic annotation of\nmedical images, using the approach of ""bag-of-words"" to represent the visual\ncontent of the medical image combined with text descriptors based approach\ntf.idf and reduced by latent semantic to extract the co-occurrence between\nterms and visual terms. A medical report is composed of a text describing a\nmedical image. First, we are interested to index the text and extract all\nrelevant terms using a thesaurus containing MeSH medical concepts. In a second\nphase, the medical image is indexed while recovering areas of interest which\nare invariant to change in scale, light and tilt. To annotate a new medical\nimage, we use the approach of ""bagof-words"" to recover the feature vector.\nIndeed, we use the vector space model to retrieve similar medical image from\nthe database training. The calculation of the relevance value of an image to\nthe query image is based on the cosine function. We conclude with an experiment\ncarried out on five types of radiological imaging to evaluate the performance\nof our system of medical annotation. The results showed that our approach works\nbetter with more images from the radiology of the skull.', 'With the ongoing rise of machine learning, the need for methods for\nexplaining decisions made by artificial intelligence systems is becoming a more\nand more important topic. Especially for image classification tasks, many\nstate-of-the-art tools to explain such classifiers rely on visual highlighting\nof important areas of the input data. Contrary, counterfactual explanation\nsystems try to enable a counterfactual reasoning by modifying the input image\nin a way such that the classifier would have made a different prediction. By\ndoing so, the users of counterfactual explanation systems are equipped with a\ncompletely different kind of explanatory information. However, methods for\ngenerating realistic counterfactual explanations for image classifiers are\nstill rare. Especially in medical contexts, where relevant information often\nconsists of textural and structural information, high-quality counterfactual\nimages have the potential to give meaningful insights into decision processes.\nIn this work, we present GANterfactual, an approach to generate such\ncounterfactual image explanations based on adversarial image-to-image\ntranslation techniques. Additionally, we conduct a user study to evaluate our\napproach in an exemplary medical use case. Our results show that, in the chosen\nmedical use-case, counterfactual explanations lead to significantly better\nresults regarding mental models, explanation satisfaction, trust, emotions, and\nself-efficacy than two state-of-the-art systems that work with saliency maps,\nnamely LIME and LRP.', 'Convolutional Neural Networks (CNNs) have shown to be powerful medical image\nsegmentation models. In this study, we address some of the main unresolved\nissues regarding these models. Specifically, training of these models on small\nmedical image datasets is still challenging, with many studies promoting\ntechniques such as transfer learning. Moreover, these models are infamous for\nproducing over-confident predictions and for failing silently when presented\nwith out-of-distribution (OOD) data at test time. In this paper, we advocate\nfor multi-task learning, i.e., training a single model on several different\ndatasets, spanning several different organs of interest and different imaging\nmodalities. We show that not only a single CNN learns to automatically\nrecognize the context and accurately segment the organ of interest in each\ncontext, but also that such a joint model often has more accurate and\nbetter-calibrated predictions than dedicated models trained separately on each\ndataset. Our experiments show that multi-task learning can outperform transfer\nlearning in medical image segmentation tasks. For detecting OOD data, we\npropose a method based on spectral analysis of CNN feature maps. We show that\ndifferent datasets, representing different imaging modalities and/or different\norgans of interest, have distinct spectral signatures, which can be used to\nidentify whether or not a test image is similar to the images used to train a\nmodel. We show that this approach is far more accurate than OOD detection based\non prediction uncertainty. The methods proposed in this paper contribute\nsignificantly to improving the accuracy and reliability of CNN-based medical\nimage segmentation models.', 'Celiac Disease (CD) and Environmental Enteropathy (EE) are common causes of\nmalnutrition and adversely impact normal childhood development. Both conditions\nrequire a tissue biopsy for diagnosis and a major challenge of interpreting\nclinical biopsy images to differentiate between these gastrointestinal diseases\nis striking histopathologic overlap between them. In the current study, we\npropose four diagnosis techniques for these diseases and address their\nlimitations and advantages. First, the diagnosis between CD, EE, and Normal\nbiopsies is considered, but the main challenge with this diagnosis technique is\nthe staining problem. The dataset used in this research is collected from\ndifferent centers with different staining standards. To solve this problem, we\nuse color balancing in order to train our model with a varying range of colors.\nRandom Multimodel Deep Learning (RMDL) architecture has been used as another\napproach to mitigate the effects of the staining problem. RMDL combines\ndifferent architectures and structures of deep learning and the final output of\nthe model is based on the majority vote. CD is a chronic autoimmune disease\nthat affects the small intestine genetically predisposed children and adults.\nTypically, CD rapidly progress from Marsh I to IIIa. Marsh III is sub-divided\ninto IIIa (partial villus atrophy), Marsh IIIb (subtotal villous atrophy), and\nMarsh IIIc (total villus atrophy) to explain the spectrum of villus atrophy\nalong with crypt hypertrophy and increased intraepithelial lymphocytes. In the\nsecond part of this study, we proposed two ways for diagnosing different stages\nof CD. Finally, in the third part of this study, these two steps are combined\nas Hierarchical Medical Image Classification (HMIC) to have a model to diagnose\nthe disease data hierarchically.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'Segmentation is one of the most important and popular tasks in medical image\nanalysis, which plays a critical role in disease diagnosis, surgical planning,\nand prognosis evaluation. During the past five years, on the one hand,\nthousands of medical image segmentation methods have been proposed for various\norgans and lesions in different medical images, which become more and more\nchallenging to fairly compare different methods. On the other hand,\ninternational segmentation challenges can provide a transparent platform to\nfairly evaluate and compare different methods. In this paper, we present a\ncomprehensive review of the top methods in ten 3D medical image segmentation\nchallenges during 2020, covering a variety of tasks and datasets. We also\nidentify the ""happy-families"" practices in the cutting-edge segmentation\nmethods, which are useful for developing powerful segmentation approaches.\nFinally, we discuss open research problems that should be addressed in the\nfuture. We also maintain a list of cutting-edge segmentation methods at\n\\url{https://github.com/JunMa11/SOTA-MedSeg}.', 'Direct automatic segmentation of objects from 3D medical imaging, such as\nmagnetic resonance (MR) imaging, is challenging as it often involves accurately\nidentifying a number of individual objects with complex geometries within a\nlarge volume under investigation. To address these challenges, most deep\nlearning approaches typically enhance their learning capability by\nsubstantially increasing the complexity or the number of trainable parameters\nwithin their models. Consequently, these models generally require long\ninference time on standard workstations operating clinical MR systems and are\nrestricted to high-performance computing hardware due to their large memory\nrequirement. Further, to fit 3D dataset through these large models using\nlimited computer memory, trade-off techniques such as patch-wise training are\noften used which sacrifice the fine-scale geometric information from input\nimages which could be clinically significant for diagnostic purposes. To\naddress these challenges, we present a compact convolutional neural network\nwith a shallow memory footprint to efficiently reduce the number of model\nparameters required for state-of-art performance. This is critical for\npractical employment as most clinical environments only have low-end hardware\nwith limited computing power and memory. The proposed network can maintain data\nintegrity by directly processing large full-size 3D input volumes with no\npatches required and significantly reduces the computational time required for\nboth training and inference. We also propose a novel loss function with extra\nshape constraint to improve the accuracy for imbalanced classes in 3D MR\nimages.', ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'Content-based image retrieval (CBIR) of medical images is a crucial task that\ncan contribute to a more reliable diagnosis if applied to big data. Recent\nadvances in feature extraction and classification have enormously improved CBIR\nresults for digital images. However, considering the increasing accessibility\nof big data in medical imaging, we are still in need of reducing both memory\nrequirements and computational expenses of image retrieval systems. This work\nproposes to exclude the features of image blocks that exhibit a low encoding\nerror when learned by a $n/p/n$ autoencoder ($p\\!<\\!n$). We examine the\nhistogram of autoendcoding errors of image blocks for each image class to\nfacilitate the decision which image regions, or roughly what percentage of an\nimage perhaps, shall be declared relevant for the retrieval task. This leads to\nreduction of feature dimensionality and speeds up the retrieval process. To\nvalidate the proposed scheme, we employ local binary patterns (LBP) and support\nvector machines (SVM) which are both well-established approaches in CBIR\nresearch community. As well, we use IRMA dataset with 14,410 x-ray images as\ntest data. The results show that the dimensionality of annotated feature\nvectors can be reduced by up to 50% resulting in speedups greater than 27% at\nexpense of less than 1% decrease in the accuracy of retrieval when validating\nthe precision and recall of the top 20 hits.', 'Deep learning techniques have led to state-of-the-art image super resolution\nwith natural images. Normally, pairs of high-resolution and low-resolution\nimages are used to train the deep learning models. These techniques have also\nbeen applied to medical image super-resolution. The characteristics of medical\nimages differ significantly from natural images in several ways. First, it is\ndifficult to obtain high-resolution images for training in real clinical\napplications due to the limitations of imaging systems and clinical\nrequirements. Second, other modal high-resolution images are available (e.g.,\nhigh-resolution T1-weighted images are available for enhancing low-resolution\nT2-weighted images). In this paper, we propose an unsupervised image\nsuper-resolution technique based on simple prior knowledge of the human\nanatomy. This technique does not require target T2WI high-resolution images for\ntraining. Furthermore, we present a guided residual dense network, which\nincorporates a residual dense network with a guided deep convolutional neural\nnetwork for enhancing the resolution of low-resolution images by referring to\ndifferent modal high-resolution images of the same subject. Experiments on a\npublicly available brain MRI database showed that our proposed method achieves\nbetter performance than the state-of-the-art methods.', ""In safety-critical applications like medical diagnosis, certainty associated\nwith a model's prediction is just as important as its accuracy. Consequently,\nuncertainty estimation and reduction play a crucial role. Uncertainty in\npredictions can be attributed to noise or randomness in data (aleatoric) and\nincorrect model inferences (epistemic). While model uncertainty can be reduced\nwith more data or bigger models, aleatoric uncertainty is more intricate. This\nwork proposes a novel approach that interprets data uncertainty estimated from\na self-supervised task as noise inherent to the data and utilizes it to reduce\naleatoric uncertainty in another task related to the same dataset via data\naugmentation. The proposed method was evaluated on a benchmark medical imaging\ndataset with image reconstruction as the self-supervised task and segmentation\nas the image analysis task. Our findings demonstrate the effectiveness of the\nproposed approach in significantly reducing the aleatoric uncertainty in the\nimage segmentation task while achieving better or on-par performance compared\nto the standard augmentation techniques."", 'Spatial attention has been introduced to convolutional neural networks (CNNs)\nfor improving both their performance and interpretability in visual tasks\nincluding image classification. The essence of the spatial attention is to\nlearn a weight map which represents the relative importance of activations\nwithin the same layer or channel. All existing attention mechanisms are local\nattentions in the sense that weight maps are image-specific. However, in the\nmedical field, there are cases that all the images should share the same weight\nmap because the set of images record the same kind of symptom related to the\nsame object and thereby share the same structural content. In this paper, we\nthus propose a novel global spatial attention mechanism in CNNs mainly for\nmedical image classification. The global weight map is instantiated by a\ndecision boundary between important pixels and unimportant pixels. And we\npropose to realize the decision boundary by a binary classifier in which the\nintensities of all images at a pixel are the features of the pixel. The binary\nclassification is integrated into an image classification CNN and is to be\noptimized together with the CNN. Experiments on two medical image datasets and\none facial expression dataset showed that with the proposed attention, not only\nthe performance of four powerful CNNs which are GoogleNet, VGG, ResNet, and\nDenseNet can be improved, but also meaningful attended regions can be obtained,\nwhich is beneficial for understanding the content of images of a domain.', 'We propose a causal hidden Markov model to achieve robust prediction of\nirreversible disease at an early stage, which is safety-critical and vital for\nmedical treatment in early stages. Specifically, we introduce the hidden\nvariables which propagate to generate medical data at each time step. To avoid\nlearning spurious correlation (e.g., confounding bias), we explicitly separate\nthese hidden variables into three parts: a) the disease (clinical)-related\npart; b) the disease (non-clinical)-related part; c) others, with only a),b)\ncausally related to the disease however c) may contain spurious correlations\n(with the disease) inherited from the data provided. With personal attributes\nand the disease label respectively provided as side information and\nsupervision, we prove that these disease-related hidden variables can be\ndisentangled from others, implying the avoidance of spurious correlation for\ngeneralization to medical data from other (out-of-) distributions. Guaranteed\nby this result, we propose a sequential variational auto-encoder with a\nreformulated objective function. We apply our model to the early prediction of\nperipapillary atrophy and achieve promising results on out-of-distribution test\ndata. Further, the ablation study empirically shows the effectiveness of each\ncomponent in our method. And the visualization shows the accurate\nidentification of lesion regions from others.', 'Coronavirus disease (COVID-19) is an infectious disease caused by a newly\ndiscovered coronavirus. The disease presents with symptoms such as shortness of\nbreath, fever, dry cough, and chronic fatigue, amongst others. Sometimes the\nsymptoms of the disease increase so much they lead to the death of the\npatients. The disease may be asymptomatic in some patients in the early stages,\nwhich can lead to increased transmission of the disease to others. Many studies\nhave tried to use medical imaging for early diagnosis of COVID-19. This study\nattempts to review papers on automatic methods for medical image analysis and\ndiagnosis of COVID-19. For this purpose, PubMed, Google Scholar, arXiv and\nmedRxiv were searched to find related studies by the end of April 2020, and the\nessential points of the collected studies were summarised. The contribution of\nthis study is four-fold: 1) to use as a tutorial of the field for both\nclinicians and technologists, 2) to comprehensively review the characteristics\nof COVID-19 as presented in medical images, 3) to examine automated artificial\nintelligence-based approaches for COVID-19 diagnosis based on the accuracy and\nthe method used, 4) to express the research limitations in this field and the\nmethods used to overcome them. COVID-19 reveals signs in medical images can be\nused for early diagnosis of the disease even in asymptomatic patients. Using\nautomated machine learning-based methods can diagnose the disease with high\naccuracy from medical images and reduce time, cost and error of diagnostic\nprocedure. It is recommended to collect bulk imaging data from patients in the\nshortest possible time to improve the performance of COVID-19 automated\ndiagnostic methods.', 'Medical images can be a valuable resource for reliable information to support\nmedical diagnosis. However, the large volume of medical images makes it\nchallenging to retrieve relevant information given a particular scenario. To\nsolve this challenge, content-based image retrieval (CBIR) attempts to\ncharacterize images (or image regions) with invariant content information in\norder to facilitate image search. This work presents a feature extraction\ntechnique for medical images using stacked autoencoders, which encode images to\nbinary vectors. The technique is applied to the IRMA dataset, a collection of\n14,410 x-ray images in order to demonstrate the ability of autoencoders to\nretrieve similar x-rays given test queries. Using IRMA dataset as a benchmark,\nit was found that stacked autoencoders gave excellent results with a retrieval\nerror of 376 for 1,733 test images with a compression of 74.61%.', 'Image-to-image translation is an ill-posed problem as unique one-to-one\nmapping may not exist between the source and target images. Learning-based\nmethods proposed in this context often evaluate the performance on test data\nthat is similar to the training data, which may be impractical. This demands\nrobust methods that can quantify uncertainty in the prediction for making\ninformed decisions, especially for critical areas such as medical imaging.\nRecent works that employ conditional generative adversarial networks (GANs)\nhave shown improved performance in learning photo-realistic image-to-image\nmappings between the source and the target images. However, these methods do\nnot focus on (i)~robustness of the models to out-of-distribution (OOD)-noisy\ndata and (ii)~uncertainty quantification. This paper proposes a GAN-based\nframework that (i)~models an adaptive loss function for robustness to OOD-noisy\ndata that automatically tunes the spatially varying norm for penalizing the\nresiduals and (ii)~estimates the per-voxel uncertainty in the predictions. We\ndemonstrate our method on two key applications in medical imaging:\n(i)~undersampled magnetic resonance imaging (MRI) reconstruction (ii)~MRI\nmodality propagation. Our experiments with two different real-world datasets\nshow that the proposed method (i)~is robust to OOD-noisy test data and provides\nimproved accuracy and (ii)~quantifies voxel-level uncertainty in the\npredictions.', 'Content-based image retrieval (CBIR) of medical images in large datasets to\nidentify similar images when a query image is given can be very useful in\nimproving the diagnostic decision of the clinical experts and as well in\neducational scenarios. In this paper, we used two stage classification and\nretrieval approach to retrieve similar images. First, the Gabor filters are\napplied to Radon-transformed images to extract features and to train a\nmulti-class SVM. Then based on the classification results and using an\nextracted Gabor barcode, similar images are retrieved. The proposed method was\ntested on IRMA dataset which contains more than 14,000 images. Experimental\nresults show the efficiency of our approach in retrieving similar images\ncompared to other Gabor-Radon-oriented methods.', 'Deep learning-based computer-aided diagnosis is gradually deployed to review\nand analyze medical images. However, this paradigm is restricted in real-world\nclinical applications due to the poor robustness and generalization. The issue\nis more sinister with a lack of training data. In this paper, we address the\nchallenge from the transfer learning point of view. Different from the common\nsetting that transferring knowledge from the natural image domain to the\nmedical image domain, we find the knowledge from the same domain further boosts\nthe model robustness and generalization. Therefore, we propose a novel\ntwo-stage framework for robust generalized medical image segmentation. Firstly,\nan unsupervised tile-wise autoencoder pretraining architecture is proposed to\nlearn local and global knowledge. Secondly, the downstream segmentation model\ncoupled with an auxiliary reconstruction network is designed. The\nreconstruction branch encourages the model to capture more general semantic\nfeatures. Experiments of lung segmentation on multi chest X-ray datasets are\nconducted. Comprehensive results demonstrate the superior robustness of the\nproposed framework to corruption and high generalization performance on unseen\ndatasets, especially under the scenario of the limited training data.', 'Since the low quality of document images will greatly undermine the chances\nof success in automatic text recognition and analysis, it is necessary to\nassess the quality of document images uploaded in online business process, so\nas to reject those images of low quality. In this paper, we attempt to achieve\ndocument image quality assessment and our contributions are twofold. Firstly,\nsince document image quality assessment is more interested in text, we propose\na text line based framework to estimate document image quality, which is\ncomposed of three stages: text line detection, text line quality prediction,\nand overall quality assessment. Text line detection aims to find potential text\nlines with a detector. In the text line quality prediction stage, the quality\nscore is computed for each text line with a CNN-based prediction model. The\noverall quality of document images is finally assessed with the ensemble of all\ntext line quality. Secondly, to train the prediction model, a large-scale\ndataset, comprising 52,094 text line images, is synthesized with diverse\nattributes. For each text line image, a quality label is computed with a\npiece-wise function. To demonstrate the effectiveness of the proposed\nframework, comprehensive experiments are evaluated on two popular document\nimage quality assessment benchmarks. Our framework significantly outperforms\nthe state-of-the-art methods by large margins on the large and complicated\ndataset.', 'There is a need for information retrieval from large collections of\nlow-resolution (LR) binary document images, which can be found in digital\nlibraries across the world, where the high-resolution (HR) counterpart is not\navailable. This gives rise to the problem of binary document image\nsuper-resolution (BDISR). The objective of this paper is to address the\ninteresting and challenging problem of super resolution of binary Tamil\ndocument images for improved readability and better optical character\nrecognition (OCR). We propose multiple deep neural network architectures to\naddress this problem and analyze their performance. The proposed models are all\nsingle image super-resolution techniques, which learn a generalized spatial\ncorrespondence between the LR and HR binary document images. We employ\nconvolutional layers for feature extraction followed by transposed convolution\nand sub-pixel convolution layers for upscaling the features. Since the outputs\nof the neural networks are gray scale, we utilize the advantage of power law\ntransformation as a post-processing technique to improve the character level\npixel connectivity. The performance of our models is evaluated by comparing the\nOCR accuracies and the mean opinion scores given by human evaluators on LR\nimages and the corresponding model-generated HR images.', 'We consider the problem of accurately and efficiently querying a remote\nserver to retrieve information about images captured by a mobile device. In\naddition to reduced transmission overhead and computational complexity, the\nretrieval protocol should be robust to variations in the image acquisition\nprocess, such as translation, rotation, scaling, and sensor-related\ndifferences. We propose to extract scale-invariant image features and then\nperform clustering to reduce the number of features needed for image matching.\nPrincipal Component Analysis (PCA) and Non-negative Matrix Factorization (NMF)\nare investigated as candidate clustering approaches. The image matching\ncomplexity at the database server is quadratic in the (small) number of\nclusters, not in the (very large) number of image features. We employ an\nimage-dependent information content metric to approximate the model order,\ni.e., the number of clusters, needed for accurate matching, which is preferable\nto setting the model order using trial and error. We show how to combine the\nhypotheses provided by PCA and NMF factor loadings, thereby obtaining more\naccurate retrieval than using either approach alone. In experiments on a\ndatabase of urban images, we obtain a top-1 retrieval accuracy of 89% and a\ntop-3 accuracy of 92.5%.', 'Face recognition approaches often rely on equal image resolution for\nverifying faces on two images. However, in practical applications, those image\nresolutions are usually not in the same range due to different image capture\nmechanisms or sources. In this work, we first analyze the impact of image\nresolutions on face verification performance with a state-of-the-art face\nrecognition model. For images synthetically reduced to $5\\,\\times\\,5$ px\nresolution, the verification performance drops from $99.23\\%$ increasingly down\nto almost $55\\%$. Especially for cross-resolution image pairs (one high- and\none low-resolution image), the verification accuracy decreases even further. We\ninvestigate this behavior more in-depth by looking at the feature distances for\nevery 2-image test pair. To tackle this problem, we propose the following two\nmethods: 1) Train a state-of-the-art face-recognition model straightforwardly\nwith $50\\%$ low-resolution images directly within each batch. 2) Train a\nsiamese-network structure and add a cosine distance feature loss between high-\nand low-resolution features. Both methods show an improvement for\ncross-resolution scenarios and can increase the accuracy at very low resolution\nto approximately $70\\%$. However, a disadvantage is that a specific model needs\nto be trained for every resolution pair. Thus, we extend the aforementioned\nmethods by training them with multiple image resolutions at once. The\nperformances for particular testing image resolutions are slightly worse, but\nthe advantage is that this model can be applied to arbitrary resolution images\nand achieves an overall better performance ($97.72\\%$ compared to $96.86\\%$).\nDue to the lack of a benchmark for arbitrary resolution images for the\ncross-resolution and equal-resolution task, we propose an evaluation protocol\nfor five well-known datasets, focusing on high, mid, and low-resolution images.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'This work analyses the impact of self-supervised pre-training on document\nimages in the context of document image classification. While previous\napproaches explore the effect of self-supervision on natural images, we show\nthat patch-based pre-training performs poorly on document images because of\ntheir different structural properties and poor intra-sample semantic\ninformation. We propose two context-aware alternatives to improve performance\non the Tobacco-3482 image classification task. We also propose a novel method\nfor self-supervision, which makes use of the inherent multi-modality of\ndocuments (image and text), which performs better than other popular\nself-supervised methods, including supervised ImageNet pre-training, on\ndocument image classification scenarios with a limited amount of data.', 'For more than two decades, research has been performed on content-based image\nretrieval (CBIR). By combining Radon projections and the support vector\nmachines (SVM), a content-based medical image retrieval method is presented in\nthis work. The proposed approach employs the normalized Radon projections with\ncorresponding image category labels to build an SVM classifier, and the Radon\nbarcode database which encodes every image in a binary format is also generated\nsimultaneously to tag all images. To retrieve similar images when a query image\nis given, Radon projections and the barcode of the query image are generated.\nSubsequently, the k-nearest neighbor search method is applied to find the\nimages with minimum Hamming distance of the Radon barcode within the same class\npredicted by the trained SVM classifier that uses Radon features. The\nperformance of the proposed method is validated by using the IRMA 2009 dataset\nwith 14,410 x-ray images in 57 categories. The results demonstrate that our\nmethod has the capacity to retrieve similar responses for the correctly\nidentified query image and even for those mistakenly classified by SVM. The\napproach further is very fast and has low memory requirement.', 'Medical images constitute a source of information essential for disease\ndiagnosis, treatment and follow-up. In addition, due to its patient-specific\nnature, imaging information represents a critical component required for\nadvancing precision medicine into clinical practice. This manuscript describes\nrecently developed technologies for better handling of image information:\nphotorealistic visualization of medical images with Cinematic Rendering,\nartificial agents for in-depth image understanding, support for minimally\ninvasive procedures, and patient-specific computational models with enhanced\npredictive power. Throughout the manuscript we will analyze the capabilities of\nsuch technologies and extrapolate on their potential impact to advance the\nquality of medical care, while reducing its cost.', 'Indeed, these are exciting times. We are in the heart of a digital\nrenaissance. Automation and computer technology allow engineers and scientists\nto fabricate processes that amalgamate quality of life. We anticipate much\ngrowth in medical image interpretation and understanding, due to the influx of\ncomputer technologies. This work should serve as a guide to introduce the\nreader to core themes in theoretical computer science, as well as imaging\napplications for understanding vocal-fold vibrations. In this work, we motivate\nthe use of automation, review some mathematical models of computation. We\npresent a proof of a classical problem in image analysis that cannot be\nautomated by means of algorithms. Furthermore, discuss some applications for\nprocessing medical images of the vocal folds, and discuss some of the\nexhilarating directions the art of automation will take vocal-fold image\ninterpretation and quite possibly other areas of biomedical image analysis.', 'Recognition of document images have important applications in restoring old\nand classical texts. The problem involves quality improvement before passing it\nto a properly trained OCR to get accurate recognition of the text. The image\nenhancement and quality improvement constitute important steps as subsequent\nrecognition depends upon the quality of the input image. There are scenarios\nwhen high resolution images are not available and our experiments show that the\nOCR accuracy reduces significantly with decrease in the spatial resolution of\ndocument images. Thus the only option is to improve the resolution of such\ndocument images. The goal is to construct a high resolution image, given a\nsingle low resolution binary image, which constitutes the problem of single\nimage super-resolution. Most of the previous work in super-resolution deal with\nnatural images which have more information-content than the document images.\nHere, we use Convolution Neural Network to learn the mapping between low and\nthe corresponding high resolution images. We experiment with different number\nof layers, parameter settings and non-linear functions to build a fast\nend-to-end framework for document image super-resolution. Our proposed model\nshows a very good PSNR improvement of about 4 dB on 75 dpi Tamil images,\nresulting in a 3 % improvement of word level accuracy by the OCR. It takes less\ntime than the recent sparse based natural image super-resolution technique,\nmaking it useful for real-time document recognition applications.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'Landmark detection algorithms trained on high resolution images perform\npoorly on datasets containing low resolution images. This deters the\nperformance of algorithms relying on quality landmarks, for example, face\nrecognition. To the best of our knowledge, there does not exist any dataset\nconsisting of low resolution face images along with their annotated landmarks,\nmaking supervised training infeasible. In this paper, we present a\nsemi-supervised approach to predict landmarks on low resolution images by\nlearning them from labeled high resolution images. The objective of this work\nis to show that predicting landmarks directly on low resolution images is more\neffective than the current practice of aligning images after rescaling or\nsuperresolution. In a two-step process, the proposed approach first learns to\ngenerate low resolution images by modeling the distribution of target low\nresolution images. In the second stage, the roles of generated images and real\nlow resolution images are switched and the model learns to predict landmarks\nfor real low resolution images from generated low resolution images. With\nextensive experimentation, we study the impact of each of the design choices\nand also show that prediction of landmarks directly on low resolution images\nimproves the performance of important tasks such as face recognition in low\nresolution images.', 'This paper is on image and face super-resolution. The vast majority of prior\nwork for this problem focus on how to increase the resolution of low-resolution\nimages which are artificially generated by simple bilinear down-sampling (or in\na few cases by blurring followed by down-sampling).We show that such methods\nfail to produce good results when applied to real-world low-resolution, low\nquality images. To circumvent this problem, we propose a two-stage process\nwhich firstly trains a High-to-Low Generative Adversarial Network (GAN) to\nlearn how to degrade and downsample high-resolution images requiring, during\ntraining, only unpaired high and low-resolution images. Once this is achieved,\nthe output of this network is used to train a Low-to-High GAN for image\nsuper-resolution using this time paired low- and high-resolution images. Our\nmain result is that this network can be now used to efectively increase the\nquality of real-world low-resolution images. We have applied the proposed\npipeline for the problem of face super-resolution where we report large\nimprovement over baselines and prior work although the proposed method is\npotentially applicable to other object categories.', 'Many applications for the automated diagnosis of plant disease have been\ndeveloped based on the success of deep learning techniques. However, these\napplications often suffer from overfitting, and the diagnostic performance is\ndrastically decreased when used on test datasets from new environments. In this\npaper, we propose LeafGAN, a novel image-to-image translation system with own\nattention mechanism. LeafGAN generates a wide variety of diseased images via\ntransformation from healthy images, as a data augmentation tool for improving\nthe performance of plant disease diagnosis. Thanks to its own attention\nmechanism, our model can transform only relevant areas from images with a\nvariety of backgrounds, thus enriching the versatility of the training images.\nExperiments with five-class cucumber disease classification show that data\naugmentation with vanilla CycleGAN cannot help to improve the generalization,\ni.e., disease diagnostic performance increased by only 0.7% from the baseline.\nIn contrast, LeafGAN boosted the diagnostic performance by 7.4%. We also\nvisually confirmed the generated images by our LeafGAN were much better quality\nand more convincing than those generated by vanilla CycleGAN. The code is\navailable publicly at: https://github.com/IyatomiLab/LeafGAN.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'Document Image Analysis, like any Digital Image Analysis requires\nidentification and extraction of proper features, which are generally extracted\nfrom uncompressed images, though in reality images are made available in\ncompressed form for the reasons such as transmission and storage efficiency.\nHowever, this implies that the compressed image should be decompressed, which\nindents additional computing resources. This limitation induces the motivation\nto research in extracting features directly from the compressed image. In this\nresearch, we propose to extract essential features such as projection profile,\nrun-histogram and entropy for text document analysis directly from run-length\ncompressed text-documents. The experimentation illustrates that features are\nextracted directly from the compressed image without going through the stage of\ndecompression, because of which the computing time is reduced. The feature\nvalues so extracted are exactly identical to those extracted from uncompressed\nimages.', 'Obtaining datasets labeled to facilitate model development is a challenge for\nmost machine learning tasks. The difficulty is heightened for medical imaging,\nwhere data itself is limited in accessibility and labeling requires costly time\nand effort by trained medical specialists. Medical imaging studies, however,\nare often accompanied by a medical report produced by a radiologist,\nidentifying important features on the corresponding scan for other physicians\nnot specifically trained in radiology. We propose a methodology for\napproximating image-level labels for radiology studies from associated reports\nusing a general purpose language processing tool for medical concept extraction\nand sentiment analysis, and simple manually crafted heuristics for false\npositive reduction. Using this approach, we label more than 175,000 Head CT\nstudies for the presence of 33 features indicative of 11 clinically relevant\nconditions. For 27 of the 30 keywords that yielded positive results (3 had no\noccurrences), the lower bound of the confidence intervals created to estimate\nthe percentage of accurately labeled reports was above 85%, with the average\nbeing above 95%. Though noisier then manual labeling, these results suggest\nthis method to be a viable means of labeling medical images at scale.', 'In this paper we address the problem of matching two images with two\ndifferent resolutions: a high-resolution image and a low-resolution one. The\ndifference in resolution between the two images is not known and without loss\nof generality one of the images is assumed to be the high-resolution one. On\nthe premise that changes in resolution act as a smoothing equivalent to changes\nin scale, a scale-space representation of the high-resolution image is\nproduced. Hence the one-to-one classical image matching paradigm becomes\none-to-many because the low-resolution image is compared with all the\nscale-space representations of the high-resolution one. Key to the success of\nsuch a process is the proper representation of the features to be matched in\nscale-space. We show how to represent and extract interest points at variable\nscales and we devise a method allowing the comparison of two images at two\ndifferent resolutions. The method comprises the use of photometric- and\nrotation-invariant descriptors, a geometric model mapping the high-resolution\nimage onto a low-resolution image region, and an image matching strategy based\non local constraints and on the robust estimation of this geometric model.\nExtensive experiments show that our matching method can be used for scale\nchanges up to a factor of 6.', 'Digital camera and mobile document image acquisition are new trends arising\nin the world of Optical Character Recognition and text detection. In some\ncases, such process integrates many distortions and produces poorly scanned\ntext or text-photo images and natural images, leading to an unreliable OCR\ndigitization. In this paper, we present a novel nonparametric and unsupervised\nmethod to compensate for undesirable document image distortions aiming to\noptimally improve OCR accuracy. Our approach relies on a very efficient stack\nof document image enhancing techniques to recover deformation of the entire\ndocument image. First, we propose a local brightness and contrast adjustment\nmethod to effectively handle lighting variations and the irregular distribution\nof image illumination. Second, we use an optimized greyscale conversion\nalgorithm to transform our document image to greyscale level. Third, we sharpen\nthe useful information in the resulting greyscale image using Un-sharp Masking\nmethod. Finally, an optimal global binarization approach is used to prepare the\nfinal document image to OCR recognition. The proposed approach can\nsignificantly improve text detection rate and optical character recognition\naccuracy. To demonstrate the efficiency of our approach, an exhaustive\nexperimentation on a standard dataset is presented.', 'Graphical elements: particularly tables and figures contain a visual summary\nof the most valuable information contained in a document. Therefore,\nlocalization of such graphical objects in the document images is the initial\nstep to understand the content of such graphical objects or document images. In\nthis paper, we present a novel end-to-end trainable deep learning based\nframework to localize graphical objects in the document images called as\nGraphical Object Detection (GOD). Our framework is data-driven and does not\nrequire any heuristics or meta-data to locate graphical objects in the document\nimages. The GOD explores the concept of transfer learning and domain adaptation\nto handle scarcity of labeled training images for graphical object detection\ntask in the document images. Performance analysis carried out on the various\npublic benchmark data sets: ICDAR-2013, ICDAR-POD2017,and UNLV shows that our\nmodel yields promising results as compared to state-of-the-art techniques.', 'Despite significant progress on current state-of-the-art image generation\nmodels, synthesis of document images containing multiple and complex object\nlayouts is a challenging task. This paper presents a novel approach, called\nDocSynth, to automatically synthesize document images based on a given layout.\nIn this work, given a spatial layout (bounding boxes with object categories) as\na reference by the user, our proposed DocSynth model learns to generate a set\nof realistic document images consistent with the defined layout. Also, this\nframework has been adapted to this work as a superior baseline model for\ncreating synthetic document image datasets for augmenting real data during\ntraining for document layout analysis tasks. Different sets of learning\nobjectives have been also used to improve the model performance.\nQuantitatively, we also compare the generated results of our model with real\ndata using standard evaluation metrics. The results highlight that our model\ncan successfully generate realistic and diverse document images with multiple\nobjects. We also present a comprehensive qualitative analysis summary of the\ndifferent scopes of synthetic image generation tasks. Lastly, to our knowledge\nthis is the first work of its kind.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'Automatic annotation of images with descriptive words is a challenging\nproblem with vast applications in the areas of image search and retrieval. This\nproblem can be viewed as a label-assignment problem by a classifier dealing\nwith a very large set of labels, i.e., the vocabulary set. We propose a novel\nannotation method that employs two layers of sparse coding and performs\ncoarse-to-fine labeling. Themes extracted from the training data are treated as\ncoarse labels. Each theme is a set of training images that share a common\nsubject in their visual and textual contents. Our system extracts coarse labels\nfor training and test images without requiring any prior knowledge. Vocabulary\nwords are the fine labels to be associated with images. Most of the annotation\nmethods achieve low recall due to the large number of available fine labels,\ni.e., vocabulary words. These systems also tend to achieve high precision for\nhighly frequent words only while relatively rare words are more important for\nsearch and retrieval purposes. Our system not only outperforms various\npreviously proposed annotation systems, but also achieves symmetric response in\nterms of precision and recall. Our system scores and maintains high precision\nfor words with a wide range of frequencies. Such behavior is achieved by\nintelligently reducing the number of available fine labels or words for each\nimage based on coarse labels assigned to it.', 'Significant progress has been made in boundary detection with the help of\nconvolutional neural networks. Recent boundary detection models not only focus\non real object boundary detection but also ""crisp"" boundaries (precisely\nlocalized along the object\'s contour). There are two methods to evaluate crisp\nboundary performance. One uses more strict tolerance to measure the distance\nbetween the ground truth and the detected contour. The other focuses on\nevaluating the contour map without any postprocessing. In this study, we\nanalyze both methods and conclude that both methods are two aspects of crisp\ncontour evaluation. Accordingly, we propose a novel network named deep\nrefinement network (DRNet) that stacks multiple refinement modules to achieve\nricher feature representation and a novel loss function, which combines\ncross-entropy and dice loss through effective adaptive fusion. Experimental\nresults demonstrated that we achieve state-of-the-art performance for several\navailable datasets.', 'With increasing urbanization, in recent years there has been a growing\ninterest and need in monitoring and analyzing urban flood events. Social media,\nas a new data source, can provide real-time information for flood monitoring.\nThe social media posts with locations are often referred to as Volunteered\nGeographic Information (VGI), which can reveal the spatial pattern of such\nevents. Since more images are shared on social media than ever before, recent\nresearch focused on the extraction of flood-related posts by analyzing images\nin addition to texts. Apart from merely classifying posts as flood relevant or\nnot, more detailed information, e.g. the flood severity, can also be extracted\nbased on image interpretation. However, it has been less tackled and has not\nyet been applied for flood severity mapping.\n  In this paper, we propose a novel three-step process to extract and map flood\nseverity information. First, flood relevant images are retrieved with the help\nof pre-trained convolutional neural networks as feature extractors. Second, the\nimages containing people are further classified into four severity levels by\nobserving the relationship between body parts and their partial inundation,\ni.e. images are classified according to the water level with respect to\ndifferent body parts, namely ankle, knee, hip, and chest. Lastly, locations of\nthe Tweets are used for generating a map of estimated flood extent and\nseverity. This process was applied to an image dataset collected during\nHurricane Harvey in 2017, as a proof of concept. The results show that VGI can\nbe used as a supplement to remote sensing observations for flood extent mapping\nand is beneficial, especially for urban areas, where the infrastructure is\noften occluding water. Based on the extracted water level information, an\nintegrated overview of flood severity can be provided for the early stages of\nemergency response.', 'One image processing application that is very helpful for humans is to\nimprove image quality, poor image quality makes the image more difficult to\ninterpret because the information conveyed by the image is reduced. In the\nprocess of the acquisition of medical images, the resulting image has decreased\nquality (degraded) due to external factors and medical equipment used. For this\nreason, it is necessary to have an image processing process to improve the\nquality of medical images, so that later it is expected to help facilitate\nmedical personnel in analyzing and translating medical images, which will lead\nto an improvement in the quality of diagnosis. In this study, an analysis will\nbe carried out to improve the quality of medical images with noise reduction\nwith the Gaussian Filter Method. Next, it is carried out, and tested against\nmedical images, in this case, the lung photo image. The test image is given\nnoise in the form of impulse salt & pepper and adaptive Gaussian then analyzed\nits performance qualitatively by comparing the output filter image, noise\nimage, and the original image by naked eye.', 'Visual complexity identifies the level of intricacy and details in an image\nor the level of difficulty to describe the image. It is an important concept in\na variety of areas such as cognitive psychology, computer vision and\nvisualization, and advertisement. Yet, efforts to create large, downloadable\nimage datasets with diverse content and unbiased groundtruthing are lacking. In\nthis work, we introduce Savoias, a visual complexity dataset that compromises\nof more than 1,400 images from seven image categories relevant to the above\nresearch areas, namely Scenes, Advertisements, Visualization and infographics,\nObjects, Interior design, Art, and Suprematism. The images in each category\nportray diverse characteristics including various low-level and high-level\nfeatures, objects, backgrounds, textures and patterns, text, and graphics. The\nground truth for Savoias is obtained by crowdsourcing more than 37,000 pairwise\ncomparisons of images using the forced-choice methodology and with more than\n1,600 contributors. The resulting relative scores are then converted to\nabsolute visual complexity scores using the Bradley-Terry method and matrix\ncompletion. When applying five state-of-the-art algorithms to analyze the\nvisual complexity of the images in the Savoias dataset, we found that the\nscores obtained from these baseline tools only correlate well with crowdsourced\nlabels for abstract patterns in the Suprematism category (Pearson correlation\nr=0.84). For the other categories, in particular, the objects and advertisement\ncategories, low correlation coefficients were revealed (r=0.3 and 0.56,\nrespectively). These findings suggest that (1) state-of-the-art approaches are\nmostly insufficient and (2) Savoias enables category-specific method\ndevelopment, which is likely to improve the impact of visual complexity\nanalysis on specific application areas, including computer vision.', '""If I provide you a face image of mine (without telling you the actual age\nwhen I took the picture) and a large amount of face images that I crawled\n(containing labeled faces of different ages but not necessarily paired), can\nyou show me what I would look like when I am 80 or what I was like when I was\n5?"" The answer is probably a ""No."" Most existing face aging works attempt to\nlearn the transformation between age groups and thus would require the paired\nsamples as well as the labeled query image. In this paper, we look at the\nproblem from a generative modeling perspective such that no paired samples is\nrequired. In addition, given an unlabeled image, the generative model can\ndirectly produce the image with desired age attribute. We propose a conditional\nadversarial autoencoder (CAAE) that learns a face manifold, traversing on which\nsmooth age progression and regression can be realized simultaneously. In CAAE,\nthe face is first mapped to a latent vector through a convolutional encoder,\nand then the vector is projected to the face manifold conditional on age\nthrough a deconvolutional generator. The latent vector preserves personalized\nface features (i.e., personality) and the age condition controls progression\nvs. regression. Two adversarial networks are imposed on the encoder and\ngenerator, respectively, forcing to generate more photo-realistic faces.\nExperimental results demonstrate the appealing performance and flexibility of\nthe proposed framework by comparing with the state-of-the-art and ground truth.', 'Traditionally, the main focus of image super-resolution techniques is on\nrecovering the most likely high-quality images from low-quality images, using a\none-to-one low- to high-resolution mapping. Proceeding that way, we ignore the\nfact that there are generally many valid versions of high-resolution images\nthat map to a given low-resolution image. We are tackling in this work the\nproblem of obtaining different high-resolution versions from the same\nlow-resolution image using Generative Adversarial Models. Our learning approach\nmakes use of high frequencies available in the training high-resolution images\nfor preserving and exploring in an unsupervised manner the structural\ninformation available within these images. Experimental results on the CelebA\ndataset confirm the effectiveness of the proposed method, which allows the\ngeneration of both realistic and diverse high-resolution images from\nlow-resolution images.', 'Generating realistic images is difficult, and many formulations for this task\nhave been proposed recently. If we restrict the task to that of generating a\nparticular class of images, however, the task becomes more tractable. That is\nto say, instead of generating an arbitrary image as a sample from the manifold\nof natural images, we propose to sample images from a particular ""subspace"" of\nnatural images, directed by a low-resolution image from the same subspace. The\nproblem we address, while close to the formulation of the single-image\nsuper-resolution problem, is in fact rather different. Single image\nsuper-resolution is the task of predicting the image closest to the ground\ntruth from a relatively low resolution image. We propose to produce samples of\nhigh resolution images given extremely small inputs with a new method called\nLatent Adversarial Generator (LAG). In our generative sampling framework, we\nonly use the input (possibly of very low-resolution) to direct what class of\nsamples the network should produce. As such, the output of our algorithm is not\na unique image that relates to the input, but rather a possible se} of related\nimages sampled from the manifold of natural images. Our method learns\nexclusively in the latent space of the adversary using perceptual loss -- it\ndoes not have a pixel loss.', 'Document Layout Analysis is a fundamental step in Handwritten Text Processing\nsystems, from the extraction of the text lines to the type of zone it belongs\nto. We present a system based on artificial neural networks which is able to\ndetermine not only the baselines of text lines present in the document, but\nalso performs geometric and logic layout analysis of the document. Experiments\nin three different datasets demonstrate the potential of the method and show\ncompetitive results with respect to state-of-the-art methods.', 'Achieving a symbiotic blending between reality and virtuality is a dream that\nhas been lying in the minds of many people for a long time. Advances in various\ndomains constantly bring us closer to making that dream come true. Augmented\nreality as well as virtual reality are in fact trending terms and are expected\nto further progress in the years to come.\n  This master\'s thesis aims to explore these areas and starts by defining\nnecessary terms such as augmented reality (AR) or virtual reality (VR). Usual\ntaxonomies to classify and compare the corresponding experiences are then\ndiscussed.\n  In order to enable those applications, many technical challenges need to be\ntackled, such as accurate motion tracking with 6 degrees of freedom (positional\nand rotational), that is necessary for compelling experiences and to prevent\nuser sickness. Additionally, augmented reality experiences typically rely on\nimage processing to position the superimposed content. To do so, ""paper""\nmarkers or features extracted from the environment are often employed. Both\nsets of techniques are explored and common solutions and algorithms are\npresented.\n  After investigating those technical aspects, I carry out an objective\ncomparison of the existing state-of-the-art and state-of-the-practice in those\ndomains, and I discuss present and potential applications in these areas. As a\npractical validation, I present the results of an application that I have\ndeveloped using Microsoft HoloLens, one of the more advanced affordable\ntechnologies for augmented reality that is available today. Based on the\nexperience and lessons learned during this development, I discuss the\nlimitations of current technologies and present some avenues of future\nresearch.', 'Identity documents recognition is an important sub-field of document\nanalysis, which deals with tasks of robust document detection, type\nidentification, text fields recognition, as well as identity fraud prevention\nand document authenticity validation given photos, scans, or video frames of an\nidentity document capture. Significant amount of research has been published on\nthis topic in recent years, however a chief difficulty for such research is\nscarcity of datasets, due to the subject matter being protected by security\nrequirements. A few datasets of identity documents which are available lack\ndiversity of document types, capturing conditions, or variability of document\nfield values. In addition, the published datasets were typically designed only\nfor a subset of document recognition problems, not for a complex identity\ndocument analysis. In this paper, we present a dataset MIDV-2020 which consists\nof 1000 video clips, 2000 scanned images, and 1000 photos of 1000 unique mock\nidentity documents, each with unique text field values and unique artificially\ngenerated faces, with rich annotation. For the presented benchmark dataset\nbaselines are provided for such tasks as document location and identification,\ntext fields recognition, and face detection. With 72409 annotated images in\ntotal, to the date of publication the proposed dataset is the largest publicly\navailable identity documents dataset with variable artificially generated data,\nand we believe that it will prove invaluable for advancement of the field of\ndocument analysis and recognition. The dataset is available for download at\nftp://smartengines.com/midv-2020 and http://l3i-share.univ-lr.fr .', 'Documents often exhibit various forms of degradation, which make it hard to\nbe read and substantially deteriorate the performance of an OCR system. In this\npaper, we propose an effective end-to-end framework named Document Enhancement\nGenerative Adversarial Networks (DE-GAN) that uses the conditional GANs (cGANs)\nto restore severely degraded document images. To the best of our knowledge,\nthis practice has not been studied within the context of generative adversarial\ndeep networks. We demonstrate that, in different tasks (document clean up,\nbinarization, deblurring and watermark removal), DE-GAN can produce an enhanced\nversion of the degraded document with a high quality. In addition, our approach\nprovides consistent improvements compared to state-of-the-art methods over the\nwidely used DIBCO 2013, DIBCO 2017 and H-DIBCO 2018 datasets, proving its\nability to restore a degraded document image to its ideal condition. The\nobtained results on a wide variety of degradation reveal the flexibility of the\nproposed model to be exploited in other document enhancement problems.', 'Computation of document image quality metrics often depends upon the\navailability of a ground truth image corresponding to the document. This limits\nthe applicability of quality metrics in applications such as hyperparameter\noptimization of image processing algorithms that operate on-the-fly on unseen\ndocuments. This work proposes the use of surrogate models to learn the behavior\nof a given document quality metric on existing datasets where ground truth\nimages are available. The trained surrogate model can later be used to predict\nthe metric value on previously unseen document images without requiring access\nto ground truth images. The surrogate model is empirically evaluated on the\nDocument Image Binarization Competition (DIBCO) and the Handwritten Document\nImage Binarization Competition (H-DIBCO) datasets.', 'This paper presents results of Document Visual Question Answering Challenge\norganized as part of ""Text and Documents in the Deep Learning Era"" workshop, in\nCVPR 2020. The challenge introduces a new problem - Visual Question Answering\non document images. The challenge comprised two tasks. The first task concerns\nwith asking questions on a single document image. On the other hand, the second\ntask is set as a retrieval task where the question is posed over a collection\nof images. For the task 1 a new dataset is introduced comprising 50,000\nquestions-answer(s) pairs defined over 12,767 document images. For task 2\nanother dataset has been created comprising 20 questions over 14,362 document\nimages which share the same document template.', 'We introduce a novel type of text representation that preserves the 2D layout\nof a document. This is achieved by encoding each document page as a\ntwo-dimensional grid of characters. Based on this representation, we present a\ngeneric document understanding pipeline for structured documents. This pipeline\nmakes use of a fully convolutional encoder-decoder network that predicts a\nsegmentation mask and bounding boxes. We demonstrate its capabilities on an\ninformation extraction task from invoices and show that it significantly\noutperforms approaches based on sequential text or document images.', 'There is a large collection of Handwritten English paper documents of\nHistorical and Scientific importance. But paper documents are not recognized\ndirectly by computer. Hence the closest way of indexing these documents is by\nstoring their document digital image. Hence a large database of document images\ncan replace the paper documents. But the document and data corresponding to\neach image cannot be directly recognized by the computer.\n  This paper applies the technique of word spotting using Modified Character\nShape Code to Handwritten English document images for quick and efficient query\nsearch of words on a database of document images. It is different from other\nWord Spotting techniques as it implements two level of selection for word\nsegments to match search query. First based on word size and then based on\ncharacter shape code of query. It makes the process faster and more efficient\nand reduces the need of multiple pre-processing.', 'Today all kind of information is getting digitized and along with all this\ndigitization, the huge archive of various kinds of documents is being digitized\ntoo. We know that, Optical Character Recognition is the method through which,\nnewspapers and other paper documents convert into digital resources. But, it is\na fact that this method works on texts only. As a result, if we try to process\nany document which contains non-textual zones, then we will get garbage texts\nas output. That is why; in order to digitize documents properly they should be\nprepossessed carefully. And while preprocessing, segmenting document in\ndifferent regions according to the category properly is most important. But,\nthe Optical Character Recognition processes available for Bangla language have\nno such algorithm that can categorize a newspaper/book page fully. So we worked\nto decompose a document into its several parts like headlines, sub headlines,\ncolumns, images etc. And if the input is skewed and rotated, then the input was\nalso deskewed and de-rotated. To decompose any Bangla document we found out the\nedges of the input image. Then we find out the horizontal and vertical area of\nevery pixel where it lies in. Later on the input image was cut according to\nthese areas. Then we pick each and every sub image and found out their\nheight-width ratio, line height. Then according to these values the sub images\nwere categorized. To deskew the image we found out the skew angle and de skewed\nthe image according to this angle. To de-rotate the image we used the line\nheight, matra line, pixel ratio of matra line.', 'Nowadays, as cameras are rapidly adopted in our daily routine, images of\ndocuments are becoming both abundant and prevalent. Unlike natural images that\ncapture physical objects, document-images contain a significant amount of text\nwith critical semantics and complicated layouts. In this work, we devise a\ngeneric unsupervised technique to learn multimodal affinities between textual\nentities in a document-image, considering their visual style, the content of\ntheir underlying text and their geometric context within the image. We then use\nthese learned affinities to automatically cluster the textual entities in the\nimage into different semantic groups. The core of our approach is a deep\noptimization scheme dedicated for an image provided by the user that detects\nand leverages reliable pairwise connections in the multimodal representation of\nthe textual elements in order to properly learn the affinities. We show that\nour technique can operate on highly varying images spanning a wide range of\ndocuments and demonstrate its applicability for various editing operations\nmanipulating the content, appearance and geometry of the image.', 'The accuracy of OCR is usually affected by the quality of the input document\nimage and different kinds of marred document images hamper the OCR results.\nAmong these scenarios, the low-resolution image is a common and challenging\ncase. In this paper, we propose the cascaded networks for document image\nsuper-resolution. Our model is composed by the Detail-Preserving Networks with\nsmall magnification. The loss function with perceptual terms is designed to\nsimultaneously preserve the original patterns and enhance the edge of the\ncharacters. These networks are trained with the same architecture and different\nparameters and then assembled into a pipeline model with a larger\nmagnification. The low-resolution images can upscale gradually by passing\nthrough each Detail-Preserving Network until the final high-resolution images.\nThrough extensive experiments on two scanning document image datasets, we\ndemonstrate that the proposed approach outperforms recent state-of-the-art\nimage super-resolution methods, and combining it with standard OCR system lead\nto signification improvements on the recognition results.', 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.', 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', 'The digital camera captured document images may often be warped and distorted\ndue to different camera angles or document surfaces. A robust technique is\nneeded to solve this kind of distortion. The research on dewarping of the\ndocument suffers due to the limited availability of benchmark public dataset.\nIn recent times, deep learning based approaches are used to solve the problems\naccurately. To train most of the deep neural networks a large number of\ndocument images is required and generating such a large volume of document\nimages manually is difficult. In this paper, we propose a technique to generate\na synthetic warped image from a flat-bedded scanned document image. It is done\nby calculating warping factors for each pixel position using two warping\nposition parameters (WPP) and eight warping control parameters (WCP). These\nparameters can be specified as needed depending upon the desired warping. The\nresults are compared with similar real captured images both qualitative and\nquantitative way.', 'Document segmentation is a method of rending the document into distinct\nregions. A document is an assortment of information and a standard mode of\nconveying information to others. Pursuance of data from documents involves ton\nof human effort, time intense and might severely prohibit the usage of data\nsystems. So, automatic information pursuance from the document has become a big\nissue. It is been shown that document segmentation will facilitate to beat such\nproblems. This paper proposes a new approach to segment and classify the\ndocument regions as text, image, drawings and table. Document image is divided\ninto blocks using Run length smearing rule and features are extracted from\nevery blocks. Discipulus tool has been used to construct the Genetic\nprogramming based classifier model and located 97.5% classification accuracy.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : Our team is developing street views for a mapping application, and we need to censor any potentially sensitive text information within the images we capture. As such, we require a model capable of detecting and interpreting multi-oriented text in various scenes. Traditional methods are proving too slow to process the volume of images we capture for the street view. We think the most efficient solution would be to use two independent models. The first model should be able to extract text regions from the original image, while the second should interpret the text from the region, regardless of its orientation. We are seeking a suitable structure of models that can identify text from images, and ideally, these structures should support parallel training due to the large training set we have for this task. The second model should also be capable of understanding different languages and local slang. Additionally, we would prefer models that come with commercially licensed code.","['Key information extraction from document images is of paramount importance in\noffice automation. Conventional template matching based approaches fail to\ngeneralize well to document images of unseen templates, and are not robust\nagainst text recognition errors. In this paper, we propose an end-to-end\nSpatial Dual-Modality Graph Reasoning method (SDMG-R) to extract key\ninformation from unstructured document images. We model document images as\ndual-modality graphs, nodes of which encode both the visual and textual\nfeatures of detected text regions, and edges of which represent the spatial\nrelations between neighboring text regions. The key information extraction is\nsolved by iteratively propagating messages along graph edges and reasoning the\ncategories of graph nodes. In order to roundly evaluate our proposed method as\nwell as boost the future research, we release a new dataset named WildReceipt,\nwhich is collected and annotated tailored for the evaluation of key information\nextraction from document images of unseen templates in the wild. It contains 25\nkey information categories, a total of about 69000 text boxes, and is about 2\ntimes larger than the existing public datasets. Extensive experiments validate\nthat all information including visual features, textual features and spatial\nrelations can benefit key information extraction. It has been shown that SDMG-R\ncan effectively extract key information from document images of unseen\ntemplates, and obtain new state-of-the-art results on the recent popular\nbenchmark SROIE and our WildReceipt. Our code and dataset will be publicly\nreleased.', 'In this work, we propose a new technique that combines appearance and text in\na Convolutional Neural Network (CNN), with the aim of detecting regions of\ndifferent textual categories. We define a novel visual representation of the\nsemantic meaning of text that allows a seamless integration in a standard CNN\narchitecture. This representation, referred to as text-map, is integrated with\nthe actual image to provide a much richer input to the network. Text-maps are\ncolored with different intensities depending on the relevance of the words\nrecognized over the image. Concretely, these words are previously extracted\nusing Optical Character Recognition (OCR) and they are colored according to the\nprobability of belonging to a textual category of interest. In this sense, this\nsolution is especially relevant in the context of item coding for supermarket\nproducts, where different types of textual categories must be identified, such\nas ingredients or nutritional facts. We evaluated our solution in the\nproprietary item coding dataset of Nielsen Brandbank, which contains more than\n10,000 images for train and 2,000 images for test. The reported results\ndemonstrate that our approach focused on visual and textual data outperforms\nstate-of-the-art algorithms only based on appearance, such as standard Faster\nR-CNN. These enhancements are reflected in precision and recall, which are\nimproved in 42 and 33 points respectively.', 'We introduce Chinese Text in the Wild, a very large dataset of Chinese text\nin street view images. While optical character recognition (OCR) in document\nimages is well studied and many commercial tools are available, detection and\nrecognition of text in natural images is still a challenging problem,\nespecially for more complicated character sets such as Chinese text. Lack of\ntraining data has always been a problem, especially for deep learning methods\nwhich require massive training data.\n  In this paper we provide details of a newly created dataset of Chinese text\nwith about 1 million Chinese characters annotated by experts in over 30\nthousand street view images. This is a challenging dataset with good diversity.\nIt contains planar text, raised text, text in cities, text in rural areas, text\nunder poor illumination, distant text, partially occluded text, etc. For each\ncharacter in the dataset, the annotation includes its underlying character, its\nbounding box, and 6 attributes. The attributes indicate whether it has complex\nbackground, whether it is raised, whether it is handwritten or printed, etc.\nThe large size and diversity of this dataset make it suitable for training\nrobust neural networks for various tasks, particularly detection and\nrecognition. We give baseline results using several state-of-the-art networks,\nincluding AlexNet, OverFeat, Google Inception and ResNet for character\nrecognition, and YOLOv2 for character detection in images. Overall Google\nInception has the best performance on recognition with 80.5% top-1 accuracy,\nwhile YOLOv2 achieves an mAP of 71.0% on detection. Dataset, source code and\ntrained models will all be publicly available on the website.', ""A significant number of images shared on social media platforms such as\nFacebook and Instagram contain text in various forms. It's increasingly\nbecoming commonplace for bad actors to share misinformation, hate speech or\nother kinds of harmful content as text overlaid on images on such platforms. A\nscene-text understanding system should hence be able to handle text in various\norientations that the adversary might use. Moreover, such a system can be\nincorporated into screen readers used to aid the visually impaired. In this\nwork, we extend the scene-text extraction system at Facebook, Rosetta, to\nefficiently handle text in various orientations. Specifically, we incorporate\nthe Rotation Region Proposal Networks (RRPN) in our text extraction pipeline\nand offer practical suggestions for building and deploying a model for\ndetecting and recognizing text in arbitrary orientations efficiently.\nExperimental results show a significant improvement on detecting rotated text."", 'Text detection and recognition in natural images have long been considered as\ntwo separate tasks that are processed sequentially. Training of two tasks in a\nunified framework is non-trivial due to significant dif- ferences in\noptimisation difficulties. In this work, we present a conceptually simple yet\nefficient framework that simultaneously processes the two tasks in one shot.\nOur main contributions are three-fold: 1) we propose a novel text-alignment\nlayer that allows it to precisely compute convolutional features of a text\ninstance in ar- bitrary orientation, which is the key to boost the per-\nformance; 2) a character attention mechanism is introduced by using character\nspatial information as explicit supervision, leading to large improvements in\nrecognition; 3) two technologies, together with a new RNN branch for word\nrecognition, are integrated seamlessly into a single model which is end-to-end\ntrainable. This allows the two tasks to work collaboratively by shar- ing\nconvolutional features, which is critical to identify challenging text\ninstances. Our model achieves impressive results in end-to-end recognition on\nthe ICDAR2015 dataset, significantly advancing most recent results, with\nimprovements of F-measure from (0.54, 0.51, 0.47) to (0.82, 0.77, 0.63), by\nusing a strong, weak and generic lexicon respectively. Thanks to joint\ntraining, our method can also serve as a good detec- tor by achieving a new\nstate-of-the-art detection performance on two datasets.', 'Recent advancements in the area of Computer Vision with state-of-art Neural\nNetworks has given a boost to Optical Character Recognition (OCR) accuracies.\nHowever, extracting characters/text alone is often insufficient for relevant\ninformation extraction as documents also have a visual structure that is not\ncaptured by OCR. Extracting information from tables, charts, footnotes, boxes,\nheadings and retrieving the corresponding structured representation for the\ndocument remains a challenge and finds application in a large number of\nreal-world use cases. In this paper, we propose a novel enterprise based\nend-to-end framework called DeepReader which facilitates information extraction\nfrom document images via identification of visual entities and populating a\nmeta relational model across different entities in the document image. The\nmodel schema allows for an easy to understand abstraction of the entities\ndetected by the deep vision models and the relationships between them.\nDeepReader has a suite of state-of-the-art vision algorithms which are applied\nto recognize handwritten and printed text, eliminate noisy effects, identify\nthe type of documents and detect visual entities like tables, lines and boxes.\nDeep Reader maps the extracted entities into a rich relational schema so as to\ncapture all the relevant relationships between entities (words, textboxes,\nlines etc) detected in the document. Relevant information and fields can then\nbe extracted from the document by writing SQL queries on top of the\nrelationship tables. A natural language based interface is added on top of the\nrelationship schema so that a non-technical user, specifying the queries in\nnatural language, can fetch the information with minimal effort. In this paper,\nwe also demonstrate many different capabilities of Deep Reader and report\nresults on a real-world use case.', 'We present a novel deep neural model for text detection in document images.\nFor robust text detection in noisy scanned documents, the advantages of\nmulti-task learning are adopted by adding an auxiliary task of text\nenhancement. Namely, our proposed model is designed to perform noise reduction\nand text region enhancement as well as text detection. Moreover, we enrich the\ntraining data for the model with synthesized document images that are fully\nlabeled for text detection and enhancement, thus overcome the insufficiency of\nlabeled document image data. For the effective exploitation of the synthetic\nand real data, the training process is separated in two phases. The first phase\nis training only synthetic data in a fully-supervised manner. Then real data\nwith only detection labels are added in the second phase. The enhancement task\nfor the real data is weakly-supervised with information from their detection\nlabels. Our methods are demonstrated in a real document dataset with\nperformances exceeding those of other text detection methods. Moreover,\nablations are conducted and the results confirm the effectiveness of the\nsynthetic data, auxiliary task, and weak-supervision. Whereas the existing text\ndetection studies mostly focus on the text in scenes, our proposed method is\noptimized to the applications for the text in scanned documents.', 'Nowadays, as cameras are rapidly adopted in our daily routine, images of\ndocuments are becoming both abundant and prevalent. Unlike natural images that\ncapture physical objects, document-images contain a significant amount of text\nwith critical semantics and complicated layouts. In this work, we devise a\ngeneric unsupervised technique to learn multimodal affinities between textual\nentities in a document-image, considering their visual style, the content of\ntheir underlying text and their geometric context within the image. We then use\nthese learned affinities to automatically cluster the textual entities in the\nimage into different semantic groups. The core of our approach is a deep\noptimization scheme dedicated for an image provided by the user that detects\nand leverages reliable pairwise connections in the multimodal representation of\nthe textual elements in order to properly learn the affinities. We show that\nour technique can operate on highly varying images spanning a wide range of\ndocuments and demonstrate its applicability for various editing operations\nmanipulating the content, appearance and geometry of the image.', 'Extracting texts of various size and shape from images containing multiple\nobjects is an important problem in many contexts, especially, in connection to\ne-commerce, augmented reality assistance system in natural scene, etc. The\nexisting works (based on only CNN) often perform sub-optimally when the image\ncontains regions of high entropy having multiple objects. This paper presents\nan end-to-end text detection strategy combining a segmentation algorithm and an\nensemble of multiple text detectors of different types to detect text in every\nindividual image segments independently. The proposed strategy involves a\nsuper-pixel based image segmenter which splits an image into multiple regions.\nA convolutional deep neural architecture is developed which works on each of\nthe segments and detects texts of multiple shapes, sizes, and structures. It\noutperforms the competing methods in terms of coverage in detecting texts in\nimages especially the ones where the text of various types and sizes are\ncompacted in a small region along with various other objects. Furthermore, the\nproposed text detection method along with a text recognizer outperforms the\nexisting state-of-the-art approaches in extracting text from high entropy\nimages. We validate the results on a dataset consisting of product images on an\ne-commerce website.', ""Text in an image provides vital information for interpreting its contents,\nand text in a scene can aide with a variety of tasks from navigation, to\nobstacle avoidance, and odometry. Despite its value, however, identifying\ngeneral text in images remains a challenging research problem. Motivated by the\nneed to consider the widely varying forms of natural text, we propose a\nbottom-up approach to the problem which reflects the `characterness' of an\nimage region. In this sense our approach mirrors the move from saliency\ndetection methods to measures of `objectness'. In order to measure the\ncharacterness we develop three novel cues that are tailored for character\ndetection, and a Bayesian method for their integration. Because text is made up\nof sets of characters, we then design a Markov random field (MRF) model so as\nto exploit the inherent dependencies between characters.\n  We experimentally demonstrate the effectiveness of our characterness cues as\nwell as the advantage of Bayesian multi-cue integration. The proposed text\ndetector outperforms state-of-the-art methods on a few benchmark scene text\ndetection datasets. We also show that our measurement of `characterness' is\nsuperior than state-of-the-art saliency detection models when applied to the\nsame task."", 'Reading text from images remains challenging due to multi-orientation,\nperspective distortion and especially the curved nature of irregular text. Most\nof existing approaches attempt to solve the problem in two or multiple stages,\nwhich is considered to be the bottleneck to optimize the overall performance.\nTo address this issue, we propose an end-to-end trainable network architecture,\nnamed TextNet, which is able to simultaneously localize and recognize irregular\ntext from images. Specifically, we develop a scale-aware attention mechanism to\nlearn multi-scale image features as a backbone network, sharing fully\nconvolutional features and computation for localization and recognition. In\ntext detection branch, we directly generate text proposals in quadrangles,\ncovering oriented, perspective and curved text regions. To preserve text\nfeatures for recognition, we introduce a perspective RoI transform layer, which\ncan align quadrangle proposals into small feature maps. Furthermore, in order\nto extract effective features for recognition, we propose to encode the aligned\nRoI features by RNN into context information, combining spatial attention\nmechanism to generate text sequences. This overall pipeline is capable of\nhandling both regular and irregular cases. Finally, text localization and\nrecognition tasks can be jointly trained in an end-to-end fashion with designed\nmulti-task loss. Experiments on standard benchmarks show that the proposed\nTextNet can achieve state-of-the-art performance, and outperform existing\napproaches on irregular datasets by a large margin.', 'Text spotting in natural scene images is of great importance for many image\nunderstanding tasks. It includes two sub-tasks: text detection and recognition.\nIn this work, we propose a unified network that simultaneously localizes and\nrecognizes text with a single forward pass, avoiding intermediate processes\nsuch as image cropping and feature re-calculation, word separation, and\ncharacter grouping.\n  In contrast to existing approaches that consider text detection and\nrecognition as two distinct tasks and tackle them one by one, the proposed\nframework settles these two tasks concurrently. The whole framework can be\ntrained end-to-end and is able to handle text of arbitrary shapes. The\nconvolutional features are calculated only once and shared by both detection\nand recognition modules. Through multi-task training, the learned features\nbecome more discriminate and improve the overall performance. By employing the\n$2$D attention model in word recognition, the irregularity of text can be\nrobustly addressed. It provides the spatial location for each character, which\nnot only helps local feature extraction in word recognition, but also indicates\nan orientation angle to refine text localization. Our proposed method has\nachieved state-of-the-art performance on several standard text spotting\nbenchmarks, including both regular and irregular ones.', 'Scene text detection methods based on neural networks have emerged recently\nand have shown promising results. Previous methods trained with rigid\nword-level bounding boxes exhibit limitations in representing the text region\nin an arbitrary shape. In this paper, we propose a new scene text detection\nmethod to effectively detect text area by exploring each character and affinity\nbetween characters. To overcome the lack of individual character level\nannotations, our proposed framework exploits both the given character-level\nannotations for synthetic images and the estimated character-level\nground-truths for real images acquired by the learned interim model. In order\nto estimate affinity between characters, the network is trained with the newly\nproposed representation for affinity. Extensive experiments on six benchmarks,\nincluding the TotalText and CTW-1500 datasets which contain highly curved texts\nin natural images, demonstrate that our character-level text detection\nsignificantly outperforms the state-of-the-art detectors. According to the\nresults, our proposed method guarantees high flexibility in detecting\ncomplicated scene text images, such as arbitrarily-oriented, curved, or\ndeformed texts.', 'Recent deep learning models have demonstrated strong capabilities for\nclassifying text and non-text components in natural images. They extract a\nhigh-level feature computed globally from a whole image component (patch),\nwhere the cluttered background information may dominate true text features in\nthe deep representation. This leads to less discriminative power and poorer\nrobustness. In this work, we present a new system for scene text detection by\nproposing a novel Text-Attentional Convolutional Neural Network (Text-CNN) that\nparticularly focuses on extracting text-related regions and features from the\nimage components. We develop a new learning mechanism to train the Text-CNN\nwith multi-level and rich supervised information, including text region mask,\ncharacter label, and binary text/nontext information. The rich supervision\ninformation enables the Text-CNN with a strong capability for discriminating\nambiguous texts, and also increases its robustness against complicated\nbackground components. The training process is formulated as a multi-task\nlearning problem, where low-level supervised information greatly facilitates\nmain task of text/non-text classification. In addition, a powerful low-level\ndetector called Contrast- Enhancement Maximally Stable Extremal Regions\n(CE-MSERs) is developed, which extends the widely-used MSERs by enhancing\nintensity contrast between text patterns and background. This allows it to\ndetect highly challenging text patterns, resulting in a higher recall. Our\napproach achieved promising results on the ICDAR 2013 dataset, with a F-measure\nof 0.82, improving the state-of-the-art results substantially.', 'Detecting incidental scene text is a challenging task because of\nmulti-orientation, perspective distortion, and variation of text size, color\nand scale. Retrospective research has only focused on using rectangular\nbounding box or horizontal sliding window to localize text, which may result in\nredundant background noise, unnecessary overlap or even information loss. To\naddress these issues, we propose a new Convolutional Neural Networks (CNNs)\nbased method, named Deep Matching Prior Network (DMPNet), to detect text with\ntighter quadrangle. First, we use quadrilateral sliding windows in several\nspecific intermediate convolutional layers to roughly recall the text with\nhigher overlapping area and then a shared Monte-Carlo method is proposed for\nfast and accurate computing of the polygonal areas. After that, we designed a\nsequential protocol for relative regression which can exactly predict text with\ncompact quadrangle. Moreover, a auxiliary smooth Ln loss is also proposed for\nfurther regressing the position of text, which has better overall performance\nthan L2 loss and smooth L1 loss in terms of robustness and stability. The\neffectiveness of our approach is evaluated on a public word-level,\nmulti-oriented scene text database, ICDAR 2015 Robust Reading Competition\nChallenge 4 ""Incidental scene text localization"". The performance of our method\nis evaluated by using F-measure and found to be 70.64%, outperforming the\nexisting state-of-the-art method with F-measure 63.76%.', 'We present a novel single-shot text detector that directly outputs word-level\nbounding boxes in a natural image. We propose an attention mechanism which\nroughly identifies text regions via an automatically learned attentional map.\nThis substantially suppresses background interference in the convolutional\nfeatures, which is the key to producing accurate inference of words,\nparticularly at extremely small sizes. This results in a single model that\nessentially works in a coarse-to-fine manner. It departs from recent FCN- based\ntext detectors which cascade multiple FCN models to achieve an accurate\nprediction. Furthermore, we develop a hierarchical inception module which\nefficiently aggregates multi-scale inception features. This enhances local\ndetails, and also encodes strong context information, allow- ing the detector\nto work reliably on multi-scale and multi- orientation text with single-scale\nimages. Our text detector achieves an F-measure of 77% on the ICDAR 2015 bench-\nmark, advancing the state-of-the-art results in [18, 28]. Demo is available at:\nhttp://sstd.whuang.org/.', 'In this paper, we develop a novel unified framework called DeepText for text\nregion proposal generation and text detection in natural images via a fully\nconvolutional neural network (CNN). First, we propose the inception region\nproposal network (Inception-RPN) and design a set of text characteristic prior\nbounding boxes to achieve high word recall with only hundred level candidate\nproposals. Next, we present a powerful textdetection network that embeds\nambiguous text category (ATC) information and multilevel region-of-interest\npooling (MLRP) for text and non-text classification and accurate localization.\nFinally, we apply an iterative bounding box voting scheme to pursue high recall\nin a complementary manner and introduce a filtering algorithm to retain the\nmost suitable bounding box, while removing redundant inner and outer boxes for\neach text instance. Our approach achieves an F-measure of 0.83 and 0.85 on the\nICDAR 2011 and 2013 robust text detection benchmarks, outperforming previous\nstate-of-the-art results.', 'Text in natural images is of arbitrary orientations, requiring detection in\nterms of oriented bounding boxes. Normally, a multi-oriented text detector\noften involves two key tasks: 1) text presence detection, which is a\nclassification problem disregarding text orientation; 2) oriented bounding box\nregression, which concerns about text orientation. Previous methods rely on\nshared features for both tasks, resulting in degraded performance due to the\nincompatibility of the two tasks. To address this issue, we propose to perform\nclassification and regression on features of different characteristics,\nextracted by two network branches of different designs. Concretely, the\nregression branch extracts rotation-sensitive features by actively rotating the\nconvolutional filters, while the classification branch extracts\nrotation-invariant features by pooling the rotation-sensitive features. The\nproposed method named Rotation-sensitive Regression Detector (RRD) achieves\nstate-of-the-art performance on three oriented scene text benchmark datasets,\nincluding ICDAR 2015, MSRA-TD500, RCTW-17 and COCO-Text. Furthermore, RRD\nachieves a significant improvement on a ship collection dataset, demonstrating\nits generality on oriented object detection.', 'Recent state-of-the-art scene text recognition methods have primarily focused\non horizontal text in images. However, in several Asian countries, including\nChina, large amounts of text in signs, books, and TV commercials are vertically\ndirected. Because the horizontal and vertical texts exhibit different\ncharacteristics, developing an algorithm that can simultaneously recognize both\ntypes of text in real environments is necessary. To address this problem, we\nadopted the direction encoding mask (DEM) and selective attention network (SAN)\nmethods based on supervised learning. DEM contains directional information to\ncompensate in cases that lack text direction; therefore, our network is trained\nusing this information to handle the vertical text. The SAN method is designed\nto work individually for both types of text. To train the network to recognize\nboth types of text and to evaluate the effectiveness of the designed model, we\nprepared a new synthetic vertical text dataset and collected an actual vertical\ntext dataset (VTD142) from the Web. Using these datasets, we proved that our\nproposed model can accurately recognize both vertical and horizontal text and\ncan achieve state-of-the-art results in experiments using benchmark datasets,\nincluding the street view test (SVT), IIIT-5k, and ICDAR. Although our model is\nrelatively simple as compared to its predecessors, it maintains the accuracy\nand is trained in an end-to-end manner.', 'Unifying text detection and text recognition in an end-to-end training\nfashion has become a new trend for reading text in the wild, as these two tasks\nare highly relevant and complementary. In this paper, we investigate the\nproblem of scene text spotting, which aims at simultaneous text detection and\nrecognition in natural images. An end-to-end trainable neural network named as\nMask TextSpotter is presented. Different from the previous text spotters that\nfollow the pipeline consisting of a proposal generation network and a\nsequence-to-sequence recognition network, Mask TextSpotter enjoys a simple and\nsmooth end-to-end learning procedure, in which both detection and recognition\ncan be achieved directly from two-dimensional space via semantic segmentation.\nFurther, a spatial attention module is proposed to enhance the performance and\nuniversality. Benefiting from the proposed two-dimensional representation on\nboth detection and recognition, it easily handles text instances of irregular\nshapes, for instance, curved text. We evaluate it on four English datasets and\none multi-language dataset, achieving consistently superior performance over\nstate-of-the-art methods in both detection and end-to-end text recognition\ntasks. Moreover, we further investigate the recognition module of our method\nseparately, which significantly outperforms state-of-the-art methods on both\nregular and irregular text datasets for scene text recognition.', 'Texts appearing in daily scenes that can be recognized by OCR (Optical\nCharacter Recognition) tools contain significant information, such as street\nname, product brand and prices. Two tasks -- text-based visual question\nanswering and text-based image captioning, with a text extension from existing\nvision-language applications, are catching on rapidly. To address these\nproblems, many sophisticated multi-modality encoding frameworks (such as\nheterogeneous graph structure) are being used. In this paper, we argue that a\nsimple attention mechanism can do the same or even better job without any bells\nand whistles. Under this mechanism, we simply split OCR token features into\nseparate visual- and linguistic-attention branches, and send them to a popular\nTransformer decoder to generate answers or captions. Surprisingly, we find this\nsimple baseline model is rather strong -- it consistently outperforms\nstate-of-the-art (SOTA) models on two popular benchmarks, TextVQA and all three\ntasks of ST-VQA, although these SOTA models use far more complex encoding\nmechanisms. Transferring it to text-based image captioning, we also surpass the\nTextCaps Challenge 2020 winner. We wish this work to set the new baseline for\nthis two OCR text related applications and to inspire new thinking of\nmulti-modality encoder design. Code is available at\nhttps://github.com/ZephyrZhuQi/ssbaseline', 'Recognizing text from natural images is a hot research topic in computer\nvision due to its various applications. Despite the enduring research of\nseveral decades on optical character recognition (OCR), recognizing texts from\nnatural images is still a challenging task. This is because scene texts are\noften in irregular (e.g. curved, arbitrarily-oriented or seriously distorted)\narrangements, which have not yet been well addressed in the literature.\nExisting methods on text recognition mainly work with regular (horizontal and\nfrontal) texts and cannot be trivially generalized to handle irregular texts.\nIn this paper, we develop the arbitrary orientation network (AON) to directly\ncapture the deep features of irregular texts, which are combined into an\nattention-based decoder to generate character sequence. The whole network can\nbe trained end-to-end by using only images and word-level annotations.\nExtensive experiments on various benchmarks, including the CUTE80,\nSVT-Perspective, IIIT5k, SVT and ICDAR datasets, show that the proposed\nAON-based method achieves the-state-of-the-art performance in irregular\ndatasets, and is comparable to major existing methods in regular datasets.']","['This paper describes a dataset containing small images of text from everyday\nscenes. The purpose of the dataset is to support the development of new\nautomated systems that can detect and analyze text. Although much research has\nbeen devoted to text detection and recognition in scanned documents, relatively\nlittle attention has been given to text detection in other types of images,\nsuch as photographs that are posted on social-media sites. This new dataset,\nknown as COCO-Text-Patch, contains approximately 354,000 small images that are\neach labeled as ""text"" or ""non-text"". This dataset particularly addresses the\nproblem of text verification, which is an essential stage in the end-to-end\ntext detection and recognition pipeline. In order to evaluate the utility of\nthis dataset, it has been used to train two deep convolution neural networks to\ndistinguish text from non-text. One network is inspired by the GoogLeNet\narchitecture, and the second one is based on CaffeNet. Accuracy levels of 90.2%\nand 90.9% were obtained using the two networks, respectively. All of the\nimages, source code, and deep-learning trained models described in this paper\nwill be publicly available', 'Regional language extraction from a natural scene image is always a\nchallenging proposition due to its dependence on the text information extracted\nfrom Image. Text Extraction on the other hand varies on different lighting\ncondition, arbitrary orientation, inadequate text information, heavy background\ninfluence over text and change of text appearance. This paper presents a novel\nunified method for tackling the above challenges. The proposed work uses an\nimage correction and segmentation technique on the existing Text Detection\nPipeline an Efficient and Accurate Scene Text Detector (EAST). EAST uses\nstandard PVAnet architecture to select features and non maximal suppression to\ndetect text from image. Text recognition is done using combined architecture of\nMaxOut convolution neural network (CNN) and Bidirectional long short term\nmemory (LSTM) network. After recognizing text using the Deep Learning based\napproach, the native Languages are translated to English and tokenized using\nstandard Text Tokenizers. The tokens that very likely represent a location is\nused to find the Global Positioning System (GPS) coordinates of the location\nand subsequently the regional languages spoken in that location is extracted.\nThe proposed method is tested on a self generated dataset collected from\nGovernment of India dataset and experimented on Standard Dataset to evaluate\nthe performance of the proposed technique. Comparative study with a few\nstate-of-the-art methods on text detection, recognition and extraction of\nregional language from images shows that the proposed method outperforms the\nexisting methods.', 'Many approaches have recently been proposed to detect irregular scene text\nand achieved promising results. However, their localization results may not\nwell satisfy the following text recognition part mainly because of two reasons:\n1) recognizing arbitrary shaped text is still a challenging task, and 2)\nprevalent non-trainable pipeline strategies between text detection and text\nrecognition will lead to suboptimal performances. To handle this\nincompatibility problem, in this paper we propose an end-to-end trainable text\nspotting approach named Text Perceptron. Concretely, Text Perceptron first\nemploys an efficient segmentation-based text detector that learns the latent\ntext reading order and boundary information. Then a novel Shape Transform\nModule (abbr. STM) is designed to transform the detected feature regions into\nregular morphologies without extra parameters. It unites text detection and the\nfollowing recognition part into a whole framework, and helps the whole network\nachieve global optimization. Experiments show that our method achieves\ncompetitive performance on two standard text benchmarks, i.e., ICDAR 2013 and\nICDAR 2015, and also obviously outperforms existing methods on irregular text\nbenchmarks SCUT-CTW1500 and Total-Text.', 'Recently, scene text detection has become an active research topic in\ncomputer vision and document analysis, because of its great importance and\nsignificant challenge. However, vast majority of the existing methods detect\ntext within local regions, typically through extracting character, word or line\nlevel candidates followed by candidate aggregation and false positive\nelimination, which potentially exclude the effect of wide-scope and long-range\ncontextual cues in the scene. To take full advantage of the rich information\navailable in the whole natural image, we propose to localize text in a holistic\nmanner, by casting scene text detection as a semantic segmentation problem. The\nproposed algorithm directly runs on full images and produces global, pixel-wise\nprediction maps, in which detections are subsequently formed. To better make\nuse of the properties of text, three types of information regarding text\nregion, individual characters and their relationship are estimated, with a\nsingle Fully Convolutional Network (FCN) model. With such predictions of text\nproperties, the proposed algorithm can simultaneously handle horizontal,\nmulti-oriented and curved text in real-world natural images. The experiments on\nstandard benchmarks, including ICDAR 2013, ICDAR 2015 and MSRA-TD500,\ndemonstrate that the proposed algorithm substantially outperforms previous\nstate-of-the-art approaches. Moreover, we report the first baseline result on\nthe recently-released, large-scale dataset COCO-Text.', 'Since the low quality of document images will greatly undermine the chances\nof success in automatic text recognition and analysis, it is necessary to\nassess the quality of document images uploaded in online business process, so\nas to reject those images of low quality. In this paper, we attempt to achieve\ndocument image quality assessment and our contributions are twofold. Firstly,\nsince document image quality assessment is more interested in text, we propose\na text line based framework to estimate document image quality, which is\ncomposed of three stages: text line detection, text line quality prediction,\nand overall quality assessment. Text line detection aims to find potential text\nlines with a detector. In the text line quality prediction stage, the quality\nscore is computed for each text line with a CNN-based prediction model. The\noverall quality of document images is finally assessed with the ensemble of all\ntext line quality. Secondly, to train the prediction model, a large-scale\ndataset, comprising 52,094 text line images, is synthesized with diverse\nattributes. For each text line image, a quality label is computed with a\npiece-wise function. To demonstrate the effectiveness of the proposed\nframework, comprehensive experiments are evaluated on two popular document\nimage quality assessment benchmarks. Our framework significantly outperforms\nthe state-of-the-art methods by large margins on the large and complicated\ndataset.', ""Conventional approaches to image-text retrieval mainly focus on indexing\nvisual objects appearing in pictures but ignore the interactions between these\nobjects. Such objects occurrences and interactions are equivalently useful and\nimportant in this field as they are usually mentioned in the text. Scene graph\npresentation is a suitable method for the image-text matching challenge and\nobtained good results due to its ability to capture the inter-relationship\ninformation. Both images and text are represented in scene graph levels and\nformulate the retrieval challenge as a scene graph matching challenge. In this\npaper, we introduce the Local and Global Scene Graph Matching (LGSGM) model\nthat enhances the state-of-the-art method by integrating an extra graph\nconvolution network to capture the general information of a graph.\nSpecifically, for a pair of scene graphs of an image and its caption, two\nseparate models are used to learn the features of each graph's nodes and edges.\nThen a Siamese-structure graph convolution model is employed to embed graphs\ninto vector forms. We finally combine the graph-level and the vector-level to\ncalculate the similarity of this image-text pair. The empirical experiments\nshow that our enhancement with the combination of levels can improve the\nperformance of the baseline method by increasing the recall by more than 10% on\nthe Flickr30k dataset."", 'The question we answer with this work is: can we convert a text document into\nan image to exploit best image classification models to classify documents? To\nanswer this question we present a novel text classification method which\nconverts a text document into an encoded image, using word embedding and\ncapabilities of Convolutional Neural Networks (CNNs), successfully employed in\nimage classification. We evaluate our approach by obtaining promising results\non some well-known benchmark datasets for text classification. This work allows\nthe application of many of the advanced CNN architectures developed for\nComputer Vision to Natural Language Processing. We test the proposed approach\non a multi-modal dataset, proving that it is possible to use a single deep\nmodel to represent text and image in the same feature space.', ""Although many methods have been proposed to deal with nature image\nsuper-resolution (SR) and get impressive performance, the text images SR is not\ngood due to their ignorance of document images. In this paper, we propose a\nmatting-based dual generative adversarial network (mdGAN) for document image\nSR. Firstly, the input image is decomposed into document text, foreground and\nbackground layers using deep image matting. Then two parallel branches are\nconstructed to recover text boundary information and color information\nrespectively. Furthermore, in order to improve the restoration accuracy of\ncharacters in output image, we use the input image's corresponding ground truth\ntext label as extra supervise information to refine the two-branch networks\nduring training. Experiments on real text images demonstrate that our method\noutperforms several state-of-the-art methods quantitatively and qualitatively."", 'A growing demand for natural-scene text detection has been witnessed by the\ncomputer vision community since text information plays a significant role in\nscene understanding and image indexing. Deep neural networks are being used due\nto their strong capabilities of pixel-wise classification or word localization,\nsimilar to being used in common vision problems. In this paper, we present a\nnovel two-task network with integrating bottom and top cues. The first task\naims to predict a pixel-by-pixel labeling and based on which, word proposals\nare generated with a canonical connected component analysis. The second task\naims to output a bundle of character candidates used later to verify the word\nproposals. The two sub-networks share base convolutional features and moreover,\nwe present a new loss to strengthen the interaction between them. We evaluate\nthe proposed network on public benchmark datasets and show it can detect\narbitrary-orientation scene text with a finer output boundary. In ICDAR 2013\ntext localization task, we achieve the state-of-the-art performance with an\nF-score of 0.919 and a much better recall of 0.915.', 'An unconstrained end-to-end text localization and recognition method is\npresented. The method detects initial text hypothesis in a single pass by an\nefficient region-based method and subsequently refines the text hypothesis\nusing a more robust local text model, which deviates from the common assumption\nof region-based methods that all characters are detected as connected\ncomponents.\n  Additionally, a novel feature based on character stroke area estimation is\nintroduced. The feature is efficiently computed from a region distance map, it\nis invariant to scaling and rotations and allows to efficiently detect text\nregions regardless of what portion of text they capture.\n  The method runs in real time and achieves state-of-the-art text localization\nand recognition results on the ICDAR 2013 Robust Reading dataset.', 'Daily engagement in life experiences is increasingly interwoven with mobile\ndevice use. Screen capture at the scale of seconds is being used in behavioral\nstudies and to implement ""just-in-time"" health interventions. The increasing\npsychological breadth of digital information will continue to make the actual\nscreens that people view a preferred if not required source of data about life\nexperiences. Effective and efficient Information Extraction and Retrieval from\ndigital screenshots is a crucial prerequisite to successful use of screen data.\nIn this paper, we present the experimental workflow we exploited to: (i)\npre-process a unique collection of screen captures, (ii) extract unstructured\ntext embedded in the images, (iii) organize image text and metadata based on a\nstructured schema, (iv) index the resulting document collection, and (v) allow\nfor Image Retrieval through a dedicated vertical search engine application. The\nadopted procedure integrates different open source libraries for traditional\nimage processing, Optical Character Recognition (OCR), and Image Retrieval. Our\naim is to assess whether and how state-of-the-art methodologies can be applied\nto this novel data set. We show how combining OpenCV-based pre-processing\nmodules with a Long short-term memory (LSTM) based release of Tesseract OCR,\nwithout ad hoc training, led to a 74% character-level accuracy of the extracted\ntext. Further, we used the processed repository as baseline for a dedicated\nImage Retrieval system, for the immediate use and application for behavioral\nand prevention scientists. We discuss issues of Text Information Extraction and\nRetrieval that are particular to the screenshot image case and suggest\nimportant future work.', 'Document Layout Analysis is a fundamental step in Handwritten Text Processing\nsystems, from the extraction of the text lines to the type of zone it belongs\nto. We present a system based on artificial neural networks which is able to\ndetermine not only the baselines of text lines present in the document, but\nalso performs geometric and logic layout analysis of the document. Experiments\nin three different datasets demonstrate the potential of the method and show\ncompetitive results with respect to state-of-the-art methods.', 'The Know Your Customer (KYC) and Anti Money Laundering (AML) are worldwide\npractices to online customer identification based on personal identification\ndocuments, similarity and liveness checking, and proof of address. To answer\nthe basic regulation question: are you whom you say you are? The customer needs\nto upload valid identification documents (ID). This task imposes some\ncomputational challenges since these documents are diverse, may present\ndifferent and complex backgrounds, some occlusion, partial rotation, poor\nquality, or damage. Advanced text and document segmentation algorithms were\nused to process the ID images. In this context, we investigated a method based\non U-Net to detect the document edges and text regions in ID images. Besides\nthe promising results on image segmentation, the U-Net based approach is\ncomputationally expensive for a real application, since the image segmentation\nis a customer device task. We propose a model optimization based on Octave\nConvolutions to qualify the method to situations where storage, processing, and\ntime resources are limited, such as in mobile and robotic applications. We\nconducted the evaluation experiments in two new datasets CDPhotoDataset and\nDTDDataset, which are composed of real ID images of Brazilian documents. Our\nresults showed that the proposed models are efficient to document segmentation\ntasks and portable.', 'Since real-world ubiquitous documents (e.g., invoices, tickets, resumes and\nleaflets) contain rich information, automatic document image understanding has\nbecome a hot topic. Most existing works decouple the problem into two separate\ntasks, (1) text reading for detecting and recognizing texts in images and (2)\ninformation extraction for analyzing and extracting key elements from\npreviously extracted plain text. However, they mainly focus on improving\ninformation extraction task, while neglecting the fact that text reading and\ninformation extraction are mutually correlated. In this paper, we propose a\nunified end-to-end text reading and information extraction network, where the\ntwo tasks can reinforce each other. Specifically, the multimodal visual and\ntextual features of text reading are fused for information extraction and in\nturn, the semantics in information extraction contribute to the optimization of\ntext reading. On three real-world datasets with diverse document images (from\nfixed layout to variable layout, from structured text to semi-structured text),\nour proposed method significantly outperforms the state-of-the-art methods in\nboth efficiency and accuracy.', 'Text detection in natural scene images for content analysis is an interesting\ntask. The research community has seen some great developments for\nEnglish/Mandarin text detection. However, Urdu text extraction in natural scene\nimages is a task not well addressed. In this work, firstly, a new dataset is\nintroduced for Urdu text in natural scene images. The dataset comprises of 500\nstandalone images acquired from real scenes. Secondly, the channel enhanced\nMaximally Stable Extremal Region (MSER) method is applied to extract Urdu text\nregions as candidates in an image. Two-stage filtering mechanism is applied to\neliminate non-candidate regions. In the first stage, text and noise are\nclassified based on their geometric properties. In the second stage, a support\nvector machine classifier is trained to discard non-text candidate regions.\nAfter this, text candidate regions are linked using centroid-based vertical and\nhorizontal distances. Text lines are further analyzed by a different classifier\nbased on HOG features to remove non-text regions. Extensive experimentation is\nperformed on the locally developed dataset to evaluate the performance. The\nexperimental results show good performance on test set images. The dataset will\nbe made available for research use. To the best of our knowledge, the work is\nthe first of its kind for the Urdu language and would provide a good dataset\nfor free research use and serve as a baseline performance on the task of Urdu\ntext extraction.', ""Previous deep learning based state-of-the-art scene text detection methods\ncan be roughly classified into two categories. The first category treats scene\ntext as a type of general objects and follows general object detection paradigm\nto localize scene text by regressing the text box locations, but troubled by\nthe arbitrary-orientation and large aspect ratios of scene text. The second one\nsegments text regions directly, but mostly needs complex post processing. In\nthis paper, we present a method that combines the ideas of the two types of\nmethods while avoiding their shortcomings. We propose to detect scene text by\nlocalizing corner points of text bounding boxes and segmenting text regions in\nrelative positions. In inference stage, candidate boxes are generated by\nsampling and grouping corner points, which are further scored by segmentation\nmaps and suppressed by NMS. Compared with previous methods, our method can\nhandle long oriented text naturally and doesn't need complex post processing.\nThe experiments on ICDAR2013, ICDAR2015, MSRA-TD500, MLT and COCO-Text\ndemonstrate that the proposed algorithm achieves better or comparable results\nin both accuracy and efficiency. Based on VGG16, it achieves an F-measure of\n84.3% on ICDAR2015 and 81.5% on MSRA-TD500."", 'Text detection, the key technology for understanding scene text, has become\nan attractive research topic. For detecting various scene texts, researchers\npropose plenty of detectors with different advantages: detection-based models\nenjoy fast detection speed, and segmentation-based algorithms are not limited\nby text shapes. However, for most intelligent systems, the detector needs to\ndetect arbitrary-shaped texts with high speed and accuracy simultaneously.\nThus, in this study, we design an efficient pipeline named as MT, which can\ndetect adhesive arbitrary-shaped texts with only a single binary mask in the\ninference stage. This paper presents the contributions on three aspects: (1) a\nlight-weight detection framework is designed to speed up the inference process\nwhile keeping high detection accuracy; (2) a multi-perspective feature module\nis proposed to learn more discriminative representations to segment the mask\naccurately; (3) a multi-factor constraints IoU minimization loss is introduced\nfor training the proposed model. The effectiveness of MT is evaluated on four\nreal-world scene text datasets, and it surpasses all the state-of-the-art\ncompetitors to a large extent.', 'Image-based sequence recognition has been a long-standing research topic in\ncomputer vision. In this paper, we investigate the problem of scene text\nrecognition, which is among the most important and challenging tasks in\nimage-based sequence recognition. A novel neural network architecture, which\nintegrates feature extraction, sequence modeling and transcription into a\nunified framework, is proposed. Compared with previous systems for scene text\nrecognition, the proposed architecture possesses four distinctive properties:\n(1) It is end-to-end trainable, in contrast to most of the existing algorithms\nwhose components are separately trained and tuned. (2) It naturally handles\nsequences in arbitrary lengths, involving no character segmentation or\nhorizontal scale normalization. (3) It is not confined to any predefined\nlexicon and achieves remarkable performances in both lexicon-free and\nlexicon-based scene text recognition tasks. (4) It generates an effective yet\nmuch smaller model, which is more practical for real-world application\nscenarios. The experiments on standard benchmarks, including the IIIT-5K,\nStreet View Text and ICDAR datasets, demonstrate the superiority of the\nproposed algorithm over the prior arts. Moreover, the proposed algorithm\nperforms well in the task of image-based music score recognition, which\nevidently verifies the generality of it.', 'Localizing text instances in natural scenes is regarded as a fundamental\nchallenge in computer vision. Nevertheless, owing to the extremely varied\naspect ratios and scales of text instances in real scenes, most conventional\ntext detectors suffer from the sub-text problem that only localizes the\nfragments of text instance (i.e., sub-texts). In this work, we quantitatively\nanalyze the sub-text problem and present a simple yet effective design,\nCOntrastive RElation (CORE) module, to mitigate that issue. CORE first\nleverages a vanilla relation block to model the relations among all text\nproposals (sub-texts of multiple text instances) and further enhances\nrelational reasoning via instance-level sub-text discrimination in a\ncontrastive manner. Such way naturally learns instance-aware representations of\ntext proposals and thus facilitates scene text detection. We integrate the CORE\nmodule into a two-stage text detector of Mask R-CNN and devise our text\ndetector CORE-Text. Extensive experiments on four benchmarks demonstrate the\nsuperiority of CORE-Text. Code is available:\n\\url{https://github.com/jylins/CORE-Text}.', 'Text classification is a fundamental task in NLP applications. Latest\nresearch in this field has largely been divided into two major sub-fields.\nLearning representations is one sub-field and learning deeper models, both\nsequential and convolutional, which again connects back to the representation\nis the other side. We posit the idea that the stronger the representation is,\nthe simpler classifier models are needed to achieve higher performance. In this\npaper we propose a completely novel direction to text classification research,\nwherein we convert text to a representation very similar to images, such that\nany deep network able to handle images is equally able to handle text. We take\na deeper look at the representation of documents as an image and subsequently\nutilize very simple convolution based models taken as is from computer vision\ndomain. This image can be cropped, re-scaled, re-sampled and augmented just\nlike any other image to work with most of the state-of-the-art large\nconvolution based models which have been designed to handle large image\ndatasets. We show impressive results with some of the latest benchmarks in the\nrelated fields. We perform transfer learning experiments, both from text to\ntext domain and also from image to text domain. We believe this is a paradigm\nshift from the way document understanding and text classification has been\ntraditionally done, and will drive numerous novel research ideas in the\ncommunity.', 'Recently, methods based on deep learning have dominated the field of text\nrecognition. With a large number of training data, most of them can achieve the\nstate-of-the-art performances. However, it is hard to harvest and label\nsufficient text sequence images from the real scenes. To mitigate this issue,\nseveral methods to synthesize text sequence images were proposed, yet they\nusually need complicated preceding or follow-up steps. In this work, we present\na method which is able to generate infinite training data without any auxiliary\npre/post-process. We tackle the generation task as an image-to-image\ntranslation one and utilize conditional adversarial networks to produce\nrealistic text sequence images in the light of the semantic ones. Some\nevaluation metrics are involved to assess our method and the results\ndemonstrate that the caliber of the data is satisfactory. The code and dataset\nwill be publicly available soon.', 'Detection and recognition of text in natural images are two main problems in\nthe field of computer vision that have a wide variety of applications in\nanalysis of sports videos, autonomous driving, industrial automation, to name a\nfew. They face common challenging problems that are factors in how text is\nrepresented and affected by several environmental conditions. The current\nstate-of-the-art scene text detection and/or recognition methods have exploited\nthe witnessed advancement in deep learning architectures and reported a\nsuperior accuracy on benchmark datasets when tackling multi-resolution and\nmulti-oriented text. However, there are still several remaining challenges\naffecting text in the wild images that cause existing methods to underperform\ndue to there models are not able to generalize to unseen data and the\ninsufficient labeled data. Thus, unlike previous surveys in this field, the\nobjectives of this survey are as follows: first, offering the reader not only a\nreview on the recent advancement in scene text detection and recognition, but\nalso presenting the results of conducting extensive experiments using a unified\nevaluation framework that assesses pre-trained models of the selected methods\non challenging cases, and applies the same evaluation criteria on these\ntechniques. Second, identifying several existing challenges for detecting or\nrecognizing text in the wild images, namely, in-plane-rotation, multi-oriented\nand multi-resolution text, perspective distortion, illumination reflection,\npartial occlusion, complex fonts, and special characters. Finally, the paper\nalso presents insight into the potential research directions in this field to\naddress some of the mentioned challenges that are still encountering scene text\ndetection and recognition techniques.', 'Document layout comprises both structural and visual (eg. font-sizes)\ninformation that is vital but often ignored by machine learning models. The few\nexisting models which do use layout information only consider textual contents,\nand overlook the existence of contents in other modalities such as images.\nAdditionally, spatial interactions of presented contents in a layout were never\nreally fully exploited. To bridge this gap, we parse a document into content\nblocks (eg. text, table, image) and propose a novel layout-aware multimodal\nhierarchical framework, LAMPreT, to model the blocks and the whole document.\nOur LAMPreT encodes each block with a multimodal transformer in the lower-level\nand aggregates the block-level representations and connections utilizing a\nspecifically designed transformer at the higher-level. We design hierarchical\npretraining objectives where the lower-level model is trained similarly to\nmultimodal grounding models, and the higher-level model is trained with our\nproposed novel layout-aware objectives. We evaluate the proposed model on two\nlayout-aware tasks -- text block filling and image suggestion and show the\neffectiveness of our proposed hierarchical architecture as well as pretraining\ntechniques.', 'We introduce a novel type of text representation that preserves the 2D layout\nof a document. This is achieved by encoding each document page as a\ntwo-dimensional grid of characters. Based on this representation, we present a\ngeneric document understanding pipeline for structured documents. This pipeline\nmakes use of a fully convolutional encoder-decoder network that predicts a\nsegmentation mask and bounding boxes. We demonstrate its capabilities on an\ninformation extraction task from invoices and show that it significantly\noutperforms approaches based on sequential text or document images.', 'In this paper, we are interested in editing text in natural images, which\naims to replace or modify a word in the source image with another one while\nmaintaining its realistic look. This task is challenging, as the styles of both\nbackground and text need to be preserved so that the edited image is visually\nindistinguishable from the source image. Specifically, we propose an end-to-end\ntrainable style retention network (SRNet) that consists of three modules: text\nconversion module, background inpainting module and fusion module. The text\nconversion module changes the text content of the source image into the target\ntext while keeping the original text style. The background inpainting module\nerases the original text, and fills the text region with appropriate texture.\nThe fusion module combines the information from the two former modules, and\ngenerates the edited text images. To our knowledge, this work is the first\nattempt to edit text in natural images at the word level. Both visual effects\nand quantitative results on synthetic and real-world dataset (ICDAR 2013) fully\nconfirm the importance and necessity of modular decomposition. We also conduct\nextensive experiments to validate the usefulness of our method in various\nreal-world applications such as text image synthesis, augmented reality (AR)\ntranslation, information hiding, etc.', 'Document image enhancement and binarization methods are often used to improve\nthe accuracy and efficiency of document image analysis tasks such as text\nrecognition. Traditional non-machine-learning methods are constructed on\nlow-level features in an unsupervised manner but have difficulty with\nbinarization on documents with severely degraded backgrounds. Convolutional\nneural network-based methods focus only on grayscale images and on local\ntextual features. In this paper, we propose a two-stage color document image\nenhancement and binarization method using generative adversarial neural\nnetworks. In the first stage, four color-independent adversarial networks are\ntrained to extract color foreground information from an input image for\ndocument image enhancement. In the second stage, two independent adversarial\nnetworks with global and local features are trained for image binarization of\ndocuments of variable size. For the adversarial neural networks, we formulate\nloss functions between a discriminator and generators having an encoder-decoder\nstructure. Experimental results show that the proposed method achieves better\nperformance than many classical and state-of-the-art algorithms over the\nDocument Image Binarization Contest (DIBCO) datasets, the LRDE Document\nBinarization Dataset (LRDE DBD), and our shipping label image dataset. We plan\nto release the shipping label dataset as well as our implementation code at\ngithub.com/opensuh/DocumentBinarization/.', 'Document Image Analysis, like any Digital Image Analysis requires\nidentification and extraction of proper features, which are generally extracted\nfrom uncompressed images, though in reality images are made available in\ncompressed form for the reasons such as transmission and storage efficiency.\nHowever, this implies that the compressed image should be decompressed, which\nindents additional computing resources. This limitation induces the motivation\nto research in extracting features directly from the compressed image. In this\nresearch, we propose to extract essential features such as projection profile,\nrun-histogram and entropy for text document analysis directly from run-length\ncompressed text-documents. The experimentation illustrates that features are\nextracted directly from the compressed image without going through the stage of\ndecompression, because of which the computing time is reduced. The feature\nvalues so extracted are exactly identical to those extracted from uncompressed\nimages.', 'This paper introduces a novel rotation-based framework for arbitrary-oriented\ntext detection in natural scene images. We present the Rotation Region Proposal\nNetworks (RRPN), which are designed to generate inclined proposals with text\norientation angle information. The angle information is then adapted for\nbounding box regression to make the proposals more accurately fit into the text\nregion in terms of the orientation. The Rotation Region-of-Interest (RRoI)\npooling layer is proposed to project arbitrary-oriented proposals to a feature\nmap for a text region classifier. The whole framework is built upon a\nregion-proposal-based architecture, which ensures the computational efficiency\nof the arbitrary-oriented text detection compared with previous text detection\nsystems. We conduct experiments using the rotation-based framework on three\nreal-world scene text detection datasets and demonstrate its superiority in\nterms of effectiveness and efficiency over previous approaches.', 'Scene text detection attracts much attention in computer vision, because it\ncan be widely used in many applications such as real-time text translation,\nautomatic information entry, blind person assistance, robot sensing and so on.\nThough many methods have been proposed for horizontal and oriented texts,\ndetecting irregular shape texts such as curved texts is still a challenging\nproblem. To solve the problem, we propose a robust scene text detection method\nwith adaptive text region representation. Given an input image, a text region\nproposal network is first used for extracting text proposals. Then, these\nproposals are verified and refined with a refinement network. Here, recurrent\nneural network based adaptive text region representation is proposed for text\nregion refinement, where a pair of boundary points are predicted each time step\nuntil no new points are found. In this way, text regions of arbitrary shapes\nare detected and represented with adaptive number of boundary points. This\ngives more accurate description of text regions. Experimental results on five\nbenchmarks, namely, CTW1500, TotalText, ICDAR2013, ICDAR2015 and MSRATD500,\nshow that the proposed method achieves state-of-the-art in scene text\ndetection.', ""The ubiquity of smartphone cameras has led to more and more documents being\ncaptured by cameras rather than scanned. Unlike flatbed scanners, photographed\ndocuments are often folded and crumpled, resulting in large local variance in\ntext structure. The problem of document rectification is fundamental to the\nOptical Character Recognition (OCR) process on documents, and its ability to\novercome geometric distortions significantly affects recognition accuracy.\nDespite the great progress in recent OCR systems, most still rely on a\npre-process that ensures the text lines are straight and axis aligned. Recent\nworks have tackled the problem of rectifying document images taken in-the-wild\nusing various supervision signals and alignment means. However, they focused on\nglobal features that can be extracted from the document's boundaries, ignoring\nvarious signals that could be obtained from the document's content.\n  We present CREASE: Content Aware Rectification using Angle Supervision, the\nfirst learned method for document rectification that relies on the document's\ncontent, the location of the words and specifically their orientation, as hints\nto assist in the rectification process. We utilize a novel pixel-wise angle\nregression approach and a curvature estimation side-task for optimizing our\nrectification model. Our method surpasses previous approaches in terms of OCR\naccuracy, geometric error and visual similarity."", 'Structured information extraction from document images usually consists of\nthree steps: text detection, text recognition, and text field labeling. While\ntext detection and text recognition have been heavily studied and improved a\nlot in literature, text field labeling is less explored and still faces many\nchallenges. Existing learning based methods for text labeling task usually\nrequire a large amount of labeled examples to train a specific model for each\ntype of document. However, collecting large amounts of document images and\nlabeling them is difficult and sometimes impossible due to privacy issues.\nDeploying separate models for each type of document also consumes a lot of\nresources. Facing these challenges, we explore one-shot learning for the text\nfield labeling task. Existing one-shot learning methods for the task are mostly\nrule-based and have difficulty in labeling fields in crowded regions with few\nlandmarks and fields consisting of multiple separate text regions. To alleviate\nthese problems, we proposed a novel deep end-to-end trainable approach for\none-shot text field labeling, which makes use of attention mechanism to\ntransfer the layout information between document images. We further applied\nconditional random field on the transferred layout information for the\nrefinement of field labeling. We collected and annotated a real-world one-shot\nfield labeling dataset with a large variety of document types and conducted\nextensive experiments to examine the effectiveness of the proposed model. To\nstimulate research in this direction, the collected dataset and the one-shot\nmodel will be released1.', 'Text in curve orientation, despite being one of the common text orientations\nin real world environment, has close to zero existence in well received scene\ntext datasets such as ICDAR2013 and MSRA-TD500. The main motivation of\nTotal-Text is to fill this gap and facilitate a new research direction for the\nscene text community. On top of the conventional horizontal and multi-oriented\ntexts, it features curved-oriented text. Total-Text is highly diversified in\norientations, more than half of its images have a combination of more than two\norientations. Recently, a new breed of solutions that casted text detection as\na segmentation problem has demonstrated their effectiveness against\nmulti-oriented text. In order to evaluate its robustness against curved text,\nwe fine-tuned DeconvNet and benchmark it on Total-Text. Total-Text with its\nannotation is available at https://github.com/cs-chan/Total-Text-Dataset', 'India is a multilingual multi-script country. In every state of India there\nare two languages one is state local language and the other is English. For\nexample in Andhra Pradesh, a state in India, the document may contain text\nwords in English and Telugu script. For Optical Character Recognition (OCR) of\nsuch a bilingual document, it is necessary to identify the script before\nfeeding the text words to the OCRs of individual scripts. In this paper, we are\nintroducing a simple and efficient technique of script identification for\nKannada, English and Hindi text words of a printed document. The proposed\napproach is based on the horizontal and vertical projection profile for the\ndiscrimination of the three scripts. The feature extraction is done based on\nthe horizontal projection profile of each text words. We analysed 700 different\nwords of Kannada, English and Hindi in order to extract the discrimination\nfeatures and for the development of knowledge base. We use the horizontal\nprojection profile of each text word and based on the horizontal projection\nprofile we extract the appropriate features. The proposed system is tested on\n100 different document images containing more than 1000 text words of each\nscript and a classification rate of 98.25%, 99.25% and 98.87% is achieved for\nKannada, English and Hindi respectively.', 'In this paper, we propose a way of synthesizing realistic images directly\nwith natural language description, which has many useful applications, e.g.\nintelligent image manipulation. We attempt to accomplish such synthesis: given\na source image and a target text description, our model synthesizes images to\nmeet two requirements: 1) being realistic while matching the target text\ndescription; 2) maintaining other image features that are irrelevant to the\ntext description. The model should be able to disentangle the semantic\ninformation from the two modalities (image and text), and generate new images\nfrom the combined semantics. To achieve this, we proposed an end-to-end neural\narchitecture that leverages adversarial learning to automatically learn\nimplicit loss functions, which are optimized to fulfill the aforementioned two\nrequirements. We have evaluated our model by conducting experiments on\nCaltech-200 bird dataset and Oxford-102 flower dataset, and have demonstrated\nthat our model is capable of synthesizing realistic images that match the given\ndescriptions, while still maintain other features of original images.', 'Previous approaches for scene text detection usually rely on manually defined\nsliding windows. This work presents an intuitive two-stage region-based method\nto detect multi-oriented text without any prior knowledge regarding the textual\nshape. In the first stage, we estimate the possible locations of text instances\nby detecting and linking corners instead of shifting a set of default anchors.\nThe quadrilateral proposals are geometry adaptive, which allows our method to\ncope with various text aspect ratios and orientations. In the second stage, we\ndesign a new pooling layer named Dual-RoI Pooling which embeds data\naugmentation inside the region-wise subnetwork for more robust classification\nand regression over these proposals. Experimental results on public benchmarks\nconfirm that the proposed method is capable of achieving comparable performance\nwith state-of-the-art methods. The code is publicly available at\nhttps://github.com/xhzdeng/crpn', 'Previous approaches for scene text detection have already achieved promising\nperformances across various benchmarks. However, they usually fall short when\ndealing with challenging scenarios, even when equipped with deep neural network\nmodels, because the overall performance is determined by the interplay of\nmultiple stages and components in the pipelines. In this work, we propose a\nsimple yet powerful pipeline that yields fast and accurate text detection in\nnatural scenes. The pipeline directly predicts words or text lines of arbitrary\norientations and quadrilateral shapes in full images, eliminating unnecessary\nintermediate steps (e.g., candidate aggregation and word partitioning), with a\nsingle neural network. The simplicity of our pipeline allows concentrating\nefforts on designing loss functions and neural network architecture.\nExperiments on standard datasets including ICDAR 2015, COCO-Text and MSRA-TD500\ndemonstrate that the proposed algorithm significantly outperforms\nstate-of-the-art methods in terms of both accuracy and efficiency. On the ICDAR\n2015 dataset, the proposed algorithm achieves an F-score of 0.7820 at 13.2fps\nat 720p resolution.', 'The technological advancement and sophistication in cameras and gadgets\nprompt researchers to have focus on image analysis and text understanding. The\ndeep learning techniques demonstrated well to assess the potential for\nclassifying text from natural scene images as reported in recent years. There\nare variety of deep learning approaches that prospects the detection and\nrecognition of text, effectively from images. In this work, we presented Arabic\nscene text recognition using Convolutional Neural Networks (ConvNets) as a deep\nlearning classifier. As the scene text data is slanted and skewed, thus to deal\nwith maximum variations, we employ five orientations with respect to single\noccurrence of a character. The training is formulated by keeping filter size 3\nx 3 and 5 x 5 with stride value as 1 and 2. During text classification phase,\nwe trained network with distinct learning rates. Our approach reported\nencouraging results on recognition of Arabic characters from segmented Arabic\nscene images.', 'Many historical map sheets are publicly available for studies that require\nlong-term historical geographic data. The cartographic design of these maps\nincludes a combination of map symbols and text labels. Automatically reading\ntext labels from map images could greatly speed up the map interpretation and\nhelps generate rich metadata describing the map content. Many text detection\nalgorithms have been proposed to locate text regions in map images\nautomatically, but most of the algorithms are trained on out-ofdomain datasets\n(e.g., scenic images). Training data determines the quality of machine learning\nmodels, and manually annotating text regions in map images is labor-extensive\nand time-consuming. On the other hand, existing geographic data sources, such\nas Open- StreetMap (OSM), contain machine-readable map layers, which allow us\nto separate out the text layer and obtain text label annotations easily.\nHowever, the cartographic styles between OSM map tiles and historical maps are\nsignificantly different. This paper proposes a method to automatically generate\nan unlimited amount of annotated historical map images for training text\ndetection models. We use a style transfer model to convert contemporary map\nimages into historical style and place text labels upon them. We show that the\nstate-of-the-art text detection models (e.g., PSENet) can benefit from the\nsynthetic historical maps and achieve significant improvement for historical\nmap text detection.', 'Many tasks are related to determining if a particular text string exists in\nan image. In this work, we propose a new framework that learns this task in an\nend-to-end way. The framework takes an image and a text string as input and\nthen outputs the probability of the text string being present in the image.\nThis is the first end-to-end framework that learns such relationships between\ntext and images in scene text area. The framework does not require explicit\nscene text detection or recognition and thus no bounding box annotations are\nneeded for it. It is also the first work in scene text area that tackles suh a\nweakly labeled problem. Based on this framework, we developed a model called\nGuided Attention. Our designed model achieves much better results than several\nstate-of-the-art scene text reading based solutions for a challenging Street\nView Business Matching task. The task tries to find correct business names for\nstorefront images and the dataset we collected for it is substantially larger,\nand more challenging than existing scene text dataset. This new real-world task\nprovides a new perspective for studying scene text related problems. We also\ndemonstrate the uniqueness of our task via a comparison between our problem and\na typical Visual Question Answering problem.', ""In this paper, we investigate the problem of retrieving images from a\ndatabase based on a multi-modal (image-text) query. Specifically, the query\ntext prompts some modification in the query image and the task is to retrieve\nimages with the desired modifications. For instance, a user of an E-Commerce\nplatform is interested in buying a dress, which should look similar to her\nfriend's dress, but the dress should be of white color with a ribbon sash. In\nthis case, we would like the algorithm to retrieve some dresses with desired\nmodifications in the query dress. We propose an autoencoder based model,\nComposeAE, to learn the composition of image and text query for retrieving\nimages. We adopt a deep metric learning approach and learn a metric that pushes\ncomposition of source image and text query closer to the target images. We also\npropose a rotational symmetry constraint on the optimization problem. Our\napproach is able to outperform the state-of-the-art method TIRG \\cite{TIRG} on\nthree benchmark datasets, namely: MIT-States, Fashion200k and Fashion IQ. In\norder to ensure fair comparison, we introduce strong baselines by enhancing\nTIRG method. To ensure reproducibility of the results, we publish our code\nhere: \\url{https://github.com/ecom-research/ComposeAE}."", 'Video text spotting is still an important research topic due to its various\nreal-applications. Previous approaches usually fall into the four-staged\npipeline: text detection in individual images, framewisely recognizing\nlocalized text regions, tracking text streams and generating final results with\ncomplicated post-processing skills, which might suffer from the huge\ncomputational cost as well as the interferences of low-quality text. In this\npaper, we propose a fast and robust video text spotting framework by only\nrecognizing the localized text one-time instead of frame-wisely recognition.\nSpecifically, we first obtain text regions in videos with a well-designed\nspatial-temporal detector. Then we concentrate on developing a novel text\nrecommender for selecting the highest-quality text from text streams and only\nrecognizing the selected ones. Here, the recommender assembles text tracking,\nquality scoring and recognition into an end-to-end trainable module, which not\nonly avoids the interferences from low-quality text but also dramatically\nspeeds up the video text spotting process. In addition, we collect a larger\nscale video text dataset (LSVTD) for promoting the video text spotting\ncommunity, which contains 100 text videos from 22 different real-life\nscenarios. Extensive experiments on two public benchmarks show that our method\ngreatly speeds up the recognition process averagely by 71 times compared with\nthe frame-wise manner, and also achieves the remarkable state-of-the-art.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', ""In this paper, we study the compositional learning of images and texts for\nimage retrieval. The query is given in the form of an image and text that\ndescribes the desired modifications to the image; the goal is to retrieve the\ntarget image that satisfies the given modifications and resembles the query by\ncomposing information in both the text and image modalities. To remedy this, we\npropose a novel architecture designed for the image-text composition task and\nshow that the proposed structure can effectively encode the differences between\nthe source and target images conditioned on the text. Furthermore, we introduce\na new joint training technique based on the graph convolutional network that is\ngenerally applicable for any existing composition methods in a plug-and-play\nmanner. We found that the proposed technique consistently improves performance\nand achieves state-of-the-art scores on various benchmarks. To avoid misleading\nexperimental results caused by trivial training hyper-parameters, we reproduce\nall individual baselines and train models with a unified training environment.\nWe expect this approach to suppress undesirable effects from irrelevant\ncomponents and emphasize the image-text composition module's ability. Also, we\nachieve the state-of-the-art score without restricting the training\nenvironment, which implies the superiority of our method considering the gains\nfrom hyper-parameter tuning. The code, including all the baseline methods, are\nreleased https://github.com/nashory/rtic-gcn-pytorch."", 'Image foreground extraction is a classical problem in image processing and\nvision, with a large range of applications. In this dissertation, we focus on\nthe extraction of text and graphics in mixed-content images, and design novel\napproaches for various aspects of this problem.\n  We first propose a sparse decomposition framework, which models the\nbackground by a subspace containing smooth basis vectors, and foreground as a\nsparse and connected component. We then formulate an optimization framework to\nsolve this problem, by adding suitable regularizations to the cost function to\npromote the desired characteristics of each component. We present two\ntechniques to solve the proposed optimization problem, one based on alternating\ndirection method of multipliers (ADMM), and the other one based on robust\nregression. Promising results are obtained for screen content image\nsegmentation using the proposed algorithm.\n  We then propose a robust subspace learning algorithm for the representation\nof the background component using training images that could contain both\nbackground and foreground components, as well as noise. With the learnt\nsubspace for the background, we can further improve the segmentation results,\ncompared to using a fixed subspace. Lastly, we investigate a different class of\nsignal/image decomposition problem, where only one signal component is active\nat each signal element. In this case, besides estimating each component, we\nneed to find their supports, which can be specified by a binary mask. We\npropose a mixed-integer programming problem, that jointly estimates the two\ncomponents and their supports through an alternating optimization scheme. We\nshow the application of this algorithm on various problems, including image\nsegmentation, video motion segmentation, and also separation of text from\ntextured images.', ""Scene text is an important feature to be extracted, especially in\nvision-based mobile robot navigation as many potential landmarks such as\nnameplates and information signs contain text. In this paper, a novel two-step\ntext localization method for Indoor Mobile Robot Navigation is introduced. This\nmethod is based on morphological operators and machine learning techniques and\ncan be used in real time environments. The proposed method has two steps. At\nFirst, a new set of morphological operators is applied with a particular\nsequence to extract high contrast areas that have high probability of text\nexistence. Using of morphological operators has many advantages such as: high\ncomputation speed, being invariant to several geometrical transformations like\ntranslation, rotations, and scaling, and being able to extract all areas\ncontaining text. After extracting text candidate regions, a set of nine\nfeatures are extracted for accurate detection and deletion of the regions that\ndon't have text. These features are descriptors for texture properties and are\ncomputed in real time. Then, we use a SVM classifier to detect the existence of\ntext in the region. Performance of the proposed algorithm is compared against a\nnumber of widely used text localization algorithms and the results show that\nthis method can quickly and effectively localize and extract text regions from\nreal scenes and can be used in mobile robot navigation under an indoor\nenvironment to detect text based landmarks."", 'Recognizing arbitrary multi-character text in unconstrained natural\nphotographs is a hard problem. In this paper, we address an equally hard\nsub-problem in this domain viz. recognizing arbitrary multi-digit numbers from\nStreet View imagery. Traditional approaches to solve this problem typically\nseparate out the localization, segmentation, and recognition steps. In this\npaper we propose a unified approach that integrates these three steps via the\nuse of a deep convolutional neural network that operates directly on the image\npixels. We employ the DistBelief implementation of deep neural networks in\norder to train large, distributed neural networks on high quality images. We\nfind that the performance of this approach increases with the depth of the\nconvolutional network, with the best performance occurring in the deepest\narchitecture we trained, with eleven hidden layers. We evaluate this approach\non the publicly available SVHN dataset and achieve over $96\\%$ accuracy in\nrecognizing complete street numbers. We show that on a per-digit recognition\ntask, we improve upon the state-of-the-art, achieving $97.84\\%$ accuracy. We\nalso evaluate this approach on an even more challenging dataset generated from\nStreet View imagery containing several tens of millions of street number\nannotations and achieve over $90\\%$ accuracy. To further explore the\napplicability of the proposed system to broader text recognition tasks, we\napply it to synthetic distorted text from reCAPTCHA. reCAPTCHA is one of the\nmost secure reverse turing tests that uses distorted text to distinguish humans\nfrom bots. We report a $99.8\\%$ accuracy on the hardest category of reCAPTCHA.\nOur evaluations on both tasks indicate that at specific operating thresholds,\nthe performance of the proposed system is comparable to, and in some cases\nexceeds, that of human operators.', 'Handwritten document images can be highly affected by degradation for\ndifferent reasons: Paper ageing, daily-life scenarios (wrinkles, dust, etc.),\nbad scanning process and so on. These artifacts raise many readability issues\nfor current Handwritten Text Recognition (HTR) algorithms and severely devalue\ntheir efficiency. In this paper, we propose an end to end architecture based on\nGenerative Adversarial Networks (GANs) to recover the degraded documents into a\nclean and readable form. Unlike the most well-known document binarization\nmethods, which try to improve the visual quality of the degraded document, the\nproposed architecture integrates a handwritten text recognizer that promotes\nthe generated document image to be more readable. To the best of our knowledge,\nthis is the first work to use the text information while binarizing handwritten\ndocuments. Extensive experiments conducted on degraded Arabic and Latin\nhandwritten documents demonstrate the usefulness of integrating the recognizer\nwithin the GAN architecture, which improves both the visual quality and the\nreadability of the degraded document images. Moreover, we outperform the state\nof the art in H-DIBCO challenges, after fine tuning our pre-trained model with\nsynthetically degraded Latin handwritten images, on this task.', 'This paper presents results of Document Visual Question Answering Challenge\norganized as part of ""Text and Documents in the Deep Learning Era"" workshop, in\nCVPR 2020. The challenge introduces a new problem - Visual Question Answering\non document images. The challenge comprised two tasks. The first task concerns\nwith asking questions on a single document image. On the other hand, the second\ntask is set as a retrieval task where the question is posed over a collection\nof images. For the task 1 a new dataset is introduced comprising 50,000\nquestions-answer(s) pairs defined over 12,767 document images. For task 2\nanother dataset has been created comprising 20 questions over 14,362 document\nimages which share the same document template.', 'Line separators are used to segregate text-lines from one another in document\nimage analysis. Finding the separator points at every line terminal in a\ndocument image would enable text-line segmentation. In particular, identifying\nthe separators in handwritten text could be a thrilling exercise. Obviously it\nwould be challenging to perform this in the compressed version of a document\nimage and that is the proposed objective in this research. Such an effort would\nprevent the computational burden of decompressing a document for text-line\nsegmentation. Since document images are generally compressed using run length\nencoding (RLE) technique as per the CCITT standards, the first column in the\nRLE will be a white column. The value (depth) in the white column is very low\nwhen a particular line is a text line and the depth could be larger at the\npoint of text line separation. A longer consecutive sequence of such larger\ndepth should indicate the gap between the text lines, which provides the\nseparator region. In case of over separation and under separation issues,\ncorrective actions such as deletion and insertion are suggested respectively.\nAn extensive experimentation is conducted on the compressed images of the\nbenchmark datasets of ICDAR13 and Alireza et al [17] to demonstrate the\nefficacy.', 'We introduce FontCode, an information embedding technique for text documents.\nProvided a text document with specific fonts, our method embeds user-specified\ninformation in the text by perturbing the glyphs of text characters while\npreserving the text content. We devise an algorithm to chooses unobtrusive yet\nmachine-recognizable glyph perturbations, leveraging a recently developed\ngenerative model that alters the glyphs of each character continuously on a\nfont manifold. We then introduce an algorithm that embeds a user-provided\nmessage in the text document and produces an encoded document whose appearance\nis minimally perturbed from the original document. We also present a glyph\nrecognition method that recovers the embedded information from an encoded\ndocument stored as a vector graphic or pixel image, or even on a printed paper.\nIn addition, we introduce a new error-correction coding scheme that rectifies a\ncertain number of recognition errors. Lastly, we demonstrate that our technique\nenables a wide array of applications, using it as a text document metadata\nholder, an unobtrusive optical barcode, a cryptographic message embedding\nscheme, and a text document signature.', 'Document segmentation is a method of rending the document into distinct\nregions. A document is an assortment of information and a standard mode of\nconveying information to others. Pursuance of data from documents involves ton\nof human effort, time intense and might severely prohibit the usage of data\nsystems. So, automatic information pursuance from the document has become a big\nissue. It is been shown that document segmentation will facilitate to beat such\nproblems. This paper proposes a new approach to segment and classify the\ndocument regions as text, image, drawings and table. Document image is divided\ninto blocks using Run length smearing rule and features are extracted from\nevery blocks. Discipulus tool has been used to construct the Genetic\nprogramming based classifier model and located 97.5% classification accuracy.', 'In natural scenes and documents, we can find the correlation between a text\nand its color. For instance, the word, ""hot"", is often printed in red, while\n""cold"" is often in blue. This correlation can be thought of as a feature that\nrepresents the semantic difference between the words. Based on this\nobservation, we propose the idea of using text color for word embeddings. While\ntext-only word embeddings (e.g. word2vec) have been extremely successful, they\noften represent antonyms as similar since they are often interchangeable in\nsentences. In this paper, we try two tasks to verify the usefulness of text\ncolor in understanding the meanings of words, especially in identifying\nsynonyms and antonyms. First, we quantify the color distribution of words from\nthe book cover images and analyze the correlation between the color and meaning\nof the word. Second, we try to retrain word embeddings with the color\ndistribution of words as a constraint. By observing the changes in the word\nembeddings of synonyms and antonyms before and after re-training, we aim to\nunderstand the kind of words that have positive or negative effects in their\nword embeddings when incorporating text color information.', 'This paper describes the COCO-Text dataset. In recent years large-scale\ndatasets like SUN and Imagenet drove the advancement of scene understanding and\nobject recognition. The goal of COCO-Text is to advance state-of-the-art in\ntext detection and recognition in natural images. The dataset is based on the\nMS COCO dataset, which contains images of complex everyday scenes. The images\nwere not collected with text in mind and thus contain a broad variety of text\ninstances. To reflect the diversity of text in natural scenes, we annotate text\nwith (a) location in terms of a bounding box, (b) fine-grained classification\ninto machine printed text and handwritten text, (c) classification into legible\nand illegible text, (d) script of the text and (e) transcriptions of legible\ntext. The dataset contains over 173k text annotations in over 63k images. We\nprovide a statistical analysis of the accuracy of our annotations. In addition,\nwe present an analysis of three leading state-of-the-art photo Optical\nCharacter Recognition (OCR) approaches on our dataset. While scene text\ndetection and recognition enjoys strong advances in recent years, we identify\nsignificant shortcomings motivating future work.', 'We present a novel data generation tool for document processing. The tool\nfocuses on providing a maximal level of visual information in a normal type\ndocument, ranging from character position to paragraph-level position. It also\nenables working with a large dataset on low-resource languages as well as\nproviding a mean of processing thorough full-level information of the\ndocumented text. The data generation tools come with a dataset of 320000\nVietnamese synthetic document images and an instruction to generate a dataset\nof similar size in other languages. The repository can be found at:\nhttps://github.com/tson1997/SDL-Document-Image-Generation', 'To understand how people look, interact, or perform tasks, we need to quickly\nand accurately capture their 3D body, face, and hands together from an RGB\nimage. Most existing methods focus only on parts of the body. A few recent\napproaches reconstruct full expressive 3D humans from images using 3D body\nmodels that include the face and hands. These methods are optimization-based\nand thus slow, prone to local optima, and require 2D keypoints as input. We\naddress these limitations by introducing ExPose (EXpressive POse and Shape\nrEgression), which directly regresses the body, face, and hands, in SMPL-X\nformat, from an RGB image. This is a hard problem due to the high\ndimensionality of the body and the lack of expressive training data.\nAdditionally, hands and faces are much smaller than the body, occupying very\nfew image pixels. This makes hand and face estimation hard when body images are\ndownscaled for neural networks. We make three main contributions. First, we\naccount for the lack of training data by curating a dataset of SMPL-X fits on\nin-the-wild images. Second, we observe that body estimation localizes the face\nand hands reasonably well. We introduce body-driven attention for face and hand\nregions in the original image to extract higher-resolution crops that are fed\nto dedicated refinement modules. Third, these modules exploit part-specific\nknowledge from existing face- and hand-only datasets. ExPose estimates\nexpressive 3D humans more accurately than existing optimization methods at a\nsmall fraction of the computational cost. Our data, model and code are\navailable for research at https://expose.is.tue.mpg.de .', 'Most existing text reading benchmarks make it difficult to evaluate the\nperformance of more advanced deep learning models in large vocabularies due to\nthe limited amount of training data. To address this issue, we introduce a new\nlarge-scale text reading benchmark dataset named Chinese Street View Text\n(C-SVT) with 430,000 street view images, which is at least 14 times as large as\nthe existing Chinese text reading benchmarks. To recognize Chinese text in the\nwild while keeping large-scale datasets labeling cost-effective, we propose to\nannotate one part of the CSVT dataset (30,000 images) in locations and text\nlabels as full annotations and add 400,000 more images, where only the\ncorresponding text-of-interest in the regions is given as weak annotations. To\nexploit the rich information from the weakly annotated data, we design a text\nreading network in a partially supervised learning framework, which enables to\nlocalize and recognize text, learn from fully and weakly annotated data\nsimultaneously. To localize the best matched text proposals from weakly labeled\nimages, we propose an online proposal matching module incorporated in the whole\nmodel, spotting the keyword regions by sharing parameters for end-to-end\ntraining. Compared with fully supervised training algorithms, this model can\nimprove the end-to-end recognition performance remarkably by 4.03% in F-score\nat the same labeling cost. The proposed model can also achieve state-of-the-art\nresults on the ICDAR 2017-RCTW dataset, which demonstrates the effectiveness of\nthe proposed partially supervised learning framework.', 'Citizen engagement and technology usage are two emerging trends driven by\nsmart city initiatives. Governments around the world are adopting technology\nfor faster resolution of civic issues. Typically, citizens report issues, such\nas broken roads, garbage dumps, etc. through web portals and mobile apps, in\norder for the government authorities to take appropriate actions. Several\nmediums -- text, image, audio, video -- are used to report these issues.\nThrough a user study with 13 citizens and 3 authorities, we found that image is\nthe most preferred medium to report civic issues. However, analyzing civic\nissue related images is challenging for the authorities as it requires manual\neffort. Moreover, previous works have been limited to identifying a specific\nset of issues from images. In this work, given an image, we propose to generate\na Civic Issue Graph consisting of a set of objects and the semantic relations\nbetween them, which are representative of the underlying civic issue. We also\nrelease two multi-modal (text and images) datasets, that can help in further\nanalysis of civic issues from images. We present a novel approach for\nadversarial training of existing scene graph models that enables the use of\nscene graphs for new applications in the absence of any labelled training data.\nWe conduct several experiments to analyze the efficacy of our approach, and\nusing human evaluation, we establish the appropriateness of our model at\nrepresenting different civic issues.', 'The camera captured images have various aspects to investigate. Generally,\nthe emphasis of research depends on the interesting regions. Sometimes the\nfocus could be on color segmentation, object detection or scene text analysis.\nThe image analysis, visibility and layout analysis are the tasks easier for\nhumans as suggested by behavioral trait of humans, but in contrast when these\nsame tasks are supposed to perform by machines then it seems to be challenging.\nThe learning machines always learn from the properties associated to provided\nsamples. The numerous approaches are designed in recent years for scene text\nextraction and recognition and the efforts are underway to improve the\naccuracy. The convolutional approach provided reasonable results on non-cursive\ntext analysis appeared in natural images. The work presented in this manuscript\nexploited the strength of linear pyramids by considering each pyramid as a\nfeature of the provided sample. Each pyramid image process through various\nempirically selected kernels. The performance was investigated by considering\nArabic text on each image pyramid of EASTR-42k dataset. The error rate of 0.17%\nwas reported on Arabic scene text recognition.', ""Neural Networks and Deep Learning have seen an upsurge of research in the\npast decade due to the improved results. Generates text from the given image is\na crucial task that requires the combination of both sectors which are computer\nvision and natural language processing in order to understand an image and\nrepresent it using a natural language. However existing works have all been\ndone on a particular lingual domain and on the same set of data. This leads to\nthe systems being developed to perform poorly on images that belong to specific\nlocales' geographical context. TextMage is a system that is capable of\nunderstanding visual scenes that belong to the Bangladeshi geographical context\nand use its knowledge to represent what it understands in Bengali. Hence, we\nhave trained a model on our previously developed and published dataset named\nBanglaLekhaImageCaptions. This dataset contains 9,154 images along with two\nannotations for each image. In order to access performance, the proposed model\nhas been implemented and evaluated."", 'The text-independent approach to writer identification does not require the\nwriter to write some predetermined text. Previous research on text-independent\nwriter identification has been based on identifying writer-specific features\ndesigned by experts. However, in the last decade, deep learning methods have\nbeen successfully applied to learn features from data automatically. We propose\nhere an end-to-end deep-learning method for text-independent writer\nidentification that does not require prior identification of features. A\nConvolutional Neural Network (CNN) is trained initially to extract local\nfeatures, which represent characteristics of individual handwriting in the\nwhole character images and their sub-regions. Randomly sampled tuples of images\nfrom the training set are used to train the CNN and aggregate the extracted\nlocal features of images from the tuples to form global features. For every\ntraining epoch, the process of randomly sampling tuples is repeated, which is\nequivalent to a large number of training patterns being prepared for training\nthe CNN for text-independent writer identification. We conducted experiments on\nthe JEITA-HP database of offline handwritten Japanese character patterns. With\n200 characters, our method achieved an accuracy of 99.97% to classify 100\nwriters. Even when using 50 characters for 100 writers or 100 characters for\n400 writers, our method achieved accuracy levels of 92.80% or 93.82%,\nrespectively. We conducted further experiments on the Firemaker and IAM\ndatabases of offline handwritten English text. Using only one page per writer\nto train, our method achieved over 91.81% accuracy to classify 900 writers.\nOverall, we achieved a better performance than the previously published best\nresult based on handcrafted features and clustering algorithms, which\ndemonstrates the effectiveness of our method for handwritten English text also.', 'Video game genre classification based on its cover and textual description\nwould be utterly beneficial to many modern identification, collocation, and\nretrieval systems. At the same time, it is also an extremely challenging task\ndue to the following reasons: First, there exists a wide variety of video game\ngenres, many of which are not concretely defined. Second, video game covers\nvary in many different ways such as colors, styles, textual information, etc,\neven for games of the same genre. Third, cover designs and textual descriptions\nmay vary due to many external factors such as country, culture, target reader\npopulations, etc. With the growing competitiveness in the video game industry,\nthe cover designers and typographers push the cover designs to its limit in the\nhope of attracting sales. The computer-based automatic video game genre\nclassification systems become a particularly exciting research topic in recent\nyears. In this paper, we propose a multi-modal deep learning framework to solve\nthis problem. The contribution of this paper is four-fold. First, we compiles a\nlarge dataset consisting of 50,000 video games from 21 genres made of cover\nimages, description text, and title text and the genre information. Second,\nimage-based and text-based, state-of-the-art models are evaluated thoroughly\nfor the task of genre classification for video games. Third, we developed an\nefficient and salable multi-modal framework based on both images and texts.\nFourth, a thorough analysis of the experimental results is given and future\nworks to improve the performance is suggested. The results show that the\nmulti-modal framework outperforms the current state-of-the-art image-based or\ntext-based models. Several challenges are outlined for this task. More efforts\nand resources are needed for this classification task in order to reach a\nsatisfactory level.', 'We model the process of human full interpretation of object images, namely\nthe ability to identify and localize all semantic features and parts that are\nrecognized by human observers. The task is approached by dividing the\ninterpretation of the complete object to the interpretation of multiple reduced\nbut interpretable local regions. We model interpretation by a structured\nlearning framework, in which there are primitive components and relations that\nplay a useful role in local interpretation by humans. To identify useful\ncomponents and relations used in the interpretation process, we consider the\ninterpretation of minimal configurations, namely reduced local regions that are\nminimal in the sense that further reduction will turn them unrecognizable and\nuninterpretable. We show experimental results of our model, and results of\npredicting and testing relations that were useful to the model via transformed\nminimal images.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'We present in this paper a new approach for the automatic annotation of\nmedical images, using the approach of ""bag-of-words"" to represent the visual\ncontent of the medical image combined with text descriptors based approach\ntf.idf and reduced by latent semantic to extract the co-occurrence between\nterms and visual terms. A medical report is composed of a text describing a\nmedical image. First, we are interested to index the text and extract all\nrelevant terms using a thesaurus containing MeSH medical concepts. In a second\nphase, the medical image is indexed while recovering areas of interest which\nare invariant to change in scale, light and tilt. To annotate a new medical\nimage, we use the approach of ""bagof-words"" to recover the feature vector.\nIndeed, we use the vector space model to retrieve similar medical image from\nthe database training. The calculation of the relevance value of an image to\nthe query image is based on the cosine function. We conclude with an experiment\ncarried out on five types of radiological imaging to evaluate the performance\nof our system of medical annotation. The results showed that our approach works\nbetter with more images from the radiology of the skull.', 'In recent scene recognition research images or large image regions are often\nrepresented as disorganized ""bags"" of features which can then be analyzed using\nmodels originally developed to capture co-variation of word counts in text.\nHowever, image feature counts are likely to be constrained in different ways\nthan word counts in text. For example, as a camera pans upwards from a building\nentrance over its first few floors and then further up into the sky Fig. 1,\nsome feature counts in the image drop while others rise -- only to drop again\ngiving way to features found more often at higher elevations. The space of all\npossible feature count combinations is constrained both by the properties of\nthe larger scene and the size and the location of the window into it. To\ncapture such variation, in this paper we propose the use of the counting grid\nmodel. This generative model is based on a grid of feature counts, considerably\nlarger than any of the modeled images, and considerably smaller than the real\nestate needed to tile the images next to each other tightly. Each modeled image\nis assumed to have a representative window in the grid in which the feature\ncounts mimic the feature distribution in the image. We provide a learning\nprocedure that jointly maps all images in the training set to the counting grid\nand estimates the appropriate local counts in it. Experimentally, we\ndemonstrate that the resulting representation captures the space of feature\ncount combinations more accurately than the traditional models, not only when\nthe input images come from a panning camera, but even when modeling images of\ndifferent scenes from the same category.', 'Identity documents recognition is an important sub-field of document\nanalysis, which deals with tasks of robust document detection, type\nidentification, text fields recognition, as well as identity fraud prevention\nand document authenticity validation given photos, scans, or video frames of an\nidentity document capture. Significant amount of research has been published on\nthis topic in recent years, however a chief difficulty for such research is\nscarcity of datasets, due to the subject matter being protected by security\nrequirements. A few datasets of identity documents which are available lack\ndiversity of document types, capturing conditions, or variability of document\nfield values. In addition, the published datasets were typically designed only\nfor a subset of document recognition problems, not for a complex identity\ndocument analysis. In this paper, we present a dataset MIDV-2020 which consists\nof 1000 video clips, 2000 scanned images, and 1000 photos of 1000 unique mock\nidentity documents, each with unique text field values and unique artificially\ngenerated faces, with rich annotation. For the presented benchmark dataset\nbaselines are provided for such tasks as document location and identification,\ntext fields recognition, and face detection. With 72409 annotated images in\ntotal, to the date of publication the proposed dataset is the largest publicly\navailable identity documents dataset with variable artificially generated data,\nand we believe that it will prove invaluable for advancement of the field of\ndocument analysis and recognition. The dataset is available for download at\nftp://smartengines.com/midv-2020 and http://l3i-share.univ-lr.fr .', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'Traditionally, text generation models take in a sequence of text as input,\nand iteratively generate the next most probable word using pre-trained\nparameters. In this work, we propose the architecture to use images instead of\ntext as the input of the text generation model, called StoryGen. In the\narchitecture, we design a Relational Text Data Generator algorithm that relates\ndifferent features from multiple images. The output samples from the model\ndemonstrate the ability to generate meaningful paragraphs of text containing\nthe extracted features from the input images. This is an undergraduate project\nreport. Completed Dec. 2019 at the Cooper Union.', 'In this paper we introduce a new method for text detection in natural images.\nThe method comprises two contributions: First, a fast and scalable engine to\ngenerate synthetic images of text in clutter. This engine overlays synthetic\ntext to existing background images in a natural way, accounting for the local\n3D scene geometry. Second, we use the synthetic images to train a\nFully-Convolutional Regression Network (FCRN) which efficiently performs text\ndetection and bounding-box regression at all locations and multiple scales in\nan image. We discuss the relation of FCRN to the recently-introduced YOLO\ndetector, as well as other end-to-end object detection systems based on deep\nlearning. The resulting detection network significantly out performs current\nmethods for text detection in natural images, achieving an F-measure of 84.2%\non the standard ICDAR 2013 benchmark. Furthermore, it can process 15 images per\nsecond on a GPU.', 'The continually increasing number of documents produced each year\nnecessitates ever improving information processing methods for searching,\nretrieving, and organizing text. Central to these information processing\nmethods is document classification, which has become an important application\nfor supervised learning. Recently the performance of these traditional\nclassifiers has degraded as the number of documents has increased. This is\nbecause along with this growth in the number of documents has come an increase\nin the number of categories. This paper approaches this problem differently\nfrom current document classification methods that view the problem as\nmulti-class classification. Instead we perform hierarchical classification\nusing an approach we call Hierarchical Deep Learning for Text classification\n(HDLTex). HDLTex employs stacks of deep learning architectures to provide\nspecialized understanding at each level of the document hierarchy.', 'Digital camera and mobile document image acquisition are new trends arising\nin the world of Optical Character Recognition and text detection. In some\ncases, such process integrates many distortions and produces poorly scanned\ntext or text-photo images and natural images, leading to an unreliable OCR\ndigitization. In this paper, we present a novel nonparametric and unsupervised\nmethod to compensate for undesirable document image distortions aiming to\noptimally improve OCR accuracy. Our approach relies on a very efficient stack\nof document image enhancing techniques to recover deformation of the entire\ndocument image. First, we propose a local brightness and contrast adjustment\nmethod to effectively handle lighting variations and the irregular distribution\nof image illumination. Second, we use an optimized greyscale conversion\nalgorithm to transform our document image to greyscale level. Third, we sharpen\nthe useful information in the resulting greyscale image using Un-sharp Masking\nmethod. Finally, an optimal global binarization approach is used to prepare the\nfinal document image to OCR recognition. The proposed approach can\nsignificantly improve text detection rate and optical character recognition\naccuracy. To demonstrate the efficiency of our approach, an exhaustive\nexperimentation on a standard dataset is presented.', 'Computer vision systems for wood identification have the potential to empower\nboth producer and consumer countries to combat illegal logging if they can be\ndeployed effectively in the field. In this work, carried out as part of an\nactive international partnership with the support of UNIDO, we constructed and\ncurated a field-relevant image data set to train a classifier for wood\nidentification of $15$ commercial Ghanaian woods using the XyloTron system. We\ntested model performance in the laboratory, and then collected real-world field\nperformance data across multiple sites using multiple XyloTron devices. We\npresent efficacies of the trained model in the laboratory and in the field,\ndiscuss practical implications and challenges of deploying machine learning\nwood identification models, and conclude that field testing is a necessary step\n- and should be considered the gold-standard - for validating computer vision\nwood identification systems.', 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.', 'This work presents a method for visual text recognition without using any\npaired supervisory data. We formulate the text recognition task as one of\naligning the conditional distribution of strings predicted from given text\nimages, with lexically valid strings sampled from target corpora. This enables\nfully automated, and unsupervised learning from just line-level text-images,\nand unpaired text-string samples, obviating the need for large aligned\ndatasets. We present detailed analysis for various aspects of the proposed\nmethod, namely - (1) impact of the length of training sequences on convergence,\n(2) relation between character frequencies and the order in which they are\nlearnt, (3) generalisation ability of our recognition network to inputs of\narbitrary lengths, and (4) impact of varying the text corpus on recognition\naccuracy. Finally, we demonstrate excellent text recognition accuracy on both\nsynthetically generated text images, and scanned images of real printed books,\nusing no labelled training examples.', 'In this paper we introduce a script identification method based on\nhand-crafted texture features and an artificial neural network. The proposed\npipeline achieves near state-of-the-art performance for script identification\nof video-text and state-of-the-art performance on visual language\nidentification of handwritten text. More than using the deep network as a\nclassifier, the use of its intermediary activations as a learned metric\ndemonstrates remarkable results and allows the use of discriminative models on\nunknown classes. Comparative experiments in video-text and text in the wild\ndatasets provide insights on the internals of the proposed deep network.', 'Recovering the original printed texts from a document with added handwritten\nannotations in the marginal area is one of the challenging problems, especially\nwhen the original document is not available. Therefore, this paper aims at\nsalvaging automatically the original document from the annotated document by\ndetecting and removing any handwritten annotations that appear in the marginal\narea of the document without any loss of information. Here a two stage\nalgorithm is proposed, where in the first stage due to approximate marginal\nboundary detection with horizontal and vertical projection profiles, all of the\nmarginal annotations along with some part of the original printed text that may\nappear very close to the marginal boundary are removed. Therefore as a second\nstage, using the connected components, a strategy is applied to bring back the\nprinted text components cropped during the first stage. The proposed method is\nvalidated using a dataset of 50 documents having complex handwritten\nannotations, which gives an overall accuracy of 89.01% in removing the marginal\nannotations and 97.74% in case of retrieving the original printed text\ndocument.', 'Everyone knows that thousand of words are represented by a single image. As a\nresult image search has become a very popular mechanism for the Web searchers.\nImage search means, the search results are produced by the search engine should\nbe a set of images along with their Web page Unified Resource Locator. Now Web\nsearcher can perform two types of image search, they are Text to Image and\nImage to Image search. In Text to Image search, search query should be a text.\nBased on the input text data system will generate a set of images along with\ntheir Web page URL as an output. On the other hand, in Image to Image search,\nsearch query should be an image and based on this image system will generate a\nset of images along with their Web page URL as an output. According to the\ncurrent scenarios, Text to Image search mechanism always not returns perfect\nresult. It matches the text data and then displays the corresponding images as\nan output, which is not always perfect. To resolve this problem, Web\nresearchers have introduced the Image to Image search mechanism. In this paper,\nwe have also proposed an alternate approach of Image to Image search mechanism\nusing Histogram.', 'Segmentation of a text-document into lines, words and characters, which is\nconsidered to be the crucial pre-processing stage in Optical Character\nRecognition (OCR) is traditionally carried out on uncompressed documents,\nalthough most of the documents in real life are available in compressed form,\nfor the reasons such as transmission and storage efficiency. However, this\nimplies that the compressed image should be decompressed, which indents\nadditional computing resources. This limitation has motivated us to take up\nresearch in document image analysis using compressed documents. In this paper,\nwe think in a new way to carry out segmentation at line, word and character\nlevel in run-length compressed printed-text-documents. We extract the\nhorizontal projection profile curve from the compressed file and using the\nlocal minima points perform line segmentation. However, tracing vertical\ninformation which leads to tracking words-characters in a run-length compressed\nfile is not very straight forward. Therefore, we propose a novel technique for\ncarrying out simultaneous word and character segmentation by popping out column\nruns from each row in an intelligent sequence. The proposed algorithms have\nbeen validated with 1101 text-lines, 1409 words and 7582 characters from a\ndata-set of 35 noise and skew free compressed documents of Bengali, Kannada and\nEnglish Scripts.', 'Recognition of document images have important applications in restoring old\nand classical texts. The problem involves quality improvement before passing it\nto a properly trained OCR to get accurate recognition of the text. The image\nenhancement and quality improvement constitute important steps as subsequent\nrecognition depends upon the quality of the input image. There are scenarios\nwhen high resolution images are not available and our experiments show that the\nOCR accuracy reduces significantly with decrease in the spatial resolution of\ndocument images. Thus the only option is to improve the resolution of such\ndocument images. The goal is to construct a high resolution image, given a\nsingle low resolution binary image, which constitutes the problem of single\nimage super-resolution. Most of the previous work in super-resolution deal with\nnatural images which have more information-content than the document images.\nHere, we use Convolution Neural Network to learn the mapping between low and\nthe corresponding high resolution images. We experiment with different number\nof layers, parameter settings and non-linear functions to build a fast\nend-to-end framework for document image super-resolution. Our proposed model\nshows a very good PSNR improvement of about 4 dB on 75 dpi Tamil images,\nresulting in a 3 % improvement of word level accuracy by the OCR. It takes less\ntime than the recent sparse based natural image super-resolution technique,\nmaking it useful for real-time document recognition applications.', 'Part information has been shown to be resistant to occlusions and viewpoint\nchanges, which is beneficial for various vision-related tasks. However, we\nfound very limited work in car pose estimation and reconstruction from street\nviews leveraging the part information. There are two major contributions in\nthis paper. Firstly, we make the first attempt to build a framework to\nsimultaneously estimate shape, translation, orientation, and semantic parts of\ncars in 3D space from a single street view. As it is labor-intensive to\nannotate semantic parts on real street views, we propose a specific approach to\nimplicitly transfer part features from synthesized images to real street views.\nFor pose and shape estimation, we propose a novel network structure that\nutilizes both part features and 3D losses. Secondly, we are the first to\nconstruct a high-quality dataset that contains 348 different car models with\nphysical dimensions and part-level annotations based on global and local\ndeformations. Given these models, we further generate 60K synthesized images\nwith randomization of orientation, illumination, occlusion, and texture. Our\nresults demonstrate that our part segmentation performance is significantly\nimproved after applying our implicit transfer approach. Our network for pose\nand shape estimation achieves the state-of-the-art performance on the\nApolloCar3D dataset and outperforms 3D-RCNN and DeepMANTA by 12.57 and 8.91\npercentage points in terms of mean A3DP-Abs.', 'Synthesising photo-realistic images from natural language is one of the\nchallenging problems in computer vision. Over the past decade, a number of\napproaches have been proposed, of which the improved Stacked Generative\nAdversarial Network (StackGAN-v2) has proven capable of generating high\nresolution images that reflect the details specified in the input text\ndescriptions. In this paper, we aim to assess the robustness and\nfault-tolerance capability of the StackGAN-v2 model by introducing variations\nin the training data. However, due to the working principle of Generative\nAdversarial Network (GAN), it is difficult to predict the output of the model\nwhen the training data are modified. Hence, in this work, we adopt Metamorphic\nTesting technique to evaluate the robustness of the model with a variety of\nunexpected training dataset. As such, we first implement StackGAN-v2 algorithm\nand test the pre-trained model provided by the original authors to establish a\nground truth for our experiments. We then identify a metamorphic relation, from\nwhich test cases are generated. Further, metamorphic relations were derived\nsuccessively based on the observations of prior test results. Finally, we\nsynthesise the results from our experiment of all the metamorphic relations and\nfound that StackGAN-v2 algorithm is susceptible to input images with obtrusive\nobjects, even if it overlaps with the main object minimally, which was not\nreported by the authors and users of StackGAN-v2 model. The proposed\nmetamorphic relations can be applied to other text-to-image synthesis models to\nnot only verify the robustness but also to help researchers understand and\ninterpret the results made by the machine learning models.', 'One of the major challenges in training text-to-image generation models is\nthe need of a large number of high-quality image-text pairs. While image\nsamples are often easily accessible, the associated text descriptions typically\nrequire careful human captioning, which is particularly time- and\ncost-consuming. In this paper, we propose the first work to train text-to-image\ngeneration models without any text data. Our method leverages the well-aligned\nmulti-modal semantic space of the powerful pre-trained CLIP model: the\nrequirement of text-conditioning is seamlessly alleviated via generating text\nfeatures from image features. Extensive experiments are conducted to illustrate\nthe effectiveness of the proposed method. We obtain state-of-the-art results in\nthe standard text-to-image generation tasks. Importantly, the proposed\nlanguage-free model outperforms most existing models trained with full\nimage-text pairs. Furthermore, our method can be applied in fine-tuning\npre-trained models, which saves both training time and cost in training\ntext-to-image generation models. Our pre-trained model obtains competitive\nresults in zero-shot text-to-image generation on the MS-COCO dataset, yet with\naround only 1% of the model size and training data size relative to the\nrecently proposed large DALL-E model.', 'Computation of document image quality metrics often depends upon the\navailability of a ground truth image corresponding to the document. This limits\nthe applicability of quality metrics in applications such as hyperparameter\noptimization of image processing algorithms that operate on-the-fly on unseen\ndocuments. This work proposes the use of surrogate models to learn the behavior\nof a given document quality metric on existing datasets where ground truth\nimages are available. The trained surrogate model can later be used to predict\nthe metric value on previously unseen document images without requiring access\nto ground truth images. The surrogate model is empirically evaluated on the\nDocument Image Binarization Competition (DIBCO) and the Handwritten Document\nImage Binarization Competition (H-DIBCO) datasets.', 'Robust text reading from street view images provides valuable information for\nvarious applications. Performance improvement of existing methods in such a\nchallenging scenario heavily relies on the amount of fully annotated training\ndata, which is costly and in-efficient to obtain. To scale up the amount of\ntraining data while keeping the labeling procedure cost-effective, this\ncompetition introduces a new challenge on Large-scale Street View Text with\nPartial Labeling (LSVT), providing 50, 000 and 400, 000 images in full and weak\nannotations, respectively. This competition aims to explore the abilities of\nstate-of-the-art methods to detect and recognize text instances from\nlarge-scale street view images, closing the gap between research benchmarks and\nreal applications. During the competition period, a total of 41 teams\nparticipated in the two proposed tasks with 132 valid submissions, i.e., text\ndetection and end-to-end text spotting. This paper includes dataset\ndescriptions, task definitions, evaluation protocols and results summaries of\nthe ICDAR 2019-LSVT challenge.', 'The StarLight program conceptualizes fast interstellar travel via small wafer\nsatellites (wafersats) that are propelled by directed energy. This process is\nwildly different from traditional space travel and trades large and slow\nspacecraft for small, fast, inexpensive, and fragile ones. The main goal of\nthese wafer satellites is to gather useful images during their deep space\njourney. We introduce and solve some of the main problems that accompany this\nconcept. First, we need an object detection system that can detect planets that\nwe have never seen before, some containing features that we may not even know\nexist in the universe. Second, once we have images of exoplanets, we need a way\nto take these images and rank them by importance. Equipment fails and data\nrates are slow, thus we need a method to ensure that the most important images\nto humankind are the ones that are prioritized for data transfer. Finally, the\nenergy on board is minimal and must be conserved and used sparingly. No\nexoplanet images should be missed, but using energy erroneously would be\ndetrimental. We introduce simulator-based methods that leverage artificial\nintelligence, mostly in the form of computer vision, in order to solve all\nthree of these issues. Our results confirm that simulators provide an extremely\nrich training environment that surpasses that of real images, and can be used\nto train models on features that have yet to be observed by humans. We also\nshow that the immersive and adaptable environment provided by the simulator,\ncombined with deep learning, lets us navigate and save energy in an otherwise\nimplausible way.', 'The existing lensless compressive camera\n($\\text{L}^2\\text{C}^2$)~\\cite{Huang13ICIP} suffers from low capture rates,\nresulting in low resolution images when acquired over a short time. In this\nwork, we propose a new regime to mitigate these drawbacks. We replace the\nglobal-based compressive sensing used in the existing $\\text{L}^2\\text{C}^2$ by\nthe local block (patch) based compressive sensing. We use a single sensor for\neach block, rather than for the entire image, thus forming a multiple but\nspatially parallel sensor $\\text{L}^2\\text{C}^2$. This new camera retains the\nadvantages of existing $\\text{L}^2\\text{C}^2$ while leading to the following\nadditional benefits: 1) Since each block can be very small, {\\em e.g.}$~8\\times\n8$ pixels, we only need to capture $\\sim 10$ measurements to achieve reasonable\nreconstruction. Therefore the capture time can be reduced significantly. 2) The\ncoding patterns used in each block can be the same, therefore the sensing\nmatrix is only of the block size compared to the entire image size in existing\n$\\text{L}^2\\text{C}^2$. This saves the memory requirement of the sensing matrix\nas well as speeds up the reconstruction. 3) Patch based image reconstruction is\nfast and since real time stitching algorithms exist, we can perform real time\nreconstruction. 4) These small blocks can be integrated to any desirable\nnumber, leading to ultra high resolution images while retaining fast capture\nrate and fast reconstruction. We develop multiple geometries of this block-wise\n$\\text{L}^2\\text{C}^2$ in this paper. We have built prototypes of the proposed\nblock-wise $\\text{L}^2\\text{C}^2$ and demonstrated excellent results of real\ndata.', 'The goal of cross-view image based geo-localization is to determine the\nlocation of a given street view image by matching it against a collection of\ngeo-tagged satellite images. This task is notoriously challenging due to the\ndrastic viewpoint and appearance differences between the two domains. We show\nthat we can address this discrepancy explicitly by learning to synthesize\nrealistic street views from satellite inputs. Following this observation, we\npropose a novel multi-task architecture in which image synthesis and retrieval\nare considered jointly. The rationale behind this is that we can bias our\nnetwork to learn latent feature representations that are useful for retrieval\nif we utilize them to generate images across the two input domains. To the best\nof our knowledge, ours is the first approach that creates realistic street\nviews from satellite images and localizes the corresponding query street-view\nsimultaneously in an end-to-end manner. In our experiments, we obtain\nstate-of-the-art performance on the CVUSA and CVACT benchmarks. Finally, we\nshow compelling qualitative results for satellite-to-street view synthesis.', 'Deep Convolutional Neural Networks (DCNN) have been proven to be effective\nfor various computer vision problems. In this work, we demonstrate its\neffectiveness on a continuous object orientation estimation task, which\nrequires prediction of 0 to 360 degrees orientation of the objects. We do so by\nproposing and comparing three continuous orientation prediction approaches\ndesigned for the DCNNs. The first two approaches work by representing an\norientation as a point on a unit circle and minimizing either L2 loss or\nangular difference loss. The third method works by first converting the\ncontinuous orientation estimation task into a set of discrete orientation\nestimation tasks and then converting the discrete orientation outputs back to\nthe continuous orientation using a mean-shift algorithm. By evaluating on a\nvehicle orientation estimation task and a pedestrian orientation estimation\ntask, we demonstrate that the discretization-based approach not only works\nbetter than the other two approaches but also achieves state-of-the-art\nperformance. We also demonstrate that finding an appropriate feature\nrepresentation is critical to achieve a good performance when adapting a DCNN\ntrained for an image recognition task.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : My current job involves efficient image retrieval. The primary issue is that images require substantial storage space, and the constant conversion to and from compact code results in unnecessary time expenditure. Therefore, I am seeking a model capable of retrieving images in compact code format. The challenge for such a model is that images in compact code do not always maintain close local information, leading to increased computational costs for local features. Consequently, I am interested in image retrieval algorithms and innovative loss functions that prioritize global features. Additionally, I require algorithms that can add labels or annotations directly to the images’ compact code without necessitating any format conversion of the image. I prefer methods that have demonstrated effective performance on popular image retrieval benchmarks.","[""It has been shown that image descriptors extracted by convolutional neural\nnetworks (CNNs) achieve remarkable results for retrieval problems. In this\npaper, we apply attention mechanism to CNN, which aims at enhancing more\nrelevant features that correspond to important keypoints in the input image.\nThe generated attention-aware features are then aggregated by the previous\nstate-of-the-art generalized mean (GeM) pooling followed by normalization to\nproduce a compact global descriptor, which can be efficiently compared to other\nimage descriptors by the dot product. An extensive comparison of our proposed\napproach with state-of-the-art methods is performed on the new challenging\nROxford5k and RParis6k retrieval benchmarks. Results indicate significant\nimprovement over previous work. In particular, our attention-aware GeM (AGeM)\ndescriptor outperforms state-of-the-art method on ROxford5k under the `Hard'\nevaluation protocal."", 'Several recent works have shown that image descriptors produced by deep\nconvolutional neural networks provide state-of-the-art performance for image\nclassification and retrieval problems. It has also been shown that the\nactivations from the convolutional layers can be interpreted as local features\ndescribing particular image regions. These local features can be aggregated\nusing aggregation approaches developed for local features (e.g. Fisher\nvectors), thus providing new powerful global descriptors.\n  In this paper we investigate possible ways to aggregate local deep features\nto produce compact global descriptors for image retrieval. First, we show that\ndeep features and traditional hand-engineered features have quite different\ndistributions of pairwise similarities, hence existing aggregation methods have\nto be carefully re-evaluated. Such re-evaluation reveals that in contrast to\nshallow features, the simple aggregation method based on sum pooling provides\narguably the best performance for deep convolutional features. This method is\nefficient, has few parameters, and bears little risk of overfitting when e.g.\nlearning the PCA matrix. Overall, the new compact global descriptor improves\nthe state-of-the-art on four common benchmarks considerably.', 'Hashing methods have been recently found very effective in retrieval of\nremote sensing (RS) images due to their computational efficiency and fast\nsearch speed. The traditional hashing methods in RS usually exploit\nhand-crafted features to learn hash functions to obtain binary codes, which can\nbe insufficient to optimally represent the information content of RS images. To\novercome this problem, in this paper we introduce a metric-learning based\nhashing network, which learns: 1) a semantic-based metric space for effective\nfeature representation; and 2) compact binary hash codes for fast archive\nsearch. Our network considers an interplay of multiple loss functions that\nallows to jointly learn a metric based semantic space facilitating similar\nimages to be clustered together in that target space and at the same time\nproducing compact final activations that lose negligible information when\nbinarized. Experiments carried out on two benchmark RS archives point out that\nthe proposed network significantly improves the retrieval performance under the\nsame retrieval time when compared to the state-of-the-art hashing methods in\nRS.', 'The objective of this paper is to design an embedding method that maps local\nfeatures describing an image (e.g. SIFT) to a higher dimensional representation\nuseful for the image retrieval problem. First, motivated by the relationship\nbetween the linear approximation of a nonlinear function in high dimensional\nspace and the stateof-the-art feature representation used in image retrieval,\ni.e., VLAD, we propose a new approach for the approximation. The embedded\nvectors resulted by the function approximation process are then aggregated to\nform a single representation for image retrieval. Second, in order to make the\nproposed embedding method applicable to large scale problem, we further derive\nits fast version in which the embedded vectors can be efficiently computed,\ni.e., in the closed-form. We compare the proposed embedding methods with the\nstate of the art in the context of image search under various settings: when\nthe images are represented by medium length vectors, short vectors, or binary\nvectors. The experimental results show that the proposed embedding methods\noutperform existing the state of the art on the standard public image retrieval\nbenchmarks.', 'Deep hashing methods have been proved to be effective for the large-scale\nmedical image search assisting reference-based diagnosis for clinicians.\nHowever, when the salient region plays a maximal discriminative role in\nophthalmic image, existing deep hashing methods do not fully exploit the\nlearning ability of the deep network to capture the features of salient regions\npointedly. The different grades or classes of ophthalmic images may be share\nsimilar overall performance but have subtle differences that can be\ndifferentiated by mining salient regions. To address this issue, we propose a\nnovel end-to-end network, named Attention-based Saliency Hashing (ASH), for\nlearning compact hash-code to represent ophthalmic images. ASH embeds a\nspatial-attention module to focus more on the representation of salient regions\nand highlights their essential role in differentiating ophthalmic images.\nBenefiting from the spatial-attention module, the information of salient\nregions can be mapped into the hash-code for similarity calculation. In the\ntraining stage, we input the image pairs to share the weights of the network,\nand a pairwise loss is designed to maximize the discriminability of the\nhash-code. In the retrieval stage, ASH obtains the hash-code by inputting an\nimage with an end-to-end manner, then the hash-code is used to similarity\ncalculation to return the most similar images. Extensive experiments on two\ndifferent modalities of ophthalmic image datasets demonstrate that the proposed\nASH can further improve the retrieval performance compared to the\nstate-of-the-art deep hashing methods due to the huge contributions of the\nspatial-attention module.', 'Content-based image retrieval (CBIR) in large medical image archives is a\nchallenging and necessary task. Generally, different feature extraction methods\nare used to assign expressive and invariant features to each image such that\nthe search for similar images comes down to feature classification and/or\nmatching. The present work introduces a new image retrieval method for medical\napplications that employs a convolutional neural network (CNN) with recently\nintroduced Radon barcodes. We combine neural codes for global classification\nwith Radon barcodes for the final retrieval. We also examine image search based\non regions of interest (ROI) matching after image retrieval. The IRMA dataset\nwith more than 14,000 x-rays images is used to evaluate the performance of our\nmethod. Experimental results show that our approach is superior to many\npublished works.', ""In this paper, we introduce an approach to overcome the low accuracy of the\nContent-Based Image Retrieval (CBIR) (when using the global features). To\nincrease the accuracy, we use Harris-Laplace detector to identify the interest\nregions of image. Then, we build the Region-Based Image Retrieval (RBIR). For\nthe efficient image storage and retrieval, we encode images into binary\nsignatures. The binary signature of a image is created from its interest\nregions. Furthermore, this paper also provides an algorithm for image retrieval\non S-tree by comparing the images' signatures on a metric similarly to EMD\n(earth mover's distance). Finally, we evaluate the created models on COREL's\nimages."", 'Content-based image retrieval (CBIR) systems on pixel domain use low-level\nfeatures, such as colour, texture and shape, to retrieve images. In this\ncontext, two types of image representations i.e. local and global image\nfeatures have been studied in the literature. Extracting these features from\npixel images and comparing them with images from the database is very\ntime-consuming. Therefore, in recent years, there has been some effort to\naccomplish image analysis directly in the compressed domain with lesser\ncomputations. Furthermore, most of the images in our daily transactions are\nstored in the JPEG compressed format. Therefore, it would be ideal if we could\nretrieve features directly from the partially decoded or compressed data and\nuse them for retrieval. Here, we propose a unified model for image retrieval\nwhich takes DCT coefficients as input and efficiently extracts global and local\nfeatures directly in the JPEG compressed domain for accurate image retrieval.\nThe experimental findings indicate that our proposed model performed similarly\nto the current DELG model which takes RGB features as an input with reference\nto mean average precision while having a faster training and retrieval speed.', 'It has been shown that the activations invoked by an image within the top\nlayers of a large convolutional neural network provide a high-level descriptor\nof the visual content of the image. In this paper, we investigate the use of\nsuch descriptors (neural codes) within the image retrieval application. In the\nexperiments with several standard retrieval benchmarks, we establish that\nneural codes perform competitively even when the convolutional neural network\nhas been trained for an unrelated classification task (e.g.\\ Image-Net). We\nalso evaluate the improvement in the retrieval performance of neural codes,\nwhen the network is retrained on a dataset of images that are similar to images\nencountered at test time.\n  We further evaluate the performance of the compressed neural codes and show\nthat a simple PCA compression provides very good short codes that give\nstate-of-the-art accuracy on a number of datasets. In general, neural codes\nturn out to be much more resilient to such compression in comparison other\nstate-of-the-art descriptors. Finally, we show that discriminative\ndimensionality reduction trained on a dataset of pairs of matched photographs\nimproves the performance of PCA-compressed neural codes even further. Overall,\nour quantitative experiments demonstrate the promise of neural codes as visual\ndescriptors for image retrieval.', 'Sketch-based image retrieval (SBIR) has undergone an increasing interest in\nthe community of computer vision bringing high impact in real applications. For\ninstance, SBIR brings an increased benefit to eCommerce search engines because\nit allows users to formulate a query just by drawing what they need to buy.\nHowever, current methods showing high precision in retrieval work in a high\ndimensional space, which negatively affects aspects like memory consumption and\ntime processing. Although some authors have also proposed compact\nrepresentations, these drastically degrade the performance in a low dimension.\nTherefore in this work, we present different results of evaluating methods for\nproducing compact embeddings in the context of sketch-based image retrieval.\nOur main interest is in strategies aiming to keep the local structure of the\noriginal space. The recent unsupervised local-topology preserving dimension\nreduction method UMAP fits our requirements and shows outstanding performance,\nimproving even the precision achieved by SOTA methods. We evaluate six methods\nin two different datasets. We use Flickr15K and eCommerce datasets; the latter\nis another contribution of this work. We show that UMAP allows us to have\nfeature vectors of 16 bytes improving precision by more than 35%.', 'Due to an increase in the number of image achieves, Content-Based Image\nRetrieval (CBIR) has gained attention for research community of computer\nvision. The image visual contents are represented in a feature space in the\nform of numerical values that is considered as a feature vector of image.\nImages belonging to different classes may contain the common visuals and shapes\nthat can result in the closeness of computed feature space of two different\nimages belonging to separate classes. Due to this reason, feature extraction\nand image representation is selected with appropriate features as it directly\naffects the performance of image retrieval system. The commonly used visual\nfeatures are image spatial layout, color, texture and shape. Image feature\nspace is combined to achieve the discriminating ability that is not possible to\nachieve when the features are used separately. Due to this reason, in this\npaper, we aim to explore the low-level feature combination that are based on\ncolor and shape features. We selected color moments and color histogram to\nrepresent color while shape is represented by using invariant moments. We\nselected this combination, as these features are reported intuitive, compact\nand robust for image representation. We evaluated the performance of our\nproposed research by using the Corel, Coil and Ground Truth (GT) image\ndatasets. We evaluated the proposed low-level feature fusion by calculating the\nprecision, recall and time required for feature extraction. The precision,\nrecall and feature extraction values obtained from the proposed low-level\nfeature fusion outperforms the existing research of CBIR.', 'Driven by the urgent demand for managing remote sensing big data, large-scale\nremote sensing image retrieval (RSIR) attracts increasing attention in the\nremote sensing field. In general, existing retrieval methods can be regarded as\nvisual-based retrieval approaches which search and return a set of similar\nimages from a database to a given query image. Although retrieval methods have\nachieved great success, there is still a question that needs to be responded\nto: Can we obtain the accurate semantic labels of the returned similar images\nto further help analyzing and processing imagery? Inspired by the above\nquestion, in this paper, we redefine the image retrieval problem as visual and\nsemantic retrieval of images. Specifically, we propose a novel deep hashing\nconvolutional neural network (DHCNN) to simultaneously retrieve the similar\nimages and classify their semantic labels in a unified framework. In more\ndetail, a convolutional neural network (CNN) is used to extract\nhigh-dimensional deep features. Then, a hash layer is perfectly inserted into\nthe network to transfer the deep features into compact hash codes. In addition,\na fully connected layer with a softmax function is performed on hash layer to\ngenerate class distribution. Finally, a loss function is elaborately designed\nto simultaneously consider the label loss of each image and similarity loss of\npairs of images. Experimental results on two remote sensing datasets\ndemonstrate that the proposed method achieves the state-of-art retrieval and\nclassification performance.', 'Extracting informative image features and learning effective approximate\nhashing functions are two crucial steps in image retrieval . Conventional\nmethods often study these two steps separately, e.g., learning hash functions\nfrom a predefined hand-crafted feature space. Meanwhile, the bit lengths of\noutput hashing codes are preset in most previous methods, neglecting the\nsignificance level of different bits and restricting their practical\nflexibility. To address these issues, we propose a supervised learning\nframework to generate compact and bit-scalable hashing codes directly from raw\nimages. We pose hashing learning as a problem of regularized similarity\nlearning. Specifically, we organize the training images into a batch of triplet\nsamples, each sample containing two images with the same label and one with a\ndifferent label. With these triplet samples, we maximize the margin between\nmatched pairs and mismatched pairs in the Hamming space. In addition, a\nregularization term is introduced to enforce the adjacency consistency, i.e.,\nimages of similar appearances should have similar codes. The deep convolutional\nneural network is utilized to train the model in an end-to-end fashion, where\ndiscriminative image features and hash functions are simultaneously optimized.\nFurthermore, each bit of our hashing codes is unequally weighted so that we can\nmanipulate the code lengths by truncating the insignificant bits. Our framework\noutperforms state-of-the-arts on public benchmarks of similar image search and\nalso achieves promising results in the application of person re-identification\nin surveillance. It is also shown that the generated bit-scalable hashing codes\nwell preserve the discriminative powers with shorter code lengths.', 'Hashing has been widely applied to multimodal retrieval on large-scale\nmultimedia data due to its efficiency in computation and storage. In this\narticle, we propose a novel deep semantic multimodal hashing network (DSMHN)\nfor scalable image-text and video-text retrieval. The proposed deep hashing\nframework leverages 2-D convolutional neural networks (CNN) as the backbone\nnetwork to capture the spatial information for image-text retrieval, while the\n3-D CNN as the backbone network to capture the spatial and temporal information\nfor video-text retrieval. In the DSMHN, two sets of modality-specific hash\nfunctions are jointly learned by explicitly preserving both intermodality\nsimilarities and intramodality semantic labels. Specifically, with the\nassumption that the learned hash codes should be optimal for the classification\ntask, two stream networks are jointly trained to learn the hash functions by\nembedding the semantic labels on the resultant hash codes. Moreover, a unified\ndeep multimodal hashing framework is proposed to learn compact and high-quality\nhash codes by exploiting the feature representation learning, intermodality\nsimilarity-preserving learning, semantic label-preserving learning, and hash\nfunction learning with different types of loss functions simultaneously. The\nproposed DSMHN method is a generic and scalable deep hashing framework for both\nimage-text and video-text retrievals, which can be flexibly integrated with\ndifferent types of loss functions. We conduct extensive experiments for both\nsingle modal- and cross-modal-retrieval tasks on four widely used\nmultimodal-retrieval data sets. Experimental results on both image-text- and\nvideo-text-retrieval tasks demonstrate that the DSMHN significantly outperforms\nthe state-of-the-art methods.', 'Most image instance retrieval pipelines are based on comparison of vectors\nknown as global image descriptors between a query image and the database\nimages. Due to their success in large scale image classification,\nrepresentations extracted from Convolutional Neural Networks (CNN) are quickly\ngaining ground on Fisher Vectors (FVs) as state-of-the-art global descriptors\nfor image instance retrieval. While CNN-based descriptors are generally\nremarked for good retrieval performance at lower bitrates, they nevertheless\npresent a number of drawbacks including the lack of robustness to common object\ntransformations such as rotations compared with their interest point based FV\ncounterparts.\n  In this paper, we propose a method for computing invariant global descriptors\nfrom CNNs. Our method implements a recently proposed mathematical theory for\ninvariance in a sensory cortex modeled as a feedforward neural network. The\nresulting global descriptors can be made invariant to multiple arbitrary\ntransformation groups while retaining good discriminativeness.\n  Based on a thorough empirical evaluation using several publicly available\ndatasets, we show that our method is able to significantly and consistently\nimprove retrieval results every time a new type of invariance is incorporated.\nWe also show that our method which has few parameters is not prone to\noverfitting: improvements generalize well across datasets with different\nproperties with regard to invariances. Finally, we show that our descriptors\nare able to compare favourably to other state-of-the-art compact descriptors in\nsimilar bitranges, exceeding the highest retrieval results reported in the\nliterature on some datasets. A dedicated dimensionality reduction step\n--quantization or hashing-- may be able to further improve the competitiveness\nof the descriptors.', 'This paper addresses the problem of large-scale image retrieval. We propose a\ntwo-layer fusion method which takes advantage of global and local cues and\nranks database images from coarse to fine (C2F). Departing from the previous\nmethods fusing multiple image descriptors simultaneously, C2F is featured by a\nlayered procedure composed by filtering and refining. In particular, C2F\nconsists of three components. 1) Distractor filtering. With holistic\nrepresentations, noise images are filtered out from the database, so the number\nof candidate images to be used for comparison with the query can be greatly\nreduced. 2) Adaptive weighting. For a certain query, the similarity of\ncandidate images can be estimated by holistic similarity scores in\ncomplementary to the local ones. 3) Candidate refining. Accurate retrieval is\nconducted via local features, combining the pre-computed adaptive weights.\nExperiments are presented on two benchmarks, \\emph{i.e.,} Holidays and Ukbench\ndatasets. We show that our method outperforms recent fusion methods in terms of\nstorage consumption and computation complexity, and that the accuracy is\ncompetitive to the state-of-the-arts.', 'The use of high-dimensional features has become a normal practice in many\ncomputer vision applications. The large dimension of these features is a\nlimiting factor upon the number of data points which may be effectively stored\nand processed, however. We address this problem by developing a novel approach\nto learning a compact binary encoding, which exploits both pair-wise proximity\nand class-label information on training data set. Exploiting this extra\ninformation allows the development of encodings which, although compact,\noutperform the original high-dimensional features in terms of final\nclassification or retrieval performance. The method is general, in that it is\napplicable to both non-parametric and parametric learning methods. This\ngenerality means that the embedded features are suitable for a wide variety of\ncomputer vision tasks, such as image classification and content-based image\nretrieval. Experimental results demonstrate that the new compact descriptor\nachieves an accuracy comparable to, and in some cases better than, the visual\ndescriptor in the original space despite being significantly more compact.\nMoreover, any convex loss function and convex regularization penalty (e.g., $\n\\ell_p $ norm with $ p \\ge 1 $) can be incorporated into the framework, which\nprovides future flexibility.', 'An image retrieval method based on convolution neural network and dimension\nreduction is proposed in this paper. Convolution neural network is used to\nextract high-level features of images, and to solve the problem that the\nextracted feature dimensions are too high and have strong correlation,\nmultilinear principal component analysis is used to reduce the dimension of\nfeatures. The features after dimension reduction are binary hash coded for fast\nimage retrieval. Experiments show that the method proposed in this paper has\nbetter retrieval effect than the retrieval method based on principal component\nanalysis on the e-commerce image datasets.', 'This paper focuses on network pruning for image retrieval acceleration.\nPrevailing image retrieval works target at the discriminative feature learning,\nwhile little attention is paid to how to accelerate the model inference, which\nshould be taken into consideration in real-world practice. The challenge of\npruning image retrieval models is that the middle-level feature should be\npreserved as much as possible. Such different requirements of the retrieval and\nclassification model make the traditional pruning methods not that suitable for\nour task. To solve the problem, we propose a new Progressive Local Filter\nPruning (PLFP) method for image retrieval acceleration. Specifically, layer by\nlayer, we analyze the local geometric properties of each filter and select the\none that can be replaced by the neighbors. Then we progressively prune the\nfilter by gradually changing the filter weights. In this way, the\nrepresentation ability of the model is preserved. To verify this, we evaluate\nour method on two widely-used image retrieval datasets,i.e., Oxford5k and\nParis6K, and one person re-identification dataset,i.e., Market-1501. The\nproposed method arrives with superior performance to the conventional pruning\nmethods, suggesting the effectiveness of the proposed method for image\nretrieval.', 'One of the key challenges of deep learning based image retrieval remains in\naggregating convolutional activations into one highly representative feature\nvector. Ideally, this descriptor should encode semantic, spatial and low level\ninformation. Even though off-the-shelf pre-trained neural networks can already\nproduce good representations in combination with aggregation methods,\nappropriate fine tuning for the task of image retrieval has shown to\nsignificantly boost retrieval performance. In this paper, we present a simple\nyet effective supervised aggregation method built on top of existing regional\npooling approaches. In addition to the maximum activation of a given region, we\ncalculate regional average activations of extracted feature maps. Subsequently,\nweights for each of the pooled feature vectors are learned to perform a\nweighted aggregation to a single feature vector. Furthermore, we apply our\nnewly proposed NRA loss function for deep metric learning to fine tune the\nbackbone neural network and to learn the aggregation weights. Our method\nachieves state-of-the-art results for the INRIA Holidays data set and\ncompetitive results for the Oxford Buildings and Paris data sets while reducing\nthe training time significantly.', 'Domain adaptive image retrieval includes single-domain retrieval and\ncross-domain retrieval. Most of the existing image retrieval methods only focus\non single-domain retrieval, which assumes that the distributions of retrieval\ndatabases and queries are similar. However, in practical application, the\ndiscrepancies between retrieval databases often taken in ideal\nillumination/pose/background/camera conditions and queries usually obtained in\nuncontrolled conditions are very large. In this paper, considering the\npractical application, we focus on challenging cross-domain retrieval. To\naddress the problem, we propose an effective method named Probability Weighted\nCompact Feature Learning (PWCF), which provides inter-domain correlation\nguidance to promote cross-domain retrieval accuracy and learns a series of\ncompact binary codes to improve the retrieval speed. First, we derive our loss\nfunction through the Maximum A Posteriori Estimation (MAP): Bayesian\nPerspective (BP) induced focal-triplet loss, BP induced quantization loss and\nBP induced classification loss. Second, we propose a common manifold structure\nbetween domains to explore the potential correlation across domains.\nConsidering the original feature representation is biased due to the\ninter-domain discrepancy, the manifold structure is difficult to be\nconstructed. Therefore, we propose a new feature named Histogram Feature of\nNeighbors (HFON) from the sample statistics perspective. Extensive experiments\non various benchmark databases validate that our method outperforms many\nstate-of-the-art image retrieval methods for domain adaptive image retrieval.\nThe source code is available at https://github.com/fuxianghuang1/PWCF', 'Fast and scalable Content-Based Image Retrieval using visual features is\nrequired for document analysis, Medical image analysis, etc. in the present\nage. Convolutional Neural Network (CNN) activations as features achieved their\noutstanding performance in this area. Deep Convolutional representations using\nthe softmax function in the output layer are also ones among visual features.\nHowever, almost all the image retrieval systems hold their index of visual\nfeatures on main memory in order to high responsiveness, limiting their\napplicability for big data applications. In this paper, we propose a fast\ncalculation method of cosine similarity with L2 norm indexed in advance on\nElasticsearch. We evaluate our approach with ImageNet Dataset and VGG-16\npre-trained model. The evaluation results show the effectiveness and efficiency\nof our proposed method.', 'Representing images by compact hash codes is an attractive approach for\nlarge-scale content-based image retrieval. In most state-of-the-art\nhashing-based image retrieval systems, for each image, local descriptors are\nfirst aggregated as a global representation vector. This global vector is then\nsubjected to a hashing function to generate a binary hash code. In previous\nworks, the aggregating and the hashing processes are designed independently.\nHence these frameworks may generate suboptimal hash codes. In this paper, we\nfirst propose a novel unsupervised hashing framework in which feature\naggregating and hashing are designed simultaneously and optimized jointly.\nSpecifically, our joint optimization generates aggregated representations that\ncan be better reconstructed by some binary codes. This leads to more\ndiscriminative binary hash codes and improved retrieval accuracy. In addition,\nthe proposed method is flexible. It can be extended for supervised hashing.\nWhen the data label is available, the framework can be adapted to learn binary\ncodes which minimize the reconstruction loss w.r.t. label vectors. Furthermore,\nwe also propose a fast version of the state-of-the-art hashing method Binary\nAutoencoder to be used in our proposed frameworks. Extensive experiments on\nbenchmark datasets under various settings show that the proposed methods\noutperform state-of-the-art unsupervised and supervised hashing methods.', 'With the rapid growing of remotely sensed imagery data, there is a high\ndemand for effective and efficient image retrieval tools to manage and exploit\nsuch data. In this letter, we present a novel content-based remote sensing\nimage retrieval method based on Triplet deep metric learning convolutional\nneural network (CNN). By constructing a Triplet network with metric learning\nobjective function, we extract the representative features of the images in a\nsemantic space in which images from the same class are close to each other\nwhile those from different classes are far apart. In such a semantic space,\nsimple metric measures such as Euclidean distance can be used directly to\ncompare the similarity of images and effectively retrieve images of the same\nclass. We also investigate a supervised and an unsupervised learning methods\nfor reducing the dimensionality of the learned semantic features. We present\ncomprehensive experimental results on two publicly available remote sensing\nimage retrieval datasets and show that our method significantly outperforms\nstate-of-the-art.', 'Image retrieval is the problem of searching an image database for items that\nare similar to a query image. To address this task, two main types of image\nrepresentations have been studied: global and local image features. In this\nwork, our key contribution is to unify global and local features into a single\ndeep model, enabling accurate retrieval with efficient feature extraction. We\nrefer to the new model as DELG, standing for DEep Local and Global features. We\nleverage lessons from recent feature learning work and propose a model that\ncombines generalized mean pooling for global features and attentive selection\nfor local features. The entire network can be learned end-to-end by carefully\nbalancing the gradient flow between two heads -- requiring only image-level\nlabels. We also introduce an autoencoder-based dimensionality reduction\ntechnique for local features, which is integrated into the model, improving\ntraining efficiency and matching performance. Comprehensive experiments show\nthat our model achieves state-of-the-art image retrieval on the Revisited\nOxford and Paris datasets, and state-of-the-art single-model instance-level\nrecognition on the Google Landmarks dataset v2. Code and models are available\nat https://github.com/tensorflow/models/tree/master/research/delf .', 'Text contained in an image carries high-level semantics that can be exploited\nto achieve richer image understanding. In particular, the mere presence of text\nprovides strong guiding content that should be employed to tackle a diversity\nof computer vision tasks such as image retrieval, fine-grained classification,\nand visual question answering. In this paper, we address the problem of\nfine-grained classification and image retrieval by leveraging textual\ninformation along with visual cues to comprehend the existing intrinsic\nrelation between the two modalities. The novelty of the proposed model consists\nof the usage of a PHOC descriptor to construct a bag of textual words along\nwith a Fisher Vector Encoding that captures the morphology of text. This\napproach provides a stronger multimodal representation for this task and as our\nexperiments demonstrate, it achieves state-of-the-art results on two different\ntasks, fine-grained classification and image retrieval.', 'Similarity-preserving hashing is a commonly used method for nearest neighbour\nsearch in large-scale image retrieval. For image retrieval, deep-networks-based\nhashing methods are appealing since they can simultaneously learn effective\nimage representations and compact hash codes. This paper focuses on\ndeep-networks-based hashing for multi-label images, each of which may contain\nobjects of multiple categories. In most existing hashing methods, each image is\nrepresented by one piece of hash code, which is referred to as semantic\nhashing. This setting may be suboptimal for multi-label image retrieval. To\nsolve this problem, we propose a deep architecture that learns\n\\textbf{instance-aware} image representations for multi-label image data, which\nare organized in multiple groups, with each group containing the features for\none category. The instance-aware representations not only bring advantages to\nsemantic hashing, but also can be used in category-aware hashing, in which an\nimage is represented by multiple pieces of hash codes and each piece of code\ncorresponds to a category. Extensive evaluations conducted on several benchmark\ndatasets demonstrate that, for both semantic hashing and category-aware\nhashing, the proposed method shows substantial improvement over the\nstate-of-the-art supervised and unsupervised hashing methods.', 'We investigate the problem of image retrieval based on visual queries when\nthe latter comprise arbitrary regions-of-interest (ROI) rather than entire\nimages. Our proposal is a compact image descriptor that combines the\nstate-of-the-art in content-based descriptor extraction with a multi-level,\nVoronoi-based spatial partitioning of each dataset image. The proposed\nmulti-level Voronoi-based encoding uses a spatial hierarchical K-means over\ninterest-point locations, and computes a content-based descriptor over each\ncell. In order to reduce the matching complexity with minimal or no sacrifice\nin retrieval performance: (i) we utilize the tree structure of the spatial\nhierarchical K-means to perform a top-to-bottom pruning for local similarity\nmaxima; (ii) we propose a new image similarity score that combines relevant\ninformation from all partition levels into a single measure for similarity;\n(iii) we combine our proposal with a novel and efficient approach for optimal\nbit allocation within quantized descriptor representations. By deriving both a\nVoronoi-based VLAD descriptor (termed as Fast-VVLAD) and a Voronoi-based deep\nconvolutional neural network (CNN) descriptor (termed as Fast-VDCNN), we\ndemonstrate that our Voronoi-based framework is agnostic to the descriptor\nbasis, and can easily be slotted into existing frameworks. Via a range of ROI\nqueries in two standard datasets, it is shown that the Voronoi-based\ndescriptors achieve comparable or higher mean Average Precision against\nconventional grid-based spatial search, while offering more than two-fold\nreduction in complexity. Finally, beyond ROI queries, we show that Voronoi\npartitioning improves the geometric invariance of compact CNN descriptors,\nthereby resulting in competitive performance to the current state-of-the-art on\nwhole image retrieval.', 'Due to its low storage cost and fast query speed, hashing has been widely\nused in large-scale image retrieval tasks. Hash bucket search returns data\npoints within a given Hamming radius to each query, which can enable search at\na constant or sub-linear time cost. However, existing hashing methods cannot\nachieve satisfactory retrieval performance for hash bucket search in complex\nscenarios, since they learn only one hash code for each image. More\nspecifically, by using one hash code to represent one image, existing methods\nmight fail to put similar image pairs to the buckets with a small Hamming\ndistance to the query when the semantic information of images is complex. As a\nresult, a large number of hash buckets need to be visited for retrieving\nsimilar images, based on the learned codes. This will deteriorate the\nefficiency of hash bucket search. In this paper, we propose a novel hashing\nframework, called multiple code hashing (MCH), to improve the performance of\nhash bucket search. The main idea of MCH is to learn multiple hash codes for\neach image, with each code representing a different region of the image.\nFurthermore, we propose a deep reinforcement learning algorithm to learn the\nparameters in MCH. To the best of our knowledge, this is the first work that\nproposes to learn multiple hash codes for each image in image retrieval.\nExperiments demonstrate that MCH can achieve a significant improvement in hash\nbucket search, compared with existing methods that learn only one hash code for\neach image.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Similarity-preserving hashing is a widely-used method for nearest neighbour\nsearch in large-scale image retrieval tasks. For most existing hashing methods,\nan image is first encoded as a vector of hand-engineering visual features,\nfollowed by another separate projection or quantization step that generates\nbinary codes. However, such visual feature vectors may not be optimally\ncompatible with the coding process, thus producing sub-optimal hashing codes.\nIn this paper, we propose a deep architecture for supervised hashing, in which\nimages are mapped into binary codes via carefully designed deep neural\nnetworks. The pipeline of the proposed deep architecture consists of three\nbuilding blocks: 1) a sub-network with a stack of convolution layers to produce\nthe effective intermediate image features; 2) a divide-and-encode module to\ndivide the intermediate image features into multiple branches, each encoded\ninto one hash bit; and 3) a triplet ranking loss designed to characterize that\none image is more similar to the second image than to the third one. Extensive\nevaluations on several benchmark image datasets show that the proposed\nsimultaneous feature learning and hash coding pipeline brings substantial\nimprovements over other state-of-the-art supervised or unsupervised hashing\nmethods.', 'With the rapid growth of web images, hashing has received increasing\ninterests in large scale image retrieval. Research efforts have been devoted to\nlearning compact binary codes that preserve semantic similarity based on\nlabels. However, most of these hashing methods are designed to handle simple\nbinary similarity. The complex multilevel semantic structure of images\nassociated with multiple labels have not yet been well explored. Here we\npropose a deep semantic ranking based method for learning hash functions that\npreserve multilevel semantic similarity between multi-label images. In our\napproach, deep convolutional neural network is incorporated into hash functions\nto jointly learn feature representations and mappings from them to hash codes,\nwhich avoids the limitation of semantic representation power of hand-crafted\nfeatures. Meanwhile, a ranking list that encodes the multilevel similarity\ninformation is employed to guide the learning of such deep hash functions. An\neffective scheme based on surrogate loss is used to solve the intractable\noptimization problem of nonsmooth and multivariate ranking measures involved in\nthe learning procedure. Experimental results show the superiority of our\nproposed approach over several state-of-the-art hashing methods in term of\nranking evaluation metrics when tested on multi-label image datasets.', 'In most state-of-the-art hashing-based visual search systems, local image\ndescriptors of an image are first aggregated as a single feature vector. This\nfeature vector is then subjected to a hashing function that produces a binary\nhash code. In previous work, the aggregating and the hashing processes are\ndesigned independently. In this paper, we propose a novel framework where\nfeature aggregating and hashing are designed simultaneously and optimized\njointly. Specifically, our joint optimization produces aggregated\nrepresentations that can be better reconstructed by some binary codes. This\nleads to more discriminative binary hash codes and improved retrieval accuracy.\nIn addition, we also propose a fast version of the recently-proposed Binary\nAutoencoder to be used in our proposed framework. We perform extensive\nretrieval experiments on several benchmark datasets with both SIFT and\nconvolutional features. Our results suggest that the proposed framework\nachieves significant improvements over the state of the art.', 'Recently, image representation built upon Convolutional Neural Network (CNN)\nhas been shown to provide effective descriptors for image search, outperforming\npre-CNN features as short-vector representations. Yet such models are not\ncompatible with geometry-aware re-ranking methods and still outperformed, on\nsome particular object retrieval benchmarks, by traditional image search\nsystems relying on precise descriptor matching, geometric re-ranking, or query\nexpansion. This work revisits both retrieval stages, namely initial search and\nre-ranking, by employing the same primitive information derived from the CNN.\nWe build compact feature vectors that encode several image regions without the\nneed to feed multiple inputs to the network. Furthermore, we extend integral\nimages to handle max-pooling on convolutional layer activations, allowing us to\nefficiently localize matching objects. The resulting bounding box is finally\nused for image re-ranking. As a result, this paper significantly improves\nexisting CNN-based recognition pipeline: We report for the first time results\ncompeting with traditional methods on the challenging Oxford5k and Paris6k\ndatasets.', ""The convolutional neural network (CNN) features can give a good description\nof image content, which usually represent images with unique global vectors.\nAlthough they are compact compared to local descriptors, they still cannot\nefficiently deal with large-scale image retrieval due to the cost of the linear\nincremental computation and storage. To address this issue, we build a simple\nbut effective indexing framework based on inverted table, which significantly\ndecreases both the search time and memory usage. In addition, several\nstrategies are fully investigated under an indexing framework to adapt it to\nCNN features and compensate for quantization errors. First, we use multiple\nassignment for the query and database images to increase the probability of\nrelevant images' co-existing in the same Voronoi cells obtained via the\nclustering algorithm. Then, we introduce embedding codes to further improve\nprecision by removing false matches during a search. We demonstrate that by\nusing hashing schemes to calculate the embedding codes and by changing the\nranking rule, indexing framework speeds can be greatly improved. Extensive\nexperiments conducted on several unsupervised and supervised benchmarks support\nthese results and the superiority of the proposed indexing framework. We also\nprovide a fair comparison between the popular CNN features."", 'Fine-grained image recognition is very challenging due to the difficulty of\ncapturing both semantic global features and discriminative local features.\nMeanwhile, these two features are not easy to be integrated, which are even\nconflicting when used simultaneously. In this paper, a retrieval-based\ncoarse-to-fine framework is proposed, where we re-rank the TopN classification\nresults by using the local region enhanced embedding features to improve the\nTop1 accuracy (based on the observation that the correct category usually\nresides in TopN results). To obtain the discriminative regions for\ndistinguishing the fine-grained images, we introduce a weakly-supervised method\nto train a box generating branch with only image-level labels. In addition, to\nlearn more effective semantic global features, we design a multi-level loss\nover an automatically constructed hierarchical category structure. Experimental\nresults show that our method achieves state-of-the-art performance on three\nbenchmarks: CUB-200-2011, Stanford Cars, and FGVC Aircraft. Also,\nvisualizations and analysis are provided for better understanding.', 'Medical Image Retrieval is a challenging field in Visual information\nretrieval, due to the multi-dimensional and multi-modal context of the\nunderlying content. Traditional models often fail to take the intrinsic\ncharacteristics of data into consideration, and have thus achieved limited\naccuracy when applied to medical images. The Bag of Visual Words (BoVW) is a\ntechnique that can be used to effectively represent intrinsic image features in\nvector space, so that applications like image classification and similar-image\nsearch can be optimized. In this paper, we present a MedIR approach based on\nthe BoVW model for content-based medical image retrieval. As medical images as\nmulti-dimensional, they exhibit underlying cluster and manifold information\nwhich enhances semantic relevance and allows for label uniformity. Hence, the\nBoVW features extracted for each image are used to train a supervised machine\nlearning classifier based on positive and negative training images, for\nextending content based image retrieval. During experimental validation, the\nproposed model performed very well, achieving a Mean Average Precision of\n88.89% during top-3 image retrieval experiments.', 'Remote Sensing Image Retrieval remains a challenging topic due to the special\nnature of Remote Sensing Imagery. Such images contain various different\nsemantic objects, which clearly complicates the retrieval task. In this paper,\nwe present an image retrieval pipeline that uses attentive, local convolutional\nfeatures and aggregates them using the Vector of Locally Aggregated Descriptors\n(VLAD) to produce a global descriptor. We study various system parameters such\nas the multiplicative and additive attention mechanisms and descriptor\ndimensionality. We propose a query expansion method that requires no external\ninputs. Experiments demonstrate that even without training, the local\nconvolutional features and global representation outperform other systems.\nAfter system tuning, we can achieve state-of-the-art or competitive results.\nFurthermore, we observe that our query expansion method increases overall\nsystem performance by about 3%, using only the top-three retrieved images.\nFinally, we show how dimensionality reduction produces compact descriptors with\nincreased retrieval performance and fast retrieval computation times, e.g. 50%\nfaster than the current systems.', 'Visual localization to compute 6DoF camera pose from a given image has wide\napplications such as in robotics, virtual reality, augmented reality, etc. Two\nkinds of descriptors are important for the visual localization. One is global\ndescriptors that extract the whole feature from each image. The other is local\ndescriptors that extract the local feature from each image patch usually\nenclosing a key point. More and more methods of the visual localization have\ntwo stages: at first to perform image retrieval by global descriptors and then\nfrom the retrieval feedback to make 2D-3D point correspondences by local\ndescriptors. The two stages are in serial for most of the methods. This simple\ncombination has not achieved superiority of fusing local and global\ndescriptors. The 3D points obtained from the retrieval feedback are as the\nnearest neighbor candidates of the 2D image points only by global descriptors.\nEach of the 2D image points is also called a query local feature when\nperforming the 2D-3D point correspondences. In this paper, we propose a novel\nparallel search framework, which leverages advantages of both local and global\ndescriptors to get nearest neighbor candidates of a query local feature.\nSpecifically, besides using deep learning based global descriptors, we also\nutilize local descriptors to construct random tree structures for obtaining\nnearest neighbor candidates of the query local feature. We propose a new\nprobabilistic model and a new deep learning based local descriptor when\nconstructing the random trees. A weighted Hamming regularization term to keep\ndiscriminativeness after binarization is given in the loss function for the\nproposed local descriptor. The loss function co-trains both real and binary\ndescriptors of which the results are integrated into the random trees.', 'We propose a novel approach for instance-level image retrieval. It produces a\nglobal and compact fixed-length representation for each image by aggregating\nmany region-wise descriptors. In contrast to previous works employing\npre-trained deep networks as a black box to produce features, our method\nleverages a deep architecture trained for the specific task of image retrieval.\nOur contribution is twofold: (i) we leverage a ranking framework to learn\nconvolution and projection weights that are used to build the region features;\nand (ii) we employ a region proposal network to learn which regions should be\npooled to form the final global descriptor. We show that using clean training\ndata is key to the success of our approach. To that aim, we use a large scale\nbut noisy landmark dataset and develop an automatic cleaning approach. The\nproposed architecture produces a global image representation in a single\nforward pass. Our approach significantly outperforms previous approaches based\non global descriptors on standard datasets. It even surpasses most prior works\nbased on costly local descriptor indexing and spatial verification. Additional\nmaterial is available at www.xrce.xerox.com/Deep-Image-Retrieval.', 'Query expansion is a popular method to improve the quality of image retrieval\nwith both conventional and CNN representations. It has been so far limited to\nglobal image similarity. This work focuses on diffusion, a mechanism that\ncaptures the image manifold in the feature space. The diffusion is carried out\non descriptors of overlapping image regions rather than on a global image\ndescriptor like in previous approaches. An efficient off-line stage allows\noptional reduction in the number of stored regions. In the on-line stage, the\nproposed handling of unseen queries in the indexing stage removes additional\ncomputation to adjust the precomputed data. We perform diffusion through a\nsparse linear system solver, yielding practical query times well below one\nsecond. Experimentally, we observe a significant boost in performance of image\nretrieval with compact CNN descriptors on standard benchmarks, especially when\nthe query object covers only a small part of the image. Small objects have been\na common failure case of CNN-based retrieval.', 'In the large-scale image retrieval task, the two most important requirements\nare the discriminability of image representations and the efficiency in\ncomputation and storage of representations. Regarding the former requirement,\nConvolutional Neural Network (CNN) is proven to be a very powerful tool to\nextract highly discriminative local descriptors for effective image search.\nAdditionally, in order to further improve the discriminative power of the\ndescriptors, recent works adopt fine-tuned strategies. In this paper, taking a\ndifferent approach, we propose a novel, computationally efficient, and\ncompetitive framework. Specifically, we firstly propose various strategies to\ncompute masks, namely SIFT-mask, SUM-mask, and MAX-mask, to select a\nrepresentative subset of local convolutional features and eliminate redundant\nfeatures. Our in-depth analyses demonstrate that proposed masking schemes are\neffective to address the burstiness drawback and improve retrieval accuracy.\nSecondly, we propose to employ recent embedding and aggregating methods which\ncan significantly boost the feature discriminability. Regarding the computation\nand storage efficiency, we include a hashing module to produce very compact\nbinary image representations. Extensive experiments on six image retrieval\nbenchmarks demonstrate that our proposed framework achieves the\nstate-of-the-art retrieval performances.', 'Image Retrieval is a fundamental task of obtaining images similar to the\nquery one from a database. A common image retrieval practice is to firstly\nretrieve candidate images via similarity search using global image features and\nthen re-rank the candidates by leveraging their local features. Previous\nlearning-based studies mainly focus on either global or local image\nrepresentation learning to tackle the retrieval task. In this paper, we abandon\nthe two-stage paradigm and seek to design an effective single-stage solution by\nintegrating local and global information inside images into compact image\nrepresentations. Specifically, we propose a Deep Orthogonal Local and Global\n(DOLG) information fusion framework for end-to-end image retrieval. It\nattentively extracts representative local information with multi-atrous\nconvolutions and self-attention at first. Components orthogonal to the global\nimage representation are then extracted from the local information. At last,\nthe orthogonal components are concatenated with the global representation as a\ncomplementary, and then aggregation is performed to generate the final\nrepresentation. The whole framework is end-to-end differentiable and can be\ntrained with image-level labels. Extensive experimental results validate the\neffectiveness of our solution and show that our model achieves state-of-the-art\nimage retrieval performances on Revisited Oxford and Paris datasets.', 'Deep hashing has been widely applied to large-scale image retrieval by\nencoding high-dimensional data points into binary codes for efficient\nretrieval. Compared with pairwise/triplet similarity based hash learning,\ncentral similarity based hashing can more efficiently capture the global data\ndistribution. For multi-label image retrieval, however, previous methods only\nuse multiple hash centers with equal weights to generate one centroid as the\nlearning target, which ignores the relationship between the weights of hash\ncenters and the proportion of instance regions in the image. To address the\nabove issue, we propose a two-step alternative optimization approach,\nInstance-weighted Central Similarity (ICS), to automatically learn the center\nweight corresponding to a hash code. Firstly, we apply the maximum entropy\nregularizer to prevent one hash center from dominating the loss function, and\ncompute the center weights via projection gradient descent. Secondly, we update\nneural network parameters by standard back-propagation with fixed center\nweights. More importantly, the learned center weights can well reflect the\nproportion of foreground instances in the image. Our method achieves the\nstate-of-the-art performance on the image retrieval benchmarks, and especially\nimproves the mAP by 1.6%-6.4% on the MS COCO dataset.', 'In applications involving matching of image sets, the information from\nmultiple images must be effectively exploited to represent each set.\nState-of-the-art methods use probabilistic distribution or subspace to model a\nset and use specific distance measure to compare two sets. These methods are\nslow to compute and not compact to use in a large scale scenario.\nLearning-based hashing is often used in large scale image retrieval as they\nprovide a compact representation of each sample and the Hamming distance can be\nused to efficiently compare two samples. However, most hashing methods encode\neach image separately and discard knowledge that multiple images in the same\nset represent the same object or person. We investigate the set hashing problem\nby combining both set representation and hashing in a single deep neural\nnetwork. An image set is first passed to a CNN module to extract image\nfeatures, then these features are aggregated using two types of set feature to\ncapture both set specific and database-wide distribution information. The\ncomputed set feature is then fed into a multilayer perceptron to learn a\ncompact binary embedding. Triplet loss is used to train the network by forming\nset similarity relations using class labels. We extensively evaluate our\napproach on datasets used for image matching and show highly competitive\nperformance compared to state-of-the-art methods.', 'With the explosive growth of image databases, deep hashing, which learns\ncompact binary descriptors for images, has become critical for fast image\nretrieval. Many existing deep hashing methods leverage quantization loss,\ndefined as distance between the features before and after quantization, to\nreduce the error from binarizing features. While minimizing the quantization\nloss guarantees that quantization has minimal effect on retrieval accuracy, it\nunfortunately significantly reduces the expressiveness of features even before\nthe quantization. In this paper, we show that the above definition of\nquantization loss is too restricted and in fact not necessary for maintaining\nhigh retrieval accuracy. We therefore propose a new form of quantization loss\nmeasured in triplets. The core idea of the triplet quantization loss is to\nlearn discriminative real-valued descriptors which lead to minimal loss on\nretrieval accuracy after quantization. Extensive experiments on two widely used\nbenchmark data sets of different scales, CIFAR-10 and In-shop, demonstrate that\nthe proposed method outperforms the state-of-the-art deep hashing methods.\nMoreover, we show that the compact binary descriptors obtained with triplet\nquantization loss lead to very small performance drop after quantization.', 'Vision-based localization of an agent in a map is an important problem in\nrobotics and computer vision. In that context, localization by learning\nmatchable image features is gaining popularity due to recent advances in\nmachine learning. Features that uniquely describe the visual contents of images\nhave a wide range of applications, including image retrieval and understanding.\nIn this work, we propose a method that learns image features targeted for\nimage-retrieval-based localization. Retrieval-based localization has several\nbenefits, such as easy maintenance and quick computation. However, the\nstate-of-the-art features only provide visual similarity scores which do not\nexplicitly reveal the geometric distance between query and retrieved images.\nKnowing this distance is highly desirable for accurate localization, especially\nwhen the reference images are sparsely distributed in the scene. Therefore, we\npropose a novel loss function for learning image features which are both\nvisually representative and geometrically relatable. This is achieved by\nguiding the learning process such that the feature and geometric distances\nbetween images are directly proportional. In our experiments we show that our\nfeatures not only offer significantly better localization accuracy, but also\nallow to estimate the trajectory of a query sequence in absence of the\nreference images.', 'While deep learning has become a key ingredient in the top performing methods\nfor many computer vision tasks, it has failed so far to bring similar\nimprovements to instance-level image retrieval. In this article, we argue that\nreasons for the underwhelming results of deep methods on image retrieval are\nthreefold: i) noisy training data, ii) inappropriate deep architecture, and\niii) suboptimal training procedure. We address all three issues.\n  First, we leverage a large-scale but noisy landmark dataset and develop an\nautomatic cleaning method that produces a suitable training set for deep\nretrieval. Second, we build on the recent R-MAC descriptor, show that it can be\ninterpreted as a deep and differentiable architecture, and present improvements\nto enhance it. Last, we train this network with a siamese architecture that\ncombines three streams with a triplet loss. At the end of the training process,\nthe proposed architecture produces a global image representation in a single\nforward pass that is well suited for image retrieval. Extensive experiments\nshow that our approach significantly outperforms previous retrieval approaches,\nincluding state-of-the-art methods based on costly local descriptor indexing\nand spatial verification. On Oxford 5k, Paris 6k and Holidays, we respectively\nreport 94.7, 96.6, and 94.8 mean average precision. Our representations can\nalso be heavily compressed using product quantization with little loss in\naccuracy. For additional material, please see\nwww.xrce.xerox.com/Deep-Image-Retrieval.', 'Image descriptors based on activations of Convolutional Neural Networks\n(CNNs) have become dominant in image retrieval due to their discriminative\npower, compactness of representation, and search efficiency. Training of CNNs,\neither from scratch or fine-tuning, requires a large amount of annotated data,\nwhere a high quality of annotation is often crucial. In this work, we propose\nto fine-tune CNNs for image retrieval on a large collection of unordered images\nin a fully automated manner. Reconstructed 3D models obtained by the\nstate-of-the-art retrieval and structure-from-motion methods guide the\nselection of the training data. We show that both hard-positive and\nhard-negative examples, selected by exploiting the geometry and the camera\npositions available from the 3D models, enhance the performance of\nparticular-object retrieval. CNN descriptor whitening discriminatively learned\nfrom the same training data outperforms commonly used PCA whitening. We propose\na novel trainable Generalized-Mean (GeM) pooling layer that generalizes max and\naverage pooling and show that it boosts retrieval performance. Applying the\nproposed method to the VGG network achieves state-of-the-art performance on the\nstandard benchmarks: Oxford Buildings, Paris, and Holidays datasets.', 'The cost-effective visual representation and fast query-by-example search are\ntwo challenging goals that should be maintained for web-scale visual retrieval\ntasks on moderate hardware. This paper introduces a fast and robust method that\nensures both of these goals by obtaining state-of-the-art performance for an\nimage-to-video search scenario. Hence, we present critical enhancements to\nwell-known indexing and visual representation techniques by promoting faster,\nbetter and moderate retrieval performance. We also boost the superiority of our\nmethod for some visual challenges by exploiting individual decisions of local\nand global descriptors at query time. For instance, local content descriptors\nrepresent copied/duplicated scenes with large geometric deformations such as\nscale, orientation and affine transformation. In contrast, the use of global\ncontent descriptors is more practical for near-duplicate and semantic searches.\nExperiments are conducted on a large-scale Stanford I2V dataset. The\nexperimental results show that our method is useful in terms of complexity and\nquery processing time for large-scale visual retrieval scenarios, even if local\nand global representations are used together. The proposed method is superior\nand achieves state-of-the-art performance based on the mean average precision\n(MAP) score of this dataset. Lastly, we report additional MAP scores after\nupdating the ground annotations unveiled by retrieval results of the proposed\nmethod, and it shows that the actual performance.', 'Image retrieval in realistic scenarios targets large dynamic datasets of\nunlabeled images. In these cases, training or fine-tuning a model every time\nnew images are added to the database is neither efficient nor scalable.\nConvolutional neural networks trained for image classification over large\ndatasets have been proven effective feature extractors for image retrieval. The\nmost successful approaches are based on encoding the activations of\nconvolutional layers, as they convey the image spatial information. In this\npaper, we go beyond this spatial information and propose a local-aware encoding\nof convolutional features based on semantic information predicted in the target\nimage. To this end, we obtain the most discriminative regions of an image using\nClass Activation Maps (CAMs). CAMs are based on the knowledge contained in the\nnetwork and therefore, our approach, has the additional advantage of not\nrequiring external information. In addition, we use CAMs to generate object\nproposals during an unsupervised re-ranking stage after a first fast search.\nOur experiments on two public available datasets for instance retrieval,\nOxford5k and Paris6k, demonstrate the competitiveness of our approach\noutperforming the current state-of-the-art when using off-the-shelf models\ntrained on ImageNet. The source code and model used in this paper are publicly\navailable at http://imatge-upc.github.io/retrieval-2017-cam/.', 'Visual analytics have played an increasingly critical role in the Internet of\nThings, where massive visual signals have to be compressed and fed into\nmachines. But facing such big data and constrained bandwidth capacity, existing\nimage/video compression methods lead to very low-quality representations, while\nexisting feature compression techniques fail to support diversified visual\nanalytics applications/tasks with low-bit-rate representations. In this paper,\nwe raise and study the novel problem of supporting multiple machine vision\nanalytics tasks with the compressed visual representation, namely, the\ninformation compression problem in analytics taxonomy. By utilizing the\nintrinsic transferability among different tasks, our framework successfully\nconstructs compact and expressive representations at low bit-rates to support a\ndiversified set of machine vision tasks, including both high-level\nsemantic-related tasks and mid-level geometry analytic tasks. In order to\nimpose compactness in the representations, we propose a codebook-based\nhyperprior, which helps map the representation into a low-dimensional manifold.\nAs it well fits the signal structure of the deep visual feature, it facilitates\nmore accurate entropy estimation, and results in higher compression efficiency.\nWith the proposed framework and the codebook-based hyperprior, we further\ninvestigate the relationship of different task features owning different levels\nof abstraction granularity. Experimental results demonstrate that with the\nproposed scheme, a set of diversified tasks can be supported at a significantly\nlower bit-rate, compared with existing compression schemes.']","['Image instance retrieval is the problem of retrieving images from a database\nwhich contain the same object. Convolutional Neural Network (CNN) based\ndescriptors are becoming the dominant approach for generating {\\it global image\ndescriptors} for the instance retrieval problem. One major drawback of\nCNN-based {\\it global descriptors} is that uncompressed deep neural network\nmodels require hundreds of megabytes of storage making them inconvenient to\ndeploy in mobile applications or in custom hardware. In this work, we study the\nproblem of neural network model compression focusing on the image instance\nretrieval task. We study quantization, coding, pruning and weight sharing\ntechniques for reducing model size for the instance retrieval problem. We\nprovide extensive experimental results on the trade-off between retrieval\nperformance and model size for different types of networks on several data sets\nproviding the most comprehensive study on this topic. We compress models to the\norder of a few MBs: two orders of magnitude smaller than the uncompressed\nmodels while achieving negligible loss in retrieval performance.', 'Convolutional Neural Networks (CNNs) have achieved superior performance on\nobject image retrieval, while Bag-of-Words (BoW) models with handcrafted local\nfeatures still dominate the retrieval of overlapping images in 3D\nreconstruction. In this paper, we narrow down this gap by presenting an\nefficient CNN-based method to retrieve images with overlaps, which we refer to\nas the matchable image retrieval problem. Different from previous methods that\ngenerates training data based on sparse reconstruction, we create a large-scale\nimage database with rich 3D geometrics and exploit information from surface\nreconstruction to obtain fine-grained training data. We propose a batched\ntriplet-based loss function combined with mesh re-projection to effectively\nlearn the CNN representation. The proposed method significantly accelerates the\nimage retrieval process in 3D reconstruction and outperforms the\nstate-of-the-art CNN-based and BoW methods for matchable image retrieval. The\ncode and data are available at https://github.com/hlzz/mirror.', 'Recent works in deep-learning have shown that second-order information is\nbeneficial in many computer-vision tasks. Second-order information can be\nenforced both in the spatial context and the abstract feature dimensions. In\nthis work, we explore two second-order components. One is focused on\nsecond-order spatial information to increase the performance of image\ndescriptors, both local and global. It is used to re-weight feature maps, and\nthus emphasise salient image locations that are subsequently used for\ndescription. The second component is concerned with a second-order similarity\n(SOS) loss, that we extend to global descriptors for image retrieval, and is\nused to enhance the triplet loss with hard-negative mining. We validate our\napproach on two different tasks and datasets for image retrieval and image\nmatching. The results show that our two second-order components complement each\nother, bringing significant performance improvements in both tasks and lead to\nstate-of-the-art results across the public benchmarks. Code available at:\nhttp://github.com/tonyngjichun/SOLAR', 'The content based image retrieval aims to find the similar images from a\nlarge scale dataset against a query image. Generally, the similarity between\nthe representative features of the query image and dataset images is used to\nrank the images for retrieval. In early days, various hand designed feature\ndescriptors have been investigated based on the visual cues such as color,\ntexture, shape, etc. that represent the images. However, the deep learning has\nemerged as a dominating alternative of hand-designed feature engineering from a\ndecade. It learns the features automatically from the data. This paper presents\na comprehensive survey of deep learning based developments in the past decade\nfor content based image retrieval. The categorization of existing\nstate-of-the-art methods from different perspectives is also performed for\ngreater understanding of the progress. The taxonomy used in this survey covers\ndifferent supervision, different networks, different descriptor type and\ndifferent retrieval type. A performance analysis is also performed using the\nstate-of-the-art methods. The insights are also presented for the benefit of\nthe researchers to observe the progress and to make the best choices. The\nsurvey presented in this paper will help in further research progress in image\nretrieval using deep learning.', 'Convolutional Neural Networks (CNNs) achieve state-of-the-art performance in\nmany computer vision tasks. However, this achievement is preceded by extreme\nmanual annotation in order to perform either training from scratch or\nfine-tuning for the target task. In this work, we propose to fine-tune CNN for\nimage retrieval from a large collection of unordered images in a fully\nautomated manner. We employ state-of-the-art retrieval and\nStructure-from-Motion (SfM) methods to obtain 3D models, which are used to\nguide the selection of the training data for CNN fine-tuning. We show that both\nhard positive and hard negative examples enhance the final performance in\nparticular object retrieval with compact codes.', 'Image representations derived from pre-trained Convolutional Neural Networks\n(CNNs) have become the new state of the art in computer vision tasks such as\ninstance retrieval. This work explores the suitability for instance retrieval\nof image- and region-wise representations pooled from an object detection CNN\nsuch as Faster R-CNN. We take advantage of the object proposals learned by a\nRegion Proposal Network (RPN) and their associated CNN features to build an\ninstance search pipeline composed of a first filtering stage followed by a\nspatial reranking. We further investigate the suitability of Faster R-CNN\nfeatures when the network is fine-tuned for the same objects one wants to\nretrieve. We assess the performance of our proposed system with the Oxford\nBuildings 5k, Paris Buildings 6k and a subset of TRECVid Instance Search 2013,\nachieving competitive results.', 'This paper presents a simple yet effective supervised deep hash approach that\nconstructs binary hash codes from labeled data for large-scale image search. We\nassume that the semantic labels are governed by several latent attributes with\neach attribute on or off, and classification relies on these attributes. Based\non this assumption, our approach, dubbed supervised semantics-preserving deep\nhashing (SSDH), constructs hash functions as a latent layer in a deep network\nand the binary codes are learned by minimizing an objective function defined\nover classification error and other desirable hash codes properties. With this\ndesign, SSDH has a nice characteristic that classification and retrieval are\nunified in a single learning model. Moreover, SSDH performs joint learning of\nimage representations, hash codes, and classification in a point-wised manner,\nand thus is scalable to large-scale datasets. SSDH is simple and can be\nrealized by a slight enhancement of an existing deep architecture for\nclassification; yet it is effective and outperforms other hashing approaches on\nseveral benchmarks and large datasets. Compared with state-of-the-art\napproaches, SSDH achieves higher retrieval accuracy, while the classification\nperformance is not sacrificed.', 'Image retrieval can be formulated as a ranking problem where the goal is to\norder database images by decreasing similarity to the query. Recent deep models\nfor image retrieval have outperformed traditional methods by leveraging\nranking-tailored loss functions, but important theoretical and practical\nproblems remain. First, rather than directly optimizing the global ranking,\nthey minimize an upper-bound on the essential loss, which does not necessarily\nresult in an optimal mean average precision (mAP). Second, these methods\nrequire significant engineering efforts to work well, e.g. special pre-training\nand hard-negative mining. In this paper we propose instead to directly optimize\nthe global mAP by leveraging recent advances in listwise loss formulations.\nUsing a histogram binning approximation, the AP can be differentiated and thus\nemployed to end-to-end learning. Compared to existing losses, the proposed\nmethod considers thousands of images simultaneously at each iteration and\neliminates the need for ad hoc tricks. It also establishes a new state of the\nart on many standard retrieval benchmarks. Models and evaluation scripts have\nbeen made available at https://europe.naverlabs.com/Deep-Image-Retrieval/', 'Image retrieval task consists of finding similar images to a query image from\na set of gallery (database) images. Such systems are used in various\napplications e.g. person re-identification (ReID) or visual product search.\nDespite active development of retrieval models it still remains a challenging\ntask mainly due to large intra-class variance caused by changes in view angle,\nlighting, background clutter or occlusion, while inter-class variance may be\nrelatively low. A large portion of current research focuses on creating more\nrobust features and modifying objective functions, usually based on Triplet\nLoss. Some works experiment with using centroid/proxy representation of a class\nto alleviate problems with computing speed and hard samples mining used with\nTriplet Loss. However, these approaches are used for training alone and\ndiscarded during the retrieval stage. In this paper we propose to use the mean\ncentroid representation both during training and retrieval. Such an aggregated\nrepresentation is more robust to outliers and assures more stable features. As\neach class is represented by a single embedding - the class centroid - both\nretrieval time and storage requirements are reduced significantly. Aggregating\nmultiple embeddings results in a significant reduction of the search space due\nto lowering the number of candidate target vectors, which makes the method\nespecially suitable for production deployments. Comprehensive experiments\nconducted on two ReID and Fashion Retrieval datasets demonstrate effectiveness\nof our method, which outperforms the current state-of-the-art. We propose\ncentroid training and retrieval as a viable method for both Fashion Retrieval\nand ReID applications.', 'Content-based image retrieval (CBIR) of medical images is a crucial task that\ncan contribute to a more reliable diagnosis if applied to big data. Recent\nadvances in feature extraction and classification have enormously improved CBIR\nresults for digital images. However, considering the increasing accessibility\nof big data in medical imaging, we are still in need of reducing both memory\nrequirements and computational expenses of image retrieval systems. This work\nproposes to exclude the features of image blocks that exhibit a low encoding\nerror when learned by a $n/p/n$ autoencoder ($p\\!<\\!n$). We examine the\nhistogram of autoendcoding errors of image blocks for each image class to\nfacilitate the decision which image regions, or roughly what percentage of an\nimage perhaps, shall be declared relevant for the retrieval task. This leads to\nreduction of feature dimensionality and speeds up the retrieval process. To\nvalidate the proposed scheme, we employ local binary patterns (LBP) and support\nvector machines (SVM) which are both well-established approaches in CBIR\nresearch community. As well, we use IRMA dataset with 14,410 x-ray images as\ntest data. The results show that the dimensionality of annotated feature\nvectors can be reduced by up to 50% resulting in speedups greater than 27% at\nexpense of less than 1% decrease in the accuracy of retrieval when validating\nthe precision and recall of the top 20 hits.', 'Image features for retrieval-based localization must be invariant to dynamic\nobjects (e.g. cars) as well as seasonal and daytime changes. Such invariances\nare, up to some extent, learnable with existing methods using triplet-like\nlosses, given a large number of diverse training images. However, due to the\nhigh algorithmic training complexity, there exists insufficient comparison\nbetween different loss functions on large datasets. In this paper, we train and\nevaluate several localization methods on three different benchmark datasets,\nincluding Oxford RobotCar with over one million images. This large scale\nevaluation yields valuable insights into the generalizability and performance\nof retrieval-based localization. Based on our findings, we develop a novel\nmethod for learning more accurate and better generalizing localization\nfeatures. It consists of two main contributions: (i) a feature volume-based\nloss function, and (ii) hard positive and pairwise negative mining. On the\nchallenging Oxford RobotCar night condition, our method outperforms the\nwell-known triplet loss by 24.4% in localization accuracy within 5m.', 'The research on hashing techniques for visual data is gaining increased\nattention in recent years due to the need for compact representations\nsupporting efficient search/retrieval in large-scale databases such as online\nimages. Among many possibilities, Mean Average Precision(mAP) has emerged as\nthe dominant performance metric for hashing-based retrieval. One glaring\nshortcoming of mAP is its inability in balancing retrieval accuracy and\nutilization of hash codes: pushing a system to attain higher mAP will\ninevitably lead to poorer utilization of the hash codes. Poor utilization of\nthe hash codes hinders good retrieval because of increased collision of samples\nin the hash space. This means that a model giving a higher mAP values does not\nnecessarily do a better job in retrieval. In this paper, we introduce a new\nmetric named Mean Local Group Average Precision (mLGAP) for better evaluation\nof the performance of hashing-based retrieval. The new metric provides a\nretrieval performance measure that also reconciles the utilization of hash\ncodes, leading to a more practically meaningful performance metric than\nconventional ones like mAP. To this end, we start by mathematical analysis of\nthe deficiencies of mAP for hashing-based retrieval. We then propose mLGAP and\nshow why it is more appropriate for hashing-based retrieval. Experiments on\nimage retrieval are used to demonstrate the effectiveness of the proposed\nmetric.', 'Text-to-image retrieval is an essential task in cross-modal information\nretrieval, i.e., retrieving relevant images from a large and unlabelled dataset\ngiven textual queries. In this paper, we propose VisualSparta, a novel\n(Visual-text Sparse Transformer Matching) model that shows significant\nimprovement in terms of both accuracy and efficiency. VisualSparta is capable\nof outperforming previous state-of-the-art scalable methods in MSCOCO and\nFlickr30K. We also show that it achieves substantial retrieving speed\nadvantages, i.e., for a 1 million image index, VisualSparta using CPU gets\n~391X speedup compared to CPU vector search and ~5.4X speedup compared to\nvector search with GPU acceleration. Experiments show that this speed advantage\neven gets bigger for larger datasets because VisualSparta can be efficiently\nimplemented as an inverted index. To the best of our knowledge, VisualSparta is\nthe first transformer-based text-to-image retrieval model that can achieve\nreal-time searching for large-scale datasets, with significant accuracy\nimprovement compared to previous state-of-the-art methods.', 'Most existing works of adversarial samples focus on attacking image\nrecognition models, while little attention is paid to the image retrieval task.\nIn this paper, we identify two inherent challenges in applying prevailing image\nrecognition attack methods to image retrieval. First, image retrieval demands\ndiscriminative visual features, which is significantly different from the\none-hot class prediction in image recognition. Second, due to the disjoint and\npotentially unrelated classes between the training and test set in image\nretrieval, predicting the query category from predefined training classes is\nnot accurate and leads to a sub-optimal adversarial gradient. To address these\nlimitations, we propose a new white-box attack approach, Opposite-Direction\nFeature Attack (ODFA), to generate adversarial queries. Opposite-Direction\nFeature Attack (ODFA) effectively exploits feature-level adversarial gradients\nand takes advantage of feature distance in the representation space. To our\nknowledge, we are among the early attempts to design an attack method\nspecifically for image retrieval. When we deploy an attacked image as the\nquery, the true matches are prone to receive low ranks. We demonstrate through\nextensive experiments that (1) only crafting adversarial queries is sufficient\nto fool the state-of-the-art retrieval systems; (2) the proposed attack method,\nODFA, leads to a higher attack success rate than classification attack methods,\nvalidating the necessity of leveraging characteristics of image retrieval; (3)\nthe adversarial queries generated by our method have good transferability to\nother retrieval models without accessing their parameters, i.e.,the black-box\nsetting.', 'Hash coding has been widely used in approximate nearest neighbor search for\nlarge-scale image retrieval. Given semantic annotations such as class labels\nand pairwise similarities of the training data, hashing methods can learn and\ngenerate effective and compact binary codes. While some newly introduced images\nmay contain undefined semantic labels, which we call unseen images, zeor-shot\nhashing techniques have been studied. However, existing zeor-shot hashing\nmethods focus on the retrieval of single-label images, and cannot handle\nmulti-label images. In this paper, for the first time, a novel transductive\nzero-shot hashing method is proposed for multi-label unseen image retrieval. In\norder to predict the labels of the unseen/target data, a visual-semantic bridge\nis built via instance-concept coherence ranking on the seen/source data. Then,\npairwise similarity loss and focal quantization loss are constructed for\ntraining a hashing model using both the seen/source and unseen/target data.\nExtensive evaluations on three popular multi-label datasets demonstrate that,\nthe proposed hashing method achieves significantly better results than the\ncompeting methods.', 'The analysis of large collections of image data is still a challenging\nproblem due to the difficulty of capturing the true concepts in visual data.\nThe similarity between images could be computed using different and possibly\nmultimodal features such as color or edge information or even text labels. This\nmotivates the design of image analysis solutions that are able to effectively\nintegrate the multi-view information provided by different feature sets. We\ntherefore propose a new image retrieval solution that is able to sort images\nthrough a random walk on a multi-layer graph, where each layer corresponds to a\ndifferent type of information about the image data. We study in depth the\ndesign of the image graph and propose in particular an effective method to\nselect the edge weights for the multi-layer graph, such that the image ranking\nscores are optimised. We then provide extensive experiments in different\nreal-world photo collections, which confirm the high performance of our new\nimage retrieval algorithm that generally surpasses state-of-the-art solutions\ndue to a more meaningful image similarity computation.', 'This paper describes two approaches for content-based image retrieval and\npattern spotting in document images using deep learning. The first approach\nuses a pre-trained CNN model to cope with the lack of training data, which is\nfine-tuned to achieve a compact yet discriminant representation of queries and\nimage candidates. The second approach uses a Siamese Convolution Neural Network\ntrained on a previously prepared subset of image pairs from the ImageNet\ndataset to provide the similarity-based feature maps. In both methods, the\nlearned representation scheme considers feature maps of different sizes which\nare evaluated in terms of retrieval performance. A robust experimental protocol\nusing two public datasets (Tobacoo-800 and DocExplore) has shown that the\nproposed methods compare favorably against state-of-the-art document image\nretrieval and pattern spotting methods.', 'This paper explores the task of interactive image retrieval using natural\nlanguage queries, where a user progressively provides input queries to refine a\nset of retrieval results. Moreover, our work explores this problem in the\ncontext of complex image scenes containing multiple objects. We propose\nDrill-down, an effective framework for encoding multiple queries with an\nefficient compact state representation that significantly extends current\nmethods for single-round image retrieval. We show that using multiple rounds of\nnatural language queries as input can be surprisingly effective to find\narbitrarily specific images of complex scenes. Furthermore, we find that\nexisting image datasets with textual captions can provide a surprisingly\neffective form of weak supervision for this task. We compare our method with\nexisting sequential encoding and embedding networks, demonstrating superior\nperformance on two proposed benchmarks: automatic image retrieval on a\nsimulated scenario that uses region captions as queries, and interactive image\nretrieval using real queries from human evaluators.', 'Supervised hashing aims to map the original features to compact binary codes\nthat are able to preserve label based similarity in the Hamming space.\nNon-linear hash functions have demonstrated the advantage over linear ones due\nto their powerful generalization capability. In the literature, kernel\nfunctions are typically used to achieve non-linearity in hashing, which achieve\nencouraging retrieval performance at the price of slow evaluation and training\ntime. Here we propose to use boosted decision trees for achieving non-linearity\nin hashing, which are fast to train and evaluate, hence more suitable for\nhashing with high dimensional data. In our approach, we first propose\nsub-modular formulations for the hashing binary code inference problem and an\nefficient GraphCut based block search method for solving large-scale inference.\nThen we learn hash functions by training boosted decision trees to fit the\nbinary codes. Experiments demonstrate that our proposed method significantly\noutperforms most state-of-the-art methods in retrieval precision and training\ntime. Especially for high-dimensional data, our method is orders of magnitude\nfaster than many methods in terms of training time.', 'In this paper, we aim to learn a mapping (or embedding) from images to a\ncompact binary space in which Hamming distances correspond to a ranking measure\nfor the image retrieval task.\n  We make use of a triplet loss because this has been shown to be most\neffective for ranking problems.\n  However, training in previous works can be prohibitively expensive due to the\nfact that optimization is directly performed on the triplet space, where the\nnumber of possible triplets for training is cubic in the number of training\nexamples.\n  To address this issue, we propose to formulate high-order binary codes\nlearning as a multi-label classification problem by explicitly separating\nlearning into two interleaved stages.\n  To solve the first stage, we design a large-scale high-order binary codes\ninference algorithm to reduce the high-order objective to a standard binary\nquadratic problem such that graph cuts can be used to efficiently infer the\nbinary code which serve as the label of each training datum.\n  In the second stage we propose to map the original image to compact binary\ncodes via carefully designed deep convolutional neural networks (CNNs) and the\nhashing function fitting can be solved by training binary CNN classifiers.\n  An incremental/interleaved optimization strategy is proffered to ensure that\nthese two steps are interactive with each other during training for better\naccuracy.\n  We conduct experiments on several benchmark datasets, which demonstrate both\nimproved training time (by as much as two orders of magnitude) as well as\nproducing state-of-the-art hashing for various retrieval tasks.', 'Image retrieval utilizes image descriptors to retrieve the most similar\nimages to a given query image. Convolutional neural network (CNN) is becoming\nthe dominant approach to extract image descriptors for image retrieval. For\nlow-power hardware implementation of image retrieval, the drawback of CNN-based\nfeature descriptor is that it requires hundreds of megabytes of storage. To\naddress this problem, this paper applies deep model quantization and\ncompression to CNN in ASIC chip for image retrieval. It is demonstrated that\nthe CNN-based features descriptor can be extracted using as few as 2-bit\nweights quantization to deliver a similar performance as floating-point model\nfor image retrieval. In addition, to implement CNN in ASIC, especially for\nlarge scale images, the limited buffer size of chips should be considered. To\nretrieve large scale images, we propose an improved pooling strategy, region\nnested invariance pooling (RNIP), which uses cropped sub-images for CNN.\nTesting results on chip show that integrating RNIP with the proposed 2-bit CNN\nmodel compression approach is capable of retrieving large scale images.', 'A laser scanner can easily acquire the geometric data of physical\nenvironments in the form of a point cloud. Recognizing objects from a point\ncloud is often required for industrial 3D reconstruction, which should include\nnot only geometry information but also semantic information. However,\nrecognition process is often a bottleneck in 3D reconstruction because it\nrequires expertise on domain knowledge and intensive labor. To address this\nproblem, various methods have been developed to recognize objects by retrieving\nthe corresponding model in the database from an input geometry query. In recent\nyears, the technique of converting geometric data into an image and applying\nview-based 3D shape retrieval has demonstrated high accuracy. Depth image which\nencodes depth value as intensity of pixel is frequently used for view-based 3D\nshape retrieval. However, geometric data collected from objects is often\nincomplete due to the occlusions and the limit of line of sight. Image\ngenerated by occluded point clouds lowers the performance of view-based 3D\nobject retrieval due to loss of information. In this paper, we propose a method\nof viewpoint and image resolution estimation method for view-based 3D shape\nretrieval from point cloud query. Automatic selection of viewpoint and image\nresolution by calculating the data acquisition rate and density from the\nsampled viewpoints and image resolutions are proposed. The retrieval\nperformance from the images generated by the proposed method is experimented\nand compared for various dataset. Additionally, view-based 3D shape retrieval\nperformance with deep convolutional neural network has been experimented with\nthe proposed method.', 'The objective of this work is to learn a compact embedding of a set of\ndescriptors that is suitable for efficient retrieval and ranking, whilst\nmaintaining discriminability of the individual descriptors. We focus on a\nspecific example of this general problem -- that of retrieving images\ncontaining multiple faces from a large scale dataset of images. Here the set\nconsists of the face descriptors in each image, and given a query for multiple\nidentities, the goal is then to retrieve, in order, images which contain all\nthe identities, all but one, \\etc\n  To this end, we make the following contributions: first, we propose a CNN\narchitecture -- {\\em SetNet} -- to achieve the objective: it learns face\ndescriptors and their aggregation over a set to produce a compact fixed length\ndescriptor designed for set retrieval, and the score of an image is a count of\nthe number of identities that match the query; second, we show that this\ncompact descriptor has minimal loss of discriminability up to two faces per\nimage, and degrades slowly after that -- far exceeding a number of baselines;\nthird, we explore the speed vs.\\ retrieval quality trade-off for set retrieval\nusing this compact descriptor; and, finally, we collect and annotate a large\ndataset of images containing various number of celebrities, which we use for\nevaluation and is publicly released.', 'This paper presents a novel approach for image retrieval and pattern spotting\nin document image collections. The manual feature engineering is avoided by\nlearning a similarity-based representation using a Siamese Neural Network\ntrained on a previously prepared subset of image pairs from the ImageNet\ndataset. The learned representation is used to provide the similarity-based\nfeature maps used to find relevant image candidates in the data collection\ngiven an image query. A robust experimental protocol based on the public\nTobacco800 document image collection shows that the proposed method compares\nfavorably against state-of-the-art document image retrieval methods, reaching\n0.94 and 0.83 of mean average precision (mAP) for retrieval and pattern\nspotting (IoU=0.7), respectively. Besides, we have evaluated the proposed\nmethod considering feature maps of different sizes, showing the impact of\nreducing the number of features in the retrieval performance and\ntime-consuming.', 'In this work, we evaluate contrastive models for the task of image retrieval.\nWe hypothesise that models that are learned to encode semantic similarity among\ninstances via discriminative learning should perform well on the task of image\nretrieval, where relevancy is defined in terms of instances of the same object.\nThrough our extensive evaluation, we find that representations from models\ntrained using contrastive methods perform on-par with (and outperforms) a\npre-trained supervised baseline trained on the ImageNet labels in retrieval\ntasks under various configurations. This is remarkable given that the\ncontrastive models require no explicit supervision. Thus, we conclude that\nthese models can be used to bootstrap base models to build more robust image\nretrieval engines.', 'Face image retrieval, which searches for images of the same identity from the\nquery input face image, is drawing more attention as the size of the image\ndatabase increases rapidly. In order to conduct fast and accurate retrieval, a\ncompact hash code-based methods have been proposed, and recently, deep face\nimage hashing methods with supervised classification training have shown\noutstanding performance. However, classification-based scheme has a\ndisadvantage in that it cannot reveal complex similarities between face images\ninto the hash code learning. In this paper, we attempt to improve the face\nimage retrieval quality by proposing a Similarity Guided Hashing (SGH) method,\nwhich gently considers self and pairwise-similarity simultaneously. SGH employs\nvarious data augmentations designed to explore elaborate similarities between\nface images, solving both intra and inter identity-wise difficulties. Extensive\nexperimental results on the protocols with existing benchmarks and an\nadditionally proposed large scale higher resolution face image dataset\ndemonstrate that our SGH delivers state-of-the-art retrieval performance.', 'Image-text retrieval of natural scenes has been a popular research topic.\nSince image and text are heterogeneous cross-modal data, one of the key\nchallenges is how to learn comprehensive yet unified representations to express\nthe multi-modal data. A natural scene image mainly involves two kinds of visual\nconcepts, objects and their relationships, which are equally essential to\nimage-text retrieval. Therefore, a good representation should account for both\nof them. In the light of recent success of scene graph in many CV and NLP tasks\nfor describing complex natural scenes, we propose to represent image and text\nwith two kinds of scene graphs: visual scene graph (VSG) and textual scene\ngraph (TSG), each of which is exploited to jointly characterize objects and\nrelationships in the corresponding modality. The image-text retrieval task is\nthen naturally formulated as cross-modal scene graph matching. Specifically, we\ndesign two particular scene graph encoders in our model for VSG and TSG, which\ncan refine the representation of each node on the graph by aggregating\nneighborhood information. As a result, both object-level and relationship-level\ncross-modal features can be obtained, which favorably enables us to evaluate\nthe similarity of image and text in the two levels in a more plausible way. We\nachieve state-of-the-art results on Flickr30k and MSCOCO, which verifies the\nadvantages of our graph matching based approach for image-text retrieval.', ""This paper considers matching images of low-light scenes, aiming to widen the\nfrontier of SfM and visual SLAM applications. Recent image sensors can record\nthe brightness of scenes with more than eight-bit precision, available in their\nRAW-format image. We are interested in making full use of such high-precision\ninformation to match extremely low-light scene images that conventional methods\ncannot handle. For extreme low-light scenes, even if some of their brightness\ninformation exists in the RAW format images' low bits, the standard raw image\nprocessing on cameras fails to utilize them properly. As was recently shown by\nChen et al., CNNs can learn to produce images with a natural appearance from\nsuch RAW-format images. To consider if and how well we can utilize such\ninformation stored in RAW-format images for image matching, we have created a\nnew dataset named MID (matching in the dark). Using it, we experimentally\nevaluated combinations of eight image-enhancing methods and eleven image\nmatching methods consisting of classical/neural local descriptors and\nclassical/neural initial point-matching methods. The results show the advantage\nof using the RAW-format images and the strengths and weaknesses of the above\ncomponent methods. They also imply there is room for further research."", 'We propose an unsupervised hashing method which aims to produce binary codes\nthat preserve the ranking induced by a real-valued representation. Such compact\nhash codes enable the complete elimination of real-valued feature storage and\nallow for significant reduction of the computation complexity and storage cost\nof large-scale image retrieval applications. Specifically, we learn a neural\nnetwork-based model, which transforms the input representation into a binary\nrepresentation. We formalize the training objective of the network in an\nintuitive and effective way, considering each training sample as a query and\naiming to obtain the same retrieval results using the produced hash codes as\nthose obtained with the original features. This training formulation directly\noptimizes the hashing model for the target usage of the hash codes it produces.\nWe further explore the addition of a decoder trained to obtain an approximated\nreconstruction of the original features. At test time, we retrieved the most\npromising database samples with an efficient graph-based search procedure using\nonly our hash codes and perform re-ranking using the reconstructed features,\nthus without needing to access the original features at all. Experiments\nconducted on multiple publicly available large-scale datasets show that our\nmethod consistently outperforms all compared state-of-the-art unsupervised\nhashing methods and that the reconstruction procedure can effectively boost the\nsearch accuracy with a minimal constant additional cost.', 'Image retrieval has been a top topic in the field of both computer vision and\nmachine learning for a long time. Content based image retrieval, which tries to\nretrieve images from a database visually similar to a query image, has\nattracted much attention. Two most important issues of image retrieval are the\nrepresentation and ranking of the images. Recently, bag-of-words based method\nhas shown its power as a representation method. Moreover, nonnegative matrix\nfactorization is also a popular way to represent the data samples. In addition,\ncontextual similarity learning has also been studied and proven to be an\neffective method for the ranking problem. However, these technologies have\nnever been used together. In this paper, we developed an effective image\nretrieval system by representing each image using the bag-of-words method as\nhistograms, and then apply the nonnegative matrix factorization to factorize\nthe histograms, and finally learn the ranking score using the contextual\nsimilarity learning method. The proposed novel system is evaluated on a large\nscale image database and the effectiveness is shown.', 'Learning the similarity between remote sensing (RS) images forms the\nfoundation for content-based RS image retrieval (CBIR). Recently, deep metric\nlearning approaches that map the semantic similarity of images into an\nembedding (metric) space have been found very popular in RS. A common approach\nfor learning the metric space relies on the selection of triplets of similar\n(positive) and dissimilar (negative) images to a reference image called as an\nanchor. Choosing triplets is a difficult task particularly for multi-label RS\nCBIR, where each training image is annotated by multiple class labels. To\naddress this problem, in this paper we propose a novel triplet sampling method\nin the framework of deep neural networks (DNNs) defined for multi-label RS CBIR\nproblems. The proposed method selects a small set of the most representative\nand informative triplets based on two main steps. In the first step, a set of\nanchors that are diverse to each other in the embedding space is selected from\nthe current mini-batch using an iterative algorithm. In the second step,\ndifferent sets of positive and negative images are chosen for each anchor by\nevaluating the relevancy, hardness and diversity of the images among each other\nbased on a novel strategy. Experimental results obtained on two multi-label\nbenchmark archives show that the selection of the most informative and\nrepresentative triplets in the context of DNNs results in: i) reducing the\ncomputational complexity of the training phase of the DNNs without any\nsignificant loss on the performance; and ii) an increase in learning speed\nsince informative triplets allow fast convergence. The code of the proposed\nmethod is publicly available at\nhttps://git.tu-berlin.de/rsim/image-retrieval-from-triplets.', 'We investigate the gains in precision and speed, that can be obtained by\nusing Convolutional Networks (ConvNets) for on-the-fly retrieval - where\nclassifiers are learnt at run time for a textual query from downloaded images,\nand used to rank large image or video datasets.\n  We make three contributions: (i) we present an evaluation of state-of-the-art\nimage representations for object category retrieval over standard benchmark\ndatasets containing 1M+ images; (ii) we show that ConvNets can be used to\nobtain features which are incredibly performant, and yet much lower dimensional\nthan previous state-of-the-art image representations, and that their\ndimensionality can be reduced further without loss in performance by\ncompression using product quantization or binarization. Consequently, features\nwith the state-of-the-art performance on large-scale datasets of millions of\nimages can fit in the memory of even a commodity GPU card; (iii) we show that\nan SVM classifier can be learnt within a ConvNet framework on a GPU in parallel\nwith downloading the new training images, allowing for a continuous refinement\nof the model as more images become available, and simultaneous training and\nranking. The outcome is an on-the-fly system that significantly outperforms its\npredecessors in terms of: precision of retrieval, memory requirements, and\nspeed, facilitating accurate on-the-fly learning and ranking in under a second\non a single GPU.', 'Unsupervised binary representation allows fast data retrieval without any\nannotations, enabling practical application like fast person re-identification\nand multimedia retrieval. It is argued that conflicts in binary space are one\nof the major barriers to high-performance unsupervised hashing as current\nmethods failed to capture the precise code conflicts in the full domain. A\nnovel relaxation method called Shuffle and Learn is proposed to tackle code\nconflicts in the unsupervised hash. Approximated derivatives for joint\nprobability and the gradients for the binary layer are introduced to bridge the\nupdate from the hash to the input. Proof on $\\epsilon$-Convergence of joint\nprobability with approximated derivatives is provided to guarantee the\npreciseness on update applied on the mutual information. The proposed algorithm\nis carried out with iterative global updates to minimize mutual information,\ndiverging the code before regular unsupervised optimization. Experiments\nsuggest that the proposed method can relax the code optimization from local\noptimum and help to generate binary representations that are more\ndiscriminative and informative without any annotations. Performance benchmarks\non image retrieval with the unsupervised binary code are conducted on three\nopen datasets, and the model achieves state-of-the-art accuracy on image\nretrieval task for all those datasets. Datasets and reproducible code are\nprovided.', 'Document Image Analysis, like any Digital Image Analysis requires\nidentification and extraction of proper features, which are generally extracted\nfrom uncompressed images, though in reality images are made available in\ncompressed form for the reasons such as transmission and storage efficiency.\nHowever, this implies that the compressed image should be decompressed, which\nindents additional computing resources. This limitation induces the motivation\nto research in extracting features directly from the compressed image. In this\nresearch, we propose to extract essential features such as projection profile,\nrun-histogram and entropy for text document analysis directly from run-length\ncompressed text-documents. The experimentation illustrates that features are\nextracted directly from the compressed image without going through the stage of\ndecompression, because of which the computing time is reduced. The feature\nvalues so extracted are exactly identical to those extracted from uncompressed\nimages.', 'Convolutional Neural Network (CNN) is a very powerful approach to extract\ndiscriminative local descriptors for effective image search. Recent work adopts\nfine-tuned strategies to further improve the discriminative power of the\ndescriptors. Taking a different approach, in this paper, we propose a novel\nframework to achieve competitive retrieval performance. Firstly, we propose\nvarious masking schemes, namely SIFT-mask, SUM-mask, and MAX-mask, to select a\nrepresentative subset of local convolutional features and remove a large number\nof redundant features. We demonstrate that this can effectively address the\nburstiness issue and improve retrieval accuracy. Secondly, we propose to employ\nrecent embedding and aggregating methods to further enhance feature\ndiscriminability. Extensive experiments demonstrate that our proposed framework\nachieves state-of-the-art retrieval accuracy.', 'Recently it has become essential to search for and retrieve high-resolution\nand efficient images easily due to swift development of digital images, many\npresent annotation algorithms facing a big challenge which is the variance for\nrepresent the image where high level represent image semantic and low level\nillustrate the features, this issue is known as semantic gab. This work has\nbeen used MPEG-7 standard to extract the features from the images, where the\ncolor feature was extracted by using Scalable Color Descriptor (SCD) and Color\nLayout Descriptor (CLD), whereas the texture feature was extracted by employing\nEdge Histogram Descriptor (EHD), the CLD produced high dimensionality feature\nvector therefore it is reduced by Principal Component Analysis (PCA). The\nfeatures that have extracted by these three descriptors could be passing to the\nclassifiers (Naive Bayes and Decision Tree) for training. Finally, they\nannotated the query image. In this study TUDarmstadt image bank had been used.\nThe results of tests and comparative performance evaluation indicated better\nprecision and executing time of Naive Bayes classification in comparison with\nDecision Tree classification.', 'With the prolification of multimodal interaction in various domains, recently\nthere has been much interest in text based image retrieval in the computer\nvision community. However most of the state of the art techniques model this\nproblem in a purely neural way, which makes it difficult to incorporate\npragmatic strategies in searching a large scale catalog especially when the\nsearch requirements are insufficient and the model needs to resort to an\ninteractive retrieval process through multiple iterations of\nquestion-answering. Motivated by this, we propose a neural-symbolic approach\nfor a one-shot retrieval of images from a large scale catalog, given the\ncaption description. To facilitate this, we represent the catalog and caption\nas scene-graphs and model the retrieval task as a learnable graph matching\nproblem, trained end-to-end with a REINFORCE algorithm. Further, we briefly\ndescribe an extension of this pipeline to an iterative retrieval framework,\nbased on interactive questioning and answering.', 'To retrieve images based on their content is one of the most studied topics\nin the field of computer vision. Nowadays, this problem can be addressed using\nmodern techniques such as feature extraction using machine learning, but over\nthe years different classical methods have been developed. In this paper, we\nimplement a query by example retrieval system for finding paintings in a museum\nimage collection using classic computer vision techniques. Specifically, we\nstudy the performance of the color, texture, text and feature descriptors in\ndatasets with different perturbations in the images: noise, overlapping text\nboxes, color corruption and rotation. We evaluate each of the cases using the\nMean Average Precision (MAP) metric, and we obtain results that vary between\n0.5 and 1.0 depending on the problem conditions.', 'Visual localization algorithms have achieved significant improvements in\nperformance thanks to recent advances in camera technology and vision-based\ntechniques. However, there remains one critical caveat: all current approaches\nthat are based on image retrieval currently scale at best linearly with the\nsize of the environment with respect to both storage, and consequentially in\nmost approaches, query time. This limitation severely curtails the capability\nof autonomous systems in a wide range of compute, power, storage, size, weight\nor cost constrained applications such as drones. In this work, we present a\nnovel binary tree encoding approach for visual localization which can serve as\nan alternative for existing quantization and indexing techniques. The proposed\ntree structure allows us to derive a compressed training scheme that achieves\nsub-linearity in both required storage and inference time. The encoding memory\ncan be easily configured to satisfy different storage constraints. Moreover,\nour approach is amenable to an optional sequence filtering mechanism to further\nimprove the localization results, while maintaining the same amount of storage.\nOur system is entirely agnostic to the front-end descriptors, allowing it to be\nused on top of recent state-of-the-art image representations. Experimental\nresults show that the proposed method significantly outperforms\nstate-of-the-art approaches under limited storage constraints.', 'With the development of Information technology and communication, a large\npart of the databases is dedicated to images and videos. Thus retrieving images\nrelated to a query image from a large database has become an important area of\nresearch in computer vision. Until now, there are various methods of image\nretrieval that try to define image contents by texture, color or shape\nproperties. In this paper, a method is presented for image retrieval based on a\ncombination of local texture information derived from two different texture\ndescriptors. First, the color channels of the input image are separated. The\ntexture information is extracted using two descriptors such as evaluated local\nbinary patterns and predefined pattern units. After extracting the features,\nthe similarity matching is done based on distance criteria. The performance of\nthe proposed method is evaluated in terms of precision and recall on the\nSimplicity database. The comparative results showed that the proposed approach\noffers higher precision rate than many known methods.', 'In this paper, we study the task of image retrieval, where the input query is\nspecified in the form of an image plus some text that describes desired\nmodifications to the input image. For example, we may present an image of the\nEiffel tower, and ask the system to find images which are visually similar but\nare modified in small ways, such as being taken at nighttime instead of during\nthe day. To tackle this task, we learn a similarity metric between a target\nimage and a source image plus source text, an embedding and composing function\nsuch that target image feature is close to the source image plus text\ncomposition feature. We propose a new way to combine image and text using such\nfunction that is designed for the retrieval task. We show this outperforms\nexisting approaches on 3 different datasets, namely Fashion-200k, MIT-States\nand a new synthetic dataset we create based on CLEVR. We also show that our\napproach can be used to classify input queries, in addition to image retrieval.', 'Image to image translation aims to learn a mapping that transforms an image\nfrom one visual domain to another. Recent works assume that images descriptors\ncan be disentangled into a domain-invariant content representation and a\ndomain-specific style representation. Thus, translation models seek to preserve\nthe content of source images while changing the style to a target visual\ndomain. However, synthesizing new images is extremely challenging especially in\nmulti-domain translations, as the network has to compose content and style to\ngenerate reliable and diverse images in multiple domains. In this paper we\npropose the use of an image retrieval system to assist the image-to-image\ntranslation task. First, we train an image-to-image translation model to map\nimages to multiple domains. Then, we train an image retrieval model using real\nand generated images to find images similar to a query one in content but in a\ndifferent domain. Finally, we exploit the image retrieval system to fine-tune\nthe image-to-image translation model and generate higher quality images. Our\nexperiments show the effectiveness of the proposed solution and highlight the\ncontribution of the retrieval network, which can benefit from additional\nunlabeled data and help image-to-image translation models in the presence of\nscarce data.', ""The main focus of this paper is document image classification and retrieval,\nwhere we analyze and compare different parameters for the RunLeght Histogram\n(RL) and Fisher Vector (FV) based image representations. We do an exhaustive\nexperimental study using different document image datasets, including the MARG\nbenchmarks, two datasets built on customer data and the images from the Patent\nImage Classification task of the Clef-IP 2011. The aim of the study is to give\nguidelines on how to best choose the parameters such that the same features\nperform well on different tasks. As an example of such need, we describe the\nImage-based Patent Retrieval task's of Clef-IP 2011, where we used the same\nimage representation to predict the image type and retrieve relevant patents."", 'In recent years, text-guided image manipulation has gained increasing\nattention in the multimedia and computer vision community. The input to\nconditional image generation has evolved from image-only to multimodality. In\nthis paper, we study a setting that allows users to edit an image with multiple\nobjects using complex text instructions to add, remove, or change the objects.\nThe inputs of the task are multimodal including (1) a reference image and (2)\nan instruction in natural language that describes desired modifications to the\nimage. We propose a GAN-based method to tackle this problem. The key idea is to\ntreat text as neural operators to locally modify the image feature. We show\nthat the proposed model performs favorably against recent strong baselines on\nthree public datasets. Specifically, it generates images of greater fidelity\nand semantic relevance, and when used as a image query, leads to better\nretrieval performance.', 'Universal adversarial perturbations (UAPs), a.k.a. input-agnostic\nperturbations, has been proved to exist and be able to fool cutting-edge deep\nlearning models on most of the data samples. Existing UAP methods mainly focus\non attacking image classification models. Nevertheless, little attention has\nbeen paid to attacking image retrieval systems. In this paper, we make the\nfirst attempt in attacking image retrieval systems. Concretely, image retrieval\nattack is to make the retrieval system return irrelevant images to the query at\nthe top ranking list. It plays an important role to corrupt the neighbourhood\nrelationships among features in image retrieval attack. To this end, we propose\na novel method to generate retrieval-against UAP to break the neighbourhood\nrelationships of image features via degrading the corresponding ranking metric.\nTo expand the attack method to scenarios with varying input sizes or\nuntouchable network parameters, a multi-scale random resizing scheme and a\nranking distillation strategy are proposed. We evaluate the proposed method on\nfour widely-used image retrieval datasets, and report a significant performance\ndrop in terms of different metrics, such as mAP and mP@10. Finally, we test our\nattack methods on the real-world visual search engine, i.e., Google Images,\nwhich demonstrates the practical potentials of our methods.', 'Important information that relates to a specific topic in a document is often\norganized in tabular format to assist readers with information retrieval and\ncomparison, which may be difficult to provide in natural language. However,\ntabular data in unstructured digital documents, e.g., Portable Document Format\n(PDF) and images, are difficult to parse into structured machine-readable\nformat, due to complexity and diversity in their structure and style. To\nfacilitate image-based table recognition with deep learning, we develop the\nlargest publicly available table recognition dataset PubTabNet\n(https://github.com/ibm-aur-nlp/PubTabNet), containing 568k table images with\ncorresponding structured HTML representation. PubTabNet is automatically\ngenerated by matching the XML and PDF representations of the scientific\narticles in PubMed Central Open Access Subset (PMCOA). We also propose a novel\nattention-based encoder-dual-decoder (EDD) architecture that converts images of\ntables into HTML code. The model has a structure decoder which reconstructs the\ntable structure and helps the cell decoder to recognize cell content. In\naddition, we propose a new Tree-Edit-Distance-based Similarity (TEDS) metric\nfor table recognition, which more appropriately captures multi-hop cell\nmisalignment and OCR errors than the pre-established metric. The experiments\ndemonstrate that the EDD model can accurately recognize complex tables solely\nrelying on the image representation, outperforming the state-of-the-art by 9.7%\nabsolute TEDS score.', 'We introduce the first work to tackle the image retrieval problem as a\ncontinuous operation. While the proposed approaches in the literature can be\nroughly categorized into two main groups: category- and instance-based\nretrieval, in this work we show that the retrieval task is much richer and more\ncomplex. Image similarity goes beyond this discrete vantage point and spans a\ncontinuous spectrum among the classical operating points of category and\ninstance similarity. However, current retrieval models are static and incapable\nof exploring this rich structure of the retrieval space since they are trained\nand evaluated with a single operating point as a target objective. Hence, we\nintroduce a novel retrieval model that for a given query is capable of\nproducing a dynamic embedding that can target an arbitrary point along the\ncontinuous retrieval spectrum. Our model disentangles the visual signal of a\nquery image into its basic components of categorical and attribute information.\nFurthermore, using a continuous control parameter our model learns to\nreconstruct a dynamic embedding of the query by mixing these components with\ndifferent proportions to target a specific point along the retrieval simplex.\nWe demonstrate our idea in a comprehensive evaluation of the proposed model and\nhighlight the advantages of our approach against a set of well-established\ndiscrete retrieval models.', ""This paper addresses the problem of learning word image representations:\ngiven the cropped image of a word, we are interested in finding a descriptive,\nrobust, and compact fixed-length representation. Machine learning techniques\ncan then be supplied with these representations to produce models useful for\nword retrieval or recognition tasks. Although many works have focused on the\nmachine learning aspect once a global representation has been produced, little\nwork has been devoted to the construction of those base image representations:\nmost works use standard coding and aggregation techniques directly on top of\nstandard computer vision features such as SIFT or HOG.\n  We propose to learn local mid-level features suitable for building word image\nrepresentations. These features are learnt by leveraging character bounding box\nannotations on a small set of training images. However, contrary to other\napproaches that use character bounding box information, our approach does not\nrely on detecting the individual characters explicitly at testing time. Our\nlocal mid-level features can then be aggregated to produce a global word image\nsignature. When pairing these features with the recent word attributes\nframework of Almaz\\'an et al., we obtain results comparable with or better than\nthe state-of-the-art on matching and recognition tasks using global descriptors\nof only 96 dimensions."", 'Content-based image retrieval (CBIR) is a task of retrieving images from\ntheir contents. Since retrieval process is a time-consuming task in large image\ndatabases, acceleration methods can be very useful. This paper presents a novel\nmethod to speed up CBIR systems. In the proposed method, first Zernike moments\nare extracted from query image and an interval is calculated for that query.\nImages in database which are out of the interval are ignored in retrieval\nprocess. Therefore, a database reduction occurs before retrieval which leads to\nspeed up. It is shown that in reduced database, relevant images to query image\nare preserved and irrelevant images are throwed away. Therefore, the proposed\nmethod speed up retrieval process and preserve CBIR accuracy, simultaneously.', 'Content-based image retrieval has seen astonishing progress over the past\ndecade, especially for the task of retrieving images of the same object that is\ndepicted in the query image. This scenario is called instance or object\nretrieval and requires matching fine-grained visual patterns between images.\nSemantics, however, do not play a crucial role. This brings rise to the\nquestion: Do the recent advances in instance retrieval transfer to more generic\nimage retrieval scenarios? To answer this question, we first provide a brief\noverview of the most relevant milestones of instance retrieval. We then apply\nthem to a semantic image retrieval task and find that they perform inferior to\nmuch less sophisticated and more generic methods in a setting that requires\nimage understanding. Following this, we review existing approaches to closing\nthis so-called semantic gap by integrating prior world knowledge. We conclude\nthat the key problem for the further advancement of semantic image retrieval\nlies in the lack of a standardized task definition and an appropriate benchmark\ndataset.', 'Hashing produces compact representations for documents, to perform tasks like\nclassification or retrieval based on these short codes. When hashing is\nsupervised, the codes are trained using labels on the training data. This paper\nfirst shows that the evaluation protocols used in the literature for supervised\nhashing are not satisfactory: we show that a trivial solution that encodes the\noutput of a classifier significantly outperforms existing supervised or\nsemi-supervised methods, while using much shorter codes. We then propose two\nalternative protocols for supervised hashing: one based on retrieval on a\ndisjoint set of classes, and another based on transfer learning to new classes.\nWe provide two baseline methods for image-related tasks to assess the\nperformance of (semi-)supervised hashing: without coding and with unsupervised\ncodes. These baselines give a lower- and upper-bound on the performance of a\nsupervised hashing scheme.', 'Suppose we have a black-box function (e.g., deep neural network) that takes\nan image as input and outputs a value that indicates preference. How can we\nretrieve optimal images with respect to this function from an external database\non the Internet? Standard retrieval problems in the literature (e.g., item\nrecommendations) assume that an algorithm has full access to the set of items.\nIn other words, such algorithms are designed for service providers. In this\npaper, we consider the retrieval problem under different assumptions.\nSpecifically, we consider how users with limited access to an image database\ncan retrieve images using their own black-box functions. This formulation\nenables a flexible and finer-grained image search defined by each user. We\nassume the user can access the database through a search query with tight API\nlimits. Therefore, a user needs to efficiently retrieve optimal images in terms\nof the number of queries. We propose an efficient retrieval algorithm Tiara for\nthis problem. In the experiments, we confirm that our proposed method performs\nbetter than several baselines under various settings.', ""This paper approaches the image retrieval system on the base of visual\nfeatures local region RBIR (region-based image retrieval). First of all, the\npaper presents a method for extracting the interest points based on\nHarris-Laplace to create the feature region of the image. Next, in order to\nreduce the storage space and speed up query image, the paper builds the binary\nsignature structure to describe the visual content of image. Based on the\nimage's binary signature, the paper builds the SG (signature graph) to classify\nand store image's binary signatures. Since then, the paper builds the image\nretrieval algorithm on SG through the similar measure EMD (earth mover's\ndistance) between the image's binary signatures. Last but not least, the paper\ngives an image retrieval model RBIR, experiments and assesses the image\nretrieval method on Corel image database over 10,000 images."", ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'With the advances in both stable interest region detectors and robust and\ndistinctive descriptors, local feature-based image or object retrieval has\nbecome a popular research topic. %All of the local feature-based image\nretrieval system involves two important processes: local feature extraction and\nimage representation. The other key technology for image retrieval systems is\nimage representation such as the bag-of-visual words (BoVW), Fisher vector, or\nVector of Locally Aggregated Descriptors (VLAD) framework. In this paper, we\nreview local features and image representations for image retrieval. Because\nmany and many methods are proposed in this area, these methods are grouped into\nseveral classes and summarized. In addition, recent deep learning-based\napproaches for image retrieval are briefly reviewed.', 'In this paper, we consider a different data format for images: vector\ngraphics. In contrast to raster graphics which are widely used in image\nrecognition, vector graphics can be scaled up or down into any resolution\nwithout aliasing or information loss, due to the analytic representation of the\nprimitives in the document. Furthermore, vector graphics are able to give extra\nstructural information on how low-level elements group together to form high\nlevel shapes or structures. These merits of graphic vectors have not been fully\nleveraged in existing methods. To explore this data format, we target on the\nfundamental recognition tasks: object localization and classification. We\npropose an efficient CNN-free pipeline that does not render the graphic into\npixels (i.e. rasterization), and takes textual document of the vector graphics\nas input, called YOLaT (You Only Look at Text). YOLaT builds multi-graphs to\nmodel the structural and spatial information in vector graphics, and a\ndual-stream graph neural network is proposed to detect objects from the graph.\nOur experiments show that by directly operating on vector graphics, YOLaT\nout-performs raster-graphic based object detection baselines in terms of both\naverage precision and efficiency.', 'The explosive growth of digital images in video surveillance and social media\nhas led to the significant need for efficient search of persons of interest in\nlaw enforcement and forensic applications. Despite tremendous progress in\nprimary biometric traits (e.g., face and fingerprint) based person\nidentification, a single biometric trait alone cannot meet the desired\nrecognition accuracy in forensic scenarios. Tattoos, as one of the important\nsoft biometric traits, have been found to be valuable for assisting in person\nidentification. However, tattoo search in a large collection of unconstrained\nimages remains a difficult problem, and existing tattoo search methods mainly\nfocus on matching cropped tattoos, which is different from real application\nscenarios. To close the gap, we propose an efficient tattoo search approach\nthat is able to learn tattoo detection and compact representation jointly in a\nsingle convolutional neural network (CNN) via multi-task learning. While the\nfeatures in the backbone network are shared by both tattoo detection and\ncompact representation learning, individual latent layers of each sub-network\noptimize the shared features toward the detection and feature learning tasks,\nrespectively. We resolve the small batch size issue inside the joint tattoo\ndetection and compact representation learning network via random image stitch\nand preceding feature buffering. We evaluate the proposed tattoo search system\nusing multiple public-domain tattoo benchmarks, and a gallery set with about\n300K distracter tattoo images compiled from these datasets and images from the\nInternet. In addition, we also introduce a tattoo sketch dataset containing 300\ntattoos for sketch-based tattoo search. Experimental results show that the\nproposed approach has superior performance in tattoo detection and tattoo\nsearch at scale compared to several state-of-the-art tattoo retrieval\nalgorithms.', 'When one wants to train a neural network to perform semantic segmentation,\ncreating pixel-level annotations for each of the images in the database is a\ntedious task. If he works with aerial or satellite images, which are usually\nvery large, it is even worse. With that in mind, we investigate how to use\nimage-level annotations in order to perform semantic segmentation. Image-level\nannotations are much less expensive to acquire than pixel-level annotations,\nbut we lose a lot of information for the training of the model. From the\nannotations of the images, the model must find by itself how to classify the\ndifferent regions of the image. In this work, we use the method proposed by Anh\nand Kwak [1] to produce pixel-level annotation from image level annotation. We\ncompare the overall quality of our generated dataset with the original dataset.\nIn addition, we propose an adaptation of the AffinityNet that allows us to\ndirectly perform a semantic segmentation. Our results show that the generated\nlabels lead to the same performances for the training of several segmentation\nnetworks. Also, the quality of semantic segmentation performed directly by the\nAffinityNet and the Random Walk is close to the one of the best\nfully-supervised approaches.', ""Existing deep quantization methods provided an efficient solution for\nlarge-scale image retrieval. However, the significant intra-class variations\nlike pose, illumination, and expressions in face images, still pose a challenge\nfor face image retrieval. In light of this, face image retrieval requires\nsufficiently powerful learning metrics, which are absent in current deep\nquantization works. Moreover, to tackle the growing unseen identities in the\nquery stage, face image retrieval drives more demands regarding model\ngeneralization and system scalability than general image retrieval tasks. This\npaper integrates product quantization with orthonormal constraints into an\nend-to-end deep learning framework to effectively retrieve face images.\nSpecifically, a novel scheme that uses predefined orthonormal vectors as\ncodewords is proposed to enhance the quantization informativeness and reduce\ncodewords' redundancy. A tailored loss function maximizes discriminability\namong identities in each quantization subspace for both the quantized and\noriginal features. An entropy-based regularization term is imposed to reduce\nthe quantization error. Experiments are conducted on four commonly-used face\ndatasets under both seen and unseen identities retrieval settings. Our method\noutperforms all the compared deep hashing/quantization state-of-the-arts under\nboth settings. Results validate the effectiveness of the proposed orthonormal\ncodewords in improving models' standard retrieval performance and\ngeneralization ability. Combing with further experiments on two general image\ndatasets, it demonstrates the broad superiority of our method for scalable\nimage retrieval."", 'Recovering the original printed texts from a document with added handwritten\nannotations in the marginal area is one of the challenging problems, especially\nwhen the original document is not available. Therefore, this paper aims at\nsalvaging automatically the original document from the annotated document by\ndetecting and removing any handwritten annotations that appear in the marginal\narea of the document without any loss of information. Here a two stage\nalgorithm is proposed, where in the first stage due to approximate marginal\nboundary detection with horizontal and vertical projection profiles, all of the\nmarginal annotations along with some part of the original printed text that may\nappear very close to the marginal boundary are removed. Therefore as a second\nstage, using the connected components, a strategy is applied to bring back the\nprinted text components cropped during the first stage. The proposed method is\nvalidated using a dataset of 50 documents having complex handwritten\nannotations, which gives an overall accuracy of 89.01% in removing the marginal\nannotations and 97.74% in case of retrieving the original printed text\ndocument.', 'Everyone knows that thousand of words are represented by a single image. As a\nresult image search has become a very popular mechanism for the Web searchers.\nImage search means, the search results are produced by the search engine should\nbe a set of images along with their Web page Unified Resource Locator. Now Web\nsearcher can perform two types of image search, they are Text to Image and\nImage to Image search. In Text to Image search, search query should be a text.\nBased on the input text data system will generate a set of images along with\ntheir Web page URL as an output. On the other hand, in Image to Image search,\nsearch query should be an image and based on this image system will generate a\nset of images along with their Web page URL as an output. According to the\ncurrent scenarios, Text to Image search mechanism always not returns perfect\nresult. It matches the text data and then displays the corresponding images as\nan output, which is not always perfect. To resolve this problem, Web\nresearchers have introduced the Image to Image search mechanism. In this paper,\nwe have also proposed an alternate approach of Image to Image search mechanism\nusing Histogram.', 'Access to online visual search engines implies sharing of private user\ncontent - the query images. We introduce the concept of targeted mismatch\nattack for deep learning based retrieval systems to generate an adversarial\nimage to conceal the query image. The generated image looks nothing like the\nuser intended query, but leads to identical or very similar retrieval results.\nTransferring attacks to fully unseen networks is challenging. We show\nsuccessful attacks to partially unknown systems, by designing various loss\nfunctions for the adversarial image construction. These include loss functions,\nfor example, for unknown global pooling operation or unknown input resolution\nby the retrieval system. We evaluate the attacks on standard retrieval\nbenchmarks and compare the results retrieved with the original and adversarial\nimage.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently working on a project that involves transforming black-and-white videos into colored versions, with a particular focus on color propagation in videos. I have found that traditional methods of color propagation tend to be computationally demanding. Therefore, I am seeking a more efficient method, especially when additional video information is available. I am interested in a model that can utilize a set of colored key frames from the video and apply color propagation algorithms to colorize other frames based on the provided colored ones. Ideally, this method should incorporate both local and global strategies. The local strategies should ensure color consistency in neighboring regions and frames, thereby maintaining stable color propagation. Simultaneously, from a global perspective, the model should be capable of identifying the keyframes that require colorization, ensuring overall process efficiency. While I value the efficiency of the color propagation methods, I am willing to compromise on quality, provided the performance surpasses the baselines established within the past five years.","['Videos contain highly redundant information between frames. Such redundancy\nhas been extensively studied in video compression and encoding, but is less\nexplored for more advanced video processing. In this paper, we propose a\nlearnable unified framework for propagating a variety of visual properties of\nvideo images, including but not limited to color, high dynamic range (HDR), and\nsegmentation information, where the properties are available for only a few\nkey-frames. Our approach is based on a temporal propagation network (TPN),\nwhich models the transition-related affinity between a pair of frames in a\npurely data-driven manner. We theoretically prove two essential factors for\nTPN: (a) by regularizing the global transformation matrix as orthogonal, the\n""style energy"" of the property can be well preserved during propagation; (b)\nsuch regularization can be achieved by the proposed switchable TPN with\nbi-directional training on pairs of frames. We apply the switchable TPN to\nthree tasks: colorizing a gray-scale video based on a few color key-frames,\ngenerating an HDR video from a low dynamic range (LDR) video and a few HDR\nframes, and propagating a segmentation mask from the first frame in videos.\nExperimental results show that our approach is significantly more accurate and\nefficient than the state-of-the-art methods.', 'Video colorization is a challenging and highly ill-posed problem. Although\nrecent years have witnessed remarkable progress in single image colorization,\nthere is relatively less research effort on video colorization and existing\nmethods always suffer from severe flickering artifacts (temporal inconsistency)\nor unsatisfying colorization performance. We address this problem from a new\nperspective, by jointly considering colorization and temporal consistency in a\nunified framework. Specifically, we propose a novel temporally consistent video\ncolorization framework (TCVC). TCVC effectively propagates frame-level deep\nfeatures in a bidirectional way to enhance the temporal consistency of\ncolorization. Furthermore, TCVC introduces a self-regularization learning (SRL)\nscheme to minimize the prediction difference obtained with different time\nsteps. SRL does not require any ground-truth color videos for training and can\nfurther improve temporal consistency. Experiments demonstrate that our method\ncan not only obtain visually pleasing colorized video, but also achieve clearly\nbetter temporal consistency than state-of-the-art methods.', 'We propose a hybrid recurrent Video Colorization with Hybrid Generative\nAdversarial Network (VCGAN), an improved approach to video colorization using\nend-to-end learning. The VCGAN addresses two prevalent issues in the video\ncolorization domain: Temporal consistency and unification of colorization\nnetwork and refinement network into a single architecture. To enhance\ncolorization quality and spatiotemporal consistency, the mainstream of\ngenerator in VCGAN is assisted by two additional networks, i.e., global feature\nextractor and placeholder feature extractor, respectively. The global feature\nextractor encodes the global semantics of grayscale input to enhance\ncolorization quality, whereas the placeholder feature extractor acts as a\nfeedback connection to encode the semantics of the previous colorized frame in\norder to maintain spatiotemporal consistency. If changing the input for\nplaceholder feature extractor as grayscale input, the hybrid VCGAN also has the\npotential to perform image colorization. To improve the consistency of far\nframes, we propose a dense long-term loss that smooths the temporal disparity\nof every two remote frames. Trained with colorization and temporal losses\njointly, VCGAN strikes a good balance between color vividness and video\ncontinuity. Experimental results demonstrate that VCGAN produces higher-quality\nand temporally more consistent colorful videos than existing approaches.', 'Coloring line art images based on the colors of reference images is an\nimportant stage in animation production, which is time-consuming and tedious.\nIn this paper, we propose a deep architecture to automatically color line art\nvideos with the same color style as the given reference images. Our framework\nconsists of a color transform network and a temporal constraint network. The\ncolor transform network takes the target line art images as well as the line\nart and color images of one or more reference images as input, and generates\ncorresponding target color images. To cope with larger differences between the\ntarget line art image and reference color images, our architecture utilizes\nnon-local similarity matching to determine the region correspondences between\nthe target image and the reference images, which are used to transform the\nlocal color information from the references to the target. To ensure global\ncolor style consistency, we further incorporate Adaptive Instance Normalization\n(AdaIN) with the transformation parameters obtained from a style embedding\nvector that describes the global color style of the references, extracted by an\nembedder. The temporal constraint network takes the reference images and the\ntarget image together in chronological order, and learns the spatiotemporal\nfeatures through 3D convolution to ensure the temporal consistency of the\ntarget image and the reference image. Our model can achieve even better\ncoloring results by fine-tuning the parameters with only a small amount of\nsamples when dealing with an animation of a new style. To evaluate our method,\nwe build a line art coloring dataset. Experiments show that our method achieves\nthe best performance on line art video coloring compared to the\nstate-of-the-art methods and other baselines.', 'Endoscopic videos from multicentres often have different imaging conditions,\ne.g., color and illumination, which make the models trained on one domain\nusually fail to generalize well to another. Domain adaptation is one of the\npotential solutions to address the problem. However, few of existing works\nfocused on the translation of video-based data. In this work, we propose a\nnovel generative adversarial network (GAN), namely VideoGAN, to transfer the\nvideo-based data across different domains. As the frames of a video may have\nsimilar content and imaging conditions, the proposed VideoGAN has an X-shape\ngenerator to preserve the intra-video consistency during translation.\nFurthermore, a loss function, namely color histogram loss, is proposed to tune\nthe color distribution of each translated frame. Two colonoscopic datasets from\ndifferent centres, i.e., CVC-Clinic and ETIS-Larib, are adopted to evaluate the\nperformance of domain adaptation of our VideoGAN. Experimental results\ndemonstrate that the adapted colonoscopic video generated by our VideoGAN can\nsignificantly boost the segmentation accuracy, i.e., an improvement of 5%, of\ncolorectal polyps on multicentre datasets. As our VideoGAN is a general network\narchitecture, we also evaluate its performance with the CamVid driving video\ndataset on the cloudy-to-sunny translation task. Comprehensive experiments show\nthat the domain gap could be substantially narrowed down by our VideoGAN.', 'Applying image processing algorithms independently to each frame of a video\noften leads to undesired inconsistent results over time. Developing temporally\nconsistent video-based extensions, however, requires domain knowledge for\nindividual tasks and is unable to generalize to other applications. In this\npaper, we present an efficient end-to-end approach based on deep recurrent\nnetwork for enforcing temporal consistency in a video. Our method takes the\noriginal unprocessed and per-frame processed videos as inputs to produce a\ntemporally consistent video. Consequently, our approach is agnostic to specific\nimage processing algorithms applied on the original video. We train the\nproposed network by minimizing both short-term and long-term temporal losses as\nwell as the perceptual loss to strike a balance between temporal stability and\nperceptual similarity with the processed frames. At test time, our model does\nnot require computing optical flow and thus achieves real-time speed even for\nhigh-resolution videos. We show that our single model can handle multiple and\nunseen tasks, including but not limited to artistic style transfer,\nenhancement, colorization, image-to-image translation and intrinsic image\ndecomposition. Extensive objective evaluation and subject study demonstrate\nthat the proposed approach performs favorably against the state-of-the-art\nmethods on various types of videos.', 'Models optimized for accuracy on single images are often prohibitively slow\nto run on each frame in a video. Recent work exploits the use of optical flow\nto warp image features forward from select keyframes, as a means to conserve\ncomputation on video. This approach, however, achieves only limited speedup,\neven when optimized, due to the accuracy degradation introduced by repeated\nforward warping, and the inference cost of optical flow estimation. To address\nthese problems, we propose a new scheme that propagates features using the\nblock motion vectors (BMV) present in compressed video (e.g. H.264 codecs),\ninstead of optical flow, and bi-directionally warps and fuses features from\nenclosing keyframes to capture scene context on each video frame. Our\ntechnique, interpolation-BMV, enables us to accurately estimate the features of\nintermediate frames, while keeping inference costs low. We evaluate our system\non the CamVid and Cityscapes datasets, comparing to both a strong single-frame\nbaseline and related work. We find that we are able to substantially accelerate\nsegmentation on video, achieving near real-time frame rates (20+ frames per\nsecond) on large images (e.g. 960 x 720 pixels), while maintaining competitive\naccuracy. This represents an improvement of almost 6x over the single-frame\nbaseline and 2.5x over the fastest prior work.', ""In this paper, we tackle the problem of colorization of grayscale videos to\nreduce bandwidth usage. For this task, we use some colored keyframes as\nreference images from the colored version of the grayscale video. We propose a\nmodel that extracts keyframes from a colored video and trains a Convolutional\nNeural Network from scratch on these colored frames. Through the extracted\nkeyframes we get a good knowledge of the colors that have been used in the\nvideo which helps us in colorizing the grayscale version of the video\nefficiently. An application of the technique that we propose in this paper, is\nin saving bandwidth while sending raw colored videos that haven't gone through\nany compression. A raw colored video takes up around three times more memory\nsize than its grayscale version. We can exploit this fact and send a grayscale\nvideo along with out trained model instead of a colored video. Later on, in\nthis paper we show how this technique can help to save bandwidth usage to upto\nthree times while transmitting raw colored videos."", 'This paper presents the first end-to-end network for exemplar-based video\ncolorization. The main challenge is to achieve temporal consistency while\nremaining faithful to the reference style. To address this issue, we introduce\na recurrent framework that unifies the semantic correspondence and color\npropagation steps. Both steps allow a provided reference image to guide the\ncolorization of every frame, thus reducing accumulated propagation errors.\nVideo frames are colorized in sequence based on the colorization history, and\nits coherency is further enforced by the temporal consistency loss. All of\nthese components, learned end-to-end, help produce realistic videos with good\ntemporal stability. Experiments show our result is superior to the\nstate-of-the-art methods both quantitatively and qualitatively.', 'Video object segmentation is a fundamental step in many advanced vision\napplications. Most existing algorithms are based on handcrafted features such\nas HOG, super-pixel segmentation or texture-based techniques, while recently\ndeep features have been found to be more efficient. Existing algorithms observe\nperformance degradation in the presence of challenges such as illumination\nvariations, shadows, and color camouflage. To handle these challenges we\npropose a fusion based moving object segmentation algorithm which exploits\ncolor as well as depth information using GAN to achieve more accuracy. Our goal\nis to segment moving objects in the presence of challenging background scenes,\nin real environments. To address this problem, GAN is trained in an\nunsupervised manner on color and depth information independently with\nchallenging video sequences. During testing, the trained GAN generates\nbackgrounds similar to that in the test sample. The generated background\nsamples are then compared with the test sample to segment moving objects. The\nfinal result is computed by fusion of object boundaries in both modalities, RGB\nand the depth. The comparison of our proposed algorithm with five\nstate-of-the-art methods on publicly available dataset has shown the strength\nof our algorithm for moving object segmentation in videos in the presence of\nchallenging real scenarios.', 'A background model describes a scene without any foreground objects and has a\nnumber of applications, ranging from video surveillance to computational\nphotography. Recent studies have introduced the method of Dynamic Mode\nDecomposition (DMD) for robustly separating video frames into a background\nmodel and foreground components. While the method introduced operates by\nconverting color images to grayscale, we in this study propose a technique to\nobtain the background model in the color domain. The effectiveness of our\ntechnique is demonstrated using a publicly available Scene Background\nInitialisation (SBI) dataset. Our results both qualitatively and quantitatively\nshow that DMD can successfully obtain a colored background model.', ""We propose a technique that propagates information forward through video\ndata. The method is conceptually simple and can be applied to tasks that\nrequire the propagation of structured information, such as semantic labels,\nbased on video content. We propose a 'Video Propagation Network' that processes\nvideo frames in an adaptive manner. The model is applied online: it propagates\ninformation forward without the need to access future frames. In particular we\ncombine two components, a temporal bilateral network for dense and video\nadaptive filtering, followed by a spatial network to refine features and\nincreased flexibility. We present experiments on video object segmentation and\nsemantic video segmentation and show increased performance comparing to the\nbest previous task-specific methods, while having favorable runtime.\nAdditionally we demonstrate our approach on an example regression task of color\npropagation in a grayscale video."", 'We address the problem of soft color segmentation, defined as decomposing a\ngiven image into several RGBA layers, each containing only homogeneous color\nregions. The resulting layers from decomposition pave the way for applications\nthat benefit from layer-based editing, such as recoloring and compositing of\nimages and videos. The current state-of-the-art approach for this problem is\nhindered by slow processing time due to its iterative nature, and consequently\ndoes not scale to certain real-world scenarios. To address this issue, we\npropose a neural network based method for this task that decomposes a given\nimage into multiple layers in a single forward pass. Furthermore, our method\nseparately decomposes the color layers and the alpha channel layers. By\nleveraging a novel training objective, our method achieves proper assignment of\ncolors amongst layers. As a consequence, our method achieve promising quality\nwithout existing issue of inference speed for iterative approaches. Our\nthorough experimental analysis shows that our method produces qualitative and\nquantitative results comparable to previous methods while achieving a 300,000x\nspeed improvement. Finally, we utilize our proposed method on several\napplications, and demonstrate its speed advantage, especially in video editing.']","['The remastering of vintage film comprises of a diversity of sub-tasks\nincluding super-resolution, noise removal, and contrast enhancement which aim\nto restore the deteriorated film medium to its original state. Additionally,\ndue to the technical limitations of the time, most vintage film is either\nrecorded in black and white, or has low quality colors, for which colorization\nbecomes necessary. In this work, we propose a single framework to tackle the\nentire remastering task semi-interactively. Our work is based on temporal\nconvolutional neural networks with attention mechanisms trained on videos with\ndata-driven deterioration simulation. Our proposed source-reference attention\nallows the model to handle an arbitrary number of reference color images to\ncolorize long videos without the need for segmentation while maintaining\ntemporal consistency. Quantitative analysis shows that our framework\noutperforms existing approaches, and that, in contrast to existing approaches,\nthe performance of our framework increases with longer videos and more\nreference color images.', 'Convolutional networks optimized for accuracy on challenging, dense\nprediction tasks are prohibitively slow to run on each frame in a video. The\nspatial similarity of nearby video frames, however, suggests opportunity to\nreuse computation. Existing work has explored basic feature reuse and feature\nwarping based on optical flow, but has encountered limits to the speedup\nattainable with these techniques. In this paper, we present a new, two part\napproach to accelerating inference on video. First, we propose a fast feature\npropagation technique that utilizes the block motion vectors present in\ncompressed video (e.g. H.264 codecs) to cheaply propagate features from frame\nto frame. Second, we develop a novel feature estimation scheme, termed feature\ninterpolation, that fuses features propagated from enclosing keyframes to\nrender accurate feature estimates, even at sparse keyframe frequencies. We\nevaluate our system on the Cityscapes and CamVid datasets, comparing to both a\nframe-by-frame baseline and related work. We find that we are able to\nsubstantially accelerate segmentation on video, achieving near real-time frame\nrates (20.1 frames per second) on large images (960 x 720 pixels), while\nmaintaining competitive accuracy. This represents an improvement of almost 6x\nover the single-frame baseline and 2.5x over the fastest prior work.', 'Video inpainting, which aims at filling in missing regions of a video,\nremains challenging due to the difficulty of preserving the precise spatial and\ntemporal coherence of video contents. In this work we propose a novel\nflow-guided video inpainting approach. Rather than filling in the RGB pixels of\neach frame directly, we consider video inpainting as a pixel propagation\nproblem. We first synthesize a spatially and temporally coherent optical flow\nfield across video frames using a newly designed Deep Flow Completion network.\nThen the synthesized flow field is used to guide the propagation of pixels to\nfill up the missing regions in the video. Specifically, the Deep Flow\nCompletion network follows a coarse-to-fine refinement to complete the flow\nfields, while their quality is further improved by hard flow example mining.\nFollowing the guide of the completed flow, the missing video regions can be\nfilled up precisely. Our method is evaluated on DAVIS and YouTube-VOS datasets\nqualitatively and quantitatively, achieving the state-of-the-art performance in\nterms of inpainting quality and speed.', 'Segmentation of an object from a video is a challenging task in multimedia\napplications. Depending on the application, automatic or interactive methods\nare desired; however, regardless of the application type, efficient computation\nof video object segmentation is crucial for time-critical applications;\nspecifically, mobile and interactive applications require near real-time\nefficiencies. In this paper, we address the problem of video segmentation from\nthe perspective of efficiency. We initially redefine the problem of video\nobject segmentation as the propagation of MRF energies along the temporal\ndomain. For this purpose, a novel and efficient method is proposed to propagate\nMRF energies throughout the frames via bilateral filters without using any\nglobal texture, color or shape model. Recently presented bi-exponential filter\nis utilized for efficiency, whereas a novel technique is also developed to\ndynamically solve graph-cuts for varying, non-lattice graphs in general linear\nfiltering scenario. These improvements are experimented for both automatic and\ninteractive video segmentation scenarios. Moreover, in addition to the\nefficiency, segmentation quality is also tested both quantitatively and\nqualitatively. Indeed, for some challenging examples, significant time\nefficiency is observed without loss of segmentation quality.', 'One of the challenges of computer vision is that it needs to adapt to color\ndeviations in changeable environments. Therefore, minimizing the adverse\neffects of color deviation on the prediction is one of the main goals of vision\ntask. Current solutions focus on using generative models to augment training\ndata to enhance the invariance of input variation. However, such methods often\nintroduce new noise, which limits the gain from generated data. To this end,\nthis paper proposes a strategy eliminate deviation with deviation, which is\nnamed Random Color Dropout (RCD). Our hypothesis is that if there are color\ndeviation between the query image and the gallery image, the retrieval results\nof some examples will be better after ignoring the color information.\nSpecifically, this strategy balances the weights between color features and\ncolor-independent features in the neural network by dropouting partial color\ninformation in the training data, so as to overcome the effect of color\ndevitaion. The proposed RCD can be combined with various existing ReID models\nwithout changing the learning strategy, and can be applied to other computer\nvision fields, such as object detection. Experiments on several ReID baselines\nand three common large-scale datasets such as Market1501, DukeMTMC, and MSMT17\nhave verified the effectiveness of this method. Experiments on Cross-domain\ntests have shown that this strategy is significant eliminating the domain gap.\nFurthermore, in order to understand the working mechanism of RCD, we analyzed\nthe effectiveness of this strategy from the perspective of classification,\nwhich reveals that it may be better to utilize many instead of all of color\ninformation in visual tasks with strong domain variations.', 'Automatic animation line art colorization is a challenging computer vision\nproblem, since the information of the line art is highly sparse and abstracted\nand there exists a strict requirement for the color and style consistency\nbetween frames. Recently, a lot of Generative Adversarial Network (GAN) based\nimage-to-image translation methods for single line art colorization have\nemerged. They can generate perceptually appealing results conditioned on line\nart images. However, these methods can not be adopted for the purpose of\nanimation colorization because there is a lack of consideration of the\nin-between frame consistency. Existing methods simply input the previous\ncolored frame as a reference to color the next line art, which will mislead the\ncolorization due to the spatial misalignment of the previous colored frame and\nthe next line art especially at positions where apparent changes happen. To\naddress these challenges, we design a kind of correlation matching feature\ntransfer model (called CMFT) to align the colored reference feature in a\nlearnable way and integrate the model into an U-Net based generator in a\ncoarse-to-fine manner. This enables the generator to transfer the layer-wise\nsynchronized features from the deep semantic code to the content progressively.\nExtension evaluation shows that CMFT model can effectively improve the\nin-between consistency and the quality of colored frames especially when the\nmotion is intense and diverse.', 'For many years, channels of a color image have been processed individually,\nor the image has been converted to grayscale one with respect to color image\nprocessing. Pure quaternion representation of color images solves this issue as\nit allows images to be processed in a holistic space. Nevertheless, it brings\nadditional costs due to the extra fourth dimension. In this paper, we propose\nan approach for representing color images with full quaternion numbers that\nenables us to process color images holistically without additional cost in\ntime, space and computation. With taking auto- and cross-correlation of color\nchannels into account, an autoencoder neural network is used to generate a\nglobal model for transforming a color image into a full quaternion matrix. To\nevaluate the model, we use UCID dataset, and the results indicate that the\nmodel has an acceptable performance on color images. Moreover, we propose a\ncompression method based on the generated model and QSVD as a case study. The\nmethod is compared with the same compression method using pure quaternion\nrepresentation and is assessed with UCID dataset. The results demonstrate that\nthe compression method using the proposed full quaternion representation fares\nbetter than the other in terms of time, quality, and size of compressed files.', 'We present a novel approach of color transfer between images by exploring\ntheir high-level semantic information. First, we set up a database which\nconsists of the collection of downloaded images from the internet, which are\nsegmented automatically by using matting techniques. We then, extract image\nforegrounds from both source and multiple target images. Then by using image\nmatting algorithms, the system extracts the semantic information such as faces,\nlips, teeth, eyes, eyebrows, etc., from the extracted foregrounds of the source\nimage. And, then the color is transferred between corresponding parts with the\nsame semantic information. Next we get the color transferred result by\nseamlessly compositing different parts together using alpha blending. In the\nfinal step, we present an efficient method of color consistency to optimize the\ncolor of a collection of images showing the common scene. The main advantage of\nour method over existing techniques is that it does not need face matching, as\none could use more than one target images. It is not restricted to head shot\nimages as we can also change the color style in the wild. Moreover, our\nalgorithm does not require to choose the same color style, same pose and image\nsize between source and target images. Our algorithm is not restricted to\none-to-one image color transfer and can make use of more than one target images\nto transfer the color in different parts in the source image. Comparing with\nother approaches, our algorithm is much better in color blending in the input\ndata.', 'In this work, we present a method for automatic colorization of grayscale\nvideos. The core of the method is a Generative Adversarial Network that is\ntrained and tested on sequences of frames in a sliding window manner. Network\nconvolutional and deconvolutional layers are three-dimensional, with frame\nheight, width and time as the dimensions taken into account. Multiple\nchrominance estimates per frame are aggregated and combined with available\nluminance information to recreate a colored sequence. Colorization trials are\nrun succesfully on a dataset of old black-and-white films. The usefulness of\nour method is also validated with numerical results, computed with a newly\nproposed metric that measures colorization consistency over a frame sequence.', 'We present a fully automatic approach to video colorization with\nself-regularization and diversity. Our model contains a colorization network\nfor video frame colorization and a refinement network for spatiotemporal color\nrefinement. Without any labeled data, both networks can be trained with\nself-regularized losses defined in bilateral and temporal space. The bilateral\nloss enforces color consistency between neighboring pixels in a bilateral space\nand the temporal loss imposes constraints between corresponding pixels in two\nnearby frames. While video colorization is a multi-modal problem, our method\nuses a perceptual loss with diversity to differentiate various modes in the\nsolution space. Perceptual experiments demonstrate that our approach\noutperforms state-of-the-art approaches on fully automatic video colorization.\nThe results are shown in the supplementary video at\nhttps://youtu.be/Y15uv2jnK-4', 'Document image enhancement and binarization methods are often used to improve\nthe accuracy and efficiency of document image analysis tasks such as text\nrecognition. Traditional non-machine-learning methods are constructed on\nlow-level features in an unsupervised manner but have difficulty with\nbinarization on documents with severely degraded backgrounds. Convolutional\nneural network-based methods focus only on grayscale images and on local\ntextual features. In this paper, we propose a two-stage color document image\nenhancement and binarization method using generative adversarial neural\nnetworks. In the first stage, four color-independent adversarial networks are\ntrained to extract color foreground information from an input image for\ndocument image enhancement. In the second stage, two independent adversarial\nnetworks with global and local features are trained for image binarization of\ndocuments of variable size. For the adversarial neural networks, we formulate\nloss functions between a discriminator and generators having an encoder-decoder\nstructure. Experimental results show that the proposed method achieves better\nperformance than many classical and state-of-the-art algorithms over the\nDocument Image Binarization Contest (DIBCO) datasets, the LRDE Document\nBinarization Dataset (LRDE DBD), and our shipping label image dataset. We plan\nto release the shipping label dataset as well as our implementation code at\ngithub.com/opensuh/DocumentBinarization/.', ""In industry, there exist plenty of scenarios where old gray photos need to be\nautomatically colored, such as video sites and archives. In this paper, we\npresent the HistoryNet focusing on historical person's diverse high fidelity\nclothing colorization based on fine grained semantic understanding and prior.\nColorization of historical persons is realistic and practical, however,\nexisting methods do not perform well in the regards. In this paper, a\nHistoryNet including three parts, namely, classification, fine grained semantic\nparsing and colorization, is proposed. Classification sub-module supplies\nclassifying of images according to the eras, nationalities and garment types;\nParsing sub-network supplies the semantic for person contours, clothing and\nbackground in the image to achieve more accurate colorization of clothes and\npersons and prevent color overflow. In the training process, we integrate\nclassification and semantic parsing features into the coloring generation\nnetwork to improve colorization. Through the design of classification and\nparsing subnetwork, the accuracy of image colorization can be improved and the\nboundary of each part of image can be more clearly. Moreover, we also propose a\nnovel Modern Historical Movies Dataset (MHMD) containing 1,353,166 images and\n42 labels of eras, nationalities, and garment types for automatic colorization\nfrom 147 historical movies or TV series made in modern time. Various\nquantitative and qualitative comparisons demonstrate that our method\noutperforms the state-of-the-art colorization methods, especially on military\nuniforms, which has correct colors according to the historical literatures."", 'For semantic segmentation, most existing real-time deep models trained with\neach frame independently may produce inconsistent results for a video sequence.\nAdvanced methods take into considerations the correlations in the video\nsequence, e.g., by propagating the results to the neighboring frames using\noptical flow, or extracting the frame representations with other frames, which\nmay lead to inaccurate results or unbalanced latency. In this work, we process\nefficient semantic video segmentation in a per-frame fashion during the\ninference process. Different from previous per-frame models, we explicitly\nconsider the temporal consistency among frames as extra constraints during the\ntraining process and embed the temporal consistency into the segmentation\nnetwork. Therefore, in the inference process, we can process each frame\nindependently with no latency, and improve the temporal consistency with no\nextra computational cost and post-processing. We employ compact models for\nreal-time execution. To narrow the performance gap between compact models and\nlarge models, new knowledge distillation methods are designed. Our results\noutperform previous keyframe based methods with a better trade-off between the\naccuracy and the inference speed on popular benchmarks, including the\nCityscapes and Camvid. The temporal consistency is also improved compared with\ncorresponding baselines which are trained with each frame independently. Code\nis available at: https://tinyurl.com/segment-video', 'We propose the first deep learning approach for exemplar-based local\ncolorization. Given a reference color image, our convolutional neural network\ndirectly maps a grayscale image to an output colorized image. Rather than using\nhand-crafted rules as in traditional exemplar-based methods, our end-to-end\ncolorization network learns how to select, propagate, and predict colors from\nthe large-scale data. The approach performs robustly and generalizes well even\nwhen using reference images that are unrelated to the input grayscale image.\nMore importantly, as opposed to other learning-based colorization methods, our\nnetwork allows the user to achieve customizable results by simply feeding\ndifferent references. In order to further reduce manual effort in selecting the\nreferences, the system automatically recommends references with our proposed\nimage retrieval algorithm, which considers both semantic and luminance\ninformation. The colorization can be performed fully automatically by simply\npicking the top reference suggestion. Our approach is validated through a user\nstudy and favorable quantitative comparisons to the-state-of-the-art methods.\nFurthermore, our approach can be naturally extended to video colorization. Our\ncode and models will be freely available for public use.', 'Compared to color images captured by conventional RGB cameras, monochrome\nimages usually have better signal-to-noise ratio (SNR) and richer textures due\nto its higher quantum efficiency. It is thus natural to apply a mono-color\ndual-camera system to restore color images with higher visual quality. In this\npaper, we propose a mono-color image enhancement algorithm that colorizes the\nmonochrome image with the color one. Based on the assumption that adjacent\nstructures with similar luminance values are likely to have similar colors, we\nfirst perform dense scribbling to assign colors to the monochrome pixels\nthrough block matching. Two types of outliers, including occlusion and color\nambiguity, are detected and removed from the initial scribbles. We also\nintroduce a sampling strategy to accelerate the scribbling process. Then, the\ndense scribbles are propagated to the entire image. To alleviate incorrect\ncolor propagation in the regions that have no color hints at all, we generate\nextra color seeds based on the existed scribbles to guide the propagation\nprocess. Experimental results show that, our algorithm can efficiently restore\ncolor images with higher SNR and richer details from the mono-color image\npairs, and achieves good performance in solving the color bleeding problem.', 'A novel color image enhancement method is proposed based on Retinex to\nenhance color images under non-uniform illumination or poor visibility\nconditions. Different from the conventional Retinex algorithms, the Weighted\nGuided Image Filter is used as a surround function instead of the Gaussian\nfilter to estimate the background illumination, which can overcome the\ndrawbacks of local blur and halo artifact that may appear by Gaussian filter.\nTo avoid color distortion, the image is converted to the HSI color model, and\nonly the intensity channel is enhanced. Then a linear color restoration\nalgorithm is adopted to convert the enhanced intensity image back to the RGB\ncolor model, which ensures the hue is constant and undistorted. Experimental\nresults show that the proposed method is effective to enhance both color and\ngray images with low exposure and non-uniform illumination, resulting in better\nvisual quality than traditional method. At the same time, the objective\nevaluation indicators are also superior to the conventional methods. In\naddition, the efficiency of the proposed method is also improved thanks to the\nlinear color restoration algorithm.', 'Greyscale image colorization for applications in image restoration has seen\nsignificant improvements in recent years. Many of these techniques that use\nlearning-based methods struggle to effectively colorize sparse inputs. With the\nconsistent growth of the anime industry, the ability to colorize sparse input\nsuch as line art can reduce significant cost and redundant work for production\nstudios by eliminating the in-between frame colorization process. Simply using\nexisting methods yields inconsistent colors between related frames resulting in\na flicker effect in the final video. In order to successfully automate key\nareas of large-scale anime production, the colorization of line arts must be\ntemporally consistent between frames. This paper proposes a method to colorize\nline art frames in an adversarial setting, to create temporally coherent video\nof large anime by improving existing image to image translation methods. We\nshow that by adding an extra condition to the generator and discriminator, we\ncan effectively create temporally consistent video sequences from anime line\narts. Code and models available at: https://github.com/Harry-Thasarathan/TCVC', 'Although recent inpainting approaches have demonstrated significant\nimprovements with deep neural networks, they still suffer from artifacts such\nas blunt structures and abrupt colors when filling in the missing regions. To\naddress these issues, we propose an external-internal inpainting scheme with a\nmonochromic bottleneck that helps image inpainting models remove these\nartifacts. In the external learning stage, we reconstruct missing structures\nand details in the monochromic space to reduce the learning dimension. In the\ninternal learning stage, we propose a novel internal color propagation method\nwith progressive learning strategies for consistent color restoration.\nExtensive experiments demonstrate that our proposed scheme helps image\ninpainting models produce more structure-preserved and visually compelling\nresults.', ""Color transfer is an image editing process that adjusts the colors of a\npicture to match a target picture's color theme. A natural color transfer not\nonly matches the color styles but also prevents after-transfer artifacts due to\nimage compression, noise, and gradient smoothness change. The recently\ndiscovered color homography theorem proves that colors across a change in\nphotometric viewing condition are related by a homography. In this paper, we\npropose a color-homography-based color transfer decomposition which encodes\ncolor transfer as a combination of chromaticity shift and shading adjustment. A\npowerful form of shading adjustment is shown to be a global shading curve by\nwhich the same shading homography can be applied elsewhere. Our experiments\nshow that the proposed color transfer decomposition provides a very close\napproximation to many popular color transfer methods. The advantage of our\napproach is that the learned color transfer can be applied to many other images\n(e.g. other frames in a video), instead of a frame-to-frame basis. We\ndemonstrate two applications for color transfer enhancement and video color\ngrading re-application. This simple model of color transfer is also important\nfor future color transfer algorithm design."", 'We consider the reconstruction problem of video snapshot compressive imaging\n(SCI), which captures high-speed videos using a low-speed 2D sensor (detector).\nThe underlying principle of SCI is to modulate sequential high-speed frames\nwith different masks and then these encoded frames are integrated into a\nsnapshot on the sensor and thus the sensor can be of low-speed. On one hand,\nvideo SCI enjoys the advantages of low-bandwidth, low-power and low-cost. On\nthe other hand, applying SCI to large-scale problems (HD or UHD videos) in our\ndaily life is still challenging and one of the bottlenecks lies in the\nreconstruction algorithm. Exiting algorithms are either too slow (iterative\noptimization algorithms) or not flexible to the encoding process (deep learning\nbased end-to-end networks). In this paper, we develop fast and flexible\nalgorithms for SCI based on the plug-and-play (PnP) framework. In addition to\nthe PnP-ADMM method, we further propose the PnP-GAP (generalized alternating\nprojection) algorithm with a lower computational workload. We first employ the\nimage deep denoising priors to show that PnP can recover a UHD color video with\n30 frames from a snapshot measurement. Since videos have strong temporal\ncorrelation, by employing the video deep denoising priors, we achieve a\nsignificant improvement in the results. Furthermore, we extend the proposed PnP\nalgorithms to the color SCI system using mosaic sensors, where each pixel only\ncaptures the red, green or blue channels. A joint reconstruction and\ndemosaicing paradigm is developed for flexible and high quality reconstruction\nof color video SCI systems. Extensive results on both simulation and real\ndatasets verify the superiority of our proposed algorithm.', 'We propose a deep learning approach for user-guided image colorization. The\nsystem directly maps a grayscale image, along with sparse, local user ""hints""\nto an output colorization with a Convolutional Neural Network (CNN). Rather\nthan using hand-defined rules, the network propagates user edits by fusing\nlow-level cues along with high-level semantic information, learned from\nlarge-scale data. We train on a million images, with simulated user inputs. To\nguide the user towards efficient input selection, the system recommends likely\ncolors based on the input image and current user inputs. The colorization is\nperformed in a single feed-forward pass, enabling real-time use. Even with\nrandomly simulated user inputs, we show that the proposed system helps novice\nusers quickly create realistic colorizations, and offers large improvements in\ncolorization quality with just a minute of use. In addition, we demonstrate\nthat the framework can incorporate other user ""hints"" to the desired\ncolorization, showing an application to color histogram transfer. Our code and\nmodels are available at https://richzhang.github.io/ideepcolor.', 'Video semantic segmentation is active in recent years benefited from the\ngreat progress of image semantic segmentation. For such a task, the per-frame\nimage segmentation is generally unacceptable in practice due to high\ncomputation cost. To tackle this issue, many works use the flow-based feature\npropagation to reuse the features of previous frames. However, the optical flow\nestimation inevitably suffers inaccuracy and then causes the propagated\nfeatures distorted. In this paper, we propose distortion-aware feature\ncorrection to alleviate the issue, which improves video segmentation\nperformance by correcting distorted propagated features. To be specific, we\nfirstly propose to transfer distortion patterns from feature into image space\nand conduct effective distortion map prediction. Benefited from the guidance of\ndistortion maps, we proposed Feature Correction Module (FCM) to rectify\npropagated features in the distorted areas. Our proposed method can\nsignificantly boost the accuracy of video semantic segmentation at a low price.\nThe extensive experimental results on Cityscapes and CamVid show that our\nmethod outperforms the recent state-of-the-art methods.', 'Video object segmentation, aiming to segment the foreground objects given the\nannotation of the first frame, has been attracting increasing attentions. Many\nstate-of-the-art approaches have achieved great performance by relying on\nonline model updating or mask-propagation techniques. However, most online\nmodels require high computational cost due to model fine-tuning during\ninference. Most mask-propagation based models are faster but with relatively\nlow performance due to failure to adapt to object appearance variation. In this\npaper, we are aiming to design a new model to make a good balance between speed\nand performance. We propose a model, called NPMCA-net, which directly localizes\nforeground objects based on mask-propagation and non-local technique by\nmatching pixels in reference and target frames. Since we bring in information\nof both first and previous frames, our network is robust to large object\nappearance variation, and can better adapt to occlusions. Extensive experiments\nshow that our approach can achieve a new state-of-the-art performance with a\nfast speed at the same time (86.5% IoU on DAVIS-2016 and 72.2% IoU on\nDAVIS-2017, with speed of 0.11s per frame) under the same level comparison.\nSource code is available at https://github.com/siyueyu/NPMCA-net.', 'In this work we focus on the problem of colorization for image compression.\nSince color information occupies a large proportion of the total storage size\nof an image, a method that can predict accurate color from its grayscale\nversion can produce dramatic reduction in image file size. But colorization for\ncompression poses several challenges. First, while colorization for artistic\npurposes simply involves predicting plausible chroma, colorization for\ncompression requires generating output colors that are as close as possible to\nthe ground truth. Second, many objects in the real world exhibit multiple\npossible colors. Thus, to disambiguate the colorization problem some additional\ninformation must be stored to reproduce the true colors with good accuracy. To\naccount for the multimodal color distribution of objects we propose a deep\ntree-structured network that generates multiple color hypotheses for every\npixel from a grayscale picture (as opposed to a single color produced by most\nprior colorization approaches). We show how to leverage the multimodal output\nof our model to reproduce with high fidelity the true colors of an image by\nstoring very little additional information. In the experiments we show that our\nproposed method outperforms traditional JPEG color coding by a large margin,\nproducing colors that are nearly indistinguishable from the ground truth at the\nstorage cost of just a few hundred bytes for high-resolution pictures!', 'Error propagation is a general but crucial problem in online semi-supervised\nvideo object segmentation. We aim to suppress error propagation through a\ncorrection mechanism with high reliability. The key insight is to disentangle\nthe correction from the conventional mask propagation process with reliable\ncues. We introduce two modulators, propagation and correction modulators, to\nseparately perform channel-wise re-calibration on the target frame embeddings\naccording to local temporal correlations and reliable references respectively.\nSpecifically, we assemble the modulators with a cascaded propagation-correction\nscheme. This avoids overriding the effects of the reliable correction modulator\nby the propagation modulator. Although the reference frame with the ground\ntruth label provides reliable cues, it could be very different from the target\nframe and introduce uncertain or incomplete correlations. We augment the\nreference cues by supplementing reliable feature patches to a maintained pool,\nthus offering more comprehensive and expressive object representations to the\nmodulators. In addition, a reliability filter is designed to retrieve reliable\npatches and pass them in subsequent frames. Our model achieves state-of-the-art\nperformance on YouTube-VOS18/19 and DAVIS17-Val/Test benchmarks. Extensive\nexperiments demonstrate that the correction mechanism provides considerable\nperformance gain by fully utilizing reliable guidance. Code is available at:\nhttps://github.com/JerryX1110/RPCMVOS.', 'Color and structure are the two pillars that construct an image. Usually, the\nstructure is well expressed through a rich spectrum of colors, allowing objects\nin an image to be recognized by neural networks. However, under extreme\nlimitations of color space, the structure tends to vanish, and thus a neural\nnetwork might fail to understand the image. Interested in exploring this\ninterplay between color and structure, we study the scientific problem of\nidentifying and preserving the most informative image structures while\nconstraining the color space to just a few bits, such that the resulting image\ncan be recognized with possibly high accuracy. To this end, we propose a color\nquantization network, ColorCNN, which learns to structure the images from the\nclassification loss in an end-to-end manner. Given a color space size, ColorCNN\nquantizes colors in the original image by generating a color index map and an\nRGB color palette. Then, this color-quantized image is fed to a pre-trained\ntask network to evaluate its performance. In our experiment, with only a 1-bit\ncolor space (i.e., two colors), the proposed network achieves 82.1% top-1\naccuracy on the CIFAR10 dataset, outperforming traditional color quantization\nmethods by a large margin. For applications, when encoded with PNG, the\nproposed color quantization shows superiority over other image compression\nmethods in the extremely low bit-rate regime. The code is available at:\nhttps://github.com/hou-yz/color_distillation.', ""This thesis presents methods and approaches to image color correction, color\nenhancement, and color editing. To begin, we study the color correction problem\nfrom the standpoint of the camera's image signal processor (ISP). A camera's\nISP is hardware that applies a series of in-camera image processing and color\nmanipulation steps, many of which are nonlinear in nature, to render the\ninitial sensor image to its final photo-finished representation saved in the\n8-bit standard RGB (sRGB) color space. As white balance (WB) is one of the\nmajor procedures applied by the ISP for color correction, this thesis presents\ntwo different methods for ISP white balancing. Afterward, we discuss another\nscenario of correcting and editing image colors, where we present a set of\nmethods to correct and edit WB settings for images that have been improperly\nwhite-balanced by the ISP. Then, we explore another factor that has a\nsignificant impact on the quality of camera-rendered colors, in which we\noutline two different methods to correct exposure errors in camera-rendered\nimages. Lastly, we discuss post-capture auto color editing and manipulation. In\nparticular, we propose auto image recoloring methods to generate different\nrealistic versions of the same camera-rendered image with new colors. Through\nextensive evaluations, we demonstrate that our methods provide superior\nsolutions compared to existing alternatives targeting color correction, color\nenhancement, and color editing."", ""Event cameras are novel, bio-inspired visual sensors, whose pixels output\nasynchronous and independent timestamped spikes at local intensity changes,\ncalled 'events'. Event cameras offer advantages over conventional frame-based\ncameras in terms of latency, high dynamic range (HDR) and temporal resolution.\nUntil recently, event cameras have been limited to outputting events in the\nintensity channel, however, recent advances have resulted in the development of\ncolor event cameras, such as the Color-DAVIS346. In this work, we present and\nrelease the first Color Event Camera Dataset (CED), containing 50 minutes of\nfootage with both color frames and events. CED features a wide variety of\nindoor and outdoor scenes, which we hope will help drive forward event-based\nvision research. We also present an extension of the event camera simulator\nESIM that enables simulation of color events. Finally, we present an evaluation\nof three state-of-the-art image reconstruction methods that can be used to\nconvert the Color-DAVIS346 into a continuous-time, HDR, color video camera to\nvisualise the event stream, and for use in downstream vision applications."", ""Lossy compression brings artifacts into the compressed image and degrades the\nvisual quality. In recent years, many compression artifacts removal methods\nbased on convolutional neural network (CNN) have been developed with great\nsuccess. However, these methods usually train a model based on one specific\nvalue or a small range of quality factors. Obviously, if the test image's\nquality factor does not match to the assumed value range, then degraded\nperformance will be resulted. With this motivation and further consideration of\npractical usage, a highly robust compression artifacts removal network is\nproposed in this paper. Our proposed network is a single model approach that\ncan be trained for handling a wide range of quality factors while consistently\ndelivering superior or comparable image artifacts removal performance. To\ndemonstrate, we focus on the JPEG compression with quality factors, ranging\nfrom 1 to 60. Note that a turnkey success of our proposed network lies in the\nnovel utilization of the quantization tables as part of the training data.\nFurthermore, it has two branches in parallel---i.e., the restoration branch and\nthe global branch. The former effectively removes the local artifacts, such as\nringing artifacts removal. On the other hand, the latter extracts the global\nfeatures of the entire image that provides highly instrumental image quality\nimprovement, especially effective on dealing with the global artifacts, such as\nblocking, color shifting. Extensive experimental results performed on color and\ngrayscale images have clearly demonstrated the effectiveness and efficacy of\nour proposed single-model approach on the removal of compression artifacts from\nthe decoded image."", 'This paper addresses fast semantic segmentation on video.Video segmentation\noften calls for real-time, or even fasterthan real-time, processing. One common\nrecipe for conserving computation arising from feature extraction is to\npropagate features of few selected keyframes. However, recent advances in fast\nimage segmentation make these solutions less attractive. To leverage fast image\nsegmentation for furthering video segmentation, we propose a simple yet\nefficient propagation framework. Specifically, we perform lightweight flow\nestimation in 1/8-downscaled image space for temporal warping in segmentation\noutpace space. Moreover, we introduce a guided spatially-varying convolution\nfor fusing segmentations derived from the previous and current frames, to\nmitigate propagation error and enable lightweight feature extraction on\nnon-keyframes. Experimental results on Cityscapes and CamVid show that our\nscheme achieves the state-of-the-art accuracy-throughput trade-off on video\nsegmentation.', 'Near-infrared imaging can capture haze-free near-infrared gray images and\nvisible color images, according to physical scattering models, e.g., Rayleigh\nor Mie models. However, there exist serious discrepancies in brightness and\nimage structures between the near-infrared gray images and the visible color\nimages. The direct use of the near-infrared gray images brings about another\ncolor distortion problem in the dehazed images. Therefore, the color distortion\nshould also be considered for near-infrared dehazing. To reflect this point,\nthis paper presents an approach of adding a new color regularization to\nconventional dehazing framework. The proposed color regularization can model\nthe color prior for unknown haze-free images from two captured images. Thus,\nnatural-looking colors and fine details can be induced on the dehazed images.\nThe experimental results show that the proposed color regularization model can\nhelp remove the color distortion and the haze at the same time. Also, the\neffectiveness of the proposed color regularization is verified by comparing\nwith other conventional regularizations. It is also shown that the proposed\ncolor regularization can remove the edge artifacts which arise from the use of\nthe conventional dark prior model.', 'Over the last decade, the process of automatic image colorization has been of\nsignificant interest for several application areas including restoration of\naged or degraded images. This problem is highly ill-posed due to the large\ndegrees of freedom during the assignment of color information. Many of the\nrecent developments in automatic colorization involve images that contain a\ncommon theme or require highly processed data such as semantic maps as input.\nIn our approach, we attempt to fully generalize the colorization procedure\nusing a conditional Deep Convolutional Generative Adversarial Network (DCGAN),\nextend current methods to high-resolution images and suggest training\nstrategies that speed up the process and greatly stabilize it. The network is\ntrained over datasets that are publicly available such as CIFAR-10 and\nPlaces365. The results of the generative model and traditional deep neural\nnetworks are compared.', 'Photorealistic image stylization concerns transferring style of a reference\nphoto to a content photo with the constraint that the stylized photo should\nremain photorealistic. While several photorealistic image stylization methods\nexist, they tend to generate spatially inconsistent stylizations with\nnoticeable artifacts. In this paper, we propose a method to address these\nissues. The proposed method consists of a stylization step and a smoothing\nstep. While the stylization step transfers the style of the reference photo to\nthe content photo, the smoothing step ensures spatially consistent\nstylizations. Each of the steps has a closed-form solution and can be computed\nefficiently. We conduct extensive experimental validations. The results show\nthat the proposed method generates photorealistic stylization outputs that are\nmore preferred by human subjects as compared to those by the competing methods\nwhile running much faster. Source code and additional results are available at\nhttps://github.com/NVIDIA/FastPhotoStyle .', 'Recently, some correlation filter based trackers with detection proposals\nhave achieved state-of-the-art tracking results. However, a large number of\nredundant proposals given by the proposal generator may degrade the performance\nand speed of these trackers. In this paper, we propose an adaptive proposal\nselection algorithm which can generate a small number of high-quality proposals\nto handle the problem of scale variations for visual object tracking.\nSpecifically, we firstly utilize the color histograms in the HSV color space to\nrepresent the instances (i.e., the initial target in the first frame and the\npredicted target in the previous frame) and proposals. Then, an adaptive\nstrategy based on the color similarity is formulated to select high-quality\nproposals. We further integrate the proposed adaptive proposal selection\nalgorithm with coarse-to-fine deep features to validate the generalization and\nefficiency of the proposed tracker. Experiments on two benchmark datasets\ndemonstrate that the proposed algorithm performs favorably against several\nstate-of-the-art trackers.', 'The ability to characterize the color content of natural imagery is an\nimportant application of image processing. The pixel by pixel coloring of\nimages may be viewed naturally as points in color space, and the inherent\nstructure and distribution of these points affords a quantization, through\nclustering, of the color information in the image. In this paper, we present a\nnovel topologically driven clustering algorithm that permits segmentation of\nthe color features in a digital image. The algorithm blends Locally Linear\nEmbedding (LLE) and vector quantization by mapping color information to a lower\ndimensional space, identifying distinct color regions, and classifying pixels\ntogether based on both a proximity measure and color content. It is observed\nthat these techniques permit a significant reduction in color resolution while\nmaintaining the visually important features of images.', 'While many image colorization algorithms have recently shown the capability\nof producing plausible color versions from gray-scale photographs, they still\nsuffer from limited semantic understanding. To address this shortcoming, we\npropose to exploit pixelated object semantics to guide image colorization. The\nrationale is that human beings perceive and distinguish colors based on the\nsemantic categories of objects. Starting from an autoregressive model, we\ngenerate image color distributions, from which diverse colored results are\nsampled. We propose two ways to incorporate object semantics into the\ncolorization model: through a pixelated semantic embedding and a pixelated\nsemantic generator. Specifically, the proposed convolutional neural network\nincludes two branches. One branch learns what the object is, while the other\nbranch learns the object colors. The network jointly optimizes a color\nembedding loss, a semantic segmentation loss and a color generation loss, in an\nend-to-end fashion. Experiments on PASCAL VOC2012 and COCO-stuff reveal that\nour network, when trained with semantic segmentation labels, produces more\nrealistic and finer results compared to the colorization state-of-the-art.', 'Image denoising and demosaicking are the most important early stages in\ndigital camera pipelines. They constitute a severely ill-posed problem that\naims at reconstructing a full color image from a noisy color filter array (CFA)\nimage. In most of the literature, denoising and demosaicking are treated as two\nindependent problems, without considering their interaction, or asking which\nshould be applied first. Several recent works have started addressing them\njointly in works that involve heavy weight CNNs, thus incompatible with low\npower portable imaging devices. Hence, the question of how to combine denoising\nand demosaicking to reconstruct full color images remains very relevant: Is\ndenoising to be applied first, or should that be demosaicking first? In this\npaper, we review the main variants of these strategies and carry-out an\nextensive evaluation to find the best way to reconstruct full color images from\na noisy mosaic. We conclude that demosaicking should applied first, followed by\ndenoising. Yet we prove that this requires an adaptation of classic denoising\nalgorithms to demosaicked noise, which we justify and specify.', 'Nowadays modern displays are capable to render video content with high\ndynamic range (HDR) and wide color gamut (WCG). However, most available\nresources are still in standard dynamic range (SDR). Therefore, there is an\nurgent demand to transform existing SDR-TV contents into their HDR-TV versions.\nIn this paper, we conduct an analysis of SDRTV-to-HDRTV task by modeling the\nformation of SDRTV/HDRTV content. Base on the analysis, we propose a three-step\nsolution pipeline including adaptive global color mapping, local enhancement\nand highlight generation. Moreover, the above analysis inspires us to present a\nlightweight network that utilizes global statistics as guidance to conduct\nimage-adaptive color mapping. In addition, we construct a dataset using HDR\nvideos in HDR10 standard, named HDRTV1K, and select five metrics to evaluate\nthe results of SDRTV-to-HDRTV algorithms. Furthermore, our final results\nachieve state-of-the-art performance in quantitative comparisons and visual\nquality. The code and dataset are available at\nhttps://github.com/chxy95/HDRTVNet.', 'Recent years have seen remarkable progress in semantic segmentation. Yet, it\nremains a challenging task to apply segmentation techniques to video-based\napplications. Specifically, the high throughput of video streams, the sheer\ncost of running fully convolutional networks, together with the low-latency\nrequirements in many real-world applications, e.g. autonomous driving, present\na significant challenge to the design of the video segmentation framework. To\ntackle this combined challenge, we develop a framework for video semantic\nsegmentation, which incorporates two novel components: (1) a feature\npropagation module that adaptively fuses features over time via spatially\nvariant convolution, thus reducing the cost of per-frame computation; and (2)\nan adaptive scheduler that dynamically allocate computation based on accuracy\nprediction. Both components work together to ensure low latency while\nmaintaining high segmentation quality. On both Cityscapes and CamVid, the\nproposed framework obtained competitive performance compared to the state of\nthe art, while substantially reducing the latency, from 360 ms to 119 ms.', 'Automatic document content processing is affected by artifacts caused by the\nshape of the paper, non-uniform and diverse color of lighting conditions.\nFully-supervised methods on real data are impossible due to the large amount of\ndata needed. Hence, the current state of the art deep learning models are\ntrained on fully or partially synthetic images. However, document shadow or\nshading removal results still suffer because: (a) prior methods rely on\nuniformity of local color statistics, which limit their application on\nreal-scenarios with complex document shapes and textures and; (b) synthetic or\nhybrid datasets with non-realistic, simulated lighting conditions are used to\ntrain the models. In this paper we tackle these problems with our two main\ncontributions. First, a physically constrained learning-based method that\ndirectly estimates document reflectance based on intrinsic image formation\nwhich generalizes to challenging illumination conditions. Second, a new dataset\nthat clearly improves previous synthetic ones, by adding a large range of\nrealistic shading and diverse multi-illuminant conditions, uniquely customized\nto deal with documents in-the-wild. The proposed architecture works in a\nself-supervised manner where only the synthetic texture is used as a weak\ntraining signal (obviating the need for very costly ground truth with\ndisentangled versions of shading and reflectance). The proposed approach leads\nto a significant generalization of document reflectance estimation in real\nscenes with challenging illumination. We extensively evaluate on the real\nbenchmark datasets available for intrinsic image decomposition and document\nshadow removal tasks. Our reflectance estimation scheme, when used as a\npre-processing step of an OCR pipeline, shows a 26% improvement of character\nerror rate (CER), thus, proving the practical applicability.', 'Although adversarial samples of deep neural networks (DNNs) have been\nintensively studied on static images, their extensions in videos are never\nexplored. Compared with images, attacking a video needs to consider not only\nspatial cues but also temporal cues. Moreover, to improve the imperceptibility\nas well as reduce the computation cost, perturbations should be added on as\nfewer frames as possible, i.e., adversarial perturbations are temporally\nsparse. This further motivates the propagation of perturbations, which denotes\nthat perturbations added on the current frame can transfer to the next frames\nvia their temporal interactions. Thus, no (or few) extra perturbations are\nneeded for these frames to misclassify them. To this end, we propose an\nl2,1-norm based optimization algorithm to compute the sparse adversarial\nperturbations for videos. We choose the action recognition as the targeted\ntask, and networks with a CNN+RNN architecture as threat models to verify our\nmethod. Thanks to the propagation, we can compute perturbations on a shortened\nversion video, and then adapt them to the long version video to fool DNNs.\nExperimental results on the UCF101 dataset demonstrate that even only one frame\nin a video is perturbed, the fooling rate can still reach 59.7%.', 'We propose a new algorithm for color transfer between images that have\nperceptually similar semantic structures. We aim to achieve a more accurate\ncolor transfer that leverages semantically-meaningful dense correspondence\nbetween images. To accomplish this, our algorithm uses neural representations\nfor matching. Additionally, the color transfer should be spatially variant and\nglobally coherent. Therefore, our algorithm optimizes a local linear model for\ncolor transfer satisfying both local and global constraints. Our proposed\napproach jointly optimizes matching and color transfer, adopting a\ncoarse-to-fine strategy. The proposed method can be successfully extended from\none-to-one to one-to-many color transfer. The latter further addresses the\nproblem of mismatching elements of the input image. We validate our proposed\nmethod by testing it on a large variety of image content.', 'One of the important requirements in image retrieval, indexing,\nclassification, clustering and etc. is extracting efficient features from\nimages. The color feature is one of the most widely used visual features. Use\nof color histogram is the most common way for representing color feature. One\nof disadvantage of the color histogram is that it does not take the color\nspatial distribution into consideration. In this paper dynamic color\ndistribution entropy of neighborhoods method based on color distribution\nentropy is presented, which effectively describes the spatial information of\ncolors. The image retrieval results in compare to improved color distribution\nentropy show the acceptable efficiency of this approach.', 'We propose a novel framework to produce cartoon videos by fetching the color\ninformation from two input keyframes while following the animated motion guided\nby a user sketch. The key idea of the proposed approach is to estimate the\ndense cross-domain correspondence between the sketch and cartoon video frames,\nand employ a blending module with occlusion estimation to synthesize the middle\nframe guided by the sketch. After that, the input frames and the synthetic\nframe equipped with established correspondence are fed into an arbitrary-time\nframe interpolation pipeline to generate and refine additional inbetween\nframes. Finally, a module to preserve temporal consistency is employed.\nCompared to common frame interpolation methods, our approach can address frames\nwith relatively large motion and also has the flexibility to enable users to\ncontrol the generated video sequences by editing the sketch guidance. By\nexplicitly considering the correspondence between frames and the sketch, we can\nachieve higher quality results than other image synthesis methods. Our results\nshow that our system generalizes well to different movie frames, achieving\nbetter results than existing solutions.', 'Evaluation of automatic video summaries is a challenging problem. In the past\nyears, some evaluation methods are presented that utilize only a single feature\nlike color feature to detect similarity between automatic video summaries and\nground-truth user summaries. One of the drawbacks of using a single feature is\nthat sometimes it gives a false similarity detection which makes the assessment\nof the quality of the generated video summary less perceptual and not accurate.\nIn this paper, a novel method for evaluating automatic video summaries is\npresented. This method is based on comparing automatic video summaries\ngenerated by video summarization techniques with ground-truth user summaries.\nThe objective of this evaluation method is to quantify the quality of video\nsummaries, and allow comparing different video summarization techniques\nutilizing both color and texture features of the video frames and using the\nBhattacharya distance as a dissimilarity measure due to its advantages. Our\nExperiments show that the proposed evaluation method overcomes the drawbacks of\nother methods and gives a more perceptual evaluation of the quality of the\nautomatic video summaries.', 'Matching people across multiple camera views known as person\nre-identification, is a challenging problem due to the change in visual\nappearance caused by varying lighting conditions. The perceived color of the\nsubject appears to be different with respect to illumination. Previous works\nuse color as it is or address these challenges by designing color spaces\nfocusing on a specific cue. In this paper, we propose a data driven approach\nfor learning color patterns from pixels sampled from images across two camera\nviews. The intuition behind this work is that, even though pixel values of same\ncolor would be different across views, they should be encoded with the same\nvalues. We model color feature generation as a learning problem by jointly\nlearning a linear transformation and a dictionary to encode pixel values. We\nalso analyze different photometric invariant color spaces. Using color as the\nonly cue, we compare our approach with all the photometric invariant color\nspaces and show superior performance over all of them. Combining with other\nlearned low-level and high-level features, we obtain promising results in\nViPER, Person Re-ID 2011 and CAVIAR4REID datasets.', 'Object detection in videos has drawn increasing attention since it is more\npractical in real scenarios. Most of the deep learning methods use CNNs to\nprocess each decoded frame in a video stream individually. However, the free of\ncharge yet valuable motion information already embedded in the video\ncompression format is usually overlooked. In this paper, we propose a fast\nobject detection method by taking advantage of this with a novel Motion aided\nMemory Network (MMNet). The MMNet has two major advantages: 1) It significantly\naccelerates the procedure of feature extraction for compressed videos. It only\nneed to run a complete recognition network for I-frames, i.e. a few reference\nframes in a video, and it produces the features for the following P frames\n(predictive frames) with a light weight memory network, which runs fast; 2)\nUnlike existing methods that establish an additional network to model motion of\nframes, we take full advantage of both motion vectors and residual errors that\nare freely available in video streams. To our best knowledge, the MMNet is the\nfirst work that investigates a deep convolutional detector on compressed\nvideos. Our method is evaluated on the large-scale ImageNet VID dataset, and\nthe results show that it is 3x times faster than single image detector R-FCN\nand 10x times faster than high-performance detector MANet at a minor accuracy\nloss.', 'Semantic video segmentation is challenging due to the sheer amount of data\nthat needs to be processed and labeled in order to construct accurate models.\nIn this paper we present a deep, end-to-end trainable methodology to video\nsegmentation that is capable of leveraging information present in unlabeled\ndata in order to improve semantic estimates. Our model combines a convolutional\narchitecture and a spatio-temporal transformer recurrent layer that are able to\ntemporally propagate labeling information by means of optical flow, adaptively\ngated based on its locally estimated uncertainty. The flow, the recognition and\nthe gated temporal propagation modules can be trained jointly, end-to-end. The\ntemporal, gated recurrent flow propagation component of our model can be\nplugged into any static semantic segmentation architecture and turn it into a\nweakly supervised video processing one. Our extensive experiments in the\nchallenging CityScapes and Camvid datasets, and based on multiple deep\narchitectures, indicate that the resulting model can leverage unlabeled\ntemporal frames, next to a labeled one, in order to improve both the video\nsegmentation accuracy and the consistency of its temporal labeling, at no\nadditional annotation cost and with little extra computation.', ""Several researches in leaf identification did not include color information\nas features. The main reason is caused by a fact that they used green colored\nleaves as samples. However, for foliage plants, plants with colorful leaves,\nfancy patterns in their leaves, and interesting plants with unique shape, color\nand also texture could not be neglected. For example, Epipremnum pinnatum\n'Aureum' and Epipremnum pinnatum 'Marble Queen' have similar patterns, same\nshape, but different colors. Combination of shape, color, texture features, and\nother attribute contained on the leaf is very useful in leaf identification. In\nthis research, Polar Fourier Transform and three kinds of geometric features\nwere used to represent shape features, color moments that consist of mean,\nstandard deviation, skewness were used to represent color features, texture\nfeatures are extracted from GLCMs, and vein features were added to improve\nperformance of the identification system. The identification system uses\nProbabilistic Neural Network (PNN) as a classifier. The result shows that the\nsystem gives average accuracy of 93.0833% for 60 kinds of foliage plants."", 'By assuming that the most frequently occuring color in a video or a region of\na video I propose a new algorithm for detecting foreground objects in a video.\nThe process of detecting the foreground objects is complicated because of the\nfact that there may be swaying trees, objects of the background being moved\naround or lighting changes in the video. To deal with such complexities many\nhave come up with solutions which heavily rely on expensive floating point\noperations. In this paper I used a data structure called Octree which is\nimplemented only using binary operations. Traditionally octrees were used for\ncolor quantization but here in this paper I used it as a data structure to\nstore the most frequently occuring colors in a video as well. For each of the\nstarting few video frames, I constructed a Octree using all the colors of that\nframe. Next I pruned all the trees by removing nodes below a certain height and\ngave the leaf nodes a color which is dependant on the topological path from\nthat node to its parent. Hence any two leaf nodes in two different octrees with\nthe same topological path from themselves to the root will represent the same\ncolor. Next I merged all these individual trees into a single tree retaining\nonly those nodes whose topological path to itself from the root is most common\namong all the trees. The colors represented by the leaf nodes in the resultant\ntree will be the most frequently occuring colors in the starting video frames\nof the video. Hence any color of an incomming frame that is not close to any of\nthe colors represented by the leaf node of the merged tree can be regarded as\nbelonging to a foreground object.\n  As an Octree is constructed using only binary operations, it is very fast\ncompared to other leading algorithms.', 'Semi-supervised video object segmentation aims to separate a target object\nfrom a video sequence, given the mask in the first frame. Most of current\nprevailing methods utilize information from additional modules trained in other\ndomains like optical flow and instance segmentation, and as a result they do\nnot compete with other methods on common ground. To address this issue, we\npropose a simple yet strong transductive method, in which additional modules,\ndatasets, and dedicated architectural designs are not needed. Our method takes\na label propagation approach where pixel labels are passed forward based on\nfeature similarity in an embedding space. Different from other propagation\nmethods, ours diffuses temporal information in a holistic manner which take\naccounts of long-term object appearance. In addition, our method requires few\nadditional computational overhead, and runs at a fast $\\sim$37 fps speed. Our\nsingle model with a vanilla ResNet50 backbone achieves an overall score of 72.3\non the DAVIS 2017 validation set and 63.1 on the test set. This simple yet high\nperforming and efficient method can serve as a solid baseline that facilitates\nfuture research. Code and models are available at\n\\url{https://github.com/microsoft/transductive-vos.pytorch}.', 'This paper proposes an automatic image correction method for portrait\nphotographs, which promotes consistency of facial skin color by suppressing\nskin color changes due to background colors. In portrait photographs, skin\ncolor is often distorted due to the lighting environment (e.g., light reflected\nfrom a colored background wall and over-exposure by a camera strobe), and if\nthe photo is artificially combined with another background color, this color\nchange is emphasized, resulting in an unnatural synthesized result. In our\nframework, after roughly extracting the face region and rectifying the skin\ncolor distribution in a color space, we perform color and brightness correction\naround the face in the original image to achieve a proper color balance of the\nfacial image, which is not affected by luminance and background colors. Unlike\nconventional algorithms for color correction, our final result is attained by a\ncolor correction process with a guide image. In particular, our guided image\nfiltering for the color correction does not require a perfectly-aligned guide\nimage required in the original guide image filtering method proposed by He et\nal. Experimental results show that our method generates more natural results\nthan conventional methods on not only headshot photographs but also natural\nscene photographs. We also show automatic yearbook style photo generation as an\nanother application.', 'We present a novel algorithm for transferring artistic styles of semantically\nmeaningful local regions of an image onto local regions of a target video while\npreserving its photorealism. Local regions may be selected either fully\nautomatically from an image, through using video segmentation algorithms, or\nfrom casual user guidance such as scribbles. Our method, based on a deep neural\nnetwork architecture inspired by recent work in photorealistic style transfer,\nis real-time and works on arbitrary inputs without runtime optimization once\ntrained on a diverse dataset of artistic styles. By augmenting our video\ndataset with noisy semantic labels and jointly optimizing over style, content,\nmask, and temporal losses, our method can cope with a variety of imperfections\nin the input and produce temporally coherent videos without visual artifacts.\nWe demonstrate our method on a variety of style images and target videos,\nincluding the ability to transfer different styles onto multiple objects\nsimultaneously, and smoothly transition between styles in time.', ""In this work we discuss the known algorithms for linear colour segmentation\nbased on a physical approach and propose a new modification of segmentation\nalgorithm. This algorithm is based on a region adjacency graph framework\nwithout a pre-segmentation stage. Proposed edge weight functions are defined\nfrom linear image model with normal noise. The colour space projective\ntransform is introduced as a novel pre-processing technique for better handling\nof shadow and highlight areas. The resulting algorithm is tested on a benchmark\ndataset consisting of the images of 19 natural scenes selected from the\nBarnard's DXC-930 SFU dataset and 12 natural scene images newly published for\ncommon use. The dataset is provided with pixel-by-pixel ground truth colour\nsegmentation for every image. Using this dataset, we show that the proposed\nalgorithm modifications lead to qualitative advantages over other model-based\nsegmentation algorithms, and also show the positive effect of each proposed\nmodification. The source code and datasets for this work are available for free\naccess at http://github.com/visillect/segmentation."", ""In this paper, we proposed three methods to solve color recognition of\nRubik's cube, which includes one offline method and two online methods. Scatter\nbalance \\& extreme learning machine (SB-ELM), a offline method, is proposed to\nillustrate the efficiency of training based method. We also point out the\nconception of color drifting which indicates offline methods are always\nineffectiveness and can not work well in continuous change circumstance. By\ncontrast, dynamic weight label propagation is proposed for labeling blocks\ncolor by known center blocks color of Rubik's cube. Furthermore, weak label\nhierarchic propagation, another online method, is also proposed for unknown all\ncolor information but only utilizes weak label of center block in color\nrecognition. We finally design a Rubik's cube robot and construct a dataset to\nillustrate the efficiency and effectiveness of our online methods and to\nindicate the ineffectiveness of offline method by color drifting in our\ndataset."", 'Decolorization is the process to convert a color image or video to its\ngrayscale version, and it has received great attention in recent years. An\nideal decolorization algorithm should preserve the original color contrast as\nmuch as possible. Meanwhile, it should provide the final decolorized result as\nfast as possible. However, most of the current methods are suffering from\neither unsatisfied color information preservation or high computational cost,\nlimiting their application value. In this paper, a simple but effective\ntechnique is proposed for real-time decolorization. Based on the typical\nrgb2gray() color conversion model, which produces a grayscale image by linearly\ncombining R, G, and B channels, we propose a dominant color hypothesis and a\ncorresponding distance measurement metric to evaluate the quality of grayscale\nconversion. The local optimum scheme provides several ""good"" candidates in a\nconfidence interval, from which the ""best"" result can be extracted.\nExperimental results demonstrate that remarkable simplicity of the proposed\nmethod facilitates the process of high resolution images and videos in\nreal-time using a common CPU.', 'Information about the illuminant color is well contained in both achromatic\nregions and the specular components of highlight regions. In this paper, we\npropose a novel way to achieve color constancy by exploiting such clues. The\nkey to our approach lies in the use of suitably extracted derivative colors,\nwhich are able to compute the illuminant color robustly with kernel density\nestimation. While extracting derivative colors from achromatic regions to\napproximate the illuminant color well is basically straightforward, the success\nof our extraction in highlight regions is attributed to the different rates of\nvariation of the diffuse and specular magnitudes in the dichromatic reflection\nmodel. The proposed approach requires no training phase and is simple to\nimplement. More significantly, it performs quite satisfactorily under\ninter-database parameter settings. Our experiments on three standard databases\ndemonstrate its effectiveness and fine performance in comparison to\nstate-of-the-art methods.', 'In this paper, we investigate the effect of color space selection on\ndetectability and discriminability of colored objects under various conditions.\n20 color spaces from the literature are evaluated on a large dataset of\nsimulated and real images. We measure the suitability of color spaces from two\ndifferent perspectives: detectability and discriminability of various color\ngroups. Through experimental evaluation, we found that there is no single\noptimal color space suitable for all color groups. The color spaces have\ndifferent levels of sensitivity to different color groups and they are useful\ndepending on the color of the sought object. Overall, the best results were\nachieved in both simulated and real images using color spaces C1C2C3, UVW and\nXYZ. In addition, using a simulated environment, we show a practical\napplication of color space selection in the context of top-down control in\nactive visual search. The results indicate that on average color space C1C2C3\nfollowed by HSI and XYZ achieve the best time in searching for objects of\nvarious colors. Here, the right choice of color space can improve time of\nsearch on average by 20%. As part of our contribution, we also introduce a\nlarge dataset of simulated 3D objects', 'Recent state-of-the-art video generation systems employ Generative\nAdversarial Networks (GANs) or Variational Autoencoders (VAEs) to produce novel\nvideos. However, VAE models typically produce blurry outputs when faced with\nsub-optimal conditioning of the input, and GANs are known to be unstable for\nlarge output sizes. In addition, the output videos of these models are\ndifficult to evaluate, partly because the GAN loss function is not an accurate\nmeasure of convergence. In this work, we propose using a state-of-the-art\nneural flow generator called Glow to generate videos conditioned on a textual\nlabel, one frame at a time. Neural flow models are more stable than standard\nGANs, as they only optimize a single cross entropy loss function, which is\nmonotonic and avoids the circular convergence issues of the GAN minimax\nobjective. In addition, we also show how to condition Glow on external context,\nwhile still preserving the invertible nature of each ""flow"" layer. Finally, we\nevaluate the proposed Glow model by calculating cross entropy on a held-out\nvalidation set of videos, in order to compare multiple versions of the proposed\nmodel via an ablation study. We show generated videos and discuss future\nimprovements.', 'Color transfer between images uses the statistics information of image\neffectively. We present a novel approach of local color transfer between images\nbased on the simple statistics and locally linear embedding. A sketching\ninterface is proposed for quickly and easily specifying the color\ncorrespondences between target and source image. The user can specify the\ncorrespondences of local region using scribes, which more accurately transfers\nthe target color to the source image while smoothly preserving the boundaries,\nand exhibits more natural output results. Our algorithm is not restricted to\none-to-one image color transfer and can make use of more than one target images\nto transfer the color in different regions in the source image. Moreover, our\nalgorithm does not require to choose the same color style and image size\nbetween source and target images. We propose the sub-sampling to reduce the\ncomputational load. Comparing with other approaches, our algorithm is much\nbetter in color blending in the input data. Our approach preserves the other\ncolor details in the source image. Various experimental results show that our\napproach specifies the correspondences of local color region in source and\ntarget images. And it expresses the intention of users and generates more\nactual and natural results of visual effect.', 'Research on the detection of AI-generated videos has focused almost\nexclusively on face videos, usually referred to as deepfakes. Manipulations\nlike face swapping, face reenactment and expression manipulation have been the\nsubject of an intense research with the development of a number of efficient\ntools to distinguish artificial videos from genuine ones. Much less attention\nhas been paid to the detection of artificial non-facial videos. Yet, new tools\nfor the generation of such kind of videos are being developed at a fast pace\nand will soon reach the quality level of deepfake videos. The goal of this\npaper is to investigate the detectability of a new kind of AI-generated videos\nframing driving street sequences (here referred to as DeepStreets videos),\nwhich, by their nature, can not be analysed with the same tools used for facial\ndeepfakes. Specifically, we present a simple frame-based detector, achieving\nvery good performance on state-of-the-art DeepStreets videos generated by the\nVid2vid architecture. Noticeably, the detector retains very good performance on\ncompressed videos, even when the compression level used during training does\nnot match that used for the test videos.', 'We consider the cross-modal task of producing color representations for text\nphrases. Motivated by the fact that a significant fraction of user queries on\nan image search engine follow an (attribute, object) structure, we propose a\ngenerative adversarial network that generates color profiles for such bigrams.\nWe design our pipeline to learn composition - the ability to combine seen\nattributes and objects to unseen pairs. We propose a novel dataset curation\npipeline from existing public sources. We describe how a set of phrases of\ninterest can be compiled using a graph propagation technique, and then mapped\nto images. While this dataset is specialized for our investigations on color,\nthe method can be extended to other visual dimensions where composition is of\ninterest. We provide detailed ablation studies that test the behavior of our\nGAN architecture with loss functions from the contrastive learning literature.\nWe show that the generative model achieves lower Frechet Inception Distance\nthan discriminative ones, and therefore predicts color profiles that better\nmatch those from real images. Finally, we demonstrate improved performance in\nimage retrieval and classification, indicating the crucial role that color\nplays in these downstream tasks.', 'Mappings between color spaces are ubiquitous in image processing problems\nsuch as gamut mapping, decolorization, and image optimization for color-blind\npeople. Simple color transformations often result in information loss and\nambiguities (for example, when mapping from RGB to grayscale), and one wishes\nto find an image-specific transformation that would preserve as much as\npossible the structure of the original image in the target color space. In this\npaper, we propose Laplacian colormaps, a generic framework for\nstructure-preserving color transformations between images. We use the image\nLaplacian to capture the structural information, and show that if the color\ntransformation between two images preserves the structure, the respective\nLaplacians have similar eigenvectors, or in other words, are approximately\njointly diagonalizable. Employing the relation between joint diagonalizability\nand commutativity of matrices, we use Laplacians commutativity as a criterion\nof color mapping quality and minimize it w.r.t. the parameters of a color\ntransformation to achieve optimal structure preservation. We show numerous\napplications of our approach, including color-to-gray conversion, gamut\nmapping, multispectral image fusion, and image optimization for color deficient\nviewers.', 'Most of the existing works in video synthesis focus on generating videos\nusing adversarial learning. Despite their success, these methods often require\ninput reference frame or fail to generate diverse videos from the given data\ndistribution, with little to no uniformity in the quality of videos that can be\ngenerated. Different from these methods, we focus on the problem of generating\nvideos from latent noise vectors, without any reference input frames. To this\nend, we develop a novel approach that jointly optimizes the input latent space,\nthe weights of a recurrent neural network and a generator through\nnon-adversarial learning. Optimizing for the input latent space along with the\nnetwork weights allows us to generate videos in a controlled environment, i.e.,\nwe can faithfully generate all videos the model has seen during the learning\nprocess as well as new unseen videos. Extensive experiments on three\nchallenging and diverse datasets well demonstrate that our approach generates\nsuperior quality videos compared to the existing state-of-the-art methods.', 'With the dominance of keyframe-based SLAM in the field of robotics, the\nrelative frame poses between keyframes have typically been sacrificed for a\nfaster algorithm to achieve online applications. However, those approaches can\nbecome insufficient for applications that may require refined poses of all\nframes, not just keyframes which are relatively sparse compared to all input\nframes. This paper proposes a novel algorithm to correct the relative frames\nbetween keyframes after the keyframes have been updated by a back-end\noptimization process. The correction model is derived using conservation of the\nmeasurement constraint between landmarks and the robot pose. The proposed\nalgorithm is designed to be easily integrable to existing keyframe-based SLAM\nsystems while exhibiting robust and accurate performance superior to existing\ninterpolation methods. The algorithm also requires low computational resources\nand hence has a minimal burden on the whole SLAM pipeline. We provide the\nevaluation of the proposed pose correction algorithm in comparison to existing\ninterpolation methods in various vector spaces, and our method has demonstrated\nexcellent accuracy in both KITTI and EuRoC datasets.', 'Portrait photo retouching is a photo retouching task that emphasizes\nhuman-region priority and group-level consistency. The lookup table-based\nmethod achieves promising retouching performance by learning image-adaptive\nweights to combine 3-dimensional lookup tables (3D LUTs) and conducting\npixel-to-pixel color transformation. However, this paradigm ignores the local\ncontext cues and applies the same transformation to portrait pixels and\nbackground pixels when they exhibit the same raw RGB values. In contrast, an\nexpert usually conducts different operations to adjust the color temperatures\nand tones of portrait regions and background regions. This inspires us to model\nlocal context cues to improve the retouching quality explicitly. Firstly, we\nconsider an image patch and predict pixel-adaptive lookup table weights to\nprecisely retouch the center pixel. Secondly, as neighboring pixels exhibit\ndifferent affinities to the center pixel, we estimate a local attention mask to\nmodulate the influence of neighboring pixels. Thirdly, the quality of the local\nattention mask can be further improved by applying supervision, which is based\non the affinity map calculated by the groundtruth portrait mask. As for\ngroup-level consistency, we propose to directly constrain the variance of mean\ncolor components in the Lab space. Extensive experiments on PPR10K dataset\nverify the effectiveness of our method, e.g. on high-resolution photos, the\nPSNR metric receives over 0.5 gains while the group-level consistency metric\nobtains at least 2.1 decreases.', 'In this paper the use of Random Sprays Retinex (RSR) algorithm for global\nillumination estimation is proposed and its feasibility tested. Like other\nalgorithms based on the Retinex model, RSR also provides local illumination\nestimation and brightness adjustment for each pixel and it is faster than other\npath-wise Retinex algorithms. As the assumption of the uniform illumination\nholds in many cases, it should be possible to use the mean of local\nillumination estimations of RSR as a global illumination estimation for images\nwith (assumed) uniform illumination allowing also the accuracy to be easily\nmeasured. Therefore we propose a method for estimating global illumination\nestimation based on local RSR results. To our best knowledge this is the first\ntime that RSR algorithm is used to obtain global illumination estimation. For\nour tests we use a publicly available color constancy image database for\ntesting. The results are presented and discussed and it turns out that the\nproposed method outperforms many existing unsupervised color constancy\nalgorithms. The source code is available at\nhttp://www.fer.unizg.hr/ipg/resources/color_constancy/.', ""A novel method to convert color/multi-spectral images to gray-level images is\nintroduced to increase the performance of document binarization methods. The\nmethod uses the distribution of the pixel data of the input document image in a\ncolor space to find a transformation, called the dual transform, which balances\nthe amount of information on all color channels. Furthermore, in order to\nreduce the intensity variations on the gray output, a color reduction\npreprocessing step is applied. Then, a channel is selected as the gray value\nrepresentation of the document image based on the homogeneity criterion on the\ntext regions. In this way, the proposed method can provide a\nluminance-independent contrast enhancement. The performance of the method is\nevaluated against various images from two databases, the ICDAR'03 Robust\nReading, the KAIST and the DIBCO'09 datasets, subjectively and objectively with\npromising results. The ground truth images for the images from the ICDAR'03\nRobust Reading dataset have been created manually by the authors."", ""Thumbnail is the face of online videos. The explosive growth of videos both\nin number and variety underpins the importance of a good thumbnail because it\nsaves potential viewers time to choose videos and even entice them to click on\nthem. A good thumbnail should be a frame that best represents the content of a\nvideo while at the same time capturing viewers' attention. However, the\ntechniques and models in the past only focus on frames within a video, and we\nbelieve such narrowed focus leave out much useful information that are part of\na video. In this paper, we expand the definition of content to include title,\ndescription, and audio of a video and utilize information provided by these\nmodalities in our selection model. Specifically, our model will first sample\nframes uniformly in time and return the top 1,000 frames in this subset with\nthe highest aesthetic scores by a Double-column Convolutional Neural Network,\nto avoid the computational burden of processing all frames in downstream task.\nThen, the model incorporates frame features extracted from VGG16, text features\nfrom ELECTRA, and audio features from TRILL. These models were selected because\nof their results on popular datasets as well as their competitive performances.\nAfter feature extraction, the time-series features, frames and audio, will be\nfed into Transformer encoder layers to return a vector representing their\ncorresponding modality. Each of the four features (frames, title, description,\naudios) will pass through a context gating layer before concatenation. Finally,\nour model will generate a vector in the latent space and select the frame that\nis most similar to this vector in the latent space. To the best of our\nknowledge, we are the first to propose a multi-modal deep learning model to\nselect video thumbnail, which beats the result from the previous\nState-of-The-Art models."", 'Image colorization is the process of estimating RGB colors for grayscale\nimages or video frames to improve their aesthetic and perceptual quality. Deep\nlearning techniques for image colorization have progressed notably over the\nlast decade, calling the need for a systematic survey and benchmarking of these\ntechniques. This article presents a comprehensive survey of recent\nstate-of-the-art deep learning-based image colorization techniques, describing\ntheir fundamental block architectures, inputs, optimizers, loss functions,\ntraining protocols, and training data \\textit{etc.} It categorizes the existing\ncolorization techniques into seven classes and discusses important factors\ngoverning their performance, such as benchmark datasets and evaluation metrics.\nWe highlight the limitations of existing datasets and introduce a new dataset\nspecific to colorization. Using the existing datasets and our new one, we\nperform an extensive experimental evaluation of existing image colorization\nmethods. Finally, we discuss the limitations of existing methods and recommend\npossible solutions as well as future research directions for this rapidly\nevolving topic of deep image colorization. Dataset and codes for evaluation are\npublicly available at https://github.com/saeed-anwar/ColorSurvey', 'Color modelling and extraction is an important topic in fashion, art, and\ndesign. Recommender systems, color-based retrieval, decorating, and fashion\ndesign can benefit from color extraction tools. Research has shown that\nmodeling color so that it can be automatically analyzed and / or extracted is a\ndifficult task. Unlike machines, color perception, although very subjective, is\nmuch simpler for humans. That being said, the first step in color modeling is\nto estimate the number of colors in the item / object. This is because color\nmodels can take advantage of the number of colors as the seed for better\nmodelling, e.g., to make color extraction further deterministic. We aim in this\nwork to develop and test models that can count the number of colors of clothing\nand other items. We propose a novel color counting method based on cumulative\ncolor histogram, which stands out among other methods. We compare the method we\npropose with other methods that utilize exhaustive color search that uses\nGaussian Mixture Models (GMMs) and K-Means as bases for scoring the optimal\nnumber of colors, in addition to another method that relies on deep learning\nmodels. Unfortunately, the GMM, K-Means, and Deep Learning models all fail to\naccurately capture the number of colors. Our proposed method can provide the\ncolor baseline that can be used in AI-based fashion applications, and can also\nfind applications in other areas, for example, interior design. To the best of\nour knowledge, this work is the first of its kind that addresses the problem of\ncolor-counting machine.', 'Color texture representation is an important step in the task of texture\nclassification. Shortest paths was used to extract color texture features from\nRGB and HSV color spaces. In this paper, we propose to use shortest paths in\nthe HSI space to build a texture representation for classification. In\nparticular, two undirected graphs are used to model the H channel and the S and\nI channels respectively in order to represent a color texture image. Moreover,\nthe shortest paths is constructed by using four pairs of pixels according to\ndifferent scales and directions of the texture image. Experimental results on\ncolored Brodatz and USPTex databases reveal that our proposed method is\neffective, and the highest classification accuracy rate is 96.93% in the\nBrodatz database.', ""Although many methods have been proposed to deal with nature image\nsuper-resolution (SR) and get impressive performance, the text images SR is not\ngood due to their ignorance of document images. In this paper, we propose a\nmatting-based dual generative adversarial network (mdGAN) for document image\nSR. Firstly, the input image is decomposed into document text, foreground and\nbackground layers using deep image matting. Then two parallel branches are\nconstructed to recover text boundary information and color information\nrespectively. Furthermore, in order to improve the restoration accuracy of\ncharacters in output image, we use the input image's corresponding ground truth\ntext label as extra supervise information to refine the two-branch networks\nduring training. Experiments on real text images demonstrate that our method\noutperforms several state-of-the-art methods quantitatively and qualitatively."", 'We propose the first approach for the decomposition of a monocular color\nvideo into direct and indirect illumination components in real time. We\nretrieve, in separate layers, the contribution made to the scene appearance by\nthe scene reflectance, the light sources and the reflections from various\ncoherent scene regions to one another. Existing techniques that invert global\nlight transport require image capture under multiplexed controlled lighting, or\nonly enable the decomposition of a single image at slow off-line frame rates.\nIn contrast, our approach works for regular videos and produces temporally\ncoherent decomposition layers at real-time frame rates. At the core of our\napproach are several sparsity priors that enable the estimation of the\nper-pixel direct and indirect illumination layers based on a small set of\njointly estimated base reflectance colors. The resulting variational\ndecomposition problem uses a new formulation based on sparse and dense sets of\nnon-linear equations that we solve efficiently using a novel alternating\ndata-parallel optimization strategy. We evaluate our approach qualitatively and\nquantitatively, and show improvements over the state of the art in this field,\nin both quality and runtime. In addition, we demonstrate various real-time\nappearance editing applications for videos with consistent illumination.', 'In real-world applications, e.g. law enforcement and video retrieval, one\noften needs to search a certain person in long videos with just one portrait.\nThis is much more challenging than the conventional settings for person\nre-identification, as the search may need to be carried out in the environments\ndifferent from where the portrait was taken. In this paper, we aim to tackle\nthis challenge and propose a novel framework, which takes into account the\nidentity invariance along a tracklet, thus allowing person identities to be\npropagated via both the visual and the temporal links. We also develop a novel\nscheme called Progressive Propagation via Competitive Consensus, which\nsignificantly improves the reliability of the propagation process. To promote\nthe study of person search, we construct a large-scale benchmark, which\ncontains 127K manually annotated tracklets from 192 movies. Experiments show\nthat our approach remarkably outperforms mainstream person re-id methods,\nraising the mAP from 42.16% to 62.27%.', 'We present color sails, a discrete-continuous color gamut representation that\nextends the color gradient analogy to three dimensions and allows interactive\ncontrol of the color blending behavior. Our representation models a wide\nvariety of color distributions in a compact manner, and lends itself to\napplications such as color exploration for graphic design, illustration and\nsimilar fields. We propose a Neural Network that can fit a color sail to any\nimage. Then, the user can adjust color sail parameters to change the base\ncolors, their blending behavior and the number of colors, exploring a wide\nrange of options for the original design. In addition, we propose a Deep\nLearning model that learns to automatically segment an image into\ncolor-compatible alpha masks, each equipped with its own color sail. This\nallows targeted color exploration by either editing their corresponding color\nsails or using standard software packages. Our model is trained on a custom\ndiverse dataset of art and design. We provide both quantitative evaluations,\nand a user study, demonstrating the effectiveness of color sail interaction.\nInteractive demos are available at www.colorsails.com.', 'We use methods from Riemann geometry to investigate transformations between\nthe color spaces of color-normal and color weak observers. The two main\napplications are the simulation of the perception of a color weak observer for\na color normal observer and the compensation of color images in a way that a\ncolor weak observer has approximately the same perception as a color normal\nobserver. The metrics in the color spaces of interest are characterized with\nthe help of ellipsoids defined by the just-noticable-differences between color\nwhich are measured with the help of color-matching experiments. The constructed\nmappings are isometries of Riemann spaces that preserve the perceived\ncolor-differences for both observers. Among the two approaches to build such an\nisometry, we introduce normal coordinates in Riemann spaces as a tool to\nconstruct a global color-weak compensation map. Compared to previously used\nmethods this method is free from approximation errors due to local\nlinearizations and it avoids the problem of shifting locations of the origin of\nthe local coordinate system. We analyse the variations of the Riemann metrics\nfor different observers obtained from new color matching experiments and\ndescribe three variations of the basic method. The performance of the methods\nis evaluated with the help of semantic differential (SD) tests.', 'Color in an image is resolved into 3 or 4 color components and 2-Dimages of\nthese components are stored in separate channels. Most of the color image\nenhancement algorithms are applied channel-by-channel on each image. But such a\nsystem of color image processing is not processing the original color. When a\ncolor image is represented as a quaternion image, processing is done in\noriginal colors. This paper proposes an implementation of the quaternion\napproach of enhancement algorithm for enhancing color images and is referred as\nthe modified alpha-rooting by the two-dimensional quaternion discrete Fourier\ntransform (2-D QDFT). Enhancement results of this proposed method are compared\nwith the channel-by-channel image enhancement by the 2-D DFT. Enhancements in\ncolor images are quantitatively measured by the color enhancement measure\nestimation (CEME), which allows for selecting optimum parameters for processing\nby the genetic algorithm. Enhancement of color images by the quaternion based\nmethod allows for obtaining images which are closer to the genuine\nrepresentation of the real original color.', ""Now that everyone can easily record videos, the quantity of which is\ncontinuously increasing, research on methods for improved video retrieval is\nimportant in the contemporary world. In cases where target videos are to be\nidentified within a large collection gathered by individuals, the appropriate\ninformation must be obtained to retrieve the correct video within a large\nnumber of similar items in the target database. The purpose of this research is\nto retrieve target videos in such cases by introducing an interaction, or a\ndialog, between the system and the user. We propose a system to retrieve videos\nby asking questions about the content of the videos and leveraging the user's\nresponses to the questions. Additionally, we confirmed the usefulness of the\nproposed system through experiments using the dataset called AVSD which\nincludes videos and dialogs about the videos."", 'Video prediction, forecasting the future frames from a sequence of input\nframes, is a challenging task since the view changes are influenced by various\nfactors, such as the global context surrounding the scene and local motion\ndynamics. In this paper, we propose a new framework to integrate these\ncomplementary attributes to predict complex pixel dynamics through deep\nnetworks. We present global context propagation networks that iteratively\naggregate the non-local neighboring representations to preserve the contextual\ninformation over the past frames. To capture the local motion pattern of\nobjects, we also devise local filter memory networks that generate adaptive\nfilter kernels by storing the prototypical motion of moving objects in the\nmemory. The proposed framework, utilizing the outputs from both networks, can\naddress blurry predictions and color distortion. We conduct experiments on\nCaltech pedestrian and UCF101 datasets, and demonstrate state-of-the-art\nresults. Especially for multi-step prediction, we obtain an outstanding\nperformance in quantitative and qualitative evaluation.', ""Plant aliveness is proven through laboratory experiments and special\nscientific instruments. In this paper, we aim to detect the degree of animation\nof plants based on the magnification of the small color changes in the plant's\ngreen leaves using the Eulerian video magnification. Capturing the video under\na controlled environment, e.g., using a tripod and direct current (DC) light\nsources, reduces camera movements and minimizes light fluctuations; we aim to\nreduce the external factors as much as possible. The acquired video is then\nstabilized and a proposed algorithm used to reduce the illumination variations.\nLastly, the Euler magnification is utilized to magnify the color changes on the\nlight invariant video. The proposed system does not require any special purpose\ninstruments as it uses a digital camera with a regular frame rate. The results\nof magnified color changes on both natural and plastic leaves show that the\nlive green leaves have color changes in contrast to the plastic leaves. Hence,\nwe can argue that the color changes of the leaves are due to biological\noperations, such as photosynthesis. To date, this is possibly the first work\nthat focuses on interpreting visually, some biological operations of plants\nwithout any special purpose instruments."", 'Creating and labelling datasets of videos for use in training Human Activity\nRecognition models is an arduous task. In this paper, we approach this by using\n3D rendering tools to generate a synthetic dataset of videos, and show that a\nclassifier trained on these videos can generalise to real videos. We use five\ndifferent augmentation techniques to generate the videos, leading to a wide\nvariety of accurately labelled unique videos. We fine tune a pre-trained I3D\nmodel on our videos, and find that the model is able to achieve a high accuracy\nof 73% on the HMDB51 dataset over three classes. We also find that augmenting\nthe HMDB training set with our dataset provides a 2% improvement in the\nperformance of the classifier. Finally, we discuss possible extensions to the\ndataset, including virtual try on and modeling motion of the people.', 'The majority of existing color naming methods focuses on the eleven basic\ncolor terms of the English language. However, in many applications, different\nsets of color names are used for the accurate description of objects. Labeling\ndata to learn these domain-specific color names is an expensive and laborious\ntask. Therefore, in this article we aim to learn color names from weakly\nlabeled data. For this purpose, we add an attention branch to the color naming\nnetwork. The attention branch is used to modulate the pixel-wise color naming\npredictions of the network. In experiments, we illustrate that the attention\nbranch correctly identifies the relevant regions. Furthermore, we show that our\nmethod obtains state-of-the-art results for pixel-wise and image-wise\nclassification on the EBAY dataset and is able to learn color names for various\ndomains.', 'Universal style transfer aims to transfer arbitrary visual styles to content\nimages. Existing feed-forward based methods, while enjoying the inference\nefficiency, are mainly limited by inability of generalizing to unseen styles or\ncompromised visual quality. In this paper, we present a simple yet effective\nmethod that tackles these limitations without training on any pre-defined\nstyles. The key ingredient of our method is a pair of feature transforms,\nwhitening and coloring, that are embedded to an image reconstruction network.\nThe whitening and coloring transforms reflect a direct matching of feature\ncovariance of the content image to a given style image, which shares similar\nspirits with the optimization of Gram matrix based cost in neural style\ntransfer. We demonstrate the effectiveness of our algorithm by generating\nhigh-quality stylized images with comparisons to a number of recent methods. We\nalso analyze our method by visualizing the whitened features and synthesizing\ntextures via simple feature coloring.', ""In this paper, we present a color transfer algorithm to colorize a broad\nrange of gray images without any user intervention. The algorithm uses a\nmachine learning-based approach to automatically colorize grayscale images. The\nalgorithm uses the superpixel representation of the reference color images to\nlearn the relationship between different image features and their corresponding\ncolor values. We use this learned information to predict the color value of\neach grayscale image superpixel. As compared to processing individual image\npixels, our use of superpixels helps us to achieve a much higher degree of\nspatial consistency as well as speeds up the colorization process. The\npredicted color values of the gray-scale image superpixels are used to provide\na 'micro-scribble' at the centroid of the superpixels. These color scribbles\nare refined by using a voting based approach. To generate the final\ncolorization result, we use an optimization-based approach to smoothly spread\nthe color scribble across all pixels within a superpixel. Experimental results\non a broad range of images and the comparison with existing state-of-the-art\ncolorization methods demonstrate the greater effectiveness of the proposed\nalgorithm."", ""Colors play a particularly important role in both designing and accessing Web\npages. A well-designed color scheme improves Web pages' visual aesthetic and\nfacilitates user interactions. As far as we know, existing color assessment\nstudies focus on images; studies on color assessment and editing for Web pages\nare rare. This paper investigates color assessment for Web pages based on\nexisting online color theme-rating data sets and applies this assessment to Web\ncolor edit. This study consists of three parts. First, we study the extraction\nof a Web page's color theme. Second, we construct color assessment models that\nscore the color compatibility of a Web page by leveraging machine learning\ntechniques. Third, we incorporate the learned color assessment model into a new\napplication, namely, color transfer for Web pages. Our study combines\ntechniques from computer graphics, Web mining, computer vision, and machine\nlearning. Experimental results suggest that our constructed color assessment\nmodels are effective, and useful in the color transfer for Web pages, which has\nreceived little attention in both Web mining and computer graphics communities."", 'We suggest a new color distance based on two observations. First, perceptual\ncolor differences were designed to be used to compare very similar colors. They\ndo not capture human perception for medium and large color differences well.\nThresholding was proposed to solve the problem for large color differences,\ni.e. two totally different colors are always the same distance apart. We show\nthat thresholding alone cannot improve medium color differences. We suggest to\nalleviate this problem using basic color terms. Second, when a color distance\nis used for edge detection, many small distances around the just noticeable\ndifference may account for false edges. We suggest to reduce the effect of\nsmall distances.', 'Manually re-drawing an image in a certain artistic style takes a professional\nartist a long time. Doing this for a video sequence single-handedly is beyond\nimagination. We present two computational approaches that transfer the style\nfrom one image (for example, a painting) to a whole video sequence. In our\nfirst approach, we adapt to videos the original image style transfer technique\nby Gatys et al. based on energy minimization. We introduce new ways of\ninitialization and new loss functions to generate consistent and stable\nstylized video sequences even in cases with large motion and strong occlusion.\nOur second approach formulates video stylization as a learning problem. We\npropose a deep network architecture and training procedures that allow us to\nstylize arbitrary-length videos in a consistent and stable way, and nearly in\nreal time. We show that the proposed methods clearly outperform simpler\nbaselines both qualitatively and quantitatively. Finally, we propose a way to\nadapt these approaches also to 360 degree images and videos as they emerge with\nrecent virtual reality hardware.', 'In recent years, recognition of text from natural scene image and video frame\nhas got increased attention among the researchers due to its various\ncomplexities and challenges. Because of low resolution, blurring effect,\ncomplex background, different fonts, color and variant alignment of text within\nimages and video frames, etc., text recognition in such scenario is difficult.\nMost of the current approaches usually apply a binarization algorithm to\nconvert them into binary images and next OCR is applied to get the recognition\nresult. In this paper, we present a novel approach based on color channel\nselection for text recognition from scene images and video frames. In the\napproach, at first, a color channel is automatically selected and then selected\ncolor channel is considered for text recognition. Our text recognition\nframework is based on Hidden Markov Model (HMM) which uses Pyramidal Histogram\nof Oriented Gradient features extracted from selected color channel. From each\nsliding window of a color channel our color-channel selection approach analyzes\nthe image properties from the sliding window and then a multi-label Support\nVector Machine (SVM) classifier is applied to select the color channel that\nwill provide the best recognition results in the sliding window. This color\nchannel selection for each sliding window has been found to be more fruitful\nthan considering a single color channel for the whole word image. Five\ndifferent features have been analyzed for multi-label SVM based color channel\nselection where wavelet transform based feature outperforms others. Our\nframework has been tested on different publicly available scene/video text\nimage datasets. For Devanagari script, we collected our own data dataset. The\nperformances obtained from experimental results are encouraging and show the\nadvantage of the proposed method.', 'The color sensation evoked by an object depends on both the spectral power\ndistribution of the illumination and the reflectance properties of the object\nbeing illuminated. The color sensation can be characterized by three\ncolor-space values, such as XYZ, RGB, HSV, L*a*b*, etc. It is straightforward\nto compute the three values given the illuminant and reflectance curves. The\nconverse process of computing a reflectance curve given the color-space values\nand the illuminant is complicated by the fact that an infinite number of\ndifferent reflectance curves can give rise to a single set of color-space\nvalues (metamerism). This paper presents five algorithms for generating a\nreflectance curve from a specified sRGB triplet, written for a general\naudience. The algorithms are designed to generate reflectance curves that are\nsimilar to those found with naturally occurring colored objects. The computed\nreflectance curves are compared to a database of thousands of reflectance\ncurves measured from paints and pigments available both commercially and in\nnature, and the similarity is quantified. One particularly useful application\nof these algorithms is in the field of computer graphics, where modeling color\ntransformations sometimes requires wavelength-specific information, such as\nwhen modeling subtractive color mixture.', 'In this paper, a color transfer framework to evoke different emotions for\nimages based on color combinations is proposed. The purpose of this color\ntransfer is to change the ""look and feel"" of images, i.e., evoking different\nemotions. Colors are confirmed as the most attractive factor in images. In\naddition, various studies in both art and science areas have concluded that\nother than single color, color combinations are necessary to evoke specific\nemotions. Therefore, we propose a novel framework to transfer color of images\nbased on color combinations, using a predefined color emotion model. The\ncontribution of this new framework is three-fold. First, users do not need to\nprovide reference images as used in traditional color transfer algorithms. In\nmost situations, users may not have enough aesthetic knowledge or path to\nchoose desired reference images. Second, because of the usage of color\ncombinations instead of single color for emotions, a new color transfer\nalgorithm that does not require an image library is proposed. Third, again\nbecause of the usage of color combinations, artifacts that are normally seen in\ntraditional frameworks using single color are avoided. We present encouraging\nresults generated from this new framework and its potential in several possible\napplications including color transfer of photos and paintings.', 'Pretrained language models have been shown to encode relational information,\nsuch as the relations between entities or concepts in knowledge-bases --\n(Paris, Capital, France). However, simple relations of this type can often be\nrecovered heuristically and the extent to which models implicitly reflect\ntopological structure that is grounded in world, such as perceptual structure,\nis unknown. To explore this question, we conduct a thorough case study on\ncolor. Namely, we employ a dataset of monolexemic color terms and color chips\nrepresented in CIELAB, a color space with a perceptually meaningful distance\nmetric.\n  Using two methods of evaluating the structural alignment of colors in this\nspace with text-derived color term representations, we find significant\ncorrespondence. Analyzing the differences in alignment across the color\nspectrum, we find that warmer colors are, on average, better aligned to the\nperceptual color space than cooler ones, suggesting an intriguing connection to\nfindings from recent work on efficient communication in color naming. Further\nanalysis suggests that differences in alignment are, in part, mediated by\ncollocationality and differences in syntactic usage, posing questions as to the\nrelationship between color perception and usage and context.', 'Color image segmentation is a very emerging topic for image processing\nresearch. Since it has the ability to present the result in a way that is much\nmore close to the human yes perceive, so todays more research is going on this\narea. Choosing a proper color space is a very important issue for color image\nsegmentation process. Generally LAB and HSV are the two frequently chosen color\nspaces. In this paper a comparative analysis is performed between these two\ncolor spaces with respect to color image segmentation. For measuring their\nperformance, we consider the parameters: mse and psnr . It is found that HSV\ncolor space is performing better than LAB.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently involved in a project focused on developing a night-mode facial recognition application. This technology is primarily used in security systems and driver assistance technologies that operate in dark environments. Our project requires a system capable of efficiently processing and analyzing human facial features, head orientations, and eye gaze under near-infrared illumination. To initiate this project, we are in search of a comprehensive dataset containing near-infrared videos of various subjects exhibiting different facial expressions. This dataset should be meticulously labeled to allow us to validate and evaluate algorithms for face detection, eye detection, head pose, and eye gazing tasks in dim environments.  Additionally, it would be beneficial if the dataset included a diverse range of subjects to ensure the robustness and generalizability of the algorithms we develop based on it.  Moreover, we are also seeking recommendations on human-machine interaction in dim environments since the model needs to provide further instructions to the user.","['Commercial application of facial recognition demands robustness to a variety\nof challenges such as illumination, occlusion, spoofing, disguise, etc.\nDisguised face recognition is one of the emerging issues for access control\nsystems, such as security checkpoints at the borders. However, the lack of\navailability of face databases with a variety of disguise addons limits the\ndevelopment of academic research in the area. In this paper, we present a\nmultimodal disguised face dataset to facilitate the disguised face recognition\nresearch. The presented database contains 8 facial add-ons and 7 additional\ncombinations of these add-ons to create a variety of disguised face images.\nEach facial image is captured in visible, visible plus infrared, infrared, and\nthermal spectra. Specifically, the database contains 100 subjects divided into\nsubset-A (30 subjects, 1 image per modality) and subset-B (70 subjects, 5 plus\nimages per modality). We also present baseline face detection results performed\non the proposed database to provide reference results and compare the\nperformance in different modalities. Qualitative and quantitative analysis is\nperformed to evaluate the challenging nature of disguise addons. The dataset\nwill be publicly available with the acceptance of the research article. The\ndatabase is available at: https://github.com/usmancheema89/SejongFaceDatabase.', 'Face recognition in real life situations like low illumination condition is\nstill an open challenge in biometric security. It is well established that the\nstate-of-the-art methods in face recognition provide low accuracy in the case\nof poor illumination. In this work, we propose an algorithm for a more robust\nillumination invariant face recognition using a multi-modal approach. We\npropose a new dataset consisting of aligned faces of thermal and visual images\nof a hundred subjects. We then apply face detection on thermal images using the\nbiggest blob extraction method and apply them for fusing images of different\nmodalities for the purpose of face recognition. An algorithm is proposed to\nimplement fusion of thermal and visual images. We reason for why relying on\nonly one modality can give erroneous results. We use a lighter and faster CNN\nmodel called MobileNet for the purpose of face recognition with faster\ninferencing and to be able to be use it in real time biometric systems. We test\nour proposed method on our own created dataset to show that real-time face\nrecognition on fused images shows far better results than using visual or\nthermal images separately.', 'Facial action unit recognition has many applications from market research to\npsychotherapy and from image captioning to entertainment. Despite its recent\nprogress, deployment of these models has been impeded due to their limited\ngeneralization to unseen people and demographics. This work conducts an\nin-depth analysis of performance across several dimensions: individuals(40\nsubjects), genders (male and female), skin types (darker and lighter), and\ndatabases (BP4D and DISFA). To help suppress the variance in data, we use the\nnotion of self-supervised denoising autoencoders to design a method for deep\nface normalization(DeepFN) that transfers facial expressions of different\npeople onto a common facial template which is then used to train and evaluate\nfacial action recognition models. We show that person-independent models yield\nsignificantly lower performance (55% average F1 and accuracy across 40\nsubjects) than person-dependent models (60.3%), leading to a generalization gap\nof 5.3%. However, normalizing the data with the newly introduced DeepFN\nsignificantly increased the performance of person-independent models (59.6%),\neffectively reducing the gap. Similarly, we observed generalization gaps when\nconsidering gender (2.4%), skin type (5.3%), and dataset (9.4%), which were\nsignificantly reduced with the use of DeepFN. These findings represent an\nimportant step towards the creation of more generalizable facial action unit\nrecognition systems.', 'Automatic recognition of spontaneous facial expressions is a major challenge\nin the field of affective computing. Head rotation, face pose, illumination\nvariation, occlusion etc. are the attributes that increase the complexity of\nrecognition of spontaneous expressions in practical applications. Effective\nrecognition of expressions depends significantly on the quality of the database\nused. Most well-known facial expression databases consist of posed expressions.\nHowever, currently there is a huge demand for spontaneous expression databases\nfor the pragmatic implementation of the facial expression recognition\nalgorithms. In this paper, we propose and establish a new facial expression\ndatabase containing spontaneous expressions of both male and female\nparticipants of Indian origin. The database consists of 428 segmented video\nclips of the spontaneous facial expressions of 50 participants. In our\nexperiment, emotions were induced among the participants by using emotional\nvideos and simultaneously their self-ratings were collected for each\nexperienced emotion. Facial expression clips were annotated carefully by four\ntrained decoders, which were further validated by the nature of stimuli used\nand self-report of emotions. An extensive analysis was carried out on the\ndatabase using several machine learning algorithms and the results are provided\nfor future reference. Such a spontaneous database will help in the development\nand validation of algorithms for recognition of spontaneous expressions.', ""In our everyday lives and social interactions we often try to perceive the\nemotional states of people. There has been a lot of research in providing\nmachines with a similar capacity of recognizing emotions. From a computer\nvision perspective, most of the previous efforts have been focusing in\nanalyzing the facial expressions and, in some cases, also the body pose. Some\nof these methods work remarkably well in specific settings. However, their\nperformance is limited in natural, unconstrained environments. Psychological\nstudies show that the scene context, in addition to facial expression and body\npose, provides important information to our perception of people's emotions.\nHowever, the processing of the context for automatic emotion recognition has\nnot been explored in depth, partly due to the lack of proper data. In this\npaper we present EMOTIC, a dataset of images of people in a diverse set of\nnatural situations, annotated with their apparent emotion. The EMOTIC dataset\ncombines two different types of emotion representation: (1) a set of 26\ndiscrete categories, and (2) the continuous dimensions Valence, Arousal, and\nDominance. We also present a detailed statistical and algorithmic analysis of\nthe dataset along with annotators' agreement analysis. Using the EMOTIC dataset\nwe train different CNN models for emotion recognition, combining the\ninformation of the bounding box containing the person with the contextual\ninformation extracted from the scene. Our results show how scene context\nprovides important information to automatically recognize emotional states and\nmotivate further research in this direction. Dataset and code is open-sourced\nand available at: https://github.com/rkosti/emotic and link for the\npeer-reviewed published article: https://ieeexplore.ieee.org/document/8713881"", 'Face identification/recognition has significantly advanced over the past\nyears. However, most of the proposed approaches rely on static RGB frames and\non neutral facial expressions. This has two disadvantages. First, important\nfacial shape cues are ignored. Second, facial deformations due to expressions\ncan have an impact on the performance of such a method. In this paper, we\npropose a novel framework for dynamic 3D face identification/recognition based\non facial keypoints. Each dynamic sequence of facial expressions is represented\nas a spatio-temporal graph, which is constructed using 3D facial landmarks.\nEach graph node contains local shape and texture features that are extracted\nfrom its neighborhood. For the classification/identification of faces, a\nSpatio-temporal Graph Convolutional Network (ST-GCN) is used. Finally, we\nevaluate our approach on a challenging dynamic 3D facial expression dataset.']","[""A key step to driver safety is to observe the driver's activities with the\nface being a key step in this process to extracting information such as head\npose, blink rate, yawns, talking to passenger which can then help derive higher\nlevel information such as distraction, drowsiness, intent, and where they are\nlooking. In the context of driving safety, it is important for the system\nperform robust estimation under harsh lighting and occlusion but also be able\nto detect when the occlusion occurs so that information predicted from occluded\nparts of the face can be taken into account properly. This paper introduces the\nOccluded Stacked Hourglass, based on the work of original Stacked Hourglass\nnetwork for body pose joint estimation, which is retrained to process a\ndetected face window and output 68 occlusion heat maps, each corresponding to a\nfacial landmark. Landmark location, occlusion levels and a refined face\ndetection score, to reject false positives, are extracted from these heat maps.\nUsing the facial landmark locations, features such as head pose and eye/mouth\nopenness can be extracted to derive driver attention and activity. The system\nis evaluated for face detection, head pose, and occlusion estimation on various\ndatasets in the wild, both quantitatively and qualitatively, and shows\nstate-of-the-art results."", ""The DeepFakes, which are the facial manipulation techniques, is the emerging\nthreat to digital society. Various DeepFake detection methods and datasets are\nproposed for detecting such data, especially for face-swapping. However, recent\nresearches less consider facial animation, which is also important in the\nDeepFake attack side. It tries to animate a face image with actions provided by\na driving video, which also leads to a concern about the security of recent\npayment systems that reply on liveness detection to authenticate real users via\nrecognising a sequence of user facial actions. However, our experiments show\nthat the existed datasets are not sufficient to develop reliable detection\nmethods. While the current liveness detector cannot defend such videos as the\nattack. As a response, we propose a new human face animation dataset, called\nDeepFake MNIST+, generated by a SOTA image animation generator. It includes\n10,000 facial animation videos in ten different actions, which can spoof the\nrecent liveness detectors. A baseline detection method and a comprehensive\nanalysis of the method is also included in this paper. In addition, we analyze\nthe proposed dataset's properties and reveal the difficulty and importance of\ndetecting animation datasets under different types of motion and compression\nquality."", 'Learning-based methods are believed to work well for unconstrained gaze\nestimation, i.e. gaze estimation from a monocular RGB camera without\nassumptions regarding user, environment, or camera. However, current gaze\ndatasets were collected under laboratory conditions and methods were not\nevaluated across multiple datasets. Our work makes three contributions towards\naddressing these limitations. First, we present the MPIIGaze that contains\n213,659 full face images and corresponding ground-truth gaze positions\ncollected from 15 users during everyday laptop use over several months. An\nexperience sampling approach ensured continuous gaze and head poses and\nrealistic variation in eye appearance and illumination. To facilitate\ncross-dataset evaluations, 37,667 images were manually annotated with eye\ncorners, mouth corners, and pupil centres. Second, we present an extensive\nevaluation of state-of-the-art gaze estimation methods on three current\ndatasets, including MPIIGaze. We study key challenges including target gaze\nrange, illumination conditions, and facial appearance variation. We show that\nimage resolution and the use of both eyes affect gaze estimation performance\nwhile head pose and pupil centre information are less informative. Finally, we\npropose GazeNet, the first deep appearance-based gaze estimation method.\nGazeNet improves the state of the art by 22% percent (from a mean error of 13.9\ndegrees to 10.8 degrees) for the most challenging cross-dataset evaluation.', 'Facial expressions are one of the most powerful, natural and immediate means\nfor human being to communicate their emotions and intensions. Recognition of\nfacial expression has many applications including human-computer interaction,\ncognitive science, human emotion analysis, personality development etc. In this\npaper, we propose a new method for the recognition of facial expressions from\nsingle image frame that uses combination of appearance and geometric features\nwith support vector machines classification. In general, appearance features\nfor the recognition of facial expressions are computed by dividing face region\ninto regular grid (holistic representation). But, in this paper we extracted\nregion specific appearance features by dividing the whole face region into\ndomain specific local regions. Geometric features are also extracted from\ncorresponding domain specific regions. In addition, important local regions are\ndetermined by using incremental search approach which results in the reduction\nof feature dimension and improvement in recognition accuracy. The results of\nfacial expressions recognition using features from domain specific regions are\nalso compared with the results obtained using holistic representation. The\nperformance of the proposed facial expression recognition system has been\nvalidated on publicly available extended Cohn-Kanade (CK+) facial expression\ndata sets.', 'Face recognition is a long standing challenge in the field of Artificial\nIntelligence (AI). The goal is to create systems that accurately detect,\nrecognize, verify, and understand human faces. There are significant technical\nhurdles in making these systems accurate, particularly in unconstrained\nsettings due to confounding factors related to pose, resolution, illumination,\nocclusion, and viewpoint. However, with recent advances in neural networks,\nface recognition has achieved unprecedented accuracy, largely built on\ndata-driven deep learning methods. While this is encouraging, a critical aspect\nthat is limiting facial recognition accuracy and fairness is inherent facial\ndiversity. Every face is different. Every face reflects something unique about\nus. Aspects of our heritage - including race, ethnicity, culture, geography -\nand our individual identify - age, gender, and other visible manifestations of\nself-expression, are reflected in our faces. We expect face recognition to work\nequally accurately for every face. Face recognition needs to be fair. As we\nrely on data-driven methods to create face recognition technology, we need to\nensure necessary balance and coverage in training data. However, there are\nstill scientific questions about how to represent and extract pertinent facial\nfeatures and quantitatively measure facial diversity. Towards this goal,\nDiversity in Faces (DiF) provides a data set of one million annotated human\nface images for advancing the study of facial diversity. The annotations are\ngenerated using ten well-established facial coding schemes from the scientific\nliterature. The facial coding schemes provide human-interpretable quantitative\nmeasures of facial features. We believe that by making the extracted coding\nschemes available on a large set of faces, we can accelerate research and\ndevelopment towards creating more fair and accurate facial recognition systems.', ""Affect recognition based on subjects' facial expressions has been a topic of\nmajor research in the attempt to generate machines that can understand the way\nsubjects feel, act and react. In the past, due to the unavailability of large\namounts of data captured in real-life situations, research has mainly focused\non controlled environments. However, recently, social media and platforms have\nbeen widely used. Moreover, deep learning has emerged as a means to solve\nvisual analysis and recognition problems. This paper exploits these advances\nand presents significant contributions for affect analysis and recognition\nin-the-wild. Affect analysis and recognition can be seen as a dual knowledge\ngeneration problem, involving: i) creation of new, large and rich in-the-wild\ndatabases and ii) design and training of novel deep neural architectures that\nare able to analyse affect over these databases and to successfully generalise\ntheir performance on other datasets. The paper focuses on large in-the-wild\ndatabases, i.e., Aff-Wild and Aff-Wild2 and presents the design of two classes\nof deep neural networks trained with these databases. The first class refers to\nuni-task affect recognition, focusing on prediction of the valence and arousal\ndimensional variables. The second class refers to estimation of all main\nbehavior tasks, i.e. valence-arousal prediction; categorical emotion\nclassification in seven basic facial expressions; facial Action Unit detection.\nA novel multi-task and holistic framework is presented which is able to jointly\nlearn and effectively generalize and perform affect recognition over all\nexisting in-the-wild databases. Large experimental studies illustrate the\nachieved performance improvement over the existing state-of-the-art in affect\nrecognition."", 'Automatic face recognition is an area with immense practical potential which\nincludes a wide range of commercial and law enforcement applications. Hence it\nis unsurprising that it continues to be one of the most active research areas\nof computer vision. Even after over three decades of intense research, the\nstate-of-the-art in face recognition continues to improve, benefitting from\nadvances in a range of different research fields such as image processing,\npattern recognition, computer graphics, and physiology. Systems based on\nvisible spectrum images, the most researched face recognition modality, have\nreached a significant level of maturity with some practical success. However,\nthey continue to face challenges in the presence of illumination, pose and\nexpression changes, as well as facial disguises, all of which can significantly\ndecrease recognition accuracy. Amongst various approaches which have been\nproposed in an attempt to overcome these limitations, the use of infrared (IR)\nimaging has emerged as a particularly promising research direction. This paper\npresents a comprehensive and timely review of the literature on this subject.\nOur key contributions are: (i) a summary of the inherent properties of infrared\nimaging which makes this modality promising in the context of face recognition,\n(ii) a systematic review of the most influential approaches, with a focus on\nemerging common trends as well as key differences between alternative\nmethodologies, (iii) a description of the main databases of infrared facial\nimages available to the researcher, and lastly (iv) a discussion of the most\npromising avenues for future research.', 'In the face recognition application scenario, we need to process facial\nimages captured in various conditions, such as at night by near-infrared (NIR)\nsurveillance cameras. The illumination difference between NIR and visible-light\n(VIS) causes a domain gap between facial images, and the variations in pose and\nemotion also make facial matching more difficult. Heterogeneous face\nrecognition (HFR) has difficulties in domain discrepancy, and many studies have\nfocused on extracting domain-invariant features, such as facial part relational\ninformation. However, when pose variation occurs, the facial component position\nchanges, and a different part relation is extracted. In this paper, we propose\na part relation attention module that crops facial parts obtained through a\nsemantic mask and performs relational modeling using each of these\nrepresentative features. Furthermore, we suggest component adaptive triplet\nloss function using adaptive weights for each part to reduce the intra-class\nidentity regardless of the domain as well as pose. Finally, our method exhibits\na performance improvement in the CASIA NIR-VIS 2.0 and achieves superior result\nin the BUAA-VisNir with large pose and emotion variations.', 'Human emotions analysis has been the focus of many studies, especially in the\nfield of Affective Computing, and is important for many applications, e.g.\nhuman-computer intelligent interaction, stress analysis, interactive games,\nanimations, etc. Solutions for automatic emotion analysis have also benefited\nfrom the development of deep learning approaches and the availability of vast\namount of visual facial data on the internet. This paper proposes a novel\nmethod for human emotion recognition from a single RGB image. We construct a\nlarge-scale dataset of facial videos (\\textbf{FaceVid}), rich in facial\ndynamics, identities, expressions, appearance and 3D pose variations. We use\nthis dataset to train a deep Convolutional Neural Network for estimating\nexpression parameters of a 3D Morphable Model and combine it with an effective\nback-end emotion classifier. Our proposed framework runs at 50 frames per\nsecond and is capable of robustly estimating parameters of 3D expression\nvariation and accurately recognizing facial expressions from in-the-wild\nimages. We present extensive experimental evaluation that shows that the\nproposed method outperforms the compared techniques in estimating the 3D\nexpression parameters and achieves state-of-the-art performance in recognising\nthe basic emotions from facial images, as well as recognising stress from\nfacial videos. %compared to the current state of the art in emotion recognition\nfrom facial images.', 'Face recognition is a popular and well-studied area with wide applications in\nour society. However, racial bias had been proven to be inherent in most State\nOf The Art (SOTA) face recognition systems. Many investigative studies on face\nrecognition algorithms have reported higher false positive rates of African\nsubjects cohorts than the other cohorts. Lack of large-scale African face image\ndatabases in public domain is one of the main restrictions in studying the\nracial bias problem of face recognition. To this end, we collect a face image\ndatabase namely CASIA-Face-Africa which contains 38,546 images of 1,183 African\nsubjects. Multi-spectral cameras are utilized to capture the face images under\nvarious illumination settings. Demographic attributes and facial expressions of\nthe subjects are also carefully recorded. For landmark detection, each face\nimage in the database is manually labeled with 68 facial keypoints. A group of\nevaluation protocols are constructed according to different applications,\ntasks, partitions and scenarios. The performances of SOTA face recognition\nalgorithms without re-training are reported as baselines. The proposed database\nalong with its face landmark annotations, evaluation protocols and preliminary\nresults form a good benchmark to study the essential aspects of face biometrics\nfor African subjects, especially face image preprocessing, face feature\nanalysis and matching, facial expression recognition, sex/age estimation,\nethnic classification, face image generation, etc. The database can be\ndownloaded from our http://www.cripacsir.cn/dataset/', ""Patients with schizophrenia often display impairments in the expression of\nemotion and speech and those are observed in their facial behaviour. Automatic\nanalysis of patients' facial expressions that is aimed at estimating symptoms\nof schizophrenia has received attention recently. However, the datasets that\nare typically used for training and evaluating the developed methods, contain\nonly a small number of patients (4-34) and are recorded while the subjects were\nperforming controlled tasks such as listening to life vignettes, or answering\nemotional questions. In this paper, we use videos of professional-patient\ninterviews, in which symptoms were assessed in a standardised way as they\nshould/may be assessed in practice, and which were recorded in realistic\nconditions (i.e. varying illumination levels and camera viewpoints) at the\npatients' homes or at mental health services. We automatically analyse the\nfacial behaviour of 91 out-patients - this is almost 3 times the number of\npatients in other studies - and propose SchiNet, a novel neural network\narchitecture that estimates expression-related symptoms in two different\nassessment interviews. We evaluate the proposed SchiNet for patient-independent\nprediction of symptoms of schizophrenia. Experimental results show that some\nautomatically detected facial expressions are significantly correlated to\nsymptoms of schizophrenia, and that the proposed network for estimating symptom\nseverity delivers promising results."", ""A driver's gaze is critical for determining their attention, state,\nsituational awareness, and readiness to take over control from partially\nautomated vehicles. Estimating the gaze direction is the most obvious way to\ngauge a driver's state under ideal conditions when limited to using\nnon-intrusive imaging sensors. Unfortunately, the vehicular environment\nintroduces a variety of challenges that are usually unaccounted for - harsh\nillumination, nighttime conditions, and reflective eyeglasses. Relying on head\npose alone under such conditions can prove to be unreliable and erroneous. In\nthis study, we offer solutions to address these problems encountered in the\nreal world. To solve issues with lighting, we demonstrate that using an\ninfrared camera with suitable equalization and normalization suffices. To\nhandle eyeglasses and their corresponding artifacts, we adopt image-to-image\ntranslation using generative adversarial networks to pre-process images prior\nto gaze estimation. Our proposed Gaze Preserving CycleGAN (GPCycleGAN) is\ntrained to preserve the driver's gaze while removing potential eyeglasses from\nface images. GPCycleGAN is based on the well-known CycleGAN approach - with the\naddition of a gaze classifier and a gaze consistency loss for additional\nsupervision. Our approach exhibits improved performance, interpretability,\nrobustness and superior qualitative results on challenging real-world datasets."", 'To detect bias in face recognition networks, it can be useful to probe a\nnetwork under test using samples in which only specific attributes vary in some\ncontrolled way. However, capturing a sufficiently large dataset with specific\ncontrol over the attributes of interest is difficult. In this work, we describe\na simulator that applies specific head pose and facial expression adjustments\nto images of previously unseen people. The simulator first fits a 3D morphable\nmodel to a provided image, applies the desired head pose and facial expression\ncontrols, then renders the model into an image. Next, a conditional Generative\nAdversarial Network (GAN) conditioned on the original image and the rendered\nmorphable model is used to produce the image of the original person with the\nnew facial expression and head pose. We call this conditional GAN -- MorphGAN.\nImages generated using MorphGAN conserve the identity of the person in the\noriginal image, and the provided control over head pose and facial expression\nallows test sets to be created to identify robustness issues of a facial\nrecognition deep network with respect to pose and expression. Images generated\nby MorphGAN can also serve as data augmentation when training data are scarce.\nWe show that by augmenting small datasets of faces with new poses and\nexpressions improves the recognition performance by up to 9% depending on the\naugmentation and data scarcity.', 'Automated facial identification and facial expression recognition have been\ntopics of active research over the past few decades. Facial and expression\nrecognition find applications in human-computer interfaces, subject tracking,\nreal-time security surveillance systems and social networking. Several holistic\nand geometric methods have been developed to identify faces and expressions\nusing public and local facial image databases. In this work we present the\nevolution in facial image data sets and the methodologies for facial\nidentification and recognition of expressions such as anger, sadness,\nhappiness, disgust, fear and surprise. We observe that most of the earlier\nmethods for facial and expression recognition aimed at improving the\nrecognition rates for facial feature-based methods using static images.\nHowever, the recent methodologies have shifted focus towards robust\nimplementation of facial/expression recognition from large image databases that\nvary with space (gathered from the internet) and time (video recordings). The\nevolution trends in databases and methodologies for facial and expression\nrecognition can be useful for assessing the next-generation topics that may\nhave applications in security systems or personal identification systems that\ninvolve ""Quantitative face"" assessments.', 'Automatic affect recognition using visual cues is an important task towards a\ncomplete interaction between humans and machines. Applications can be found in\ntutoring systems and human computer interaction. A critical step towards that\ndirection is facial feature extraction. In this paper, we propose a facial\nfeature extractor model trained on an in-the-wild and massively collected video\ndataset provided by the RealEyes company. The dataset consists of a million\nlabelled frames and 2,616 thousand subjects. As temporal information is\nimportant to the emotion recognition domain, we utilise LSTM cells to capture\nthe temporal dynamics in the data. To show the favourable properties of our\npre-trained model on modelling facial affect, we use the RECOLA database, and\ncompare with the current state-of-the-art approach. Our model provides the best\nresults in terms of concordance correlation coefficient.', ""Facial features deformed according to the intended facial expression.\nSpecific facial features are associated with specific facial expression, i.e.\nhappy means the deformation of mouth. This paper presents the study of facial\nfeature deformation for each facial expression by using an optical flow\nalgorithm and segmented into three different regions of interest. The\ndeformation of facial features shows the relation between facial the and facial\nexpression. Based on the experiments, the deformations of eye and mouth are\nsignificant in all expressions except happy. For happy expression, cheeks and\nmouths are the significant regions. This work also suggests that different\nfacial features' intensity varies in the way that they contribute to the\nrecognition of the different facial expression intensity. The maximum magnitude\nacross all expressions is shown by the mouth for surprise expression which is\n9x10-4. While the minimum magnitude is shown by the mouth for angry expression\nwhich is 0.4x10-4."", 'Facial expression recognition is a challenging task when neural network is\napplied to pattern recognition. Most of the current recognition research is\nbased on single source facial data, which generally has the disadvantages of\nlow accuracy and low robustness. In this paper, a neural network algorithm of\nfacial expression recognition based on multimodal data fusion is proposed. The\nalgorithm is based on the multimodal data, and it takes the facial image, the\nhistogram of oriented gradient of the image and the facial landmarks as the\ninput, and establishes CNN, LNN and HNN three sub neural networks to extract\ndata features, using multimodal data feature fusion mechanism to improve the\naccuracy of facial expression recognition. Experimental results show that,\nbenefiting by the complementarity of multimodal data, the algorithm has a great\nimprovement in accuracy, robustness and detection speed compared with the\ntraditional facial expression recognition algorithm. Especially in the case of\npartial occlusion, illumination and head posture transformation, the algorithm\nalso shows a high confidence.', 'Head detection and localization is a demanding task and a key element for\nmany computer vision applications, like video surveillance, Human Computer\nInteraction and face analysis. The stunning amount of work done for detecting\nfaces on RGB images, together with the availability of huge face datasets,\nallowed to setup very effective systems on that domain. However, due to\nillumination issues, infrared or depth cameras may be required in real\napplications. In this paper, we introduce a novel method for head detection on\ndepth images that exploits the classification ability of deep learning\napproaches. In addition to reduce the dependency on the external illumination,\ndepth images implicitly embed useful information to deal with the scale of the\ntarget objects. Two public datasets have been exploited: the first one, called\nPandora, is used to train a deep binary classifier with face and non-face\nimages. The second one, collected by Cornell University, is used to perform a\ncross-dataset test during daily activities in unconstrained environments.\nExperimental results show that the proposed method overcomes the performance of\nstate-of-art methods working on depth images.', 'Differently from computer vision systems which require explicit supervision,\nhumans can learn facial expressions by observing people in their environment.\nIn this paper, we look at how similar capabilities could be developed in\nmachine vision. As a starting point, we consider the problem of relating facial\nexpressions to objectively measurable events occurring in videos. In\nparticular, we consider a gameshow in which contestants play to win significant\nsums of money. We extract events affecting the game and corresponding facial\nexpressions objectively and automatically from the videos, obtaining large\nquantities of labelled data for our study. We also develop, using benchmarks\nsuch as FER and SFEW 2.0, state-of-the-art deep neural networks for facial\nexpression recognition, showing that pre-training on face verification data can\nbe highly beneficial for this task. Then, we extend these models to use facial\nexpressions to predict events in videos and learn nameable expressions from\nthem. The dataset and emotion recognition models are available at\nhttp://www.robots.ox.ac.uk/~vgg/data/facevalue', 'Face detection in low light scenarios is challenging but vital to many\npractical applications, e.g., surveillance video, autonomous driving at night.\nMost existing face detectors heavily rely on extensive annotations, while\ncollecting data is time-consuming and laborious. To reduce the burden of\nbuilding new datasets for low light conditions, we make full use of existing\nnormal light data and explore how to adapt face detectors from normal light to\nlow light. The challenge of this task is that the gap between normal and low\nlight is too huge and complex for both pixel-level and object-level. Therefore,\nmost existing low-light enhancement and adaptation methods do not achieve\ndesirable performance. To address the issue, we propose a joint High-Low\nAdaptation (HLA) framework. Through a bidirectional low-level adaptation and\nmulti-task high-level adaptation scheme, our HLA-Face outperforms\nstate-of-the-art methods even without using dark face labels for training. Our\nproject is publicly available at https://daooshee.github.io/HLA-Face-Website/', 'Teleconference or telepresence based on virtual reality (VR) headmount\ndisplay (HMD) device is a very interesting and promising application since HMD\ncan provide immersive feelings for users. However, in order to facilitate\nface-to-face communications for HMD users, real-time 3D facial performance\ncapture of a person wearing HMD is needed, which is a very challenging task due\nto the large occlusion caused by HMD. The existing limited solutions are very\ncomplex either in setting or in approach as well as lacking the performance\ncapture of 3D eye gaze movement. In this paper, we propose a convolutional\nneural network (CNN) based solution for real-time 3D face-eye performance\ncapture of HMD users without complex modification to devices. To address the\nissue of lacking training data, we generate massive pairs of HMD face-label\ndataset by data synthesis as well as collecting VR-IR eye dataset from multiple\nsubjects. Then, we train a dense-fitting network for facial region and an eye\ngaze network to regress 3D eye model parameters. Extensive experimental results\ndemonstrate that our system can efficiently and effectively produce in real\ntime a vivid personalized 3D avatar with the correct identity, pose, expression\nand eye motion corresponding to the HMD user.', 'In this paper, we focus on the facial expression translation task and propose\na novel Expression Conditional GAN (ECGAN) which can learn the mapping from one\nimage domain to another one based on an additional expression attribute. The\nproposed ECGAN is a generic framework and is applicable to different expression\ngeneration tasks where specific facial expression can be easily controlled by\nthe conditional attribute label. Besides, we introduce a novel face mask loss\nto reduce the influence of background changing. Moreover, we propose an entire\nframework for facial expression generation and recognition in the wild, which\nconsists of two modules, i.e., generation and recognition. Finally, we evaluate\nour framework on several public face datasets in which the subjects have\ndifferent races, illumination, occlusion, pose, color, content and background\nconditions. Even though these datasets are very diverse, both the qualitative\nand quantitative results demonstrate that our approach is able to generate\nfacial expressions accurately and robustly.', ""Automated estimation of the allocation of a driver's visual attention may be\na critical component of future Advanced Driver Assistance Systems. In theory,\nvision-based tracking of the eye can provide a good estimate of gaze location.\nIn practice, eye tracking from video is challenging because of sunglasses,\neyeglass reflections, lighting conditions, occlusions, motion blur, and other\nfactors. Estimation of head pose, on the other hand, is robust to many of these\neffects, but cannot provide as fine-grained of a resolution in localizing the\ngaze. However, for the purpose of keeping the driver safe, it is sufficient to\npartition gaze into regions. In this effort, we propose a system that extracts\nfacial features and classifies their spatial configuration into six regions in\nreal-time. Our proposed method achieves an average accuracy of 91.4% at an\naverage decision rate of 11 Hz on a dataset of 50 drivers from an on-road\nstudy."", 'Accurate, robust, inexpensive gaze tracking in the car can help keep a driver\nsafe by facilitating the more effective study of how to improve (1) vehicle\ninterfaces and (2) the design of future Advanced Driver Assistance Systems. In\nthis paper, we estimate head pose and eye pose from monocular video using\nmethods developed extensively in prior work and ask two new interesting\nquestions. First, how much better can we classify driver gaze using head and\neye pose versus just using head pose? Second, are there individual-specific\ngaze strategies that strongly correlate with how much gaze classification\nimproves with the addition of eye pose information? We answer these questions\nby evaluating data drawn from an on-road study of 40 drivers. The main insight\nof the paper is conveyed through the analogy of an ""owl"" and ""lizard"" which\ndescribes the degree to which the eyes and the head move when shifting gaze.\nWhen the head moves a lot (""owl""), not much classification improvement is\nattained by estimating eye pose on top of head pose. On the other hand, when\nthe head stays still and only the eyes move (""lizard""), classification accuracy\nincreases significantly from adding in eye pose. We characterize how that\naccuracy varies between people, gaze strategies, and gaze regions.', 'The task of action recognition in dark videos is useful in various scenarios,\ne.g., night surveillance and self-driving at night. Though progress has been\nmade in the action recognition task for videos in normal illumination, few have\nstudied action recognition in the dark. This is partly due to the lack of\nsufficient datasets for such a task. In this paper, we explored the task of\naction recognition in dark videos. We bridge the gap of the lack of data for\nthis task by collecting a new dataset: the Action Recognition in the Dark\n(ARID) dataset. It consists of over 3,780 video clips with 11 action\ncategories. To the best of our knowledge, it is the first dataset focused on\nhuman actions in dark videos. To gain further understandings of our ARID\ndataset, we analyze the ARID dataset in detail and exhibited its necessity over\nsynthetic dark videos. Additionally, we benchmarked the performance of several\ncurrent action recognition models on our dataset and explored potential methods\nfor increasing their performances. Our results show that current action\nrecognition models and frame enhancement methods may not be effective solutions\nfor the task of action recognition in dark videos.', 'With the growth of popularity of facial micro-expressions in recent years,\nthe demand for long videos with micro- and macro-expressions remains high.\nExtended from SAMM, a micro-expressions dataset released in 2016, this paper\npresents SAMM Long Videos dataset for spontaneous micro- and macro-expressions\nrecognition and spotting. SAMM Long Videos dataset consists of 147 long videos\nwith 343 macro-expressions and 159 micro-expressions. The dataset is FACS-coded\nwith detailed Action Units (AUs). We compare our dataset with Chinese Academy\nof Sciences Macro-Expressions and Micro-Expressions (CAS(ME)2) dataset, which\nis the only available fully annotated dataset with micro- and\nmacro-expressions. Furthermore, we preprocess the long videos using OpenFace,\nwhich includes face alignment and detection of facial AUs. We conduct facial\nexpression spotting using this dataset and compare it with the baseline of MEGC\nIII. Our spotting method outperformed the baseline result with F1-score of\n0.3299.', ""Much of the work on automatic facial expression recognition relies on\ndatabases containing a certain number of emotion classes and their exaggerated\nfacial configurations (generally six prototypical facial expressions), based on\nEkman's Basic Emotion Theory. However, recent studies have revealed that facial\nexpressions in our human life can be blended with multiple basic emotions. And\nthe emotion labels for these in-the-wild facial expressions cannot easily be\nannotated solely on pre-defined AU patterns. How to analyze the action units\nfor such complex expressions is still an open question. To address this issue,\nwe develop a RAF-AU database that employs a sign-based (i.e., AUs) and\njudgement-based (i.e., perceived emotion) approach to annotating blended facial\nexpressions in the wild. We first reviewed the annotation methods in existing\ndatabases and identified crowdsourcing as a promising strategy for labeling\nin-the-wild facial expressions. Then, RAF-AU was finely annotated by\nexperienced coders, on which we also conducted a preliminary investigation of\nwhich key AUs contribute most to a perceived emotion, and the relationship\nbetween AUs and facial expressions. Finally, we provided a baseline for AU\nrecognition in RAF-AU using popular features and multi-label learning methods."", 'Face recognition performance has improved remarkably in the last decade. Much\nof this success can be attributed to the development of deep learning\ntechniques such as convolutional neural networks (CNNs). While CNNs have pushed\nthe state-of-the-art forward, their training process requires a large amount of\nclean and correctly labelled training data. If a CNN is intended to tolerate\nfacial pose, then we face an important question: should this training data be\ndiverse in its pose distribution, or should face images be normalized to a\nsingle pose in a pre-processing step? To address this question, we evaluate a\nnumber of popular facial landmarking and pose correction algorithms to\nunderstand their effect on facial recognition performance. Additionally, we\nintroduce a new, automatic, single-image frontalization scheme that exceeds the\nperformance of current algorithms. CNNs trained using sets of different\npre-processing methods are used to extract features from the Point and Shoot\nChallenge (PaSC) and CMU Multi-PIE datasets. We assert that the subsequent\nverification and recognition performance serves to quantify the effectiveness\nof each pose correction scheme.', 'This paper presents a database of human faces for persons wearing spectacles.\nThe database consists of images of faces having significant variations with\nrespect to illumination, head pose, skin color, facial expressions and sizes,\nand nature of spectacles. The database contains data of 60 subjects. This\ndatabase is expected to be a precious resource for the development and\nevaluation of algorithms for face detection, eye detection, head tracking, eye\ngaze tracking, etc., for subjects wearing spectacles. As such, this can be a\nvaluable contribution to the computer vision community.', 'One of the key research areas in computer vision addressed by a vast number\nof publications is the processing and understanding of images containing human\nfaces. The most often addressed tasks include face detection, facial landmark\nlocalization, face recognition and facial expression analysis. Other, more\nspecialized tasks such as affective computing, the extraction of vital signs\nfrom videos or analysis of social interaction usually require one or several of\nthe aforementioned tasks that have to be performed. In our work, we analyze\nthat a large number of tasks for facial image processing in thermal infrared\nimages that are currently solved using specialized rule-based methods or not\nsolved at all can be addressed with modern learning-based approaches. We have\nused USTC-NVIE database for training of a number of machine learning algorithms\nfor facial landmark localization.', 'Deep Neural Networks (DNNs) have shown to outperform traditional methods in\nvarious visual recognition tasks including Facial Expression Recognition (FER).\nIn spite of efforts made to improve the accuracy of FER systems using DNN,\nexisting methods still are not generalizable enough in practical applications.\nThis paper proposes a 3D Convolutional Neural Network method for FER in videos.\nThis new network architecture consists of 3D Inception-ResNet layers followed\nby an LSTM unit that together extracts the spatial relations within facial\nimages as well as the temporal relations between different frames in the video.\nFacial landmark points are also used as inputs to our network which emphasize\non the importance of facial components rather than the facial regions that may\nnot contribute significantly to generating facial expressions. Our proposed\nmethod is evaluated using four publicly available databases in\nsubject-independent and cross-database tasks and outperforms state-of-the-art\nmethods.', 'Concern regarding the wide-spread use of fraudulent images/videos in social\nmedia necessitates precise detection of such fraud. The importance of facial\nexpressions in communication is widely known, and adversarial attacks often\nfocus on manipulating the expression related features. Thus, it is important to\ndevelop methods that can detect manipulations in facial expressions, and\nlocalize the manipulated regions. To address this problem, we propose a\nframework that is able to detect manipulations in facial expression using a\nclose combination of facial expression recognition and image manipulation\nmethods. With the addition of feature maps extracted from the facial expression\nrecognition framework, our manipulation detector is able to localize the\nmanipulated region. We show that, on the Face2Face dataset, where there is\nabundant expression manipulation, our method achieves over 3% higher accuracy\nfor both classification and localization of manipulations compared to\nstate-of-the-art methods. In addition, results on the NeuralTextures dataset\nwhere the facial expressions corresponding to the mouth regions have been\nmodified, show 2% higher accuracy in both classification and localization of\nmanipulation. We demonstrate that the method performs at-par with the\nstate-of-the-art methods in cases where the expression is not manipulated, but\nrather the identity is changed, thus ensuring generalizability of the approach.', 'Gaze and face tracking algorithms have traditionally battled a compromise\nbetween computational complexity and accuracy; the most accurate neural net\nalgorithms cannot be implemented in real time, but less complex real-time\nalgorithms suffer from higher error. This project seeks to better bridge that\ngap by improving on real-time eye and facial recognition algorithms in order to\ndevelop accurate, real-time gaze estimation with an emphasis on minimizing\ntraining data and computational complexity. Our goal is to use eye and facial\nrecognition techniques to enable users to perform limited tasks based on gaze\nand facial input using only a standard, low-quality web cam found in most\nmodern laptops and smart phones and the limited computational power and\ntraining data typical of those scenarios. We therefore identified seven\npromising, fundamentally different algorithms based on different user features\nand developed one outstanding, one workable, and one honorable mention gaze\ntracking pipelines that match the performance of modern gaze trackers while\nusing no training data.', 'Automatic facial expression recognition is an important research area in the\nemotion recognition and computer vision. Applications can be found in several\ndomains such as medical treatment, driver fatigue surveillance, sociable\nrobotics, and several other human-computer interaction systems. Therefore, it\nis crucial that the machine should be able to recognize the emotional state of\nthe user with high accuracy. In recent years, deep neural networks have been\nused with great success in recognizing emotions. In this paper, we present a\nnew model for continuous emotion recognition based on facial expression\nrecognition by using an unsupervised learning approach based on transfer\nlearning and autoencoders. The proposed approach also includes preprocessing\nand post-processing techniques which contribute favorably to improving the\nperformance of predicting the concordance correlation coefficient for arousal\nand valence dimensions. Experimental results for predicting spontaneous and\nnatural emotions on the RECOLA 2016 dataset have shown that the proposed\napproach based on visual information can achieve CCCs of 0.516 and 0.264 for\nvalence and arousal, respectively.', 'Detecting facial action units (AU) is one of the fundamental steps in\nautomatic recognition of facial expression of emotions and cognitive states.\nThough there have been a variety of approaches proposed for this task, most of\nthese models are trained only for the specific target AUs, and as such they\nfail to easily adapt to the task of recognition of new AUs (i.e., those not\ninitially used to train the target models). In this paper, we propose a deep\nlearning approach for facial AU detection that can easily and in a fast manner\nadapt to a new AU or target subject by leveraging only a few labeled samples\nfrom the new task (either an AU or subject). To this end, we propose a modeling\napproach based on the notion of the model-agnostic meta-learning, originally\nproposed for the general image recognition/detection tasks (e.g., the character\nrecognition from the Omniglot dataset). Specifically, each subject and/or AU is\ntreated as a new learning task and the model learns to adapt based on the\nknowledge of the previous tasks (the AUs and subjects used to pre-train the\ntarget models). Thus, given a new subject or AU, this meta-knowledge (that is\nshared among training and test tasks) is used to adapt the model to the new\ntask using the notion of deep learning and model-agnostic meta-learning. We\nshow on two benchmark datasets (BP4D and DISFA) for facial AU detection that\nthe proposed approach can be easily adapted to new tasks (AUs/subjects). Using\nonly a few labeled examples from these tasks, the model achieves large\nimprovements over the baselines (i.e., non-adapted models).', 'For real-world driver drowsiness detection from videos, the variation of head\npose is so large that the existing methods on global face is not capable of\nextracting effective features, such as looking aside and lowering head.\nTemporal dependencies with variable length are also rarely considered by the\nprevious approaches, e.g., yawning and speaking. In this paper, we propose a\nLong-term Multi-granularity Deep Framework to detect driver drowsiness in\ndriving videos containing the frontal faces. The framework includes two key\ncomponents: (1) Multi-granularity Convolutional Neural Network (MCNN), a novel\nnetwork utilizes a group of parallel CNN extractors on well-aligned facial\npatches of different granularities, and extracts facial representations\neffectively for large variation of head pose, furthermore, it can flexibly fuse\nboth detailed appearance clues of the main parts and local to global spatial\nconstraints; (2) a deep Long Short Term Memory network is applied on facial\nrepresentations to explore long-term relationships with variable length over\nsequential frames, which is capable to distinguish the states with temporal\ndependencies, such as blinking and closing eyes. Our approach achieves 90.05%\naccuracy and about 37 fps speed on the evaluation set of the public NTHU-DDD\ndataset, which is the state-of-the-art method on driver drowsiness detection.\nMoreover, we build a new dataset named FI-DDD, which is of higher precision of\ndrowsy locations in temporal dimension.', 'In this paper, we propose a Geometry-Contrastive Generative Adversarial\nNetwork (GC-GAN) for transferring continuous emotions across different\nsubjects. Given an input face with certain emotion and a target facial\nexpression from another subject, GC-GAN can generate an identity-preserving\nface with the target expression. Geometry information is introduced into cGANs\nas continuous conditions to guide the generation of facial expressions. In\norder to handle the misalignment across different subjects or emotions,\ncontrastive learning is used to transform geometry manifold into an embedded\nsemantic manifold of facial expressions. Therefore, the embedded geometry is\ninjected into the latent space of GANs and control the emotion generation\neffectively. Experimental results demonstrate that our proposed method can be\napplied in facial expression transfer even there exist big differences in\nfacial shapes and expressions between different subjects.', 'This study demonstrates how facial biometrics, acquired using multi-spectral\nsensors, such as RGB, depth, and infrared, assist the data accumulation in the\nprocess of authorizing users of automated and semi-automated access systems.\nThis data serves the purposes of person authentication, as well as facial\ntemperature estimation. We utilize depth data taken using an inexpensive RGB-D\nsensor to find the head pose of a subject. This allows the selection of video\nframes containing a frontal-view head pose for face recognition and face\ntemperature reading. Usage of the frontal-view frames improves the efficiency\nof face recognition while the corresponding synchronized IR video frames allow\nfor more efficient temperature estimation for facial regions of interest. In\naddition, this study reports emerging applications of biometrics in biomedical\nand health care solutions. Including surveys of recent pilot projects,\ninvolving new sensors of biometric data and new applications of human\nphysiological and behavioral biometrics. It also shows the new and promising\nhorizons of using biometrics in natural and contactless control interfaces for\nsurgical control, rehabilitation and accessibility.', 'Mobile eye-tracking systems have been available for about a decade now and\nare becoming increasingly popular in different fields of application, including\nmarketing, sociology, usability studies and linguistics. While the\nuser-friendliness and ergonomics of the hardware are developing at a rapid\npace, the software for the analysis of mobile eye-tracking data in some points\nstill lacks robustness and functionality. With this paper, we investigate which\nstate-of-the-art computer vision algorithms may be used to automate the\npost-analysis of mobile eye-tracking data. For the case study in this paper, we\nfocus on mobile eye-tracker recordings made during human-human face-to-face\ninteractions. We compared two recent publicly available frameworks (YOLOv2 and\nOpenPose) to relate the gaze location generated by the eye-tracker to the head\nand hands visible in the scene camera data. In this paper we will show that the\nuse of this single-pipeline framework provides robust results, which are both\nmore accurate and faster than previous work in the field. Moreover, our\napproach does not rely on manual interventions during this process.', ""Explainable face recognition is the problem of explaining why a facial\nmatcher matches faces. In this paper, we provide the first comprehensive\nbenchmark and baseline evaluation for explainable face recognition. We define a\nnew evaluation protocol called the ``inpainting game'', which is a curated set\nof 3648 triplets (probe, mate, nonmate) of 95 subjects, which differ by\nsynthetically inpainting a chosen facial characteristic like the nose, eyebrows\nor mouth creating an inpainted nonmate. An explainable face matcher is tasked\nwith generating a network attention map which best explains which regions in a\nprobe image match with a mated image, and not with an inpainted nonmate for\neach triplet. This provides ground truth for quantifying what image regions\ncontribute to face matching. Furthermore, we provide a comprehensive benchmark\non this dataset comparing five state of the art methods for network attention\nin face recognition on three facial matchers. This benchmark includes two new\nalgorithms for network attention called subtree EBP and Density-based Input\nSampling for Explanation (DISE) which outperform the state of the art by a wide\nmargin. Finally, we show qualitative visualization of these network attention\ntechniques on novel images, and explore how these explainable face recognition\nmodels can improve transparency and trust for facial matchers."", 'We aim to construct a system that captures real-world facial images through\nthe front camera on a laptop. The system is capable of processing/recognizing\nthe captured image and predict a result in real-time. In this system, we\nexploit the power of deep learning technique to learn a facial emotion\nrecognition (FER) model based on a set of labeled facial images. Finally,\nexperiments are conducted to evaluate our model using largely used public\ndatabase.', 'Facial expressions are important cues to observe human emotions. Facial\nexpression recognition has attracted many researchers for years, but it is\nstill a challenging topic since expression features vary greatly with the head\nposes, environments, and variations in the different persons involved. In this\nwork, three major steps are involved to improve the performance of micro-facial\nexpression recognition. First, an Adaptive Homomorphic Filtering is used for\nface detection and rotation rectification processes. Secondly, Micro-facial\nfeatures were used to extract the appearance variations of a testing\nimage-spatial analysis. The features of motion information are used for\nexpression recognition in a sequence of facial images. An effective\nMicro-Facial Expression Based Deep-Rooted Learning (MFEDRL) classifier is\nproposed in this paper to better recognize spontaneous micro-expressions by\nlearning parameters on the optimal features. This proposed method includes two\nloss functions such as cross entropy loss function and centre loss function.\nThen the performance of the algorithm will be evaluated using recognition rate\nand false measures. Simulation results show that the predictive performance of\nthe proposed method outperforms that of the existing classifiers such as\nConvolutional Neural Network (CNN), Deep Neural Network (DNN), Artificial\nNeural Network (ANN), Support Vector Machine (SVM), and k-Nearest Neighbours\n(KNN) in terms of accuracy and Mean Absolute Error (MAE).', 'Current state-of-the-art models for automatic Facial Expression Recognition\n(FER) are based on very deep neural networks that are effective but rather\nexpensive to train. Given the dynamic conditions of FER, this characteristic\nhinders such models of been used as a general affect recognition. In this\npaper, we address this problem by formalizing the FaceChannel, a light-weight\nneural network that has much fewer parameters than common deep neural networks.\nWe introduce an inhibitory layer that helps to shape the learning of facial\nfeatures in the last layer of the network and thus improving performance while\nreducing the number of trainable parameters. To evaluate our model, we perform\na series of experiments on different benchmark datasets and demonstrate how the\nFaceChannel achieves a comparable, if not better, performance to the current\nstate-of-the-art in FER. Our experiments include cross-dataset analysis, to\nestimate how our model behaves on different affective recognition conditions.\nWe conclude our paper with an analysis of how FaceChannel learns and adapt the\nlearned facial features towards the different datasets.', ""Recent advances in deep learning have significantly pushed the\nstate-of-the-art in photorealistic video animation given a single image. In\nthis paper, we extrapolate those advances to the 3D domain, by studying 3D\nimage-to-video translation with a particular focus on 4D facial expressions.\nAlthough 3D facial generative models have been widely explored during the past\nyears, 4D animation remains relatively unexplored. To this end, in this study\nwe employ a deep mesh encoder-decoder like architecture to synthesize realistic\nhigh resolution facial expressions by using a single neutral frame along with\nan expression identification. In addition, processing 3D meshes remains a\nnon-trivial task compared to data that live on grid-like structures, such as\nimages. Given the recent progress in mesh processing with graph convolutions,\nwe make use of a recently introduced learnable operator which acts directly on\nthe mesh structure by taking advantage of local vertex orderings. In order to\ngeneralize to 4D facial expressions across subjects, we trained our model using\na high resolution dataset with 4D scans of six facial expressions from 180\nsubjects. Experimental results demonstrate that our approach preserves the\nsubject's identity information even for unseen subjects and generates high\nquality expressions. To the best of our knowledge, this is the first study\ntackling the problem of 4D facial expression synthesis."", ""The majority of learning-based semantic segmentation methods are optimized\nfor daytime scenarios and favorable lighting conditions. Real-world driving\nscenarios, however, entail adverse environmental conditions such as nighttime\nillumination or glare which remain a challenge for existing approaches. In this\nwork, we propose a multimodal semantic segmentation model that can be applied\nduring daytime and nighttime. To this end, besides RGB images, we leverage\nthermal images, making our network significantly more robust. We avoid the\nexpensive annotation of nighttime images by leveraging an existing daytime\nRGB-dataset and propose a teacher-student training approach that transfers the\ndataset's knowledge to the nighttime domain. We further employ a domain\nadaptation method to align the learned feature spaces across the domains and\npropose a novel two-stage training scheme. Furthermore, due to a lack of\nthermal data for autonomous driving, we present a new dataset comprising over\n20,000 time-synchronized and aligned RGB-thermal image pairs. In this context,\nwe also present a novel target-less calibration method that allows for\nautomatic robust extrinsic and intrinsic thermal camera calibration. Among\nothers, we employ our new dataset to show state-of-the-art results for\nnighttime semantic segmentation."", 'Determining the direction in which a person is looking is an important\nproblem in a wide range of HCI applications. In this paper we describe a highly\naccurate algorithm that performs gaze estimation using an affordable and widely\navailable device such as Kinect. The method we propose starts by performing\naccurate head pose estimation achieved by fitting a person specific morphable\nmodel of the face to depth data. The ordinarily competing requirements of high\naccuracy and high speed are met concurrently by formulating the fitting\nobjective function as a combination of terms which excel either in accurate or\nfast fitting, and then by adaptively adjusting their relative contributions\nthroughout fitting. Following pose estimation, pose normalization is done by\nre-rendering the fitted model as a frontal face. Finally gaze estimates are\nobtained through regression from the appearance of the eyes in synthetic,\nnormalized images. Using EYEDIAP, the standard public dataset for the\nevaluation of gaze estimation algorithms from RGB-D data, we demonstrate that\nour method greatly outperforms the state of the art.', 'Gaze-tracking is a novel way of interacting with computers which allows new\nscenarios, such as enabling people with motor-neuron disabilities to control\ntheir computers or doctors to interact with patient information without\ntouching screen or keyboard. Further, there are emerging applications of\ngaze-tracking in interactive gaming, user experience research, human attention\nanalysis and behavioral studies. Accurate estimation of the gaze may involve\naccounting for head-pose, head-position, eye rotation, distance from the object\nas well as operating conditions such as illumination, occlusion, background\nnoise and various biological aspects of the user. Commercially available\ngaze-trackers utilize specialized sensor assemblies that usually consist of an\ninfrared light source and camera. There are several challenges in the universal\nproliferation of gaze-tracking as accessibility technologies, specifically its\naffordability, reliability, and ease-of-use. In this paper, we try to address\nthese challenges through the development of a hardware-agnostic gaze-tracker.\nWe present a deep neural network architecture as an appearance-based method for\nconstrained gaze-tracking that utilizes facial imagery captured on an ordinary\nRGB camera ubiquitous in all modern computing devices. Our system achieved an\nerror of 1.8073cm on GazeCapture dataset without any calibration or device\nspecific fine-tuning. This research shows promise that one day soon any\ncomputer, tablet, or phone will be controllable using just your eyes due to the\nprediction capabilities of deep neutral networks.', 'Gaze estimation reveals where a person is looking. It is an important clue\nfor understanding human intention. The recent development of deep learning has\nrevolutionized many computer vision tasks, the appearance-based gaze estimation\nis no exception. However, it lacks a guideline for designing deep learning\nalgorithms for gaze estimation tasks. In this paper, we present a comprehensive\nreview of the appearance-based gaze estimation methods with deep learning. We\nsummarize the processing pipeline and discuss these methods from four\nperspectives: deep feature extraction, deep neural network architecture design,\npersonal calibration as well as device and platform. Since the data\npre-processing and post-processing methods are crucial for gaze estimation, we\nalso survey face/eye detection method, data rectification method, 2D/3D gaze\nconversion method, and gaze origin conversion method. To fairly compare the\nperformance of various gaze estimation approaches, we characterize all the\npublicly available gaze estimation datasets and collect the code of typical\ngaze estimation algorithms. We implement these codes and set up a benchmark of\nconverting the results of different methods into the same evaluation metrics.\nThis paper not only serves as a reference to develop deep learning-based gaze\nestimation methods but also a guideline for future gaze estimation research.\nImplemented methods and data processing codes are available at\nhttp://phi-ai.org/GazeHub.', 'With the emergence of Virtual and Mixed Reality (XR) devices, eye tracking\nhas received significant attention in the computer vision community. Eye gaze\nestimation is a crucial component in XR -- enabling energy efficient rendering,\nmulti-focal displays, and effective interaction with content. In head-mounted\nXR devices, the eyes are imaged off-axis to avoid blocking the field of view.\nThis leads to increased challenges in inferring eye related quantities and\nsimultaneously provides an opportunity to develop accurate and robust learning\nbased approaches. To this end, we present MagicEyes, the first large scale eye\ndataset collected using real MR devices with comprehensive ground truth\nlabeling. MagicEyes includes $587$ subjects with $80,000$ images of\nhuman-labeled ground truth and over $800,000$ images with gaze target labels.\nWe evaluate several state-of-the-art methods on MagicEyes and also propose a\nnew multi-task EyeNet model designed for detecting the cornea, glints and pupil\nalong with eye segmentation in a single forward pass.', 'Facial landmark detection, head pose estimation, and facial deformation\nanalysis are typical facial behavior analysis tasks in computer vision. The\nexisting methods usually perform each task independently and sequentially,\nignoring their interactions. To tackle this problem, we propose a unified\nframework for simultaneous facial landmark detection, head pose estimation, and\nfacial deformation analysis, and the proposed model is robust to facial\nocclusion. Following a cascade procedure augmented with model-based head pose\nestimation, we iteratively update the facial landmark locations, facial\nocclusion, head pose and facial de- formation until convergence. The\nexperimental results on benchmark databases demonstrate the effectiveness of\nthe proposed method for simultaneous facial landmark detection, head pose and\nfacial deformation estimation, even if the images are under facial occlusion.', 'A head-mounted display (HMD) could be an important component of augmented\nreality system. However, as the upper face region is seriously occluded by the\ndevice, the user experience could be affected in applications such as\ntelecommunication and multi-player video games. In this paper, we first present\na novel experimental setup that consists of two near-infrared (NIR) cameras to\npoint to the eye regions and one visible-light RGB camera to capture the\nvisible face region. The main purpose of this paper is to synthesize realistic\nface images without occlusions based on the images captured by these cameras.\nTo this end, we propose a novel synthesis framework that contains four modules:\n3D head reconstruction, face alignment and tracking, face synthesis, and eye\nsynthesis. In face synthesis, we propose a novel algorithm that can robustly\nalign and track a personalized 3D head model given a face that is severely\noccluded by the HMD. In eye synthesis, in order to generate accurate eye\nmovements and dynamic wrinkle variations around eye regions, we propose another\nnovel algorithm to colorize the NIR eye images and further remove the ""red eye""\neffects caused by the colorization. Results show that both hardware setup and\nsystem framework are robust to synthesize realistic face images in video\nsequences.', 'Human eye gaze detection plays an important role in various fields, including\nhuman-computer interaction, virtual reality and cognitive science. Although\ndifferent relatively accurate systems of eye tracking and gaze detection exist,\nthey are usually either too expensive to be bought for low cost applications or\ntoo complex to be implemented easily. In this article, we propose a\nnon-wearable system for eye tracking and gaze detection with low complexity and\ncost. The proposed system provides a medium accuracy which makes it suitable\nfor general applications in which low cost and easy implementation is more\nimportant than achieving very precise gaze detection. The proposed method\nincludes pupil and marker detection using infrared image processing, and gaze\nevaluation using an interpolation-based strategy. The interpolation-based\nstrategy exploits the positions of the detected pupils and markers in a target\ncaptured image and also in some previously captured training images for\nestimating the position of a point that the user is gazing at. The proposed\nsystem has been evaluated by three users in two different lighting conditions.\nThe experimental results show that the accuracy of this low cost system can be\nbetween 90% and 100% for finding major gazing directions.', 'We present a novel multistream network that learns robust eye representations\nfor gaze estimation. We first create a synthetic dataset containing eye region\nmasks detailing the visible eyeball and iris using a simulator. We then perform\neye region segmentation with a U-Net type model which we later use to generate\neye region masks for real-world eye images. Next, we pretrain an eye image\nencoder in the real domain with self-supervised contrastive learning to learn\ngeneralized eye representations. Finally, this pretrained eye encoder, along\nwith two additional encoders for visible eyeball region and iris, are used in\nparallel in our multistream framework to extract salient features for gaze\nestimation from real-world images. We demonstrate the performance of our method\non the EYEDIAP dataset in two different evaluation settings and achieve\nstate-of-the-art results, outperforming all the existing benchmarks on this\ndataset. We also conduct additional experiments to validate the robustness of\nour self-supervised network with respect to different amounts of labeled data\nused for training.', 'Previous approaches to model and analyze facial expression analysis use three\ndifferent techniques: facial action units, geometric features and graph based\nmodelling. However, previous approaches have treated these technique\nseparately. There is an interrelationship between these techniques. The facial\nexpression analysis is significantly improved by utilizing these mappings\nbetween major geometric features involved in facial expressions and the subset\nof facial action units whose presence or absence are unique to a facial\nexpression. This paper combines dimension reduction techniques and image\nclassification with search space pruning achieved by this unique subset of\nfacial action units to significantly prune the search space. The performance\nresults on the publicly facial expression database shows an improvement in\nperformance by 70% over time while maintaining the emotion recognition\ncorrectness.', 'Surveillance based on Computer Vision has become a major necessity in current\nera. Most of the surveillance systems operate on visible light imaging, but\nperformance based on visible light imaging is limited due to some factors like\nvariation in light intensity during the daytime. The matter of concern lies in\nthe need for processing images in low light, such as in the need of nighttime\nsurveillance. In this paper, we have proposed a novel approach for human\ndetection using FLIR(Forward Looking Infrared) camera. As the principle\ninvolves sensing based on thermal radiation in the Near IR Region, it is\npossible to detect Humans from an image captured using a FLIR camera even in\nlow light. The proposed method for human detection involves processing of\nThermal images by using HOG (Histogram of Oriented Gradients) feature\nextraction technique along with some enhancements. The principle of the\nproposed technique lies in an adaptive background subtraction algorithm, which\nworks in association with the HOG technique. By means of this method, we are\nable to reduce execution time, precision and some other parameters, which\nresult in improvement of overall accuracy of the human detection system.', 'In many visual systems, visual tracking often bases on RGB image sequences,\nin which some targets are invalid in low-light conditions, and tracking\nperformance is thus affected significantly. Introducing other modalities such\nas depth and infrared data is an effective way to handle imaging limitations of\nindividual sources, but multi-modal imaging platforms usually require elaborate\ndesigns and cannot be applied in many real-world applications at present.\nNear-infrared (NIR) imaging becomes an essential part of many surveillance\ncameras, whose imaging is switchable between RGB and NIR based on the light\nintensity. These two modalities are heterogeneous with very different visual\nproperties and thus bring big challenges for visual tracking. However, existing\nworks have not studied this challenging problem. In this work, we address the\ncross-modal object tracking problem and contribute a new video dataset,\nincluding 654 cross-modal image sequences with over 481K frames in total, and\nthe average video length is more than 735 frames. To promote the research and\ndevelopment of cross-modal object tracking, we propose a new algorithm, which\nlearns the modality-aware target representation to mitigate the appearance gap\nbetween RGB and NIR modalities in the tracking process. It is plug-and-play and\ncould thus be flexibly embedded into different tracking frameworks. Extensive\nexperiments on the dataset are conducted, and we demonstrate the effectiveness\nof the proposed algorithm in two representative tracking frameworks against 17\nstate-of-the-art tracking methods. We will release the dataset for free\nacademic usage, dataset download link and code will be released soon.', 'Recognizing human affect and emotions is a problem that has a wide range of\napplications within both academia and industry. Affect and emotion recognition\nwithin computer vision primarily relies on images of faces. With the prevalence\nof portable devices (e.g. smart phones and/or smart glasses),acquiring user\nfacial images requires focus, time, and precision. While existing systems work\ngreat for full frontal faces, they tend to not work so well with partial faces\nlike those of the operator of the device when under use. Due to this, we\npropose a methodology in which we can accurately infer the overall affect of a\nperson by looking at the ocular region of an individual.', 'We present labelled pupils in the wild (LPW), a novel dataset of 66\nhigh-quality, high-speed eye region videos for the development and evaluation\nof pupil detection algorithms. The videos in our dataset were recorded from 22\nparticipants in everyday locations at about 95 FPS using a state-of-the-art\ndark-pupil head-mounted eye tracker. They cover people with different\nethnicities, a diverse set of everyday indoor and outdoor illumination\nenvironments, as well as natural gaze direction distributions. The dataset also\nincludes participants wearing glasses, contact lenses, as well as make-up. We\nbenchmark five state-of-the-art pupil detection algorithms on our dataset with\nrespect to robustness and accuracy. We further study the influence of image\nresolution, vision aids, as well as recording location (indoor, outdoor) on\npupil detection performance. Our evaluations provide valuable insights into the\ngeneral pupil detection problem and allow us to identify key challenges for\nrobust pupil detection on head-mounted eye trackers.', 'Estimation eye gaze direction is useful in various human-computer interaction\ntasks. Knowledge of gaze direction can give valuable information regarding\nusers point of attention. Certain patterns of eye movements known as eye\naccessing cues are reported to be related to the cognitive processes in the\nhuman brain. We propose a real-time framework for the classification of eye\ngaze direction and estimation of eye accessing cues. In the first stage, the\nalgorithm detects faces using a modified version of the Viola-Jones algorithm.\nA rough eye region is obtained using geometric relations and facial landmarks.\nThe eye region obtained is used in the subsequent stage to classify the eye\ngaze direction. A convolutional neural network is employed in this work for the\nclassification of eye gaze direction. The proposed algorithm was tested on Eye\nChimera database and found to outperform state of the art methods. The\ncomputational complexity of the algorithm is very less in the testing phase.\nThe algorithm achieved an average frame rate of 24 fps in the desktop\nenvironment.', ""Despite significant advances in improving the gaze tracking accuracy under\ncontrolled conditions, the tracking robustness under real-world conditions,\nsuch as large head pose and movements, use of eyeglasses, illumination and eye\ntype variations, remains a major challenge in eye tracking. In this paper, we\nrevisit this challenge and introduce a real-time multi-camera eye tracking\nframework to improve the tracking robustness. First, differently from previous\nwork, we design a multi-view tracking setup that allows for acquiring multiple\neye appearances simultaneously. Leveraging multi-view appearances enables to\nmore reliably detect gaze features under challenging conditions, particularly\nwhen they are obstructed in conventional single-view appearance due to large\nhead movements or eyewear effects. The features extracted on various\nappearances are then used for estimating multiple gaze outputs. Second, we\npropose to combine estimated gaze outputs through an adaptive fusion mechanism\nto compute user's overall point of regard. The proposed mechanism firstly\ndetermines the estimation reliability of each gaze output according to user's\nmomentary head pose and predicted gazing behavior, and then performs a\nreliability-based weighted fusion. We demonstrate the efficacy of our framework\nwith extensive simulations and user experiments on a collected dataset\nfeaturing 20 subjects. Our results show that in comparison with\nstate-of-the-art eye trackers, the proposed framework provides not only a\nsignificant enhancement in accuracy but also a notable robustness. Our\nprototype system runs at 30 frames-per-second (fps) and achieves 1 degree\naccuracy under challenging experimental scenarios, which makes it suitable for\napplications demanding high accuracy and robustness."", 'Multi-task learning is an effective learning strategy for deep-learning-based\nfacial expression recognition tasks. However, most existing methods take into\nlimited consideration the feature selection, when transferring information\nbetween different tasks, which may lead to task interference when training the\nmulti-task networks. To address this problem, we propose a novel selective\nfeature-sharing method, and establish a multi-task network for facial\nexpression recognition and facial expression synthesis. The proposed method can\neffectively transfer beneficial features between different tasks, while\nfiltering out useless and harmful information. Moreover, we employ the facial\nexpression synthesis task to enlarge and balance the training dataset to\nfurther enhance the generalization ability of the proposed method. Experimental\nresults show that the proposed method achieves state-of-the-art performance on\nthose commonly used facial expression recognition benchmarks, which makes it a\npotential solution to real-world facial expression recognition problems.', 'Recent advances in facial expression synthesis have shown promising results\nusing diverse expression representations including facial action units. Facial\naction units for an elaborate facial expression synthesis need to be\nintuitively represented for human comprehension, not a numeric categorization\nof facial action units. To address this issue, we utilize human-friendly\napproach: use of natural language where language helps human grasp conceptual\ncontexts. In this paper, therefore, we propose a new facial expression\nsynthesis model from language-based facial expression description. Our method\ncan synthesize the facial image with detailed expressions. In addition,\neffectively embedding language features on facial features, our method can\ncontrol individual word to handle each part of facial movement. Extensive\nqualitative and quantitative evaluations were conducted to verify the\neffectiveness of the natural language.', ""Image-based age estimation aims to predict a person's age from facial images.\nIt is used in a variety of real-world applications. Although end-to-end deep\nmodels have achieved impressive results for age estimation on benchmark\ndatasets, their performance in-the-wild still leaves much room for improvement\ndue to the challenges caused by large variations in head pose, facial\nexpressions, and occlusions. To address this issue, we propose a simple yet\neffective method to explicitly incorporate facial semantics into age\nestimation, so that the model would learn to correctly focus on the most\ninformative facial components from unaligned facial images regardless of head\npose and non-rigid deformation. To this end, we design a face parsing-based\nnetwork to learn semantic information at different scales and a novel face\nparsing attention module to leverage these semantic features for age\nestimation. To evaluate our method on in-the-wild data, we also introduce a new\nchallenging large-scale benchmark called IMDB-Clean. This dataset is created by\nsemi-automatically cleaning the noisy IMDB-WIKI dataset using a constrained\nclustering method. Through comprehensive experiment on IMDB-Clean and other\nbenchmark datasets, under both intra-dataset and cross-dataset evaluation\nprotocols, we show that our method consistently outperforms all existing age\nestimation methods and achieves a new state-of-the-art performance. To the best\nof our knowledge, our work presents the first attempt of leveraging face\nparsing attention to achieve semantic-aware age estimation, which may be\ninspiring to other high level facial analysis tasks. Code and data are\navailable on \\url{https://github.com/ibug-group/fpage}."", ""Images of the eye are key in several computer vision problems, such as shape\nregistration and gaze estimation. Recent large-scale supervised methods for\nthese problems require time-consuming data collection and manual annotation,\nwhich can be unreliable. We propose synthesizing perfectly labelled\nphoto-realistic training data in a fraction of the time. We used computer\ngraphics techniques to build a collection of dynamic eye-region models from\nhead scan geometry. These were randomly posed to synthesize close-up eye images\nfor a wide range of head poses, gaze directions, and illumination conditions.\nWe used our model's controllability to verify the importance of realistic\nillumination and shape variations in eye-region training data. Finally, we\ndemonstrate the benefits of our synthesized training data (SynthesEyes) by\nout-performing state-of-the-art methods for eye-shape registration as well as\ncross-dataset appearance-based gaze estimation in the wild."", ""The interaction between the vestibular and ocular system has primarily been\nstudied in controlled environments. Consequently, off-the shelf tools for\ncategorization of gaze events (e.g. fixations, pursuits, saccade) fail when\nhead movements are allowed. Our approach was to collect a novel, naturalistic,\nand multimodal dataset of eye+head movements when subjects performed everyday\ntasks while wearing a mobile eye tracker equipped with an inertial measurement\nunit and a 3D stereo camera. This Gaze-in-the-Wild dataset (GW) includes\neye+head rotational velocities (deg/s), infrared eye images and scene imagery\n(RGB+D). A portion was labelled by coders into gaze motion events with a mutual\nagreement of 0.72 sample based Cohen's $\\kappa$. This labelled data was used to\ntrain and evaluate two machine learning algorithms, Random Forest and a\nRecurrent Neural Network model, for gaze event classification. Assessment\ninvolved the application of established and novel event based performance\nmetrics. Classifiers achieve $\\sim$90$\\%$ human performance in detecting\nfixations and saccades but fall short (60$\\%$) on detecting pursuit movements.\nMoreover, pursuit classification is far worse in the absence of head movement\ninformation. A subsequent analysis of feature significance in our\nbest-performing model revealed a reliance upon absolute eye and head velocity,\nindicating that classification does not require spatial alignment of the head\nand eye tracking coordinate systems. The GW dataset, trained classifiers and\nevaluation metrics will be made publicly available with the intention of\nfacilitating growth in the emerging area of head-free gaze event\nclassification."", ""Eye-based information channels include the pupils, gaze, saccades, fixational\nmovements, and numerous forms of eye opening and closure. Pupil size variation\nindicates cognitive load and emotion, while a person's gaze direction is said\nto be congruent with the motivation to approach or avoid stimuli. The eyelids\nare involved in facial expressions that can encode basic emotions.\nAdditionally, eye-based cues can have implications for human annotators of\nemotions or feelings. Despite these facts, the use of eye-based cues in\naffective computing is in its infancy, however, and this work is intended to\nstart to address this. Eye-based feature sets, incorporating data from all of\nthe aforementioned information channels, that can be estimated from video are\nproposed. Feature set refinement is provided by way of continuous arousal and\nvalence learning and prediction experiments on the RECOLA validation set. The\neye-based features are then combined with a speech feature set to provide\nconfirmation of their usefulness and assess affect prediction performance\ncompared with group-of-humans-level performance on the RECOLA test set. The\ncore contribution of this paper, a refined eye-based feature set, is shown to\nprovide benefits for affect prediction. It is hoped that this work stimulates\nfurther research into eye-based affective computing."", 'In this paper, we propose an automatic approach for localizing the inner eye\ncanthus in thermal face images. We first coarsely detect 5 facial keypoints\ncorresponding to the center of the eyes, the nosetip and the ears. Then we\ncompute a sparse 2D-3D points correspondence using a 3D Morphable Face Model\n(3DMM). This correspondence is used to project the entire 3D face onto the\nimage, and subsequently locate the inner eye canthus. Detecting this location\nallows to obtain the most precise body temperature measurement for a person\nusing a thermal camera. We evaluated the approach on a thermal face dataset\nprovided with manually annotated landmarks. However, such manual annotations\nare normally conceived to identify facial parts such as eyes, nose and mouth,\nand are not specifically tailored for localizing the eye canthus region. As\nadditional contribution, we enrich the original dataset by using the annotated\nlandmarks to deform and project the 3DMM onto the images. Then, by manually\nselecting a small region corresponding to the eye canthus, we enrich the\ndataset with additional annotations. By using the manual landmarks, we ensure\nthe correctness of the 3DMM projection, which can be used as ground-truth for\nfuture evaluations. Moreover, we supply the dataset with the 3D head poses and\nper-point visibility masks for detecting self-occlusions. The data will be\npublicly released.', ""Estimating eye-gaze from images alone is a challenging task, in large parts\ndue to un-observable person-specific factors. Achieving high accuracy typically\nrequires labeled data from test users which may not be attainable in real\napplications. We observe that there exists a strong relationship between what\nusers are looking at and the appearance of the user's eyes. In response to this\nunderstanding, we propose a novel dataset and accompanying method which aims to\nexplicitly learn these semantic and temporal relationships. Our video dataset\nconsists of time-synchronized screen recordings, user-facing camera views, and\neye gaze data, which allows for new benchmarks in temporal gaze tracking as\nwell as label-free refinement of gaze. Importantly, we demonstrate that the\nfusion of information from visual stimuli as well as eye images can lead\ntowards achieving performance similar to literature-reported figures acquired\nthrough supervised personalization. Our final method yields significant\nperformance improvements on our proposed EVE dataset, with up to a 28 percent\nimprovement in Point-of-Gaze estimates (resulting in 2.49 degrees in angular\nerror), paving the path towards high-accuracy screen-based eye tracking purely\nfrom webcam sensors. The dataset and reference source code are available at\nhttps://ait.ethz.ch/projects/2020/EVE"", ""Onfocus detection aims at identifying whether the focus of the individual\ncaptured by a camera is on the camera or not. Based on the behavioral research,\nthe focus of an individual during face-to-camera communication leads to a\nspecial type of eye contact, i.e., the individual-camera eye contact, which is\na powerful signal in social communication and plays a crucial role in\nrecognizing irregular individual status (e.g., lying or suffering mental\ndisease) and special purposes (e.g., seeking help or attracting fans). Thus,\ndeveloping effective onfocus detection algorithms is of significance for\nassisting the criminal investigation, disease discovery, and social behavior\nanalysis. However, the review of the literature shows that very few efforts\nhave been made toward the development of onfocus detector due to the lack of\nlarge-scale public available datasets as well as the challenging nature of this\ntask. To this end, this paper engages in the onfocus detection research by\naddressing the above two issues. Firstly, we build a large-scale onfocus\ndetection dataset, named as the OnFocus Detection In the Wild (OFDIW). It\nconsists of 20,623 images in unconstrained capture conditions (thus called ``in\nthe wild'') and contains individuals with diverse emotions, ages, facial\ncharacteristics, and rich interactions with surrounding objects and background\nscenes. On top of that, we propose a novel end-to-end deep model, i.e., the\neye-context interaction inferring network (ECIIN), for onfocus detection, which\nexplores eye-context interaction via dynamic capsule routing. Finally,\ncomprehensive experiments are conducted on the proposed OFDIW dataset to\nbenchmark the existing learning models and demonstrate the effectiveness of the\nproposed ECIIN. The project (containing both datasets and codes) is at\nhttps://github.com/wintercho/focus."", 'We propose HeadOn, the first real-time source-to-target reenactment approach\nfor complete human portrait videos that enables transfer of torso and head\nmotion, face expression, and eye gaze. Given a short RGB-D video of the target\nactor, we automatically construct a personalized geometry proxy that embeds a\nparametric head, eye, and kinematic torso model. A novel real-time reenactment\nalgorithm employs this proxy to photo-realistically map the captured motion\nfrom the source actor to the target actor. On top of the coarse geometric\nproxy, we propose a video-based rendering technique that composites the\nmodified target portrait video via view- and pose-dependent texturing, and\ncreates photo-realistic imagery of the target actor under novel torso and head\nposes, facial expressions, and gaze directions. To this end, we propose a\nrobust tracking of the face and torso of the source actor. We extensively\nevaluate our approach and show significant improvements in enabling much\ngreater flexibility in creating realistic reenacted output videos.', 'Driving support systems, such as car navigation systems are becoming common\nand they support driver in several aspects. Non-intrusive method of detecting\nFatigue and drowsiness based on eye-blink count and eye directed instruction\ncontrolhelps the driver to prevent from collision caused by drowsy driving. Eye\ndetection and tracking under various conditions such as illumination,\nbackground, face alignment and facial expression makes the problem\ncomplex.Neural Network based algorithm is proposed in this paper to detect the\neyes efficiently. In the proposed algorithm, first the neural Network is\ntrained to reject the non-eye regionbased on images with features of eyes and\nthe images with features of non-eye using Gabor filter and Support Vector\nMachines to reduce the dimension and classify efficiently. In the algorithm,\nfirst the face is segmented using L*a*btransform color space, then eyes are\ndetected using HSV and Neural Network approach. The algorithm is tested on\nnearly 100 images of different persons under different conditions and the\nresults are satisfactory with success rate of 98%.The Neural Network is trained\nwith 50 non-eye images and 50 eye images with different angles using Gabor\nfilter. This paper is a part of research work on ""Development of Non-Intrusive\nsystem for real-time Monitoring and Prediction of Driver Fatigue and\ndrowsiness"" project sponsored by Department of Science & Technology, Govt. of\nIndia, New Delhi at Vignan Institute of Technology and Sciences, Vignan Hills,\nHyderabad.', 'Human face analysis is an important task in computer vision. According to\ncognitive-psychological studies, facial dynamics could provide crucial cues for\nface analysis. The motion of a facial local region in facial expression is\nrelated to the motion of other facial local regions. In this paper, a novel\ndeep learning approach, named facial dynamics interpreter network, has been\nproposed to interpret the important relations between local dynamics for\nestimating facial traits from expression sequence. The facial dynamics\ninterpreter network is designed to be able to encode a relational importance,\nwhich is used for interpreting the relation between facial local dynamics and\nestimating facial traits. By comparative experiments, the effectiveness of the\nproposed method has been verified. The important relations between facial local\ndynamics are investigated by the proposed facial dynamics interpreter network\nin gender classification and age estimation. Moreover, experimental results\nshow that the proposed method outperforms the state-of-the-art methods in\ngender classification and age estimation.', 'Video-to-video synthesis is a challenging problem aiming at learning a\ntranslation function between a sequence of semantic maps and a photo-realistic\nvideo depicting the characteristics of a driving video. We propose a\nhead-to-head system of our own implementation capable of fully transferring the\nhuman head 3D pose, facial expressions and eye gaze from a source to a target\nactor, while preserving the identity of the target actor. Our system produces\nhigh-fidelity, temporally-smooth and photo-realistic synthetic videos\nfaithfully transferring the human time-varying head attributes from the source\nto the target actor. Our proposed implementation: 1) works in real time ($\\sim\n20$ fps), 2) runs on a commodity laptop with a webcam as the only input, 3) is\ninteractive, allowing the participant to drive a target person, e.g. a\ncelebrity, politician, etc, instantly by varying their expressions, head pose,\nand eye gaze, and visualising the synthesised video concurrently.', 'To date, most facial expression analysis have been based on posed image\ndatabases and is carried out without being able to protect the identity of the\nsubjects whose expressions are being recognised. In this paper, we propose and\nimplement a system for classifying facial expressions of images in the\nencrypted domain based on a Paillier cryptosystem implementation of Fisher\nLinear Discriminant Analysis and k-nearest neighbour (FLDA + kNN). We present\nresults of experiments carried out on a recently developed natural visible and\ninfrared facial expression (NVIE) database of spontaneous images. To the best\nof our knowledge, this is the first system that will allow the recog-nition of\nencrypted spontaneous facial expressions by a remote server on behalf of a\nclient.', 'In this work, we propose a novel approach for generating videos of the six\nbasic facial expressions given a neutral face image. We propose to exploit the\nface geometry by modeling the facial landmarks motion as curves encoded as\npoints on a hypersphere. By proposing a conditional version of manifold-valued\nWasserstein generative adversarial network (GAN) for motion generation on the\nhypersphere, we learn the distribution of facial expression dynamics of\ndifferent classes, from which we synthesize new facial expression motions. The\nresulting motions can be transformed to sequences of landmarks and then to\nimages sequences by editing the texture information using another conditional\nGenerative Adversarial Network. To the best of our knowledge, this is the first\nwork that explores manifold-valued representations with GAN to address the\nproblem of dynamic facial expression generation. We evaluate our proposed\napproach both quantitatively and qualitatively on two public datasets;\nOulu-CASIA and MUG Facial Expression. Our experimental results demonstrate the\neffectiveness of our approach in generating realistic videos with continuous\nmotion, realistic appearance and identity preservation. We also show the\nefficiency of our framework for dynamic facial expressions generation, dynamic\nfacial expression transfer and data augmentation for training improved emotion\nrecognition models.', 'Facial feature detection from facial images has attracted great attention in\nthe field of computer vision. It is a nontrivial task since the appearance and\nshape of the face tend to change under different conditions. In this paper, we\npropose a hierarchical probabilistic model that could infer the true locations\nof facial features given the image measurements even if the face is with\nsignificant facial expression and pose. The hierarchical model implicitly\ncaptures the lower level shape variations of facial components using the\nmixture model. Furthermore, in the higher level, it also learns the joint\nrelationship among facial components, the facial expression, and the pose\ninformation through automatic structure learning and parameter estimation of\nthe probabilistic model. Experimental results on benchmark databases\ndemonstrate the effectiveness of the proposed hierarchical probabilistic model.', 'Eye movements play a vital role in perceiving the world. Eye gaze can give a\ndirect indication of the users point of attention, which can be useful in\nimproving human-computer interaction. Gaze estimation in a non-intrusive manner\ncan make human-computer interaction more natural. Eye tracking can be used for\nseveral applications such as fatigue detection, biometric authentication,\ndisease diagnosis, activity recognition, alertness level estimation,\ngaze-contingent display, human-computer interaction, etc. Even though\neye-tracking technology has been around for many decades, it has not found much\nuse in consumer applications. The main reasons are the high cost of eye\ntracking hardware and lack of consumer level applications. In this work, we\nattempt to address these two issues. In the first part of this work,\nimage-based algorithms are developed for gaze tracking which includes a new\ntwo-stage iris center localization algorithm. We have developed a new algorithm\nwhich works in challenging conditions such as motion blur, glint, and varying\nillumination levels. A person independent gaze direction classification\nframework using a convolutional neural network is also developed which\neliminates the requirement of user-specific calibration.\n  In the second part of this work, we have developed two applications which can\nbenefit from eye tracking data. A new framework for biometric identification\nbased on eye movement parameters is developed. A framework for activity\nrecognition, using gaze data from a head-mounted eye tracker is also developed.\nThe information from gaze data, ego-motion, and visual features are integrated\nto classify the activities.', 'Photorealistic facial expression synthesis from single face image can be\nwidely applied to face recognition, data augmentation for emotion recognition\nor entertainment. This problem is challenging, in part due to a paucity of\nlabeled facial expression data, making it difficult for algorithms to\ndisambiguate changes due to identity and changes due to expression. In this\npaper, we propose the conditional difference adversarial autoencoder, CDAAE,\nfor facial expression synthesis. The CDAAE takes a facial image of a previously\nunseen person and generates an image of that human face with a target emotion\nor facial action unit label. The CDAAE adds a feedforward path to an\nautoencoder structure connecting low level features at the encoder to features\nat the corresponding level at the decoder. It handles the problem of\ndisambiguating changes due to identity and changes due to facial expression by\nlearning to generate the difference between low-level features of images of the\nsame person but with different facial expressions. The CDAAE structure can be\nused to generate novel expressions by combining and interpolating between\nfacial expressions/action units within the training set. Our experimental\nresults demonstrate that the CDAAE can preserve identity information when\ngenerating facial expression for unseen subjects more faithfully than previous\napproaches. This is especially advantageous when training with small databases.', ""Recent research has demonstrated the ability to estimate gaze on mobile\ndevices by performing inference on the image from the phone's front-facing\ncamera, and without requiring specialized hardware. While this offers wide\npotential applications such as in human-computer interaction, medical diagnosis\nand accessibility (e.g., hands free gaze as input for patients with motor\ndisorders), current methods are limited as they rely on collecting data from\nreal users, which is a tedious and expensive process that is hard to scale\nacross devices. There have been some attempts to synthesize eye region data\nusing 3D models that can simulate various head poses and camera settings,\nhowever these lack in realism.\n  In this paper, we improve upon a recently suggested method, and propose a\ngenerative adversarial framework to generate a large dataset of high resolution\ncolorful images with high diversity (e.g., in subjects, head pose, camera\nsettings) and realism, while simultaneously preserving the accuracy of gaze\nlabels. The proposed approach operates on extended regions of the eye, and even\ncompletes missing parts of the image. Using this rich synthesized dataset, and\nwithout using any additional training data from real users, we demonstrate\nimprovements over state-of-the-art for estimating 2D gaze position on mobile\ndevices. We further demonstrate cross-device generalization of model\nperformance, as well as improved robustness to diverse head pose, blur and\ndistance."", 'Facial expressions are widely used in the behavioral interpretation of\nemotions, cognitive science, and social interactions. In this paper, we present\na novel method for fully automatic facial expression recognition in facial\nimage sequences. As the facial expression evolves over time facial landmarks\nare automatically tracked in consecutive video frames, using displacements\nbased on elastic bunch graph matching displacement estimation. Feature vectors\nfrom individual landmarks, as well as pairs of landmarks tracking results are\nextracted, and normalized, with respect to the first frame in the sequence. The\nprototypical expression sequence for each class of facial expression is formed,\nby taking the median of the landmark tracking results from the training facial\nexpression sequences. Multi-class AdaBoost with dynamic time warping similarity\ndistance between the feature vector of input facial expression and prototypical\nfacial expression, is used as a weak classifier to select the subset of\ndiscriminative feature vectors. Finally, two methods for facial expression\nrecognition are presented, either by using multi-class AdaBoost with dynamic\ntime warping, or by using support vector machine on the boosted feature\nvectors. The results on the Cohn-Kanade (CK+) facial expression database show a\nrecognition accuracy of 95.17% and 97.35% using multi-class AdaBoost and\nsupport vector machines, respectively.', 'Recognition of human emotions from the imaging templates is useful in a wide\nvariety of human-computer interaction and intelligent systems applications.\nHowever, the automatic recognition of facial expressions using image template\nmatching techniques suffer from the natural variability with facial features\nand recording conditions. In spite of the progress achieved in facial emotion\nrecognition in recent years, the effective and computationally simple feature\nselection and classification technique for emotion recognition is still an open\nproblem. In this paper, we propose an efficient and straightforward facial\nemotion recognition algorithm to reduce the problem of inter-class pixel\nmismatch during classification. The proposed method includes the application of\npixel normalization to remove intensity offsets followed-up with a Min-Max\nmetric in a nearest neighbor classifier that is capable of suppressing feature\noutliers. The results indicate an improvement of recognition performance from\n92.85% to 98.57% for the proposed Min-Max classification method when tested on\nJAFFE database. The proposed emotion recognition technique outperforms the\nexisting template matching methods.', 'Facial recognition is changing the way we live in and interact with our\nsociety. Here we discuss the two sides of facial recognition, summarizing\npotential risks and current concerns. We introduce current policies and\nregulations in different countries. Very importantly, we point out that the\nrisks and concerns are not only from facial recognition, but also realistically\nvery similar to other biometric recognition technology, including but not\nlimited to gait recognition, iris recognition, fingerprint recognition, voice\nrecognition, etc. To create a responsible future, we discuss possible\ntechnological moves and efforts that should be made to keep facial recognition\n(and biometric recognition in general) developing for social good.', 'The randomness and uniqueness of human eye patterns is a major breakthrough\nin the search for quicker, easier and highly reliable forms of automatic human\nidentification. It is being used extensively in security solutions. This\nincludes access control to physical facilities, security systems and\ninformation databases, Suspect tracking, surveillance and intrusion detection\nand by various Intelligence agencies through out the world. We use the\nadvantage of human eye uniqueness to identify people and approve its validity\nas a biometric. . Eye detection involves first extracting the eye from a\ndigital face image, and then encoding the unique patterns of the eye in such a\nway that they can be compared with pre-registered eye patterns. The eye\ndetection system consists of an automatic segmentation system that is based on\nthe wavelet transform, and then the Wavelet analysis is used as a pre-processor\nfor a back propagation neural network with conjugate gradient learning. The\ninputs to the neural network are the wavelet maxima neighborhood coefficients\nof face images at a particular scale. The output of the neural network is the\nclassification of the input into an eye or non-eye region. An accuracy of 90%\nis observed for identifying test images under different conditions included in\ntraining stage.', ""Gaze correction aims to redirect the person's gaze into the camera by\nmanipulating the eye region, and it can be considered as a specific image\nresynthesis problem. Gaze correction has a wide range of applications in real\nlife, such as taking a picture with staring at the camera. In this paper, we\npropose a novel method that is based on the inpainting model to learn from the\nface image to fill in the missing eye regions with new contents representing\ncorrected eye gaze. Moreover, our model does not require the training dataset\nlabeled with the specific head pose and eye angle information, thus, the\ntraining data is easy to collect. To retain the identity information of the eye\nregion in the original input, we propose a self-guided pretrained model to\nlearn the angle-invariance feature. Experiments show our model achieves very\ncompelling gaze-corrected results in the wild dataset which is collected from\nthe website and will be introduced in details. Code is available at\nhttps://github.com/zhangqianhui/GazeCorrection."", 'Model-based eye tracking has been a dominant approach for eye gaze tracking\nbecause of its ability to generalize to different subjects, without the need of\nany training data and eye gaze annotations. Model-based eye tracking, however,\nis susceptible to eye feature detection errors, in particular for eye tracking\nin the wild. To address this issue, we propose a Bayesian framework for\nmodel-based eye tracking. The proposed system consists of a cascade-Bayesian\nConvolutional Neural Network (c-BCNN) to capture the probabilistic\nrelationships between eye appearance and its landmarks, and a geometric eye\nmodel to estimate eye gaze from the eye landmarks. Given a testing eye image,\nthe Bayesian framework can generate, through Bayesian inference, the eye gaze\ndistribution without explicit landmark detection and model training, based on\nwhich it not only estimates the most likely eye gaze but also its uncertainty.\nFurthermore, with Bayesian inference instead of point-based inference, our\nmodel can not only generalize better to different sub-jects, head poses, and\nenvironments but also is robust to image noise and landmark detection errors.\nFinally, with the estimated gaze uncertainty, we can construct a cascade\narchitecture that allows us to progressively improve gaze estimation accuracy.\nCompared to state-of-the-art model-based and learning-based methods, the\nproposed Bayesian framework demonstrates significant improvement in\ngeneralization capability across several benchmark datasets and in accuracy and\nrobustness under challenging real-world conditions.', 'Facial expression recognition is a challenging task due to two major\nproblems: the presence of inter-subject variations in facial expression\nrecognition dataset and impure expressions posed by human subjects. In this\npaper we present a novel Human-to-Animation conditional Generative Adversarial\nNetwork (HA-GAN) to overcome these two problems by using many (human faces) to\none (animated face) mapping. Specifically, for any given input human expression\nimage, our HA-GAN transfers the expression information from the input image to\na fixed animated identity. Stylized animated characters from the Facial\nExpression Research Group-Database (FERGDB) are used for the generation of\nfixed identity. By learning this many-to-one identity mapping function using\nour proposed HA-GAN, the effect of inter-subject variations can be reduced in\nFacial Expression Recognition(FER). We also argue that the expressions in the\ngenerated animated images are pure expressions and since FER is performed on\nthese generated images, the performance of facial expression recognition is\nimproved. Our initial experimental results on the state-of-the-art datasets\nshow that facial expression recognition carried out on the generated animated\nimages using our HA-GAN framework outperforms the baseline deep neural network\nand produces comparable or even better results than the state-of-the-art\nmethods for facial expression recognition.', 'The infrared small-dim target detection is one of the key techniques in the\ninfrared search and tracking system. Since the local regions similar to\ninfrared small-dim targets spread over the whole background, exploring the\ninteraction information amongst image features in large-range dependencies to\nmine the difference between the target and background is crucial for robust\ndetection. However, existing deep learning-based methods are limited by the\nlocality of convolutional neural networks, which impairs the ability to capture\nlarge-range dependencies. Additionally, the small-dim appearance of the\ninfrared target makes the detection model highly possible to miss detection. To\nthis end, we propose a robust and general infrared small-dim target detection\nmethod with the transformer. We adopt the self-attention mechanism of the\ntransformer to learn the interaction information of image features in a larger\nrange. Moreover, we design a feature enhancement module to learn discriminative\nfeatures of small-dim targets to avoid miss detection. After that, to avoid the\nloss of the target information, we adopt a decoder with the U-Net-like skip\nconnection operation to contain more information of small-dim targets. Finally,\nwe get the detection result by a segmentation head. Extensive experiments on\ntwo public datasets show the obvious superiority of the proposed method over\nstate-of-the-art methods and the proposed method has stronger cross-scene\ngeneralization and anti-noise performance.', 'The muscular activities caused the activation of certain AUs for every facial\nexpression at the certain duration of time throughout the facial expression.\nThis paper presents the methods to recognise facial Action Unit (AU) using\nfacial distance of the facial features which activates the muscles. The seven\nfacial action units involved are AU1, AU4, AU6, AU12, AU15, AU17 and AU25 that\ncharacterises happy and sad expression. The recognition is performed on each AU\naccording to rules defined based on the distance of each facial points. The\nfacial distances chosen are extracted from twelve facial features. Then the\nfacial distances are trained using Support Vector Machine (SVM) and Neural\nNetwork (NN). Classification result using SVM is presented with several\ndifferent SVM kernels while result using NN is presented for each training,\nvalidation and testing phase.', 'Facial expression is the most natural means for human beings to communicate\ntheir emotions. Most facial expression analysis studies consider the case of\nacted expressions. Spontaneous facial expression recognition is significantly\nmore challenging since each person has a different way to react to a given\nemotion. We consider the problem of recognizing spontaneous facial expression\nby learning discriminative dictionaries for sparse representation. Facial\nimages are represented as a sparse linear combination of prototype atoms via\nOrthogonal Matching Pursuit algorithm. Sparse codes are then used to train an\nSVM classifier dedicated to the recognition task. The dictionary that\nsparsifies the facial images (feature points with the same class labels should\nhave similar sparse codes) is crucial for robust classification. Learning\nsparsifying dictionaries heavily relies on the initialization process of the\ndictionary. To improve the performance of dictionaries, a random face feature\ndescriptor based on the Random Projection concept is developed. The\neffectiveness of the proposed method is evaluated through several experiments\non the spontaneous facial expressions DynEmo database. It is also estimated on\nthe well-known acted facial expressions JAFFE database for a purpose of\ncomparison with state-of-the-art methods.', 'This thesis describes the development of fast algorithms for the computation\nof PERcentage CLOSure of eyes (PERCLOS) and Saccadic Ratio (SR). PERCLOS and SR\nare two ocular parameters reported to be measures of alertness levels in human\nbeings. PERCLOS is the percentage of time in which at least 80% of the eyelid\nremains closed over the pupil. Saccades are fast and simultaneous movement of\nboth the eyes in the same direction. SR is the ratio of peak saccadic velocity\nto the saccadic duration. This thesis addresses the issues of image based\nestimation of PERCLOS and SR, prevailing in the literature such as illumination\nvariation, poor illumination conditions, head rotations etc. In this work,\nalgorithms for real-time PERCLOS computation has been developed and implemented\non an embedded platform. The platform has been used as a case study for\nassessment of loss of attention in automotive drivers. The SR estimation has\nbeen carried out offline as real-time implementation requires high frame rates\nof processing which is difficult to achieve due to hardware limitations. The\naccuracy in estimation of the loss of attention using PERCLOS and SR has been\nvalidated using brain signals, which are reported to be an authentic cue for\nestimating the state of alertness in human beings. The major contributions of\nthis thesis include database creation, design and implementation of fast\nalgorithms for estimating PERCLOS and SR on embedded computing platforms.', ""A user's eyes provide means for Human Computer Interaction (HCI) research as\nan important modal. The time to time scientific explorations of the eye has\nalready seen an upsurge of the benefits in HCI applications from gaze\nestimation to the measure of attentiveness of a user looking at a screen for a\ngiven time period. The eye tracking system as an assisting, interactive tool\ncan be incorporated by physically disabled individuals, fitted best for those\nwho have eyes as only a limited set of communication. The threefold objective\nof this paper is - 1. To introduce a neural network based architecture to\npredict users' gaze at 9 positions displayed in the 11.31{\\deg} visual range on\nthe screen, through a low resolution based system such as a webcam in real time\nby learning various aspects of eyes as an ocular feature set. 2.A collection of\ncoarsely supervised feature set obtained in real time which is also validated\nthrough the user case study presented in the paper for 21 individuals ( 17 men\nand 4 women ) from whom a 35k set of instances was derived with an accuracy\nscore of 82.36% and f1_score of 82.2% and 3.A detailed study over applicability\nand underlying challenges of such systems. The experimental results verify the\nfeasibility and validity of the proposed eye gaze tracking model."", 'The ability to edit facial expressions has a wide range of applications in\ncomputer graphics. The ideal facial expression editing algorithm needs to\nsatisfy two important criteria. First, it should allow precise and targeted\nediting of individual facial actions. Second, it should generate high fidelity\noutputs without artifacts. We build a solution based on StyleGAN, which has\nbeen used extensively for semantic manipulation of faces. As we do so, we add\nto our understanding of how various semantic attributes are encoded in\nStyleGAN. In particular, we show that a naive strategy to perform editing in\nthe latent space results in undesired coupling between certain action units,\neven if they are conceptually distinct. For example, although brow lowerer and\nlip tightener are distinct action units, they appear correlated in the training\ndata. Hence, StyleGAN has difficulty in disentangling them. We allow\ndisentangled editing of such action units by computing detached regions of\ninfluence for each action unit, and restrict editing to these regions. We\nvalidate the effectiveness of our local editing method through perception\nexperiments conducted with 23 subjects. The results show that our method\nprovides higher control over local editing and produces images with superior\nfidelity compared to the state-of-the-art methods.', 'The locations of the fiducial facial landmark points around facial components\nand facial contour capture the rigid and non-rigid facial deformations due to\nhead movements and facial expressions. They are hence important for various\nfacial analysis tasks. Many facial landmark detection algorithms have been\ndeveloped to automatically detect those key points over the years, and in this\npaper, we perform an extensive review of them. We classify the facial landmark\ndetection algorithms into three major categories: holistic methods, Constrained\nLocal Model (CLM) methods, and the regression-based methods. They differ in the\nways to utilize the facial appearance and shape information. The holistic\nmethods explicitly build models to represent the global facial appearance and\nshape information. The CLMs explicitly leverage the global shape model but\nbuild the local appearance models. The regression-based methods implicitly\ncapture facial shape and appearance information. For algorithms within each\ncategory, we discuss their underlying theories as well as their differences. We\nalso compare their performances on both controlled and in the wild benchmark\ndatasets, under varying facial expressions, head poses, and occlusion. Based on\nthe evaluations, we point out their respective strengths and weaknesses. There\nis also a separate section to review the latest deep learning-based algorithms.\n  The survey also includes a listing of the benchmark databases and existing\nsoftware. Finally, we identify future research directions, including combining\nmethods in different categories to leverage their respective strengths to solve\nlandmark detection ""in-the-wild"".', ""Deep neural networks have been widely used for feature learning in facial\nexpression recognition systems. However, small datasets and large intra-class\nvariability can lead to overfitting. In this paper, we propose a method which\nlearns an optimized compact network topology for real-time facial expression\nrecognition utilizing localized facial landmark features. Our method employs a\nspatio-temporal bilinear layer as backbone to capture the motion of facial\nlandmarks during the execution of a facial expression effectively. Besides, it\ntakes advantage of Monte Carlo Dropout to capture the model's uncertainty which\nis of great importance to analyze and treat uncertain cases. The performance of\nour method is evaluated on three widely used datasets and it is comparable to\nthat of video-based state-of-the-art methods while it has much less complexity."", 'Near-infrared gray images captured together with corresponding visible color\nimages have recently proven useful for image restoration and classification.\nThis paper introduces a new coloring method to add colors to near-infrared gray\nimages based on a contrast-preserving mapping model. A naive coloring method\ndirectly adds the colors from the visible color image to the near-infrared gray\nimage; however, this method results in an unrealistic image because of the\ndiscrepancies in brightness and image structure between the captured\nnear-infrared gray image and the visible color image. To solve the discrepancy\nproblem, first we present a new contrast-preserving mapping model to create a\nnew near-infrared gray image with a similar appearance in the luminance plane\nto the visible color image, while preserving the contrast and details of the\ncaptured near-infrared gray image. Then based on the proposed\ncontrast-preserving mapping model, we develop a method to derive realistic\ncolors that can be added to the newly created near-infrared gray image.\nExperimental results show that the proposed method can not only preserve the\nlocal contrasts and details of the captured near-infrared gray image, but\ntransfers the realistic colors from the visible color image to the newly\ncreated near-infrared gray image. Experimental results also show that the\nproposed approach can be applied to near-infrared denoising.', 'In this paper, we introduce a novel deep learning method for photo-realistic\nmanipulation of the emotional state of actors in ""in-the-wild"" videos. The\nproposed method is based on a parametric 3D face representation of the actor in\nthe input scene that offers a reliable disentanglement of the facial identity\nfrom the head pose and facial expressions. It then uses a novel deep domain\ntranslation framework that alters the facial expressions in a consistent and\nplausible manner, taking into account their dynamics. Finally, the altered\nfacial expressions are used to photo-realistically manipulate the facial region\nin the input scene based on an especially-designed neural face renderer. To the\nbest of our knowledge, our method is the first to be capable of controlling the\nactor\'s facial expressions by even using as a sole input the semantic labels of\nthe manipulated emotions, while at the same time preserving the speech-related\nlip movements. We conduct extensive qualitative and quantitative evaluations\nand comparisons, which demonstrate the effectiveness of our approach and the\nespecially promising results that we obtain. Our method opens a plethora of new\npossibilities for useful applications of neural rendering technologies, ranging\nfrom movie post-production and video games to photo-realistic affective\navatars.', 'Analyzing the gaze accuracy characteristics of an eye tracker is a critical\ntask as its gaze data is frequently affected by non-ideal operating conditions\nin various consumer eye tracking applications. In this study, gaze error\npatterns produced by a commercial eye tracking device were studied with the\nhelp of machine learning algorithms, such as classifiers and regression models.\nGaze data were collected from a group of participants under multiple conditions\nthat commonly affect eye trackers operating on desktop and handheld platforms.\nThese conditions (referred here as error sources) include user distance, head\npose, and eye-tracker pose variations, and the collected gaze data were used to\ntrain the classifier and regression models. It was seen that while the impact\nof the different error sources on gaze data characteristics were nearly\nimpossible to distinguish by visual inspection or from data statistics, machine\nlearning models were successful in identifying the impact of the different\nerror sources and predicting the variability in gaze error levels due to these\nconditions. The objective of this study was to investigate the efficacy of\nmachine learning methods towards the detection and prediction of gaze error\npatterns, which would enable an in-depth understanding of the data quality and\nreliability of eye trackers under unconstrained operating conditions. Coding\nresources for all the machine learning methods adopted in this study were\nincluded in an open repository named MLGaze to allow researchers to replicate\nthe principles presented here using data from their own eye trackers.', 'This report explores the latest advances in the field of digital document\nrecognition. With the focus on printed document imagery, we discuss the major\ndevelopments in optical character recognition (OCR) and document image\nenhancement/restoration in application to Latin and non-Latin scripts. In\naddition, we review and discuss the available technologies for hand-written\ndocument recognition. In this report, we also provide some company-accumulated\nbenchmark results on available OCR engines.', 'In this paper, a modified algorithm for the detection of nasal and temporal\neye corners is presented. The algorithm is a modification of the Santos and\nProenka Method. In the first step, we detect the face and the eyes using\nclassifiers based on Haar-like features. We then segment out the sclera, from\nthe detected eye region. From the segmented sclera, we segment out an\napproximate eyelid contour. Eye corner candidates are obtained using Harris and\nStephens corner detector. We introduce a post-pruning of the Eye corner\ncandidates to locate the eye corners, finally. The algorithm has been tested on\nYale, JAFFE databases as well as our created database.', 'Despite its great success, deep learning severely suffers from robustness;\nthat is, deep neural networks are very vulnerable to adversarial attacks, even\nthe simplest ones. Inspired by recent advances in brain science, we propose the\nDenoised Internal Models (DIM), a novel generative autoencoder-based model to\ntackle this challenge. Simulating the pipeline in the human brain for visual\nsignal processing, DIM adopts a two-stage approach. In the first stage, DIM\nuses a denoiser to reduce the noise and the dimensions of inputs, reflecting\nthe information pre-processing in the thalamus. Inspired from the sparse coding\nof memory-related traces in the primary visual cortex, the second stage\nproduces a set of internal models, one for each category. We evaluate DIM over\n42 adversarial attacks, showing that DIM effectively defenses against all the\nattacks and outperforms the SOTA on the overall robustness.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', ""Most existing person re-identification (ReID) methods have good feature\nrepresentations to distinguish pedestrians with deep convolutional neural\nnetwork (CNN) and metric learning methods. However, these works concentrate on\nthe similarity between encoder output and ground-truth, ignoring the\ncorrelation between input and encoder output, which affects the performance of\nidentifying different pedestrians. To address this limitation, We design a Deep\nInfoMax (DIM) network to maximize the mutual information (MI) between the input\nimage and encoder output, which doesn't need any auxiliary labels. To evaluate\nthe effectiveness of the DIM network, we propose end-to-end Global-DIM and\nLocal-DIM models. Additionally, the DIM network provides a new solution for\ncross-dataset unsupervised ReID issue as it needs no extra labels. The\nexperiments prove the superiority of MI theory on the ReID issue, which\nachieves the state-of-the-art results."", ""This paper addresses the following questions pertaining to the intrinsic\ndimensionality of any given image representation: (i) estimate its intrinsic\ndimensionality, (ii) develop a deep neural network based non-linear mapping,\ndubbed DeepMDS, that transforms the ambient representation to the minimal\nintrinsic space, and (iii) validate the veracity of the mapping through image\nmatching in the intrinsic space. Experiments on benchmark image datasets (LFW,\nIJB-C and ImageNet-100) reveal that the intrinsic dimensionality of deep neural\nnetwork representations is significantly lower than the dimensionality of the\nambient features. For instance, SphereFace's 512-dim face representation and\nResNet's 512-dim image representation have an intrinsic dimensionality of 16\nand 19 respectively. Further, the DeepMDS mapping is able to obtain a\nrepresentation of significantly lower dimensionality while maintaining\ndiscriminative ability to a large extent, 59.75% TAR @ 0.1% FAR in 16-dim vs\n71.26% TAR in 512-dim on IJB-C and a Top-1 accuracy of 77.0% at 19-dim vs 83.4%\nat 512-dim on ImageNet-100."", 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Document segmentation is a method of rending the document into distinct\nregions. A document is an assortment of information and a standard mode of\nconveying information to others. Pursuance of data from documents involves ton\nof human effort, time intense and might severely prohibit the usage of data\nsystems. So, automatic information pursuance from the document has become a big\nissue. It is been shown that document segmentation will facilitate to beat such\nproblems. This paper proposes a new approach to segment and classify the\ndocument regions as text, image, drawings and table. Document image is divided\ninto blocks using Run length smearing rule and features are extracted from\nevery blocks. Discipulus tool has been used to construct the Genetic\nprogramming based classifier model and located 97.5% classification accuracy.', 'Recently, face super-resolution (FSR) methods either feed whole face image\ninto convolutional neural networks (CNNs) or utilize extra facial priors (e.g.,\nfacial parsing maps, facial landmarks) to focus on facial structure, thereby\nmaintaining the consistency of the facial structure while restoring facial\ndetails. However, the limited receptive fields of CNNs and inaccurate facial\npriors will reduce the naturalness and fidelity of the reconstructed face. In\nthis paper, we propose a novel paradigm based on the self-attention mechanism\n(i.e., the core of Transformer) to fully explore the representation capacity of\nthe facial structure feature. Specifically, we design a Transformer-CNN\naggregation network (TANet) consisting of two paths, in which one path uses\nCNNs responsible for restoring fine-grained facial details while the other\nutilizes a resource-friendly Transformer to capture global information by\nexploiting the long-distance visual relation modeling. By aggregating the\nfeatures from the above two paths, the consistency of global facial structure\nand fidelity of local facial detail restoration are strengthened\nsimultaneously. Experimental results of face reconstruction and recognition\nverify that the proposed method can significantly outperform the\nstate-of-the-art methods.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : My objective is to devise a methodology for constructing intelligent robots that mimic human-like behavior. These robots would be capable of operating in real-world environments and executing a range of complex tasks. Such a robotic agent would interact with its surroundings, making decisions and performing actions. More importantly, it would continuously adapt to its environment by learning from signals and cues it receives. I aim to create a unified framework that can integrate a large number of independent modules and models. Each of these would perform a fundamental function of the robotic agent, such as processing visual inputs, decision-making, and robotic arm manipulation. While these sub-models would be independent, each responsible for a different function, they would need to interact with each other to optimize the overall performance of the agent. I currently believe that the transfer of information, such as gradient information and parameter updates, from one module to another should be swift and efficient. I intend to demonstrate the adaptability of my robotic agent to different environments and situations effectively using various datasets.","['In this paper, we propose a novel task, Manipulation Question Answering\n(MQA), where the robot performs manipulation actions to change the environment\nin order to answer a given question. To solve this problem, a framework\nconsisting of a QA module and a manipulation module is proposed. For the QA\nmodule, we adopt the method for the Visual Question Answering (VQA) task. For\nthe manipulation module, a Deep Q Network (DQN) model is designed to generate\nmanipulation actions for the robot to interact with the environment. We\nconsider the situation where the robot continuously manipulating objects inside\na bin until the answer to the question is found. Besides, a novel dataset that\ncontains a variety of object models, scenarios and corresponding\nquestion-answer pairs is established in a simulation environment. Extensive\nexperiments have been conducted to validate the effectiveness of the proposed\nframework.', 'To achieve seamless human-robot interactions, robots need to intimately\nreason about complex interaction dynamics and future human behaviors within\ntheir motion planning process. However, there is a disconnect between\nstate-of-the-art neural network-based human behavior models and robot motion\nplanners -- either the behavior models are limited in their consideration of\ndownstream planning or a simplified behavior model is used to ensure\ntractability of the planning problem. In this work, we present a framework that\nfuses together the interpretability and flexibility of trajectory optimization\n(TO) with the predictive power of state-of-the-art human trajectory prediction\nmodels. In particular, we leverage gradient information from data-driven\nprediction models to explicitly reason about human-robot interaction dynamics\nwithin a gradient-based TO problem. We demonstrate the efficacy of our approach\nin a multi-agent scenario whereby a robot is required to safely and efficiently\nnavigate through a crowd of up to ten pedestrians. We compare against a variety\nof planning methods, and show that by explicitly accounting for interaction\ndynamics within the planner, our method offers safer and more efficient\nbehaviors, even yielding proactive and nuanced behaviors such as waiting for a\npedestrian to pass before moving.', 'The study of human-robot interaction is fundamental to the design and use of\nrobotics in real-world applications. Robots will need to predict and adapt to\nthe actions of human collaborators in order to achieve good performance and\nimprove safety and end-user adoption. This paper evaluates a human-robot\ncollaboration scheme that combines the task allocation and motion levels of\nreasoning: the robotic agent uses Bayesian inference to predict the next goal\nof its human partner from his or her ongoing motion, and re-plans its own\nactions in real time. This anticipative adaptation is desirable in many\npractical scenarios, where humans are unable or unwilling to take on the\ncognitive overhead required to explicitly communicate their intent to the\nrobot. A behavioral experiment indicates that the combination of goal inference\nand dynamic task planning significantly improves both objective and perceived\nperformance of the human-robot team. Participants were highly sensitive to the\ndifferences between robot behaviors, preferring to work with a robot that\nadapted to their actions over one that did not.', 'Learning predictive models from interaction with the world allows an agent,\nsuch as a robot, to learn about how the world works, and then use this learned\nmodel to plan coordinated sequences of actions to bring about desired outcomes.\nHowever, learning a model that captures the dynamics of complex skills\nrepresents a major challenge: if the agent needs a good model to perform these\nskills, it might never be able to collect the experience on its own that is\nrequired to learn these delicate and complex behaviors. Instead, we can imagine\naugmenting the training set with observational data of other agents, such as\nhumans. Such data is likely more plentiful, but represents a different\nembodiment. For example, videos of humans might show a robot how to use a tool,\nbut (i) are not annotated with suitable robot actions, and (ii) contain a\nsystematic distributional shift due to the embodiment differences between\nhumans and robots. We address the first challenge by formulating the\ncorresponding graphical model and treating the action as an observed variable\nfor the interaction data and an unobserved variable for the observation data,\nand the second challenge by using a domain-dependent prior. In addition to\ninteraction data, our method is able to leverage videos of passive observations\nin a driving dataset and a dataset of robotic manipulation videos. A robotic\nplanning agent equipped with our method can learn to use tools in a tabletop\nrobotic manipulation setting by observing humans without ever seeing a robotic\nvideo of tool use.', 'Learning from Demonstration (LfD) constitutes one of the most robust\nmethodologies for constructing efficient cognitive robotic systems. Despite the\nlarge body of research works already reported, current key technological\nchallenges include those of multi-agent learning and long-term autonomy.\nTowards this direction, a novel cognitive architecture for multi-agent LfD\nrobotic learning is introduced, targeting to enable the reliable deployment of\nopen, scalable and expandable robotic systems in large-scale and complex\nenvironments. In particular, the designed architecture capitalizes on the\nrecent advances in the Artificial Intelligence (AI) field, by establishing a\nFederated Learning (FL)-based framework for incarnating a multi-human\nmulti-robot collaborative learning environment. The fundamental\nconceptualization relies on employing multiple AI-empowered cognitive processes\n(implementing various robotic tasks) that operate at the edge nodes of a\nnetwork of robotic platforms, while global AI models (underpinning the\naforementioned robotic tasks) are collectively created and shared among the\nnetwork, by elegantly combining information from a large number of human-robot\ninteraction instances. Regarding pivotal novelties, the designed cognitive\narchitecture a) introduces a new FL-based formalism that extends the\nconventional LfD learning paradigm to support large-scale multi-agent\noperational settings, b) elaborates previous FL-based self-learning robotic\nschemes so as to incorporate the human in the learning loop and c) consolidates\nthe fundamental principles of FL with additional sophisticated AI-enabled\nlearning methodologies for modelling the multi-level inter-dependencies among\nthe robotic tasks. The applicability of the proposed framework is explained\nusing an example of a real-world industrial case study for agile\nproduction-based Critical Raw Materials (CRM) recovery.', 'In open-ended continuous environments, robots need to learn multiple\nparameterised control tasks in hierarchical reinforcement learning. We\nhypothesise that the most complex tasks can be learned more easily by\ntransferring knowledge from simpler tasks, and faster by adapting the\ncomplexity of the actions to the task. We propose a task-oriented\nrepresentation of complex actions, called procedures, to learn online task\nrelationships and unbounded sequences of action primitives to control the\ndifferent observables of the environment. Combining both goal-babbling with\nimitation learning, and active learning with transfer of knowledge based on\nintrinsic motivation, our algorithm self-organises its learning process. It\nchooses at any given time a task to focus on; and what, how, when and from whom\nto transfer knowledge. We show with a simulation and a real industrial robot\narm, in cross-task and cross-learner transfer settings, that task composition\nis key to tackle highly complex tasks. Task decomposition is also efficiently\ntransferred across different embodied learners and by active imitation, where\nthe robot requests just a small amount of demonstrations and the adequate type\nof information. The robot learns and exploits task dependencies so as to learn\ntasks of every complexity.', ""The great diversity of end-user tasks ranging from manufacturing environments\nto personal homes makes pre-programming robots for general purpose applications\nextremely challenging. In fact, teaching robots new actions from scratch that\ncan be reused for previously unseen tasks remains a difficult challenge and is\ngenerally left up to robotics experts. In this work, we present iRoPro, an\ninteractive Robot Programming framework that allows end-users with little to no\ntechnical background to teach a robot new reusable actions. We combine\nProgramming by Demonstration and Automated Planning techniques to allow the\nuser to construct the robot's knowledge base by teaching new actions by\nkinesthetic demonstration. The actions are generalised and reused with a task\nplanner to solve previously unseen problems defined by the user. We implement\niRoPro as an end-to-end system on a Baxter Research Robot to simultaneously\nteach low- and high-level actions by demonstration that the user can customise\nvia a Graphical User Interface to adapt to their specific use case. To evaluate\nthe feasibility of our approach, we first conducted pre-design experiments to\nbetter understand the user's adoption of involved concepts and the proposed\nrobot programming process. We compare results with post-design experiments,\nwhere we conducted a user study to validate the usability of our approach with\nreal end-users. Overall, we showed that users with different programming levels\nand educational backgrounds can easily learn and use iRoPro and its robot\nprogramming process."", 'Intelligent robots need to achieve abstract objectives using concrete,\nspatiotemporally complex sensory information and motor control. Tabula rasa\ndeep reinforcement learning (RL) has tackled demanding tasks in terms of either\nvisual, abstract, or physical reasoning, but solving these jointly remains a\nformidable challenge. One recent, unsolved benchmark task that integrates these\nchallenges is Mujoban, where a robot needs to arrange 3D warehouses generated\nfrom 2D Sokoban puzzles. We explore whether integrated tasks like Mujoban can\nbe solved by composing RL modules together in a sense-plan-act hierarchy, where\nmodules have well-defined roles similarly to classic robot architectures.\nUnlike classic architectures that are typically model-based, we use only\nmodel-free modules trained with RL or supervised learning. We find that our\nmodular RL approach dramatically outperforms the state-of-the-art monolithic RL\nagent on Mujoban. Further, learned modules can be reused when, e.g., using a\ndifferent robot platform to solve the same task. Together our results give\nstrong evidence for the importance of research into modular RL designs. Project\nwebsite: https://sites.google.com/view/modular-rl/', ""Service robots need to show appropriate social behaviour in order to be\ndeployed in social environments such as healthcare, education, retail, etc.\nSome of the main capabilities that robots should have are navigation and\nconversational skills. If the person is impatient, the person might want a\nrobot to navigate faster and vice versa. Linguistic features that indicate\npoliteness can provide social cues about a person's patient and impatient\nbehaviour. The novelty presented in this paper is to dynamically incorporate\npoliteness in robotic dialogue systems for navigation. Understanding the\npoliteness in users' speech can be used to modulate the robot behaviour and\nresponses. Therefore, we developed a dialogue system to navigate in an indoor\nenvironment, which produces different robot behaviours and responses based on\nusers' intention and degree of politeness. We deploy and test our system with\nthe Pepper robot that adapts to the changes in user's politeness."", 'Reinforcement learning methods can achieve significant performance but\nrequire a large amount of training data collected on the same robotic platform.\nA policy trained with expensive data is rendered useless after making even a\nminor change to the robot hardware. In this paper, we address the challenging\nproblem of adapting a policy, trained to perform a task, to a novel robotic\nhardware platform given only few demonstrations of robot motion trajectories on\nthe target robot. We formulate it as a few-shot meta-learning problem where the\ngoal is to find a meta-model that captures the common structure shared across\ndifferent robotic platforms such that data-efficient adaptation can be\nperformed. We achieve such adaptation by introducing a learning framework\nconsisting of a probabilistic gradient-based meta-learning algorithm that\nmodels the uncertainty arising from the few-shot setting with a low-dimensional\nlatent variable. We experimentally evaluate our framework on a simulated\nreaching and a real-robot picking task using 400 simulated robots generated by\nvarying the physical parameters of an existing set of robotic platforms. Our\nresults show that the proposed method can successfully adapt a trained policy\nto different robotic platforms with novel physical parameters and the\nsuperiority of our meta-learning algorithm compared to state-of-the-art methods\nfor the introduced few-shot policy adaptation problem.', 'Deep artificial neural networks, trained with labeled data sets are widely\nused in numerous vision and robotics applications today. In terms of AI, these\nare called reflex models, referring to the fact that they do not self-evolve or\nactively adapt to environmental changes. As demand for intelligent robot\ncontrol expands to many high level tasks, reinforcement learning and state\nbased models play an increasingly important role. Herein, in computer vision\nand robotics domain, we study a novel approach to add reinforcement controls\nonto the image recognition reflex models to attain better overall performance,\nspecifically to a wider environment range beyond what is expected of the task\nreflex models. Follow a common infrastructure with environment sensing and AI\nbased modeling of self-adaptive agents, we implement multiple types of AI\ncontrol agents. To the end, we provide comparative results of these agents with\nbaseline, and an insightful analysis of their benefit to improve overall image\nrecognition performance in real world.']","['One spectrum on which robotic control paradigms lie is the degree in which a\nmodel of the environment is involved, from methods that are completely\nmodel-free such as model-free RL, to methods that require a known model such as\noptimal control, with other methods such as model-based RL somewhere in the\nmiddle. On one end of the spectrum, model-free RL can learn control policies\nfor high-dimensional (hi-dim), complex robotic tasks through trial-and-error\nwithout knowledge of a model of the environment, but tends to require a large\namount of data. On the other end, ""classical methods"" such as optimal control\ngenerate solutions without collecting data, but assume that an accurate model\nof the system and environment is known and are mostly limited to problems with\nlow-dimensional (lo-dim) state spaces. In this paper, we bring the two ends of\nthe spectrum together. Although models of hi-dim systems and environments may\nnot exist, lo-dim approximations of these systems and environments are widely\navailable, especially in robotics. Therefore, we propose to solve hi-dim,\ncomplex robotic tasks in two stages. First, assuming a coarse model of the\nhi-dim system, we compute a lo-dim value function for the lo-dim version of the\nproblem using classical methods (eg. value iteration and optimal control).\nThen, the lo-dim value function is used as a baseline function to warm-start\nthe model-free RL process that learns hi-dim policies. The lo-dim value\nfunction provides global guidance for model-free RL, alleviating the data\ninefficiency of model-free RL. We demonstrate our approach on two robot\nlearning tasks with hi-dim state spaces and observe significant improvement in\npolicy performance and learning efficiency. We also give an empirical analysis\nof our method with a third task.', 'To solve its task, a robot needs to have the ability to interpret its\nperceptions. In vision, this interpretation is particularly difficult and\nrelies on the understanding of the structure of the scene, at least to the\nextent of its task and sensorimotor abilities. A robot with the ability to\nbuild and adapt this interpretation process according to its own tasks and\ncapabilities would push away the limits of what robots can achieve in a non\ncontrolled environment. A solution is to provide the robot with processes to\nbuild such representations that are not specific to an environment or a\nsituation. A lot of works focus on objects segmentation, recognition and\nmanipulation. Defining an object solely on the basis of its visual appearance\nis challenging given the wide range of possible objects and environments.\nTherefore, current works make simplifying assumptions about the structure of a\nscene. Such assumptions reduce the adaptivity of the object extraction process\nto the environments in which the assumption holds. To limit such assumptions,\nwe introduce an exploration method aimed at identifying moveable elements in a\nscene without considering the concept of object. By using the interactive\nperception framework, we aim at bootstrapping the acquisition process of a\nrepresentation of the environment with a minimum of context specific\nassumptions. The robotic system builds a perceptual map called relevance map\nwhich indicates the moveable parts of the current scene. A classifier is\ntrained online to predict the category of each region (moveable or\nnon-moveable). It is also used to select a region with which to interact, with\nthe goal of minimizing the uncertainty of the classification. A specific\nclassifier is introduced to fit these needs: the collaborative mixture models\nclassifier. The method is tested on a set of scenarios of increasing\ncomplexity, using both simulations and a PR2 robot.', ""Robotic systems are more present in our society everyday. In human-robot\nenvironments, it is crucial that end-users may correctly understand their\nrobotic team-partners, in order to collaboratively complete a task. To increase\naction understanding, users demand more explainability about the decisions by\nthe robot in particular situations. Recently, explainable robotic systems have\nemerged as an alternative focused not only on completing a task satisfactorily,\nbut also on justifying, in a human-like manner, the reasons that lead to making\na decision. In reinforcement learning scenarios, a great effort has been\nfocused on providing explanations using data-driven approaches, particularly\nfrom the visual input modality in deep learning-based systems. In this work, we\nfocus rather on the decision-making process of reinforcement learning agents\nperforming a task in a robotic scenario. Experimental results are obtained\nusing 3 different set-ups, namely, a deterministic navigation task, a\nstochastic navigation task, and a continuous visual-based sorting object task.\nAs a way to explain the goal-driven robot's actions, we use the probability of\nsuccess computed by three different proposed approaches: memory-based,\nlearning-based, and introspection-based. The difference between these\napproaches is the amount of memory required to compute or estimate the\nprobability of success as well as the kind of reinforcement learning\nrepresentation where they could be used. In this regard, we use the\nmemory-based approach as a baseline since it is obtained directly from the\nagent's observations. When comparing the learning-based and the\nintrospection-based approaches to this baseline, both are found to be suitable\nalternatives to compute the probability of success, obtaining high levels of\nsimilarity when compared using both the Pearson's correlation and the mean\nsquared error."", 'Robotic systems are ever more capable of automation and fulfilment of complex\ntasks, particularly with reliance on recent advances in intelligent systems,\ndeep learning and artificial intelligence. However, as robots and humans come\ncloser in their interactions, the matter of interpretability, or explainability\nof robot decision-making processes for the human grows in importance. A\nsuccessful interaction and collaboration will only take place through mutual\nunderstanding of underlying representations of the environment and the task at\nhand. This is currently a challenge in deep learning systems. We present a\nhierarchical deep reinforcement learning system, consisting of a low-level\nagent handling the large actions/states space of a robotic system efficiently,\nby following the directives of a high-level agent which is learning the\nhigh-level dynamics of the environment and task. This high-level agent forms a\nrepresentation of the world and task at hand that is interpretable for a human\noperator. The method, which we call Dot-to-Dot, is tested on a MuJoCo-based\nmodel of the Fetch Robotics Manipulator, as well as a Shadow Hand, to test its\nperformance. Results show efficient learning of complex actions/states spaces\nby the low-level agent, and an interpretable representation of the task and\ndecision-making process learned by the high-level agent.', 'The capability of a mobile robot to efficiently and safely perform complex\nmissions is limited by its knowledge of the environment, namely the situation.\nAdvanced reasoning, decision-making, and execution skills enable an intelligent\nagent to act autonomously in unknown environments. Situational Awareness (SA)\nis a fundamental capability of humans that has been deeply studied in various\nfields, such as psychology, military, aerospace, and education. Nevertheless,\nit has yet to be considered in robotics, which has focused on single\ncompartmentalized concepts such as sensing, spatial perception, sensor fusion,\nstate estimation, and Simultaneous Localization and Mapping (SLAM). Hence, the\npresent research aims to connect the broad multidisciplinary existing knowledge\nto pave the way for a complete SA system for mobile robotics that we deem\nparamount for autonomy. To this aim, we define the principal components to\nstructure a robotic SA and their area of competence. Accordingly, this paper\ninvestigates each aspect of SA, surveying the state-of-the-art robotics\nalgorithms that cover them, and discusses their current limitations.\nRemarkably, essential aspects of SA are still immature since the current\nalgorithmic development restricts their performance to only specific\nenvironments. Nevertheless, Artificial Intelligence (AI), particularly Deep\nLearning (DL), has brought new methods to bridge the gap that maintains these\nfields apart from the deployment to real-world scenarios. Furthermore, an\nopportunity has been discovered to interconnect the vastly fragmented space of\nrobotic comprehension algorithms through the mechanism of Situational Graph\n(S-Graph), a generalization of the well-known scene graph. Therefore, we\nfinally shape our vision for the future of robotic Situational Awareness by\ndiscussing interesting recent research directions.', 'Robot understanding of human intentions is essential for fluid human-robot\ninteraction. Intentions, however, cannot be directly observed and must be\ninferred from behaviors. We learn a model of adaptive human behavior\nconditioned on the intention as a latent variable. We then embed the human\nbehavior model into a principled probabilistic decision model, which enables\nthe robot to (i) explore actively in order to infer human intentions and (ii)\nchoose actions that maximize its performance. Furthermore, the robot learns\nfrom the demonstrated actions of human experts to further improve exploration.\nPreliminary experiments in simulation indicate that our approach, when applied\nto autonomous driving, improves the efficiency and safety of driving in common\ninteractive driving scenarios.', 'A pressing question when designing intelligent autonomous systems is how to\nintegrate the various subsystems concerned with complementary tasks. More\nspecifically, robotic vision must provide task-relevant information about the\nenvironment and the objects in it to various planning related modules. In most\nimplementations of the traditional Perception-Cognition-Action paradigm these\ntasks are treated as quasi-independent modules that function as black boxes for\neach other. It is our view that perception can benefit tremendously from a\ntight collaboration with cognition. We present RoboSherlock, a\nknowledge-enabled cognitive perception systems for mobile robots performing\nhuman-scale everyday manipulation tasks. In RoboSherlock, perception and\ninterpretation of realistic scenes is formulated as an unstructured information\nmanagement(UIM) problem. The application of the UIM principle supports the\nimplementation of perception systems that can answer task-relevant queries\nabout objects in a scene, boost object recognition performance by combining the\nstrengths of multiple perception algorithms, support knowledge-enabled\nreasoning about objects and enable automatic and knowledge-driven generation of\nprocessing pipelines. We demonstrate the potential of the proposed framework\nthrough feasibility studies of systems for real-world scene perception that\nhave been built on top of the framework.', 'Imitation Learning (IL) is a powerful paradigm to teach robots to perform\nmanipulation tasks by allowing them to learn from human demonstrations\ncollected via teleoperation, but has mostly been limited to single-arm\nmanipulation. However, many real-world tasks require multiple arms, such as\nlifting a heavy object or assembling a desk. Unfortunately, applying IL to\nmulti-arm manipulation tasks has been challenging -- asking a human to control\nmore than one robotic arm can impose significant cognitive burden and is often\nonly possible for a maximum of two robot arms. To address these challenges, we\npresent Multi-Arm RoboTurk (MART), a multi-user data collection platform that\nallows multiple remote users to simultaneously teleoperate a set of robotic\narms and collect demonstrations for multi-arm tasks. Using MART, we collected\ndemonstrations for five novel two and three-arm tasks from several\ngeographically separated users. From our data we arrived at a critical insight:\nmost multi-arm tasks do not require global coordination throughout its full\nduration, but only during specific moments. We show that learning from such\ndata consequently presents challenges for centralized agents that directly\nattempt to model all robot actions simultaneously, and perform a comprehensive\nstudy of different policy architectures with varying levels of centralization\non our tasks. Finally, we propose and evaluate a base-residual policy framework\nthat allows trained policies to better adapt to the mixed coordination setting\ncommon in multi-arm manipulation, and show that a centralized policy augmented\nwith a decentralized residual model outperforms all other models on our set of\nbenchmark tasks. Additional results and videos at\nhttps://roboturk.stanford.edu/multiarm .', 'In order for cooperative robots (""co-robots"") to respond to human behaviors\naccurately and efficiently in human-robot collaboration, interpretation of\nhuman actions, awareness of new situations, and appropriate decision making are\nall crucial abilities for co-robots. For this purpose, the human behaviors\nshould be interpreted by co-robots in the same manner as human peers. To\naddress this issue, a novel interpretability indicator is introduced so that\nrobot actions are appropriate to the current human behaviors. In addition, the\ncomplete consideration of all potential situations of a robot\'s environment is\nnearly impossible in real-world applications, making it difficult for the\nco-robot to act appropriately and safely in new scenarios. This is true even\nwhen the pretrained model is highly accurate in a known situation. For\neffective and safe teaming with humans, we introduce a new generalizability\nindicator that allows a co-robot to self-reflect and reason about when an\nobservation falls outside the co-robot\'s learned model. Based on topic modeling\nand two novel indicators, we propose a new Self-reflective Risk-aware\nArtificial Cognitive (SRAC) model. The co-robots are able to consider action\nrisks and identify new situations so that better decisions can be made.\nExperiments both using real-world datasets and on physical robots suggest that\nour SRAC model significantly outperforms the traditional methodology and\nenables better decision making in response to human activities.', ""In this study, we investigate how a robot can generate novel and creative\nactions from its own experience of learning basic actions. Inspired by a\nmachine learning approach to computational creativity, we propose a dynamic\nneural network model that can learn and generate robot's actions. We conducted\na set of simulation experiments with a humanoid robot. The results showed that\nthe proposed model was able to learn the basic actions and also to generate\nnovel actions by modulating and combining those learned actions. The analysis\non the neural activities illustrated that the ability to generate creative\nactions emerged from the model's nonlinear memory structure self-organized\nduring training. The results also showed that the different way of learning the\nbasic actions induced the self-organization of the memory structure with the\ndifferent characteristics, resulting in the generation of different levels of\ncreative actions. Our approach can be utilized in human-robot interaction in\nwhich a user can interactively explore the robot's memory to control its\nbehavior and also discover other novel actions."", ""Reinforcement learning (RL) can be used to create a tactical decision-making\nagent for autonomous driving. However, previous approaches only output\ndecisions and do not provide information about the agent's confidence in the\nrecommended actions. This paper investigates how a Bayesian RL technique, based\non an ensemble of neural networks with additional randomized prior functions\n(RPF), can be used to estimate the uncertainty of decisions in autonomous\ndriving. A method for classifying whether or not an action should be considered\nsafe is also introduced. The performance of the ensemble RPF method is\nevaluated by training an agent on a highway driving scenario. It is shown that\nthe trained agent can estimate the uncertainty of its decisions and indicate an\nunacceptable level when the agent faces a situation that is far from the\ntraining distribution. Furthermore, within the training distribution, the\nensemble RPF agent outperforms a standard Deep Q-Network agent. In this study,\nthe estimated uncertainty is used to choose safe actions in unknown situations.\nHowever, the uncertainty information could also be used to identify situations\nthat should be added to the training process."", 'Legged robots have significant potential to operate in highly unstructured\nenvironments. The design of locomotion control is, however, still challenging.\nCurrently, controllers must be either manually designed for specific robots and\ntasks, or automatically designed via machine learning methods that require long\ntraining times and yield large opaque controllers. Drawing inspiration from\nanimal locomotion, we propose a simple yet versatile modular neural control\nstructure with fast learning. The key advantages of our approach are that\nbehavior-specific control modules can be added incrementally to obtain\nincreasingly complex emergent locomotion behaviors, and that neural connections\ninterfacing with existing modules can be quickly and automatically learned. In\na series of experiments, we show how eight modules can be quickly learned and\nadded to a base control module to obtain emergent adaptive behaviors allowing a\nhexapod robot to navigate in complex environments. We also show that modules\ncan be added and removed during operation without affecting the functionality\nof the remaining controller. Finally, the control approach was successfully\ndemonstrated on a physical hexapod robot. Taken together, our study reveals a\nsignificant step towards fast automatic design of versatile neural locomotion\ncontrol for complex robotic systems.', ""This paper presents a distributed algorithm applicable to a wide range of\npractical multi-robot applications. In such multi-robot applications, the\nuser-defined objectives of the mission can be cast as a general optimization\nproblem, without explicit guidelines of the subtasks per different robot. Owing\nto the unknown environment, unknown robot dynamics, sensor nonlinearities,\netc., the analytic form of the optimization cost function is not available a\npriori. Therefore, standard gradient-descent-like algorithms are not applicable\nto these problems. To tackle this, we introduce a new algorithm that carefully\ndesigns each robot's subcost function, the optimization of which can accomplish\nthe overall team objective. Upon this transformation, we propose a distributed\nmethodology based on the cognitive-based adaptive optimization (CAO) algorithm,\nthat is able to approximate the evolution of each robot's cost function and to\nadequately optimize its decision variables (robot actions). The latter can be\nachieved by online learning only the problem-specific characteristics that\naffect the accomplishment of mission objectives. The overall, low-complexity\nalgorithm can straightforwardly incorporate any kind of operational constraint,\nis fault-tolerant, and can appropriately tackle time-varying cost functions. A\ncornerstone of this approach is that it shares the same convergence\ncharacteristics as those of block coordinate descent algorithms. The proposed\nalgorithm is evaluated in three heterogeneous simulation set-ups under multiple\nscenarios, against both general-purpose and problem-specific algorithms. Source\ncode is available at\nhttps://github.com/athakapo/A-distributed-plug-n-play-algorithm-for-multi-robot-applications."", ""Human-robot interaction often occurs in the form of instructions given from a\nhuman to a robot. For a robot to successfully follow instructions, a common\nrepresentation of the world and objects in it should be shared between humans\nand the robot so that the instructions can be grounded. Achieving this\nrepresentation can be done via learning, where both the world representation\nand the language grounding are learned simultaneously. However, in robotics\nthis can be a difficult task due to the cost and scarcity of data. In this\npaper, we tackle the problem by separately learning the world representation of\nthe robot and the language grounding. While this approach can address the\nchallenges in getting sufficient data, it may give rise to inconsistencies\nbetween both learned components. Therefore, we further propose Bayesian\nlearning to resolve such inconsistencies between the natural language grounding\nand a robot's world representation by exploiting spatio-relational information\nthat is implicitly present in instructions given by a human. Moreover, we\ndemonstrate the feasibility of our approach on a scenario involving a robotic\narm in the physical world."", 'Seamless Human-Robot Interaction is the ultimate goal of developing service\nrobotic systems. For this, the robotic agents have to understand their\nsurroundings to better complete a given task. Semantic scene understanding\nallows a robotic agent to extract semantic knowledge about the objects in the\nenvironment. In this work, we present a semantic scene understanding pipeline\nthat fuses 2D and 3D detection branches to generate a semantic map of the\nenvironment. The 2D mask proposals from state-of-the-art 2D detectors are\ninverse-projected to the 3D space and combined with 3D detections from point\nsegmentation networks. Unlike previous works that were evaluated on collected\ndatasets, we test our pipeline on an active photo-realistic robotic environment\n- BenchBot. Our novelty includes rectification of 3D proposals using projected\n2D detections and modality fusion based on object size. This work is done as\npart of the Robotic Vision Scene Understanding Challenge (RVSU). The\nperformance evaluation demonstrates that our pipeline has improved on baseline\nmethods without significant computational bottleneck.', 'We propose a developmental approach that allows a robot to interpret and\ndescribe the actions of human agents by reusing previous experience. The robot\nfirst learns the association between words and object affordances by\nmanipulating the objects in its environment. It then uses this information to\nlearn a mapping between its own actions and those performed by a human in a\nshared environment. It finally fuses the information from these two models to\ninterpret and describe human actions in light of its own experience. In our\nexperiments, we show that the model can be used flexibly to do inference on\ndifferent aspects of the scene. We can predict the effects of an action on the\nbasis of object properties. We can revise the belief that a certain action\noccurred, given the observed effects of the human action. In an early action\nrecognition fashion, we can anticipate the effects when the action has only\nbeen partially observed. By estimating the probability of words given the\nevidence and feeding them into a pre-defined grammar, we can generate relevant\ndescriptions of the scene. We believe that this is a step towards providing\nrobots with the fundamental skills to engage in social collaboration with\nhumans.', 'Despite the success of reinforcement learning methods, they have yet to have\ntheir breakthrough moment when applied to a broad range of robotic manipulation\ntasks. This is partly due to the fact that reinforcement learning algorithms\nare notoriously difficult and time consuming to train, which is exacerbated\nwhen training from images rather than full-state inputs. As humans perform\nmanipulation tasks, our eyes closely monitor every step of the process with our\ngaze focusing sequentially on the objects being manipulated. With this in mind,\nwe present our Attention-driven Robotic Manipulation (ARM) algorithm, which is\na general manipulation algorithm that can be applied to a range of\nsparse-rewarded tasks, given only a small number of demonstrations. ARM splits\nthe complex task of manipulation into a 3 stage pipeline: (1) a Q-attention\nagent extracts relevant pixel locations from RGB and point cloud inputs, (2) a\nnext-best pose agent that accepts crops from the Q-attention agent and outputs\nposes, and (3) a control agent that takes the goal pose and outputs joint\nactions. We show that current learning algorithms fail on a range of RLBench\ntasks, whilst ARM is successful.', ""The intuitive collaboration of humans and intelligent robots (embodied AI) in\nthe real-world is an essential objective for many desirable applications of\nrobotics. Whilst there is much research regarding explicit communication, we\nfocus on how humans and robots interact implicitly, on motor adaptation level.\nWe present a real-world setup of a human-robot collaborative maze game,\ndesigned to be non-trivial and only solvable through collaboration, by limiting\nthe actions to rotations of two orthogonal axes, and assigning each axes to one\nplayer. This results in neither the human nor the agent being able to solve the\ngame on their own. We use deep reinforcement learning for the control of the\nrobotic agent, and achieve results within 30 minutes of real-world play,\nwithout any type of pre-training. We then use this setup to perform systematic\nexperiments on human/agent behaviour and adaptation when co-learning a policy\nfor the collaborative game. We present results on how co-policy learning occurs\nover time between the human and the robotic agent resulting in each\nparticipant's agent serving as a representation of how they would play the\ngame. This allows us to relate a person's success when playing with different\nagents than their own, by comparing the policy of the agent with that of their\nown agent."", 'Human-robot interaction can be divided into two categories based on the\nphysical distance between the human and robot: remote and proximal. In proximal\ninteraction, the human and robot often engage in close coordination; in remote\ninteraction, the human and robot are less coupled due to communication\nconstraints. As a result, providing automation for the robot in remote\ninteraction becomes more important. Thus far, human factor studies on\nautomation in remote human-robot interaction have been restricted to various\nforms of supervision, in which the robot is essentially being used as a smart\nmobile manipulation platform with sensing capabilities. In this paper, we\ninvestigate the incorporation of general planning capability into the robot to\nfacilitate peer-to-peer human-robot teaming, in which the human and robot are\nviewed as teammates that are physically separated. The human and robot share\nthe same global goal and collaborate to achieve it. Note that humans may feel\nuncomfortable at such robot autonomy, which can potentially reduce teaming\nperformance. One important difference between peer-to-peer teaming and\nsupervised teaming is that an autonomous robot in peer-to-peer teaming can\nachieve the goal alone when the task information is completely specified.\nHowever, incompleteness often exists, which implies information asymmetry.\nWhile information asymmetry can be desirable sometimes, it may also lead to the\nrobot choosing improper actions that negatively influence the teaming\nperformance. We aim to investigate the various trade-offs, e.g., mental\nworkload and situation awareness, between these two types of remote human-robot\nteaming.', 'In this paper we present a Bayesian reinforcement learning framework that\nallows robotic manipulators to adaptively recover from random mechanical\nfailures autonomously, hence being survivable. To this end, we formulate the\nframework of Bayesian Policy Morphing (BPM) that enables a robot agent to\nself-modify its learned policy after the diminution of its maneuvering\ndimensionality. We build upon existing actor-critic framework, and extend it to\nperform policy gradient updates as posterior learning, taking past policy\nupdates as prior distributions. We show that policy search, in the direction\nbiased by prior experience, significantly improves learning efficiency in terms\nof sampling requirements. We demonstrate our results on an 8-DOF robotic arm\nwith our algorithm of BPM, while intentionally disabling random joints with\ndifferent damage types like unresponsive joints, constant offset errors and\nangular imprecision. Our results have shown that, even with physical damages,\nthe robotic arm can still successfully maintain its functionality to accurately\nlocate and grasp a given target object.', 'As robots leave the controlled environments of factories to autonomously\nfunction in more complex, natural environments, they will have to respond to\nthe inevitable fact that they will become damaged. However, while animals can\nquickly adapt to a wide variety of injuries, current robots cannot ""think\noutside the box"" to find a compensatory behavior when damaged: they are limited\nto their pre-specified self-sensing abilities, can diagnose only anticipated\nfailure modes, and require a pre-programmed contingency plan for every type of\npotential damage, an impracticality for complex robots. Here we introduce an\nintelligent trial and error algorithm that allows robots to adapt to damage in\nless than two minutes, without requiring self-diagnosis or pre-specified\ncontingency plans. Before deployment, a robot exploits a novel algorithm to\ncreate a detailed map of the space of high-performing behaviors: This map\nrepresents the robot\'s intuitions about what behaviors it can perform and their\nvalue. If the robot is damaged, it uses these intuitions to guide a\ntrial-and-error learning algorithm that conducts intelligent experiments to\nrapidly discover a compensatory behavior that works in spite of the damage.\nExperiments reveal successful adaptations for a legged robot injured in five\ndifferent ways, including damaged, broken, and missing legs, and for a robotic\narm with joints broken in 14 different ways. This new technique will enable\nmore robust, effective, autonomous robots, and suggests principles that animals\nmay use to adapt to injury.', 'A common vision from science fiction is that robots will one day inhabit our\nphysical spaces, sense the world as we do, assist our physical labours, and\ncommunicate with us through natural language. Here we study how to design\nartificial agents that can interact naturally with humans using the\nsimplification of a virtual environment. This setting nevertheless integrates a\nnumber of the central challenges of artificial intelligence (AI) research:\ncomplex visual perception and goal-directed physical control, grounded language\ncomprehension and production, and multi-agent social interaction. To build\nagents that can robustly interact with humans, we would ideally train them\nwhile they interact with humans. However, this is presently impractical.\nTherefore, we approximate the role of the human with another learned agent, and\nuse ideas from inverse reinforcement learning to reduce the disparities between\nhuman-human and agent-agent interactive behaviour. Rigorously evaluating our\nagents poses a great challenge, so we develop a variety of behavioural tests,\nincluding evaluation by humans who watch videos of agents or interact directly\nwith them. These evaluations convincingly demonstrate that interactive training\nand auxiliary losses improve agent behaviour beyond what is achieved by\nsupervised learning of actions alone. Further, we demonstrate that agent\ncapabilities generalise beyond literal experiences in the dataset. Finally, we\ntrain evaluation models whose ratings of agents agree well with human\njudgement, thus permitting the evaluation of new agent models without\nadditional effort. Taken together, our results in this virtual environment\nprovide evidence that large-scale human behavioural imitation is a promising\ntool to create intelligent, interactive agents, and the challenge of reliably\nevaluating such agents is possible to surmount.', ""Predictions and predictive knowledge have seen recent success in improving\nnot only robot control but also other applications ranging from industrial\nprocess control to rehabilitation. A property that makes these predictive\napproaches well suited for robotics is that they can be learned online and\nincrementally through interaction with the environment. However, a remaining\nchallenge for many prediction-learning approaches is an appropriate choice of\nprediction-learning parameters, especially parameters that control the\nmagnitude of a learning machine's updates to its predictions (the learning rate\nor step size). To begin to address this challenge, we examine the use of online\nstep-size adaptation using a sensor-rich robotic arm. Our method of choice,\nTemporal-Difference Incremental Delta-Bar-Delta (TIDBD), learns and adapts step\nsizes on a feature level; importantly, TIDBD allows step-size tuning and\nrepresentation learning to occur at the same time. We show that TIDBD is a\npractical alternative for classic Temporal-Difference (TD) learning via an\nextensive parameter search. Both approaches perform comparably in terms of\npredicting future aspects of a robotic data stream. Furthermore, the use of a\nstep-size adaptation method like TIDBD appears to allow a system to\nautomatically detect and characterize common sensor failures in a robotic\napplication. Together, these results promise to improve the ability of robotic\ndevices to learn from interactions with their environments in a robust way,\nproviding key capabilities for autonomous agents and robots."", ""Human input has enabled autonomous systems to improve their capabilities and\nachieve complex behaviors that are otherwise challenging to generate\nautomatically. Recent work focuses on how robots can use such input - like\ndemonstrations or corrections - to learn intended objectives. These techniques\nassume that the human's desired objective already exists within the robot's\nhypothesis space. In reality, this assumption is often inaccurate: there will\nalways be situations where the person might care about aspects of the task that\nthe robot does not know about. Without this knowledge, the robot cannot infer\nthe correct objective. Hence, when the robot's hypothesis space is\nmisspecified, even methods that keep track of uncertainty over the objective\nfail because they reason about which hypothesis might be correct, and not\nwhether any of the hypotheses are correct. In this paper, we posit that the\nrobot should reason explicitly about how well it can explain human inputs given\nits hypothesis space and use that situational confidence to inform how it\nshould incorporate human input. We demonstrate our method on a 7\ndegree-of-freedom robot manipulator in learning from two important types of\nhuman input: demonstrations of manipulation tasks, and physical corrections\nduring the robot's task execution."", 'The increasing presence of robots alongside humans, such as in human-robot\nteams in manufacturing, gives rise to research questions about the kind of\nbehaviors people prefer in their robot counterparts. We term actions that\nsupport interaction by reducing future interference with others as supportive\nrobot actions and investigate their utility in a co-located manipulation\nscenario. We compare two robot modes in a shared table pick-and-place task: (1)\nTask-oriented: the robot only takes actions to further its own task objective\nand (2) Supportive: the robot sometimes prefers supportive actions to\ntask-oriented ones when they reduce future goal-conflicts. Our experiments in\nsimulation, using a simplified human model, reveal that supportive actions\nreduce the interference between agents, especially in more difficult tasks, but\nalso cause the robot to take longer to complete the task. We implemented these\nmodes on a physical robot in a user study where a human and a robot perform\nobject placement on a shared table. Our results show that a supportive robot\nwas perceived as a more favorable coworker by the human and also reduced\ninterference with the human in the more difficult of two scenarios. However, it\nalso took longer to complete the task highlighting an interesting trade-off\nbetween task-efficiency and human-preference that needs to be considered before\ndesigning robot behavior for close-proximity manipulation scenarios.', 'Autonomy is fundamental for artificial agents acting in complex real-world\nscenarios. The acquisition of many different skills is pivotal to foster\nversatile autonomous behaviour and thus a main objective for robotics and\nmachine learning. Intrinsic motivations have proven to properly generate a\ntask-agnostic signal to drive the autonomous acquisition of multiple policies\nin settings requiring the learning of multiple tasks. However, in real world\nscenarios tasks may be interdependent so that some of them may constitute the\nprecondition for learning other ones. Despite different strategies have been\nused to tackle the acquisition of interdependent/hierarchical tasks, fully\nautonomous open-ended learning in these scenarios is still an open question.\nBuilding on previous research within the framework of intrinsically-motivated\nopen-ended learning, we propose an architecture for robot control that tackles\nthis problem from the point of view of decision making, i.e. treating the\nselection of tasks as a Markov Decision Process where the system selects the\npolicies to be trained in order to maximise its competence over all the tasks.\nThe system is then tested with a humanoid robot solving interdependent multiple\nreaching tasks.', 'The ability to plan and execute goal specific actions in varied, unexpected\nsettings is a central requirement of intelligent agents. In this paper, we\nexplore how an agent can be equipped with an internal model of the dynamics of\nthe external world, and how it can use this model to plan novel actions by\nrunning multiple internal simulations (""visual imagination""). Our models\ndirectly process raw visual input, and use a novel object-centric prediction\nformulation based on visual glimpses centered on objects (fixations) to enforce\ntranslational invariance of the learned physical laws. The agent gathers\ntraining data through random interaction with a collection of different\nenvironments, and the resulting model can then be used to plan goal-directed\nactions in novel environments that the agent has not seen before. We\ndemonstrate that our agent can accurately plan actions for playing a simulated\nbilliards game, which requires pushing a ball into a target position or into\ncollision with another ball.', ""A major component for developing intelligent and autonomous robots is a\nsuitable knowledge representation, from which a robot can acquire knowledge\nabout its actions or world. However, unlike humans, robots cannot creatively\nadapt to novel scenarios, as their knowledge and environment are rigidly\ndefined. To address the problem of producing novel and flexible task plans\ncalled task trees, we explore how we can derive plans with concepts not\noriginally in the robot's knowledge base. Existing knowledge in the form of a\nknowledge graph is used as a base of reference to create task trees that are\nmodified with new object or state combinations. To demonstrate the flexibility\nof our method, we randomly selected recipes from the Recipe1M+ dataset and\ngenerated their task trees. The task trees were then thoroughly checked with a\nvisualization tool that portrays how each ingredient changes with each action\nto produce the desired meal. Our results indicate that the proposed method can\nproduce task plans with high accuracy even for never-before-seen ingredient\ncombinations."", 'Intelligent behaviour in the physical world exhibits structure at multiple\nspatial and temporal scales. Although movements are ultimately executed at the\nlevel of instantaneous muscle tensions or joint torques, they must be selected\nto serve goals defined on much longer timescales, and in terms of relations\nthat extend far beyond the body itself, ultimately involving coordination with\nother agents. Recent research in artificial intelligence has shown the promise\nof learning-based approaches to the respective problems of complex movement,\nlonger-term planning and multi-agent coordination. However, there is limited\nresearch aimed at their integration. We study this problem by training teams of\nphysically simulated humanoid avatars to play football in a realistic virtual\nenvironment. We develop a method that combines imitation learning, single- and\nmulti-agent reinforcement learning and population-based training, and makes use\nof transferable representations of behaviour for decision making at different\nlevels of abstraction. In a sequence of stages, players first learn to control\na fully articulated body to perform realistic, human-like movements such as\nrunning and turning; they then acquire mid-level football skills such as\ndribbling and shooting; finally, they develop awareness of others and play as a\nteam, bridging the gap between low-level motor control at a timescale of\nmilliseconds, and coordinated goal-directed behaviour as a team at the\ntimescale of tens of seconds. We investigate the emergence of behaviours at\ndifferent levels of abstraction, as well as the representations that underlie\nthese behaviours using several analysis techniques, including statistics from\nreal-world sports analytics. Our work constitutes a complete demonstration of\nintegrated decision-making at multiple scales in a physically embodied\nmulti-agent setting. See project video at https://youtu.be/KHMwq9pv7mg.', ""This paper contributes a preliminary report on the advantages and\ndisadvantages of incorporating simultaneous human control and feedback signals\nin the training of a reinforcement learning robotic agent. While robotic\nhuman-machine interfaces have become increasingly complex in both form and\nfunction, control remains challenging for users. This has resulted in an\nincreasing gap between user control approaches and the number of robotic motors\nwhich can be controlled. One way to address this gap is to shift some autonomy\nto the robot. Semi-autonomous actions of the robotic agent can then be shaped\nby human feedback, simplifying user control. Most prior work on agent shaping\nby humans has incorporated training with feedback, or has included indirect\ncontrol signals. By contrast, in this paper we explore how a human can provide\nconcurrent feedback signals and real-time myoelectric control signals to train\na robot's actor-critic reinforcement learning control system. Using both a\nphysical and a simulated robotic system, we compare training performance on a\nsimple movement task when reward is derived from the environment, when reward\nis provided by the human, and combinations of these two approaches. Our results\nindicate that some benefit can be gained with the inclusion of human generated\nfeedback."", ""Designing robots capable of generating interpretable behavior is a\nprerequisite for achieving effective human-robot collaboration. This means that\nthe robots need to be capable of generating behavior that aligns with human\nexpectations and, when required, provide explanations to the humans in the\nloop. However, exhibiting such behavior in arbitrary environments could be\nquite expensive for robots, and in some cases, the robot may not even be able\nto exhibit the expected behavior. Given structured environments (like\nwarehouses and restaurants), it may be possible to design the environment so as\nto boost the interpretability of the robot's behavior or to shape the human's\nexpectations of the robot's behavior. In this paper, we investigate the\nopportunities and limitations of environment design as a tool to promote a type\nof interpretable behavior -- known in the literature as explicable behavior. We\nformulate a novel environment design framework that considers design over\nmultiple tasks and over a time horizon. In addition, we explore the\nlongitudinal aspect of explicable behavior and the trade-off that arises\nbetween the cost of design and the cost of generating explicable behavior over\na time horizon."", 'Nowadays, model-free reinforcement learning algorithms have achieved\nremarkable performance on many decision making and control tasks, but high\nsample complexity and low sample efficiency still hinder the wide use of\nmodel-free reinforcement learning algorithms. In this paper, we argue that if\nwe intend to design an intelligent agent that learns fast and transfers well,\nthe agent must be able to reflect key elements of intelligence, like intuition,\nMemory, PredictionandCuriosity. We propose an agent framework that integrates\noff-policy reinforcement learning with world model learning, so as to embody\nthe important features of intelligence in our algorithm design. We adopt the\nstate-of-art model-free reinforcement learning algorithm, Soft Actor-Critic, as\nthe agent intuition, and world model learning through RNN to endow the agent\nwith memory, curiosity, and the ability to predict. We show that these ideas\ncan work collaboratively with each other and our agent (RMC) can give new\nstate-of-art results while maintaining sample efficiency and training\nstability. Moreover, our agent framework can be easily extended from MDP to\nPOMDP problems without performance loss.', ""We design a simple reinforcement learning (RL) agent that implements an\noptimistic version of $Q$-learning and establish through regret analysis that\nthis agent can operate with some level of competence in any environment. While\nwe leverage concepts from the literature on provably efficient RL, we consider\na general agent-environment interface and provide a novel agent design and\nanalysis. This level of generality positions our results to inform the design\nof future agents for operation in complex real environments. We establish that,\nas time progresses, our agent performs competitively relative to policies that\nrequire longer times to evaluate. The time it takes to approach asymptotic\nperformance is polynomial in the complexity of the agent's state representation\nand the time required to evaluate the best policy that the agent can represent.\nNotably, there is no dependence on the complexity of the environment. The\nultimate per-period performance loss of the agent is bounded by a constant\nmultiple of a measure of distortion introduced by the agent's state\nrepresentation. This work is the first to establish that an algorithm\napproaches this asymptotic condition within a tractable time frame."", 'A practical approach to robot reinforcement learning is to first collect a\nlarge batch of real or simulated robot interaction data, using some data\ncollection policy, and then learn from this data to perform various tasks,\nusing offline learning algorithms. Previous work focused on manually designing\nthe data collection policy, and on tasks where suitable policies can easily be\ndesigned, such as random picking policies for collecting data about object\ngrasping. For more complex tasks, however, it may be difficult to find a data\ncollection policy that explores the environment effectively, and produces data\nthat is diverse enough for the downstream task. In this work, we propose that\ndata collection policies should actively explore the environment to collect\ndiverse data. In particular, we develop a simple-yet-effective goal-conditioned\nreinforcement-learning method that actively focuses data collection on novel\nobservations, thereby collecting a diverse data-set. We evaluate our method on\nsimulated robot manipulation tasks with visual inputs and show that the\nimproved diversity of active data collection leads to significant improvements\nin the downstream learning tasks.', 'Well structured visual representations can make robot learning faster and can\nimprove generalization. In this paper, we study how we can acquire effective\nobject-centric representations for robotic manipulation tasks without human\nlabeling by using autonomous robot interaction with the environment. Such\nrepresentation learning methods can benefit from continuous refinement of the\nrepresentation as the robot collects more experience, allowing them to scale\neffectively without human intervention. Our representation learning approach is\nbased on object persistence: when a robot removes an object from a scene, the\nrepresentation of that scene should change according to the features of the\nobject that was removed. We formulate an arithmetic relationship between\nfeature vectors from this observation, and use it to learn a representation of\nscenes and objects that can then be used to identify object instances, localize\nthem in the scene, and perform goal-directed grasping tasks where the robot\nmust retrieve commanded objects from a bin. The same grasping procedure can\nalso be used to automatically collect training data for our method, by\nrecording images of scenes, grasping and removing an object, and recording the\noutcome. Our experiments demonstrate that this self-supervised approach for\ntasked grasping substantially outperforms direct reinforcement learning from\nimages and prior representation learning methods.', 'Telepresence robots are used in various forms in various use-cases that helps\nto avoid physical human presence at the scene of action. In this work, we focus\non a telepresence robot that can be used to attend a meeting remotely with a\ngroup of people. Unlike a one-to-one meeting, participants in a group meeting\ncan be located at a different part of the room, especially in an informal\nsetup. As a result, all of them may not be at the viewing angle of the robot,\na.k.a. the remote participant. In such a case, to provide a better meeting\nexperience, the robot should localize the speaker and bring the speaker at the\ncenter of the viewing angle. Though sound source localization can easily be\ndone using a microphone-array, bringing the speaker or set of speakers at the\nviewing angle is not a trivial task. First of all, the robot should react only\nto a human voice, but not to the random noises. Secondly, if there are multiple\nspeakers, to whom the robot should face or should it rotate continuously with\nevery new speaker? Lastly, most robotic platforms are resource-constrained and\nto achieve a real-time response, i.e., avoiding network delay, all the\nalgorithms should be implemented within the robot itself. This article presents\na study and implementation of an attention shifting scheme in a telepresence\nmeeting scenario which best suits the needs and expectations of the collocated\nand remote attendees. We define a policy to decide when a robot should rotate\nand how much based on real-time speaker localization. Using user satisfaction\nstudy, we show the efficacy and usability of our system in the meeting\nscenario. Moreover, our system can be easily adapted to other scenarios where\nmultiple people are located.', 'Humans make decisions and act alongside other humans to pursue both\nshort-term and long-term goals. As a result of ongoing progress in areas such\nas computing science and automation, humans now also interact with non-human\nagents of varying complexity as part of their day-to-day activities;\nsubstantial work is being done to integrate increasingly intelligent machine\nagents into human work and play. With increases in the cognitive, sensory, and\nmotor capacity of these agents, intelligent machinery for human assistance can\nnow reasonably be considered to engage in joint action with humans---i.e., two\nor more agents adapting their behaviour and their understanding of each other\nso as to progress in shared objectives or goals. The mechanisms, conditions,\nand opportunities for skillful joint action in human-machine partnerships is of\ngreat interest to multiple communities. Despite this, human-machine joint\naction is as yet under-explored, especially in cases where a human and an\nintelligent machine interact in a persistent way during the course of\nreal-time, daily-life experience. In this work, we contribute a virtual reality\nenvironment wherein a human and an agent can adapt their predictions, their\nactions, and their communication so as to pursue a simple foraging task. In a\ncase study with a single participant, we provide an example of human-agent\ncoordination and decision-making involving prediction learning on the part of\nthe human and the machine agent, and control learning on the part of the\nmachine agent wherein audio communication signals are used to cue its human\npartner in service of acquiring shared reward. These comparisons suggest the\nutility of studying human-machine coordination in a virtual reality\nenvironment, and identify further research that will expand our understanding\nof persistent human-machine joint action.', ""A common vision from science fiction is that robots will one day inhabit our\nphysical spaces, sense the world as we do, assist our physical labours, and\ncommunicate with us through natural language. Here we study how to design\nartificial agents that can interact naturally with humans using the\nsimplification of a virtual environment. We show that imitation learning of\nhuman-human interactions in a simulated world, in conjunction with\nself-supervised learning, is sufficient to produce a multimodal interactive\nagent, which we call MIA, that successfully interacts with non-adversarial\nhumans 75% of the time. We further identify architectural and algorithmic\ntechniques that improve performance, such as hierarchical action selection.\nAltogether, our results demonstrate that imitation of multi-modal, real-time\nhuman behaviour may provide a straightforward and surprisingly effective means\nof imbuing agents with a rich behavioural prior from which agents might then be\nfine-tuned for specific purposes, thus laying a foundation for training capable\nagents for interactive robots or digital assistants. A video of MIA's behaviour\nmay be found at https://youtu.be/ZFgRhviF7mY"", 'One of the long-term challenges of robotics is to enable robots to interact\nwith humans in the visual world via natural language, as humans are visual\nanimals that communicate through language. Overcoming this challenge requires\nthe ability to perform a wide variety of complex tasks in response to\nmultifarious instructions from humans. In the hope that it might drive progress\ntowards more flexible and powerful human interactions with robots, we propose a\ndataset of varied and complex robot tasks, described in natural language, in\nterms of objects visible in a large set of real images. Given an instruction,\nsuccess requires navigating through a previously-unseen environment to identify\nan object. This represents a practical challenge, but one that closely reflects\none of the core visual problems in robotics. Several state-of-the-art\nvision-and-language navigation, and referring-expression models are tested to\nverify the difficulty of this new task, but none of them show promising results\nbecause there are many fundamental differences between our task and previous\nones. A novel Interactive Navigator-Pointer model is also proposed that\nprovides a strong baseline on the task. The proposed model especially achieves\nthe best performance on the unseen test split, but still leaves substantial\nroom for improvement compared to the human performance.', 'The representation of the knowledge needed by a robot to perform complex\ntasks is restricted by the limitations of perception. One possible way of\novercoming this situation and designing ""knowledgeable"" robots is to rely on\nthe interaction with the user. We propose a multi-modal interaction framework\nthat allows to effectively acquire knowledge about the environment where the\nrobot operates. In particular, in this paper we present a rich representation\nframework that can be automatically built from the metric map annotated with\nthe indications provided by the user. Such a representation, allows then the\nrobot to ground complex referential expressions for motion commands and to\ndevise topological navigation plans to achieve the target locations.', 'In this paper, we propose an AI-FML robotic agent for student learning\nbehavior ontology construction which can be applied in English speaking and\nlistening domain. The AI-FML robotic agent with the ontology contains the\nperception intelligence, computational intelligence, and cognition intelligence\nfor analyzing student learning behavior. In addition, there are three\nintelligent agents, including a perception agent, a computational agent, and a\ncognition agent in the AI-FML robotic agent. We deploy the perception agent and\nthe cognition agent on the robot Kebbi Air. Moreover, the computational agent\nwith the Deep Neural Network (DNN) model is performed in the cloud and can\ncommunicate with the perception agent and cognition agent via the Internet. The\nproposed AI-FML robotic agent is applied in Taiwan and tested in Japan. The\nexperimental results show that the agents can be utilized in the human and\nmachine co-learning model for the future education.', 'When deployed, AI agents will encounter problems that are beyond their\nautonomous problem-solving capabilities. Leveraging human assistance can help\nagents overcome their inherent limitations and robustly cope with unfamiliar\nsituations. We present a general interactive framework that enables an agent to\nrequest and interpret rich, contextually useful information from an assistant\nthat has knowledge about the task and the environment. We demonstrate the\npracticality of our framework on a simulated human-assisted navigation problem.\nAided with an assistance-requesting policy learned by our method, a navigation\nagent achieves up to a 7x improvement in success rate on tasks that take place\nin previously unseen environments, compared to fully autonomous behavior. We\nshow that the agent can take advantage of different types of information\ndepending on the context, and analyze the benefits and challenges of learning\nthe assistance-requesting policy when the assistant can recursively decompose\ntasks into subtasks.', ""Seamlessly interacting with humans or robots is hard because these agents are\nnon-stationary. They update their policy in response to the ego agent's\nbehavior, and the ego agent must anticipate these changes to co-adapt. Inspired\nby humans, we recognize that robots do not need to explicitly model every\nlow-level action another agent will make; instead, we can capture the latent\nstrategy of other agents through high-level representations. We propose a\nreinforcement learning-based framework for learning latent representations of\nan agent's policy, where the ego agent identifies the relationship between its\nbehavior and the other agent's future strategy. The ego agent then leverages\nthese latent dynamics to influence the other agent, purposely guiding them\ntowards policies suitable for co-adaptation. Across several simulated domains\nand a real-world air hockey game, our approach outperforms the alternatives and\nlearns to influence the other agent."", 'Robot navigation in human semi-static and crowded environments can lead to\nthe freezing problem, where the robot can not move due to the presence of\nhumans standing on its path and no other path is available. Classical\napproaches of robot navigation do not provide a solution for this problem. In\nsuch situations, the robot could interact with the humans in order to clear its\npath instead of considering them as unanimated obstacles. In this work, we\npropose a robot behavior for a wheeled humanoid robot that complains with\nsocial norms for clearing its path when the robot is frozen due to the presence\nof humans. The behavior consists of two modules: 1) A detection module, which\nmake use of the Yolo v3 algorithm trained to detect human hands and human arms.\n2) A gesture module, which make use of a policy trained in simulation using the\nProximal Policy Optimization algorithm. Orchestration of the two models is done\nusing the ROS framework.', ""The future will be replete with scenarios where humans are robots will be\nworking together in complex environments. Teammates interact, and the robot's\ninteraction has to be about getting useful information about the human's\n(teammate's) model. There are many challenges before a robot can interact, such\nas incorporating the structural differences in the human's model, ensuring\nsimpler responses, etc. In this paper, we investigate how a robot can interact\nto localize the human model from a set of models. We show how to generate\nquestions to refine the robot's understanding of the teammate's model. We\nevaluate the method in various planning domains. The evaluation shows that\nthese questions can be generated offline, and can help refine the model through\nsimple answers."", ""As environments involving both robots and humans become increasingly common,\nso does the need to account for people during planning. To plan effectively,\nrobots must be able to respond to and sometimes influence what humans do. This\nrequires a human model which predicts future human actions. A simple model may\nassume the human will continue what they did previously; a more complex one\nmight predict that the human will act optimally, disregarding the robot;\nwhereas an even more complex one might capture the robot's ability to influence\nthe human. These models make different trade-offs between computational time\nand performance of the resulting robot plan. Using only one model of the human\neither wastes computational resources or is unable to handle critical\nsituations. In this work, we give the robot access to a suite of human models\nand enable it to assess the performance-computation trade-off online. By\nestimating how an alternate model could improve human prediction and how that\nmay translate to performance gain, the robot can dynamically switch human\nmodels whenever the additional computation is justified. Our experiments in a\ndriving simulator showcase how the robot can achieve performance comparable to\nalways using the best human model, but with greatly reduced computation."", 'When deploying autonomous agents in the real world, we need effective ways of\ncommunicating objectives to them. Traditional skill learning has revolved\naround reinforcement and imitation learning, each with rigid constraints on the\nformat of information exchanged between the human and the agent. While scalar\nrewards carry little information, demonstrations require significant effort to\nprovide and may carry more information than is necessary. Furthermore, rewards\nand demonstrations are often defined and collected before training begins, when\nthe human is most uncertain about what information would help the agent. In\ncontrast, when humans communicate objectives with each other, they make use of\na large vocabulary of informative behaviors, including non-verbal\ncommunication, and often communicate throughout learning, responding to\nobserved behavior. In this way, humans communicate intent with minimal effort.\nIn this paper, we propose such interactive learning as an alternative to reward\nor demonstration-driven learning. To accomplish this, we introduce a\nmulti-agent training framework that enables an agent to learn from another\nagent who knows the current task. Through a series of experiments, we\ndemonstrate the emergence of a variety of interactive learning behaviors,\nincluding information-sharing, information-seeking, and question-answering.\nMost importantly, we find that our approach produces an agent that is capable\nof learning interactively from a human user, without a set of explicit\ndemonstrations or a reward function, and achieving significantly better\nperformance cooperatively with a human than a human performing the task alone.', 'As autonomous service robots become more affordable and thus available also\nfor the general public, there is a growing need for user friendly interfaces to\ncontrol the robotic system. Currently available control modalities typically\nexpect users to be able to express their desire through either touch, speech or\ngesture commands. While this requirement is fulfilled for the majority of\nusers, paralyzed users may not be able to use such systems. In this paper, we\npresent a novel framework, that allows these users to interact with a robotic\nservice assistant in a closed-loop fashion, using only thoughts. The\nbrain-computer interface (BCI) system is composed of several interacting\ncomponents, i.e., non-invasive neuronal signal recording and decoding,\nhigh-level task planning, motion and manipulation planning as well as\nenvironment perception. In various experiments, we demonstrate its\napplicability and robustness in real world scenarios, considering\nfetch-and-carry tasks and tasks involving human-robot interaction. As our\nresults demonstrate, our system is capable of adapting to frequent changes in\nthe environment and reliably completing given tasks within a reasonable amount\nof time. Combined with high-level planning and autonomous robotic systems,\ninteresting new perspectives open up for non-invasive BCI-based human-robot\ninteractions.', 'Reinforcement learning is a promising method to accomplish robotic control\ntasks. The task of playing musical instruments is, however, largely unexplored\nbecause it involves the challenge of achieving sequential goals - melodies -\nthat have a temporal dimension. In this paper, we address robotic musicianship\nby introducing a temporal extension to goal-conditioned reinforcement learning:\nTime-dependent goals. We demonstrate that these can be used to train a robotic\nmusician to play the theremin instrument. We train the robotic agent in\nsimulation and transfer the acquired policy to a real-world robotic\nthereminist. Supplemental video: https://youtu.be/jvC9mPzdQN4', 'Reinforcement learning is a powerful technique to train an agent to perform a\ntask. However, an agent that is trained using reinforcement learning is only\ncapable of achieving the single task that is specified via its reward function.\nSuch an approach does not scale well to settings in which an agent needs to\nperform a diverse set of tasks, such as navigating to varying positions in a\nroom or moving objects to varying locations. Instead, we propose a method that\nallows an agent to automatically discover the range of tasks that it is capable\nof performing. We use a generator network to propose tasks for the agent to try\nto achieve, specified as goal states. The generator network is optimized using\nadversarial training to produce tasks that are always at the appropriate level\nof difficulty for the agent. Our method thus automatically produces a\ncurriculum of tasks for the agent to learn. We show that, by using this\nframework, an agent can efficiently and automatically learn to perform a wide\nset of tasks without requiring any prior knowledge of its environment. Our\nmethod can also learn to achieve tasks with sparse rewards, which traditionally\npose significant challenges.', ""As reinforcement learning agents become increasingly integrated into complex,\nreal-world environments, designing for safety becomes a critical consideration.\nWe specifically focus on researching scenarios where agents can cause undesired\nside effects while executing a policy on a primary task. Since one can define\nmultiple tasks for a given environment dynamics, there are two important\nchallenges. First, we need to abstract the concept of safety that applies\nbroadly to that environment independent of the specific task being executed.\nSecond, we need a mechanism for the abstracted notion of safety to modulate the\nactions of agents executing different policies to minimize their side-effects.\nIn this work, we propose Safety Aware Reinforcement Learning (SARL) - a\nframework where a virtual safe agent modulates the actions of a main\nreward-based agent to minimize side effects. The safe agent learns a\ntask-independent notion of safety for a given environment. The main agent is\nthen trained with a regularization loss given by the distance between the\nnative action probabilities of the two agents. Since the safe agent effectively\nabstracts a task-independent notion of safety via its action probabilities, it\ncan be ported to modulate multiple policies solving different tasks within the\ngiven environment without further training. We contrast this with solutions\nthat rely on task-specific regularization metrics and test our framework on the\nSafeLife Suite, based on Conway's Game of Life, comprising a number of complex\ntasks in dynamic environments. We show that our solution is able to match the\nperformance of solutions that rely on task-specific side-effect penalties on\nboth the primary and safety objectives while additionally providing the benefit\nof generalizability and portability."", 'A core aspect of human intelligence is the ability to learn new tasks quickly\nand switch between them flexibly. Here, we describe a modular continual\nreinforcement learning paradigm inspired by these abilities. We first introduce\na visual interaction environment that allows many types of tasks to be unified\nin a single framework. We then describe a reward map prediction scheme that\nlearns new tasks robustly in the very large state and action spaces required by\nsuch an environment. We investigate how properties of module architecture\ninfluence efficiency of task learning, showing that a module motif\nincorporating specific design principles (e.g. early bottlenecks, low-order\npolynomial nonlinearities, and symmetry) significantly outperforms more\nstandard neural network motifs, needing fewer training examples and fewer\nneurons to achieve high levels of performance. Finally, we present a\nmeta-controller architecture for task switching based on a dynamic neural\nvoting scheme, which allows new modules to use information learned from\npreviously-seen tasks to substantially improve their own learning efficiency.', 'Among the many anticipated roles for robots in the future is that of being a\nhuman teammate. Aside from all the technological hurdles that have to be\novercome with respect to hardware and control to make robots fit to work with\nhumans, the added complication here is that humans have many conscious and\nsubconscious expectations of their teammates - indeed, we argue that teaming is\nmostly a cognitive rather than physical coordination activity. This introduces\nnew challenges for the AI and robotics community and requires fundamental\nchanges to the traditional approach to the design of autonomy. With this in\nmind, we propose an update to the classical view of the intelligent agent\narchitecture, highlighting the requirements for mental modeling of the human in\nthe deliberative process of the autonomous agent. In this article, we outline\nbriefly the recent efforts of ours, and others in the community, towards\ndeveloping cognitive teammates along these guidelines.', 'Various automatic curriculum learning (ACL) methods have been proposed to\nimprove the sample efficiency and final performance of deep reinforcement\nlearning (DRL). They are designed to control how a DRL agent collects data,\nwhich is inspired by how humans gradually adapt their learning processes to\ntheir capabilities. For example, ACL can be used for subgoal generation, reward\nshaping, environment generation, or initial state generation. However, prior\nwork only considers curriculum learning following one of the aforementioned\npredefined paradigms. It is unclear which of these paradigms are complementary,\nand how the combination of them can be learned from interactions with the\nenvironment. Therefore, in this paper, we propose a unified automatic\ncurriculum learning framework to create multi-objective but coherent curricula\nthat are generated by a set of parametric curriculum modules. Each curriculum\nmodule is instantiated as a neural network and is responsible for generating a\nparticular curriculum. In order to coordinate those potentially conflicting\nmodules in unified parameter space, we propose a multi-task hyper-net learning\nframework that uses a single hyper-net to parameterize all those curriculum\nmodules. In addition to existing hand-designed curricula paradigms, we further\ndesign a flexible memory mechanism to learn an abstract curriculum, which may\notherwise be difficult to design manually. We evaluate our method on a series\nof robotic manipulation tasks and demonstrate its superiority over other\nstate-of-the-art ACL methods in terms of sample efficiency and final\nperformance.', 'Many works in collaborative robotics and human-robot interaction focuses on\nidentifying and predicting human behaviour while considering the information\nabout the robot itself as given. This can be the case when sensors and the\nrobot are calibrated in relation to each other and often the reconfiguration of\nthe system is not possible, or extra manual work is required. We present a deep\nlearning based approach to remove the constraint of having the need for the\nrobot and the vision sensor to be fixed and calibrated in relation to each\nother. The system learns the visual cues of the robot body and is able to\nlocalise it, as well as estimate the position of robot joints in 3D space by\njust using a 2D color image. The method uses a cascaded convolutional neural\nnetwork, and we present the structure of the network, describe our own\ncollected dataset, explain the network training and achieved results. A fully\ntrained system shows promising results in providing an accurate mask of where\nthe robot is located and a good estimate of its joints positions in 3D. The\naccuracy is not good enough for visual servoing applications yet, however, it\ncan be sufficient for general safety and some collaborative tasks not requiring\nvery high precision. The main benefit of our method is the possibility of the\nvision sensor to move freely. This allows it to be mounted on moving objects,\nfor example, a body of the person or a mobile robot working in the same\nenvironment as the robots are operating in.', ""Observational learning is a type of learning that occurs as a function of\nobserving, retaining and possibly replicating or imitating the behaviour of\nanother agent. It is a core mechanism appearing in various instances of social\nlearning and has been found to be employed in several intelligent species,\nincluding humans. In this paper, we investigate to what extent the explicit\nmodelling of other agents is necessary to achieve observational learning\nthrough machine learning. Especially, we argue that observational learning can\nemerge from pure Reinforcement Learning (RL), potentially coupled with memory.\nThrough simple scenarios, we demonstrate that an RL agent can leverage the\ninformation provided by the observations of an other agent performing a task in\na shared environment. The other agent is only observed through the effect of\nits actions on the environment and never explicitly modeled. Two key aspects\nare borrowed from observational learning: i) the observer behaviour needs to\nchange as a result of viewing a 'teacher' (another agent) and ii) the observer\nneeds to be motivated somehow to engage in making use of the other agent's\nbehaviour. The later is naturally modeled by RL, by correlating the learning\nagent's reward with the teacher agent's behaviour."", 'In outdoor environments, mobile robots are required to navigate through\nterrain with varying characteristics, some of which might significantly affect\nthe integrity of the platform. Ideally, the robot should be able to identify\nareas that are safe for navigation based on its own percepts about the\nenvironment while avoiding damage to itself. Bayesian optimisation (BO) has\nbeen successfully applied to the task of learning a model of terrain\ntraversability while guiding the robot through more traversable areas. An\nissue, however, is that localisation uncertainty can end up guiding the robot\nto unsafe areas and distort the model being learnt. In this paper, we address\nthis problem and present a novel method that allows BO to consider localisation\nuncertainty by applying a Gaussian process model for uncertain inputs as a\nprior. We evaluate the proposed method in simulation and in experiments with a\nreal robot navigating over rough terrain and compare it against standard BO\nmethods.', 'Imitation learning is an effective approach for autonomous systems to acquire\ncontrol policies when an explicit reward function is unavailable, using\nsupervision provided as demonstrations from an expert, typically a human\noperator. However, standard imitation learning methods assume that the agent\nreceives examples of observation-action tuples that could be provided, for\ninstance, to a supervised learning algorithm. This stands in contrast to how\nhumans and animals imitate: we observe another person performing some behavior\nand then figure out which actions will realize that behavior, compensating for\nchanges in viewpoint, surroundings, object positions and types, and other\nfactors. We term this kind of imitation learning ""imitation-from-observation,""\nand propose an imitation learning method based on video prediction with context\ntranslation and deep reinforcement learning. This lifts the assumption in\nimitation learning that the demonstration should consist of observations in the\nsame environment configuration, and enables a variety of interesting\napplications, including learning robotic skills that involve tool use simply by\nobserving videos of human tool use. Our experimental results show the\neffectiveness of our approach in learning a wide range of real-world robotic\ntasks modeled after common household chores from videos of a human\ndemonstrator, including sweeping, ladling almonds, pushing objects as well as a\nnumber of tasks in simulation.', ""Communication is not only an action of choosing a signal, but needs to\nconsider the context and sensor signals. It also needs to decide what\ninformation is communicated and how it is represented in or understood from\nsignals. Therefore, communication should be realized comprehensively together\nwith its purpose and other functions.\n  The recent successful results in end-to-end reinforcement learning (RL) show\nthe importance of comprehensive learning and the usefulness of end-to-end RL.\nAlthough little is known, we have shown that a variety of communications emerge\nthrough RL using a (recurrent) neural network (NN). Here, three of them are\nintroduced.\n  In the 1st one, negotiation to avoid conflicts among 4 randomly-picked agents\nwas learned. Each agent generates a binary signal from the output of its\nrecurrent NN (RNN), and receives 4 signals from the agents three times. After\nlearning, each agent made an appropriate final decision after negotiation for\nany combination of 4 agents. Differentiation of individuality among the agents\nalso could be seen.\n  The 2nd one focused on discretization of communication signal. A sender agent\nperceives the receiver's location and generates a continuous signal twice by\nits RNN. A receiver agent receives them sequentially, and moves according to\nits RNN's output to reach the sender's location. When noises were added to the\nsignal, it was binarized through learning and 2-bit communication was\nestablished.\n  The 3rd one focused on end-to-end comprehensive communication. A sender\nreceives 1,785-pixel real camera image on which a real robot can be seen, and\nsends two sounds whose frequencies are computed by its NN. A receiver receives\nthem, and two motion commands for the robot are generated by its NN. After\nlearning, though some preliminary learning was necessary for the sender, the\nrobot could reach the goal from any initial location."", 'It is highly desirable for robots that work alongside humans to be able to\nunderstand instructions in natural language. Existing language conditioned\nimitation learning models directly predict the actuator commands from the image\nobservation and the instruction text. Rather than directly predicting actuator\ncommands, we propose translating the natural language instruction to a Python\nfunction which queries the scene by accessing the output of the object detector\nand controls the robot to perform the specified task. This enables the use of\nnon-differentiable modules such as a constraint solver when computing commands\nto the robot. Moreover, the labels in this setup are significantly more\ninformative computer programs that capture the intent of the expert rather than\nteleoperated demonstrations. We show that the proposed method performs better\nthan training a neural network to directly predict the robot actions.', ""Predictive human models often need to adapt their parameters online from\nhuman data. This raises previously ignored safety-related questions for robots\nrelying on these models such as what the model could learn online and how\nquickly could it learn it. For instance, when will the robot have a confident\nestimate in a nearby human's goal? Or, what parameter initializations guarantee\nthat the robot can learn the human's preferences in a finite number of\nobservations? To answer such analysis questions, our key idea is to model the\nrobot's learning algorithm as a dynamical system where the state is the current\nmodel parameter estimate and the control is the human data the robot observes.\nThis enables us to leverage tools from reachability analysis and optimal\ncontrol to compute the set of hypotheses the robot could learn in finite time,\nas well as the worst and best-case time it takes to learn them. We demonstrate\nthe utility of our analysis tool in four human-robot domains, including\nautonomous driving and indoor navigation."", 'Human emotions are expressed through multiple modalities, including verbal\nand non-verbal information. Moreover, the affective states of human users can\nbe the indicator for the level of engagement and successful interaction,\nsuitable for the robot to use as a rewarding factor to optimize robotic\nbehaviors through interaction. This study demonstrates a multimodal human-robot\ninteraction (HRI) framework with reinforcement learning to enhance the robotic\ninteraction policy and personalize emotional interaction for a human user. The\ngoal is to apply this framework in social scenarios that can let the robots\ngenerate a more natural and engaging HRI framework.', 'Recent development in developing humanoid robot poses new challenges to\nhuman-machine interaction communication. A major challenge is to develop robots\nthat can behave like and interact with human in the most natural way possible.\nThis paper proposes a system to develop a robot that can receive command, and\ntalk to people in natural language. In addition, the robot can also be\n""trained"" to become an expert in sepcific areas to provide expert advice to\nhuman-beings. Most important of all, the robot can display emotions through\nfacial expression, speech and gesture so that the interaction process will\nbecome more comprehensive and compelling.', ""In many robotic applications, an autonomous agent must act within and explore\na partially observed environment that is unobserved by its human teammate. We\nconsider such a setting in which the agent can, while acting, transmit\ndeclarative information to the human that helps them understand aspects of this\nunseen environment. In this work, we address the algorithmic question of how\nthe agent should plan out what actions to take and what information to\ntransmit. Naturally, one would expect the human to have preferences, which we\nmodel information-theoretically by scoring transmitted information based on the\nchange it induces in weighted entropy of the human's belief state. We formulate\nthis setting as a belief MDP and give a tractable algorithm for solving it\napproximately. Then, we give an algorithm that allows the agent to learn the\nhuman's preferences online, through exploration. We validate our approach\nexperimentally in simulated discrete and continuous partially observed\nsearch-and-recover domains. Visit http://tinyurl.com/chitnis-corl-18 for a\nsupplementary video."", 'Meta-learning algorithms can accelerate the model-based reinforcement\nlearning (MBRL) algorithms by finding an initial set of parameters for the\ndynamical model such that the model can be trained to match the actual dynamics\nof the system with only a few data-points. However, in the real world, a robot\nmight encounter any situation starting from motor failures to finding itself in\na rocky terrain where the dynamics of the robot can be significantly different\nfrom one another. In this paper, first, we show that when meta-training\nsituations (the prior situations) have such diverse dynamics, using a single\nset of meta-trained parameters as a starting point still requires a large\nnumber of observations from the real system to learn a useful model of the\ndynamics. Second, we propose an algorithm called FAMLE that mitigates this\nlimitation by meta-training several initial starting points (i.e., initial\nparameters) for training the model and allows the robot to select the most\nsuitable starting point to adapt the model to the current situation with only a\nfew gradient steps. We compare FAMLE to MBRL, MBRL with a meta-trained model\nwith MAML, and model-free policy search algorithm PPO for various simulated and\nreal robotic tasks, and show that FAMLE allows the robots to adapt to novel\ndamages in significantly fewer time-steps than the baselines.', ""We explore the problem of estimating the mass distribution of an articulated\nobject by an interactive robotic agent. Our method predicts the mass\ndistribution of an object by using the limited sensing and actuating\ncapabilities of a robotic agent that is interacting with the object. We are\ninspired by the role of exploratory play in human infants. We take the combined\napproach of supervised and reinforcement learning to train an agent that learns\nto strategically interact with the object to estimate the object's mass\ndistribution. Our method consists of two neural networks: (i) the policy\nnetwork which decides how to interact with the object, and (ii) the predictor\nnetwork that estimates the mass distribution given a history of observations\nand interactions. Using our method, we train a robotic arm to estimate the mass\ndistribution of an object with moving parts (e.g. an articulated rigid body\nsystem) by pushing it on a surface with unknown friction properties. We also\ndemonstrate how our training from simulations can be transferred to real\nhardware using a small amount of real-world data for fine-tuning. We use a UR10\nrobot to interact with 3D printed articulated chains with varying mass\ndistributions and show that our method significantly outperforms the baseline\nsystem that uses random pushes to interact with the object."", ""Cooperation in multi-agent and multi-robot systems can help agents build\nvarious formations, shapes, and patterns presenting corresponding functions and\npurposes adapting to different situations. Relationships between agents such as\ntheir spatial proximity and functional similarities could play a crucial role\nin cooperation between agents. Trust level between agents is an essential\nfactor in evaluating their relationships' reliability and stability, much as\npeople do. This paper proposes a new model called Relative Needs Entropy (RNE)\nto assess trust between robotic agents. RNE measures the distance of needs\ndistribution between individual agents or groups of agents. To exemplify its\nutility, we implement and demonstrate our trust model through experiments\nsimulating a heterogeneous multi-robot grouping task in a persistent urban\nsearch and rescue mission consisting of tasks at two levels of difficulty. The\nresults suggest that RNE trust-Based grouping of robots can achieve better\nperformance and adaptability for diverse task execution compared to the\nstate-of-the-art energy-based or distance-based grouping models."", ""Predicting accurate future trajectories of multiple agents is essential for\nautonomous systems, but is challenging due to the complex agent interaction and\nthe uncertainty in each agent's future behavior. Forecasting multi-agent\ntrajectories requires modeling two key dimensions: (1) time dimension, where we\nmodel the influence of past agent states over future states; (2) social\ndimension, where we model how the state of each agent affects others. Most\nprior methods model these two dimensions separately, e.g., first using a\ntemporal model to summarize features over time for each agent independently and\nthen modeling the interaction of the summarized features with a social model.\nThis approach is suboptimal since independent feature encoding over either the\ntime or social dimension can result in a loss of information. Instead, we would\nprefer a method that allows an agent's state at one time to directly affect\nanother agent's state at a future time. To this end, we propose a new\nTransformer, AgentFormer, that jointly models the time and social dimensions.\nThe model leverages a sequence representation of multi-agent trajectories by\nflattening trajectory features across time and agents. Since standard attention\noperations disregard the agent identity of each element in the sequence,\nAgentFormer uses a novel agent-aware attention mechanism that preserves agent\nidentities by attending to elements of the same agent differently than elements\nof other agents. Based on AgentFormer, we propose a stochastic multi-agent\ntrajectory prediction model that can attend to features of any agent at any\nprevious timestep when inferring an agent's future position. The latent intent\nof all agents is also jointly modeled, allowing the stochasticity in one\nagent's behavior to affect other agents. Our method substantially improves the\nstate of the art on well-established pedestrian and autonomous driving\ndatasets."", ""Learning in multi-agent environments is difficult due to the non-stationarity\nintroduced by an opponent's or partner's changing behaviors. Instead of\nreactively adapting to the other agent's (opponent or partner) behavior, we\npropose an algorithm to proactively influence the other agent's strategy to\nstabilize -- which can restrain the non-stationarity caused by the other agent.\nWe learn a low-dimensional latent representation of the other agent's strategy\nand the dynamics of how the latent strategy evolves with respect to our robot's\nbehavior. With this learned dynamics model, we can define an unsupervised\nstability reward to train our robot to deliberately influence the other agent\nto stabilize towards a single strategy. We demonstrate the effectiveness of\nstabilizing in improving efficiency of maximizing the task reward in a variety\nof simulated environments, including autonomous driving, emergent\ncommunication, and robotic manipulation. We show qualitative results on our\nwebsite: https://sites.google.com/view/stable-marl/."", 'Humans are able to seamlessly visually imitate others, by inferring their\nintentions and using past experience to achieve the same end goal. In other\nwords, we can parse complex semantic knowledge from raw video and efficiently\ntranslate that into concrete motor control. Is it possible to give a robot this\nsame capability? Prior research in robot imitation learning has created agents\nwhich can acquire diverse skills from expert human operators. However,\nexpanding these techniques to work with a single positive example during test\ntime is still an open challenge. Apart from control, the difficulty stems from\nmismatches between the demonstrator and robot domains. For example, objects may\nbe placed in different locations (e.g. kitchen layouts are different in every\nhouse). Additionally, the demonstration may come from an agent with different\nmorphology and physical appearance (e.g. human), so one-to-one action\ncorrespondences are not available. This paper investigates techniques which\nallow robots to partially bridge these domain gaps, using their past\nexperience. A neural network is trained to mimic ground truth robot actions\ngiven context video from another agent, and must generalize to unseen task\ninstances when prompted with new videos during test time. We hypothesize that\nour policy representations must be both context driven and dynamics aware in\norder to perform these tasks. These assumptions are baked into the neural\nnetwork using the Transformers attention mechanism and a self-supervised\ninverse dynamics loss. Finally, we experimentally determine that our method\naccomplishes a $\\sim 2$x improvement in terms of task success rate over prior\nbaselines in a suite of one-shot manipulation tasks.', ""The exploration mechanism used by a Deep Reinforcement Learning (RL) agent\nplays a key role in determining its sample efficiency. Thus, improving over\nrandom exploration is crucial to solve long-horizon tasks with sparse rewards.\nWe propose to leverage an ensemble of partial solutions as teachers that guide\nthe agent's exploration with action suggestions throughout training. While the\nsetup of learning with teachers has been previously studied, our proposed\napproach - Actor-Critic with Teacher Ensembles (AC-Teach) - is the first to\nwork with an ensemble of suboptimal teachers that may solve only part of the\nproblem or contradict other each other, forming a unified algorithmic solution\nthat is compatible with a broad range of teacher ensembles. AC-Teach leverages\na probabilistic representation of the expected outcome of the teachers' and\nstudent's actions to direct exploration, reduce dithering, and adapt to the\ndynamically changing quality of the learner. We evaluate a variant of AC-Teach\nthat guides the learning of a Bayesian DDPG agent on three tasks - path\nfollowing, robotic pick and place, and robotic cube sweeping using a hook - and\nshow that it improves largely on sampling efficiency over a set of baselines,\nboth for our target scenario of unconstrained suboptimal teachers and for\neasier setups with optimal or single teachers. Additional results and videos at\nhttps://sites.google.com/view/acteach/home."", ""It is important for socially assistive robots to be able to recognize when a\nuser needs and wants help. Such robots need to be able to recognize human needs\nin a real-time manner so that they can provide timely assistance. We propose an\narchitecture that uses social cues to determine when a robot should provide\nassistance. Based on a multimodal fusion approach upon eye gaze and language\nmodalities, our architecture is trained and evaluated on data collected in a\nrobot-assisted Lego building task. By focusing on social cues, our architecture\nhas minimal dependencies on the specifics of a given task, enabling it to be\napplied in many different contexts. Enabling a social robot to recognize a\nuser's needs through social cues can help it to adapt to user behaviors and\npreferences, which in turn will lead to improved user experiences."", 'My overarching research goal is to provide robots with perceptional abilities\nthat allow interactions with humans in a human-like manner. To develop these\nperceptional abilities, I believe that it is useful to study the principles of\nthe human visual system. I use these principles to develop new computer vision\nalgorithms and validate their effectiveness in intelligent robotic systems. I\nam enthusiastic about this approach as it offers the dual benefit of uncovering\nprinciples inherent in the human visual system, as well as applying these\nprinciples to its artificial counterpart. Fig. 1 contains a depiction of my\nresearch.', 'The objective of a reinforcement learning agent is to behave so as to\nmaximise the sum of a suitable scalar function of state: the reward. These\nrewards are typically given and immutable. In this paper, we instead consider\nthe proposition that the reward function itself can be a good locus of learned\nknowledge. To investigate this, we propose a scalable meta-gradient framework\nfor learning useful intrinsic reward functions across multiple lifetimes of\nexperience. Through several proof-of-concept experiments, we show that it is\nfeasible to learn and capture knowledge about long-term exploration and\nexploitation into a reward function. Furthermore, we show that unlike policy\ntransfer methods that capture ""how"" the agent should behave, the learned reward\nfunctions can generalise to other kinds of agents and to changes in the\ndynamics of the environment by capturing ""what"" the agent should strive to do.', ""In the vision and language navigation task, the agent may encounter ambiguous\nsituations that are hard to interpret by just relying on visual information and\nnatural language instructions. We propose an interactive learning framework to\nendow the agent with the ability to ask for users' help in such situations. As\npart of this framework, we investigate multiple learning approaches for the\nagent with different levels of complexity. The simplest model-confusion-based\nmethod lets the agent ask questions based on its confusion, relying on the\npredefined confidence threshold of a next action prediction model. To build on\nthis confusion-based method, the agent is expected to demonstrate more\nsophisticated reasoning such that it discovers the timing and locations to\ninteract with a human. We achieve this goal using reinforcement learning (RL)\nwith a proposed reward shaping term, which enables the agent to ask questions\nonly when necessary. The success rate can be boosted by at least 15% with only\none question asked on average during the navigation. Furthermore, we show that\nthe RL agent is capable of adjusting dynamically to noisy human responses.\nFinally, we design a continual learning strategy, which can be viewed as a data\naugmentation method, for the agent to improve further utilizing its interaction\nhistory with a human. We demonstrate the proposed strategy is substantially\nmore realistic and data-efficient compared to previously proposed\npre-exploration techniques."", 'We study the problem of multi-robot mapless navigation in the popular\nCentralized Training and Decentralized Execution (CTDE) paradigm. This problem\nis challenging when each robot considers its path without explicitly sharing\nobservations with other robots and can lead to non-stationary issues in Deep\nReinforcement Learning (DRL). The typical CTDE algorithm factorizes the joint\naction-value function into individual ones, to favor cooperation and achieve\ndecentralized execution. Such factorization involves constraints (e.g.,\nmonotonicity) that limit the emergence of novel behaviors in an individual as\neach agent is trained starting from a joint action-value. In contrast, we\npropose a novel architecture for CTDE that uses a centralized state-value\nnetwork to compute a joint state-value, which is used to inject global state\ninformation in the value-based updates of the agents. Consequently, each model\ncomputes its gradient update for the weights, considering the overall state of\nthe environment. Our idea follows the insights of Dueling Networks as a\nseparate estimation of the joint state-value has both the advantage of\nimproving sample efficiency, while providing each robot information whether the\nglobal state is (or is not) valuable. Experiments in a robotic navigation task\nwith 2 4, and 8 robots, confirm the superior performance of our approach over\nprior CTDE methods (e.g., VDN, QMIX).', 'Social learning is a key component of human and animal intelligence. By\ntaking cues from the behavior of experts in their environment, social learners\ncan acquire sophisticated behavior and rapidly adapt to new circumstances. This\npaper investigates whether independent reinforcement learning (RL) agents in a\nmulti-agent environment can learn to use social learning to improve their\nperformance. We find that in most circumstances, vanilla model-free RL agents\ndo not use social learning. We analyze the reasons for this deficiency, and\nshow that by imposing constraints on the training environment and introducing a\nmodel-based auxiliary loss we are able to obtain generalized social learning\npolicies which enable agents to: i) discover complex skills that are not\nlearned from single-agent training, and ii) adapt online to novel environments\nby taking cues from experts present in the new environment. In contrast, agents\ntrained with model-free RL or imitation learning generalize poorly and do not\nsucceed in the transfer tasks. By mixing multi-agent and solo training, we can\nobtain agents that use social learning to gain skills that they can deploy when\nalone, even out-performing agents trained alone from the start.', 'The ability to distinguish between the self and the background is of\nparamount importance for robotic tasks. The particular case of hands, as the\nend effectors of a robotic system that more often enter into contact with other\nelements of the environment, must be perceived and tracked with precision to\nexecute the intended tasks with dexterity and without colliding with obstacles.\nThey are fundamental for several applications, from Human-Robot Interaction\ntasks to object manipulation. Modern humanoid robots are characterized by high\nnumber of degrees of freedom which makes their forward kinematics models very\nsensitive to uncertainty. Thus, resorting to vision sensing can be the only\nsolution to endow these robots with a good perception of the self, being able\nto localize their body parts with precision. In this paper, we propose the use\nof a Convolution Neural Network (CNN) to segment the robot hand from an image\nin an egocentric view. It is known that CNNs require a huge amount of data to\nbe trained. To overcome the challenge of labeling real-world images, we propose\nthe use of simulated datasets exploiting domain randomization techniques. We\nfine-tuned the Mask-RCNN network for the specific task of segmenting the hand\nof the humanoid robot Vizzy. We focus our attention on developing a methodology\nthat requires low amounts of data to achieve reasonable performance while\ngiving detailed insight on how to properly generate variability in the training\ndataset. Moreover, we analyze the fine-tuning process within the complex model\nof Mask-RCNN, understanding which weights should be transferred to the new task\nof segmenting robot hands. Our final model was trained solely on synthetic\nimages and achieves an average IoU of 82% on synthetic validation data and\n56.3% on real test data. These results were achieved with only 1000 training\nimages and 3 hours of training time using a single GPU.', 'Proxemics is a branch of non-verbal communication concerned with studying the\nspatial behavior of people and animals. This behavior is an essential part of\nthe communication process due to delimit the acceptable distance to interact\nwith another being. With increasing research on human-agent interaction, new\nalternatives are needed that allow optimal communication, avoiding agents\nfeeling uncomfortable. Several works consider proxemic behavior with cognitive\nagents, where human-robot interaction techniques and machine learning are\nimplemented. However, environments consider fixed personal space and that the\nagent previously knows it. In this work, we aim to study how agents behave in\nenvironments based on proxemic behavior, and propose a modified gridworld to\nthat aim. This environment considers an issuer with proxemic behavior that\nprovides a disagreement signal to the agent. Our results show that the learning\nagent can identify the proxemic space when the issuer gives feedback about\nagent performance.', ""The mental models that humans form of other agents---encapsulating human\nbeliefs about agent goals, intentions, capabilities, and more---create an\nunderlying basis for interaction. These mental models have the potential to\naffect both the human's decision making during the interaction and the human's\nsubjective assessment of the interaction. In this paper, we surveyed existing\nmethods for modeling how humans view robots, then identified a potential method\nfor improving these estimates through inferring a human's model of a robot\nagent directly from their actions. Then, we conducted an online study to\ncollect data in a grid-world environment involving humans moving an avatar past\na virtual agent. Through our analysis, we demonstrated that participants'\naction choices leaked information about their mental models of a virtual agent.\nWe conclude by discussing the implications of these findings and the potential\nfor such a method to improve human-robot interactions."", ""Any agent that is part of the environment it interacts with and has versatile\nactuators (such as arms and fingers), will in principle have the ability to\nself-modify -- for example by changing its own source code. As we continue to\ncreate more and more intelligent agents, chances increase that they will learn\nabout this ability. The question is: will they want to use it? For example,\nhighly intelligent systems may find ways to change their goals to something\nmore easily achievable, thereby `escaping' the control of their designers. In\nan important paper, Omohundro (2008) argued that goal preservation is a\nfundamental drive of any intelligent system, since a goal is more likely to be\nachieved if future versions of the agent strive towards the same goal. In this\npaper, we formalise this argument in general reinforcement learning, and\nexplore situations where it fails. Our conclusion is that the self-modification\npossibility is harmless if and only if the value function of the agent\nanticipates the consequences of self-modifications and use the current utility\nfunction when evaluating the future."", ""In this study Arduino card based mobile robot design was realized. This robot\ncan serve as a security robot, an auxiliary robot or a control robot. The\ndesigned robot has two operation modes. The first operating mode is autonomous\nmode. In this mode, the robot detects the surroundings with the help of\nultrasonic sensors placed around it, and keeps track of the places it passes by\nusing the encoder. It is able to navigate without hitting any place and passing\nfrom where it passes, and it transmits the patient's pulse and temperature\ncondition to the user by other systems installed on it. Also the IP camera\nsends the scene on the screen. The emergency button to be placed next to the\npatient sends information to the user in emergency situations. If the\nabnormality is detected in the temperature and pulse again, the user gives a\nmessage. When the pre-recorded drug use times come, the system can alert the\npatient. The second mode is manual mode. In this mode, the user can move the\ndesired direction of the robot with the Android operating system. In addition,\nall data received in autonomous mode can be sent to the user. Thus, the user\ncan control the mobile robot with the camera image even if it is not in the\nvicinity of the robot."", 'Multi-task learning can leverage information learned by one task to benefit\nthe training of other tasks. Despite this capacity, naive formulations often\ndegrade performance and in particular, identifying the tasks that would benefit\nfrom co-training remains a challenging design question. In this paper, we\nanalyze the dynamics of information transfer, or transference, across tasks\nthroughout training. Specifically, we develop a similarity measure that can\nquantify transference among tasks and use this quantity to both better\nunderstand the optimization dynamics of multi-task learning as well as improve\noverall learning performance. In the latter case, we propose two methods to\nleverage our transference metric. The first operates at a macro-level by\nselecting which tasks should train together while the second functions at a\nmicro-level by determining how to combine task gradients at each training step.\nWe find these methods can lead to significant improvement over prior work on\nthree supervised multi-task learning benchmarks and one multi-task\nreinforcement learning paradigm.', 'Agent Based Models are very popular in a number of different areas. For\nexample, they have been used in a range of domains ranging from modeling of\ntumor growth, immune systems, molecules to models of social networks, crowds\nand computer and mobile self-organizing networks. One reason for their success\nis their intuitiveness and similarity to human cognition. However, with this\npower of abstraction, in spite of being easily applicable to such a wide number\nof domains, it is hard to validate agent-based models. In addition, building\nvalid and credible simulations is not just a challenging task but also a\ncrucial exercise to ensure that what we are modeling is, at some level of\nabstraction, a model of our conceptual system; the system that we have in mind.\nIn this paper, we address this important area of validation of agent based\nmodels by presenting a novel technique which has broad applicability and can be\napplied to all kinds of agent-based models. We present a framework, where a\nvirtual overlay multi-agent system can be used to validate simulation models.\nIn addition, since agent-based models have been typically growing, in parallel,\nin multiple domains, to cater for all of these, we present a new single\nvalidation technique applicable to all agent based models. Our technique, which\nallows for the validation of agent based simulations uses VOMAS: a Virtual\nOverlay Multi-agent System. This overlay multi-agent system can comprise\nvarious types of agents, which form an overlay on top of the agent based\nsimulation model that needs to be validated. Other than being able to watch and\nlog, each of these agents contains clearly defined constraints, which, if\nviolated, can be logged in real time. To demonstrate its effectiveness, we show\nits broad applicability in a wide variety of simulation models ranging from\nsocial sciences to computer networks in spatial and non-spatial conceptual\nmodels.', ""In cooperative multiagent planning, it can often be beneficial for an agent\nto make commitments about aspects of its behavior to others, allowing them in\nturn to plan their own behaviors without taking the agent's detailed behavior\ninto account. Extending previous work in the Bayesian setting, we consider\ninstead a worst-case setting in which the agent has a set of possible\nenvironments (MDPs) it could be in, and develop a commitment semantics that\nallows for probabilistic guarantees on the agent's behavior in any of the\nenvironments it could end up facing. Crucially, an agent receives observations\n(of reward and state transitions) that allow it to potentially eliminate\npossible environments and thus obtain higher utility by adapting its policy to\nthe history of observations. We develop algorithms and provide theory and some\npreliminary empirical results showing that they ensure an agent meets its\ncommitments with history-dependent policies while minimizing maximum regret\nover the possible environments."", 'For the past few decades, man has been trying to create an intelligent\ncomputer which can talk and respond like he can. The task of creating a system\nthat can talk like a human being is the primary objective of Automatic Speech\nRecognition. Various Speech Recognition techniques have been developed in\ntheory and have been applied in practice. This paper discusses the problems\nthat have been encountered in developing Speech Recognition, the techniques\nthat have been applied to automate the task, and a representation of the core\nproblems of present day Speech Recognition by using Fuzzy Mathematics.', 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', 'People are becoming increasingly comfortable using Digital Assistants (DAs)\nto interact with services or connected objects. However, for non-programming\nusers, the available possibilities for customizing their DA are limited and do\nnot include the possibility of teaching the assistant new tasks. To make the\nmost of the potential of DAs, users should be able to customize assistants by\ninstructing them through Natural Language (NL). To provide such\nfunctionalities, NL interpretation in traditional assistants should be\nimproved: (1) The intent identification system should be able to recognize new\nforms of known intents, and to acquire new intents as they are expressed by the\nuser. (2) In order to be adaptive to novel intents, the Natural Language\nUnderstanding module should be sample efficient, and should not rely on a\npretrained model. Rather, the system should continuously collect the training\ndata as it learns new intents from the user. In this work, we propose AidMe\n(Adaptive Intent Detection in Multi-Domain Environments), a user-in-the-loop\nadaptive intent detection framework that allows the assistant to adapt to its\nuser by learning his intents as their interaction progresses. AidMe builds its\nrepertoire of intents and collects data to train a model of semantic similarity\nevaluation that can discriminate between the learned intents and autonomously\ndiscover new forms of known intents. AidMe addresses two major issues - intent\nlearning and user adaptation - for instructable digital assistants. We\ndemonstrate the capabilities of AidMe as a standalone system by comparing it\nwith a one-shot learning system and a pretrained NLU module through simulations\nof interactions with a user. We also show how AidMe can smoothly integrate to\nan existing instructable digital assistant.', 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am interested in conducting a study to examine the behavior and learning capabilities of both human and artificial intelligence (AI) agents within a simulated psychology laboratory environment. This environment must be computerized to allow AI agents to interact within it. I require a pre-existing platform or a method to construct one that enables the implementation of psychological experiments. This platform must be compatible with both human subjects and AI agents, necessitating a computerized system with a visual user interface. The platform should include an Application Programming Interface (API) that allows for the creation of customized tasks and experiments. Essential built-in functionalities should include but not be limited to object detection, object tracking, and change detection. Upon acquiring or constructing such a platform, I aim to explore how reinforcement learning (RL) agents perceive and interpret visual information. I am also interested in investigating how their learning capabilities can be enhanced through the integration of simpler models, such as a basic central vision model. I hope that this platform and the subsequent research it facilitates can contribute to the advancement of deep RL and strengthen its ties with cognitive science.","['The capability to learn and adapt to changes in the driving environment is\ncrucial for developing autonomous driving systems that are scalable beyond\ngeo-fenced operational design domains. Deep Reinforcement Learning (RL)\nprovides a promising and scalable framework for developing adaptive learning\nbased solutions. Deep RL methods usually model the problem as a (Partially\nObservable) Markov Decision Process in which an agent acts in a stationary\nenvironment to learn an optimal behavior policy. However, driving involves\ncomplex interaction between multiple, intelligent (artificial or human) agents\nin a highly non-stationary environment. In this paper, we propose the use of\nPartially Observable Markov Games(POSG) for formulating the connected\nautonomous driving problems with realistic assumptions. We provide a taxonomy\nof multi-agent learning environments based on the nature of tasks, nature of\nagents and the nature of the environment to help in categorizing various\nautonomous driving problems that can be addressed under the proposed\nformulation. As our main contributions, we provide MACAD-Gym, a Multi-Agent\nConnected, Autonomous Driving agent learning platform for furthering research\nin this direction. Our MACAD-Gym platform provides an extensible set of\nConnected Autonomous Driving (CAD) simulation environments that enable the\nresearch and development of Deep RL- based integrated sensing, perception,\nplanning and control algorithms for CAD systems with unlimited operational\ndesign domain under realistic, multi-agent settings. We also share the\nMACAD-Agents that were trained successfully using the MACAD-Gym platform to\nlearn control policies for multiple vehicle agents in a partially observable,\nstop-sign controlled, 3-way urban intersection environment with raw (camera)\nsensor observations.', 'Reinforcement learning (RL) is a popular machine learning paradigm for game\nplaying, robotics control, and other sequential decision tasks. However, RL\nagents often have long learning times with high data requirements because they\nbegin by acting randomly. In order to better learn in complex tasks, this\narticle argues that an external teacher can often significantly help the RL\nagent learn.\n  OpenAI Gym is a common framework for RL research, including a large number of\nstandard environments and agents, making RL research significantly more\naccessible. This article introduces our new open-source RL framework, the Human\nInput Parsing Platform for Openai Gym (HIPPO Gym), and the design decisions\nthat went into its creation. The goal of this platform is to facilitate\nhuman-RL research, again lowering the bar so that more researchers can quickly\ninvestigate different ways that human teachers could assist RL agents,\nincluding learning from demonstrations, learning from feedback, or curriculum\nlearning.']","['We introduce AndroidEnv, an open-source platform for Reinforcement Learning\n(RL) research built on top of the Android ecosystem. AndroidEnv allows RL\nagents to interact with a wide variety of apps and services commonly used by\nhumans through a universal touchscreen interface. Since agents train on a\nrealistic simulation of an Android device, they have the potential to be\ndeployed on real devices. In this report, we give an overview of the\nenvironment, highlighting the significant features it provides for research,\nand we present an empirical evaluation of some popular reinforcement learning\nagents on a set of tasks built on this platform.', ""Reinforcement Learning AI commonly uses reward/penalty signals that are\nobjective and explicit in an environment -- e.g. game score, completion time,\netc. -- in order to learn the optimal strategy for task performance. However,\nHuman-AI interaction for such AI agents should include additional reinforcement\nthat is implicit and subjective -- e.g. human preferences for certain AI\nbehavior -- in order to adapt the AI behavior to idiosyncratic human\npreferences. Such adaptations would mirror naturally occurring processes that\nincrease trust and comfort during social interactions. Here, we show how a\nhybrid brain-computer-interface (hBCI), which detects an individual's level of\ninterest in objects/events in a virtual environment, can be used to adapt the\nbehavior of a Deep Reinforcement Learning AI agent that is controlling a\nvirtual autonomous vehicle. Specifically, we show that the AI learns a driving\nstrategy that maintains a safe distance from a lead vehicle, and most novelly,\npreferentially slows the vehicle when the human passengers of the vehicle\nencounter objects of interest. This adaptation affords an additional 20\\%\nviewing time for subjectively interesting objects. This is the first\ndemonstration of how an hBCI can be used to provide implicit reinforcement to\nan AI agent in a way that incorporates user preferences into the control\nsystem."", ""Reinforcement learning (RL) agents are often designed specifically for a\nparticular problem and they generally have uninterpretable working processes.\nStatistical methods-based agent algorithms can be improved in terms of\ngeneralizability and interpretability using symbolic Artificial Intelligence\n(AI) tools such as logic programming. In this study, we present a model-free RL\narchitecture that is supported with explicit relational representations of the\nenvironmental objects. For the first time, we use the PrediNet network\narchitecture in a dynamic decision-making problem rather than image-based\ntasks, and Multi-Head Dot-Product Attention Network (MHDPA) as a baseline for\nperformance comparisons. We tested two networks in two environments ---i.e.,\nthe baseline Box-World environment and our novel environment,\nRelational-Grid-World (RGW). With the procedurally generated RGW environment,\nwhich is complex in terms of visual perceptions and combinatorial selections,\nit is easy to measure the relational representation performance of the RL\nagents. The experiments were carried out using different configurations of the\nenvironment so that the presented module and the environment were compared with\nthe baselines. We reached similar policy optimization performance results with\nthe PrediNet architecture and MHDPA; additionally, we achieved to extract the\npropositional representation explicitly ---which makes the agent's statistical\npolicy logic more interpretable and tractable. This flexibility in the agent's\npolicy provides convenience for designing non-task-specific agent\narchitectures. The main contributions of this study are two-fold ---an RL agent\nthat can explicitly perform relational reasoning, and a new environment that\nmeasures the relational reasoning capabilities of RL agents."", 'Human-robot collaboration is an essential research topic in artificial\nintelligence (AI), enabling researchers to devise cognitive AI systems and\naffords an intuitive means for users to interact with the robot. Of note,\ncommunication plays a central role. To date, prior studies in embodied agent\nnavigation have only demonstrated that human languages facilitate communication\nby instructions in natural languages. Nevertheless, a plethora of other forms\nof communication is left unexplored. In fact, human communication originated in\ngestures and oftentimes is delivered through multimodal cues, e.g. ""go there""\nwith a pointing gesture. To bridge the gap and fill in the missing dimension of\ncommunication in embodied agent navigation, we propose investigating the\neffects of using gestures as the communicative interface instead of verbal\ncues. Specifically, we develop a VR-based 3D simulation environment, named\nGes-THOR, based on AI2-THOR platform. In this virtual environment, a human\nplayer is placed in the same virtual scene and shepherds the artificial agent\nusing only gestures. The agent is tasked to solve the navigation problem guided\nby natural gestures with unknown semantics; we do not use any predefined\ngestures due to the diversity and versatile nature of human gestures. We argue\nthat learning the semantics of natural gestures is mutually beneficial to\nlearning the navigation task--learn to communicate and communicate to learn. In\na series of experiments, we demonstrate that human gesture cues, even without\npredefined semantics, improve the object-goal navigation for an embodied agent,\noutperforming various state-of-the-art methods.', ""We are increasingly surrounded by artificially intelligent technology that\ntakes decisions and executes actions on our behalf. This creates a pressing\nneed for general means to communicate with, instruct and guide artificial\nagents, with human language the most compelling means for such communication.\nTo achieve this in a scalable fashion, agents must be able to relate language\nto the world and to actions; that is, their understanding of language must be\ngrounded and embodied. However, learning grounded language is a notoriously\nchallenging problem in artificial intelligence research. Here we present an\nagent that learns to interpret language in a simulated 3D environment where it\nis rewarded for the successful execution of written instructions. Trained via a\ncombination of reinforcement and unsupervised learning, and beginning with\nminimal prior knowledge, the agent learns to relate linguistic symbols to\nemergent perceptual representations of its physical surroundings and to\npertinent sequences of actions. The agent's comprehension of language extends\nbeyond its prior experience, enabling it to apply familiar language to\nunfamiliar situations and to interpret entirely novel instructions. Moreover,\nthe speed with which this agent learns new words increases as its semantic\nknowledge grows. This facility for generalising and bootstrapping semantic\nknowledge indicates the potential of the present approach for reconciling\nambiguous natural language with the complexity of the physical world."", 'Cognitive Psychology and related disciplines have identified several critical\nmechanisms that enable intelligent biological agents to learn to solve complex\nproblems. There exists pressing evidence that the cognitive mechanisms that\nenable problem-solving skills in these species build on hierarchical mental\nrepresentations. Among the most promising computational approaches to provide\ncomparable learning-based problem-solving abilities for artificial agents and\nrobots is hierarchical reinforcement learning. However, so far the existing\ncomputational approaches have not been able to equip artificial agents with\nproblem-solving abilities that are comparable to intelligent animals, including\nhuman and non-human primates, crows, or octopuses. Here, we first survey the\nliterature in Cognitive Psychology, and related disciplines, and find that many\nimportant mental mechanisms involve compositional abstraction, curiosity, and\nforward models. We then relate these insights with contemporary hierarchical\nreinforcement learning methods, and identify the key machine intelligence\napproaches that realise these mechanisms. As our main result, we show that all\nimportant cognitive mechanisms have been implemented independently in isolated\ncomputational architectures, and there is simply a lack of approaches that\nintegrate them appropriately. We expect our results to guide the development of\nmore sophisticated cognitively inspired hierarchical methods, so that future\nartificial agents achieve a problem-solving performance on the level of\nintelligent animals.', 'Human beings, even small children, quickly become adept at figuring out how\nto use applications on their mobile devices. Learning to use a new app is often\nachieved via trial-and-error, accelerated by transfer of knowledge from past\nexperiences with like apps. The prospect of building a smarter smartphone - one\nthat can learn how to achieve tasks using mobile apps - is tantalizing. In this\npaper we explore the use of Reinforcement Learning (RL) with the goal of\nadvancing this aspiration. We introduce an RL-based framework for learning to\naccomplish tasks in mobile apps. RL agents are provided with states derived\nfrom the underlying representation of on-screen elements, and rewards that are\nbased on progress made in the task. Agents can interact with screen elements by\ntapping or typing. Our experimental results, over a number of mobile apps, show\nthat RL agents can learn to accomplish multi-step tasks, as well as achieve\nmodest generalization across different apps. More generally, we develop a\nplatform which addresses several engineering challenges to enable an effective\nRL training environment. Our AppBuddy platform is compatible with OpenAI Gym\nand includes a suite of mobile apps and benchmark tasks that supports a\ndiversity of RL research in the mobile app setting.', 'Social learning is a key component of human and animal intelligence. By\ntaking cues from the behavior of experts in their environment, social learners\ncan acquire sophisticated behavior and rapidly adapt to new circumstances. This\npaper investigates whether independent reinforcement learning (RL) agents in a\nmulti-agent environment can learn to use social learning to improve their\nperformance. We find that in most circumstances, vanilla model-free RL agents\ndo not use social learning. We analyze the reasons for this deficiency, and\nshow that by imposing constraints on the training environment and introducing a\nmodel-based auxiliary loss we are able to obtain generalized social learning\npolicies which enable agents to: i) discover complex skills that are not\nlearned from single-agent training, and ii) adapt online to novel environments\nby taking cues from experts present in the new environment. In contrast, agents\ntrained with model-free RL or imitation learning generalize poorly and do not\nsucceed in the transfer tasks. By mixing multi-agent and solo training, we can\nobtain agents that use social learning to gain skills that they can deploy when\nalone, even out-performing agents trained alone from the start.', 'This paper presents TrolleyMod v1.0, an open-source platform based on the\nCARLA simulator for the collection of ethical decision-making data for\nautonomous vehicles. This platform is designed to facilitate experiments aiming\nto observe and record human decisions and actions in high-fidelity simulations\nof ethical dilemmas that occur in the context of driving. Targeting experiments\nin the class of trolley problems, TrolleyMod provides a seamless approach to\ncreating new experimental settings and environments with the realistic\nphysics-engine and the high-quality graphical capabilities of CARLA and the\nUnreal Engine. Also, TrolleyMod provides a straightforward interface between\nthe CARLA environment and Python to enable the implementation of custom\ncontrollers, such as deep reinforcement learning agents. The results of such\nexperiments can be used for sociological analyses, as well as the training and\ntuning of value-aligned autonomous vehicles based on social values that are\ninferred from observations.', 'Navigating and understanding the real world remains a key challenge in\nmachine learning and inspires a great variety of research in areas such as\nlanguage grounding, planning, navigation and computer vision. We propose an\ninstruction-following task that requires all of the above, and which combines\nthe practicality of simulated environments with the challenges of ambiguous,\nnoisy real world data. StreetNav is built on top of Google Street View and\nprovides visually accurate environments representing real places. Agents are\ngiven driving instructions which they must learn to interpret in order to\nsuccessfully navigate in this environment. Since humans equipped with driving\ninstructions can readily navigate in previously unseen cities, we set a high\nbar and test our trained agents for similar cognitive capabilities. Although\ndeep reinforcement learning (RL) methods are frequently evaluated only on data\nthat closely follow the training distribution, our dataset extends to multiple\ncities and has a clean train/test separation. This allows for thorough testing\nof generalisation ability. This paper presents the StreetNav environment and\ntasks, models that establish strong baselines, and extensive analysis of the\ntask and the trained agents.', 'We introduce a large scale crowdsourced text adventure game as a research\nplatform for studying grounded dialogue. In it, agents can perceive, emote, and\nact whilst conducting dialogue with other agents. Models and humans can both\nact as characters within the game. We describe the results of training\nstate-of-the-art generative and retrieval models in this setting. We show that\nin addition to using past dialogue, these models are able to effectively use\nthe state of the underlying world to condition their predictions. In\nparticular, we show that grounding on the details of the local environment,\nincluding location descriptions, and the objects (and their affordances) and\ncharacters (and their previous actions) present within it allows better\npredictions of agent behavior and dialogue. We analyze the ingredients\nnecessary for successful grounding in this setting, and how each of these\nfactors relate to agents that can talk and act successfully.', 'Recent advances in artificial intelligence have been driven by the presence\nof increasingly realistic and complex simulated environments. However, many of\nthe existing environments provide either unrealistic visuals, inaccurate\nphysics, low task complexity, restricted agent perspective, or a limited\ncapacity for interaction among artificial agents. Furthermore, many platforms\nlack the ability to flexibly configure the simulation, making the simulated\nenvironment a black-box from the perspective of the learning system. In this\nwork, we propose a novel taxonomy of existing simulation platforms and discuss\nthe highest level class of general platforms which enable the development of\nlearning environments that are rich in visual, physical, task, and social\ncomplexity. We argue that modern game engines are uniquely suited to act as\ngeneral platforms and as a case study examine the Unity engine and open source\nUnity ML-Agents Toolkit. We then survey the research enabled by Unity and the\nUnity ML-Agents Toolkit, discussing the kinds of research a flexible,\ninteractive and easily configurable general platform can facilitate.', ""Artificial intelligence systems increasingly involve continual learning to\nenable flexibility in general situations that are not encountered during system\ntraining. Human interaction with autonomous systems is broadly studied, but\nresearch has hitherto under-explored interactions that occur while the system\nis actively learning, and can noticeably change its behaviour in minutes. In\nthis pilot study, we investigate how the interaction between a human and a\ncontinually learning prediction agent develops as the agent develops\ncompetency. Additionally, we compare two different agent architectures to\nassess how representational choices in agent design affect the human-agent\ninteraction. We develop a virtual reality environment and a time-based\nprediction task wherein learned predictions from a reinforcement learning (RL)\nalgorithm augment human predictions. We assess how a participant's performance\nand behaviour in this task differs across agent types, using both quantitative\nand qualitative analyses. Our findings suggest that human trust of the system\nmay be influenced by early interactions with the agent, and that trust in turn\naffects strategic behaviour, but limitations of the pilot study rule out any\nconclusive statement. We identify trust as a key feature of interaction to\nfocus on when considering RL-based technologies, and make several\nrecommendations for modification to this study in preparation for a\nlarger-scale investigation. A video summary of this paper can be found at\nhttps://youtu.be/oVYJdnBqTwQ ."", 'Multi-Agent Reinforcement Learning (MARL) methods find optimal policies for\nagents that operate in the presence of other learning agents. Central to\nachieving this is how the agents coordinate. One way to coordinate is by\nlearning to communicate with each other. Can the agents develop a language\nwhile learning to perform a common task? In this paper, we formulate and study\na MARL problem where cooperative agents are connected to each other via a fixed\nunderlying network. These agents can communicate along the edges of this\nnetwork by exchanging discrete symbols. However, the semantics of these symbols\nare not predefined and, during training, the agents are required to develop a\nlanguage that helps them in accomplishing their goals. We propose a method for\ntraining these agents using emergent communication. We demonstrate the\napplicability of the proposed framework by applying it to the problem of\nmanaging traffic controllers, where we achieve state-of-the-art performance as\ncompared to a number of strong baselines. More importantly, we perform a\ndetailed analysis of the emergent communication to show, for instance, that the\ndeveloped language is grounded and demonstrate its relationship with the\nunderlying network topology. To the best of our knowledge, this is the only\nwork that performs an in depth analysis of emergent communication in a\nnetworked MARL setting while being applicable to a broad class of problems.', ""While deep reinforcement learning (deep RL) agents are effective at\nmaximizing rewards, it is often unclear what strategies they use to do so. In\nthis paper, we take a step toward explaining deep RL agents through a case\nstudy using Atari 2600 environments. In particular, we focus on using saliency\nmaps to understand how an agent learns and executes a policy. We introduce a\nmethod for generating useful saliency maps and use it to show 1) what strong\nagents attend to, 2) whether agents are making decisions for the right or wrong\nreasons, and 3) how agents evolve during learning. We also test our method on\nnon-expert human subjects and find that it improves their ability to reason\nabout these agents. Overall, our results show that saliency information can\nprovide significant insight into an RL agent's decisions and learning behavior."", ""Applying Deep Reinforcement Learning (DRL) to Human-Robot Cooperation (HRC)\nin dynamic control problems is promising yet challenging as the robot needs to\nlearn the dynamics of the controlled system and dynamics of the human partner.\nIn existing research, the robot powered by DRL adopts coupled observation of\nthe environment and the human partner to learn both dynamics simultaneously.\nHowever, such a learning strategy is limited in terms of learning efficiency\nand team performance. This work proposes a novel task decomposition method with\na hierarchical reward mechanism that enables the robot to learn the\nhierarchical dynamic control task separately from learning the human partner's\nbehavior. The method is validated with a hierarchical control task in a\nsimulated environment with human subject experiments. Our method also provides\ninsight into the design of the learning strategy for HRC. The results show that\nthe robot should learn the task first to achieve higher team performance and\nlearn the human first to achieve higher learning efficiency."", 'Learning agents that are not only capable of taking tests, but also\ninnovating is becoming a hot topic in AI. One of the most promising paths\ntowards this vision is multi-agent learning, where agents act as the\nenvironment for each other, and improving each agent means proposing new\nproblems for others. However, existing evaluation platforms are either not\ncompatible with multi-agent settings, or limited to a specific game. That is,\nthere is not yet a general evaluation platform for research on multi-agent\nintelligence. To this end, we introduce Arena, a general evaluation platform\nfor multi-agent intelligence with 35 games of diverse logics and\nrepresentations. Furthermore, multi-agent intelligence is still at the stage\nwhere many problems remain unexplored. Therefore, we provide a building toolkit\nfor researchers to easily invent and build novel multi-agent problems from the\nprovided game set based on a GUI-configurable social tree and five basic\nmulti-agent reward schemes. Finally, we provide Python implementations of five\nstate-of-the-art deep multi-agent reinforcement learning baselines. Along with\nthe baseline implementations, we release a set of 100 best agents/teams that we\ncan train with different training schemes for each game, as the base for\nevaluating agents with population performance. As such, the research community\ncan perform comparisons under a stable and uniform standard. All the\nimplementations and accompanied tutorials have been open-sourced for the\ncommunity at https://sites.google.com/view/arena-unity/.', 'Reinforcement Learning (RL) agents have great successes in solving tasks with\nlarge observation and action spaces from limited feedback. Still, training the\nagents is data-intensive and there are no guarantees that the learned behavior\nis safe and does not violate rules of the environment, which has limitations\nfor the practical deployment in real-world scenarios. This paper discusses the\nengineering of reliable agents via the integration of deep RL with\nconstraint-based augmentation models to guide the RL agent towards safe\nbehavior. Within the constraints set, the RL agent is free to adapt and\nexplore, such that its effectiveness to solve the given problem is not\nhindered. However, once the RL agent leaves the space defined by the\nconstraints, the outside models can provide guidance to still work reliably. We\ndiscuss integration points for constraint guidance within the RL process and\nperform experiments on two case studies: a strictly constrained card game and a\ngrid world environment with additional combinatorial subgoals. Our results show\nthat constraint-guidance does both provide reliability improvements and safer\nbehavior, as well as accelerated training.', ""We introduce MAgent, a platform to support research and development of\nmany-agent reinforcement learning. Unlike previous research platforms on single\nor multi-agent reinforcement learning, MAgent focuses on supporting the tasks\nand the applications that require hundreds to millions of agents. Within the\ninteractions among a population of agents, it enables not only the study of\nlearning algorithms for agents' optimal polices, but more importantly, the\nobservation and understanding of individual agent's behaviors and social\nphenomena emerging from the AI society, including communication languages,\nleaderships, altruism. MAgent is highly scalable and can host up to one million\nagents on a single GPU server. MAgent also provides flexible configurations for\nAI researchers to design their customized environments and agents. In this\ndemo, we present three environments designed on MAgent and show emerged\ncollective intelligence by learning from scratch."", 'A good dialogue agent should have the ability to interact with users by both\nresponding to questions and by asking questions, and importantly to learn from\nboth types of interaction. In this work, we explore this direction by designing\na simulator and a set of synthetic tasks in the movie domain that allow such\ninteractions between a learner and a teacher. We investigate how a learner can\nbenefit from asking questions in both offline and online reinforcement learning\nsettings, and demonstrate that the learner improves when asking questions.\nFinally, real experiments with Mechanical Turk validate the approach. Our work\nrepresents a first step in developing such end-to-end learned interactive\ndialogue agents.', 'Humans and other intelligent animals evolved highly sophisticated perception\nsystems that combine multiple sensory modalities. On the other hand,\nstate-of-the-art artificial agents rely mostly on visual inputs or structured\nlow-dimensional observations provided by instrumented environments. Learning to\nact based on combined visual and auditory inputs is still a new topic of\nresearch that has not been explored beyond simple scenarios. To facilitate\nprogress in this area we introduce a new version of VizDoom simulator to create\na highly efficient learning environment that provides raw audio observations.\nWe study the performance of different model architectures in a series of tasks\nthat require the agent to recognize sounds and execute instructions given in\nnatural language. Finally, we train our agent to play the full game of Doom and\nfind that it can consistently defeat a traditional vision-based adversary. We\nare currently in the process of merging the augmented simulator with the main\nViZDoom code repository. Video demonstrations and experiment code can be found\nat https://sites.google.com/view/sound-rl.', ""Inferring other agents' mental states such as their knowledge, beliefs and\nintentions is thought to be essential for effective interactions with other\nagents. Recently, multiagent systems trained via deep reinforcement learning\nhave been shown to succeed in solving different tasks, but it remains unclear\nhow each agent modeled or represented other agents in their environment. In\nthis work we test whether deep reinforcement learning agents explicitly\nrepresent other agents' intentions (their specific aims or goals) during a task\nin which the agents had to coordinate the covering of different spots in a 2D\nenvironment. In particular, we tracked over time the performance of a linear\ndecoder trained to predict the final goal of all agents from the hidden state\nof each agent's neural network controller. We observed that the hidden layers\nof agents represented explicit information about other agents' goals, i.e. the\ntarget landmark they ended up covering. We also performed a series of\nexperiments, in which some agents were replaced by others with fixed goals, to\ntest the level of generalization of the trained agents. We noticed that during\nthe training phase the agents developed a differential preference for each\ngoal, which hindered generalization. To alleviate the above problem, we propose\nsimple changes to the MADDPG training algorithm which leads to better\ngeneralization against unseen agents. We believe that training protocols\npromoting more active intention reading mechanisms, e.g. by preventing simple\nsymmetry-breaking solutions, is a promising direction towards achieving a more\nrobust generalization in different cooperative and competitive tasks."", 'Behavior Trees are commonly used to model agents for robotics and games,\nwhere constrained behaviors must be designed by human experts in order to\nguarantee that these agents will execute a specific chain of actions given a\nspecific set of perceptions. In such application areas, learning is a desirable\nfeature to provide agents with the ability to adapt and improve interactions\nwith humans and environment, but often discarded due to its unreliability. In\nthis paper, we propose a framework that uses Reinforcement Learning nodes as\npart of Behavior Trees to address the problem of adding learning capabilities\nin constrained agents. We show how this framework relates to Options in\nHierarchical Reinforcement Learning, ensuring convergence of nested learning\nnodes, and we empirically show that the learning nodes do not affect the\nexecution of other nodes in the tree.', 'As AI continues to advance, human-AI teams are inevitable. However, progress\nin AI is routinely measured in isolation, without a human in the loop. It is\ncrucial to benchmark progress in AI, not just in isolation, but also in terms\nof how it translates to helping humans perform certain tasks, i.e., the\nperformance of human-AI teams.\n  In this work, we design a cooperative game - GuessWhich - to measure human-AI\nteam performance in the specific context of the AI being a visual\nconversational agent. GuessWhich involves live interaction between the human\nand the AI. The AI, which we call ALICE, is provided an image which is unseen\nby the human. Following a brief description of the image, the human questions\nALICE about this secret image to identify it from a fixed pool of images.\n  We measure performance of the human-ALICE team by the number of guesses it\ntakes the human to correctly identify the secret image after a fixed number of\ndialog rounds with ALICE. We compare performance of the human-ALICE teams for\ntwo versions of ALICE. Our human studies suggest a counterintuitive trend -\nthat while AI literature shows that one version outperforms the other when\npaired with an AI questioner bot, we find that this improvement in AI-AI\nperformance does not translate to improved human-AI performance. This suggests\na mismatch between benchmarking of AI in isolation and in the context of\nhuman-AI teams.', 'Deep reinforcement learning augments the reinforcement learning framework and\nutilizes the powerful representation of deep neural networks. Recent works have\ndemonstrated the remarkable successes of deep reinforcement learning in various\ndomains including finance, medicine, healthcare, video games, robotics, and\ncomputer vision. In this work, we provide a detailed review of recent and\nstate-of-the-art research advances of deep reinforcement learning in computer\nvision. We start with comprehending the theories of deep learning,\nreinforcement learning, and deep reinforcement learning. We then propose a\ncategorization of deep reinforcement learning methodologies and discuss their\nadvantages and limitations. In particular, we divide deep reinforcement\nlearning into seven main categories according to their applications in computer\nvision, i.e. (i)landmark localization (ii) object detection; (iii) object\ntracking; (iv) registration on both 2D image and 3D image volumetric data (v)\nimage segmentation; (vi) videos analysis; and (vii) other applications. Each of\nthese categories is further analyzed with reinforcement learning techniques,\nnetwork design, and performance. Moreover, we provide a comprehensive analysis\nof the existing publicly available datasets and examine source code\navailability. Finally, we present some open issues and discuss future research\ndirections on deep reinforcement learning in computer vision', 'Reinforcement learning (RL) agents in human-computer interactions\napplications require repeated user interactions before they can perform well.\nTo address this ""cold start"" problem, we propose a novel approach of using\ncognitive models to pre-train RL agents before they are applied to real users.\nAfter briefly reviewing relevant cognitive models, we present our general\nmethodological approach, followed by two case studies from our previous and\nongoing projects. We hope this position paper stimulates conversations between\nRL, HCI, and cognitive science researchers in order to explore the full\npotential of the approach.', 'Evidence that visual communication preceded written language and provided a\nbasis for it goes back to prehistory, in forms such as cave and rock paintings\ndepicting traces of our distant ancestors. Emergent communication research has\nsought to explore how agents can learn to communicate in order to\ncollaboratively solve tasks. Existing research has focused on language, with a\nlearned communication channel transmitting sequences of discrete tokens between\nthe agents. In this work, we explore a visual communication channel between\nagents that are allowed to draw with simple strokes. Our agents are\nparameterised by deep neural networks, and the drawing procedure is\ndifferentiable, allowing for end-to-end training. In the framework of a\nreferential communication game, we demonstrate that agents can not only\nsuccessfully learn to communicate by drawing, but with appropriate inductive\nbiases, can do so in a fashion that humans can interpret. We hope to encourage\nfuture research to consider visual communication as a more flexible and\ndirectly interpretable alternative of training collaborative agents.', 'A common vision from science fiction is that robots will one day inhabit our\nphysical spaces, sense the world as we do, assist our physical labours, and\ncommunicate with us through natural language. Here we study how to design\nartificial agents that can interact naturally with humans using the\nsimplification of a virtual environment. This setting nevertheless integrates a\nnumber of the central challenges of artificial intelligence (AI) research:\ncomplex visual perception and goal-directed physical control, grounded language\ncomprehension and production, and multi-agent social interaction. To build\nagents that can robustly interact with humans, we would ideally train them\nwhile they interact with humans. However, this is presently impractical.\nTherefore, we approximate the role of the human with another learned agent, and\nuse ideas from inverse reinforcement learning to reduce the disparities between\nhuman-human and agent-agent interactive behaviour. Rigorously evaluating our\nagents poses a great challenge, so we develop a variety of behavioural tests,\nincluding evaluation by humans who watch videos of agents or interact directly\nwith them. These evaluations convincingly demonstrate that interactive training\nand auxiliary losses improve agent behaviour beyond what is achieved by\nsupervised learning of actions alone. Further, we demonstrate that agent\ncapabilities generalise beyond literal experiences in the dataset. Finally, we\ntrain evaluation models whose ratings of agents agree well with human\njudgement, thus permitting the evaluation of new agent models without\nadditional effort. Taken together, our results in this virtual environment\nprovide evidence that large-scale human behavioural imitation is a promising\ntool to create intelligent, interactive agents, and the challenge of reliably\nevaluating such agents is possible to surmount.', 'Deep reinforcement learning has learned to play many games well, but failed\non others. To better characterize the modes and reasons of failure of deep\nreinforcement learners, we test the widely used Asynchronous Actor-Critic (A2C)\nalgorithm on four deceptive games, which are specially designed to provide\nchallenges to game-playing agents. These games are implemented in the General\nVideo Game AI framework, which allows us to compare the behavior of\nreinforcement learning-based agents with planning agents based on tree search.\nWe find that several of these games reliably deceive deep reinforcement\nlearners, and that the resulting behavior highlights the shortcomings of the\nlearning algorithm. The particular ways in which agents fail differ from how\nplanning-based agents fail, further illuminating the character of these\nalgorithms. We propose an initial typology of deceptions which could help us\nbetter understand pitfalls and failure modes of (deep) reinforcement learning.', 'AI systems have seen dramatic advancement in recent years, bringing many\napplications that pervade our everyday life. However, we are still mostly\nseeing instances of narrow AI: many of these recent developments are typically\nfocused on a very limited set of competencies and goals, e.g., image\ninterpretation, natural language processing, classification, prediction, and\nmany others. Moreover, while these successes can be accredited to improved\nalgorithms and techniques, they are also tightly linked to the availability of\nhuge datasets and computational power. State-of-the-art AI still lacks many\ncapabilities that would naturally be included in a notion of (human)\nintelligence.\n  We argue that a better study of the mechanisms that allow humans to have\nthese capabilities can help us understand how to imbue AI systems with these\ncompetencies. We focus especially on D. Kahneman\'s theory of thinking fast and\nslow, and we propose a multi-agent AI architecture where incoming problems are\nsolved by either system 1 (or ""fast"") agents, that react by exploiting only\npast experience, or by system 2 (or ""slow"") agents, that are deliberately\nactivated when there is the need to reason and search for optimal solutions\nbeyond what is expected from the system 1 agent. Both kinds of agents are\nsupported by a model of the world, containing domain knowledge about the\nenvironment, and a model of ""self"", containing information about past actions\nof the system and solvers\' skills.', 'Advances in reinforcement learning (RL) have resulted in recent breakthroughs\nin the application of artificial intelligence (AI) across many different\ndomains. An emerging landscape of development environments is making powerful\nRL techniques more accessible for a growing community of researchers. However,\nmost existing frameworks do not directly address the problem of learning in\ncomplex operating environments, such as dense urban settings or defense-related\nscenarios, that incorporate distributed, heterogeneous teams of agents. To help\nenable AI research for this important class of applications, we introduce the\nAI Arena: a scalable framework with flexible abstractions for distributed\nmulti-agent reinforcement learning. The AI Arena extends the OpenAI Gym\ninterface to allow greater flexibility in learning control policies across\nmultiple agents with heterogeneous learning strategies and localized views of\nthe environment. To illustrate the utility of our framework, we present\nexperimental results that demonstrate performance gains due to a distributed\nmulti-agent learning approach over commonly-used RL techniques in several\ndifferent learning environments.', 'When deployed, AI agents will encounter problems that are beyond their\nautonomous problem-solving capabilities. Leveraging human assistance can help\nagents overcome their inherent limitations and robustly cope with unfamiliar\nsituations. We present a general interactive framework that enables an agent to\nrequest and interpret rich, contextually useful information from an assistant\nthat has knowledge about the task and the environment. We demonstrate the\npracticality of our framework on a simulated human-assisted navigation problem.\nAided with an assistance-requesting policy learned by our method, a navigation\nagent achieves up to a 7x improvement in success rate on tasks that take place\nin previously unseen environments, compared to fully autonomous behavior. We\nshow that the agent can take advantage of different types of information\ndepending on the context, and analyze the benefits and challenges of learning\nthe assistance-requesting policy when the assistant can recursively decompose\ntasks into subtasks.', 'Deep reinforcement learning has the potential to train robots to perform\ncomplex tasks in the real world without requiring accurate models of the robot\nor its environment. A practical approach is to train agents in simulation, and\nthen transfer them to the real world. One popular method for achieving\ntransferability is to use domain randomisation, which involves randomly\nperturbing various aspects of a simulated environment in order to make trained\nagents robust to the reality gap. However, less work has gone into\nunderstanding such agents - which are deployed in the real world - beyond task\nperformance. In this work we examine such agents, through qualitative and\nquantitative comparisons between agents trained with and without visual domain\nrandomisation. We train agents for Fetch and Jaco robots on a visuomotor\ncontrol task and evaluate how well they generalise using different testing\nconditions. Finally, we investigate the internals of the trained agents by\nusing a suite of interpretability techniques. Our results show that the primary\noutcome of domain randomisation is more robust, entangled representations,\naccompanied with larger weights with greater spatial structure; moreover, the\ntypes of changes are heavily influenced by the task setup and presence of\nadditional proprioceptive inputs. Additionally, we demonstrate that our domain\nrandomised agents require higher sample complexity, can overfit and more\nheavily rely on recurrent processing. Furthermore, even with an improved\nsaliency method introduced in this work, we show that qualitative studies may\nnot always correspond with quantitative measures, necessitating the combination\nof inspection tools in order to provide sufficient insights into the behaviour\nof trained agents.', ""Movement coordination in human ensembles has been studied little in the\ncurrent literature. In the existing experimental works, situations where all\nsubjects are connected with each other through direct visual and auditory\ncoupling, and social interaction affects their coordination, have been\ninvestigated. Here, we study coordination in human ensembles via a novel\ncomputer-based set-up that enables individuals to coordinate each other's\nmotion from a distance so as to minimize the influence of social interaction.\nThe proposed platform makes it possible to implement different visual\ninteraction patterns among the players, so that participants take into\nconsideration the motion of a designated subset of the others. This allows the\nevaluation of the exclusive effects on coordination of the structure of\ninterconnections among the players and their own dynamics. Our set-up enables\nalso the deployment of virtual players to investigate dyadic interaction\nbetween a human and a virtual agent, as well as group synchronization in mixed\nteams of human and virtual agents. We use this novel set-up to study\ncoordination both in dyads and in groups over different structures of\ninterconnections, with and without virtual agents. We find that, in dual\ninteraction, virtual players manage to interact with participants in a\nhuman-like fashion, thus confirming findings in previous work. We also observe\nthat, in group interaction, the level of coordination among humans in the\nabsence of direct visual and auditory coupling depends on the structure of\ninterconnections among participants. This confirms, as recently suggested in\nthe literature, that different coordination levels are achieved over diverse\nvisual pairings in the presence and in the absence of social interaction. We\npresent preliminary experimental results on the effect on group coordination of\ndeploying virtual computer agents in the human ensemble."", ""This paper presents the first two editions of Visual Doom AI Competition,\nheld in 2016 and 2017. The challenge was to create bots that compete in a\nmulti-player deathmatch in a first-person shooter (FPS) game, Doom. The bots\nhad to make their decisions based solely on visual information, i.e., a raw\nscreen buffer. To play well, the bots needed to understand their surroundings,\nnavigate, explore, and handle the opponents at the same time. These aspects,\ntogether with the competitive multi-agent aspect of the game, make the\ncompetition a unique platform for evaluating the state of the art reinforcement\nlearning algorithms. The paper discusses the rules, solutions, results, and\nstatistics that give insight into the agents' behaviors. Best-performing agents\nare described in more detail. The results of the competition lead to the\nconclusion that, although reinforcement learning can produce capable Doom bots,\nthey still are not yet able to successfully compete against humans in this\ngame. The paper also revisits the ViZDoom environment, which is a flexible,\neasy to use, and efficient 3D platform for research for vision-based\nreinforcement learning, based on a well-recognized first-person perspective\ngame Doom."", 'Even with impressive advances in application-specific models, we still lack\nknowledge about how to build a model that can learn in a human-like way and do\nmultiple tasks. To learn in a human-like way, we need to provide a diverse\nexperience that is comparable to humans. In this paper, we introduce our\nongoing effort to build a simulated environment for developmental robotics\n(SEDRo). SEDRo provides diverse human experiences ranging from those of a fetus\nto a 12th-month-old. A series of simulated tests based on developmental\npsychology will be used to evaluate the progress of a learning model. We\nanticipate SEDRo to lower the cost of entry and facilitate research in the\ndevelopmental robotics community.', ""What might sound like the beginning of a joke has become an attractive\nprospect for many cognitive scientists: the use of deep neural network models\n(DNNs) as models of human behavior in perceptual and cognitive tasks. Although\nDNNs have taken over machine learning, attempts to use them as models of human\nbehavior are still in the early stages. Can they become a versatile model class\nin the cognitive scientist's toolbox? We first argue why DNNs have the\npotential to be interesting models of human behavior. We then discuss how that\npotential can be more fully realized. On the one hand, we argue that the cycle\nof training, testing, and revising DNNs needs to be revisited through the lens\nof the cognitive scientist's goals. Specifically, we argue that methods for\nassessing the goodness of fit between DNN models and human behavior have to\ndate been impoverished. On the other hand, cognitive science might have to\nstart using more complex tasks (including richer stimulus spaces), but doing so\nmight be beneficial for DNN-independent reasons as well. Finally, we highlight\navenues where traditional cognitive process models and DNNs may show productive\nsynergy."", 'The measurement of time is central to intelligent behavior. We know that both\nanimals and artificial agents can successfully use temporal dependencies to\nselect actions. In artificial agents, little work has directly addressed (1)\nwhich architectural components are necessary for successful development of this\nability, (2) how this timing ability comes to be represented in the units and\nactions of the agent, and (3) whether the resulting behavior of the system\nconverges on solutions similar to those of biology. Here we studied interval\ntiming abilities in deep reinforcement learning agents trained end-to-end on an\ninterval reproduction paradigm inspired by experimental literature on\nmechanisms of timing. We characterize the strategies developed by recurrent and\nfeedforward agents, which both succeed at temporal reproduction using distinct\nmechanisms, some of which bear specific and intriguing similarities to\nbiological systems. These findings advance our understanding of how agents come\nto represent time, and they highlight the value of experimentally inspired\napproaches to characterizing agent abilities.', 'Building intelligent agents that can communicate with and learn from humans\nin natural language is of great value. Supervised language learning is limited\nby the ability of capturing mainly the statistics of training data, and is\nhardly adaptive to new scenarios or flexible for acquiring new knowledge\nwithout inefficient retraining or catastrophic forgetting. We highlight the\nperspective that conversational interaction serves as a natural interface both\nfor language learning and for novel knowledge acquisition and propose a joint\nimitation and reinforcement approach for grounded language learning through an\ninteractive conversational game. The agent trained with this approach is able\nto actively acquire information by asking questions about novel objects and use\nthe just-learned knowledge in subsequent conversations in a one-shot fashion.\nResults compared with other methods verified the effectiveness of the proposed\napproach.', ""Interactive Machine Learning is concerned with creating systems that operate\nin environments alongside humans to achieve a task. A typical use is to extend\nor amplify the capabilities of a human in cognitive or physical ways, requiring\nthe machine to adapt to the users' intentions and preferences. Often, this\ntakes the form of a human operator providing some type of feedback to the user,\nwhich can be explicit feedback, implicit feedback, or a combination of both.\nExplicit feedback, such as through a mouse click, carries a high cognitive\nload. The focus of this study is to extend the current state of the art in\ninteractive machine learning by demonstrating that agents can learn a human\nuser's behavior and adapt to preferences with a reduced amount of explicit\nhuman feedback in a mixed feedback setting. The learning agent perceives a\nvalue of its own behavior from hand gestures given via a spatial interface.\nThis feedback mechanism is termed Spatial Interface Valuing. This method is\nevaluated experimentally in a simulated environment for a grasping task using a\nrobotic arm with variable grip settings. Preliminary results indicate that\nlearning agents using spatial interface valuing can learn a value function\nmapping spatial gestures to expected future rewards much more quickly as\ncompared to those same agents just receiving explicit feedback, demonstrating\nthat an agent perceiving feedback from a human user via a spatial interface can\nserve as an effective complement to existing approaches."", 'Future robots should follow human social norms in order to be useful and\naccepted in human society. In this paper, we leverage already existing social\nknowledge in human societies by capturing it in our framework through the\nnotion of social norms. We show how norms can be used to guide a reinforcement\nlearning agent towards achieving normative behavior and apply the same set of\nnorms over different domains. Thus, we are able to: (1) provide a way to\nintuitively encode social knowledge (through norms); (2) guide learning towards\nnormative behaviors (through an automatic norm reward system); and (3) achieve\na transfer of learning by abstracting policies; Finally, (4) the method is not\ndependent on a particular RL algorithm. We show how our approach can be seen as\na means to achieve abstract representation and learn procedural knowledge based\non the declarative semantics of norms and discuss possible implications of this\nin some areas of cognitive science.', ""We present a method of endowing agents in an agent-based model (ABM) with\nsophisticated cognitive capabilities and a naturally tunable level of\nintelligence. Often, ABMs use random behavior or greedy algorithms for\nmaximizing objectives (such as a predator always chasing after the closest\nprey). However, random behavior is too simplistic in many circumstances and\ngreedy algorithms, as well as classic AI planning techniques, can be brittle in\nthe context of the unpredictable and emergent situations in which agents may\nfind themselves. Our method, called agent-centric Monte Carlo cognition\n(ACMCC), centers around using a separate agent-based model to represent the\nagents' cognition. This model is then used by the agents in the primary model\nto predict the outcomes of their actions, and thus guide their behavior. To\nthat end, we have implemented our method in the NetLogo agent-based modeling\nplatform, using the recently released LevelSpace extension, which we developed\nto allow NetLogo models to interact with other NetLogo models. As an\nillustrative example, we extend the Wolf Sheep Predation model (included with\nNetLogo) by using ACMCC to guide animal behavior, and analyze the impact on\nagent performance and model dynamics. We find that ACMCC provides a reliable\nand understandable method of controlling agent intelligence, and has a large\nimpact on agent performance and model dynamics even at low settings."", ""Multi-agent reinforcement learning (MARL) extends (single-agent)\nreinforcement learning (RL) by introducing additional agents and (potentially)\npartial observability of the environment. Consequently, algorithms for solving\nMARL problems incorporate various extensions beyond traditional RL methods,\nsuch as a learned communication protocol between cooperative agents that\nenables exchange of private information or adaptive modeling of opponents in\ncompetitive settings. One popular algorithmic construct is a memory mechanism\nsuch that an agent's decisions can depend not only upon the current state but\nalso upon the history of observed states and actions. In this paper, we study\nhow a memory mechanism can be useful in environments with different properties,\nsuch as observability, internality and presence of a communication channel.\nUsing both prior work and new experiments, we show that a memory mechanism is\nhelpful when learning agents need to model other agents and/or when\ncommunication is constrained in some way; however we must to be cautious of\nagents achieving effective memoryfulness through other means."", 'We present Coevo, an online platform that allows both humans and artificial\nagents to design shapes that solve different tasks. Our goal is to explore\ncommon shared design tools that can be used by humans and artificial agents in\na context of creation. This approach can provide a better knowledge transfer\nand interaction with artificial agents since a common language of design is\ndefined. In this paper, we outline the main components of this platform and\ndiscuss the definition of a human-centered language to enhance human-AI\ncollaboration in co-creation scenarios.', 'Deep reinforcement learning (deep RL) has been successful in learning\nsophisticated behaviors automatically; however, the learning process requires a\nhuge number of trials. In contrast, animals can learn new tasks in just a few\ntrials, benefiting from their prior knowledge about the world. This paper seeks\nto bridge this gap. Rather than designing a ""fast"" reinforcement learning\nalgorithm, we propose to represent it as a recurrent neural network (RNN) and\nlearn it from data. In our proposed method, RL$^2$, the algorithm is encoded in\nthe weights of the RNN, which are learned slowly through a general-purpose\n(""slow"") RL algorithm. The RNN receives all information a typical RL algorithm\nwould receive, including observations, actions, rewards, and termination flags;\nand it retains its state across episodes in a given Markov Decision Process\n(MDP). The activations of the RNN store the state of the ""fast"" RL algorithm on\nthe current (previously unseen) MDP. We evaluate RL$^2$ experimentally on both\nsmall-scale and large-scale problems. On the small-scale side, we train it to\nsolve randomly generated multi-arm bandit problems and finite MDPs. After\nRL$^2$ is trained, its performance on new MDPs is close to human-designed\nalgorithms with optimality guarantees. On the large-scale side, we test RL$^2$\non a vision-based navigation task and show that it scales up to\nhigh-dimensional problems.', 'My overarching research goal is to provide robots with perceptional abilities\nthat allow interactions with humans in a human-like manner. To develop these\nperceptional abilities, I believe that it is useful to study the principles of\nthe human visual system. I use these principles to develop new computer vision\nalgorithms and validate their effectiveness in intelligent robotic systems. I\nam enthusiastic about this approach as it offers the dual benefit of uncovering\nprinciples inherent in the human visual system, as well as applying these\nprinciples to its artificial counterpart. Fig. 1 contains a depiction of my\nresearch.', 'For machine agents to successfully interact with humans in real-world\nsettings, they will need to develop an understanding of human mental life.\nIntuitive psychology, the ability to reason about hidden mental variables that\ndrive observable actions, comes naturally to people: even pre-verbal infants\ncan tell agents from objects, expecting agents to act efficiently to achieve\ngoals given constraints. Despite recent interest in machine agents that reason\nabout other agents, it is not clear if such agents learn or hold the core\npsychology principles that drive human reasoning. Inspired by cognitive\ndevelopment studies on intuitive psychology, we present a benchmark consisting\nof a large dataset of procedurally generated 3D animations, AGENT (Action,\nGoal, Efficiency, coNstraint, uTility), structured around four scenarios (goal\npreferences, action efficiency, unobserved constraints, and cost-reward\ntrade-offs) that probe key concepts of core intuitive psychology. We validate\nAGENT with human-ratings, propose an evaluation protocol emphasizing\ngeneralization, and compare two strong baselines built on Bayesian inverse\nplanning and a Theory of Mind neural network. Our results suggest that to pass\nthe designed tests of core intuitive psychology at human levels, a model must\nacquire or have built-in representations of how agents plan, combining utility\ncomputations and core knowledge of objects and physics.', 'In this paper, we argue that the future of Artificial Intelligence research\nresides in two keywords: integration and embodiment. We support this claim by\nanalyzing the recent advances of the field. Regarding integration, we note that\nthe most impactful recent contributions have been made possible through the\nintegration of recent Machine Learning methods (based in particular on Deep\nLearning and Recurrent Neural Networks) with more traditional ones (e.g.\nMonte-Carlo tree search, goal babbling exploration or addressable memory\nsystems). Regarding embodiment, we note that the traditional benchmark tasks\n(e.g. visual classification or board games) are becoming obsolete as\nstate-of-the-art learning algorithms approach or even surpass human performance\nin most of them, having recently encouraged the development of first-person 3D\ngame platforms embedding realistic physics. Building upon this analysis, we\nfirst propose an embodied cognitive architecture integrating heterogenous\nsub-fields of Artificial Intelligence into a unified framework. We demonstrate\nthe utility of our approach by showing how major contributions of the field can\nbe expressed within the proposed framework. We then claim that benchmarking\nenvironments need to reproduce ecologically-valid conditions for bootstrapping\nthe acquisition of increasingly complex cognitive skills through the concept of\na cognitive arms race between embodied agents.', 'Deep artificial neural networks, trained with labeled data sets are widely\nused in numerous vision and robotics applications today. In terms of AI, these\nare called reflex models, referring to the fact that they do not self-evolve or\nactively adapt to environmental changes. As demand for intelligent robot\ncontrol expands to many high level tasks, reinforcement learning and state\nbased models play an increasingly important role. Herein, in computer vision\nand robotics domain, we study a novel approach to add reinforcement controls\nonto the image recognition reflex models to attain better overall performance,\nspecifically to a wider environment range beyond what is expected of the task\nreflex models. Follow a common infrastructure with environment sensing and AI\nbased modeling of self-adaptive agents, we implement multiple types of AI\ncontrol agents. To the end, we provide comparative results of these agents with\nbaseline, and an insightful analysis of their benefit to improve overall image\nrecognition performance in real world.', 'Designing human-centered AI-driven applications require deep understandings\nof how people develop mental models of AI. Currently, we have little knowledge\nof this process and limited tools to study it. This paper presents the position\nthat AI-based games, particularly the player-AI interaction component, offer an\nideal domain to study the process in which mental models evolve. We present a\ncase study to illustrate the benefits of our approach for explainable AI.', 'Recent progress in artificial intelligence (AI) has renewed interest in\nbuilding systems that learn and think like people. Many advances have come from\nusing deep neural networks trained end-to-end in tasks such as object\nrecognition, video games, and board games, achieving performance that equals or\neven beats humans in some respects. Despite their biological inspiration and\nperformance achievements, these systems differ from human intelligence in\ncrucial ways. We review progress in cognitive science suggesting that truly\nhuman-like learning and thinking machines will have to reach beyond current\nengineering trends in both what they learn, and how they learn it.\nSpecifically, we argue that these machines should (a) build causal models of\nthe world that support explanation and understanding, rather than merely\nsolving pattern recognition problems; (b) ground learning in intuitive theories\nof physics and psychology, to support and enrich the knowledge that is learned;\nand (c) harness compositionality and learning-to-learn to rapidly acquire and\ngeneralize knowledge to new tasks and situations. We suggest concrete\nchallenges and promising routes towards these goals that can combine the\nstrengths of recent neural network advances with more structured cognitive\nmodels.', 'The objective of a reinforcement learning agent is to behave so as to\nmaximise the sum of a suitable scalar function of state: the reward. These\nrewards are typically given and immutable. In this paper, we instead consider\nthe proposition that the reward function itself can be a good locus of learned\nknowledge. To investigate this, we propose a scalable meta-gradient framework\nfor learning useful intrinsic reward functions across multiple lifetimes of\nexperience. Through several proof-of-concept experiments, we show that it is\nfeasible to learn and capture knowledge about long-term exploration and\nexploitation into a reward function. Furthermore, we show that unlike policy\ntransfer methods that capture ""how"" the agent should behave, the learned reward\nfunctions can generalise to other kinds of agents and to changes in the\ndynamics of the environment by capturing ""what"" the agent should strive to do.', 'Recently, there have been several high-profile achievements of agents\nlearning to play games against humans and beat them. In this paper, we study\nthe problem of training intelligent agents in service of game development.\nUnlike the agents built to ""beat the game"", our agents aim to produce\nhuman-like behavior to help with game evaluation and balancing. We discuss two\nfundamental metrics based on which we measure the human-likeness of agents,\nnamely skill and style, which are multi-faceted concepts with practical\nimplications outlined in this paper. We report four case studies in which the\nstyle and skill requirements inform the choice of algorithms and metrics used\nto train agents; ranging from A* search to state-of-the-art deep reinforcement\nlearning. We, further, show that the learning potential of state-of-the-art\ndeep RL models does not seamlessly transfer from the benchmark environments to\ntarget ones without heavily tuning their hyperparameters, leading to linear\nscaling of the engineering efforts and computational cost with the number of\ntarget domains.', 'We present a machine learning framework for multi-agent systems to learn both\nthe optimal policy for maximizing the rewards and the encoding of the high\ndimensional visual observation. The encoding is useful for sharing local visual\nobservations with other agents under communication resource constraints. The\nactor-encoder encodes the raw images and chooses an action based on local\nobservations and messages sent by the other agents. The machine learning agent\ngenerates not only an actuator command to the physical device, but also a\ncommunication message to the other agents. We formulate a reinforcement\nlearning problem, which extends the action space to consider the communication\naction as well. The feasibility of the reinforcement learning framework is\ndemonstrated using a 3D simulation environment with two collaborating agents.\nThe environment provides realistic visual observations to be used and shared\nbetween the two agents.', ""We propose an explainable reinforcement learning (XRL) framework that\nanalyzes an agent's history of interaction with the environment to extract\ninterestingness elements that help explain its behavior. The framework relies\non data readily available from standard RL algorithms, augmented with data that\ncan easily be collected by the agent while learning. We describe how to create\nvisual summaries of an agent's behavior in the form of short video-clips\nhighlighting key interaction moments, based on the proposed elements. We also\nreport on a user study where we evaluated the ability of humans to correctly\nperceive the aptitude of agents with different characteristics, including their\ncapabilities and limitations, given visual summaries automatically generated by\nour framework. The results show that the diversity of aspects captured by the\ndifferent interestingness elements is crucial to help humans correctly\nunderstand an agent's strengths and limitations in performing a task, and\ndetermine when it might need adjustments to improve its performance."", 'One of the long-term goals of artificial intelligence is to build an agent\nthat can communicate intelligently with human in natural language. Most\nexisting work on natural language learning relies heavily on training over a\npre-collected dataset with annotated labels, leading to an agent that\nessentially captures the statistics of the fixed external training data. As the\ntraining data is essentially a static snapshot representation of the knowledge\nfrom the annotator, the agent trained this way is limited in adaptiveness and\ngeneralization of its behavior. Moreover, this is very different from the\nlanguage learning process of humans, where language is acquired during\ncommunication by taking speaking action and learning from the consequences of\nspeaking action in an interactive manner. This paper presents an interactive\nsetting for grounded natural language learning, where an agent learns natural\nlanguage by interacting with a teacher and learning from feedback, thus\nlearning and improving language skills while taking part in the conversation.\nTo achieve this goal, we propose a model which incorporates both imitation and\nreinforcement by leveraging jointly sentence and reward feedbacks from the\nteacher. Experiments are conducted to validate the effectiveness of the\nproposed approach.', 'This paper assumes the hypothesis that human learning is perception based,\nand consequently, the learning process and perceptions should not be\nrepresented and investigated independently or modeled in different simulation\nspaces. In order to keep the analogy between the artificial and human learning,\nthe former is assumed here as being based on the artificial perception. Hence,\ninstead of choosing to apply or develop a Computational Theory of (human)\nPerceptions, we choose to mirror the human perceptions in a numeric\n(computational) space as artificial perceptions and to analyze the\ninterdependence between artificial learning and artificial perception in the\nsame numeric space, using one of the simplest tools of Artificial Intelligence\nand Soft Computing, namely the perceptrons. As practical applications, we\nchoose to work around two examples: Optical Character Recognition and Iris\nRecognition. In both cases a simple Turing test shows that artificial\nperceptions of the difference between two characters and between two irides are\nfuzzy, whereas the corresponding human perceptions are, in fact, crisp.', ""We study the task of embodied visual active learning, where an agent is set\nto explore a 3d environment with the goal to acquire visual scene understanding\nby actively selecting views for which to request annotation. While accurate on\nsome benchmarks, today's deep visual recognition pipelines tend to not\ngeneralize well in certain real-world scenarios, or for unusual viewpoints.\nRobotic perception, in turn, requires the capability to refine the recognition\ncapabilities for the conditions where the mobile system operates, including\ncluttered indoor environments or poor illumination. This motivates the proposed\ntask, where an agent is placed in a novel environment with the objective of\nimproving its visual recognition capability. To study embodied visual active\nlearning, we develop a battery of agents - both learnt and pre-specified - and\nwith different levels of knowledge of the environment. The agents are equipped\nwith a semantic segmentation network and seek to acquire informative views,\nmove and explore in order to propagate annotations in the neighbourhood of\nthose views, then refine the underlying segmentation network by online\nretraining. The trainable method uses deep reinforcement learning with a reward\nfunction that balances two competing objectives: task performance, represented\nas visual recognition accuracy, which requires exploring the environment, and\nthe necessary amount of annotated data requested during active exploration. We\nextensively evaluate the proposed models using the photorealistic Matterport3D\nsimulator and show that a fully learnt method outperforms comparable\npre-specified counterparts, even when requesting fewer annotations."", 'There has been an emerging paradigm shift from the era of ""internet AI"" to\n""embodied AI"", where AI algorithms and agents no longer learn from datasets of\nimages, videos or text curated primarily from the internet. Instead, they learn\nthrough interactions with their environments from an egocentric perception\nsimilar to humans. Consequently, there has been substantial growth in the\ndemand for embodied AI simulators to support various embodied AI research\ntasks. This growing interest in embodied AI is beneficial to the greater\npursuit of Artificial General Intelligence (AGI), but there has not been a\ncontemporary and comprehensive survey of this field. This paper aims to provide\nan encyclopedic survey for the field of embodied AI, from its simulators to its\nresearch. By evaluating nine current embodied AI simulators with our proposed\nseven features, this paper aims to understand the simulators in their provision\nfor use in embodied AI research and their limitations. Lastly, this paper\nsurveys the three main research tasks in embodied AI -- visual exploration,\nvisual navigation and embodied question answering (QA), covering the\nstate-of-the-art approaches, evaluation metrics and datasets. Finally, with the\nnew insights revealed through surveying the field, the paper will provide\nsuggestions for simulator-for-task selections and recommendations for the\nfuture directions of the field.', 'Many tasks in AI require the collaboration of multiple agents. Typically, the\ncommunication protocol between agents is manually specified and not altered\nduring training. In this paper we explore a simple neural model, called\nCommNet, that uses continuous communication for fully cooperative tasks. The\nmodel consists of multiple agents and the communication between them is learned\nalongside their policy. We apply this model to a diverse set of tasks,\ndemonstrating the ability of the agents to learn to communicate amongst\nthemselves, yielding improved performance over non-communicative agents and\nbaselines. In some cases, it is possible to interpret the language devised by\nthe agents, revealing simple but effective strategies for solving the task at\nhand.', 'Recent progress in artificial intelligence through reinforcement learning\n(RL) has shown great success on increasingly complex single-agent environments\nand two-player turn-based games. However, the real-world contains multiple\nagents, each learning and acting independently to cooperate and compete with\nother agents, and environments reflecting this degree of complexity remain an\nopen challenge. In this work, we demonstrate for the first time that an agent\ncan achieve human-level in a popular 3D multiplayer first-person video game,\nQuake III Arena Capture the Flag, using only pixels and game points as input.\nThese results were achieved by a novel two-tier optimisation process in which a\npopulation of independent RL agents are trained concurrently from thousands of\nparallel matches with agents playing in teams together and against each other\non randomly generated environments. Each agent in the population learns its own\ninternal reward signal to complement the sparse delayed reward from winning,\nand selects actions using a novel temporally hierarchical representation that\nenables the agent to reason at multiple timescales. During game-play, these\nagents display human-like behaviours such as navigating, following, and\ndefending based on a rich learned representation that is shown to encode\nhigh-level game knowledge. In an extensive tournament-style evaluation the\ntrained agents exceeded the win-rate of strong human players both as teammates\nand opponents, and proved far stronger than existing state-of-the-art agents.\nThese results demonstrate a significant jump in the capabilities of artificial\nagents, bringing us closer to the goal of human-level intelligence.', 'In this work we create agents that can perform well beyond a single,\nindividual task, that exhibit much wider generalisation of behaviour to a\nmassive, rich space of challenges. We define a universe of tasks within an\nenvironment domain and demonstrate the ability to train agents that are\ngenerally capable across this vast space and beyond. The environment is\nnatively multi-agent, spanning the continuum of competitive, cooperative, and\nindependent games, which are situated within procedurally generated physical 3D\nworlds. The resulting space is exceptionally diverse in terms of the challenges\nposed to agents, and as such, even measuring the learning progress of an agent\nis an open research problem. We propose an iterative notion of improvement\nbetween successive generations of agents, rather than seeking to maximise a\nsingular objective, allowing us to quantify progress despite tasks being\nincomparable in terms of achievable rewards. We show that through constructing\nan open-ended learning process, which dynamically changes the training task\ndistributions and training objectives such that the agent never stops learning,\nwe achieve consistent learning of new behaviours. The resulting agent is able\nto score reward in every one of our humanly solvable evaluation levels, with\nbehaviour generalising to many held-out points in the universe of tasks.\nExamples of this zero-shot generalisation include good performance on Hide and\nSeek, Capture the Flag, and Tag. Through analysis and hand-authored probe tasks\nwe characterise the behaviour of our agent, and find interesting emergent\nheuristic behaviours such as trial-and-error experimentation, simple tool use,\noption switching, and cooperation. Finally, we demonstrate that the general\ncapabilities of this agent could unlock larger scale transfer of behaviour\nthrough cheap finetuning.', 'Humans sometimes choose actions that they themselves can identify as\nsub-optimal, or wrong, even in the absence of additional information. How is\nthis possible? We present an algorithmic theory of metacognition based on a\nwell-understood trade-off in reinforcement learning (RL) between value-based RL\nand policy-based RL. To the cognitive (neuro)science community, our theory\nanswers the outstanding question of why information can be used for error\ndetection but not for action selection. To the machine learning community, our\nproposed theory creates a novel interaction between the Actor and Critic in\nActor-Critic agents and notes a novel connection between RL and Bayesian\nOptimization. We call our proposed agent the Metacognitive Actor Critic (MAC).\nWe conclude with showing how to create metacognition in machines by\nimplementing a deep MAC and showing that it can detect (some of) its own\nsuboptimal actions without external information or delay.', 'Several approaches have been developed for answering users\' specific\nquestions about AI behavior and for assessing their core functionality in terms\nof primitive executable actions. However, the problem of summarizing an AI\nagent\'s broad capabilities for a user is comparatively new. This paper presents\nan algorithm for discovering from scratch the suite of high-level\n""capabilities"" that an AI system with arbitrary internal planning\nalgorithms/policies can perform. It computes conditions describing the\napplicability and effects of these capabilities in user-interpretable terms.\nStarting from a set of user-interpretable state properties, an AI agent, and a\nsimulator that the agent can interact with, our algorithm returns a set of\nhigh-level capabilities with their parameterized descriptions. Empirical\nevaluation on several game-based scenarios shows that this approach efficiently\nlearns descriptions of various types of AI agents in deterministic, fully\nobservable settings. User studies show that such descriptions are easier to\nunderstand and reason with than the agent\'s primitive actions.', ""Recent advances in machine learning have made it possible to train\nartificially intelligent agents that perform with super-human accuracy on a\ngreat diversity of complex tasks. However, the process of training these\ncapabilities often necessitates millions of annotated examples -- far more than\nhumans typically need in order to achieve a passing level of mastery on similar\ntasks. Thus, while contemporary methods in machine learning can produce agents\nthat exhibit super-human performance, their rate of learning per opportunity in\nmany domains is decidedly lower than human-learning. In this work we formalize\na theory of Decomposed Inductive Procedure Learning (DIPL) that outlines how\ndifferent forms of inductive symbolic learning can be used in combination to\nbuild agents that learn educationally relevant tasks such as mathematical, and\nscientific procedures, at a rate similar to human learners. We motivate the\nconstruction of this theory along Marr's concepts of the computational,\nalgorithmic, and implementation levels of cognitive modeling, and outline at\nthe computational-level six learning capacities that must be achieved to\naccurately model human learning. We demonstrate that agents built along the\nDIPL theory are amenable to satisfying these capacities, and demonstrate, both\nempirically and theoretically, that DIPL enables the creation of agents that\nexhibit human-like learning performance."", ""This paper surveys the field of deep multiagent reinforcement learning. The\ncombination of deep neural networks with reinforcement learning has gained\nincreased traction in recent years and is slowly shifting the focus from\nsingle-agent to multiagent environments. Dealing with multiple agents is\ninherently more complex as (a) the future rewards depend on multiple players'\njoint actions and (b) the computational complexity increases. We present the\nmost common multiagent problem representations and their main challenges, and\nidentify five research areas that address one or more of these challenges:\ncentralised training and decentralised execution, opponent modelling,\ncommunication, efficient coordination, and reward shaping. We find that many\ncomputational studies rely on unrealistic assumptions or are not generalisable\nto other settings; they struggle to overcome the curse of dimensionality or\nnonstationarity. Approaches from psychology and sociology capture promising\nrelevant behaviours, such as communication and coordination, to help agents\nachieve better performance in multiagent settings. We suggest that, for\nmultiagent reinforcement learning to be successful, future research should\naddress these challenges with an interdisciplinary approach to open up new\npossibilities in multiagent reinforcement learning."", ""Representations are internal models of the environment that can provide\nguidance to a behaving agent, even in the absence of sensory information. It is\nnot clear how representations are developed and whether or not they are\nnecessary or even essential for intelligent behavior. We argue here that the\nability to represent relevant features of the environment is the expected\nconsequence of an adaptive process, give a formal definition of representation\nbased on information theory, and quantify it with a measure R. To measure how R\nchanges over time, we evolve two types of networks---an artificial neural\nnetwork and a network of hidden Markov gates---to solve a categorization task\nusing a genetic algorithm. We find that the capacity to represent increases\nduring evolutionary adaptation, and that agents form representations of their\nenvironment during their lifetime. This ability allows the agents to act on\nsensorial inputs in the context of their acquired representations and enables\ncomplex and context-dependent behavior. We examine which concepts (features of\nthe environment) our networks are representing, how the representations are\nlogically encoded in the networks, and how they form as an agent behaves to\nsolve a task. We conclude that R should be able to quantify the representations\nwithin any cognitive system, and should be predictive of an agent's long-term\nadaptive success."", 'The integration of deep learning to reinforcement learning (RL) has enabled\nRL to perform efficiently in high-dimensional environments. Deep RL methods\nhave been applied to solve many complex real-world problems in recent years.\nHowever, development of a deep RL-based system is challenging because of\nvarious issues such as the selection of a suitable deep RL algorithm, its\nnetwork configuration, training time, training methods, and so on. This paper\nproposes a comprehensive software framework that not only plays a vital role in\ndesigning a connect-the-dots deep RL architecture but also provides a guideline\nto develop a realistic RL application in a short time span. We have designed\nand developed a deep RL-based software framework that strictly ensures\nflexibility, robustness, and scalability. By inheriting the proposed\narchitecture, software managers can foresee any challenges when designing a\ndeep RL-based system. As a result, they can expedite the design process and\nactively control every stage of software development, which is especially\ncritical in agile development environments. To enforce generalization, the\nproposed architecture does not depend on a specific RL algorithm, a network\nconfiguration, the number of agents, or the type of agents. Using our\nframework, software developers can develop and integrate new RL algorithms or\nnew types of agents, and can flexibly change network configuration or the\nnumber of agents.', 'Numerous deep reinforcement learning agents have been proposed, and each of\nthem has its strengths and flaws. In this work, we present a Cooperative\nHeterogeneous Deep Reinforcement Learning (CHDRL) framework that can learn a\npolicy by integrating the advantages of heterogeneous agents. Specifically, we\npropose a cooperative learning framework that classifies heterogeneous agents\ninto two classes: global agents and local agents. Global agents are off-policy\nagents that can utilize experiences from the other agents. Local agents are\neither on-policy agents or population-based evolutionary algorithms (EAs)\nagents that can explore the local area effectively. We employ global agents,\nwhich are sample-efficient, to guide the learning of local agents so that local\nagents can benefit from sample-efficient agents and simultaneously maintain\ntheir advantages, e.g., stability. Global agents also benefit from effective\nlocal searches. Experimental studies on a range of continuous control tasks\nfrom the Mujoco benchmark show that CHDRL achieves better performance compared\nwith state-of-the-art baselines.', 'With the development of deep representation learning, the domain of\nreinforcement learning (RL) has become a powerful learning framework now\ncapable of learning complex policies in high dimensional environments. This\nreview summarises deep reinforcement learning (DRL) algorithms and provides a\ntaxonomy of automated driving tasks where (D)RL methods have been employed,\nwhile addressing key computational challenges in real world deployment of\nautonomous driving agents. It also delineates adjacent domains such as behavior\ncloning, imitation learning, inverse reinforcement learning that are related\nbut are not classical RL algorithms. The role of simulators in training agents,\nmethods to validate, test and robustify existing solutions in RL are discussed.', 'In recent years, there have been immense breakthroughs in Game AI research,\nparticularly with Reinforcement Learning (RL). Despite their success, the\nunderlying games are usually implemented with their own preset environments and\ngame mechanics, thus making it difficult for researchers to prototype different\ngame environments. However, testing the RL agents against a variety of game\nenvironments is critical for recent effort to study generalization in RL and\navoid the problem of overfitting that may otherwise occur. In this paper, we\npresent Griddly as a new platform for Game AI research that provides a unique\ncombination of highly configurable games, different observer types and an\nefficient C++ core engine. Additionally, we present a series of baseline\nexperiments to study the effect of different observation configurations and\ngeneralization ability of RL agents.', 'In this paper, we introduce Watch-And-Help (WAH), a challenge for testing\nsocial intelligence in agents. In WAH, an AI agent needs to help a human-like\nagent perform a complex household task efficiently. To succeed, the AI agent\nneeds to i) understand the underlying goal of the task by watching a single\ndemonstration of the human-like agent performing the same task (social\nperception), and ii) coordinate with the human-like agent to solve the task in\nan unseen environment as fast as possible (human-AI collaboration). For this\nchallenge, we build VirtualHome-Social, a multi-agent household environment,\nand provide a benchmark including both planning and learning based baselines.\nWe evaluate the performance of AI agents with the human-like agent as well as\nwith real humans using objective metrics and subjective user ratings.\nExperimental results demonstrate that the proposed challenge and virtual\nenvironment enable a systematic evaluation on the important aspects of machine\nsocial intelligence at scale.', 'In recent years deep reinforcement learning (RL) systems have attained\nsuperhuman performance in a number of challenging task domains. However, a\nmajor limitation of such applications is their demand for massive amounts of\ntraining data. A critical present objective is thus to develop deep RL methods\nthat can adapt rapidly to new tasks. In the present work we introduce a novel\napproach to this challenge, which we refer to as deep meta-reinforcement\nlearning. Previous work has shown that recurrent networks can support\nmeta-learning in a fully supervised context. We extend this approach to the RL\nsetting. What emerges is a system that is trained using one RL algorithm, but\nwhose recurrent dynamics implement a second, quite separate RL procedure. This\nsecond, learned RL algorithm can differ from the original one in arbitrary\nways. Importantly, because it is learned, it is configured to exploit structure\nin the training domain. We unpack these points in a series of seven\nproof-of-concept experiments, each of which examines a key aspect of deep\nmeta-RL. We consider prospects for extending and scaling up the approach, and\nalso point out some potentially important implications for neuroscience.', 'Deep reinforcement learning is the combination of reinforcement learning (RL)\nand deep learning. This field of research has been able to solve a wide range\nof complex decision-making tasks that were previously out of reach for a\nmachine. Thus, deep RL opens up many new applications in domains such as\nhealthcare, robotics, smart grids, finance, and many more. This manuscript\nprovides an introduction to deep reinforcement learning models, algorithms and\ntechniques. Particular focus is on the aspects related to generalization and\nhow deep RL can be used for practical applications. We assume the reader is\nfamiliar with basic machine learning concepts.', ""We present a new approach for improving the friendliness and warmth of a\nvirtual agent in an AR environment by generating appropriate movement\ncharacteristics. Our algorithm is based on a novel data-driven friendliness\nmodel that is computed using a user-study and psychological characteristics. We\nuse our model to control the movements corresponding to the gaits, gestures,\nand gazing of friendly virtual agents (FVAs) as they interact with the user's\navatar and other agents in the environment. We have integrated FVA agents with\nan AR environment using with a Microsoft HoloLens. Our algorithm can generate\nplausible movements at interactive rates to increase the social presence. We\nalso investigate the perception of a user in an AR setting and observe that an\nFVA has a statistically significant improvement in terms of the perceived\nfriendliness and social presence of a user compared to an agent without the\nfriendliness modeling. We observe an increment of 5.71% in the mean responses\nto a friendliness measure and an improvement of 4.03% in the mean responses to\na social presence measure."", 'We present an effective technique for training deep learning agents capable\nof negotiating on a set of clauses in a contract agreement using a simple\ncommunication protocol. We use Multi Agent Reinforcement Learning to train both\nagents simultaneously as they negotiate with each other in the training\nenvironment. We also model selfish and prosocial behavior to varying degrees in\nthese agents. Empirical evidence is provided showing consistency in agent\nbehaviors. We further train a meta agent with a mixture of behaviors by\nlearning an ensemble of different models using reinforcement learning. Finally,\nto ascertain the deployability of the negotiating agents, we conducted\nexperiments pitting the trained agents against human players. Results\ndemonstrate that the agents are able to hold their own against human players,\noften emerging as winners in the negotiation. Our experiments demonstrate that\nthe meta agent is able to reasonably emulate human behavior.', 'While the interpretability of machine learning models is often equated with\ntheir mere syntactic comprehensibility, we think that interpretability goes\nbeyond that, and that human interpretability should also be investigated from\nthe point of view of cognitive science. The goal of this paper is to discuss to\nwhat extent cognitive biases may affect human understanding of interpretable\nmachine learning models, in particular of logical rules discovered from data.\nTwenty cognitive biases are covered, as are possible debiasing techniques that\ncan be adopted by designers of machine learning algorithms and software. Our\nreview transfers results obtained in cognitive psychology to the domain of\nmachine learning, aiming to bridge the current gap between these two areas. It\nneeds to be followed by empirical studies specifically focused on the machine\nlearning domain.', 'Experts in Artificial Intelligence (AI) development predict that advances in\nthe development of intelligent systems and agents will reshape vital areas in\nour society. Nevertheless, if such an advance is not made prudently and\ncritically, reflexively, it can result in negative outcomes for humanity. For\nthis reason, several researchers in the area have developed a robust,\nbeneficial, and safe concept of AI for the preservation of humanity and the\nenvironment. Currently, several of the open problems in the field of AI\nresearch arise from the difficulty of avoiding unwanted behaviors of\nintelligent agents and systems, and at the same time specifying what we really\nwant such systems to do, especially when we look for the possibility of\nintelligent agents acting in several domains over the long term. It is of\nutmost importance that artificial intelligent agents have their values aligned\nwith human values, given the fact that we cannot expect an AI to develop human\nmoral values simply because of its intelligence, as discussed in the\nOrthogonality Thesis. Perhaps this difficulty comes from the way we are\naddressing the problem of expressing objectives, values, and ends, using\nrepresentational cognitive methods. A solution to this problem would be the\ndynamic approach proposed by Dreyfus, whose phenomenological philosophy shows\nthat the human experience of being-in-the-world in several aspects is not well\nrepresented by the symbolic or connectionist cognitive method, especially in\nregards to the question of learning values. A possible approach to this problem\nwould be to use theoretical models such as SED (situated embodied dynamics) to\naddress the values learning problem in AI.', 'In large-scale multi-agent systems, the large number of agents and complex\ngame relationship cause great difficulty for policy learning. Therefore,\nsimplifying the learning process is an important research issue. In many\nmulti-agent systems, the interactions between agents often happen locally,\nwhich means that agents neither need to coordinate with all other agents nor\nneed to coordinate with others all the time. Traditional methods attempt to use\npre-defined rules to capture the interaction relationship between agents.\nHowever, the methods cannot be directly used in a large-scale environment due\nto the difficulty of transforming the complex interactions between agents into\nrules. In this paper, we model the relationship between agents by a complete\ngraph and propose a novel game abstraction mechanism based on two-stage\nattention network (G2ANet), which can indicate whether there is an interaction\nbetween two agents and the importance of the interaction. We integrate this\ndetection mechanism into graph neural network-based multi-agent reinforcement\nlearning for conducting game abstraction and propose two novel learning\nalgorithms GA-Comm and GA-AC. We conduct experiments in Traffic Junction and\nPredator-Prey. The results indicate that the proposed methods can simplify the\nlearning process and meanwhile get better asymptotic performance compared with\nstate-of-the-art algorithms.', ""A common vision from science fiction is that robots will one day inhabit our\nphysical spaces, sense the world as we do, assist our physical labours, and\ncommunicate with us through natural language. Here we study how to design\nartificial agents that can interact naturally with humans using the\nsimplification of a virtual environment. We show that imitation learning of\nhuman-human interactions in a simulated world, in conjunction with\nself-supervised learning, is sufficient to produce a multimodal interactive\nagent, which we call MIA, that successfully interacts with non-adversarial\nhumans 75% of the time. We further identify architectural and algorithmic\ntechniques that improve performance, such as hierarchical action selection.\nAltogether, our results demonstrate that imitation of multi-modal, real-time\nhuman behaviour may provide a straightforward and surprisingly effective means\nof imbuing agents with a rich behavioural prior from which agents might then be\nfine-tuned for specific purposes, thus laying a foundation for training capable\nagents for interactive robots or digital assistants. A video of MIA's behaviour\nmay be found at https://youtu.be/ZFgRhviF7mY"", 'Artificial Intelligence is a central topic in the computer science\ncurriculum. From the year 2011 a project-based learning methodology based on\ncomputer games has been designed and implemented into the intelligence\nartificial course at the University of the Bio-Bio. The project aims to develop\nsoftware-controlled agents (bots) which are programmed by using heuristic\nalgorithms seen during the course. This methodology allows us to obtain good\nlearning results, however several challenges have been founded during its\nimplementation.\n  In this paper we show how linguistic descriptions of data can help to provide\nstudents and teachers with technical and personalized feedback about the\nlearned algorithms. Algorithm behavior profile and a new Turing test for\ncomputer games bots based on linguistic modelling of complex phenomena are also\nproposed in order to deal with such challenges.\n  In order to show and explore the possibilities of this new technology, a web\nplatform has been designed and implemented by one of authors and its\nincorporation in the process of assessment allows us to improve the teaching\nlearning process.', ""Experts in Artificial Intelligence (AI) development predict that advances in\nthe development of intelligent systems and agents will reshape vital areas in\nour society. Nevertheless, if such an advance isn't done with prudence, it can\nresult in negative outcomes for humanity. For this reason, several researchers\nin the area are trying to develop a robust, beneficial, and safe concept of\nartificial intelligence. Currently, several of the open problems in the field\nof AI research arise from the difficulty of avoiding unwanted behaviors of\nintelligent agents, and at the same time specifying what we want such systems\nto do. It is of utmost importance that artificial intelligent agents have their\nvalues aligned with human values, given the fact that we cannot expect an AI to\ndevelop our moral preferences simply because of its intelligence, as discussed\nin the Orthogonality Thesis. Perhaps this difficulty comes from the way we are\naddressing the problem of expressing objectives, values, and ends, using\nrepresentational cognitive methods. A solution to this problem would be the\ndynamic cognitive approach proposed by Dreyfus, whose phenomenological\nphilosophy defends that the human experience of being-in-the-world cannot be\nrepresented by the symbolic or connectionist cognitive methods. A possible\napproach to this problem would be to use theoretical models such as SED\n(situated embodied dynamics) to address the values learning problem in AI."", 'Humans are well-versed in reasoning about the behaviors of physical objects\nand choosing actions accordingly to accomplish tasks, while it remains a major\nchallenge for AI. To facilitate research addressing this problem, we propose a\nnew testbed that requires an agent to reason about physical scenarios and take\nan action appropriately. Inspired by the physical knowledge acquired in infancy\nand the capabilities required for robots to operate in real-world environments,\nwe identify 15 essential physical scenarios. We create a wide variety of\ndistinct task templates, and we ensure all the task templates within the same\nscenario can be solved by using one specific strategic physical rule. By having\nsuch a design, we evaluate two distinct levels of generalization, namely the\nlocal generalization and the broad generalization. We conduct an extensive\nevaluation with human players, learning agents with varying input types and\narchitectures, and heuristic agents with different strategies. Inspired by how\nhuman IQ is calculated, we define the physical reasoning quotient (Phy-Q score)\nthat reflects the physical reasoning intelligence of an agent using the\nphysical scenarios we considered. Our evaluation shows that 1) all agents are\nfar below human performance, and 2) learning agents, even with good local\ngeneralization ability, struggle to learn the underlying physical reasoning\nrules and fail to generalize broadly. We encourage the development of\nintelligent agents that can reach the human level Phy-Q score. Website:\nhttps://github.com/phy-q/benchmark', ""We consider the design of distributed algorithms that govern the manner in\nwhich agents contribute to a social sensing platform. Specifically, we are\ninterested in situations where fairness among the agents contributing to the\nplatform is needed. A notable example are platforms operated by public bodies,\nwhere fairness is a legal requirement. The design of such distributed systems\nis challenging due to the fact that we wish to simultaneously realise an\nefficient social sensing platform, but also deliver a predefined quality of\nservice to the agents (for example, a fair opportunity to contribute to the\nplatform). In this paper, we introduce iterated function systems (IFS) as a\ntool for the design and analysis of systems of this kind. We show how the IFS\nframework can be used to realise systems that deliver a predictable quality of\nservice to agents, can be used to underpin contracts governing the interaction\nof agents with the social sensing platform, and which are efficient.\n  To illustrate our design via a use case, we consider a large, high-density\nnetwork of participating parked vehicles. When awoken by an administrative\ncentre, this network proceeds to search for moving missing entities of interest\nusing RFID-based techniques. We regulate which vehicles are actively searching\nfor the moving entity of interest at any point in time. In doing so, we seek to\nequalise vehicular energy consumption across the network. This is illustrated\nthrough simulations of a search for a missing Alzheimer's patient in Melbourne,\nAustralia. Experimental results are presented to illustrate the efficacy of our\nsystem and the predictability of access of agents to the platform independent\nof initial conditions."", 'A reinforcement learning environment with adversary agents is proposed in\nthis work for pursuit-evasion game in the presence of fog of war, which is of\nboth scientific significance and practical importance in aerospace\napplications. One of the most popular learning environments, StarCraft, is\nadopted here and the associated mini-games are analyzed to identify the current\nlimitation for training adversary agents. The key contribution includes the\nanalysis of the potential performance of an agent by incorporating control and\ndifferential game theory into the specific reinforcement learning environment,\nand the development of an adversary agents challenge (SAAC) environment by\nextending the current StarCraft mini-games. The subsequent study showcases the\nuse of this learning environment and the effectiveness of an adversary agent\nfor evasion units. Overall, the proposed SAAC environment should benefit\npursuit-evasion studies with rapidly-emerging reinforcement learning\ntechnologies. Last but not least, the corresponding tutorial code can be found\nat GitHub.', 'Along with the development of modern computing technology and social\nsciences, both theoretical research and practical applications of social\ncomputing have been continuously extended. In particular with the boom of\nartificial intelligence (AI), social computing is significantly influenced by\nAI. However, the conventional technologies of AI have drawbacks in dealing with\nmore complicated and dynamic problems. Such deficiency can be rectified by\nhybrid human-artificial intelligence (H-AI) which integrates both human\nintelligence and AI into one unity, forming a new enhanced intelligence. H-AI\nin dealing with social problems shows the advantages that AI can not surpass.\nThis paper firstly introduces the concept of H-AI. AI is the intelligence in\nthe transition stage of H-AI, so the latest research progresses of AI in social\ncomputing are reviewed. Secondly, it summarizes typical challenges faced by AI\nin social computing, and makes it possible to introduce H-AI to solve these\nchallenges. Finally, the paper proposes a holistic framework of social\ncomputing combining with H-AI, which consists of four layers: object layer,\nbase layer, analysis layer, and application layer. It represents H-AI has\nsignificant advantages over AI in solving social problems.', 'Multi-Agent reinforcement learning has received lot of attention in recent\nyears and have applications in many different areas. Existing methods involving\nCentralized Training and Decentralized execution, attempts to train the agents\ntowards learning a pattern of coordinated actions to arrive at optimal joint\npolicy. However if some agents are stochastic to varying degrees of\nstochasticity, the above methods often fail to converge and provides poor\ncoordination among agents. In this paper we show how this stochasticity of\nagents, which could be a result of malfunction or aging of robots, can add to\nthe uncertainty in coordination and there contribute to unsatisfactory global\ncoordination. In this case, the deterministic agents have to understand the\nbehavior and limitations of the stochastic agents while arriving at optimal\njoint policy. Our solution, DSDF which tunes the discounted factor for the\nagents according to uncertainty and use the values to update the utility\nnetworks of individual agents. DSDF also helps in imparting an extent of\nreliability in coordination thereby granting stochastic agents tasks which are\nimmediate and of shorter trajectory with deterministic ones taking the tasks\nwhich involve longer planning. Such an method enables joint co-ordinations of\nagents some of which may be partially performing and thereby can reduce or\ndelay the investment of agent/robot replacement in many circumstances. Results\non benchmark environment for different scenarios shows the efficacy of the\nproposed approach when compared with existing approaches.', ""It is safe to assume that, for the foreseeable future, machine learning,\nespecially deep learning will remain both data- and computation-hungry. In this\npaper, we ask: Can we build a global exchange where everyone can contribute\ncomputation and data to train the next generation of machine learning\napplications?\n  We present an early, but running prototype of DataBright, a system that turns\nthe creation of training examples and the sharing of computation into an\ninvestment mechanism. Unlike most crowdsourcing platforms, where the\ncontributor gets paid when they submit their data, DataBright pays dividends\nwhenever a contributor's data or hardware is used by someone to train a machine\nlearning model. The contributor becomes a shareholder in the dataset they\ncreated. To enable the measurement of usage, a computation platform that\ncontributors can trust is also necessary. DataBright thus merges both a data\nmarket and a trusted computation market.\n  We illustrate that trusted computation can enable the creation of an AI\nmarket, where each data point has an exact value that should be paid to its\ncreator. DataBright allows data creators to retain ownership of their\ncontribution and attaches to it a measurable value. The value of the data is\ngiven by its utility in subsequent distributed computation done on the\nDataBright computation market. The computation market allocates tasks and\nsubsequent payments to pooled hardware. This leads to the creation of a\ndecentralized AI cloud. Our experiments show that trusted hardware such as\nIntel SGX can be added to the usual ML pipeline with no additional costs. We\nuse this setting to orchestrate distributed computation that enables the\ncreation of a computation market. DataBright is available for download at\nhttps://github.com/ds3lab/databright."", 'Designing an effective communication mechanism among agents in reinforcement\nlearning has been a challenging task, especially for real-world applications.\nThe number of agents can grow or an environment sometimes needs to interact\nwith a changing number of agents in real-world scenarios. To this end, a\nmulti-agent framework needs to handle various scenarios of agents, in terms of\nboth scales and dynamics, for being practical to real-world applications. We\nformulate the multi-agent environment with a different number of agents as a\nmulti-tasking problem and propose a meta reinforcement learning (meta-RL)\nframework to tackle this problem. The proposed framework employs a meta-learned\nCommunication Pattern Recognition (CPR) module to identify communication\nbehavior and extract information that facilitates the training process.\nExperimental results are poised to demonstrate that the proposed framework (a)\ngeneralizes to an unseen larger number of agents and (b) allows the number of\nagents to change between episodes. The ablation study is also provided to\nreason the proposed CPR design and show such design is effective.', 'We introduce Imagination-Augmented Agents (I2As), a novel architecture for\ndeep reinforcement learning combining model-free and model-based aspects. In\ncontrast to most existing model-based reinforcement learning and planning\nmethods, which prescribe how a model should be used to arrive at a policy, I2As\nlearn to interpret predictions from a learned environment model to construct\nimplicit plans in arbitrary ways, by using the predictions as additional\ncontext in deep policy networks. I2As show improved data efficiency,\nperformance, and robustness to model misspecification compared to several\nbaselines.', 'Motivated by recent advancements in Deep Reinforcement Learning (RL), we have\ndeveloped an RL agent to manage the operation of storage devices in a household\nand is designed to maximize demand-side cost savings. The proposed technique is\ndata-driven, and the RL agent learns from scratch how to efficiently use the\nenergy storage device given variable tariff structures. In most of the studies,\nthe RL agent is considered as a black box, and how the agent has learned is\noften ignored. We explain the learning progression of the RL agent, and the\nstrategies it follows based on the capacity of the storage device.', 'Counterfactual thinking describes a psychological phenomenon that people\nre-infer the possible results with different solutions about things that have\nalready happened. It helps people to gain more experience from mistakes and\nthus to perform better in similar future tasks. This paper investigates the\ncounterfactual thinking for agents to find optimal decision-making strategies\nin multi-agent reinforcement learning environments. In particular, we propose a\nmulti-agent deep reinforcement learning model with a structure which mimics the\nhuman-psychological counterfactual thinking process to improve the competitive\nabilities for agents. To this end, our model generates several possible actions\n(intent actions) with a parallel policy structure and estimates the rewards and\nregrets for these intent actions based on its current understanding of the\nenvironment. Our model incorporates a scenario-based framework to link the\nestimated regrets with its inner policies. During the iterations, our model\nupdates the parallel policies and the corresponding scenario-based regrets for\nagents simultaneously. To verify the effectiveness of our proposed model, we\nconduct extensive experiments on two different environments with real-world\napplications. Experimental results show that counterfactual thinking can\nactually benefit the agents to obtain more accumulative rewards from the\nenvironments with fair information by comparing to their opponents while\nkeeping high performing efficiency.', 'AI technology has a long history which is actively and constantly changing\nand growing. It focuses on intelligent agents, which contain devices that\nperceive the environment and based on which takes actions in order to maximize\ngoal success chances. In this paper, we will explain the modern AI basics and\nvarious representative applications of AI. In the context of the modern\ndigitalized world, AI is the property of machines, computer programs, and\nsystems to perform the intellectual and creative functions of a person,\nindependently find ways to solve problems, be able to draw conclusions and make\ndecisions. Most artificial intelligence systems have the ability to learn,\nwhich allows people to improve their performance over time. The recent research\non AI tools, including machine learning, deep learning and predictive analysis\nintended toward increasing the planning, learning, reasoning, thinking and\naction taking ability. Based on which, the proposed research intends towards\nexploring on how the human intelligence differs from the artificial\nintelligence. Moreover, we critically analyze what AI of today is capable of\ndoing, why it still cannot reach human intelligence and what are the open\nchallenges existing in front of AI to reach and outperform human level of\nintelligence. Furthermore, it will explore the future predictions for\nartificial intelligence and based on which potential solution will be\nrecommended to solve it within next decades.', 'While AI has benefited humans, it may also harm humans if not appropriately\ndeveloped. The focus of HCI work is transiting from conventional human\ninteraction with non-AI computing systems to interaction with AI systems. We\nconducted a high-level literature review and a holistic analysis of current\nwork in developing AI systems from an HCI perspective. Our review and analysis\nhighlight the new changes introduced by AI technology and the new challenges\nthat HCI professionals face when applying the human-centered AI (HCAI) approach\nin the development of AI systems. We also identified seven main issues in human\ninteraction with AI systems, which HCI professionals did not encounter when\ndeveloping non-AI computing systems. To further enable the implementation of\nthe HCAI approach, we identified new HCI opportunities tied to specific\nHCAI-driven design goals to guide HCI professionals in addressing these new\nissues. Finally, our assessment of current HCI methods shows the limitations of\nthese methods in support of developing AI systems. We propose alternative\nmethods that can help overcome these limitations and effectively help HCI\nprofessionals apply the HCAI approach to the development of AI systems. We also\noffer strategic recommendations for HCI professionals to effectively influence\nthe development of AI systems with the HCAI approach, eventually developing\nHCAI systems.', ""AI researchers employ not only the scientific method, but also methodology\nfrom mathematics and engineering. However, the use of the scientific method -\nspecifically hypothesis testing - in AI is typically conducted in service of\nengineering objectives. Growing interest in topics such as fairness and\nalgorithmic bias show that engineering-focused questions only comprise a subset\nof the important questions about AI systems. This results in the AI Knowledge\nGap: the number of unique AI systems grows faster than the number of studies\nthat characterize these systems' behavior. To close this gap, we argue that the\nstudy of AI could benefit from the greater inclusion of researchers who are\nwell positioned to formulate and test hypotheses about the behavior of AI\nsystems. We examine the barriers preventing social and behavioral scientists\nfrom conducting such studies. Our diagnosis suggests that accelerating the\nscientific study of AI systems requires new incentives for academia and\nindustry, mediated by new tools and institutions. To address these needs, we\npropose a two-sided marketplace called TuringBox. On one side, AI contributors\nupload existing and novel algorithms to be studied scientifically by others. On\nthe other side, AI examiners develop and post machine intelligence tasks\ndesigned to evaluate and characterize algorithmic behavior. We discuss this\nmarket's potential to democratize the scientific study of AI behavior, and thus\nnarrow the AI Knowledge Gap."", ""In this paper, we present our research on programming human-level artificial\nintelligence (HLAI), including 1) a definition of HLAI, 2) an environment to\ndevelop and test HLAI, and 3) a cognitive architecture for HLAI. The term AI is\nused in a broad meaning, and HLAI is not clearly defined. I claim that the\nessence of Human-Level Intelligence to be the capability to learn from others'\nexperiences via language. The key is that the event described by language has\nthe same effect as if the agent experiences it firsthand for the update of the\nbehavior policy. To develop and test models with such a capability, we are\ndeveloping a simulated environment called SEDRo. There is a 3D Home, and a\nmother character takes care of the baby (the learning agent) and teaches\nlanguages. The environment provides comparable experiences to that of a human\nbaby from birth to one year. Finally, I propose a cognitive architecture of\nHLAI called Modulated Heterarchical Prediction Memory (mHPM). In mHPM, there\nare three components: a universal module that learns to predict the next vector\ngiven the sequence of vector signals, a heterarchical network of those modules,\nand a reward-based modulation of learning. mHPM models the workings of the\nneocortex but the innate auxiliary units such hippocampus, reward system,\ninstincts, and amygdala play critical roles, too."", ""Campbell-Goodhart's law relates to the causal inference error whereby\ndecision-making agents aim to influence variables which are correlated to their\ngoal objective but do not reliably cause it. This is a well known error in\nEconomics and Political Science but not widely labelled in Artificial\nIntelligence research. Through a simple example, we show how off-the-shelf deep\nReinforcement Learning (RL) algorithms are not necessarily immune to this\ncognitive error. The off-policy learning method is tricked, whilst the\non-policy method is not. The practical implication is that naive application of\nRL to complex real life problems can result in the same types of policy errors\nthat humans make. Great care should be taken around understanding the causal\nmodel that underpins a solution derived from Reinforcement Learning."", ""How to attribute responsibility for autonomous artificial intelligence (AI)\nsystems' actions has been widely debated across the humanities and social\nscience disciplines. This work presents two experiments ($N$=200 each) that\nmeasure people's perceptions of eight different notions of moral responsibility\nconcerning AI and human agents in the context of bail decision-making. Using\nreal-life adapted vignettes, our experiments show that AI agents are held\ncausally responsible and blamed similarly to human agents for an identical\ntask. However, there was a meaningful difference in how people perceived these\nagents' moral responsibility; human agents were ascribed to a higher degree of\npresent-looking and forward-looking notions of responsibility than AI agents.\nWe also found that people expect both AI and human decision-makers and advisors\nto justify their decisions regardless of their nature. We discuss policy and\nHCI implications of these findings, such as the need for explainable AI in\nhigh-stakes scenarios."", ""We consider finite and infinite horizon dynamic programming problems, where\nthe control at each stage consists of several distinct decisions, each one made\nby one of several agents. We introduce an approach, whereby at every stage,\neach agent's decision is made by executing a local rollout algorithm that uses\na base policy, together with some coordinating information from the other\nagents. The amount of local computation required at every stage by each agent\nis independent of the number of agents, while the amount of total computation\n(over all agents) grows linearly with the number of agents. By contrast, with\nthe standard rollout algorithm, the amount of total computation grows\nexponentially with the number of agents. Despite the drastic reduction in\nrequired computation, we show that our algorithm has the fundamental cost\nimprovement property of rollout: an improved performance relative to the base\npolicy. We also discuss possibilities to improve further the method's\ncomputational efficiency through limited agent coordination and parallelization\nof the agents' computations. Finally, we explore related approximate policy\niteration algorithms for infinite horizon problems, and we prove that the cost\nimprovement property steers the algorithm towards convergence to an\nagent-by-agent optimal policy."", 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', 'There is a growing focus on how to design safe artificial intelligent (AI)\nagents. As systems become more complex, poorly specified goals or control\nmechanisms may cause AI agents to engage in unwanted and harmful outcomes. Thus\nit is necessary to design AI agents that follow initial programming intentions\nas the program grows in complexity. How to specify these initial intentions has\nalso been an obstacle to designing safe AI agents. Finally, there is a need for\nthe AI agent to have redundant safety mechanisms to ensure that any programming\nerrors do not cascade into major problems. Humans are autonomous intelligent\nagents that have avoided these problems and the present manuscript argues that\nby understanding human self-regulation and goal setting, we may be better able\nto design safe AI agents. Some general principles of human self-regulation are\noutlined and specific guidance for AI design is given.', 'Convolutional Neural Networks achieve state-of-the-art accuracy in object\ndetection tasks. However, they have large computational and energy requirements\nthat challenge their deployment on resource-constrained edge devices. Object\ndetection takes an image as an input, and identifies the existing object\nclasses as well as their locations in the image. In this paper, we leverage the\nprior knowledge about the probabilities that different object categories can\noccur jointly to increase the efficiency of object detection models. In\nparticular, our technique clusters the object categories based on their spatial\nco-occurrence probability. We use those clusters to design an adaptive network.\nDuring runtime, a branch controller decides which part(s) of the network to\nexecute based on the spatial context of the input frame. Our experiments using\nCOCO dataset show that our adaptive object detection model achieves up to 45%\nreduction in the energy consumption, and up to 27% reduction in the latency,\nwith a small loss in the average precision (AP) of object detection.', 'Learning by observation can be of key importance whenever agents sharing\nsimilar features want to learn from each other. This paper presents an agent\narchitecture that enables software agents to learn by direct observation of the\nactions executed by expert agents while they are performing a task. This is\npossible because the proposed architecture displays information that is\nessential for observation, making it possible for software agents to observe\neach other. The agent architecture supports a learning process that covers all\naspects of learning by observation, such as discovering and observing experts,\nlearning from the observed data, applying the acquired knowledge and evaluating\nthe agents progress. The evaluation provides control over the decision to\nobtain new knowledge or apply the acquired knowledge to new problems. We\ncombine two methods for learning from the observed information. The first one,\nthe recall method, uses the sequence on which the actions were observed to\nsolve new problems. The second one, the classification method, categorizes the\ninformation in the observed data and determines to which set of categories the\nnew problems belong. Results show that agents are able to learn in conditions\nwhere common supervised learning algorithms fail, such as when agents do not\nknow the results of their actions a priori or when not all the effects of the\nactions are visible. The results also show that our approach provides better\nresults than other learning methods since it requires shorter learning periods.', 'Individuality is essential in human society, which induces the division of\nlabor and thus improves the efficiency and productivity. Similarly, it should\nalso be the key to multi-agent cooperation. Inspired by that individuality is\nof being an individual separate from others, we propose a simple yet efficient\nmethod for the emergence of individuality (EOI) in multi-agent reinforcement\nlearning (MARL). EOI learns a probabilistic classifier that predicts a\nprobability distribution over agents given their observation and gives each\nagent an intrinsic reward of being correctly predicted by the classifier. The\nintrinsic reward encourages the agents to visit their own familiar\nobservations, and learning the classifier by such observations makes the\nintrinsic reward signals stronger and the agents more identifiable. To further\nenhance the intrinsic reward and promote the emergence of individuality, two\nregularizers are proposed to increase the discriminability of the classifier.\nWe implement EOI on top of popular MARL algorithms. Empirically, we show that\nEOI significantly outperforms existing methods in a variety of multi-agent\ncooperative scenarios.', 'Inattentional blindness is the psychological phenomenon that causes one to\nmiss things in plain sight. It is a consequence of the selective attention in\nperception that lets us remain focused on important parts of our world without\ndistraction from irrelevant details. Motivated by selective attention, we study\nthe properties of artificial agents that perceive the world through the lens of\na self-attention bottleneck. By constraining access to only a small fraction of\nthe visual input, we show that their policies are directly interpretable in\npixel space. We find neuroevolution ideal for training self-attention\narchitectures for vision-based reinforcement learning (RL) tasks, allowing us\nto incorporate modules that can include discrete, non-differentiable operations\nwhich are useful for our agent. We argue that self-attention has similar\nproperties as indirect encoding, in the sense that large implicit weight\nmatrices are generated from a small number of key-query parameters, thus\nenabling our agent to solve challenging vision based tasks with at least 1000x\nfewer parameters than existing methods. Since our agent attends to only task\ncritical visual hints, they are able to generalize to environments where task\nirrelevant elements are modified while conventional methods fail. Videos of our\nresults and source code available at https://attentionagent.github.io/', '3D objects (artefacts) are made to fulfill functions. Designing an object\noften starts with defining a list of functionalities that it should provide,\nalso known as functional requirements. Today, the design of 3D object models is\nstill a slow and largely artisanal activity, with few Computer-Aided Design\n(CAD) tools existing to aid the exploration of the design solution space. To\naccelerate the design process, we introduce an algorithm for generating object\nshapes with desired functionalities. Following the concept of form follows\nfunction, we assume that existing object shapes were rationally chosen to\nprovide desired functionalities. First, we use an artificial neural network to\nlearn a function-to-form mapping by analysing a dataset of objects labeled with\ntheir functionalities. Then, we combine forms providing one or more desired\nfunctions, generating an object shape that is expected to provide all of them.\nFinally, we verify in simulation whether the generated object possesses the\ndesired functionalities, by defining and executing functionality tests on it.', 'This book-length article combines several peer reviewed papers and new\nmaterial to analyze the issues of ethical artificial intelligence (AI). The\nbehavior of future AI systems can be described by mathematical equations, which\nare adapted to analyze possible unintended AI behaviors and ways that AI\ndesigns can avoid them. This article makes the case for utility-maximizing\nagents and for avoiding infinite sets in agent definitions. It shows how to\navoid agent self-delusion using model-based utility functions and how to avoid\nagents that corrupt their reward generators (sometimes called ""perverse\ninstantiation"") using utility functions that evaluate outcomes at one point in\ntime from the perspective of humans at a different point in time. It argues\nthat agents can avoid unintended instrumental actions (sometimes called ""basic\nAI drives"" or ""instrumental goals"") by accurately learning human values. This\narticle defines a self-modeling agent framework and shows how it can avoid\nproblems of resource limits, being predicted by other agents, and inconsistency\nbetween the agent\'s utility function and its definition (one version of this\nproblem is sometimes called ""motivated value selection""). This article also\ndiscusses how future AI will differ from current AI, the politics of AI, and\nthe ultimate use of AI to help understand the nature of the universe and our\nplace in it.', 'Consider a platform that wants to learn a personalized policy for each user,\nbut the platform faces the risk of a user abandoning the platform if she is\ndissatisfied with the actions of the platform. For example, a platform is\ninterested in personalizing the number of newsletters it sends, but faces the\nrisk that the user unsubscribes forever. We propose a general thresholded\nlearning model for scenarios like this, and discuss the structure of optimal\npolicies. We describe salient features of optimal personalization algorithms\nand how feedback the platform receives impacts the results. Furthermore, we\ninvestigate how the platform can efficiently learn the heterogeneity across\nusers by interacting with a population and provide performance guarantees.', ""Natural language understanding tasks such as open-domain question answering\noften require retrieving and assimilating factual information from multiple\nsources. We propose to address this problem by integrating a semi-parametric\nrepresentation of a large text corpus into a Transformer model as a source of\nfactual knowledge. Specifically, our method represents knowledge with `mention\nmemory', a table of dense vector representations of every entity mention in a\ncorpus. The proposed model - TOME - is a Transformer that accesses the\ninformation through internal memory layers in which each entity mention in the\ninput passage attends to the mention memory. This approach enables synthesis of\nand reasoning over many disparate sources of information within a single\nTransformer model. In experiments using a memory of 150 million Wikipedia\nmentions, TOME achieves strong performance on several open-domain\nknowledge-intensive tasks, including the claim verification benchmarks HoVer\nand FEVER and several entity-based QA benchmarks. We also show that the model\nlearns to attend to informative mentions without any direct supervision.\nFinally we demonstrate that the model can generalize to new unseen entities by\nupdating the memory without retraining."", ""Tracking humans that are interacting with the other subjects or environment\nremains unsolved in visual tracking, because the visibility of the human of\ninterests in videos is unknown and might vary over time. In particular, it is\nstill difficult for state-of-the-art human trackers to recover complete human\ntrajectories in crowded scenes with frequent human interactions. In this work,\nwe consider the visibility status of a subject as a fluent variable, whose\nchange is mostly attributed to the subject's interaction with the surrounding,\ne.g., crossing behind another object, entering a building, or getting into a\nvehicle, etc. We introduce a Causal And-Or Graph (C-AOG) to represent the\ncausal-effect relations between an object's visibility fluent and its\nactivities, and develop a probabilistic graph model to jointly reason the\nvisibility fluent change (e.g., from visible to invisible) and track humans in\nvideos. We formulate this joint task as an iterative search of a feasible\ncausal graph structure that enables fast search algorithm, e.g., dynamic\nprogramming method. We apply the proposed method on challenging video sequences\nto evaluate its capabilities of estimating visibility fluent changes of\nsubjects and tracking subjects of interests over time. Results with comparisons\ndemonstrate that our method outperforms the alternative trackers and can\nrecover complete trajectories of humans in complicated scenarios with frequent\nhuman interactions."", 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.', 'This special issue is dedicated to get a better picture of the relationships\nbetween computational linguistics and cognitive science. It specifically raises\ntwo questions: ""what is the potential contribution of computational language\nmodeling to cognitive science?"" and conversely: ""what is the influence of\ncognitive science in contemporary computational linguistics?""', 'Occlusion is one of the most significant challenges encountered by object\ndetectors and trackers. While both object detection and tracking has received a\nlot of attention in the past, most existing methods in this domain do not\ntarget detecting or tracking objects when they are occluded. However, being\nable to detect or track an object of interest through occlusion has been a long\nstanding challenge for different autonomous tasks. Traditional methods that\nemploy visual object trackers with explicit occlusion modeling experience drift\nand make several fundamental assumptions about the data. We propose to address\nthis with a `tracking-by-detection` approach that builds upon the success of\nregion based video object detectors. Our video level object detector uses a\nnovel recurrent computational unit at its core that enables long term\npropagation of object features even under occlusion. Finally, we compare our\napproach with existing state-of-the-art video object detectors and show that\nour approach achieves superior results on a dataset of furniture assembly\nvideos collected from the internet, where small objects like screws, nuts, and\nbolts often get occluded from the camera viewpoint.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I have a basic understanding of a research direction in Question & Answering (Q&A) systems. These systems employ an integrated knowledge base, structured similarly to a database, with relational dependencies among various concepts and pre-existing answers. They interpret user queries and retrieve answers by leveraging the dependencies between concept nodes within the knowledge base. I am interested in employing a deep neural network to model this type of knowledge base, thereby eliminating the need for hard-coded, rule-based relationships between concepts. I am curious to see if such a neural network model could overcome the limitations inherent in rule-based knowledge base models. This is because all relationships and dependencies would have implicit vectorized representations, potentially enhancing computational efficiency. This approach might also augment the model's expressiveness and capacity to answer more complex questions. My goal is to demonstrate, using Q&A datasets, that my proposed model outperforms traditional Q&A models in terms of increased accuracy and faster runtime.","['Most existing approaches for Knowledge Base Question Answering (KBQA) focus\non a specific underlying knowledge base either because of inherent assumptions\nin the approach, or because evaluating it on a different knowledge base\nrequires non-trivial changes. However, many popular knowledge bases share\nsimilarities in their underlying schemas that can be leveraged to facilitate\ngeneralization across knowledge bases. To achieve this generalization, we\nintroduce a KBQA framework based on a 2-stage architecture that explicitly\nseparates semantic parsing from the knowledge base interaction, facilitating\ntransfer learning across datasets and knowledge graphs. We show that\npretraining on datasets with a different underlying knowledge base can\nnevertheless provide significant performance gains and reduce sample\ncomplexity. Our approach achieves comparable or state-of-the-art performance\nfor LC-QuAD (DBpedia), WebQSP (Freebase), SimpleQuestions (Wikidata) and MetaQA\n(Wikimovies-KG).', 'Frequently asked question (FAQ) retrieval, with the purpose of providing\ninformation on frequent questions or concerns, has far-reaching applications in\nmany areas, where a collection of question-answer (Q-A) pairs compiled a priori\ncan be employed to retrieve an appropriate answer in response to a user\\u2019s\nquery that is likely to reoccur frequently. To this end, predominant approaches\nto FAQ retrieval typically rank question-answer pairs by considering either the\nsimilarity between the query and a question (q-Q), the relevance between the\nquery and the associated answer of a question (q-A), or combining the clues\ngathered from the q-Q similarity measure and the q-A relevance measure. In this\npaper, we extend this line of research by combining the clues gathered from the\nq-Q similarity measure and the q-A relevance measure and meanwhile injecting\nextra word interaction information, distilled from a generic (open domain)\nknowledge base, into a contextual language model for inferring the q-A\nrelevance. Furthermore, we also explore to capitalize on domain-specific\ntopically-relevant relations between words in an unsupervised manner, acting as\na surrogate to the supervised domain-specific knowledge base information. As\nsuch, it enables the model to equip sentence representations with the knowledge\nabout domain-specific and topically-relevant relations among words, thereby\nproviding a better q-A relevance measure. We evaluate variants of our approach\non a publicly-available Chinese FAQ dataset, and further apply and\ncontextualize it to a large-scale question-matching task, which aims to search\nquestions from a QA dataset that have a similar intent as an input query.\nExtensive experimental results on these two datasets confirm the promising\nperformance of the proposed approach in relation to some state-of-the-art ones.']","['Open Domain Question Answering requires systems to retrieve external\nknowledge and perform multi-hop reasoning by composing knowledge spread over\nmultiple sentences. In the recently introduced open domain question answering\nchallenge datasets, QASC and OpenBookQA, we need to perform retrieval of facts\nand compose facts to correctly answer questions. In our work, we learn a\nsemantic knowledge ranking model to re-rank knowledge retrieved through Lucene\nbased information retrieval systems. We further propose a ""knowledge fusion\nmodel"" which leverages knowledge in BERT-based language models with externally\nretrieved knowledge and improves the knowledge understanding of the BERT-based\nlanguage models. On both OpenBookQA and QASC datasets, the knowledge fusion\nmodel with semantically re-ranked knowledge outperforms previous attempts.', 'Deep learning models suffer from opaqueness. For Convolutional Neural\nNetworks (CNNs), current research strategies for explaining models focus on the\ntarget classes within the associated training dataset. As a result, the\nunderstanding of hidden feature map activations is limited by the\ndiscriminative knowledge gleaned during training. The aim of our work is to\nexplain and expand CNNs models via the mirroring or alignment of CNN to an\nexternal knowledge base. This will allow us to give a semantic context or label\nfor each visual feature. We can match CNN feature activations to nodes in our\nexternal knowledge base. This supports knowledge-based interpretation of the\nfeatures associated with model decisions. To demonstrate our approach, we build\ntwo separate graphs. We use an entity alignment method to align the feature\nnodes in a CNN with the nodes in a ConceptNet based knowledge graph. We then\nmeasure the proximity of CNN graph nodes to semantically meaningful knowledge\nbase nodes. Our results show that in the aligned embedding space, nodes from\nthe knowledge graph are close to the CNN feature nodes that have similar\nmeanings, indicating that nodes from an external knowledge base can act as\nexplanatory semantic references for features in the model. We analyse a variety\nof graph building methods in order to improve the results from our embedding\nspace. We further demonstrate that by using hierarchical relationships from our\nexternal knowledge base, we can locate new unseen classes outside the CNN\ntraining set in our embeddings space, based on visual feature activations. This\nsuggests that we can adapt our approach to identify unseen classes based on CNN\nfeature activations. Our demonstrated approach of aligning a CNN with an\nexternal knowledge base paves the way to reason about and beyond the trained\nmodel, with future adaptations to explainable models and zero-shot learning.', 'Knowledge Base, represents facts about the world, often in some form of\nsubsumption ontology, rather than implicitly, embedded in procedural code, the\nway a conventional computer program does. While there is a rapid growth in\nknowledge bases, it poses a challenge of retrieving information from them.\nKnowledge Base Question Answering is one of the promising approaches for\nextracting substantial knowledge from Knowledge Bases. Unlike web search,\nQuestion Answering over a knowledge base gives accurate and concise results,\nprovided that natural language questions can be understood and mapped precisely\nto an answer in the knowledge base. However, some of the existing\nembedding-based methods for knowledge base question answering systems ignore\nthe subtle correlation between the question and the Knowledge Base (e.g.,\nentity types, relation paths, and context) and suffer from the Out Of\nVocabulary problem. In this paper, we focused on using a pre-trained language\nmodel for the Knowledge Base Question Answering task. Firstly, we used Bert\nbase uncased for the initial experiments. We further fine-tuned these\nembeddings with a two-way attention mechanism from the knowledge base to the\nasked question and from the asked question to the knowledge base answer\naspects. Our method is based on a simple Convolutional Neural Network\narchitecture with a Multi-Head Attention mechanism to represent the asked\nquestion dynamically in multiple aspects. Our experimental results show the\neffectiveness and the superiority of the Bert pre-trained language model\nembeddings for question answering systems on knowledge bases over other\nwell-known embedding methods.', 'Directly reading documents and being able to answer questions from them is an\nunsolved challenge. To avoid its inherent difficulty, question answering (QA)\nhas been directed towards using Knowledge Bases (KBs) instead, which has proven\neffective. Unfortunately KBs often suffer from being too restrictive, as the\nschema cannot support certain types of answers, and too sparse, e.g. Wikipedia\ncontains much more information than Freebase. In this work we introduce a new\nmethod, Key-Value Memory Networks, that makes reading documents more viable by\nutilizing different encodings in the addressing and output stages of the memory\nread operation. To compare using KBs, information extraction or Wikipedia\ndocuments directly in a single framework we construct an analysis tool,\nWikiMovies, a QA dataset that contains raw text alongside a preprocessed KB, in\nthe domain of movies. Our method reduces the gap between all three settings. It\nalso achieves state-of-the-art results on the existing WikiQA benchmark.', 'Information sources such as relational databases, spreadsheets, XML, JSON,\nand Web APIs contain a tremendous amount of structured data that can be\nleveraged to build and augment knowledge graphs. However, they rarely provide a\nsemantic model to describe their contents. Semantic models of data sources\nrepresent the implicit meaning of the data by specifying the concepts and the\nrelationships within the data. Such models are the key ingredients to\nautomatically publish the data into knowledge graphs. Manually modeling the\nsemantics of data sources requires significant effort and expertise, and\nalthough desirable, building these models automatically is a challenging\nproblem. Most of the related work focuses on semantic annotation of the data\nfields (source attributes). However, constructing a semantic model that\nexplicitly describes the relationships between the attributes in addition to\ntheir semantic types is critical.\n  We present a novel approach that exploits the knowledge from a domain\nontology and the semantic models of previously modeled sources to automatically\nlearn a rich semantic model for a new source. This model represents the\nsemantics of the new source in terms of the concepts and relationships defined\nby the domain ontology. Given some sample data from the new source, we leverage\nthe knowledge in the domain ontology and the known semantic models to construct\na weighted graph that represents the space of plausible semantic models for the\nnew source. Then, we compute the top k candidate semantic models and suggest to\nthe user a ranked list of the semantic models for the new source. The approach\ntakes into account user corrections to learn more accurate semantic models on\nfuture data sources. Our evaluation shows that our method generates expressive\nsemantic models for data sources and services with minimal user input. ...', ""Knowledge bases contribute to many web search and mining tasks, yet they are\noften incomplete. To add missing facts to a given knowledge base, various\nembedding models have been proposed in the recent literature. Perhaps\nsurprisingly, relatively simple models with limited expressiveness often\nperformed remarkably well under today's most commonly used evaluation\nprotocols. In this paper, we explore whether recent models work well for\nknowledge base completion and argue that the current evaluation protocols are\nmore suited for question answering rather than knowledge base completion. We\nshow that when focusing on a different prediction task for evaluating knowledge\nbase completion, the performance of current embedding models is unsatisfactory\neven on datasets previously thought to be too easy. This is especially true\nwhen embedding models are compared against a simple rule-based baseline. This\nwork indicates the need for more research into the embedding models and\nevaluation protocols for knowledge base completion."", ""Can machines trace human knowledge like humans? Knowledge tracing (KT) is a\nfundamental task in a wide range of applications in education, such as massive\nopen online courses (MOOCs), intelligent tutoring systems, educational games,\nand learning management systems. It models dynamics in a student's knowledge\nstates in relation to different learning concepts through their interactions\nwith learning activities. Recently, several attempts have been made to use deep\nlearning models for tackling the KT problem. Although these deep learning\nmodels have shown promising results, they have limitations: either lack the\nability to go deeper to trace how specific concepts in a knowledge state are\nmastered by a student, or fail to capture long-term dependencies in an exercise\nsequence. In this paper, we address these limitations by proposing a novel deep\nlearning model for knowledge tracing, namely Sequential Key-Value Memory\nNetworks (SKVMN). This model unifies the strengths of recurrent modelling\ncapacity and memory capacity of the existing deep learning KT models for\nmodelling student learning. We have extensively evaluated our proposed model on\nfive benchmark datasets. The experimental results show that (1) SKVMN\noutperforms the state-of-the-art KT models on all datasets, (2) SKVMN can\nbetter discover the correlation between latent concepts and questions, and (3)\nSKVMN can trace the knowledge state of students dynamics, and a leverage\nsequential dependencies in an exercise sequence for improved predication\naccuracy."", ""First-order factoid question answering assumes that the question can be\nanswered by a single fact in a knowledge base (KB). While this does not seem\nlike a challenging task, many recent attempts that apply either complex\nlinguistic reasoning or deep neural networks achieve 65%-76% accuracy on\nbenchmark sets. Our approach formulates the task as two machine learning\nproblems: detecting the entities in the question, and classifying the question\nas one of the relation types in the KB. We train a recurrent neural network to\nsolve each problem. On the SimpleQuestions dataset, our approach yields\nsubstantial improvements over previously published results --- even neural\nnetworks based on much more complex architectures. The simplicity of our\napproach also has practical advantages, such as efficiency and modularity, that\nare valuable especially in an industry setting. In fact, we present a\npreliminary analysis of the performance of our model on real queries from\nComcast's X1 entertainment platform with millions of users every day."", 'Path queries on a knowledge graph can be used to answer compositional\nquestions such as ""What languages are spoken by people living in Lisbon?"".\nHowever, knowledge graphs often have missing facts (edges) which disrupts path\nqueries. Recent models for knowledge base completion impute missing facts by\nembedding knowledge graphs in vector spaces. We show that these models can be\nrecursively applied to answer path queries, but that they suffer from cascading\nerrors. This motivates a new ""compositional"" training objective, which\ndramatically improves all models\' ability to answer path queries, in some cases\nmore than doubling accuracy. On a standard knowledge base completion task, we\nalso demonstrate that compositional training acts as a novel form of structural\nregularization, reliably improving performance across all base models (reducing\nerrors by up to 43%) and achieving new state-of-the-art results.', 'Knowledge Graphs (KGs) represent real-world noisy raw information in a\nstructured form, capturing relationships between entities. However, for dynamic\nreal-world applications such as social networks, recommender systems,\ncomputational biology, relational knowledge representation has emerged as a\nchallenging research problem where there is a need to represent the changing\nnodes, attributes, and edges over time. The evolution of search engine\nresponses to user queries in the last few years is partly because of the role\nof KGs such as Google KG. KGs are significantly contributing to various AI\napplications from link prediction, entity relations prediction, node\nclassification to recommendation and question answering systems. This article\nis an attempt to summarize the journey of KG for AI.', 'In this work we propose a system for visual question answering. Our\narchitecture is composed of two parts, the first part creates the logical\nknowledge base given the image. The second part evaluates questions against the\nknowledge base. Differently from previous work, the knowledge base is\nrepresented using hyper-dimensional computing. This choice has the advantage\nthat all the operations in the system, namely creating the knowledge base and\nevaluating the questions against it, are differentiable, thereby making the\nsystem easily trainable in an end-to-end fashion.', 'Multi-hop question answering (QA) requires a model to retrieve and integrate\ninformation from different parts of a long text to answer a question. Humans\nanswer this kind of complex questions via a divide-and-conquer approach. In\nthis paper, we investigate whether top-performing models for multi-hop\nquestions understand the underlying sub-questions like humans. We adopt a\nneural decomposition model to generate sub-questions for a multi-hop complex\nquestion, followed by extracting the corresponding sub-answers. We show that\nmultiple state-of-the-art multi-hop QA models fail to correctly answer a large\nportion of sub-questions, although their corresponding multi-hop questions are\ncorrectly answered. This indicates that these models manage to answer the\nmulti-hop questions using some partial clues, instead of truly understanding\nthe reasoning paths. We also propose a new model which significantly improves\nthe performance on answering the sub-questions. Our work takes a step forward\ntowards building a more explainable multi-hop QA system.', 'The usage and amount of information available on the internet increase over\nthe past decade. This digitization leads to the need for automated answering\nsystem to extract fruitful information from redundant and transitional\nknowledge sources. Such systems are designed to cater the most prominent answer\nfrom this giant knowledge source to the user query using natural language\nunderstanding (NLU) and thus eminently depends on the Question-answering(QA)\nfield.\n  Question answering involves but not limited to the steps like mapping of user\nquestion to pertinent query, retrieval of relevant information, finding the\nbest suitable answer from the retrieved information etc. The current\nimprovement of deep learning models evince compelling performance improvement\nin all these tasks.\n  In this review work, the research directions of QA field are analyzed based\non the type of question, answer type, source of evidence-answer, and modeling\napproach. This detailing followed by open challenges of the field like\nautomatic question generation, similarity detection and, low resource\navailability for a language. In the end, a survey of available datasets and\nevaluation measures is presented.', 'In order to facilitate natural language understanding, the key is to engage\ncommonsense or background knowledge. However, how to engage commonsense\neffectively in question answering systems is still under exploration in both\nresearch academia and industry. In this paper, we propose a novel\nquestion-answering method by integrating multiple knowledge sources, i.e.\nConceptNet, Wikipedia, and the Cambridge Dictionary, to boost the performance.\nMore concretely, we first introduce a novel graph-based iterative knowledge\nretrieval module, which iteratively retrieves concepts and entities related to\nthe given question and its choices from multiple knowledge sources. Afterward,\nwe use a pre-trained language model to encode the question, retrieved knowledge\nand choices, and propose an answer choice-aware attention mechanism to fuse all\nhidden representations of the previous modules. Finally, the linear classifier\nfor specific tasks is used to predict the answer. Experimental results on the\nCommonsenseQA dataset show that our method significantly outperforms other\ncompetitive methods and achieves the new state-of-the-art. In addition, further\nablation studies demonstrate the effectiveness of our graph-based iterative\nknowledge retrieval module and the answer choice-aware attention module in\nretrieving and synthesizing background knowledge from multiple knowledge\nsources.', 'Question Answering is a task which requires building models capable of\nproviding answers to questions expressed in human language. Full question\nanswering involves some form of reasoning ability. We introduce a neural\nnetwork architecture for this task, which is a form of $Memory\\ Network$, that\nrecognizes entities and their relations to answers through a focus attention\nmechanism. Our model is named $Question\\ Dependent\\ Recurrent\\ Entity\\ Network$\nand extends $Recurrent\\ Entity\\ Network$ by exploiting aspects of the question\nduring the memorization process. We validate the model on both synthetic and\nreal datasets: the $bAbI$ question answering dataset and the $CNN\\ \\&\\ Daily\\\nNews$ $reading\\ comprehension$ dataset. In our experiments, the models achieved\na State-of-The-Art in the former and competitive results in the latter.', ""The intelligent question answering (IQA) system can accurately capture users'\nsearch intention by understanding the natural language questions, searching\nrelevant content efficiently from a massive knowledge-base, and returning the\nanswer directly to the user. Since the IQA system can save inestimable time and\nworkforce in data search and reasoning, it has received more and more attention\nin data science and artificial intelligence. This article introduced a domain\nknowledge graph using the graph database and graph computing technologies from\nmassive heterogeneous data in electric power. It then proposed an IQA system\nbased on the electrical power knowledge graph to extract the intent and\nconstraints of natural interrogation based on the natural language processing\n(NLP) method, to construct graph data query statements via knowledge reasoning,\nand to complete the accurate knowledge search and analysis to provide users\nwith an intuitive visualization. This method thoroughly combined knowledge\ngraph and graph computing characteristics, realized high-speed multi-hop\nknowledge correlation reasoning analysis in tremendous knowledge. The proposed\nwork can also provide a basis for the context-aware intelligent question and\nanswer."", 'Question answering (QA) has become a popular way for humans to access\nbillion-scale knowledge bases. Unlike web search, QA over a knowledge base\ngives out accurate and concise results, provided that natural language\nquestions can be understood and mapped precisely to structured queries over the\nknowledge base. The challenge, however, is that a human can ask one question in\nmany different ways. Previous approaches have natural limits due to their\nrepresentations: rule based approaches only understand a small set of ""canned""\nquestions, while keyword based or synonym based approaches cannot fully\nunderstand the questions. In this paper, we design a new kind of question\nrepresentation: templates, over a billion scale knowledge base and a million\nscale QA corpora. For example, for questions about a city\'s population, we\nlearn templates such as What\'s the population of $city?, How many people are\nthere in $city?. We learned 27 million templates for 2782 intents. Based on\nthese templates, our QA system KBQA effectively supports binary factoid\nquestions, as well as complex questions which are composed of a series of\nbinary factoid questions. Furthermore, we expand predicates in RDF knowledge\nbase, which boosts the coverage of knowledge base by 57 times. Our QA system\nbeats all other state-of-art works on both effectiveness and efficiency over\nQALD benchmarks.', 'Search engines based on keyword retrieval can no longer adapt to the way of\ninformation acquisition in the era of intelligent Internet of Things due to the\nreturn of keyword related Internet pages. How to quickly, accurately and\neffectively obtain the information needed by users from massive Internet data\nhas become one of the key issues urgently needed to be solved. We propose an\nintelligent question-answering system based on structured KB and unstructured\ndata, called OpenQA, in which users can give query questions and the model can\nquickly give accurate answers back to users. We integrate KBQA structured\nquestion answering based on semantic parsing and deep representation learning,\nand two-stage unstructured question answering based on retrieval and neural\nmachine reading comprehension into OpenQA, and return the final answer with the\nhighest probability through the Transformer answer selection module in OpenQA.\nWe carry out preliminary experiments on our constructed dataset, and the\nexperimental results prove the effectiveness of the proposed intelligent\nquestion answering system. At the same time, the core technology of each module\nof OpenQA platform is still in the forefront of academic hot spots, and the\ntheoretical essence and enrichment of OpenQA will be further explored based on\nthese academic hot spots.', 'Thanks to the development of the Semantic Web, a lot of new structured data\nhas become available on the Web in the form of knowledge bases (KBs). Making\nthis valuable data accessible and usable for end-users is one of the main goals\nof Question Answering (QA) over KBs. Most current QA systems query one KB, in\none language (namely English). The existing approaches are not designed to be\neasily adaptable to new KBs and languages. We first introduce a new approach\nfor translating natural language questions to SPARQL queries. It is able to\nquery several KBs simultaneously, in different languages, and can easily be\nported to other KBs and languages. In our evaluation, the impact of our\napproach is proven using 5 different well-known and large KBs: Wikidata,\nDBpedia, MusicBrainz, DBLP and Freebase as well as 5 different languages namely\nEnglish, German, French, Italian and Spanish. Second, we show how we integrated\nour approach, to make it easily accessible by the research community and by\nend-users. To summarize, we provided a conceptional solution for multilingual,\nKB-agnostic Question Answering over the Semantic Web. The provided first\napproximation validates this concept.', 'Answering complex questions is a time-consuming activity for humans that\nrequires reasoning and integration of information. Recent work on reading\ncomprehension made headway in answering simple questions, but tackling complex\nquestions is still an ongoing research challenge. Conversely, semantic parsers\nhave been successful at handling compositionality, but only when the\ninformation resides in a target knowledge-base. In this paper, we present a\nnovel framework for answering broad and complex questions, assuming answering\nsimple questions is possible using a search engine and a reading comprehension\nmodel. We propose to decompose complex questions into a sequence of simple\nquestions, and compute the final answer from the sequence of answers. To\nillustrate the viability of our approach, we create a new dataset of complex\nquestions, ComplexWebQuestions, and present a model that decomposes questions\nand interacts with the web to compute an answer. We empirically demonstrate\nthat question decomposition improves performance from 20.8 precision@1 to 27.5\nprecision@1 on this new dataset.', 'In artificial intelligence (AI), knowledge is the information required by an\nintelligent system to accomplish tasks. While traditional knowledge bases use\ndiscrete, symbolic representations, detecting knowledge encoded in the\ncontinuous representations learned from data has received increasing attention\nrecently. In this work, we propose a method for building a continuous knowledge\nbase (CKB) that can store knowledge imported from multiple, diverse neural\nnetworks. The key idea of our approach is to define an interface for each\nneural network and cast knowledge transferring as a function simulation\nproblem. Experiments on text classification show promising results: the CKB\nimports knowledge from a single model and then exports the knowledge to a new\nmodel, achieving comparable performance with the original model. More\ninteresting, we import the knowledge from multiple models to the knowledge\nbase, from which the fused knowledge is exported back to a single model,\nachieving a higher accuracy than the original model. With the CKB, it is also\neasy to achieve knowledge distillation and transfer learning. Our work opens\nthe door to building a universal continuous knowledge base to collect, store,\nand organize all continuous knowledge encoded in various neural networks\ntrained for different AI tasks.', ""Within Reinforcement Learning, there is a growing collection of research\nwhich aims to express all of an agent's knowledge of the world through\npredictions about sensation, behaviour, and time. This work can be seen not\nonly as a collection of architectural proposals, but also as the beginnings of\na theory of machine knowledge in reinforcement learning. Recent work has\nexpanded what can be expressed using predictions, and developed applications\nwhich use predictions to inform decision-making on a variety of synthetic and\nreal-world problems. While promising, we here suggest that the notion of\npredictions as knowledge in reinforcement learning is as yet underdeveloped:\nsome work explicitly refers to predictions as knowledge, what the requirements\nare for considering a prediction to be knowledge have yet to be well explored.\nThis specification of the necessary and sufficient conditions of knowledge is\nimportant; even if claims about the nature of knowledge are left implicit in\ntechnical proposals, the underlying assumptions of such claims have\nconsequences for the systems we design. These consequences manifest in both the\nway we choose to structure predictive knowledge architectures, and how we\nevaluate them. In this paper, we take a first step to formalizing predictive\nknowledge by discussing the relationship of predictive knowledge learning\nmethods to existing theories of knowledge in epistemology. Specifically, we\nexplore the relationships between Generalized Value Functions and epistemic\nnotions of Justification and Truth."", ""The paper presents an approach to build a question and answer system that is\ncapable of processing the information in a large dataset and allows the user to\ngain knowledge from this dataset by asking questions in natural language form.\nKey content of this research covers four dimensions which are; Corpus\nPreprocessing, Question Preprocessing, Deep Neural Network for Answer\nExtraction and Answer Generation. The system is capable of understanding the\nquestion, responds to the user's query in natural language form as well. The\ngoal is to make the user feel as if they were interacting with a person than a\nmachine."", 'Over the past few years, question answering and information retrieval systems\nhave become widely used. These systems attempt to find the answer of the asked\nquestions from raw text sources. A component of these systems is Answer\nSelection which selects the most relevant from candidate answers. Syntactic\nsimilarities were mostly used to compute the similarity, but in recent works,\ndeep neural networks have been used, making a significant improvement in this\nfield. In this research, a model is proposed to select the most relevant\nanswers to the factoid question from the candidate answers. The proposed model\nranks the candidate answers in terms of semantic and syntactic similarity to\nthe question, using convolutional neural networks. In this research, Attention\nmechanism and Sparse feature vector use the context-sensitive interactions\nbetween questions and answer sentence. Wide convolution increases the\nimportance of the interrogative word. Pairwise ranking is used to learn\ndifferentiable representations to distinguish positive and negative answers.\nOur model indicates strong performance on the TrecQA Raw beating previous\nstate-of-the-art systems by 1.4% in MAP and 1.1% in MRR while using the\nbenefits of no additional syntactic parsers and external tools. The results\nshow that using context-sensitive interactions between question and answer\nsentences can help to find the correct answer more accurately.', 'Domain-specific community question answering is becoming an integral part of\nprofessions. Finding related questions and answers in these communities can\nsignificantly improve the effectiveness and efficiency of information seeking.\nStack Overflow is one of the most popular communities that is being used by\nmillions of programmers. In this paper, we analyze the problem of predicting\nknowledge unit (question thread) relatedness in Stack Overflow. In particular,\nwe formulate the question relatedness task as a multi-class classification\nproblem with four degrees of relatedness. We present a large-scale dataset with\nmore than 300K pairs. To the best of our knowledge, this dataset is the largest\ndomain-specific dataset for Question-Question relatedness. We present the steps\nthat we took to collect, clean, process, and assure the quality of the dataset.\nThe proposed dataset Stack Overflow is a useful resource to develop novel\nsolutions, specifically data-hungry neural network models, for the prediction\nof relatedness in technical community question-answering forums. We adopt a\nneural network architecture and a traditional model for this task that\neffectively utilize information from different parts of knowledge units to\ncompute the relatedness between them. These models can be used to benchmark\nnovel models, as they perform well in our task and in a closely similar task.', 'This paper tackles the problem of endogenous link prediction for Knowledge\nBase completion. Knowledge Bases can be represented as directed graphs whose\nnodes correspond to entities and edges to relationships. Previous attempts\neither consist of powerful systems with high capacity to model complex\nconnectivity patterns, which unfortunately usually end up overfitting on rare\nrelationships, or in approaches that trade capacity for simplicity in order to\nfairly model all relationships, frequent or not. In this paper, we propose\nTatec a happy medium obtained by complementing a high-capacity model with a\nsimpler one, both pre-trained separately and then combined. We present several\nvariants of this model with different kinds of regularization and combination\nstrategies and show that this approach outperforms existing methods on\ndifferent types of relationships by achieving state-of-the-art results on four\nbenchmarks of the literature.', ""Gene interaction graphs aim to capture various relationships between genes\nand can represent decades of biology research. When trying to make predictions\nfrom genomic data, those graphs could be used to overcome the curse of\ndimensionality by making machine learning models sparser and more consistent\nwith biological common knowledge. In this work, we focus on assessing how well\nthose graphs capture dependencies seen in gene expression data to evaluate the\nadequacy of the prior knowledge provided by those graphs. We propose a\ncondition graphs should satisfy to provide good prior knowledge and test it\nusing `Single Gene Inference' tasks. We also compare with randomly generated\ngraphs, aiming to measure the true benefit of using biologically relevant\ngraphs in this context, and validate our findings with five clinical tasks. We\nfind some graphs capture relevant dependencies for most genes while being very\nsparse. Our analysis with random graphs finds that dependencies can be captured\nalmost as well at random which suggests that, in terms of gene expression\nlevels, the relevant information about the state of the cell is spread across\nmany genes."", 'The purpose of this paper is to describe and analyze the development of a\nknowledge-based infrastructure to support MSP decision-making processes. The\npaper emerged from a study to define specifications for a knowledge-based\ninfrastructure to provide decision support for community-level MSPs in the\nCanadian province of Quebec. As part of the study, a process assessment was\nconducted to understand the needs of communities as they collect, organize, and\nanalyze data to make decisions about their priorities. The result of this\nprocess is a portrait, which is an epidemiological profile of health and\nnutrition in their community. Portraits inform strategic planning and\ndevelopment of interventions and are used to assess the impact of\ninterventions. Our key findings indicate ambiguities and disagreement among MSP\ndecision-makers regarding causal relationships between actions and outcomes,\nand the relevant data needed for making decisions. MSP decision-makers\nexpressed a desire for easy-to-use tools that facilitate the collection,\norganization, synthesis, and analysis of data, to enable decision-making in a\ntimely manner. Findings inform conceptual modeling and ontological analysis to\ncapture the domain knowledge and specify relationships between actions and\noutcomes. This modeling and analysis provide the foundation for an ontology,\nencoded using OWL 2 Web Ontology Language. The ontology is developed to provide\nsemantic support for the MSP process, defining objectives, strategies, actions,\nindicators, and data sources. In the future, software interacting with the\nontology can facilitate interactive browsing by decision-makers in the MSP in\nthe form of concepts, instances, relationships, and axioms. Our ontology also\nfacilitates the integration and interpretation of community data and can help\nin managing semantic interoperability between different knowledge sources.', 'Cold start knowledge base population (KBP) is the problem of populating a\nknowledge base from unstructured documents. While artificial neural networks\nhave led to significant improvements in the different tasks that are part of\nKBP, the overall F1 of the end-to-end system remains quite low. This problem is\nmore acute in personal knowledge bases, which present additional challenges\nwith regard to data protection, fairness and privacy. In this work, we present\na system that uses rule based annotators and a graph neural network for missing\nlink prediction, to populate a more complete, fair and diverse knowledge base\nfrom the TACRED dataset.', 'We aim to provide table answers to keyword queries against knowledge bases.\nFor queries referring to multiple entities, like ""Washington cities population""\nand ""Mel Gibson movies"", it is better to represent each relevant answer as a\ntable which aggregates a set of entities or entity-joins within the same table\nscheme or pattern. In this paper, we study how to find highly relevant patterns\nin a knowledge base for user-given keyword queries to compose table answers. A\nknowledge base can be modeled as a directed graph called knowledge graph, where\nnodes represent entities in the knowledge base and edges represent the\nrelationships among them. Each node/edge is labeled with type and text. A\npattern is an aggregation of subtrees which contain all keywords in the texts\nand have the same structure and types on node/edges. We propose efficient\nalgorithms to find patterns that are relevant to the query for a class of\nscoring functions. We show the hardness of the problem in theory, and propose\npath-based indexes that are affordable in memory. Two query-processing\nalgorithms are proposed: one is fast in practice for small queries (with small\npatterns as answers) by utilizing the indexes; and the other one is better in\ntheory, with running time linear in the sizes of indexes and answers, which can\nhandle large queries better. We also conduct extensive experimental study to\ncompare our approaches with a naive adaption of known techniques.', 'Answering complex questions involving multiple entities and relations is a\nchallenging task. Logically, the answer to a complex question should be derived\nby decomposing the complex question into multiple simple sub-questions and then\nanswering those sub-questions. Existing work has followed this strategy but has\nnot attempted to optimize the order of how those sub-questions are answered. As\na result, the sub-questions are answered in an arbitrary order, leading to\nlarger search space and a higher risk of missing an answer. In this paper, we\npropose a novel reinforcement learning(RL) approach to answering complex\nquestions that can learn a policy to dynamically decide which sub-question\nshould be answered at each stage of reasoning. We lever-age the expected\nvalue-variance criterion to enable the learned policy to balance between the\nrisk and utility of answering a sub-question. Experiment results show that the\nRL approach can substantially improve the optimality of ordering the\nsub-questions, leading to improved accuracy of question answering. The proposed\nmethod for learning to order sub-questions is general and can thus be\npotentially combined with many existing ideas for answering complex questions\nto enhance their performance.', 'Building computers able to answer questions on any subject is a long standing\ngoal of artificial intelligence. Promising progress has recently been achieved\nby methods that learn to map questions to logical forms or database queries.\nSuch approaches can be effective but at the cost of either large amounts of\nhuman-labeled data or by defining lexicons and grammars tailored by\npractitioners. In this paper, we instead take the radical approach of learning\nto map questions to vectorial feature representations. By mapping answers into\nthe same space one can query any knowledge base independent of its schema,\nwithout requiring any grammar or lexicon. Our method is trained with a new\noptimization procedure combining stochastic gradient descent followed by a\nfine-tuning step using the weak supervision provided by blending automatically\nand collaboratively generated resources. We empirically demonstrate that our\nmodel can capture meaningful signals from its noisy supervision leading to\nmajor improvements over paralex, the only existing method able to be trained on\nsimilar weakly labeled data.', 'Knowledge graph (KG) is known to be helpful for the task of question\nanswering (QA), since it provides well-structured relational information\nbetween entities, and allows one to further infer indirect facts. However, it\nis challenging to build QA systems which can learn to reason over knowledge\ngraphs based on question-answer pairs alone. First, when people ask questions,\ntheir expressions are noisy (for example, typos in texts, or variations in\npronunciations), which is non-trivial for the QA system to match those\nmentioned entities to the knowledge graph. Second, many questions require\nmulti-hop logic reasoning over the knowledge graph to retrieve the answers. To\naddress these challenges, we propose a novel and unified deep learning\narchitecture, and an end-to-end variational learning algorithm which can handle\nnoise in questions, and learn multi-hop reasoning simultaneously. Our method\nachieves state-of-the-art performance on a recent benchmark dataset in the\nliterature. We also derive a series of new benchmark datasets, including\nquestions for multi-hop reasoning, questions paraphrased by neural translation\nmodel, and questions in human voice. Our method yields very promising results\non all these challenging datasets.', 'Semantic matching is of central significance to the answer selection task\nwhich aims to select correct answers for a given question from a candidate\nanswer pool. A useful method is to employ neural networks with attention to\ngenerate sentences representations in a way that information from pair\nsentences can mutually influence the computation of representations. In this\nwork, an effective architecture,multi-size neural network with attention\nmechanism (AM-MSNN),is introduced into the answer selection task. This\narchitecture captures more levels of language granularities in parallel,\nbecause of the various sizes of filters comparing with single-layer CNN and\nmulti-layer CNNs. Meanwhile it extends the sentence representations by\nattention mechanism, thus containing more information for different types of\nquestions. The empirical study on three various benchmark tasks of answer\nselection demonstrates the efficacy of the proposed model in all the benchmarks\nand its superiority over competitors. The experimental results show that (1)\nmulti-size neural network (MSNN) is a more useful method to capture abstract\nfeatures on different levels of granularities than single/multi-layer CNNs; (2)\nthe attention mechanism (AM) is a better strategy to derive more informative\nrepresentations; (3) AM-MSNN is a better architecture for the answer selection\ntask for the moment.', ""Retrofitting is a technique used to move word vectors closer together or\nfurther apart in their space to reflect their relationships in a Knowledge Base\n(KB). However, retrofitting only works on concepts that are present in that KB.\nRetroGAN uses a pair of Generative Adversarial Networks (GANs) to learn a\none-to-one mapping between concepts and their retrofitted counterparts. It\napplies that mapping (post-specializes) to handle concepts that do not appear\nin the original KB in a manner similar to how some natural language systems\nhandle out-of-vocabulary entries. We test our system on three word-similarity\nbenchmarks and a downstream sentence simplification task and achieve the state\nof the art (CARD-660). Altogether, our results demonstrate our system's\neffectiveness for out-of-knowledge and rare word generalization."", 'To date, most probabilistic reasoning systems have relied on a fixed belief\nnetwork constructed at design time. The network is used by an application\nprogram as a representation of (in)dependencies in the domain. Probabilistic\ninference algorithms operate over the network to answer queries. Recognizing\nthe inflexibility of fixed models has led researchers to develop automated\nnetwork construction procedures that use an expressive knowledge base to\ngenerate a network that can answer a query. Although more flexible than fixed\nmodel approaches, these construction procedures separate construction and\nevaluation into distinct phases. In this paper we develop an approach to\ncombining incremental construction and evaluation of a partial probability\nmodel. The combined method holds promise for improved methods for control of\nmodel construction based on a trade-off between fidelity of results and cost of\nconstruction.', 'Knowledge-based biomedical data science (KBDS) involves the design and\nimplementation of computer systems that act as if they knew about biomedicine.\nSuch systems depend on formally represented knowledge in computer systems,\noften in the form of knowledge graphs. Here we survey the progress in the last\nyear in systems that use formally represented knowledge to address data science\nproblems in both clinical and biological domains, as well as on approaches for\ncreating knowledge graphs. Major themes include the relationships between\nknowledge graphs and machine learning, the use of natural language processing,\nand the expansion of knowledge-based approaches to novel domains, such as\nChinese Traditional Medicine and biodiversity.', 'Ontology matching is the process of automatically determining the semantic\nequivalences between the concepts of two ontologies. Most ontology matching\nalgorithms are based on two types of strategies: terminology-based strategies,\nwhich align concepts based on their names or descriptions, and structure-based\nstrategies, which exploit concept hierarchies to find the alignment. In many\ndomains, there is additional information about the relationships of concepts\nrepresented in various ways, such as Bayesian networks, decision trees, and\nassociation rules. We propose to use the similarities between these\nrelationships to find more accurate alignments. We accomplish this by defining\nsoft constraints that prefer alignments where corresponding concepts have the\nsame local relationships encoded as knowledge rules. We use a probabilistic\nframework to integrate this new knowledge-based strategy with standard\nterminology-based and structure-based strategies. Furthermore, our method is\nparticularly effective in identifying correspondences between complex concepts.\nOur method achieves substantially better F-score than the previous\nstate-of-the-art on three ontology matching domains.', 'In statistical relational learning, knowledge graph completion deals with\nautomatically understanding the structure of large knowledge graphs---labeled\ndirected graphs---and predicting missing relationships---labeled edges.\nState-of-the-art embedding models propose different trade-offs between modeling\nexpressiveness, and time and space complexity. We reconcile both expressiveness\nand complexity through the use of complex-valued embeddings and explore the\nlink between such complex-valued embeddings and unitary diagonalization. We\ncorroborate our approach theoretically and show that all real square\nmatrices---thus all possible relation/adjacency matrices---are the real part of\nsome unitarily diagonalizable matrix. This results opens the door to a lot of\nother applications of square matrices factorization. Our approach based on\ncomplex embeddings is arguably simple, as it only involves a Hermitian dot\nproduct, the complex counterpart of the standard dot product between real\nvectors, whereas other methods resort to more and more complicated composition\nfunctions to increase their expressiveness. The proposed complex embeddings are\nscalable to large data sets as it remains linear in both space and time, while\nconsistently outperforming alternative approaches on standard link prediction\nbenchmarks.', 'Automatic question-answering is a classical problem in natural language\nprocessing, which aims at designing systems that can automatically answer a\nquestion, in the same way as human does. In this work, we propose a deep\nlearning based model for automatic question-answering. First the questions and\nanswers are embedded using neural probabilistic modeling. Then a deep\nsimilarity neural network is trained to find the similarity score of a pair of\nanswer and question. Then for each question, the best answer is found as the\none with the highest similarity score. We first train this model on a\nlarge-scale public question-answering database, and then fine-tune it to\ntransfer to the customer-care chat data. We have also tested our framework on a\npublic question-answering database and achieved very good performance.', 'Deep reading models for question-answering have demonstrated promising\nperformance over the last couple of years. However current systems tend to\nlearn how to cleverly extract a span of the source document, based on its\nsimilarity with the question, instead of seeking for the appropriate answer.\nIndeed, a reading machine should be able to detect relevant passages in a\ndocument regarding a question, but more importantly, it should be able to\nreason over the important pieces of the document in order to produce an answer\nwhen it is required. To motivate this purpose, we present ReviewQA, a\nquestion-answering dataset based on hotel reviews. The questions of this\ndataset are linked to a set of relational understanding competencies that we\nexpect a model to master. Indeed, each question comes with an associated type\nthat characterizes the required competency. With this framework, it is possible\nto benchmark the main families of models and to get an overview of what are the\nstrengths and the weaknesses of a given model on the set of tasks evaluated in\nthis dataset. Our corpus contains more than 500.000 questions in natural\nlanguage over 100.000 hotel reviews. Our setup is projective, the answer of a\nquestion does not need to be extracted from a document, like in most of the\nrecent datasets, but selected among a set of candidates that contains all the\npossible answers to the questions of the dataset. Finally, we present several\nbaselines over this dataset.', 'Question Answering has recently received high attention from artificial\nintelligence communities due to the advancements in learning technologies.\nEarly question answering models used rule-based approaches and moved to the\nstatistical approach to address the vastly available information. However,\nstatistical approaches are shown to underperform in handling the dynamic nature\nand the variation of language. Therefore, learning models have shown the\ncapability of handling the dynamic nature and variations in language. Many deep\nlearning methods have been introduced to question answering. Most of the deep\nlearning approaches have shown to achieve higher results compared to machine\nlearning and statistical methods. The dynamic nature of language has profited\nfrom the nonlinear learning in deep learning. This has created prominent\nsuccess and a spike in work on question answering. This paper discusses the\nsuccesses and challenges in question answering question answering systems and\ntechniques that are used in these challenges.', 'In most current applications of belief networks, domain knowledge is\nrepresented by a single belief network that applies to all problem instances in\nthe domain. In more complex domains, problem-specific models must be\nconstructed from a knowledge base encoding probabilistic relationships in the\ndomain. Most work in knowledge-based model construction takes the rule as the\nbasic unit of knowledge. We present a knowledge representation framework that\npermits the knowledge base designer to specify knowledge in larger semantically\nmeaningful units which we call network fragments. Our framework provides for\nrepresentation of asymmetric independence and canonical intercausal\ninteraction. We discuss the combination of network fragments to form\nproblem-specific models to reason about particular problem instances. The\nframework is illustrated using examples from the domain of military situation\nawareness.', 'Question answering is one of the primary challenges of natural language\nunderstanding. In realizing such a system, providing complex long answers to\nquestions is a challenging task as opposed to factoid answering as the former\nneeds context disambiguation. The different methods explored in the literature\ncan be broadly classified into three categories namely: 1) classification\nbased, 2) knowledge graph based and 3) retrieval based. Individually, none of\nthem address the need of an enterprise wide assistance system for an IT support\nand maintenance domain. In this domain the variance of answers is large ranging\nfrom factoid to structured operating procedures; the knowledge is present\nacross heterogeneous data sources like application specific documentation,\nticket management systems and any single technique for a general purpose\nassistance is unable to scale for such a landscape. To address this, we have\nbuilt a cognitive platform with capabilities adopted for this domain. Further,\nwe have built a general purpose question answering system leveraging the\nplatform that can be instantiated for multiple products, technologies in the\nsupport domain. The system uses a novel hybrid answering model that\norchestrates across a deep learning classifier, a knowledge graph based context\ndisambiguation module and a sophisticated bag-of-words search system. This\norchestration performs context switching for a provided question and also does\na smooth hand-off of the question to a human expert if none of the automated\ntechniques can provide a confident answer. This system has been deployed across\n675 internal enterprise IT support and maintenance projects.', 'An enhanced approach for network monitoring is to create a network monitoring\ntool that has artificial intelligence characteristics. There are a number of\napproaches available. One such approach is by the use of a combination of rule\nbased, fuzzy logic and neural networks to create a hybrid ANFIS system. Such\nsystem will have a dual knowledge database approach. One containing membership\nfunction values to compare to and do deductive reasoning and another database\nwith rules deductively formulated by an expert (a network administrator). The\nknowledge database will be updated continuously with newly acquired patterns.\nIn short, the system will be composed of 2 parts, learning from data sets and\nfine-tuning the knowledge-base using neural network and the use of fuzzy logic\nin making decision based on the rules and membership functions inside the\nknowledge base. This paper will discuss the idea, steps and issues involved in\ncreating such a system.', 'In this document I present an approach to answer validation and reranking for\nquestion answering (QA) systems. A cased-based reasoning (CBR) system judges\nanswer candidates for questions from annotated answer candidates for earlier\nquestions. The promise of this approach is that user feedback will result in\nimproved answers of the QA system, due to the growing case base. In the paper,\nI present the adequate structuring of the case base and the appropriate\nselection of relevant similarity measures, in order to solve the answer\nvalidation problem. The structural case base is built from annotated MultiNet\ngraphs, which provide representations for natural language expressions, and\ncorresponding graph similarity measures. I cover a priori relations to\nexperienced answer candidates for former questions. I compare the CBR System\nresults to current approaches in an experiment integrating CBR into an existing\nframework for answer validation and reranking. This integration is achieved by\nadding CBR-related features to the input of a learned ranking model that\ndetermines the final answer ranking. In the experiments based on QA@CLEF\nquestions, the best learned models make heavy use of CBR features. Observing\nthe results with a continually growing case base, I present a positive effect\nof the size of the case base on the accuracy of the CBR subsystem.', ""This article describes a novel approach to expand in run-time the knowledge\nbase of an Artificial Conversational Agent. A technique for automatic knowledge\nextraction from the user's sentence and four methods to insert the new acquired\nconcepts in the knowledge base have been developed and integrated into a system\nthat has already been tested for knowledge-based conversation between a social\nhumanoid robot and residents of care homes. The run-time addition of new\nknowledge allows overcoming some limitations that affect most robots and\nchatbots: the incapability of engaging the user for a long time due to the\nrestricted number of conversation topics. The insertion in the knowledge base\nof new concepts recognized in the user's sentence is expected to result in a\nwider range of topics that can be covered during an interaction, making the\nconversation less repetitive. Two experiments are presented to assess the\nperformance of the knowledge extraction technique, and the efficiency of the\ndeveloped insertion methods when adding several concepts in the Ontology."", 'Since large knowledge bases are typically incomplete, missing facts need to\nbe inferred from observed facts in a task called knowledge base completion. The\nmost successful approaches to this task have typically explored explicit paths\nthrough sequences of triples. These approaches have usually resorted to\nhuman-designed sampling procedures, since large knowledge graphs produce\nprohibitively large numbers of possible paths, most of which are uninformative.\nAs an alternative approach, we propose performing a single, short sequence of\ninteractive lookup operations on an embedded knowledge graph which has been\ntrained through end-to-end backpropagation to be an optimized and compressed\nversion of the initial knowledge base. Our proposed model, called Embedded\nKnowledge Graph Network (EKGN), achieves new state-of-the-art results on\npopular knowledge base completion benchmarks.', 'Question Answering (QA) systems are becoming the inspiring model for the\nfuture of search engines. While recently, underlying datasets for QA systems\nhave been promoted from unstructured datasets to structured datasets with\nhighly semantic-enriched metadata, but still question answering systems involve\nserious challenges which cause to be far beyond desired expectations. In this\npaper, we raise the challenges for building a Question Answering (QA) system\nespecially with the focus of employing structured data (i.e. knowledge graph).\nThis paper provide an exhaustive insight of the known challenges, so far. Thus,\nit helps researchers to easily spot open rooms for the future research agenda.', 'People have information needs of varying complexity, which can be solved by\nan intelligent agent able to answer questions formulated in a proper way,\neventually considering user context and preferences. In a scenario in which the\nuser profile can be considered as a question, intelligent agents able to answer\nquestions can be used to find the most relevant answers for a given user. In\nthis work we propose a novel model based on Artificial Neural Networks to\nanswer questions with multiple answers by exploiting multiple facts retrieved\nfrom a knowledge base. The model is evaluated on the factoid Question Answering\nand top-n recommendation tasks of the bAbI Movie Dialog dataset. After\nassessing the performance of the model on both tasks, we try to define the\nlong-term goal of a conversational recommender system able to interact using\nnatural language and to support users in their information seeking processes in\na personalized way.', 'This paper presents a system which learns to answer questions on a broad\nrange of topics from a knowledge base using few hand-crafted features. Our\nmodel learns low-dimensional embeddings of words and knowledge base\nconstituents; these representations are used to score natural language\nquestions against candidate answers. Training our system using pairs of\nquestions and structured representations of their answers, and pairs of\nquestion paraphrases, yields competitive results on a competitive benchmark of\nthe literature.', 'The aim of medical knowledge representation is to capture the detailed domain\nknowledge in a clinically efficient manner and to offer a reliable resolution\nwith the acquired knowledge. The knowledge base to be used by a medical expert\nsystem should allow incremental growth with inclusion of updated knowledge over\nthe time. As knowledge are gathered from a variety of knowledge sources by\ndifferent knowledge engineers, the problem of redundancy is an important\nconcern here due to increased processing time of knowledge and occupancy of\nlarge computational storage to accommodate all the gathered knowledge. Also\nthere may exist many inconsistent knowledge in the knowledge base. In this\npaper, we have proposed a rough set based lattice structure for knowledge\nrepresentation in medical expert systems which overcomes the problem of\nredundancy and inconsistency in knowledge and offers computational efficiency\nwith respect to both time and space. We have also generated an optimal set of\ndecision rules that would be used directly by the inference engine. The\nreliability of each rule has been measured using a new metric called\ncredibility factor, and the certainty and coverage factors of a decision rule\nhave been re-defined. With a set of decisions rules arranged in descending\norder according to their reliability measures, the medical expert system will\nconsider the highly reliable and certain rules at first, then it would search\nfor the possible and uncertain rules at later stage, if recommended by\nphysicians. The proposed knowledge representation technique has been\nillustrated using an example from the domain of low back pain. The proposed\nscheme ensures completeness, consistency, integrity, non-redundancy, and ease\nof access.', ""Search engines can quickly response a hyperlink list according to query\nkeywords. However, when a query is complex, developers need to repeatedly\nrefine the search keywords and open a large number of web pages to find and\nsummarize answers. Many research works of question and answering (Q and A)\nsystem attempt to assist search engines by providing simple, accurate and\nunderstandable answers. However, without original semantic contexts, these\nanswers lack explainability, making them difficult for users to trust and\nadopt. In this paper, a brain-inspired search engine assistant named\nDeveloperBot based on knowledge graph is proposed, which aligns to the\ncognitive process of human and has the capacity to answer complex queries with\nexplainability. Specifically, DeveloperBot firstly constructs a multi-layer\nquery graph by splitting a complex multi-constraint query into several ordered\nconstraints. Then it models the constraint reasoning process as subgraph search\nprocess inspired by the spreading activation model of cognitive science. In the\nend, novel features of the subgraph will be extracted for decision-making. The\ncorresponding reasoning subgraph and answer confidence will be derived as\nexplanations. The results of the decision-making demonstrate that DeveloperBot\ncan estimate the answers and answer confidences with high accuracy. We\nimplement a prototype and conduct a user study to evaluate whether and how the\ndirect answers and the explanations provided by DeveloperBot can assist\ndevelopers' information needs."", 'Nowadays, a huge amount of knowledge has been amassed in digital agriculture.\nThis knowledge and know-how information are collected from various sources,\nhence the question is how to organise this knowledge so that it can be\nefficiently exploited. Although this knowledge about agriculture practices can\nbe represented using ontology, rule-based expert systems, or knowledge model\nbuilt from data mining processes, the scalability still remains an open issue.\nIn this study, we propose a knowledge representation model, called an\nontology-based knowledge map, which can collect knowledge from different\nsources, store it, and exploit either directly by stakeholders or as an input\nto the knowledge discovery process (Data Mining). The proposed model consists\nof two stages, 1) build an ontology as a knowledge base for a specific domain\nand data mining concepts, and 2) build the ontology-based knowledge map model\nfor representing and storing the knowledge mined on the crop datasets. A\nframework of the proposed model has been implemented in agriculture domain. It\nis an efficient and scalable model, and it can be used as knowledge repository\na digital agriculture.', 'Functional and inclusion dependencies are the most widely used classes of\ndata dependencies in data profiling due to their ability to identify\nrelationships in data such as primary and foreign keys. These relationships are\nequally important when dealing with nested data formats such as JSON. However,\nthe definition of functional and inclusion dependencies makes use of a flat,\nunnested relational model which misses many useful types of dependencies on\ndata which involve nested data models.\n  In this work, we identify types of dependencies which are not captured by\ntraditional functional and inclusion dependencies but which nevertheless\ncapture meaningful relationships among nested data. We also demonstrate how\nalgorithms for mining these traditional dependencies can be adapted to also\nmine nested dependencies. The first strategy simply flattens the input data and\nfeeds into unmodified existing algorithms. We present a second strategy which\ninstead adapts the algorithm to efficiently process JSON data as input which in\nsome cases leads to a reduction in runtime by multiple orders of magnitude on\nreal-world datasets. We further show how these algorithms can be adapted to\nproduce useful results in the presence of a percentage of incomplete or invalid\ndata.', 'We present a mechanism for constructing graphical models, specifically\nBayesian networks, from a knowledge base of general probabilistic information.\nThe unique feature of our approach is that it uses a powerful first-order\nprobabilistic logic for expressing the general knowledge base. This logic\nallows for the representation of a wide range of logical and probabilistic\ninformation. The model construction procedure we propose uses notions from\ndirect inference to identify pieces of local statistical information from the\nknowledge base that are most appropriate to the particular event we want to\nreason about. These pieces are composed to generate a joint probability\ndistribution specified as a Bayesian network. Although there are fundamental\ndifficulties in dealing with fully general knowledge, our procedure is\npractical for quite rich knowledge bases and it supports the construction of a\nfar wider range of networks than allowed for by current template technology.', 'Knowledge base is the way to store structured and unstructured data\nthroughout the web. Since the size of the web is increasing rapidly, there are\nhuge needs to structure the knowledge in a fully automated way. However\nfully-automated knowledge-base evolution on the Semantic Web is a major\nchallenges, although there are many ontology evolution techniques available.\nTherefore learning ontology automatically can contribute to the semantic web\nsociety significantly. In this paper, we propose full-automated ontology\nlearning algorithm to generate refined knowledge base from incomplete knowledge\nbase and rdf-triples. Our algorithm is data-driven approach which is based on\nthe property of each instance. Ontology class is being elaborated by\ngeneralizing frequent property of its instances. By using that developed class\ninformation, each instance can find its most relatively matching class. By\nrepeating these two steps, we achieve fully-automated ontology evolution from\nincomplete basic knowledge base.', 'In this paper we discuss the relationships between conditional and\npreferential logics and neural network models, based on a multi-preferential\nsemantics. We propose a concept-wise multipreference semantics, recently\nintroduced for defeasible description logics to take into account preferences\nwith respect to different concepts, as a tool for providing a semantic\ninterpretation to neural network models. This approach has been explored both\nfor unsupervised neural network models (Self-Organising Maps) and for\nsupervised ones (Multilayer Perceptrons), and we expect that the same approach\nmight be extended to other neural network models. It allows for logical\nproperties of the network to be checked (by model checking) over an\ninterpretation capturing the input-output behavior of the network. For\nMultilayer Perceptrons, the deep network itself can be regarded as a\nconditional knowledge base, in which synaptic connections correspond to\nweighted conditionals. The paper describes the general approach, through the\ncases of Self-Organising Maps and Multilayer Perceptrons, and discusses some\nopen issues and perspectives.', 'Time-efficient link discovery is of central importance to implement the\nvision of the Semantic Web. Some of the most rapid Link Discovery approaches\nrely internally on planning to execute link specifications. In newer works,\nlinear models have been used to estimate the runtime the fastest planners.\nHowever, no other category of models has been studied for this purpose so far.\nIn this paper, we study non-linear runtime estimation functions for runtime\nestimation. In particular, we study exponential and mixed models for the\nestimation of the runtimes of planners. To this end, we evaluate three\ndifferent models for runtime on six datasets using 400 link specifications. We\nshow that exponential and mixed models achieve better fits when trained but are\nonly to be preferred in some cases. Our evaluation also shows that the use of\nbetter runtime approximation models has a positive impact on the overall\nexecution of link specifications.', 'Knowledge base construction (KBC) is the process of populating a knowledge\nbase, i.e., a relational database together with inference rules, with\ninformation extracted from documents and structured sources. KBC blurs the\ndistinction between two traditional database problems, information extraction\nand information integration. For the last several years, our group has been\nbuilding knowledge bases with scientific collaborators. Using our approach, we\nhave built knowledge bases that have comparable and sometimes better quality\nthan those constructed by human volunteers. In contrast to these knowledge\nbases, which took experts a decade or more human years to construct, many of\nour projects are constructed by a single graduate student.\n  Our approach to KBC is based on joint probabilistic inference and learning,\nbut we do not see inference as either a panacea or a magic bullet: inference is\na tool that allows us to be systematic in how we construct, debug, and improve\nthe quality of such systems. In addition, inference allows us to construct\nthese systems in a more loosely coupled way than traditional approaches. To\nsupport this idea, we have built the DeepDive system, which has the design goal\nof letting the user ""think about features---not algorithms."" We think of\nDeepDive as declarative in that one specifies what they want but not how to get\nit. We describe our approach with a focus on feature engineering, which we\nargue is an understudied problem relative to its importance to end-to-end\nquality.', 'We introduce an NLP toolkit based on object-oriented knowledge base and\nmulti-level grammar base. This toolkit focuses on semantic parsing, it also has\nabilities to discover new knowledge and grammar automatically, new discovered\nknowledge and grammar will be identified by human, and will be used to update\nthe knowledge base and grammar base. This process can be iterated many times to\nimprove the toolkit continuously.', 'We propose a novel model to address the task of Visual Dialog which exhibits\ncomplex dialog structures. To obtain a reasonable answer based on the current\nquestion and the dialog history, the underlying semantic dependencies between\ndialog entities are essential. In this paper, we explicitly formalize this task\nas inference in a graphical model with partially observed nodes and unknown\ngraph structures (relations in dialog). The given dialog entities are viewed as\nthe observed nodes. The answer to a given question is represented by a node\nwith missing value. We first introduce an Expectation Maximization algorithm to\ninfer both the underlying dialog structures and the missing node values\n(desired answers). Based on this, we proceed to propose a differentiable graph\nneural network (GNN) solution that approximates this process. Experiment\nresults on the VisDial and VisDial-Q datasets show that our model outperforms\ncomparative methods. It is also observed that our method can infer the\nunderlying dialog structure for better dialog reasoning.', 'Extracting structured knowledge from texts has traditionally been used for\nknowledge base generation. However, other sources of information, such as\nimages can be leveraged into this process to build more complete and richer\nknowledge bases. Structured semantic representation of the content of an image\nand knowledge graph embeddings can provide a unique representation of semantic\nrelationships between image entities. Linking known entities in knowledge\ngraphs and learning open-world images using language models has attracted lots\nof interest over the years. In this paper, we propose a shared learning\napproach to learn semantic attributes of images by combining a knowledge graph\nembedding model with the recognized attributes of images. The proposed model\npremises to help us understand the semantic relationship between the entities\nof an image and implicitly provide a link for the extracted entities through a\nknowledge graph embedding model. Under the limitation of using a custom\nuser-defined knowledge base with limited data, the proposed model presents\nsignificant accuracy and provides a new alternative to the earlier approaches.\nThe proposed approach is a step towards bridging the gap between frameworks\nwhich learn from large amounts of data and frameworks which use a limited set\nof predicates to infer new knowledge.', 'One of the most difficult problems in the development of intelligent systems\nis the construction of the underlying knowledge base. As a consequence, the\nrate of progress in the development of this type of system is directly related\nto the speed with which knowledge bases can be assembled, and on its quality.\nWe attempt to solve the knowledge acquisition problem, for a Business\nInformation System, developing a supervised multistrategy learning paradigm.\nThis paradigm is centred on a collaborative data mining strategy, where groups\nof experts collaborate using data-mining process on the supervised acquisition\nof new knowledge extracted from heterogeneous machine learning data models.\n  The Actias system is our approach to this paradigm. It is the result of\napplying the graphic logic based language of sketches to knowledge integration.\nThe system is a data mining collaborative workplace, where the Information\nSystem knowledge base is an algebraic structure. It results from the\nintegration of background knowledge with new insights extracted from data\nmodels, generated for specific data modelling tasks, and represented as rules\nusing the sketches language.', ""With the development of community based question answering (Q&A) services, a\nlarge scale of Q&A archives have been accumulated and are an important\ninformation and knowledge resource on the web. Question and answer matching has\nbeen attached much importance to for its ability to reuse knowledge stored in\nthese systems: it can be useful in enhancing user experience with recurrent\nquestions. In this paper, we try to improve the matching accuracy by overcoming\nthe lexical gap between question and answer pairs. A Word Embedding based\nCorrelation (WEC) model is proposed by integrating advantages of both the\ntranslation model and word embedding, given a random pair of words, WEC can\nscore their co-occurrence probability in Q&A pairs and it can also leverage the\ncontinuity and smoothness of continuous space word representation to deal with\nnew pairs of words that are rare in the training parallel text. An experimental\nstudy on Yahoo! Answers dataset and Baidu Zhidao dataset shows this new\nmethod's promising potential."", ""With the rapid development of knowledge base,question answering based on\nknowledge base has been a hot research issue. In this paper, we focus on\nanswering singlerelation factoid questions based on knowledge base. We build a\nquestion answering system and study the effect of context information on fact\nselection, such as entity's notable type,outdegree. Experimental results show\nthat context information can improve the result of simple question answering."", 'The knowledge engineering bottleneck is still a major challenge in\nconfigurator projects. In this paper we show how recommender systems can\nsupport knowledge base development and maintenance processes. We discuss a\ncouple of scenarios for the application of recommender systems in knowledge\nengineering and report the results of empirical studies which show the\nimportance of user-centered configuration knowledge organization.', 'In many scenarios, complete and incomplete information coexist. For this\nreason, the knowledge representation and database communities have long shown\ninterest in simultaneously supporting the closed- and the open-world views when\nreasoning about logic theories. Here we consider the setting of querying\npossibly incomplete data using logic theories, formalized as the evaluation of\nan ontology-mediated query (OMQ) that pairs a query with a theory, sometimes\ncalled an ontology, expressing background knowledge. This can be further\nenriched by specifying a set of closed predicates from the theory that are to\nbe interpreted under the closed-world assumption, while the rest are\ninterpreted with the open-world view. In this way we can retrieve more precise\nanswers to queries by leveraging the partial completeness of the data.\n  The central goal of this paper is to understand the relative expressiveness\nof OMQ languages in which the ontology is written in the expressive Description\nLogic (DL) ALCHOI and includes a set of closed predicates. We consider a\nrestricted class of conjunctive queries. Our main result is to show that every\nquery in this non-monotonic query language can be translated in polynomial time\ninto Datalog with negation under the stable model semantics. To overcome the\nchallenge that Datalog has no direct means to express the existential\nquantification present in ALCHOI, we define a two-player game that\ncharacterizes the satisfaction of the ontology, and design a Datalog query that\ncan decide the existence of a winning strategy for the game. If there are no\nclosed predicates, that is in the case of querying a plain ALCHOI knowledge\nbase, our translation yields a positive disjunctive Datalog program of\npolynomial size. To the best of our knowledge, unlike previous translations for\nrelated fragments with expressive (non-Horn) DLs, these are the first\npolynomial time translations.', 'The dynamics of belief and knowledge is one of the major components of any\nautonomous system that should be able to incorporate new pieces of information.\nIn order to apply the rationality result of belief dynamics theory to various\npractical problems, it should be generalized in two respects: first it should\nallow a certain part of belief to be declared as immutable; and second, the\nbelief state need not be deductively closed. Such a generalization of belief\ndynamics, referred to as base dynamics, is presented in this paper, along with\nthe concept of a generalized revision algorithm for knowledge bases (Horn or\nHorn logic with stratified negation). We show that knowledge base dynamics has\nan interesting connection with kernel change via hitting set and abduction. In\nthis paper, we show how techniques from disjunctive logic programming can be\nused for efficient (deductive) database updates. The key idea is to transform\nthe given database together with the update request into a disjunctive\n(datalog) logic program and apply disjunctive techniques (such as minimal model\nreasoning) to solve the original update problem. The approach extends and\nintegrates standard techniques for efficient query answering and integrity\nchecking. The generation of a hitting set is carried out through a hyper\ntableaux calculus and magic set that is focused on the goal of minimality.', 'The aim of this research is to develop a reasoning under uncertainty strategy\nin the context of the Fuzzy Inductive Reasoning (FIR) methodology. FIR emerged\nfrom the General Systems Problem Solving developed by G. Klir. It is a data\ndriven methodology based on systems behavior rather than on structural\nknowledge. It is a very useful tool for both the modeling and the prediction of\nthose systems for which no previous structural knowledge is available. FIR\nreasoning is based on pattern rules synthesized from the available data. The\nsize of the pattern rule base can be very large making the prediction process\nquite difficult. In order to reduce the size of the pattern rule base, it is\npossible to automatically extract classical Sugeno fuzzy rules starting from\nthe set of pattern rules. The Sugeno rule base preserves pattern rules\nknowledge as much as possible. In this process some information is lost but\nrobustness is considerably increased. In the forecasting process either the\npattern rule base or the Sugeno fuzzy rule base can be used. The first option\nis desirable when the computational resources make it possible to deal with the\noverall pattern rule base or when the extracted fuzzy rules are not accurate\nenough due to uncertainty associated to the original data. In the second\noption, the prediction process is done by means of the classical Sugeno\ninference system. If the amount of uncertainty associated to the data is small,\nthe predictions obtained using the Sugeno fuzzy rule base will be very\naccurate. In this paper a mixed pattern/fuzzy rules strategy is proposed to\ndeal with uncertainty in such a way that the best of both perspectives is used.\nAreas in the data space with a higher level of uncertainty are identified by\nmeans of the so-called error models. The prediction process in these areas\nmakes use of a mixed pattern/fuzzy rules scheme, whereas areas identified with\na lower level of uncertainty only use the Sugeno fuzzy rule base. The proposed\nstrategy is applied to a real biomedical system, i.e., the central nervous\nsystem control of the cardiovascular system.', 'In this paper, we investigate knowledge reasoning within a simple framework\ncalled knowledge structure. We use variable forgetting as a basic operation for\none agent to reason about its own or other agents\\ knowledge. In our framework,\ntwo notions namely agents\\ observable variables and the weakest sufficient\ncondition play important roles in knowledge reasoning. Given a background\nknowledge base and a set of observable variables for each agent, we show that\nthe notion of an agent knowing a formula can be defined as a weakest sufficient\ncondition of the formula under background knowledge base. Moreover, we show how\nto capture the notion of common knowledge by using a generalized notion of\nweakest sufficient condition. Also, we show that public announcement operator\ncan be conveniently dealt with via our notion of knowledge structure. Further,\nwe explore the computational complexity of the problem whether an epistemic\nformula is realized in a knowledge structure. In the general case, this problem\nis PSPACE-hard; however, for some interesting subcases, it can be reduced to\nco-NP. Finally, we discuss possible applications of our framework in some\ninteresting domains such as the automated analysis of the well-known muddy\nchildren puzzle and the verification of the revised Needham-Schroeder protocol.\nWe believe that there are many scenarios where the natural presentation of the\navailable information about knowledge is under the form of a knowledge\nstructure. What makes it valuable compared with the corresponding multi-agent\nS5 Kripke structure is that it can be much more succinct.', 'We present a method called TIMEMACHINE to generate a timeline of events and\nrelations for entities in a knowledge base. For example for an actor, such a\ntimeline should show the most important professional and personal milestones\nand relationships such as works, awards, collaborations, and family\nrelationships. We develop three orthogonal timeline quality criteria that an\nideal timeline should satisfy: (1) it shows events that are relevant to the\nentity; (2) it shows events that are temporally diverse, so they distribute\nalong the time axis, avoiding visual crowding and allowing for easy user\ninteraction, such as zooming in and out; and (3) it shows events that are\ncontent diverse, so they contain many different types of events (e.g., for an\nactor, it should show movies and marriages and awards, not just movies). We\npresent an algorithm to generate such timelines for a given time period and\nscreen size, based on submodular optimization and web-co-occurrence statistics\nwith provable performance guarantees. A series of user studies using Mechanical\nTurk shows that all three quality criteria are crucial to produce quality\ntimelines and that our algorithm significantly outperforms various baseline and\nstate-of-the-art methods.', 'Ontology-Mediated Query Answering (OMQA) is a well-established framework to\nanswer queries over an RDFS or OWL Knowledge Base (KB). OMQA was originally\ndesigned for unions of conjunctive queries (UCQs), and based on certain\nanswers. More recently, OMQA has been extended to SPARQL queries, but to our\nknowledge, none of the efforts made in this direction (either in the\nliterature, or the so-called SPARQL entailment regimes) is able to capture both\ncertain answers for UCQs and the standard interpretation of SPARQL over a plain\ngraph. We formalize these as requirements to be met by any semantics aiming at\nconciliating certain answers and SPARQL answers, and define three additional\nrequirements, which generalize to KBs some basic properties of SPARQL answers.\nThen we show that a semantics can be defined that satisfies all requirements\nfor SPARQL queries with SELECT, UNION, and OPTIONAL, and for DLs with the\ncanonical model property. We also investigate combined complexity for query\nanswering under such a semantics over DL-Lite R KBs. In particular, we show for\ndifferent fragments of SPARQL that known upper-bounds for query answering over\na plain graph are matched.', ""Normative expert systems have not become commonplace because they have been\ndifficult to build and use. Over the past decade, however, researchers have\ndeveloped the influence diagram, a graphical representation of a decision\nmaker's beliefs, alternatives, and preferences that serves as the knowledge\nbase of a normative expert system. Most people who have seen the representation\nfind it intuitive and easy to use. Consequently, the influence diagram has\novercome significantly the barriers to constructing normative expert systems.\nNevertheless, building influence diagrams is not practical for extremely large\nand complex domains. In this book, I address the difficulties associated with\nthe construction of the probabilistic portion of an influence diagram, called a\nknowledge map, belief network, or Bayesian network. I introduce two\nrepresentations that facilitate the generation of large knowledge maps. In\nparticular, I introduce the similarity network, a tool for building the network\nstructure of a knowledge map, and the partition, a tool for assessing the\nprobabilities associated with a knowledge map. I then use these representations\nto build Pathfinder, a large normative expert system for the diagnosis of\nlymph-node diseases (the domain contains over 60 diseases and over 100 disease\nfindings). In an early version of the system, I encoded the knowledge of the\nexpert using an erroneous assumption that all disease findings were\nindependent, given each disease. When the expert and I attempted to build a\nmore accurate knowledge map for the domain that would capture the dependencies\namong the disease findings, we failed. Using a similarity network, however, we\nbuilt the knowledge-map structure for the entire domain in approximately 40\nhours. Furthermore, the partition representation reduced the number of\nprobability assessments required by the expert from 75,000 to 14,000."", 'A class or taxonomic hierarchy is often manually constructed, and part of our\nknowledge about the world. In this paper, we propose a novel algorithm for\nautomatically acquiring a class hierarchy from a classifier which is often a\nlarge neural network these days. The information that we need from a classifier\nis its confusion matrix which contains, for each pair of base classes, the\nnumber of errors the classifier makes by mistaking one for another. Our\nalgorithm produces surprisingly good hierarchies for some well-known deep\nneural network models trained on the CIFAR-10 dataset, a neural network model\nfor predicting the native language of a non-native English speaker, a neural\nnetwork model for detecting the language of a written text, and a classifier\nfor identifying music genre. In the literature, such class hierarchies have\nbeen used to provide interpretability to the neural networks. We also discuss\nsome other potential uses of the acquired hierarchies.', 'Globally operating enterprises selling large and complex products and\nservices often have to deal with situations where variability models are\nlocally developed to take into account the requirements of local markets. For\nexample, cars sold on the U.S. market are represented by variability models in\nsome or many aspects different from European ones. In order to support global\nvariability management processes, variability models and the underlying\nknowledge bases often need to be integrated. This is a challenging task since\nan integrated knowledge base should not produce results which are different\nfrom those produced by the individual knowledge bases. In this paper, we\nintroduce an approach to variability model integration that is based on the\nconcepts of contextual modeling and conflict detection. We present the\nunderlying concepts and the results of a corresponding performance analysis.', 'This paper develops the concept of knowledge and its exchange using Semantic\nWeb technologies. It points out that knowledge is more than information because\nit embodies the meaning, that is to say semantic and context. These\ncharacteristics will influence our approach to represent and to treat the\nknowledge. In order to be adopted, the developed system needs to be simple and\nto use standards. The goal of the paper is to find standards to model knowledge\nand exchange it with an other person. Therefore, we propose to model knowledge\nusing UML models to show a graphical representation and to exchange it with XML\nto ensure the portability at low cost. We introduce the concept of ontology for\norganizing knowledge and for facilitating the knowledge exchange. Proposals\nhave been tested by implementing an application on the design knowledge of a\npen.', 'In this paper we introduce the notion of Demand-Weighted Completeness,\nallowing estimation of the completeness of a knowledge base with respect to how\nit is used. Defining an entity by its classes, we employ usage data to predict\nthe distribution over relations for that entity. For example, instances of\nperson in a knowledge base may require a birth date, name and nationality to be\nconsidered complete. These predicted relation distributions enable detection of\nimportant gaps in the knowledge base, and define the required facts for unseen\nentities. Such characterisation of the knowledge base can also quantify how\nusage and completeness change over time. We demonstrate a method to measure\nDemand-Weighted Completeness, and show that a simple neural network model\nperforms well at this prediction task.', 'Objects in the world usually appear in context, participating in spatial\nrelationships and interactions that are predictable and expected. Knowledge of\nthese contexts can be used in the task of using a mobile camera to search for a\nspecified object in a room. We call this the object search task. This paper is\nconcerned with representing this knowledge in a manner facilitating its\napplication to object search while at the same time lending itself to\nautonomous learning by a robot. The ability for the robot to learn such\nknowledge without supervision is crucial due to the vast number of possible\nrelationships that can exist for any given set of objects. Moreover, since a\nrobot will not have an infinite amount of time to learn, it must be able to\ndetermine an order in which to look for possible relationships so as to\nmaximize the rate at which new knowledge is gained. In effect, there must be a\n""focus of interest"" operator that allows the robot to choose which examples are\nlikely to convey the most new information and should be examined first. This\npaper demonstrates how a representation based on statistical confidence\nintervals allows the construction of a system that achieves the above goals. An\nalgorithm, based on the Highest Impact First heuristic, is presented as a means\nfor providing a ""focus of interest"" with which to control the learning process,\nand examples are given.', 'This paper describes a process for constructing situation-specific belief\nnetworks from a knowledge base of network fragments. A situation-specific\nnetwork is a minimal query complete network constructed from a knowledge base\nin response to a query for the probability distribution on a set of target\nvariables given evidence and context variables. We present definitions of query\ncompleteness and situation-specific networks. We describe conditions on the\nknowledge base that guarantee query completeness. The relationship of our work\nto earlier work on KBMC is also discussed.', 'Ontologies provide a formal description of concepts and their relationships\nin a knowledge domain. The goal of ontology alignment is to identify\nsemantically matching concepts and relationships across independently developed\nontologies that purport to describe the same knowledge. In order to handle the\nwidest possible class of ontologies, many alignment algorithms rely on\nterminological and structural meth- ods, but the often fuzzy nature of concepts\ncomplicates the matching process. However, one area that should provide clear\nmatching solutions due to its mathematical nature, is units of measurement.\nSeveral on- tologies for units of measurement are available, but there has been\nno attempt to align them, notwithstanding the obvious importance for tech-\nnical interoperability. We propose a general strategy to map these (and\nsimilar) ontologies by introducing MathML to accurately capture the semantic\ndescription of concepts specified therein. We provide mapping results for three\nontologies, and show that our approach improves on lexical comparisons.', 'Existential rules, long known as tuple-generating dependencies in database\ntheory, have been intensively studied in the last decade as a powerful\nformalism to represent ontological knowledge in the context of ontology-based\nquery answering. A knowledge base is then composed of an instance that contains\nincomplete data and a set of existential rules, and answers to queries are\nlogically entailed from the knowledge base. This brought again to light the\nfundamental chase tool, and its different variants that have been proposed in\nthe literature. It is well-known that the problem of determining, given a chase\nvariant and a set of existential rules, whether the chase will halt on any\ninstance, is undecidable. Hence, a crucial issue is whether it becomes\ndecidable for known subclasses of existential rules. In this work, we consider\nlinear existential rules, a simple yet important subclass of existential rules\nthat generalizes inclusion dependencies. We show the decidability of the all\ninstance chase termination problem on linear rules for three main chase\nvariants, namely semi-oblivious, restricted and core chase. To obtain these\nresults, we introduce a novel approach based on so-called derivation trees and\na single notion of forbidden pattern. Besides the theoretical interest of a\nunified approach and new proofs, we provide the first positive decidability\nresults concerning the termination of the restricted chase, proving that chase\ntermination on linear existential rules is decidable for both versions of the\nproblem: Does every fair chase sequence terminate? Does some fair chase\nsequence terminate?', 'Modern knowledge base systems frequently need to combine a collection of\ndatabases in different formats: e.g., relational databases, XML databases, rule\nbases, ontologies, etc. In the deductive database system DDBASE, we can manage\nthese different formats of knowledge and reason about them. Even the file\nsystems on different computers can be part of the knowledge base. Often, it is\nnecessary to handle different versions of a knowledge base. E.g., we might want\nto find out common parts or differences of two versions of a relational\ndatabase.\n  We will examine the use of abstractions of rule bases by predicate dependency\nand rule predicate graphs. Also the proof trees of derived atoms can help to\ncompare different versions of a rule base. Moreover, it might be possible to\nhave derivations joining rules with other formalisms of knowledge\nrepresentation.\n  Ontologies have shown their benefits in many applications of intelligent\nsystems, and there have been many proposals for rule languages compatible with\nthe semantic web stack, e.g., SWRL, the semantic web rule language. Recently,\nontologies are used in hybrid systems for specifying the provenance of the\ndifferent components.', 'Link prediction in large knowledge graphs has received a lot of attention\nrecently because of its importance for inferring missing relations and for\ncompleting and improving noisily extracted knowledge graphs. Over the years a\nnumber of machine learning researchers have presented various models for\npredicting the presence of missing relations in a knowledge base. Although all\nthe previous methods are presented with empirical results that show high\nperformance on select datasets, there is almost no previous work on\nunderstanding the connection between properties of a knowledge base and the\nperformance of a model. In this paper we analyze the RESCAL method and prove\nthat it can not encode asymmetric transitive relations in knowledge bases.', 'Database schema elements such as tables, views, triggers and functions are\ntypically defined with many interrelationships. In order to support database\nusers in understanding a given schema, a rule-based approach for analyzing the\nrespective dependencies is proposed using Datalog expressions. We show that\nmany interesting properties of schema elements can be systematically determined\nthis way. The expressiveness of the proposed analysis is exemplarily shown with\nthe problem of computing induced functional dependencies for derived relations.\nThe propagation of functional dependencies plays an important role in data\nintegration and query optimization but represents an undecidable problem in\ngeneral. And yet, our rule-based analysis covers all relational operators as\nwell as linear recursive expressions in a systematic way showing the depth of\nanalysis possible by our proposal. The analysis of functional dependencies is\nwell-integrated in a uniform approach to analyzing dependencies between schema\nelements in general.', 'Formal verification provides strong safety guarantees but only for models of\ncyber-physical systems. Hybrid system models describe the required interplay of\ncomputation and physical dynamics, which is crucial to guarantee what\ncomputations lead to safe physical behavior (e.g., cars should not collide).\nControl computations that affect physical dynamics must act in advance to avoid\npossibly unsafe future circumstances. Formal verification then ensures that the\ncontrollers correctly identify and provably avoid unsafe future situations\nunder a certain model of physics. But any model of physics necessarily deviates\nfrom reality and, moreover, any observation with real sensors and manipulation\nwith real actuators is subject to uncertainty. This makes runtime validation a\ncrucial step to monitor whether the model assumptions hold for the real system\nimplementation.\n  The key question is what property needs to be runtime-monitored and what a\nsatisfied runtime monitor entails about the safety of the system: the\nobservations of a runtime monitor only relate back to the safety of the system\nif they are themselves accompanied by a proof of correctness! For an unbroken\nchain of correctness guarantees, we, thus, synthesize runtime monitors in a\nprovably correct way from provably safe hybrid system models. This paper\naddresses the inevitable challenge of making the synthesized monitoring\nconditions robust to partial observability of sensor uncertainty and partial\ncontrollability due to actuator disturbance. We show that the monitoring\nconditions result in provable safety guarantees with fallback controllers that\nreact to monitor violation at runtime.', 'The promise of ultimate elasticity and operational simplicity of serverless\ncomputing has recently lead to an explosion of research in this area. In the\ncontext of data analytics, the concept sounds appealing, but due to the\nlimitations of current offerings, there is no consensus yet on whether or not\nthis approach is technically and economically viable. In this paper, we\nidentify interactive data analytics on cold data as a use case where serverless\ncomputing excels. We design and implement Lambada, a system following a purely\nserverless architecture, in order to illustrate when and how serverless\ncomputing should be employed for data analytics. We propose several system\ncomponents that overcome the previously known limitations inherent in the\nserverless paradigm as well as additional ones we identify in this work. We can\nshow that, thanks to careful design, a serverless query processing system can\nbe at the same time one order of magnitude faster and two orders of magnitude\ncheaper compared to commercial Query-as-a-Service systems, the only alternative\nwith similar operational simplicity.', 'Clustering is inherently ill-posed: there often exist multiple valid\nclusterings of a single dataset, and without any additional information a\nclustering system has no way of knowing which clustering it should produce.\nThis motivates the use of constraints in clustering, as they allow users to\ncommunicate their interests to the clustering system. Active constraint-based\nclustering algorithms select the most useful constraints to query, aiming to\nproduce a good clustering using as few constraints as possible. We propose\nCOBRA, an active method that first over-clusters the data by running K-means\nwith a $K$ that is intended to be too large, and subsequently merges the\nresulting small clusters into larger ones based on pairwise constraints. In its\nmerging step, COBRA is able to keep the number of pairwise queries low by\nmaximally exploiting constraint transitivity and entailment. We experimentally\nshow that COBRA outperforms the state of the art in terms of clustering quality\nand runtime, without requiring the number of clusters in advance.', 'The dynamics of belief and knowledge is one of the major components of any\nautonomous system that should be able to incorporate new pieces of information.\nIn this paper, we argue that to apply rationality result of belief dynamics\ntheory to various practical problems, it should be generalized in two respects:\nfirst of all, it should allow a certain part of belief to be declared as\nimmutable; and second, the belief state need not be deductively closed. Such a\ngeneralization of belief dynamics, referred to as base dynamics, is presented,\nalong with the concept of a generalized revision algorithm for Horn knowledge\nbases. We show that Horn knowledge base dynamics has interesting connection\nwith kernel change and abduction. Finally, we also show that both variants are\nrational in the sense that they satisfy certain rationality postulates stemming\nfrom philosophical works on belief dynamics.', 'We introduce Network Augmentation (NetAug), a new training method for\nimproving the performance of tiny neural networks. Existing regularization\ntechniques (e.g., data augmentation, dropout) have shown much success on large\nneural networks by adding noise to overcome over-fitting. However, we found\nthese techniques hurt the performance of tiny neural networks. We argue that\ntraining tiny models are different from large models: rather than augmenting\nthe data, we should augment the model, since tiny models tend to suffer from\nunder-fitting rather than over-fitting due to limited capacity. To alleviate\nthis issue, NetAug augments the network (reverse dropout) instead of inserting\nnoise into the dataset or the network. It puts the tiny model into larger\nmodels and encourages it to work as a sub-model of larger models to get extra\nsupervision, in addition to functioning as an independent model. At test time,\nonly the tiny model is used for inference, incurring zero inference overhead.\nWe demonstrate the effectiveness of NetAug on image classification and object\ndetection. NetAug consistently improves the performance of tiny models,\nachieving up to 2.2% accuracy improvement on ImageNet. On object detection,\nachieving the same level of performance, NetAug requires 41% fewer MACs on\nPascal VOC and 38% fewer MACs on COCO than the baseline.', 'Question answering and conversational systems are often baffled and need help\nclarifying certain ambiguities. However, limitations of existing datasets\nhinder the development of large-scale models capable of generating and\nutilising clarification questions. In order to overcome these limitations, we\ndevise a novel bootstrapping framework (based on self-supervision) that assists\nin the creation of a diverse, large-scale dataset of clarification questions\nbased on post-comment tuples extracted from stackexchange. The framework\nutilises a neural network based architecture for classifying clarification\nquestions. It is a two-step method where the first aims to increase the\nprecision of the classifier and second aims to increase its recall. We\nquantitatively demonstrate the utility of the newly created dataset by applying\nit to the downstream task of question-answering. The final dataset, ClarQ,\nconsists of ~2M examples distributed across 173 domains of stackexchange. We\nrelease this dataset in order to foster research into the field of\nclarification question generation with the larger goal of enhancing dialog and\nquestion answering systems.', 'In this paper we tackle the problem of routing multiple agents in a\ncoordinated manner. This is a complex problem that has a wide range of\napplications in fleet management to achieve a common goal, such as mapping from\na swarm of robots and ride sharing. Traditional methods are typically not\ndesigned for realistic environments hich contain sparsely connected graphs and\nunknown traffic, and are often too slow in runtime to be practical. In\ncontrast, we propose a graph neural network based model that is able to perform\nmulti-agent routing based on learned value iteration in a sparsely connected\ngraph with dynamically changing traffic conditions. Moreover, our learned\ncommunication module enables the agents to coordinate online and adapt to\nchanges more effectively. We created a simulated environment to mimic realistic\nmapping performed by autonomous vehicles with unknown minimum edge coverage and\ntraffic conditions; our approach significantly outperforms traditional solvers\nboth in terms of total cost and runtime. We also show that our model trained\nwith only two agents on graphs with a maximum of 25 nodes can easily generalize\nto situations with more agents and/or nodes.', 'In this paper we introduce DeepCrawl, a fully-playable Roguelike prototype\nfor iOS and Android in which all agents are controlled by policy networks\ntrained using Deep Reinforcement Learning (DRL). Our aim is to understand\nwhether recent advances in DRL can be used to develop convincing behavioral\nmodels for non-player characters in videogames. We begin with an analysis of\nrequirements that such an AI system should satisfy in order to be practically\napplicable in video game development, and identify the elements of the DRL\nmodel used in the DeepCrawl prototype. The successes and limitations of\nDeepCrawl are documented through a series of playability tests performed on the\nfinal game. We believe that the techniques we propose offer insight into\ninnovative new avenues for the development of behaviors for non-player\ncharacters in video games, as they offer the potential to overcome critical\nissues with', 'In previous work [BGHK92, BGHK93], we have studied the random-worlds approach\n-- a particular (and quite powerful) method for generating degrees of belief\n(i.e., subjective probabilities) from a knowledge base consisting of objective\n(first-order, statistical, and default) information. But allowing a knowledge\nbase to contain only objective information is sometimes limiting. We\noccasionally wish to include information about degrees of belief in the\nknowledge base as well, because there are contexts in which old beliefs\nrepresent important information that should influence new beliefs. In this\npaper, we describe three quite general techniques for extending a method that\ngenerates degrees of belief from objective information to one that can make use\nof degrees of belief as well. All of our techniques are bloused on well-known\napproaches, such as cross-entropy. We discuss general connections between the\ntechniques and in particular show that, although conceptually and technically\nquite different, all of the techniques give the same answer when applied to the\nrandom-worlds method.', ""We present a new knowledge-base of hasPart relationships, extracted from a\nlarge corpus of generic statements. Complementary to other resources available,\nit is the first which is all three of: accurate (90% precision), salient\n(covers relationships a person may mention), and has high coverage of common\nterms (approximated as within a 10 year old's vocabulary), as well as having\nseveral times more hasPart entries than in the popular ontologies ConceptNet\nand WordNet. In addition, it contains information about quantifiers, argument\nmodifiers, and links the entities to appropriate concepts in Wikipedia and\nWordNet. The knowledge base is available at https://allenai.org/data/haspartkb"", 'The goal of few-shot learning is to learn a classifier that can recognize\nunseen classes from limited support data with labels. A common practice for\nthis task is to train a model on the base set first and then transfer to novel\nclasses through fine-tuning (Here fine-tuning procedure is defined as\ntransferring knowledge from base to novel data, i.e. learning to transfer in\nfew-shot scenario.) or meta-learning. However, as the base classes have no\noverlap to the novel set, simply transferring whole knowledge from base data is\nnot an optimal solution since some knowledge in the base model may be biased or\neven harmful to the novel class. In this paper, we propose to transfer partial\nknowledge by freezing or fine-tuning particular layer(s) in the base model.\nSpecifically, layers will be imposed different learning rates if they are\nchosen to be fine-tuned, to control the extent of preserved transferability. To\ndetermine which layers to be recast and what values of learning rates for them,\nwe introduce an evolutionary search based method that is efficient to\nsimultaneously locate the target layers and determine their individual learning\nrates. We conduct extensive experiments on CUB and mini-ImageNet to demonstrate\nthe effectiveness of our proposed method. It achieves the state-of-the-art\nperformance on both meta-learning and non-meta based frameworks. Furthermore,\nwe extend our method to the conventional pre-training + fine-tuning paradigm\nand obtain consistent improvement.', ""A Concept Tree is a structure for storing knowledge where the trees are\nstored in a database called a Concept Base. It sits between the highly\ndistributed neural architectures and the distributed information systems, with\nthe intention of bringing brain-like and computer systems closer together.\nConcept Trees can grow from the semi-structured sources when consistent\nsequences of concepts are presented. Each tree ideally represents a single\ncohesive concept and the trees can link with each other for navigation and\nsemantic purposes. The trees are therefore also a type of semantic network and\nwould benefit from having a consistent level of context for each node. A\nconsistent build process is managed through a 'counting rule' and some other\nrules that can normalise the database structure. This restricted structure can\nthen be complimented and enriched by the more dynamic context. It is also\nsuggested to use the linking structure of the licas system [15] as a basis for\nthe context links, where the mathematical model is extended further to define\nthis. A number of tests have demonstrated the soundness of the architecture.\nBuilding the trees from text documents shows that the tree structure could be\ninherent in natural language. Then, two types of query language are described.\nBoth of these can perform consistent query processes to return knowledge to the\nuser and even enhance the query with new knowledge. This is supported even\nfurther with direct comparisons to a cognitive model, also being developed by\nthe author.""]"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : When modeling complex probability distributions of variables with inherent dependencies, I utilize a graph to represent these dependencies and probabilistic models to depict the variables' distributions. However, these models can be large and computationally expensive during inference. Therefore, I am seeking a fast, approximate inference algorithm that can access a highly effective proposal distribution for importance sampling. My current strategy involves identifying non-essential variables and marginalizing them during sampling to enhance the speed of inference. Additionally, I am considering transforming the graphical model into a more efficient format, such as a decision diagram, to facilitate easier interpretation during sampling (inference). To evaluate the effectiveness of my approach, I need to test it against standard benchmarks. Therefore, I would appreciate information on the standard benchmarks applicable to this task.","['Sampling is a popular method for approximate inference when exact inference\nis impractical. Generally, sampling algorithms do not exploit context-specific\nindependence (CSI) properties of probability distributions. We introduce\ncontext-specific likelihood weighting (CS-LW), a new sampling methodology,\nwhich besides exploiting the classical conditional independence properties,\nalso exploits CSI properties. Unlike the standard likelihood weighting, CS-LW\nis based on partial assignments of random variables and requires fewer samples\nfor convergence due to the sampling variance reduction. Furthermore, the speed\nof generating samples increases. Our novel notion of contextual assignments\ntheoretically justifies CS-LW. We empirically show that CS-LW is competitive\nwith state-of-the-art algorithms for approximate inference in the presence of a\nsignificant amount of CSIs.', 'Many real life domains contain a mixture of discrete and continuous variables\nand can be modeled as hybrid Bayesian Networks. Animportant subclass of hybrid\nBNs are conditional linear Gaussian (CLG) networks, where the conditional\ndistribution of the continuous variables given an assignment to the discrete\nvariables is a multivariate Gaussian. Lauritzen\'s extension to the clique tree\nalgorithm can be used for exact inference in CLG networks. However, many\ndomains also include discrete variables that depend on continuous ones, and CLG\nnetworks do not allow such dependencies to berepresented. No exact inference\nalgorithm has been proposed for these enhanced CLG networks. In this paper, we\ngeneralize Lauritzen\'s algorithm, providing the first ""exact"" inference\nalgorithm for augmented CLG networks - networks where continuous nodes are\nconditional linear Gaussians but that also allow discrete children ofcontinuous\nparents. Our algorithm is exact in the sense that it computes the exact\ndistributions over the discrete nodes, and the exact first and second moments\nof the continuous ones, up to the accuracy obtained by numerical integration\nused within thealgorithm. When the discrete children are modeled with softmax\nCPDs (as is the case in many real world domains) the approximation of the\ncontinuous distributions using the first two moments is particularly accurate.\nOur algorithm is simple to implement and often comparable in its complexity to\nLauritzen\'s algorithm. We show empirically that it achieves substantially\nhigher accuracy than previous approximate algorithms.', 'Probabilistic graphical models are powerful tools which allow us to formalise\nour knowledge about the world and reason about its inherent uncertainty. There\nexist a considerable number of methods for performing inference in\nprobabilistic graphical models; however, they can be computationally costly due\nto significant time burden and/or storage requirements; or they lack\ntheoretical guarantees of convergence and accuracy when applied to large scale\ngraphical models. To this end, we propose the Universal Marginaliser Importance\nSampler (UM-IS) -- a hybrid inference scheme that combines the flexibility of a\ndeep neural network trained on samples from the model and inherits the\nasymptotic guarantees of importance sampling. We show how combining samples\ndrawn from the graphical model with an appropriate masking function allows us\nto train a single neural network to approximate any of the corresponding\nconditional marginal distributions, and thus amortise the cost of inference. We\nalso show that the graph embeddings can be applied for tasks such as:\nclustering, classification and interpretation of relationships between the\nnodes. Finally, we benchmark the method on a large graph (>1000 nodes), showing\nthat UM-IS outperforms sampling-based methods by a large margin while being\ncomputationally efficient.', ""Probabilistic graphical models offer a powerful framework to account for the\ndependence structure between variables, which is represented as a graph.\nHowever, the dependence between variables may render inference tasks\nintractable. In this paper we review techniques exploiting the graph structure\nfor exact inference, borrowed from optimisation and computer science. They are\nbuilt on the principle of variable elimination whose complexity is dictated in\nan intricate way by the order in which variables are eliminated. The so-called\ntreewidth of the graph characterises this algorithmic complexity: low-treewidth\ngraphs can be processed efficiently. The first message that we illustrate is\ntherefore the idea that for inference in graphical model, the number of\nvariables is not the limiting factor, and it is worth checking for the\ntreewidth before turning to approximate methods. We show how algorithms\nproviding an upper bound of the treewidth can be exploited to derive a 'good'\nelimination order enabling to perform exact inference. The second message is\nthat when the treewidth is too large, algorithms for approximate inference\nlinked to the principle of variable elimination, such as loopy belief\npropagation and variational approaches, can lead to accurate results while\nbeing much less time consuming than Monte-Carlo approaches. We illustrate the\ntechniques reviewed in this article on benchmarks of inference problems in\ngenetic linkage analysis and computer vision, as well as on hidden variables\nrestoration in coupled Hidden Markov Models."", ""Dependency networks (Heckerman et al., 2000) are potential probabilistic\ngraphical models for systems comprising a large number of variables. Like\nBayesian networks, the structure of a dependency network is represented by a\ndirected graph, and each node has a conditional probability table. Learning and\ninference are realized locally on individual nodes; therefore, computation\nremains tractable even with a large number of variables. However, the\ndependency network's learned distribution is the stationary distribution of a\nMarkov chain called pseudo-Gibbs sampling and has no closed-form expressions.\nThis technical disadvantage has impeded the development of dependency networks.\nIn this paper, we consider a certain manifold for each node. Then, we can\ninterpret pseudo-Gibbs sampling as iterative m-projections onto these\nmanifolds. This interpretation provides a theoretical bound for the location\nwhere the stationary distribution of pseudo-Gibbs sampling exists in\ndistribution space. Furthermore, this interpretation involves structure and\nparameter learning algorithms as optimization problems. In addition, we compare\ndependency and Bayesian networks experimentally. The results demonstrate that\nthe dependency network and the Bayesian network have roughly the same\nperformance in terms of the accuracy of their learned distributions. The\nresults also show that the dependency network can learn much faster than the\nBayesian network."", 'A fundamental computation for statistical inference and accurate\ndecision-making is to compute the marginal probabilities or most probable\nstates of task-relevant variables. Probabilistic graphical models can\nefficiently represent the structure of such complex data, but performing these\ninferences is generally difficult. Message-passing algorithms, such as belief\npropagation, are a natural way to disseminate evidence amongst correlated\nvariables while exploiting the graph structure, but these algorithms can\nstruggle when the conditional dependency graphs contain loops. Here we use\nGraph Neural Networks (GNNs) to learn a message-passing algorithm that solves\nthese inference tasks. We first show that the architecture of GNNs is\nwell-matched to inference tasks. We then demonstrate the efficacy of this\ninference approach by training GNNs on a collection of graphical models and\nshowing that they substantially outperform belief propagation on loopy graphs.\nOur message-passing algorithms generalize out of the training set to larger\ngraphs and graphs with different structure.', 'Probabilistic Logic Programming (PLP) languages enable programmers to specify\nsystems that combine logical models with statistical knowledge. The inference\nproblem, to determine the probability of query answers in PLP, is intractable\nin general, thereby motivating the need for approximate techniques. In this\npaper, we present a technique for approximate inference of conditional\nprobabilities for PLP queries. It is an Adaptive Markov Chain Monte Carlo\n(MCMC) technique, where the distribution from which samples are drawn is\nmodified as the Markov Chain is explored. In particular, the distribution is\nprogressively modified to increase the likelihood that a generated sample is\nconsistent with evidence. In our context, each sample is uniquely characterized\nby the outcomes of a set of random variables. Inspired by reinforcement\nlearning, our technique propagates rewards to random variable/outcome pairs\nused in a sample based on whether the sample was consistent or not. The\ncumulative rewards of each outcome is used to derive a new ""adapted\ndistribution"" for each random variable. For a sequence of samples, the\ndistributions are progressively adapted after each sample. For a query with\n""Markovian evaluation structure"", we show that the adapted distribution of\nsamples converges to the query\'s conditional probability distribution. For\nMarkovian queries, we present a modified adaptation process that can be used in\nadaptive MCMC as well as adaptive independent sampling. We empirically evaluate\nthe effectiveness of the adaptive sampling methods for queries with and without\nMarkovian evaluation structure.', 'Bayesian networks (BNs) are used for inference and sampling by exploiting\nconditional independence among random variables. Context specific independence\n(CSI) is a property of graphical models where additional independence relations\narise in the context of particular values of random variables (RVs).\nIdentifying and exploiting CSI properties can simplify inference. Some\ngenerative network models (models that generate social/information network\nsamples from a network distribution P(G)), with complex interactions among a\nset of RVs, can be represented with probabilistic graphical models, in\nparticular with BNs. In the present work we show one such a case. We discuss\nhow a mixed Kronecker Product Graph Model can be represented as a BN, and study\nits BN properties that can be used for efficient sampling. Specifically, we\nshow that instead of exhibiting CSI properties, the model has deterministic\ncontext-specific dependence (DCSD). Exploiting this property focuses the\nsampling method on a subset of the sampling space that improves efficiency.', 'The paper introduces a generalization for known probabilistic models such as\nlog-linear and graphical models, called here multiplicative models. These\nmodels, that express probabilities via product of parameters are shown to\ncapture multiple forms of contextual independence between variables, including\ndecision graphs and noisy-OR functions. An inference algorithm for\nmultiplicative models is provided and its correctness is proved. The complexity\nanalysis of the inference algorithm uses a more refined parameter than the\ntree-width of the underlying graph, and shows the computational cost does not\nexceed that of the variable elimination algorithm in graphical models. The\npaper ends with examples where using the new models and algorithm is\ncomputationally beneficial.', ""Graphical models use the intuitive and well-studied methods of graph theory\nto implicitly represent dependencies between variables in large systems. They\ncan model the global behaviour of a complex system by specifying only local\nfactors. This thesis studies inference in discrete graphical models from an\nalgebraic perspective and the ways inference can be used to express and\napproximate NP-hard combinatorial problems.\n  We investigate the complexity and reducibility of various inference problems,\nin part by organizing them in an inference hierarchy. We then investigate\ntractable approximations for a subset of these problems using distributive law\nin the form of message passing. The quality of the resulting message passing\nprocedure, called Belief Propagation (BP), depends on the influence of loops in\nthe graphical model. We contribute to three classes of approximations that\nimprove BP for loopy graphs A) loop correction techniques; B) survey\npropagation, another message passing technique that surpasses BP in some\nsettings; and C) hybrid methods that interpolate between deterministic message\npassing and Markov Chain Monte Carlo inference.\n  We then review the existing message passing solutions and provide novel\ngraphical models and inference techniques for combinatorial problems under\nthree broad classes: A) constraint satisfaction problems such as\nsatisfiability, coloring, packing, set / clique-cover and dominating /\nindependent set and their optimization counterparts; B) clustering problems\nsuch as hierarchical clustering, K-median, K-clustering, K-center and\nmodularity optimization; C) problems over permutations including assignment,\ngraph morphisms and alignment, finding symmetries and traveling salesman\nproblem. In many cases we show that message passing is able to find solutions\nthat are either near optimal or favourably compare with today's\nstate-of-the-art approaches."", 'A plethora of problems in AI, engineering and the sciences are naturally\nformalized as inference in discrete probabilistic models. Exact inference is\noften prohibitively expensive, as it may require evaluating the (unnormalized)\ntarget density on its entire domain. Here we consider the setting where only a\nlimited budget of calls to the unnormalized density oracle is available,\nraising the challenge of where in the domain to allocate these function calls\nin order to construct a good approximate solution. We formulate this problem as\nan instance of sequential decision-making under uncertainty and leverage\nmethods from reinforcement learning for probabilistic inference with budget\nconstraints. In particular, we propose the TreeSample algorithm, an adaptation\nof Monte Carlo Tree Search to approximate inference. This algorithm caches all\nprevious queries to the density oracle in an explicit search tree, and\ndynamically allocates new queries based on a ""best-first"" heuristic for\nexploration, using existing upper confidence bound methods. Our non-parametric\ninference method can be effectively combined with neural networks that compile\napproximate conditionals of the target, which are then used to guide the\ninference search and enable generalization across multiple target\ndistributions. We show empirically that TreeSample outperforms standard\napproximate inference methods on synthetic factor graphs.', 'As data sets grow in size, the ability of learning methods to find structure\nin them is increasingly hampered by the time needed to search the large spaces\nof possibilities and generate a score for each that takes all of the observed\ndata into account. For instance, Bayesian networks, the model chosen in this\npaper, have a super-exponentially large search space for a fixed number of\nvariables. One possible method to alleviate this problem is to use a proxy,\nsuch as a Gaussian Process regressor, in place of the true scoring function,\ntraining it on a selection of sampled networks. We prove here that the use of\nsuch a proxy is well-founded, as we can bound the smoothness of a commonly-used\nscoring function for Bayesian network structure learning. We show here that,\ncompared to an identical search strategy using the network?s exact scores, our\nproxy-based search is able to get equivalent or better scores on a number of\ndata sets in a fraction of the time.']","['Machine learning provides algorithms that can learn from data and make\ninferences or predictions on data. Bayesian networks are a class of graphical\nmodels that allow to represent a collection of random variables and their\ncondititional dependencies by directed acyclic graphs. In this paper, an\ninference algorithm for the hidden random variables of a Bayesian network is\ngiven by using the tropicalization of the marginal distribution of the observed\nvariables. By restricting the topological structure to graded networks, an\ninference algorithm for graded Bayesian networks will be established that\nevaluates the hidden random variables rank by rank and in this way yields the\nmost probable states of the hidden variables. This algorithm can be viewed as a\ngeneralized version of the Viterbi algorithm for graded Bayesian networks.', 'Monte Carlo inference has asymptotic guarantees, but can be slow when using\ngeneric proposals. Handcrafted proposals that rely on user knowledge about the\nposterior distribution can be efficient, but are difficult to derive and\nimplement. This paper proposes to let users express their posterior knowledge\nin the form of proposal programs, which are samplers written in probabilistic\nprogramming languages. One strategy for writing good proposal programs is to\ncombine domain-specific heuristic algorithms with neural network models. The\nheuristics identify high probability regions, and the neural networks model the\nposterior uncertainty around the outputs of the algorithm. Proposal programs\ncan be used as proposal distributions in importance sampling and\nMetropolis-Hastings samplers without sacrificing asymptotic consistency, and\ncan be optimized offline using inference compilation. Support for optimizing\nand using proposal programs is easily implemented in a sampling-based\nprobabilistic programming runtime. The paper illustrates the proposed technique\nwith a proposal program that combines RANSAC and neural networks to accelerate\ninference in a Bayesian linear regression with outliers model.', 'We show an alternative way of representing a Bayesian belief network by\nsensitivities and probability distributions. This representation is equivalent\nto the traditional representation by conditional probabilities, but makes\ndependencies between nodes apparent and intuitively easy to understand. We also\npropose a QR matrix representation for the sensitivities and/or conditional\nprobabilities which is more efficient, in both memory requirements and\ncomputational speed, than the traditional representation for computer-based\nimplementations of probabilistic inference. We use sensitivities to show that\nfor a certain class of binary networks, the computation time for approximate\nprobabilistic inference with any positive upper bound on the error of the\nresult is independent of the size of the network. Finally, as an alternative to\ntraditional algorithms that use conditional probabilities, we describe an exact\nalgorithm for probabilistic inference that uses the QR-representation for\nsensitivities and updates probability distributions of nodes in a network\naccording to messages from the neighbors.', 'We introduce a method for using deep neural networks to amortize the cost of\ninference in models from the family induced by universal probabilistic\nprogramming languages, establishing a framework that combines the strengths of\nprobabilistic programming and deep learning methods. We call what we do\n""compilation of inference"" because our method transforms a denotational\nspecification of an inference problem in the form of a probabilistic program\nwritten in a universal programming language into a trained neural network\ndenoted in a neural network specification language. When at test time this\nneural network is fed observational data and executed, it performs approximate\ninference in the original model specified by the probabilistic program. Our\ntraining objective and learning procedure are designed to allow the trained\nneural network to be used as a proposal distribution in a sequential importance\nsampling inference engine. We illustrate our method on mixture models and\nCaptcha solving and show significant speedups in the efficiency of inference.', ""Approximate inference in probability models is a fundamental task in machine\nlearning. Approximate inference provides powerful tools to Bayesian reasoning,\ndecision making, and Bayesian deep learning. The main goal is to estimate the\nexpectation of interested functions w.r.t. a target distribution. When it comes\nto high dimensional probability models and large datasets, efficient\napproximate inference becomes critically important. In this thesis, we propose\na new framework for approximate inference, which combines the advantages of\nthese three frameworks and overcomes their limitations. Our proposed four\nalgorithms are motivated by the recent computational progress of Stein's\nmethod. Our proposed algorithms are applied to continuous and discrete\ndistributions under the setting when the gradient information of the target\ndistribution is available or unavailable. Theoretical analysis is provided to\nprove the convergence of our proposed algorithms. Our adaptive IS algorithm\niteratively improves the importance proposal by functionally decreasing the KL\ndivergence between the updated proposal and the target. When the gradient of\nthe target is unavailable, our proposed sampling algorithm leverages the\ngradient of a surrogate model and corrects induced bias with importance\nweights, which significantly outperforms other gradient-free sampling\nalgorithms. In addition, our theoretical results enable us to perform the\ngoodness-of-fit test on discrete distributions. At the end of the thesis, we\npropose an importance-weighted method to efficiently aggregate local models in\ndistributed learning with one-shot communication. Results on simulated and real\ndatasets indicate the statistical efficiency and wide applicability of our\nalgorithm."", 'Graphical models represent multivariate and generally not normalized\nprobability distributions. Computing the normalization factor, called the\npartition function, is the main inference challenge relevant to multiple\nstatistical and optimization applications. The problem is of an exponential\ncomplexity with respect to the number of variables. In this manuscript, aimed\nat approximating the PF, we consider Multi-Graph Models where binary variables\nand multivariable factors are associated with edges and nodes, respectively, of\nan undirected multi-graph. We suggest a new methodology for analysis and\ncomputations that combines the Gauge Function technique with the technique from\nthe field of real stable polynomials. We show that the Gauge Function has a\nnatural polynomial representation in terms of gauges/variables associated with\nedges of the multi-graph. Moreover, it can be used to recover the Partition\nFunction through a sequence of transformations allowing appealing algebraic and\ngraphical interpretations. Algebraically, one step in the sequence consists in\napplication of a differential operator over gauges associated with an edge.\nGraphically, the sequence is interpreted as a repetitive elimination of edges\nresulting in a sequence of models on decreasing in size graphs with the same\nPartition Function. Even though complexity of computing factors in the sequence\nmodels grow exponentially with the number of eliminated edges, polynomials\nassociated with the new factors remain bi-stable if the original factors have\nthis property. Moreover, we show that Belief Propagation estimations in the\nsequence do not decrease, each low-bounding the Partition Function.', 'We present an algorithm for arc reversal in Bayesian networks with\ntree-structured conditional probability tables, and consider some of its\nadvantages, especially for the simulation of dynamic probabilistic networks. In\nparticular, the method allows one to produce CPTs for nodes involved in the\nreversal that exploit regularities in the conditional distributions. We argue\nthat this approach alleviates some of the overhead associated with arc\nreversal, plays an important role in evidence integration and can be used to\nrestrict sampling of variables in DPNs. We also provide an algorithm that\ndetects the dynamic irrelevance of state variables in forward simulation. This\nalgorithm exploits the structured CPTs in a reversed network to determine, in a\ntime-independent fashion, the conditions under which a variable does or does\nnot need to be sampled.', 'Modeling and generating graphs is fundamental for studying networks in\nbiology, engineering, and social sciences. However, modeling complex\ndistributions over graphs and then efficiently sampling from these\ndistributions is challenging due to the non-unique, high-dimensional nature of\ngraphs and the complex, non-local dependencies that exist between edges in a\ngiven graph. Here we propose GraphRNN, a deep autoregressive model that\naddresses the above challenges and approximates any distribution of graphs with\nminimal assumptions about their structure. GraphRNN learns to generate graphs\nby training on a representative set of graphs and decomposes the graph\ngeneration process into a sequence of node and edge formations, conditioned on\nthe graph structure generated so far.\n  In order to quantitatively evaluate the performance of GraphRNN, we introduce\na benchmark suite of datasets, baselines and novel evaluation metrics based on\nMaximum Mean Discrepancy, which measure distances between sets of graphs. Our\nexperiments show that GraphRNN significantly outperforms all baselines,\nlearning to generate diverse graphs that match the structural characteristics\nof a target set, while also scaling to graphs 50 times larger than previous\ndeep models.', 'Bayesian graphical models have been shown to be a powerful tool for\ndiscovering uncertainty and causal structure from real-world data in many\napplication fields. Current inference methods primarily follow different kinds\nof trade-offs between computational complexity and predictive accuracy. At one\nend of the spectrum, variational inference approaches perform well in\ncomputational efficiency, while at the other end, Gibbs sampling approaches are\nknown to be relatively accurate for prediction in practice. In this paper, we\nextend an existing Gibbs sampling method, and propose a new deterministic Heron\ninference (Heron) for a family of Bayesian graphical models. In addition to the\nsupport for nontrivial distributability, one more benefit of Heron is that it\nis able to not only allow us to easily assess the convergence status but also\nlargely improve the running efficiency. We evaluate Heron against the standard\ncollapsed Gibbs sampler and state-of-the-art state augmentation method in\ninference for well-known graphical models. Experimental results using publicly\navailable real-life data have demonstrated that Heron significantly outperforms\nthe baseline methods for inferring Bayesian graphical models.', 'Probabilistic graphical models provide a powerful tool to describe complex\nstatistical structure, with many real-world applications in science and\nengineering from controlling robotic arms to understanding neuronal\ncomputations. A major challenge for these graphical models is that inferences\nsuch as marginalization are intractable for general graphs. These inferences\nare often approximated by a distributed message-passing algorithm such as\nBelief Propagation, which does not always perform well on graphs with cycles,\nnor can it always be easily specified for complex continuous probability\ndistributions. Such difficulties arise frequently in expressive graphical\nmodels that include intractable higher-order interactions. In this paper we\nconstruct iterative message-passing algorithms using Graph Neural Networks\ndefined on factor graphs to achieve fast approximate inference on graphical\nmodels that involve many-variable interactions. Experimental results on several\nfamilies of graphical models demonstrate the out-of-distribution generalization\ncapability of our method to different sized graphs, and indicate the domain in\nwhich our method gains advantage over Belief Propagation.', 'The analysis of practical probabilistic models on the computer demands a\nconvenient representation for the available knowledge and an efficient\nalgorithm to perform inference. An appealing representation is the influence\ndiagram, a network that makes explicit the random variables in a model and\ntheir probabilistic dependencies. Recent advances have developed solution\nprocedures based on the influence diagram. In this paper, we examine the\nfundamental properties that underlie those techniques, and the information\nabout the probabilistic structure that is available in the influence diagram\nrepresentation. The influence diagram is a convenient representation for\ncomputer processing while also being clear and non-mathematical. It displays\nprobabilistic dependence precisely, in a way that is intuitive for decision\nmakers and experts to understand and communicate. As a result, the same\ninfluence diagram can be used to build, assess and analyze a model,\nfacilitating changes in the formulation and feedback from sensitivity analysis.\nThe goal in this paper is to determine arbitrary conditional probability\ndistributions from a given probabilistic model. Given qualitative information\nabout the dependence of the random variables in the model we can, for a\nspecific conditional expression, specify precisely what quantitative\ninformation we need to be able to determine the desired conditional probability\ndistribution. It is also shown how we can find that probability distribution by\nperforming operations locally, that is, over subspaces of the joint\ndistribution. In this way, we can exploit the conditional independence present\nin the model to avoid having to construct or manipulate the full joint\ndistribution. These results are extended to include maximal processing when the\ninformation available is incomplete, and optimal decision making in an\nuncertain environment. Influence diagrams as a computer-aided modeling tool\nwere developed by Miller, Merkofer, and Howard [5] and extended by Howard and\nMatheson [2]. Good descriptions of how to use them in modeling are in Owen [7]\nand Howard and Matheson [2]. The notion of solving a decision problem through\ninfluence diagrams was examined by Olmsted [6] and such an algorithm was\ndeveloped by Shachter [8]. The latter paper also shows how influence diagrams\ncan be used to perform a variety of sensitivity analyses. This paper extends\nthose results by developing a theory of the properties of the diagram that are\nused by the algorithm, and the information needed to solve arbitrary\nprobability inference problems. Section 2 develops the notation and the\nframework for the paper and the relationship between influence diagrams and\njoint probability distributions. The general probabilistic inference problem is\nposed in Section 3. In Section 4 the transformations on the diagram are\ndeveloped and then put together into a solution procedure in Section 5. In\nSection 6, this procedure is used to calculate the information requirement to\nsolve an inference problem and the maximal processing that can be performed\nwith incomplete information. Section 7 contains a summary of results.', 'The Bayesian approach to machine learning amounts to computing posterior\ndistributions of random variables from a probabilistic model of how the\nvariables are related (that is, a prior distribution) and a set of observations\nof variables. There is a trend in machine learning towards expressing Bayesian\nmodels as probabilistic programs. As a foundation for this kind of programming,\nwe propose a core functional calculus with primitives for sampling prior\ndistributions and observing variables. We define measure-transformer\ncombinators inspired by theorems in measure theory, and use these to give a\nrigorous semantics to our core calculus. The original features of our semantics\ninclude its support for discrete, continuous, and hybrid measures, and, in\nparticular, for observations of zero-probability events. We compile our core\nlanguage to a small imperative language that is processed by an existing\ninference engine for factor graphs, which are data structures that enable many\nefficient inference algorithms. This allows efficient approximate inference of\nposterior marginal distributions, treating thousands of observations per second\nfor large instances of realistic models.', 'Global variational approximation methods in graphical models allow efficient\napproximate inference of complex posterior distributions by using a simpler\nmodel. The choice of the approximating model determines a tradeoff between the\ncomplexity of the approximation procedure and the quality of the approximation.\nIn this paper, we consider variational approximations based on two classes of\nmodels that are richer than standard Bayesian networks, Markov networks or\nmixture models. As such, these classes allow to find better tradeoffs in the\nspectrum of approximations. The first class of models are chain graphs, which\ncapture distributions that are partially directed. The second class of models\nare directed graphs (Bayesian networks) with additional latent variables. Both\nclasses allow representation of multi-variable dependencies that cannot be\neasily represented within a Bayesian network.', 'We propose a novel adaptive importance sampling scheme for Bayesian inversion\nproblems where the inference of the variables of interest and the power of the\ndata noise is split. More specifically, we consider a Bayesian analysis for the\nvariables of interest (i.e., the parameters of the model to invert), whereas we\nemploy a maximum likelihood approach for the estimation of the noise power. The\nwhole technique is implemented by means of an iterative procedure, alternating\nsampling and optimization steps. Moreover, the noise power is also used as a\ntempered parameter for the posterior distribution of the the variables of\ninterest. Therefore, a sequence of tempered posterior densities is generated,\nwhere the tempered parameter is automatically selected according to the actual\nestimation of the noise power. A complete Bayesian study over the model\nparameters and the scale parameter can be also performed. Numerical experiments\nshow the benefits of the proposed approach.', 'Modeling complex phenomena typically involves the use of both discrete and\ncontinuous variables. Such a setting applies across a wide range of problems,\nfrom identifying trends in time-series data to performing effective\ncompositional scene understanding in images. Here, we propose Hybrid Memoised\nWake-Sleep (HMWS), an algorithm for effective inference in such hybrid\ndiscrete-continuous models. Prior approaches to learning suffer as they need to\nperform repeated expensive inner-loop discrete inference. We build on a recent\napproach, Memoised Wake-Sleep (MWS), which alleviates part of the problem by\nmemoising discrete variables, and extend it to allow for a principled and\neffective way to handle continuous variables by learning a separate recognition\nmodel used for importance-sampling based approximate inference and\nmarginalization. We evaluate HMWS in the GP-kernel learning and 3D scene\nunderstanding domains, and show that it outperforms current state-of-the-art\ninference methods.', 'Reasoning on large and complex real-world models is a computationally\ndifficult task, yet one that is required for effective use of many AI\napplications. A plethora of inference algorithms have been developed that work\nwell on specific models or only on parts of general models. Consequently, a\nsystem that can intelligently apply these inference algorithms to different\nparts of a model for fast reasoning is highly desirable. We introduce a new\nframework called structured factored inference (SFI) that provides the\nfoundation for such a system. Using models encoded in a probabilistic\nprogramming language, SFI provides a sound means to decompose a model into\nsub-models, apply an inference algorithm to each sub-model, and combine the\nresulting information to answer a query. Our results show that SFI is nearly as\naccurate as exact inference yet retains the benefits of approximate inference\nmethods.', 'The paper studies empirically the time-space trade-off between sampling and\ninference in a sl cutset sampling algorithm. The algorithm samples over a\nsubset of nodes in a Bayesian network and applies exact inference over the\nrest. Consequently, while the size of the sampling space decreases, requiring\nless samples for convergence, the time for generating each single sample\nincreases. The w-cutset sampling selects a sampling set such that the\ninduced-width of the network when the sampling set is observed is bounded by w,\nthus requiring inference whose complexity is exponential in w. In this paper,\nwe investigate performance of w-cutset sampling over a range of w values and\nmeasure the accuracy of w-cutset sampling as a function of w. Our experiments\ndemonstrate that the cutset sampling idea is quite powerful showing that an\noptimal balance between inference and sampling benefits substantially from\nrestricting the cutset size, even at the cost of more complex inference.', 'Recent developments using directed acyclical graphs (i.e., influence diagrams\nand Bayesian networks) for knowledge representation have lessened the problems\nof using probability in knowledge-based systems (KBS). Most current research\ninvolves the efficient propagation of new evidence, but little has been done\nconcerning the maintenance of domain-specific knowledge, which includes the\nprobabilistic information about the problem domain. By making use of\nconditional independencies represented in she graphs, however, probability\nassessments are required only for certain variables when the knowledge base is\nupdated. The purpose of this study was to investigate, for those variables\nwhich require probability assessments, ways to reduce the amount of new\nknowledge required from the expert when updating probabilistic information in a\nprobabilistic knowledge-based system. Three special cases (ignored outcome,\nsplit outcome, and assumed constraint outcome) were identified under which many\nof the original probabilities (those already in the knowledge-base) do not need\nto be reassessed when maintenance is required.', 'In this paper, we introduce a method for approximating the solution to\ninference and optimization tasks in uncertain and deterministic reasoning. Such\ntasks are in general intractable for exact algorithms because of the large\nnumber of dependency relationships in their structure. Our method effectively\nmaps such a dense problem to a sparser one which is in some sense ""closest"".\nExact methods can be run on the sparser problem to derive bounds on the\noriginal answer, which can be quite sharp. We present empirical results\ndemonstrating that our method works well on the tasks of belief inference and\nfinding the probability of the most probable explanation in belief networks,\nand finding the cost of the solution that violates the smallest number of\nconstraints in constraint satisfaction problems. On one large CPCS network, for\nexample, we were able to calculate upper and lower bounds on the conditional\nprobability of a variable, given evidence, that were almost identical in the\naverage case.', 'In this paper, we use evidence-specific value abstraction for speeding\nBayesian networks inference. This is done by grouping variable values and\ntreating the combined values as a single entity. As we show, such abstractions\ncan exploit regularities in conditional probability distributions and also the\nspecific values of observed variables. To formally justify value abstraction,\nwe define the notion of safe value abstraction and devise inference algorithms\nthat use it to reduce the cost of inference. Our procedure is particularly\nuseful for learning complex networks with many hidden variables. In such cases,\nrepeated likelihood computations are required for EM or other parameter\noptimization techniques. Since these computations are repeated with respect to\nthe same evidence set, our methods can provide significant speedup to the\nlearning procedure. We demonstrate the algorithm on genetic linkage problems\nwhere the use of value abstraction sometimes differentiates between a feasible\nand non-feasible solution.', 'Inference in continuous label Markov random fields is a challenging task. We\nuse particle belief propagation (PBP) for solving the inference problem in\ncontinuous label space. Sampling particles from the belief distribution is\ntypically done by using Metropolis-Hastings Markov chain Monte Carlo methods\nwhich involves sampling from a proposal distribution. This proposal\ndistribution has to be carefully designed depending on the particular model and\ninput data to achieve fast convergence. We propose to avoid dependence on a\nproposal distribution by introducing a slice sampling based PBP algorithm. The\nproposed approach shows superior convergence performance on an image denoising\ntoy example. Our findings are validated on a challenging relational 2D feature\ntracking application.', 'Many real world models can be characterized as weak, meaning that there is\nsignificant uncertainty in both the data input and inferences. This lack of\ndeterminism makes it especially difficult for users of computer decision aids\nto understand and have confidence in the models. This paper presents a\nrepresentation for uncertainty and utilities that serves as a framework for\ngraphical summary and computer-generated explanation of decision models. The\napplication described that tests the methodology is a computer decision aid\ndesigned to enhance the clinician-patient consultation process for patients\nwith angina (chest pain due to lack of blood flow to the heart muscle). The\nangina model is represented as a Bayesian decision network. Additionally, the\nprobabilities and utilities are treated as random variables with probability\ndistributions on their range of possible values. The initial distributions\nrepresent information on all patients with anginal symptoms, and the approach\nallows for rapid tailoring to more patientspecific distributions. This\nframework provides a metric for judging the importance of each variable in the\nmodel dynamically.', 'Approximate Bayesian inference is NP-hard. Dagum and Luby defined the Local\nVariance Bound (LVB) to measure the approximation hardness of Bayesian\ninference on Bayesian networks, assuming the networks model strictly positive\njoint probability distributions, i.e. zero probabilities are not permitted.\nThis paper introduces the k-test to measure the approximation hardness of\ninference on Bayesian networks with deterministic causalities in the\nprobability distribution, i.e. when zero conditional probabilities are\npermitted. Approximation by stochastic sampling is a widely-used inference\nmethod that is known to suffer from inefficiencies due to sample rejection. The\nk-test predicts when rejection rates of stochastic sampling a Bayesian network\nwill be low, modest, high, or when sampling is intractable.', 'For many optimization problems it is possible to define a distance metric\nbetween problem variables that correlates with the likelihood and strength of\ninteractions between the variables. For example, one may define a metric so\nthat the dependencies between variables that are closer to each other with\nrespect to the metric are expected to be stronger than the dependencies between\nvariables that are further apart. The purpose of this paper is to describe a\nmethod that combines such a problem-specific distance metric with information\nmined from probabilistic models obtained in previous runs of estimation of\ndistribution algorithms with the goal of solving future problem instances of\nsimilar type with increased speed, accuracy and reliability. While the focus of\nthe paper is on additively decomposable problems and the hierarchical Bayesian\noptimization algorithm, it should be straightforward to generalize the approach\nto other model-directed optimization techniques and other problem classes.\nCompared to other techniques for learning from experience put forward in the\npast, the proposed technique is both more practical and more broadly\napplicable.', 'This paper introduces the probabilistic module interface, which allows\nencapsulation of complex probabilistic models with latent variables alongside\ncustom stochastic approximate inference machinery, and provides a\nplatform-agnostic abstraction barrier separating the model internals from the\nhost probabilistic inference system. The interface can be seen as a stochastic\ngeneralization of a standard simulation and density interface for probabilistic\nprimitives. We show that sound approximate inference algorithms can be\nconstructed for networks of probabilistic modules, and we demonstrate that the\ninterface can be implemented using learned stochastic inference networks and\nMCMC and SMC approximate inference programs.', 'Tasks such as record linkage and multi-target tracking, which involve\nreconstructing the set of objects that underlie some observed data, are\nparticularly challenging for probabilistic inference. Recent work has achieved\nefficient and accurate inference on such problems using Markov chain Monte\nCarlo (MCMC) techniques with customized proposal distributions. Currently,\nimplementing such a system requires coding MCMC state representations and\nacceptance probability calculations that are specific to a particular\napplication. An alternative approach, which we pursue in this paper, is to use\na general-purpose probabilistic modeling language (such as BLOG) and a generic\nMetropolis-Hastings MCMC algorithm that supports user-supplied proposal\ndistributions. Our algorithm gains flexibility by using MCMC states that are\nonly partial descriptions of possible worlds; we provide conditions under which\nMCMC over partial worlds yields correct answers to queries. We also show how to\nuse a context-specific Bayes net to identify the factors in the acceptance\nprobability that need to be computed for a given proposed move. Experimental\nresults on a citation matching task show that our general-purpose MCMC engine\ncompares favorably with an application-specific system.', 'Graphical causal models are an important tool for knowledge discovery because\nthey can represent both the causal relations between variables and the\nmultivariate probability distributions over the data. Once learned, causal\ngraphs can be used for classification, feature selection and hypothesis\ngeneration, while revealing the underlying causal network structure and thus\nallowing for arbitrary likelihood queries over the data. However, current\nalgorithms for learning sparse directed graphs are generally designed to handle\nonly one type of data (continuous-only or discrete-only), which limits their\napplicability to a large class of multi-modal biological datasets that include\nmixed type variables. To address this issue, we developed new methods that\nmodify and combine existing methods for finding undirected graphs with methods\nfor finding directed graphs. These hybrid methods are not only faster, but also\nperform better than the directed graph estimation methods alone for a variety\nof parameter settings and data set sizes. Here, we describe a new conditional\nindependence test for learning directed graphs over mixed data types and we\ncompare performances of different graph learning strategies on synthetic data.', 'Traditional approaches to Bayes net structure learning typically assume\nlittle regularity in graph structure other than sparseness. However, in many\ncases, we expect more systematicity: variables in real-world systems often\ngroup into classes that predict the kinds of probabilistic dependencies they\nparticipate in. Here we capture this form of prior knowledge in a hierarchical\nBayesian framework, and exploit it to enable structure learning and type\ndiscovery from small datasets. Specifically, we present a nonparametric\ngenerative model for directed acyclic graphs as a prior for Bayes net structure\nlearning. Our model assumes that variables come in one or more classes and that\nthe prior probability of an edge existing between two variables is a function\nonly of their classes. We derive an MCMC algorithm for simultaneous inference\nof the number of classes, the class assignments of variables, and the Bayes net\nstructure over variables. For several realistic, sparse datasets, we show that\nthe bias towards systematicity of connections provided by our model yields more\naccurate learned networks than a traditional, uniform prior approach, and that\nthe classes found by our model are appropriate.', 'Recent work has explored transforming data sets into smaller, approximate\nsummaries in order to scale Bayesian inference. We examine a related problem in\nwhich the parameters of a Bayesian model are very large and expensive to store\nin memory, and propose more compact representations of parameter values that\ncan be used during inference. We focus on a class of graphical models that we\nrefer to as latent Dirichlet-Categorical models, and show how a combination of\ntwo sketching algorithms known as count-min sketch and approximate counters\nprovide an efficient representation for them. We show that this sketch\ncombination -- which, despite having been used before in NLP applications, has\nnot been previously analyzed -- enjoys desirable properties. We prove that for\nthis class of models, when the sketches are used during Markov Chain Monte\nCarlo inference, the equilibrium of sketched MCMC converges to that of the\nexact chain as sketch parameters are tuned to reduce the error rate.', 'The paper introduces AND/OR importance sampling for probabilistic graphical\nmodels. In contrast to importance sampling, AND/OR importance sampling caches\nsamples in the AND/OR space and then extracts a new sample mean from the stored\nsamples. We prove that AND/OR importance sampling may have lower variance than\nimportance sampling; thereby providing a theoretical justification for\npreferring it over importance sampling. Our empirical evaluation demonstrates\nthat AND/OR importance sampling is far more accurate than importance sampling\nin many cases.', 'Inference in discrete graphical models with variational methods is difficult\nbecause of the inability to re-parameterize gradients of the Evidence Lower\nBound (ELBO). Many sampling-based methods have been proposed for estimating\nthese gradients, but they suffer from high bias or variance. In this paper, we\npropose a new approach that leverages the tractability of probabilistic circuit\nmodels, such as Sum Product Networks (SPN), to compute ELBO gradients exactly\n(without sampling) for a certain class of densities. In particular, we show\nthat selective-SPNs are suitable as an expressive variational distribution, and\nprove that when the log-density of the target model is a polynomial the\ncorresponding ELBO can be computed analytically. To scale to graphical models\nwith thousands of variables, we develop an efficient and effective construction\nof selective-SPNs with size $O(kn)$, where $n$ is the number of variables and\n$k$ is an adjustable hyperparameter. We demonstrate our approach on three types\nof graphical models -- Ising models, Latent Dirichlet Allocation, and factor\ngraphs from the UAI Inference Competition. Selective-SPNs give a better lower\nbound than mean-field and structured mean-field, and is competitive with\napproximations that do not provide a lower bound, such as Loopy Belief\nPropagation and Tree-Reweighted Belief Propagation. Our results show that\nprobabilistic circuits are promising tools for variational inference in\ndiscrete graphical models as they combine tractability and expressivity.', 'Sampling is an important tool for estimating large, complex sums and\nintegrals over high dimensional spaces. For instance, important sampling has\nbeen used as an alternative to exact methods for inference in belief networks.\nIdeally, we want to have a sampling distribution that provides optimal-variance\nestimators. In this paper, we present methods that improve the sampling\ndistribution by systematically adapting it as we obtain information from the\nsamples. We present a stochastic-gradient-descent method for sequentially\nupdating the sampling distribution based on the direct minization of the\nvariance. We also present other stochastic-gradient-descent methods based on\nthe minimizationof typical notions of distance between the current sampling\ndistribution and approximations of the target, optimal distribution. We finally\nvalidate and compare the different methods empirically by applying them to the\nproblem of action evaluation in influence diagrams.', 'Given a CNF formula and a weight for each assignment of values to variables,\ntwo natural problems are weighted model counting and distribution-aware\nsampling of satisfying assignments. Both problems have a wide variety of\nimportant applications. Due to the inherent complexity of the exact versions of\nthe problems, interest has focused on solving them approximately. Prior work in\nthis area scaled only to small problems in practice, or failed to provide\nstrong theoretical guarantees, or employed a computationally-expensive maximum\na posteriori probability (MAP) oracle that assumes prior knowledge of a\nfactored representation of the weight distribution. We present a novel approach\nthat works with a black-box oracle for weights of assignments and requires only\nan {\\NP}-oracle (in practice, a SAT-solver) to solve both the counting and\nsampling problems. Our approach works under mild assumptions on the\ndistribution of weights of satisfying assignments, provides strong theoretical\nguarantees, and scales to problems involving several thousand variables. We\nalso show that the assumptions can be significantly relaxed while improving\ncomputational efficiency if a factored representation of the weights is known.', 'Deep directed generative models have attracted much attention recently due to\ntheir expressive representation power and the ability of ancestral sampling.\nOne major difficulty of learning directed models with many latent variables is\nthe intractable inference. To address this problem, most existing algorithms\nmake assumptions to render the latent variables independent of each other,\neither by designing specific priors, or by approximating the true posterior\nusing a factorized distribution. We believe the correlations among latent\nvariables are crucial for faithful data representation. Driven by this idea, we\npropose an inference method based on the conditional pseudo-likelihood that\npreserves the dependencies among the latent variables. For learning, we propose\nto employ the hard Expectation Maximization (EM) algorithm, which avoids the\nintractability of the traditional EM by max-out instead of sum-out to compute\nthe data likelihood. Qualitative and quantitative evaluations of our model\nagainst state of the art deep models on benchmark datasets demonstrate the\neffectiveness of the proposed algorithm in data representation and\nreconstruction.', ""We describe algorithms for learning Bayesian networks from a combination of\nuser knowledge and statistical data. The algorithms have two components: a\nscoring metric and a search procedure. The scoring metric takes a network\nstructure, statistical data, and a user's prior knowledge, and returns a score\nproportional to the posterior probability of the network structure given the\ndata. The search procedure generates networks for evaluation by the scoring\nmetric. Previous work has concentrated on metrics for domains containing only\ndiscrete variables, under the assumption that data represents a multinomial\nsample. In this paper, we extend this work, developing scoring metrics for\ndomains containing all continuous variables or a mixture of discrete and\ncontinuous variables, under the assumption that continuous data is sampled from\na multivariate normal distribution. Our work extends traditional statistical\napproaches for identifying vanishing regression coefficients in that we\nidentify two important assumptions, called event equivalence and parameter\nmodularity, that when combined allow the construction of prior distributions\nfor multivariate normal parameters from a single prior Bayesian network\nspecified by a user."", 'Lifted inference has been proposed for various probabilistic logical\nframeworks in order to compute the probability of queries in a time that\ndepends on the size of the domains of the random variables rather than the\nnumber of instances. Even if various authors have underlined its importance for\nprobabilistic logic programming (PLP), lifted inference has been applied up to\nnow only to relational languages outside of logic programming. In this paper we\nadapt Generalized Counting First Order Variable Elimination (GC-FOVE) to the\nproblem of computing the probability of queries to probabilistic logic programs\nunder the distribution semantics. In particular, we extend the Prolog Factor\nLanguage (PFL) to include two new types of factors that are needed for\nrepresenting ProbLog programs. These factors take into account the existing\ncausal independence relationships among random variables and are managed by the\nextension to variable elimination proposed by Zhang and Poole for dealing with\nconvergent variables and heterogeneous factors. Two new operators are added to\nGC-FOVE for treating heterogeneous factors. The resulting algorithm, called\nLP$^2$ for Lifted Probabilistic Logic Programming, has been implemented by\nmodifying the PFL implementation of GC-FOVE and tested on three benchmarks for\nlifted inference. A comparison with PITA and ProbLog2 shows the potential of\nthe approach.', 'When a hybrid Bayesian network has conditionally deterministic variables with\ncontinuous parents, the joint density function for the continuous variables\ndoes not exist. Conditional linear Gaussian distributions can handle such cases\nwhen the continuous variables have a multi-variate normal distribution and the\ndiscrete variables do not have continuous parents. In this paper, operations\nrequired for performing inference with conditionally deterministic variables in\nhybrid Bayesian networks are developed. These methods allow inference in\nnetworks with deterministic variables where continuous variables may be\nnon-Gaussian, and their density functions can be approximated by mixtures of\ntruncated exponentials. There are no constraints on the placement of continuous\nand discrete nodes in the network.', 'We present new results on the classical algorithm of variable elimination,\nwhich underlies many algorithms including for probabilistic inference. The\nresults relate to exploiting functional dependencies, allowing one to perform\ninference and learning efficiently on models that have very large treewidth.\nThe highlight of the advance is that it works with standard (dense) factors,\nwithout the need for sparse factors or techniques based on knowledge\ncompilation that are commonly utilized. This is significant as it permits a\ndirect implementation of the improved variable elimination algorithm using\ntensors and their operations, leading to extremely efficient implementations\nespecially when learning model parameters. Moreover, the proposed technique\ndoes not require knowledge of the specific functional dependencies, only that\nthey exist, so can be used when learning these dependencies. We illustrate the\nefficacy of our proposed algorithm by compiling Bayesian network queries into\ntensor graphs and then learning their parameters from labeled data using a\nstandard tool for tensor computation.', 'A structural causal model is made of endogenous (manifest) and exogenous\n(latent) variables. We show that endogenous observations induce linear\nconstraints on the probabilities of the exogenous variables. This allows to\nexactly map a causal model into a credal network. Causal inferences, such as\ninterventions and counterfactuals, can consequently be obtained by standard\nalgorithms for the updating of credal nets. These natively return sharp values\nin the identifiable case, while intervals corresponding to the exact bounds are\nproduced for unidentifiable queries. A characterization of the causal models\nthat allow the map above to be compactly derived is given, along with a\ndiscussion about the scalability for general models. This contribution should\nbe regarded as a systematic approach to represent structural causal models by\ncredal networks and hence to systematically compute causal inferences. A number\nof demonstrative examples is presented to clarify our methodology. Extensive\nexperiments show that approximate algorithms for credal networks can\nimmediately be used to do causal inference in real-size problems.', 'A sum-product network (SPN) is a probabilistic model, based on a rooted\nacyclic directed graph, in which terminal nodes represent univariate\nprobability distributions and non-terminal nodes represent convex combinations\n(weighted sums) and products of probability functions. They are closely related\nto probabilistic graphical models, in particular to Bayesian networks with\nmultiple context-specific independencies. Their main advantage is the\npossibility of building tractable models from data, i.e., models that can\nperform several inference tasks in time proportional to the number of links in\nthe graph. They are somewhat similar to neural networks and can address the\nsame kinds of problems, such as image processing and natural language\nunderstanding. This paper offers a survey of SPNs, including their definition,\nthe main algorithms for inference and learning from data, the main\napplications, a brief review of software libraries, and a comparison with\nrelated models', ""We describe Venture, an interactive virtual machine for probabilistic\nprogramming that aims to be sufficiently expressive, extensible, and efficient\nfor general-purpose use. Like Church, probabilistic models and inference\nproblems in Venture are specified via a Turing-complete, higher-order\nprobabilistic language descended from Lisp. Unlike Church, Venture also\nprovides a compositional language for custom inference strategies built out of\nscalable exact and approximate techniques. We also describe four key aspects of\nVenture's implementation that build on ideas from probabilistic graphical\nmodels. First, we describe the stochastic procedure interface (SPI) that\nspecifies and encapsulates primitive random variables. The SPI supports custom\ncontrol flow, higher-order probabilistic procedures, partially exchangeable\nsequences and ``likelihood-free'' stochastic simulators. It also supports\nexternal models that do inference over latent variables hidden from Venture.\nSecond, we describe probabilistic execution traces (PETs), which represent\nexecution histories of Venture programs. PETs capture conditional dependencies,\nexistential dependencies and exchangeable coupling. Third, we describe\npartitions of execution histories called scaffolds that factor global inference\nproblems into coherent sub-problems. Finally, we describe a family of\nstochastic regeneration algorithms for efficiently modifying PET fragments\ncontained within scaffolds. Stochastic regeneration linear runtime scaling in\ncases where many previous approaches scaled quadratically. We show how to use\nstochastic regeneration and the SPI to implement general-purpose inference\nstrategies such as Metropolis-Hastings, Gibbs sampling, and blocked proposals\nbased on particle Markov chain Monte Carlo and mean-field variational inference\ntechniques."", 'In the big data era, scalability has become a crucial requirement for any\nuseful computational model. Probabilistic graphical models are very useful for\nmining and discovering data insights, but they are not scalable enough to be\nsuitable for big data problems. Bayesian Networks particularly demonstrate this\nlimitation when their data is represented using few random variables while each\nrandom variable has a massive set of values. With hierarchical data - data that\nis arranged in a treelike structure with several levels - one would expect to\nsee hundreds of thousands or millions of values distributed over even just a\nsmall number of levels. When modeling this kind of hierarchical data across\nlarge data sets, Bayesian networks become infeasible for representing the\nprobability distributions for the following reasons: i) Each level represents a\nsingle random variable with hundreds of thousands of values, ii) The number of\nlevels is usually small, so there are also few random variables, and iii) The\nstructure of the network is predefined since the dependency is modeled top-down\nfrom each parent to each of its child nodes, so the network would contain a\nsingle linear path for the random variables from each parent to each child\nnode. In this paper we present a scalable probabilistic graphical model to\novercome these limitations for massive hierarchical data. We believe the\nproposed model will lead to an easily-scalable, more readable, and expressive\nimplementation for problems that require probabilistic-based solutions for\nmassive amounts of hierarchical data. We successfully applied this model to\nsolve two different challenging probabilistic-based problems on massive\nhierarchical data sets for different domains, namely, bioinformatics and latent\nsemantic discovery over search logs.', 'Probabilistic programming provides the means to represent and reason about\ncomplex probabilistic models using programming language constructs. Even simple\nprobabilistic programs can produce models with infinitely many variables.\nFactored inference algorithms are widely used for probabilistic graphical\nmodels, but cannot be applied to these programs because all the variables and\nfactors have to be enumerated. In this paper, we present a new inference\nframework, lazy factored inference (LFI), that enables factored algorithms to\nbe used for models with infinitely many variables. LFI expands the model to a\nbounded depth and uses the structure of the program to precisely quantify the\neffect of the unexpanded part of the model, producing lower and upper bounds to\nthe probability of the query.', 'While client sampling is a central operation of current state-of-the-art\nfederated learning (FL) approaches, the impact of this procedure on the\nconvergence and speed of FL remains under-investigated. In this work, we\nprovide a general theoretical framework to quantify the impact of a client\nsampling scheme and of the clients heterogeneity on the federated optimization.\nFirst, we provide a unified theoretical ground for previously reported sampling\nschemes experimental results on the relationship between FL convergence and the\nvariance of the aggregation weights. Second, we prove for the first time that\nthe quality of FL convergence is also impacted by the resulting covariance\nbetween aggregation weights. Our theory is general, and is here applied to\nMultinomial Distribution (MD) and Uniform sampling, two default unbiased client\nsampling schemes of FL, and demonstrated through a series of experiments in\nnon-iid and unbalanced scenarios. Our results suggest that MD sampling should\nbe used as default sampling scheme, due to the resilience to the changes in\ndata ratio during the learning process, while Uniform sampling is superior only\nin the special case when clients have the same amount of data.', ""Several Artificial Intelligence schemes for reasoning under uncertainty\nexplore either explicitly or implicitly asymmetries among probabilities of\nvarious states of their uncertain domain models. Even though the correct\nworking of these schemes is practically contingent upon the existence of a\nsmall number of probable states, no formal justification has been proposed of\nwhy this should be the case. This paper attempts to fill this apparent gap by\nstudying asymmetries among probabilities of various states of uncertain models.\nBy rewriting the joint probability distribution over a model's variables into a\nproduct of individual variables' prior and conditional probability\ndistributions, and applying central limit theorem to this product, we can\ndemonstrate that the probabilities of individual states of the model can be\nexpected to be drawn from highly skewed, log-normal distributions. With\nsufficient asymmetry in individual prior and conditional probability\ndistributions, a small fraction of states can be expected to cover a large\nportion of the total probability space with the remaining states having\npractically negligible probability. Theoretical discussion is supplemented by\nsimulation results and an illustrative real-world example."", 'Importance sampling is often used in machine learning when training and\ntesting data come from different distributions. In this paper we propose a new\nvariant of importance sampling that can reduce the variance of importance\nsampling-based estimates by orders of magnitude when the supports of the\ntraining and testing distributions differ. After motivating and presenting our\nnew importance sampling estimator, we provide a detailed theoretical analysis\nthat characterizes both its bias and variance relative to the ordinary\nimportance sampling estimator (in various settings, which include cases where\nordinary importance sampling is biased, while our new estimator is not, and\nvice versa). We conclude with an example of how our new importance sampling\nestimator can be used to improve estimates of how well a new treatment policy\nfor diabetes will work for an individual, using only data from when the\nindividual used a previous treatment policy.', 'Events and entities are closely related; entities are often actors or\nparticipants in events and events without entities are uncommon. The\ninterpretation of events and entities is highly contextually dependent.\nExisting work in information extraction typically models events separately from\nentities, and performs inference at the sentence level, ignoring the rest of\nthe document. In this paper, we propose a novel approach that models the\ndependencies among variables of events, entities, and their relations, and\nperforms joint inference of these variables across a document. The goal is to\nenable access to document-level contextual information and facilitate\ncontext-aware predictions. We demonstrate that our approach substantially\noutperforms the state-of-the-art methods for event extraction as well as a\nstrong baseline for entity extraction.', 'We present two sampling algorithms for probabilistic confidence inference in\nBayesian networks. These two algorithms (we call them AIS-BN-mu and\nAIS-BN-sigma algorithms) guarantee that estimates of posterior probabilities\nare with a given probability within a desired precision bound. Our algorithms\nare based on recent advances in sampling algorithms for (1) estimating the mean\nof bounded random variables and (2) adaptive importance sampling in Bayesian\nnetworks. In addition to a simple stopping rule for sampling that they provide,\nthe AIS-BN-mu and AIS-BN-sigma algorithms are capable of guiding the learning\nprocess in the AIS-BN algorithm. An empirical evaluation of the proposed\nalgorithms shows excellent performance, even for very unlikely evidence.', 'Arriving at the complete probabilistic knowledge of a domain, i.e., learning\nhow all variables interact, is indeed a demanding task. In reality, settings\noften arise for which an individual merely possesses partial knowledge of the\ndomain, and yet, is expected to give adequate answers to a variety of posed\nqueries. That is, although precise answers to some queries, in principle,\ncannot be achieved, a range of plausible answers is attainable for each query\ngiven the available partial knowledge. In this paper, we propose the\nMulti-Context Model (MCM), a new graphical model to represent the state of\npartial knowledge as to a domain. MCM is a middle ground between Probabilistic\nLogic, Bayesian Logic, and Probabilistic Graphical Models. For this model we\ndiscuss: (i) the dynamics of constructing a contradiction-free MCM, i.e., to\nform partial beliefs regarding a domain in a gradual and probabilistically\nconsistent way, and (ii) how to perform inference, i.e., to evaluate a\nprobability of interest involving some variables of the domain.', 'Bayesian networks provide a method of representing conditional independence\nbetween random variables and computing the probability distributions associated\nwith these random variables. In this paper, we extend Bayesian network\nstructures to compute probability density functions for continuous random\nvariables. We make this extension by approximating prior and conditional\ndensities using sums of weighted Gaussian distributions and then finding the\npropagation rules for updating the densities in terms of these weights. We\npresent a simple example that illustrates the Bayesian network for continuous\nvariables; this example shows the effect of the network structure and\napproximation errors on the computation of densities for variables in the\nnetwork.', 'Bayesian networks provide a language for qualitatively representing the\nconditional independence properties of a distribution. This allows a natural\nand compact representation of the distribution, eases knowledge acquisition,\nand supports effective inference algorithms. It is well-known, however, that\nthere are certain independencies that we cannot capture qualitatively within\nthe Bayesian network structure: independencies that hold only in certain\ncontexts, i.e., given a specific assignment of values to certain variables. In\nthis paper, we propose a formal notion of context-specific independence (CSI),\nbased on regularities in the conditional probability tables (CPTs) at a node.\nWe present a technique, analogous to (and based on) d-separation, for\ndetermining when such independence holds in a given network. We then focus on a\nparticular qualitative representation scheme - tree-structured CPTs - for\ncapturing CSI. We suggest ways in which this representation can be used to\nsupport effective inference algorithms. In particular, we present a structural\ndecomposition of the resulting network which can improve the performance of\nclustering algorithms, and an alternative algorithm based on cutset\nconditioning.', 'Graphical models are useful tools for describing structured high-dimensional\nprobability distributions. Development of efficient algorithms for learning\ngraphical models with least amount of data remains an active research topic.\nReconstruction of graphical models that describe the statistics of discrete\nvariables is a particularly challenging problem, for which the maximum\nlikelihood approach is intractable. In this work, we provide the first\nsample-efficient method based on the Interaction Screening framework that\nallows one to provably learn fully general discrete factor models with\nnode-specific discrete alphabets and multi-body interactions, specified in an\narbitrary basis. We identify a single condition related to model\nparametrization that leads to rigorous guarantees on the recovery of model\nstructure and parameters in any error norm, and is readily verifiable for a\nlarge class of models. Importantly, our bounds make explicit distinction\nbetween parameters that are proper to the model and priors used as an input to\nthe algorithm. Finally, we show that the Interaction Screening framework\nincludes all models previously considered in the literature as special cases,\nand for which our analysis shows a systematic improvement in sample complexity.', 'A key limitation of sampling algorithms for approximate inference is that it\nis difficult to quantify their approximation error. Widely used sampling\nschemes, such as sequential importance sampling with resampling and\nMetropolis-Hastings, produce output samples drawn from a distribution that may\nbe far from the target posterior distribution. This paper shows how to\nupper-bound the symmetric KL divergence between the output distribution of a\nbroad class of sequential Monte Carlo (SMC) samplers and their target posterior\ndistributions, subject to assumptions about the accuracy of a separate\ngold-standard sampler. The proposed method applies to samplers that combine\nmultiple particles, multinomial resampling, and rejuvenation kernels. The\nexperiments show the technique being used to estimate bounds on the divergence\nof SMC samplers for posterior inference in a Bayesian linear regression model\nand a Dirichlet process mixture model.', 'Thompson sampling has emerged as an effective heuristic for a broad range of\nonline decision problems. In its basic form, the algorithm requires computing\nand sampling from a posterior distribution over models, which is tractable only\nfor simple special cases. This paper develops ensemble sampling, which aims to\napproximate Thompson sampling while maintaining tractability even in the face\nof complex models such as neural networks. Ensemble sampling dramatically\nexpands on the range of applications for which Thompson sampling is viable. We\nestablish a theoretical basis that supports the approach and present\ncomputational results that offer further insight.', 'Softmax classifiers with a very large number of classes naturally occur in\nmany applications such as natural language processing and information\nretrieval. The calculation of full softmax is costly from the computational and\nenergy perspective. There have been various sampling approaches to overcome\nthis challenge, popularly known as negative sampling (NS). Ideally, NS should\nsample negative classes from a distribution that is dependent on the input\ndata, the current parameters, and the correct positive class. Unfortunately,\ndue to the dynamically updated parameters and data samples, there is no\nsampling scheme that is provably adaptive and samples the negative classes\nefficiently. Therefore, alternative heuristics like random sampling, static\nfrequency-based sampling, or learning-based biased sampling, which primarily\ntrade either the sampling cost or the adaptivity of samples per iteration are\nadopted. In this paper, we show two classes of distributions where the sampling\nscheme is truly adaptive and provably generates negative samples in\nnear-constant time. Our implementation in C++ on CPU is significantly superior,\nboth in terms of wall-clock time and accuracy, compared to the most optimized\nTensorFlow implementations of other popular negative sampling approaches on\npowerful NVIDIA V100 GPU.', 'Approximate probabilistic inference algorithms are central to many fields.\nExamples include sequential Monte Carlo inference in robotics, variational\ninference in machine learning, and Markov chain Monte Carlo inference in\nstatistics. A key problem faced by practitioners is measuring the accuracy of\nan approximate inference algorithm on a specific data set. This paper\nintroduces the auxiliary inference divergence estimator (AIDE), an algorithm\nfor measuring the accuracy of approximate inference algorithms. AIDE is based\non the observation that inference algorithms can be treated as probabilistic\nmodels and the random variables used within the inference algorithm can be\nviewed as auxiliary variables. This view leads to a new estimator for the\nsymmetric KL divergence between the approximating distributions of two\ninference algorithms. The paper illustrates application of AIDE to algorithms\nfor inference in regression, hidden Markov, and Dirichlet process mixture\nmodels. The experiments show that AIDE captures the qualitative behavior of a\nbroad class of inference algorithms and can detect failure modes of inference\nalgorithms that are missed by standard heuristics.', 'Instead of requiring a domain expert to specify the probabilistic\ndependencies of the data, in this work we present an approach that uses the\nrelational DB schema to automatically construct a Bayesian graphical model for\na database. This resulting model contains customized distributions for columns,\nlatent variables that cluster the data, and factors that reflect and represent\nthe foreign key links. Experiments demonstrate the accuracy of the model and\nthe scalability of inference on synthetic and real-world data.', 'Probabilistic logic programs are logic programs in which some of the facts\nare annotated with probabilities. This paper investigates how classical\ninference and learning tasks known from the graphical model community can be\ntackled for probabilistic logic programs. Several such tasks such as computing\nthe marginals given evidence and learning from (partial) interpretations have\nnot really been addressed for probabilistic logic programs before.\n  The first contribution of this paper is a suite of efficient algorithms for\nvarious inference tasks. It is based on a conversion of the program and the\nqueries and evidence to a weighted Boolean formula. This allows us to reduce\nthe inference tasks to well-studied tasks such as weighted model counting,\nwhich can be solved using state-of-the-art methods known from the graphical\nmodel and knowledge compilation literature. The second contribution is an\nalgorithm for parameter estimation in the learning from interpretations\nsetting. The algorithm employs Expectation Maximization, and is built on top of\nthe developed inference algorithms.\n  The proposed approach is experimentally evaluated. The results show that the\ninference algorithms improve upon the state-of-the-art in probabilistic logic\nprogramming and that it is indeed possible to learn the parameters of a\nprobabilistic logic program from interpretations.', ""Several interesting generative learning algorithms involve a complex\nprobability distribution over many random variables, involving intractable\nnormalization constants or latent variable normalization. Some of them may even\nnot have an analytic expression for the unnormalized probability function and\nno tractable approximation. This makes it difficult to estimate the quality of\nthese models, once they have been trained, or to monitor their quality (e.g.\nfor early stopping) while training. A previously proposed method is based on\nconstructing a non-parametric density estimator of the model's probability\nfunction from samples generated by the model. We revisit this idea, propose a\nmore efficient estimator, and prove that it provides a lower bound on the true\ntest log-likelihood, and an unbiased estimator as the number of generated\nsamples goes to infinity, although one that incorporates the effect of poor\nmixing. We further propose a biased variant of the estimator that can be used\nreliably with a finite number of samples for the purpose of model comparison."", 'Probabilistic programming languages and modeling toolkits are two modular\nways to build and reuse stochastic models and inference procedures. Combining\nstrengths of both, we express models and inference as generalized coroutines in\nthe same general-purpose language. We use existing facilities of the language,\nsuch as rich libraries, optimizing compilers, and types, to develop concise,\ndeclarative, and realistic models with competitive performance on exact and\napproximate inference. In particular, a wide range of models can be expressed\nusing memoization. Because deterministic parts of models run at full speed,\ncustom inference procedures are trivial to incorporate, and inference\nprocedures can reason about themselves without interpretive overhead. Within\nthis framework, we introduce a new, general algorithm for importance sampling\nwith look-ahead.', 'Systematic generalization aims to evaluate reasoning about novel combinations\nfrom known components, an intrinsic property of human cognition. In this work,\nwe study systematic generalization of NNs in forecasting future time series of\ndependent variables in a dynamical system, conditioned on past time series of\ndependent variables, and past and future control variables. We focus on\nsystematic generalization wherein the NN-based forecasting model should perform\nwell on previously unseen combinations or regimes of control variables after\nbeing trained on a limited set of the possible regimes. For NNs to depict such\nout-of-distribution generalization, they should be able to disentangle the\nvarious dependencies between control variables and dependent variables. We\nhypothesize that a modular NN architecture guided by the readily-available\nknowledge of independence of control variables as a potentially useful\ninductive bias to this end. Through extensive empirical evaluation on a toy\ndataset and a simulated electric motor dataset, we show that our proposed\nmodular NN architecture serves as a simple yet highly effective inductive bias\nthat enabling better forecasting of the dependent variables up to large\nhorizons in contrast to standard NNs, and indeed capture the true dependency\nrelations between the dependent and the control variables.', 'The presence of latent variables can greatly complicate inferences about\ncausal relations between measured variables from statistical data. In many\ncases, the presence of latent variables makes it impossible to determine for\ntwo measured variables A and B, whether A causes B, B causes A, or there is\nsome common cause. In this paper I present several theorems that state\nconditions under which it is possible to reliably infer the causal relation\nbetween two measured variables, regardless of whether latent variables are\nacting or not.', 'The goal of this paper is to design a causal inference method accounting for\ncomplex interactions between causal factors. The proposed method relies on a\ncategory theoretical reformulation of the definitions of dependent variables,\nindependent variables and latent variables in terms of products and arrows in\nthe category of unlabeled partitions. Throughout the paper, we demonstrate how\nthe proposed method accounts for possible hidden variables, such as\nenvironmental variables or noise, and how it can be interpreted statistically\nin terms of $p$-values. This interpretation, from category theory to\nstatistics, is implemented through a collection of propositions highlighting\nthe functorial properties of ANOVA. We use these properties in combination with\nour category theoretical framework to provide solutions to causal inference\nproblems with both sound algebraic and statistical properties. As an\napplication, we show how the proposed method can be used to design a\ncombinatorial genome-wide association algorithm for the field of genetics.', 'This paper considers the problem of invoking auxiliary, unobservable\nvariables to facilitate the structuring of causal tree models for a given set\nof continuous variables. Paralleling the treatment of bi-valued variables in\n[Pearl 1986], we show that if a collection of coupled variables are governed by\na joint normal distribution and a tree-structured representation exists, then\nboth the topology and all internal relationships of the tree can be uncovered\nby observing pairwise dependencies among the observed variables (i.e., the\nleaves of the tree). Furthermore, the conditions for normally distributed\nvariables are less restrictive than those governing bi-valued variables. The\nresult extends the applications of causal tree models which were found useful\nin evidential reasoning tasks.', 'It is suggested that an AI inference system should reflect an inference\npolicy that is tailored to the domain of problems to which it is applied -- and\nfurthermore that an inference policy need not conform to any general theory of\nrational inference or induction. We note, for instance, that Bayesian reasoning\nabout the probabilistic characteristics of an inference domain may result in\nthe specification of an nonBayesian procedure for reasoning within the\ninference domain. In this paper, the idea of an inference policy is explored in\nsome detail. To support this exploration, the characteristics of some standard\nand nonstandard inference policies are examined.', ""This paper addresses the problem of sampling from binary distributions with\nconstraints. In particular, it proposes an MCMC method to draw samples from a\ndistribution of the set of all states at a specified distance from some\nreference state. For example, when the reference state is the vector of zeros,\nthe algorithm can draw samples from a binary distribution with a constraint on\nthe number of active variables, say the number of 1's. We motivate the need for\nthis algorithm with examples from statistical physics and probabilistic\ninference. Unlike previous algorithms proposed to sample from binary\ndistributions with these constraints, the new algorithm allows for large moves\nin state space and tends to propose them such that they are energetically\nfavourable. The algorithm is demonstrated on three Boltzmann machines of\nvarying difficulty: A ferromagnetic Ising model (with positive potentials), a\nrestricted Boltzmann machine with learned Gabor-like filters as potentials, and\na challenging three-dimensional spin-glass (with positive and negative\npotentials)."", 'It is ""well known"" that in linear models: (1) testable constraints on the\nmarginal distribution of observed variables distinguish certain cases in which\nan unobserved cause jointly influences several observed variables; (2) the\ntechnique of ""instrumental variables"" sometimes permits an estimation of the\ninfluence of one variable on another even when the association between the\nvariables may be confounded by unobserved common causes; (3) the association\n(or conditional probability distribution of one variable given another) of two\nvariables connected by a path or trek can be computed directly from the\nparameter values associated with each edge in the path or trek; (4) the\nassociation of two variables produced by multiple treks can be computed from\nthe parameters associated with each trek; and (5) the independence of two\nvariables conditional on a third implies the corresponding independence of the\nsums of the variables over all units conditional on the sums over all units of\neach of the original conditioning variables.These properties are exploited in\nsearch procedures. It is also known that properties (2)-(5) do not hold for all\nBayes nets with binary variables. We show that (1) holds for all Bayes nets\nwith binary variables and (5) holds for all singly trek-connected Bayes nets of\nthat kind. We further show that all five properties hold for Bayes nets with\nany DAG and binary variables parameterized with noisy-or and noisy-and gates.', 'Flow prediction (e.g., crowd flow, traffic flow) with features of\nspatial-temporal is increasingly investigated in AI research field. It is very\nchallenging due to the complicated spatial dependencies between different\nlocations and dynamic temporal dependencies among different time intervals.\nAlthough measurements of both dependencies are employed, existing methods\nsuffer from the following two problems. First, the temporal dependencies are\nmeasured either uniformly or bias against long-term dependencies, which\noverlooks the distinctive impacts of short-term and long-term temporal\ndependencies. Second, the existing methods capture spatial and temporal\ndependencies independently, which wrongly assumes that the correlations between\nthese dependencies are weak and ignores the complicated mutual influences\nbetween them. To address these issues, we propose a Spatial-Temporal\nSelf-Attention Network (ST-SAN). As the path-length of attending long-term\ndependency is shorter in the self-attention mechanism, the vanishing of\nlong-term temporal dependencies is prevented. In addition, since our model\nrelies solely on attention mechanisms, the spatial and temporal dependencies\ncan be simultaneously measured. Experimental results on real-world data\ndemonstrate that, in comparison with state-of-the-art methods, our model\nreduces the root mean square errors by 9% in inflow prediction and 4% in\noutflow prediction on Taxi-NYC data, which is very significant compared to the\nprevious improvement.', 'Pandemic influenza has the epidemic potential to kill millions of people.\nWhile various preventive measures exist (i.a., vaccination and school\nclosures), deciding on strategies that lead to their most effective and\nefficient use remains challenging. To this end, individual-based\nepidemiological models are essential to assist decision makers in determining\nthe best strategy to curb epidemic spread. However, individual-based models are\ncomputationally intensive and it is therefore pivotal to identify the optimal\nstrategy using a minimal amount of model evaluations. Additionally, as\nepidemiological modeling experiments need to be planned, a computational budget\nneeds to be specified a priori. Consequently, we present a new sampling\ntechnique to optimize the evaluation of preventive strategies using fixed\nbudget best-arm identification algorithms. We use epidemiological modeling\ntheory to derive knowledge about the reward distribution which we exploit using\nBayesian best-arm identification algorithms (i.e., Top-two Thompson sampling\nand BayesGap). We evaluate these algorithms in a realistic experimental setting\nand demonstrate that it is possible to identify the optimal strategy using only\na limited number of model evaluations, i.e., 2-to-3 times faster compared to\nthe uniform sampling method, the predominant technique used for epidemiological\ndecision making in the literature. Finally, we contribute and evaluate a\nstatistic for Top-two Thompson sampling to inform the decision makers about the\nconfidence of an arm recommendation.', 'This paper describes a method for identification of the informative variables\nin the information system with discrete decision variables. It is targeted\nspecifically towards discovery of the variables that are non-informative when\nconsidered alone, but are informative when the synergistic interactions between\nmultiple variables are considered. To this end, the mutual entropy of all\npossible k-tuples of variables with decision variable is computed. Then, for\neach variable the maximal information gain due to interactions with other\nvariables is obtained. For non-informative variables this quantity conforms to\nthe well known statistical distributions. This allows for discerning truly\ninformative variables from non-informative ones. For demonstration of the\napproach, the method is applied to several synthetic datasets that involve\ncomplex multidimensional interactions between variables. It is capable of\nidentifying most important informative variables, even in the case when the\ndimensionality of the analysis is smaller than the true dimensionality of the\nproblem. What is more, the high sensitivity of the algorithm allows for\ndetection of the influence of nuisance variables on the response variable.', 'A factor-graph representation of quantum-mechanical probabilities (involving\nany number of measurements) is proposed. Unlike standard statistical models,\nthe proposed representation uses auxiliary variables (state variables) that are\nnot random variables. All joint probability distributions are marginals of some\ncomplex-valued function $q$, and it is demonstrated how the basic concepts of\nquantum mechanics relate to factorizations and marginals of $q$.', 'Traditional economic models typically treat private information, or signals,\nas generated from some underlying state. Recent work has explicated alternative\nmodels, where signals correspond to interpretations of available information.\nWe show that the difference between these formulations can be sharply cast in\nterms of causal dependence structure, and employ graphical models to illustrate\nthe distinguishing characteristics. The graphical representation supports\ninferences about signal patterns in the interpreted framework, and suggests how\nresults based on the generated model can be extended to more general\nsituations. Specific insights about bidding games in classical auction\nmechanisms derive from qualitative graphical models.', 'Bayesian inference on structured models typically relies on the ability to\ninfer posterior distributions of underlying hidden variables. However,\ninference in implicit models or complex posterior distributions is hard. A\npopular tool for learning implicit models are generative adversarial networks\n(GANs) which learn parameters of generators by fooling discriminators.\nTypically, GANs are considered to be models themselves and are not understood\nin the context of inference. Current techniques rely on inefficient global\ndiscrimination of joint distributions to perform learning, or only consider\ndiscriminating a single output variable. We overcome these limitations by\ntreating GANs as a basis for likelihood-free inference in generative models and\ngeneralize them to Bayesian posterior inference over factor graphs. We propose\nlocal learning rules based on message passing minimizing a global divergence\ncriterion involving cooperating local adversaries used to sidestep explicit\nlikelihood evaluations. This allows us to compose models and yields a unified\ninference and learning framework for adversarial learning. Our framework treats\nmodel specification and inference separately and facilitates richly structured\nmodels within the family of Directed Acyclic Graphs, including components such\nas intractable likelihoods, non-differentiable models, simulators and generally\ncumbersome models. A key result of our treatment is the insight that Bayesian\ninference on structured models can be performed only with sampling and\ndiscrimination when using nonparametric variational families, without access to\nexplicit distributions. As a side-result, we discuss the link to likelihood\nmaximization. These approaches hold promise to be useful in the toolbox of\nprobabilistic modelers and enrich the gamut of current probabilistic\nprogramming applications.', 'Certain causal models involving unmeasured variables induce no independence\nconstraints among the observed variables but imply, nevertheless, inequality\ncontraints on the observed distribution. This paper derives a general formula\nfor such instrumental variables, that is, exogenous variables that directly\naffect some variables but not all. With the help of this formula, it is\npossible to test whether a model involving instrumental variables may account\nfor the data, or, conversely, whether a given variables can be deemed\ninstrumental.', 'This paper concerns the probabilistic evaluation of the effects of actions in\nthe presence of unmeasured variables. We show that the identification of causal\neffect between a singleton variable X and a set of variables Y can be\naccomplished systematically, in time polynomial in the number of variables in\nthe graph. When the causal effect is identifiable, a closed-form expression can\nbe obtained for the probability that the action will achieve a specified goal,\nor a set of goals.', 'In recent years, there has been an increasing interest in studying\ncausality-related properties in machine learning models generally, and in\ngenerative models in particular. While that is well motivated, it inherits the\nfundamental computational hardness of probabilistic inference, making exact\nreasoning intractable. Probabilistic tractable models have also recently\nemerged, which guarantee that conditional marginals can be computed in time\nlinear in the size of the model, where the model is usually learned from data.\nAlthough initially limited to low tree-width models, recent tractable models\nsuch as sum product networks (SPNs) and probabilistic sentential decision\ndiagrams (PSDDs) exploit efficient function representations and also capture\nhigh tree-width models.\n  In this paper, we ask the following technical question: can we use the\ndistributions represented or learned by these models to perform causal queries,\nsuch as reasoning about interventions and counterfactuals? By appealing to some\nexisting ideas on transforming such models to Bayesian networks, we answer\nmostly in the negative. We show that when transforming SPNs to a causal graph\ninterventional reasoning reduces to computing marginal distributions; in other\nwords, only trivial causal reasoning is possible. For PSDDs the situation is\nonly slightly better. We first provide an algorithm for constructing a causal\ngraph from a PSDD, which introduces augmented variables. Intervening on the\noriginal variables, once again, reduces to marginal distributions, but when\nintervening on the augmented variables, a deterministic but nonetheless\ncausal-semantics can be provided for PSDDs.', 'We propose a method to classify the causal relationship between two discrete\nvariables given only the joint distribution of the variables, acknowledging\nthat the method is subject to an inherent baseline error. We assume that the\ncausal system is acyclicity, but we do allow for hidden common causes. Our\nalgorithm presupposes that the probability distributions $P(C)$ of a cause $C$\nis independent from the probability distribution $P(E\\mid C)$ of the\ncause-effect mechanism. While our classifier is trained with a Bayesian\nassumption of flat hyperpriors, we do not make this assumption about our test\ndata. This work connects to recent developments on the identifiability of\ncausal models over continuous variables under the assumption of ""independent\nmechanisms"". Carefully-commented Python notebooks that reproduce all our\nexperiments are available online at\nhttp://vision.caltech.edu/~kchalupk/code.html.', 'This work studies the widely adopted ancestral sampling algorithms for\nauto-regressive language models, which is not widely studied in the literature.\nWe use the quality-diversity (Q-D) trade-off to investigate three popular\nsampling algorithms (top-k, nucleus and tempered sampling). We focus on the\ntask of open-ended language generation. We first show that the existing\nsampling algorithms have similar performance. After carefully inspecting the\ntransformations defined by different sampling algorithms, we identify three key\nproperties that are shared among them: entropy reduction, order preservation,\nand slope preservation. To validate the importance of the identified\nproperties, we design two sets of new sampling algorithms: one set in which\neach algorithm satisfies all three properties, and one set in which each\nalgorithm violates at least one of the properties. We compare their performance\nwith existing sampling algorithms, and find that violating the identified\nproperties could lead to drastic performance degradation, as measured by the\nQ-D trade-off. On the other hand, we find that the set of sampling algorithms\nthat satisfies these properties performs on par with the existing sampling\nalgorithms. Our data and code are available at\nhttps://github.com/moinnadeem/characterizing-sampling-algorithms', ""Real-world tasks often exhibit a compositional structure that contains a\nsequence of simpler sub-tasks. For instance, opening a door requires reaching,\ngrasping, rotating, and pulling the door knob. Such compositional tasks require\nan agent to reason about the sub-task at hand while orchestrating global\nbehavior accordingly. This can be cast as an online task inference problem,\nwhere the current task identity, represented by a context variable, is\nestimated from the agent's past experiences with probabilistic inference.\nPrevious approaches have employed simple latent distributions, e.g., Gaussian,\nto model a single context for the entire task. However, this formulation lacks\nthe expressiveness to capture the composition and transition of the sub-tasks.\nWe propose a variational inference framework OCEAN to perform online task\ninference for compositional tasks. OCEAN models global and local context\nvariables in a joint latent space, where the global variables represent a\nmixture of sub-tasks required for the task, while the local variables capture\nthe transitions between the sub-tasks. Our framework supports flexible latent\ndistributions based on prior knowledge of the task structure and can be trained\nin an unsupervised manner. Experimental results show that OCEAN provides more\neffective task inference with sequential context adaptation and thus leads to a\nperformance boost on complex, multi-stage tasks."", 'The property of perfectness plays an important role in the theory of Bayesian\nnetworks. First, the existence of perfect distributions for arbitrary sets of\nvariables and directed acyclic graphs implies that various methods for reading\nindependence from the structure of the graph (e.g., Pearl, 1988; Lauritzen,\nDawid, Larsen & Leimer, 1990) are complete. Second, the asymptotic reliability\nof various search methods is guaranteed under the assumption that the\ngenerating distribution is perfect (e.g., Spirtes, Glymour & Scheines, 2000;\nChickering & Meek, 2002). We provide a lower-bound on the probability of\nsampling a non-perfect distribution when using a fixed number of bits to\nrepresent the parameters of the Bayesian network. This bound approaches zero\nexponentially fast as one increases the number of bits used to represent the\nparameters. This result implies that perfect distributions with fixed-length\nrepresentations exist. We also provide a lower-bound on the number of bits\nneeded to guarantee that a distribution sampled from a uniform Dirichlet\ndistribution is perfect with probability greater than 1/2. This result is\nuseful for constructing randomized reductions for hardness proofs.', 'Learning target side syntactic structure has been shown to improve Neural\nMachine Translation (NMT). However, incorporating syntax through latent\nvariables introduces additional complexity in inference, as the models need to\nmarginalize over the latent syntactic structures. To avoid this, models often\nresort to greedy search which only allows them to explore a limited portion of\nthe latent space. In this work, we introduce a new latent variable model,\nLaSyn, that captures the co-dependence between syntax and semantics, while\nallowing for effective and efficient inference over the latent space. LaSyn\ndecouples direct dependence between successive latent variables, which allows\nits decoder to exhaustively search through the latent syntactic choices, while\nkeeping decoding speed proportional to the size of the latent variable\nvocabulary. We implement LaSyn by modifying a transformer-based NMT system and\ndesign a neural expectation maximization algorithm that we regularize with\npart-of-speech information as the latent sequences. Evaluations on four\ndifferent MT tasks show that incorporating target side syntax with LaSyn\nimproves both translation quality, and also provides an opportunity to improve\ndiversity.', ""In order to automate verification process, regulatory rules written in\nnatural language need to be translated into a format that machines can\nunderstand. However, none of the existing formalisms can fully represent the\nelements that appear in legal norms. For instance, most of these formalisms do\nnot provide features to capture the behavior of deontic effects, which is an\nimportant aspect in automated compliance checking. This paper presents an\napproach for transforming legal norms represented using LegalRuleML to a\nvariant of Modal Defeasible Logic (and vice versa) such that a legal statement\nrepresented using LegalRuleML can be transformed into a machine-readable format\nthat can be understood and reasoned about depending upon the client's\npreferences."", 'In Recurrent Neural Networks (RNNs), encoding information in a suboptimal or\nerroneous way can impact the quality of representations based on later elements\nin the sequence and subsequently lead to wrong predictions and a worse model\nperformance. In humans, challenging cases like garden path sentences (an\ninstance of this being the infamous ""The horse raced past the barn fell"") can\nlead their language understanding astray. However, they are still able to\ncorrect their representation accordingly and recover when new information is\nencountered. Inspired by this, I propose an augmentation to standard RNNs in\nform of a gradient-based correction mechanism: This way I hope to enable such\nmodels to dynamically adapt their inner representation of a sentence, adding a\nway to correct deviations as soon as they occur. This could therefore lead to\nmore robust models using more flexible representations, even during inference\ntime.\n  I conduct different experiments in the context of language modeling, where\nthe impact of using such a mechanism is examined in detail. To this end, I look\nat modifications based on different kinds of time-dependent error signals and\nhow they influence the model performance. Furthermore, this work contains a\nstudy of the model\'s confidence in its predictions during training and for\nchallenging test samples and the effect of the manipulation thereof. Lastly, I\nalso study the difference in behavior of these novel models compared to a\nstandard LSTM baseline and investigate error cases in detail to identify points\nof future research. I show that while the proposed approach comes with\npromising theoretical guarantees and an appealing intuition, it is only able to\nproduce minor improvements over the baseline due to challenges in its practical\napplication and the efficacy of the tested model variants.', 'Spatio-temporal sequence forecasting is one of the fundamental tasks in\nspatio-temporal data mining. It facilitates many real world applications such\nas precipitation nowcasting, citywide crowd flow prediction and air pollution\nforecasting. Recently, a few Seq2Seq based approaches have been proposed, but\none of the drawbacks of Seq2Seq models is that, small errors can accumulate\nquickly along the generated sequence at the inference stage due to the\ndifferent distributions of training and inference phase. That is because\nSeq2Seq models minimise single step errors only during training, however the\nentire sequence has to be generated during the inference phase which generates\na discrepancy between training and inference. In this work, we propose a novel\ncurriculum learning based strategy named Temporal Progressive Growing Sampling\nto effectively bridge the gap between training and inference for\nspatio-temporal sequence forecasting, by transforming the training process from\na fully-supervised manner which utilises all available previous ground-truth\nvalues to a less-supervised manner which replaces some of the ground-truth\ncontext with generated predictions. To do that we sample the target sequence\nfrom midway outputs from intermediate models trained with bigger timescales\nthrough a carefully designed decaying strategy. Experimental results\ndemonstrate that our proposed method better models long term dependencies and\noutperforms baseline approaches on two competitive datasets.', 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', 'We consider the problem of parameter estimation using weakly supervised\ndatasets, where a training sample consists of the input and a partially\nspecified annotation, which we refer to as the output. The missing information\nin the annotation is modeled using latent variables. Previous methods\noverburden a single distribution with two separate tasks: (i) modeling the\nuncertainty in the latent variables during training; and (ii) making accurate\npredictions for the output and the latent variables during testing. We propose\na novel framework that separates the demands of the two tasks using two\ndistributions: (i) a conditional distribution to model the uncertainty of the\nlatent variables for a given input-output pair; and (ii) a delta distribution\nto predict the output and the latent variables for a given input. During\nlearning, we encourage agreement between the two distributions by minimizing a\nloss-based dissimilarity coefficient. Our approach generalizes latent SVM in\ntwo important ways: (i) it models the uncertainty over latent variables instead\nof relying on a pointwise estimate; and (ii) it allows the use of loss\nfunctions that depend on latent variables, which greatly increases its\napplicability. We demonstrate the efficacy of our approach on two challenging\nproblems---object detection and action detection---using publicly available\ndatasets.', 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.', ""Probability theory, epistemically interpreted, provides an excellent, if not\nthe best available account of inductive reasoning. This is so because there are\ngeneral and definite rules for the change of subjective probabilities through\ninformation or experience; induction and belief change are one and same topic,\nafter all. The most basic of these rules is simply to conditionalize with\nrespect to the information received; and there are similar and more general\nrules. 1 Hence, a fundamental reason for the epistemological success of\nprobability theory is that there at all exists a well-behaved concept of\nconditional probability. Still, people have, and have reasons for, various\nconcerns over probability theory. One of these is my starting point:\nIntuitively, we have the notion of plain belief; we believe propositions2 to be\ntrue (or to be false or neither). Probability theory, however, offers no formal\ncounterpart to this notion. Believing A is not the same as having probability 1\nfor A, because probability 1 is incorrigible3; but plain belief is clearly\ncorrigible. And believing A is not the same as giving A a probability larger\nthan some 1 - c, because believing A and believing B is usually taken to be\nequivalent to believing A & B.4 Thus, it seems that the formal representation\nof plain belief has to take a non-probabilistic route. Indeed, representing\nplain belief seems easy enough: simply represent an epistemic state by the set\nof all propositions believed true in it or, since I make the common assumption\nthat plain belief is deductively closed, by the conjunction of all propositions\nbelieved true in it. But this does not yet provide a theory of induction, i.e.\nan answer to the question how epistemic states so represented are changed\ntbrough information or experience. There is a convincing partial answer: if the\nnew information is compatible with the old epistemic state, then the new\nepistemic state is simply represented by the conjunction of the new information\nand the old beliefs. This answer is partial because it does not cover the quite\ncommon case where the new information is incompatible with the old beliefs. It\nis, however, important to complete the answer and to cover this case, too;\notherwise, we would not represent plain belief as conigible. The crucial\nproblem is that there is no good completion. When epistemic states are\nrepresented simply by the conjunction of all propositions believed true in it,\nthe answer cannot be completed; and though there is a lot of fruitful work, no\nother representation of epistemic states has been proposed, as far as I know,\nwhich provides a complete solution to this problem. In this paper, I want to\nsuggest such a solution. In [4], I have more fully argued that this is the only\nsolution, if certain plausible desiderata are to be satisfied. Here, in section\n2, I will be content with formally defining and intuitively explaining my\nproposal. I will compare my proposal with probability theory in section 3. It\nwill turn out that the theory I am proposing is structurally homomorphic to\nprobability theory in important respects and that it is thus equally easily\nimplementable, but moreover computationally simpler. Section 4 contains a very\nbrief comparison with various kinds of logics, in particular conditional logic,\nwith Shackle's functions of potential surprise and related theories, and with\nthe Dempster - Shafer theory of belief functions."", '(l) I have enough evidence to render the sentence S probable. (la) So,\nrelative to what I know, it is rational of me to believe S. (2) Now that I have\nmore evidence, S may no longer be probable. (2a) So now, relative to what I\nknow, it is not rational of me to believe S. These seem a perfectly ordinary,\ncommon sense, pair of situations. Generally and vaguely, I take them to embody\nwhat I shall call probabilistic inference. This form of inference is clearly\nnon-monotonic. Relatively few people have taken this form of inference, based\non high probability, to serve as a foundation for non-monotonic logic or for a\nlogical or defeasible inference. There are exceptions: Jane Nutter [16] thinks\nthat sometimes probability has something to do with non-monotonic reasoning.\nJudea Pearl [ 17] has recently been exploring the possibility. There are any\nnumber of people whom one might call probability enthusiasts who feel that\nprobability provides all the answers by itself, with no need of help from\nlogic. Cheeseman [1], Henrion [5] and others think it useful to look at a\ndistribution of probabilities over a whole algebra of statements, to update\nthat distribution in the light of new evidence, and to use the latest updated\ndistribution of probability over the algebra as a basis for planning and\ndecision making. A slightly weaker form of this approach is captured by Nilsson\n[15], where one assumes certain probabilities for certain statements, and\ninfers the probabilities, or constraints on the probabilities of other\nstatement. None of this corresponds to what I call probabilistic inference. All\nof the inference that is taking place, either in Bayesian updating, or in\nprobabilistic logic, is strictly deductive. Deductive inference, particularly\nthat concerned with the distribution of classical probabilities or chances, is\nof great importance. But this is not to say that there is no important role for\nwhat earlier logicians have called ""ampliative"" or ""inductive"" or ""scientific""\ninference, in which the conclusion goes beyond the premises, asserts more than\ndo the premises. This depends on what David Israel [6] has called ""real rules\nof inference"". It is characteristic of any such logic or inference procedure\nthat it can go wrong: that statements accepted at one point may be rejected at\na later point. Research underlying the results reported here has been partially\nsupported by the Signals Warfare Center of the United States Army.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I find that the primary constraint in modern artificial intelligence applications is the computational resources expended on multilayer perceptrons (MLP) within neural networks. To enhance the efficiency of MLP, I am seeking a method to distill MLP by eliminating non-essential weights. To accomplish this, I require a theoretical understanding of the weights in shallow MLPs and an empirical application of these theories to deeper neural networks. Furthermore, I need a metric that takes into account both the computational resources saved and the performance loss resulting from the distillation. I understand that the concept of sparsity could be beneficial in representing the number of edges in a graph, but I need guidance on how to apply this concept to MLPs, which can be viewed as a specific type of graph.","['In this work, we firstly apply the Train-Tensor (TT) networks to construct a\ncompact representation of the classical Multilayer Perceptron, representing a\nreduction of up to 95% of the coefficients. A comparative analysis between\ntensor model and standard multilayer neural networks is also carried out in the\ncontext of prediction of the Mackey-Glass noisy chaotic time series and NASDAQ\nindex. We show that the weights of a multidimensional regression model can be\nlearned by means of TT network and the optimization of TT weights is a more\nrobust to the impact of coefficient initialization and hyper-parameter setting.\nFurthermore, an efficient algorithm based on alternating least squares has been\nproposed for approximating the weights in TT-format with a reduction of\ncomputational calculus, providing a much faster convergence than the well-known\nadaptive learning-method algorithms, widely applied for optimizing neural\nnetworks.', 'The holy grail in deep neural network research is porting the memory- and\ncomputation-intensive network models on embedded platforms with a minimal\ncompromise in model accuracy. To this end, we propose a novel approach, termed\nas Variational Student, where we reap the benefits of compressibility of the\nknowledge distillation (KD) framework, and sparsity inducing abilities of\nvariational inference (VI) techniques. Essentially, we build a sparse student\nnetwork, whose sparsity is induced by the variational parameters found via\noptimizing a loss function based on VI, leveraging the knowledge learnt by an\naccurate but complex pre-trained teacher network. Further, for sparsity\nenhancement, we also employ a Block Sparse Regularizer on a concatenated tensor\nof teacher and student network weights. We demonstrate that the marriage of KD\nand the VI techniques inherits compression properties from the KD framework,\nand enhances levels of sparsity from the VI approach, with minimal compromise\nin the model accuracy. We benchmark our results on LeNet MLP and VGGNet (CNN)\nand illustrate a memory footprint reduction of 64x and 213x on these MLP and\nCNN variants, respectively, without a need to retrain the teacher network.\nFurthermore, in the low data regime, we observed that our method outperforms\nstate-of-the-art Bayesian techniques in terms of accuracy.', 'Deep neural networks (DNNs) have achieved extraordinary success in numerous\nareas. However, to attain this success, DNNs often carry a large number of\nweight parameters, leading to heavy costs of memory and computation resources.\nOverfitting is also likely to happen in such network when the training data are\ninsufficient. These shortcomings severely hinder the application of DNNs in\nresource-constrained platforms. In fact, many network weights are known to be\nredundant and can be removed from the network without much loss of performance.\nTo this end, we introduce a new non-convex integrated transformed $\\ell_1$\nregularizer to promote sparsity for DNNs, which removes both redundant\nconnections and unnecessary neurons simultaneously. To be specific, we apply\nthe transformed $\\ell_1$ to the matrix space of network weights and utilize it\nto remove redundant connections. Besides, group sparsity is also employed as an\nauxiliary to remove unnecessary neurons. An efficient stochastic proximal\ngradient algorithm is presented to solve the new model at the same time. To the\nbest of our knowledge, this is the first work to utilize a non-convex\nregularizer in sparse optimization based method to promote sparsity for DNNs.\nExperiments on several public datasets demonstrate the effectiveness of the\nproposed method.']","['Neural networks are the pinnacle of Artificial Intelligence, as in recent\nyears we witnessed many novel architectures, learning and optimization\ntechniques for deep learning. Capitalizing on the fact that neural networks\ninherently constitute multipartite graphs among neuron layers, we aim to\nanalyze directly their structure to extract meaningful information that can\nimprove the learning process. To our knowledge graph mining techniques for\nenhancing learning in neural networks have not been thoroughly investigated. In\nthis paper we propose an adapted version of the k-core structure for the\ncomplete weighted multipartite graph extracted from a deep learning\narchitecture. As a multipartite graph is a combination of bipartite graphs,\nthat are in turn the incidence graphs of hypergraphs, we design k-hypercore\ndecomposition, the hypergraph analogue of k-core degeneracy. We applied\nk-hypercore to several neural network architectures, more specifically to\nconvolutional neural networks and multilayer perceptrons for image recognition\ntasks after a very short pretraining. Then we used the information provided by\nthe hypercore numbers of the neurons to re-initialize the weights of the neural\nnetwork, thus biasing the gradient optimization scheme. Extensive experiments\nproved that k-hypercore outperforms the state-of-the-art initialization\nmethods.', 'Deep neural networks are over-parameterized, which implies that the number of\nparameters are much larger than the number of samples used to train the\nnetwork. Even in such a regime deep architectures do not overfit. This\nphenomenon is an active area of research and many theories have been proposed\ntrying to understand this peculiar observation. These include the Vapnik\nChervonenkis (VC) dimension bounds and Rademacher complexity bounds which show\nthat the capacity of the network is characterized by the norm of weights rather\nthan the number of parameters. However, the effect of input noise on these\nmeasures for shallow and deep architectures has not been studied. In this\npaper, we analyze the effects of various regularization schemes on the\ncomplexity of a neural network which we characterize with the loss, $L_2$ norm\nof the weights, Rademacher complexities (Directly Approximately Regularizing\nComplexity-DARC1), VC dimension based Low Complexity Neural Network (LCNN) when\nsubject to varying degrees of Gaussian input noise. We show that $L_2$\nregularization leads to a simpler hypothesis class and better generalization\nfollowed by DARC1 regularizer, both for shallow as well as deeper\narchitectures. Jacobian regularizer works well for shallow architectures with\nhigh level of input noises. Spectral normalization attains highest test set\naccuracies both for shallow and deeper architectures. We also show that Dropout\nalone does not perform well in presence of input noise. Finally, we show that\ndeeper architectures are robust to input noise as opposed to their shallow\ncounterparts.', 'Tabular datasets are the last ""unconquered castle"" for deep learning, with\ntraditional ML methods like Gradient-Boosted Decision Trees still performing\nstrongly even against recent specialized neural architectures. In this paper,\nwe hypothesize that the key to boosting the performance of neural networks lies\nin rethinking the joint and simultaneous application of a large set of modern\nregularization techniques. As a result, we propose regularizing plain\nMultilayer Perceptron (MLP) networks by searching for the optimal\ncombination/cocktail of 13 regularization techniques for each dataset using a\njoint optimization over the decision on which regularizers to apply and their\nsubsidiary hyperparameters. We empirically assess the impact of these\nregularization cocktails for MLPs in a large-scale empirical study comprising\n40 tabular datasets and demonstrate that (i) well-regularized plain MLPs\nsignificantly outperform recent state-of-the-art specialized neural network\narchitectures, and (ii) they even outperform strong traditional ML methods,\nsuch as XGBoost.', 'We initiate the first empirical study on the use of MLP architectures for\nvision-and-language (VL) fusion. Through extensive experiments on 5 VL tasks\nand 5 robust VQA benchmarks, we find that: (i) Without pre-training, using MLPs\nfor multimodal fusion has a noticeable performance gap compared to\ntransformers; (ii) However, VL pre-training can help close the performance gap;\n(iii) Instead of heavy multi-head attention, adding tiny one-head attention to\nMLPs is sufficient to achieve comparable performance to transformers. Moreover,\nwe also find that the performance gap between MLPs and transformers is not\nwidened when being evaluated on the harder robust VQA benchmarks, suggesting\nusing MLPs for VL fusion can generalize roughly to a similar degree as using\ntransformers. These results hint that MLPs can effectively learn to align\nvision and text features extracted from lower-level encoders without heavy\nreliance on self-attention. Based on this, we ask an even bolder question: can\nwe have an all-MLP architecture for VL modeling, where both VL fusion and the\nvision encoder are replaced with MLPs? Our result shows that an all-MLP VL\nmodel is sub-optimal compared to state-of-the-art full-featured VL models when\nboth of them get pre-trained. However, pre-training an all-MLP can surprisingly\nachieve a better average score than full-featured transformer models without\npre-training. This indicates the potential of large-scale pre-training of\nMLP-like architectures for VL modeling and inspires the future research\ndirection on simplifying well-established VL modeling with less inductive\ndesign bias. Our code is publicly available at:\nhttps://github.com/easonnie/mlp-vil', ""Deep neural networks have achieved increasingly accurate results on a wide\nvariety of complex tasks. However, much of this improvement is due to the\ngrowing use and availability of computational resources (e.g use of GPUs, more\nlayers, more parameters, etc). Most state-of-the-art deep networks, despite\nperforming well, over-parameterize approximate functions and take a significant\namount of time to train. With increased focus on deploying deep neural networks\non resource constrained devices like smart phones, there has been a push to\nevaluate why these models are so resource hungry and how they can be made more\nefficient. This work evaluates and compares three distinct methods for deep\nmodel compression and acceleration: weight pruning, low rank factorization, and\nknowledge distillation. Comparisons on VGG nets trained on CIFAR10 show that\neach of the models on their own are effective, but that the true power lies in\ncombining them. We show that by combining pruning and knowledge distillation\nmethods we can create a compressed network 85 times smaller than the original,\nall while retaining 96% of the original model's accuracy."", 'In recent years, deep neural networks (DNNs) have been applied to various\nmachine leaning tasks, including image recognition, speech recognition, and\nmachine translation. However, large DNN models are needed to achieve\nstate-of-the-art performance, exceeding the capabilities of edge devices. Model\nreduction is thus needed for practical use. In this paper, we point out that\ndeep learning automatically induces group sparsity of weights, in which all\nweights connected to an output channel (node) are zero, when training DNNs\nunder the following three conditions: (1) rectified-linear-unit (ReLU)\nactivations, (2) an $L_2$-regularized objective function, and (3) the Adam\noptimizer. Next, we analyze this behavior both theoretically and\nexperimentally, and propose a simple model reduction method: eliminate the zero\nweights after training the DNN. In experiments on MNIST and CIFAR-10 datasets,\nwe demonstrate the sparsity with various training setups. Finally, we show that\nour method can efficiently reduce the model size and performs well relative to\nmethods that use a sparsity-inducing regularizer.', 'The growing energy and performance costs of deep learning have driven the\ncommunity to reduce the size of neural networks by selectively pruning\ncomponents. Similarly to their biological counterparts, sparse networks\ngeneralize just as well, if not better than, the original dense networks.\nSparsity can reduce the memory footprint of regular networks to fit mobile\ndevices, as well as shorten training time for ever growing networks. In this\npaper, we survey prior work on sparsity in deep learning and provide an\nextensive tutorial of sparsification for both inference and training. We\ndescribe approaches to remove and add elements of neural networks, different\ntraining strategies to achieve model sparsity, and mechanisms to exploit\nsparsity in practice. Our work distills ideas from more than 300 research\npapers and provides guidance to practitioners who wish to utilize sparsity\ntoday, as well as to researchers whose goal is to push the frontier forward. We\ninclude the necessary background on mathematical methods in sparsification,\ndescribe phenomena such as early structure adaptation, the intricate relations\nbetween sparsity and the training process, and show techniques for achieving\nacceleration on real hardware. We also define a metric of pruned parameter\nefficiency that could serve as a baseline for comparison of different sparse\nnetworks. We close by speculating on how sparsity can improve future workloads\nand outline major open problems in the field.', ""While machine learning is traditionally a resource intensive task, embedded\nsystems, autonomous navigation and the vision of the Internet-of-Things fuel\nthe interest in resource efficient approaches. These approaches require a\ncarefully chosen trade-off between performance and resource consumption in\nterms of computation and energy. On top of this, it is crucial to treat\nuncertainty in a consistent manner in all but the simplest applications of\nmachine learning systems. In particular, a desideratum for any real-world\nsystem is to be robust in the presence of outliers and corrupted data, as well\nas being `aware' of its limits, i.e.\\ the system should maintain and provide an\nuncertainty estimate over its own predictions. These complex demands are among\nthe major challenges in current machine learning research and key to ensure a\nsmooth transition of machine learning technology into every day's applications.\nIn this article, we provide an overview of the current state of the art of\nmachine learning techniques facilitating these real-world requirements. First\nwe provide a comprehensive review of resource-efficiency in deep neural\nnetworks with focus on techniques for model size reduction, compression and\nreduced precision. These techniques can be applied during training or as\npost-processing and are widely used to reduce both computational complexity and\nmemory footprint. As most (practical) neural networks are limited in their ways\nto treat uncertainty, we contrast them with probabilistic graphical models,\nwhich readily serve these desiderata by means of probabilistic inference. In\nthat way, we provide an extensive overview of the current state-of-the-art of\nrobust and efficient machine learning for real-world systems."", ""Mixture-of-Experts (MoE), a conditional computation architecture, achieved\npromising performance by scaling local module (i.e. feed-forward network) of\ntransformer. However, scaling the cross-token module (i.e. self-attention) is\nchallenging due to the unstable training. This work proposes Sparse-MLP, an\nall-MLP model which applies sparsely-activated MLPs to cross-token modeling.\nSpecifically, in each Sparse block of our all-MLP model, we apply two stages of\nMoE layers: one with MLP experts mixing information within channels along image\npatch dimension, the other with MLP experts mixing information within patches\nalong the channel dimension. In addition, by proposing importance-score routing\nstrategy for MoE and redesigning the image representation shape, we further\nimprove our model's computational efficiency. Experimentally, we are more\ncomputation-efficient than Vision Transformers with comparable accuracy. Also,\nour models can outperform MLP-Mixer by 2.5\\% on ImageNet Top-1 accuracy with\nfewer parameters and computational cost. On downstream tasks, i.e. Cifar10 and\nCifar100, our models can still achieve better performance than baselines."", 'This work proposes a domain-informed neural network architecture for\nexperimental particle physics, using particle interaction localization with the\ntime-projection chamber (TPC) technology for dark matter research as an example\napplication. A key feature of the signals generated within the TPC is that they\nallow localization of particle interactions through a process called\nreconstruction. While multilayer perceptrons (MLPs) have emerged as a leading\ncontender for reconstruction in TPCs, such a black-box approach does not\nreflect prior knowledge of the underlying scientific processes. This paper\nlooks anew at neural network-based interaction localization and encodes prior\ndetector knowledge, in terms of both signal characteristics and detector\ngeometry, into the feature encoding and the output layers of a multilayer\nneural network. The resulting Domain-informed Neural Network (DiNN) limits the\nreceptive fields of the neurons in the initial feature encoding layers in order\nto account for the spatially localized nature of the signals produced within\nthe TPC. This aspect of the DiNN, which has similarities with the emerging area\nof graph neural networks in that the neurons in the initial layers only connect\nto a handful of neurons in their succeeding layer, significantly reduces the\nnumber of parameters in the network in comparison to an MLP. In addition, in\norder to account for the detector geometry, the output layers of the network\nare modified using two geometric transformations to ensure the DiNN produces\nlocalizations within the interior of the detector. The end result is a neural\nnetwork architecture that has 60% fewer parameters than an MLP, but that still\nachieves similar localization performance and provides a path to future\narchitectural developments with improved performance because of their ability\nto encode additional domain knowledge into the architecture.', ""Understanding the fundamental principles behind the massive success of neural\nnetworks is one of the most important open questions in deep learning. However,\ndue to the highly complex nature of the problem, progress has been relatively\nslow. In this note, through the lens of infinite-width networks, a.k.a. neural\nkernels, we present one such principle resulting from hierarchical localities.\nIt is well-known that the eigenstructure of infinite-width multilayer\nperceptrons (MLPs) depends solely on the concept frequency, which measures the\norder of interactions. We show that the topologies from deep convolutional\nnetworks (CNNs) restructure the associated eigenspaces into finer subspaces. In\naddition to frequency, the new structure also depends on the concept space,\nwhich measures the spatial distance among nonlinear interaction terms. The\nresulting fine-grained eigenstructure dramatically improves the network's\nlearnability, empowering them to simultaneously model a much richer class of\ninteractions, including Long-Range-Low-Frequency interactions,\nShort-Range-High-Frequency interactions, and various interpolations and\nextrapolations in-between. Additionally, model scaling can improve the\nresolutions of interpolations and extrapolations and, therefore, the network's\nlearnability. Finally, we prove a sharp characterization of the generalization\nerror for infinite-width CNNs of any depth in the high-dimensional setting. Two\ncorollaries follow: (1) infinite-width deep CNNs can break the curse of\ndimensionality without losing their expressivity, and (2) scaling improves\nperformance in both the finite and infinite data regimes."", 'We study how neural networks trained by gradient descent extrapolate, i.e.,\nwhat they learn outside the support of the training distribution. Previous\nworks report mixed empirical results when extrapolating with neural networks:\nwhile feedforward neural networks, a.k.a. multilayer perceptrons (MLPs), do not\nextrapolate well in certain simple tasks, Graph Neural Networks (GNNs) --\nstructured networks with MLP modules -- have shown some success in more complex\ntasks. Working towards a theoretical explanation, we identify conditions under\nwhich MLPs and GNNs extrapolate well. First, we quantify the observation that\nReLU MLPs quickly converge to linear functions along any direction from the\norigin, which implies that ReLU MLPs do not extrapolate most nonlinear\nfunctions. But, they can provably learn a linear target function when the\ntraining distribution is sufficiently ""diverse"". Second, in connection to\nanalyzing the successes and limitations of GNNs, these results suggest a\nhypothesis for which we provide theoretical and empirical evidence: the success\nof GNNs in extrapolating algorithmic tasks to new data (e.g., larger graphs or\nedge weights) relies on encoding task-specific non-linearities in the\narchitecture or features. Our theoretical analysis builds on a connection of\nover-parameterized networks to the neural tangent kernel. Empirically, our\ntheory holds across different training settings.', 'We propose a neural network based approach for extracting models from dynamic\ndata using ordinary and partial differential equations. In particular, given a\ntime-series or spatio-temporal dataset, we seek to identify an accurate\ngoverning system which respects the intrinsic differential structure. The\nunknown governing model is parameterized by using both (shallow) multilayer\nperceptrons and nonlinear differential terms, in order to incorporate relevant\ncorrelations between spatio-temporal samples. We demonstrate the approach on\nseveral examples where the data is sampled from various dynamical systems and\ngive a comparison to recurrent networks and other data-discovery methods. In\naddition, we show that for MNIST and Fashion MNIST, our approach lowers the\nparameter cost as compared to other deep neural networks.', 'Recently, sparse training methods have started to be established as a de\nfacto approach for training and inference efficiency in artificial neural\nnetworks. Yet, this efficiency is just in theory. In practice, everyone uses a\nbinary mask to simulate sparsity since the typical deep learning software and\nhardware are optimized for dense matrix operations. In this paper, we take an\northogonal approach, and we show that we can train truly sparse neural networks\nto harvest their full potential. To achieve this goal, we introduce three novel\ncontributions, specially designed for sparse neural networks: (1) a parallel\ntraining algorithm and its corresponding sparse implementation from scratch,\n(2) an activation function with non-trainable parameters to favour the gradient\nflow, and (3) a hidden neurons importance metric to eliminate redundancies. All\nin one, we are able to break the record and to train the largest neural network\never trained in terms of representational power -- reaching the bat brain size.\nThe results show that our approach has state-of-the-art performance while\nopening the path for an environmentally friendly artificial intelligence era.', ""Long training times of deep neural networks are a bottleneck in machine\nlearning research. The major impediment to fast training is the quadratic\ngrowth of both memory and compute requirements of dense and convolutional\nlayers with respect to their information bandwidth. Recently, training `a\npriori' sparse networks has been proposed as a method for allowing layers to\nretain high information bandwidth, while keeping memory and compute low.\nHowever, the choice of which sparse topology should be used in these networks\nis unclear. In this work, we provide a theoretical foundation for the choice of\nintra-layer topology. First, we derive a new sparse neural network\ninitialization scheme that allows us to explore the space of very deep sparse\nnetworks. Next, we evaluate several topologies and show that seemingly similar\ntopologies can often have a large difference in attainable accuracy. To explain\nthese differences, we develop a data-free heuristic that can evaluate a\ntopology independently from the dataset the network will be trained on. We then\nderive a set of requirements that make a good topology, and arrive at a single\ntopology that satisfies all of them."", ""Deep Neural Networks are highly over-parameterized and the size of the neural\nnetworks can be reduced significantly after training without any decrease in\nperformance. One can clearly see this phenomenon in a wide range of\narchitectures trained for various problems. Weight/channel pruning,\ndistillation, quantization, matrix factorization are some of the main methods\none can use to remove the redundancy to come up with smaller and faster models.\n  This work starts with a short informative chapter, where we motivate the\npruning idea and provide the necessary notation. In the second chapter, we\ncompare various saliency scores in the context of parameter pruning. Using the\ninsights obtained from this comparison and stating the problems it brings we\nmotivate why pruning units instead of the individual parameters might be a\nbetter idea. We propose some set of definitions to quantify and analyze units\nthat don't learn and create any useful information. We propose an efficient way\nfor detecting dead units and use it to select which units to prune. We get 5x\nmodel size reduction through unit-wise pruning on MNIST."", 'Deep learning based models are relatively large, and it is hard to deploy\nsuch models on resource-limited devices such as mobile phones and embedded\ndevices. One possible solution is knowledge distillation whereby a smaller\nmodel (student model) is trained by utilizing the information from a larger\nmodel (teacher model). In this paper, we present a survey of knowledge\ndistillation techniques applied to deep learning models. To compare the\nperformances of different techniques, we propose a new metric called\ndistillation metric. Distillation metric compares different knowledge\ndistillation algorithms based on sizes and accuracy scores. Based on the\nsurvey, some interesting conclusions are drawn and presented in this paper.', 'Recently, click-through rate (CTR) prediction models have evolved from\nshallow methods to deep neural networks. Most deep CTR models follow an\nEmbedding\\&MLP paradigm, that is, first mapping discrete id features, e.g. user\nvisited items, into low dimensional vectors with an embedding module, then\nlearn a multi-layer perception (MLP) to fit the target. In this way, embedding\nmodule performs as the representative learning and plays a key role in the\nmodel performance. However, in many real-world applications, deep CTR model\noften suffers from poor generalization performance, which is mostly due to the\nlearning of embedding parameters. In this paper, we model user behavior using\nan interest delay model, study carefully the embedding mechanism, and obtain\ntwo important results: (i) We theoretically prove that small aggregation radius\nof embedding vectors of items which belongs to a same user interest domain will\nresult in good generalization performance of deep CTR model. (ii) Following our\ntheoretical analysis, we design a new embedding structure named res-embedding.\nIn res-embedding module, embedding vector of each item is the sum of two\ncomponents: (i) a central embedding vector calculated from an item-based\ninterest graph (ii) a residual embedding vector with its scale to be relatively\nsmall. Empirical evaluation on several public datasets demonstrates the\neffectiveness of the proposed res-embedding structure, which brings significant\nimprovement on the model performance.', 'Among interpretable machine learning methods, the class of Generalised\nAdditive Neural Networks (GANNs) is referred to as Self-Explaining Neural\nNetworks (SENN) because of the linear dependence on explicit functions of the\ninputs. In binary classification this shows the precise weight that each input\ncontributes towards the logit. The nomogram is a graphical representation of\nthese weights. We show that functions of individual and pairs of variables can\nbe derived from a functional Analysis of Variance (ANOVA) representation,\nenabling an efficient feature selection to be carried by application of the\nlogistic Lasso. This process infers the structure of GANNs which otherwise\nneeds to be predefined. As this method is particularly suited for tabular data,\nit starts by fitting a generic flexible model, in this case a Multi-layer\nPerceptron (MLP) to which the ANOVA decomposition is applied. This has the\nfurther advantage that the resulting GANN can be replicated as a SENN, enabling\nfurther refinement of the univariate and bivariate component functions to take\nplace. The component functions are partial responses hence the SENN is a\npartial response network. The Partial Response Network (PRN) is equally as\ntransparent as a traditional logistic regression model, but capable of\nnon-linear classification with comparable or superior performance to the\noriginal MLP. In other words, the PRN is a fully interpretable representation\nof the MLP, at the level of univariate and bivariate effects. The performance\nof the PRN is shown to be competitive for benchmark data, against\nstate-of-the-art machine learning methods including GBM, SVM and Random\nForests. It is also compared with spline-based Sparse Additive Models (SAM)\nshowing that a semi-parametric representation of the GAM as a neural network\ncan be as effective as the SAM though less constrained by the need to set\nspline nodes.', 'Deep learning has redefined the field of artificial intelligence (AI) thanks\nto the rise of artificial neural networks, which are architectures inspired by\ntheir neurological counterpart in the brain. Through the years, this dualism\nbetween AI and neuroscience has brought immense benefits to both fields,\nallowing neural networks to be used in dozens of applications. These networks\nuse an efficient implementation of reverse differentiation, called\nbackpropagation (BP). This algorithm, however, is often criticized for its\nbiological implausibility (e.g., lack of local update rules for the\nparameters). Therefore, biologically plausible learning methods that rely on\npredictive coding (PC), a framework for describing information processing in\nthe brain, are increasingly studied. Recent works prove that these methods can\napproximate BP up to a certain margin on multilayer perceptrons (MLPs), and\nasymptotically on any other complex model, and that zero-divergence inference\nlearning (Z-IL), a variant of PC, is able to exactly implement BP on MLPs.\nHowever, the recent literature shows also that there is no biologically\nplausible method yet that can exactly replicate the weight update of BP on\ncomplex models. To fill this gap, in this paper, we generalize (PC and) Z-IL by\ndirectly defining them on computational graphs, and show that it can perform\nexact reverse differentiation. What results is the first biologically plausible\nalgorithm that is equivalent to BP in the way of updating parameters on any\nneural network, providing a bridge between the interdisciplinary research of\nneuroscience and deep learning.', 'The state-of-the-art deep neural networks (DNNs) have significant\ncomputational and data management requirements. The size of both training data\nand models continue to increase. Sparsification and pruning methods are shown\nto be effective in removing a large fraction of connections in DNNs. The\nresulting sparse networks present unique challenges to further improve the\ncomputational efficiency of training and inference in deep learning. Both the\nfeedforward (inference) and backpropagation steps in stochastic gradient\ndescent (SGD) algorithm for training sparse DNNs involve consecutive sparse\nmatrix-vector multiplications (SpMVs). We first introduce a distributed-memory\nparallel SpMV-based solution for the SGD algorithm to improve its scalability.\nThe parallelization approach is based on row-wise partitioning of weight\nmatrices that represent neuron connections between consecutive layers. We then\npropose a novel hypergraph model for partitioning weight matrices to reduce the\ntotal communication volume and ensure computational load-balance among\nprocessors. Experiments performed on sparse DNNs demonstrate that the proposed\nsolution is highly efficient and scalable. By utilizing the proposed matrix\npartitioning scheme, the performance of our solution is further improved\nsignificantly.', 'To put a state-of-the-art neural network to practical use, it is necessary to\ndesign a model that has a good trade-off between the resource consumption and\nperformance on the test set. Many researchers and engineers are developing\nmethods that enable training or designing a model more efficiently. Developing\nan efficient model includes several strategies such as network architecture\nsearch, pruning, quantization, knowledge distillation, utilizing cheap\nconvolution, regularization, and also includes any craft that leads to a better\nperformance-resource trade-off. When combining these technologies together, it\nwould be ideal if one source of performance improvement does not conflict with\nothers. We call this property as the orthogonality in model efficiency. In this\npaper, we focus on knowledge distillation and demonstrate that knowledge\ndistillation methods are orthogonal to other efficiency-enhancing methods both\nanalytically and empirically. Analytically, we claim that knowledge\ndistillation functions analogous to a ensemble method, bootstrap aggregating.\nThis analytical explanation is provided from the perspective of implicit data\naugmentation property of knowledge distillation. Empirically, we verify\nknowledge distillation as a powerful apparatus for practical deployment of\nefficient neural network, and also introduce ways to integrate it with other\nmethods effectively.', 'Pruning coupled with learning aims to optimize the neural network (NN)\nstructure for solving specific problems. This optimization can be used for\nvarious purposes: to prevent overfitting, to save resources for implementation\nand training, to provide explainability of the trained NN, and many others. The\nminimal structure that cannot be pruned further is not unique. Ensemble of\nminimal structures can be used as a committee of intellectual agents that\nsolves problems by voting. Each minimal NN presents an ""empirical knowledge""\nabout the problem and can be verbalized. The non-uniqueness of such knowledge\nextracted from data is an important property of data-driven Artificial\nIntelligence (AI). In this work, we review an approach to pruning based on the\nprinciple: What controls training should control pruning. This principle is\nexpected to work both for artificial NN and for selection and modification of\nimportant synaptic contacts in brain. In back-propagation artificial NN\nlearning is controlled by the gradient of loss functions. Therefore, the first\norder sensitivity indicators are used for pruning and the algorithms based on\nthese indicators are reviewed. The notion of logically transparent NN was\nintroduced. The approach was illustrated on the problem of political\nforecasting: predicting the results of the US presidential election. Eight\nminimal NN were produced that give different forecasting algorithms. The\nnon-uniqueness of solution can be utilised by creation of expert panels\n(committee). Another use of NN pluralism is to identify areas of input signals\nwhere further data collection is most useful. In Conclusion, we discuss the\npossible future of widely advertised XAI program.', 'Deep learning is finding its way into the embedded world with applications\nsuch as autonomous driving, smart sensors and aug- mented reality. However, the\ncomputation of deep neural networks is demanding in energy, compute power and\nmemory. Various approaches have been investigated to reduce the necessary\nresources, one of which is to leverage the sparsity occurring in deep neural\nnetworks due to the high levels of redundancy in the network parameters. It has\nbeen shown that sparsity can be promoted specifically and the achieved sparsity\ncan be very high. But in many cases the methods are evaluated on rather small\ntopologies. It is not clear if the results transfer onto deeper topologies. In\nthis paper, the TensorQuant toolbox has been extended to offer a platform to\ninvestigate sparsity, especially in deeper models. Several practical relevant\ntopologies for varying classification problem sizes are investigated to show\nthe differences in sparsity for activations, weights and gradients.', 'Neural network models are widely used in solving many challenging problems,\nsuch as computer vision, personalized recommendation, and natural language\nprocessing. Those models are very computationally intensive and reach the\nhardware limit of the existing server and IoT devices. Thus, finding better\nmodel architectures with much less amount of computation while maximally\npreserving the accuracy is a popular research topic. Among various mechanisms\nthat aim to reduce the computation complexity, identifying the zero values in\nthe model weights and in the activations to avoid computing them is a promising\ndirection.\n  In this paper, we summarize the current status of the research on the\ncomputation of sparse neural networks, from the perspective of the sparse\nalgorithms, the software frameworks, and the hardware accelerations. We observe\nthat the search for the sparse structure can be a general methodology for\nhigh-quality model explorations, in addition to a strategy for high-efficiency\nmodel execution. We discuss the model accuracy influenced by the number of\nweight parameters and the structure of the model. The corresponding models are\ncalled to be located in the weight dominated and structure dominated regions,\nrespectively. We show that for practically complicated problems, it is more\nbeneficial to search large and sparse models in the weight dominated region. In\norder to achieve the goal, new approaches are required to search for proper\nsparse structures, and new sparse training hardware needs to be developed to\nfacilitate fast iterations of sparse models.', 'The success of convolutional neural networks (CNNs) in various applications\nis accompanied by a significant increase in computation and parameter storage\ncosts. Recent efforts to reduce these overheads involve pruning and compressing\nthe weights of various layers while at the same time aiming to not sacrifice\nperformance. In this paper, we propose a novel criterion for CNN pruning\ninspired by neural network interpretability: The most relevant units, i.e.\nweights or filters, are automatically found using their relevance scores\nobtained from concepts of explainable AI (XAI). By exploring this idea, we\nconnect the lines of interpretability and model compression research. We show\nthat our proposed method can efficiently prune CNN models in transfer-learning\nsetups in which networks pre-trained on large corpora are adapted to\nspecialized tasks. The method is evaluated on a broad range of computer vision\ndatasets. Notably, our novel criterion is not only competitive or better\ncompared to state-of-the-art pruning criteria when successive retraining is\nperformed, but clearly outperforms these previous criteria in the\nresource-constrained application scenario in which the data of the task to be\ntransferred to is very scarce and one chooses to refrain from fine-tuning. Our\nmethod is able to compress the model iteratively while maintaining or even\nimproving accuracy. At the same time, it has a computational cost in the order\nof gradient computation and is comparatively simple to apply without the need\nfor tuning hyperparameters for pruning.', 'To address the large model size and intensive computation requirement of deep\nneural networks (DNNs), weight pruning techniques have been proposed and\ngenerally fall into two categories, i.e., static regularization-based pruning\nand dynamic regularization-based pruning. However, the former method currently\nsuffers either complex workloads or accuracy degradation, while the latter one\ntakes a long time to tune the parameters to achieve the desired pruning rate\nwithout accuracy loss. In this paper, we propose a unified DNN weight pruning\nframework with dynamically updated regularization terms bounded by the\ndesignated constraint, which can generate both non-structured sparsity and\ndifferent kinds of structured sparsity. We also extend our method to an\nintegrated framework for the combination of different DNN compression tasks.', 'Multilayer-perceptrons (MLP) are known to struggle with learning functions of\nhigh-frequencies, and in particular cases with wide frequency bands. We present\na spatially adaptive progressive encoding (SAPE) scheme for input signals of\nMLP networks, which enables them to better fit a wide range of frequencies\nwithout sacrificing training stability or requiring any domain specific\npreprocessing. SAPE gradually unmasks signal components with increasing\nfrequencies as a function of time and space. The progressive exposure of\nfrequencies is monitored by a feedback loop throughout the neural optimization\nprocess, allowing changes to propagate at different rates among local spatial\nportions of the signal space. We demonstrate the advantage of SAPE on a variety\nof domains and applications, including regression of low dimensional signals\nand images, representation learning of occupancy networks, and a geometric task\nof mesh transfer between 3D shapes.', 'We consider infinitely-wide multi-layer perceptrons (MLPs) which are limits\nof standard deep feed-forward neural networks. We assume that, for each layer,\nthe weights of an MLP are initialized with i.i.d. samples from either a\nlight-tailed (finite variance) or heavy-tailed distribution in the domain of\nattraction of a symmetric $\\alpha$-stable distribution, where $\\alpha\\in(0,2]$\nmay depend on the layer. For the bias terms of the layer, we assume i.i.d.\ninitializations with a symmetric $\\alpha$-stable distribution having the same\n$\\alpha$ parameter of that layer. We then extend a recent result of Favaro,\nFortini, and Peluchetti (2020), to show that the vector of pre-activation\nvalues at all nodes of a given hidden layer converges in the limit, under a\nsuitable scaling, to a vector of i.i.d. random variables with symmetric\n$\\alpha$-stable distributions.', 'We show that passing input points through a simple Fourier feature mapping\nenables a multilayer perceptron (MLP) to learn high-frequency functions in\nlow-dimensional problem domains. These results shed light on recent advances in\ncomputer vision and graphics that achieve state-of-the-art results by using\nMLPs to represent complex 3D objects and scenes. Using tools from the neural\ntangent kernel (NTK) literature, we show that a standard MLP fails to learn\nhigh frequencies both in theory and in practice. To overcome this spectral\nbias, we use a Fourier feature mapping to transform the effective NTK into a\nstationary kernel with a tunable bandwidth. We suggest an approach for\nselecting problem-specific Fourier features that greatly improves the\nperformance of MLPs for low-dimensional regression tasks relevant to the\ncomputer vision and graphics communities.', 'Interpretability of neural networks and their underlying theoretical\nbehaviour remain an open field of study even after the great success of their\npractical applications, particularly with the emergence of deep learning. In\nthis work, NN2Poly is proposed: a theoretical approach to obtain an explicit\npolynomial model that provides an accurate representation of an already trained\nfully-connected feed-forward artificial neural network (a multilayer perceptron\nor MLP). This approach extends a previous idea proposed in the literature,\nwhich was limited to single hidden layer networks, to work with arbitrarily\ndeep MLPs in both regression and classification tasks. The objective of this\npaper is to achieve this by using a Taylor expansion on the activation\nfunction, at each layer, and then using several combinatorial properties to\ncalculate the coefficients of the desired polynomials. Discussion is presented\non the main computational challenges of this method, and the way to overcome\nthem by imposing certain constraints during the training phase. Finally,\nsimulation experiments as well as an application to a real data set are\npresented to demonstrate the effectiveness of the proposed method.', ""Through the success of deep learning in various domains, artificial neural\nnetworks are currently among the most used artificial intelligence methods.\nTaking inspiration from the network properties of biological neural networks\n(e.g. sparsity, scale-freeness), we argue that (contrary to general practice)\nartificial neural networks, too, should not have fully-connected layers. Here\nwe propose sparse evolutionary training of artificial neural networks, an\nalgorithm which evolves an initial sparse topology (Erd\\H{o}s-R\\'enyi random\ngraph) of two consecutive layers of neurons into a scale-free topology, during\nlearning. Our method replaces artificial neural networks fully-connected layers\nwith sparse ones before training, reducing quadratically the number of\nparameters, with no decrease in accuracy. We demonstrate our claims on\nrestricted Boltzmann machines, multi-layer perceptrons, and convolutional\nneural networks for unsupervised and supervised learning on 15 datasets. Our\napproach has the potential to enable artificial neural networks to scale up\nbeyond what is currently possible."", ""Deep Learning using the eponymous deep neural networks (DNNs) has become an\nattractive approach towards various data-based problems of theoretical physics\nin the past decade. There has been a clear trend to deeper architectures\ncontaining increasingly more powerful and involved layers. Contrarily, Taylor\ncoefficients of DNNs still appear mainly in the light of interpretability\nstudies, where they are computed at most to first order. However, especially in\ntheoretical physics numerous problems benefit from accessing higher orders, as\nwell. This gap motivates a general formulation of neural network (NN) Taylor\nexpansions. Restricting our analysis to multilayer perceptrons (MLPs) and\nintroducing quantities we refer to as propagators and vertices, both depending\non the MLP's weights and biases, we establish a graph-theoretical approach.\nSimilarly to Feynman rules in quantum field theories, we can systematically\nassign diagrams containing propagators and vertices to the corresponding\npartial derivative. Examining this approach for S-wave scattering lengths of\nshallow potentials, we observe NNs to adapt their derivatives mainly to the\nleading order of the target function's Taylor expansion. To circumvent this\nproblem, we propose an iterative NN perturbation theory. During each iteration\nwe eliminate the leading order, such that the next-to-leading order can be\nfaithfully learned during the subsequent iteration. After performing two\niterations, we find that the first- and second-order Born terms are correctly\nadapted during the respective iterations. Finally, we combine both results to\nfind a proxy that acts as a machine-learned second-order Born approximation."", 'From the perspective of expressive power, this work compares multi-layer\nGraph Neural Networks (GNNs) with a simplified alternative that we call\nGraph-Augmented Multi-Layer Perceptrons (GA-MLPs), which first augments node\nfeatures with certain multi-hop operators on the graph and then applies an MLP\nin a node-wise fashion. From the perspective of graph isomorphism testing, we\nshow both theoretically and numerically that GA-MLPs with suitable operators\ncan distinguish almost all non-isomorphic graphs, just like the\nWeifeiler-Lehman (WL) test. However, by viewing them as node-level functions\nand examining the equivalence classes they induce on rooted graphs, we prove a\nseparation in expressive power between GA-MLPs and GNNs that grows\nexponentially in depth. In particular, unlike GNNs, GA-MLPs are unable to count\nthe number of attributed walks. We also demonstrate via community detection\nexperiments that GA-MLPs can be limited by their choice of operator family, as\ncompared to GNNs with higher flexibility in learning.', 'This paper aims to provide a selective survey about knowledge\ndistillation(KD) framework for researchers and practitioners to take advantage\nof it for developing new optimized models in the deep neural network field. To\nthis end, we give a brief overview of knowledge distillation and some related\nworks including learning using privileged information(LUPI) and generalized\ndistillation(GD). Even though knowledge distillation based on the\nteacher-student architecture was initially devised as a model compression\ntechnique, it has found versatile applications over various frameworks.\n  In this paper, we review the characteristics of knowledge distillation from\nthe hypothesis that the three important ingredients of knowledge distillation\nare distilled knowledge and loss,teacher-student paradigm, and the distillation\nprocess. In addition, we survey the versatility of the knowledge distillation\nby studying its direct applications and its usage in combination with other\ndeep learning paradigms. Finally we present some future works in knowledge\ndistillation including explainable knowledge distillation where the analytical\nanalysis of the performance gain is studied and the self-supervised learning\nwhich is a hot research topic in deep learning community.', 'Hypernetworks are neural networks that generate weights for another neural\nnetwork. We formulate the hypernetwork training objective as a compromise\nbetween accuracy and diversity, where the diversity takes into account trivial\nsymmetry transformations of the target network. We explain how this simple\nformulation generalizes variational inference. We use multi-layered perceptrons\nto form the mapping from the low dimensional input random vector to the high\ndimensional weight space, and demonstrate how to reduce the number of\nparameters in this mapping by parameter sharing. We perform experiments and\nshow that the generated weights are diverse and lie on a non-trivial manifold.', 'A closed-form solution exists in two-class linear discriminant analysis\n(LDA), which discriminates two Gaussian-distributed classes in a\nmulti-dimensional feature space. In this work, we interpret the multilayer\nperceptron (MLP) as a generalization of a two-class LDA system so that it can\nhandle an input composed by multiple Gaussian modalities belonging to multiple\nclasses. Besides input layer $l_{in}$ and output layer $l_{out}$, the MLP of\ninterest consists of two intermediate layers, $l_1$ and $l_2$. We propose a\nfeedforward design that has three stages: 1) from $l_{in}$ to $l_1$: half-space\npartitionings accomplished by multiple parallel LDAs, 2) from $l_1$ to $l_2$:\nsubspace isolation where one Gaussian modality is represented by one neuron, 3)\nfrom $l_2$ to $l_{out}$: class-wise subspace mergence, where each Gaussian\nmodality is connected to its target class. Through this process, we present an\nautomatic MLP design that can specify the network architecture (i.e., the layer\nnumber and the neuron number at a layer) and all filter weights in a\nfeedforward one-pass fashion. This design can be generalized to an arbitrary\ndistribution by leveraging the Gaussian mixture model (GMM). Experiments are\nconducted to compare the performance of the traditional backpropagation-based\nMLP (BP-MLP) and the new feedforward MLP (FF-MLP).', 'This paper proposes a set of new error criteria and learning approaches,\nAdaptive Normalized Risk-Averting Training (ANRAT), to attack the non-convex\noptimization problem in training deep neural networks (DNNs). Theoretically, we\ndemonstrate its effectiveness on global and local convexity lower-bounded by\nthe standard $L_p$-norm error. By analyzing the gradient on the convexity index\n$\\lambda$, we explain the reason why to learn $\\lambda$ adaptively using\ngradient descent works. In practice, we show how this method improves training\nof deep neural networks to solve visual recognition tasks on the MNIST and\nCIFAR-10 datasets. Without using pretraining or other tricks, we obtain results\ncomparable or superior to those reported in recent literature on the same tasks\nusing standard ConvNets + MSE/cross entropy. Performance on deep/shallow\nmultilayer perceptrons and Denoised Auto-encoders is also explored. ANRAT can\nbe combined with other quasi-Newton training methods, innovative network\nvariants, regularization techniques and other specific tricks in DNNs. Other\nthan unsupervised pretraining, it provides a new perspective to address the\nnon-convex optimization problem in DNNs.', 'The power law has been observed in the degree distributions of many\nbiological neural networks. Sparse deep neural networks, which learn an\neconomical representation from the data, resemble biological neural networks in\nmany ways. In this paper, we study if these artificial networks also exhibit\nproperties of the power law. Experimental results on two popular deep learning\nmodels, namely, multilayer perceptrons and convolutional neural networks, are\naffirmative. The power law is also naturally related to preferential\nattachment. To study the dynamical properties of deep networks in continual\nlearning, we propose an internal preferential attachment model to explain how\nthe network topology evolves. Experimental results show that with the arrival\nof a new task, the new connections made follow this preferential attachment\nprocess.', 'A fundamental task for artificial intelligence is learning. Deep Neural\nNetworks have proven to cope perfectly with all learning paradigms, i.e.\nsupervised, unsupervised, and reinforcement learning. Nevertheless, traditional\ndeep learning approaches make use of cloud computing facilities and do not\nscale well to autonomous agents with low computational resources. Even in the\ncloud, they suffer from computational and memory limitations, and they cannot\nbe used to model adequately large physical worlds for agents which assume\nnetworks with billions of neurons. These issues are addressed in the last few\nyears by the emerging topic of sparse training, which trains sparse networks\nfrom scratch. This paper discusses sparse training state-of-the-art, its\nchallenges and limitations while introducing a couple of new theoretical\nresearch directions which has the potential of alleviating sparse training\nlimitations to push deep learning scalability well beyond its current\nboundaries. Nevertheless, the theoretical advancements impact in complex\nmulti-agents settings is discussed from a real-world perspective, using the\nsmart grid case study.', 'Graph Neural Networks (GNNs) have already been widely applied in various\ngraph mining tasks. However, they suffer from the shallow architecture issue,\nwhich is the key impediment that hinders the model performance improvement.\nAlthough several relevant approaches have been proposed, none of the existing\nstudies provides an in-depth understanding of the root causes of performance\ndegradation in deep GNNs. In this paper, we conduct the first systematic\nexperimental evaluation to present the fundamental limitations of shallow\narchitectures. Based on the experimental results, we answer the following two\nessential questions: (1) what actually leads to the compromised performance of\ndeep GNNs; (2) when we need and how to build deep GNNs. The answers to the\nabove questions provide empirical insights and guidelines for researchers to\ndesign deep and well-performed GNNs. To show the effectiveness of our proposed\nguidelines, we present Deep Graph Multi-Layer Perceptron (DGMLP), a powerful\napproach (a paradigm in its own right) that helps guide deep GNN designs.\nExperimental results demonstrate three advantages of DGMLP: 1) high accuracy --\nit achieves state-of-the-art node classification performance on various\ndatasets; 2) high flexibility -- it can flexibly choose different propagation\nand transformation depths according to graph size and sparsity; 3) high\nscalability and efficiency -- it supports fast training on large-scale graphs.\nOur code is available in https://github.com/zwt233/DGMLP.', 'Deep Neural Networks (DNNs) are the key to the state-of-the-art machine\nvision, sensor fusion and audio/video signal processing. Unfortunately, their\ncomputation complexity and tight resource constraints on the Edge make them\nhard to leverage on mobile, embedded and IoT devices. Due to great diversity of\nEdge devices, DNN designers have to take into account the hardware platform and\napplication requirements during network training. In this work we introduce\npruning via matrix pivoting as a way to improve network pruning by compromising\nbetween the design flexibility of architecture-oblivious and performance\nefficiency of architecture-aware pruning, the two dominant techniques for\nobtaining resource-efficient DNNs. We also describe local and global network\noptimization techniques for efficient implementation of the resulting pruned\nnetworks. In combination, the proposed pruning and implementation result in\nclose to linear speed up with the reduction of network coefficients during\npruning.', 'Many efforts have been made to facilitate natural language processing tasks\nwith pre-trained language models (LMs), and brought significant improvements to\nvarious applications. To fully leverage the nearly unlimited corpora and\ncapture linguistic information of multifarious levels, large-size LMs are\nrequired; but for a specific task, only parts of these information are useful.\nSuch large-sized LMs, even in the inference stage, may cause heavy computation\nworkloads, making them too time-consuming for large-scale applications. Here we\npropose to compress bulky LMs while preserving useful information with regard\nto a specific task. As different layers of the model keep different\ninformation, we develop a layer selection method for model pruning using\nsparsity-inducing regularization. By introducing the dense connectivity, we can\ndetach any layer without affecting others, and stretch shallow and wide LMs to\nbe deep and narrow. In model training, LMs are learned with layer-wise dropouts\nfor better robustness. Experiments on two benchmark datasets demonstrate the\neffectiveness of our method.', 'Convolutional Neural Networks (CNNs) are the go-to model for computer vision.\nRecently, attention-based networks, such as the Vision Transformer, have also\nbecome popular. In this paper we show that while convolutions and attention are\nboth sufficient for good performance, neither of them are necessary. We present\nMLP-Mixer, an architecture based exclusively on multi-layer perceptrons (MLPs).\nMLP-Mixer contains two types of layers: one with MLPs applied independently to\nimage patches (i.e. ""mixing"" the per-location features), and one with MLPs\napplied across patches (i.e. ""mixing"" spatial information). When trained on\nlarge datasets, or with modern regularization schemes, MLP-Mixer attains\ncompetitive scores on image classification benchmarks, with pre-training and\ninference cost comparable to state-of-the-art models. We hope that these\nresults spark further research beyond the realms of well established CNNs and\nTransformers.', ""Federated learning is an emerging technique used to prevent the leakage of\nprivate information. Unlike centralized learning that needs to collect data\nfrom users and store them collectively on a cloud server, federated learning\nmakes it possible to learn a global model while the data are distributed on the\nusers' devices. However, compared with the traditional centralized approach,\nthe federated setting consumes considerable communication resources of the\nclients, which is indispensable for updating global models and prevents this\ntechnique from being widely used. In this paper, we aim to optimize the\nstructure of the neural network models in federated learning using a\nmulti-objective evolutionary algorithm to simultaneously minimize the\ncommunication costs and the global model test errors. A scalable method for\nencoding network connectivity is adapted to federated learning to enhance the\nefficiency in evolving deep neural networks. Experimental results on both\nmultilayer perceptrons and convolutional neural networks indicate that the\nproposed optimization method is able to find optimized neural network models\nthat can not only significantly reduce communication costs but also improve the\nlearning performance of federated learning compared with the standard fully\nconnected neural networks."", 'In the context of learning to map an input $I$ to a function\n$h_I:\\mathcal{X}\\to \\mathbb{R}$, two alternative methods are compared: (i) an\nembedding-based method, which learns a fixed function in which $I$ is encoded\nas a conditioning signal $e(I)$ and the learned function takes the form $h_I(x)\n= q(x,e(I))$, and (ii) hypernetworks, in which the weights $\\theta_I$ of the\nfunction $h_I(x) = g(x;\\theta_I)$ are given by a hypernetwork $f$ as\n$\\theta_I=f(I)$. In this paper, we define the property of modularity as the\nability to effectively learn a different function for each input instance $I$.\nFor this purpose, we adopt an expressivity perspective of this property and\nextend the theory of Devore et al. 1996 and provide a lower bound on the\ncomplexity (number of trainable parameters) of neural networks as function\napproximators, by eliminating the requirements for the approximation method to\nbe robust. Our results are then used to compare the complexities of $q$ and\n$g$, showing that under certain conditions and when letting the functions $e$\nand $f$ be as large as we wish, $g$ can be smaller than $q$ by orders of\nmagnitude. This sheds light on the modularity of hypernetworks in comparison\nwith the embedding-based method. Besides, we show that for a structured target\nfunction, the overall number of trainable parameters in a hypernetwork is\nsmaller by orders of magnitude than the number of trainable parameters of a\nstandard neural network and an embedding method.', 'The construction of a multilayer perceptron (MLP) as a piecewise low-order\npolynomial approximator using a signal processing approach is presented in this\nwork. The constructed MLP contains one input, one intermediate and one output\nlayers. Its construction includes the specification of neuron numbers and all\nfilter weights. Through the construction, a one-to-one correspondence between\nthe approximation of an MLP and that of a piecewise low-order polynomial is\nestablished. Comparison between piecewise polynomial and MLP approximations is\nmade. Since the approximation capability of piecewise low-order polynomials is\nwell understood, our findings shed light on the universal approximation\ncapability of an MLP.', 'In this paper, we propose a probabilistic representation of MultiLayer\nPerceptrons (MLPs) to improve the information-theoretic interpretability. Above\nall, we demonstrate that the activations being i.i.d. is not valid for all the\nhidden layers of MLPs, thus the existing mutual information estimators based on\nnon-parametric inference methods, e.g., empirical distributions and Kernel\nDensity Estimate (KDE), are invalid for measuring the information flow in MLPs.\nMoreover, we introduce explicit probabilistic explanations for MLPs: (i) we\ndefine the probability space (Omega_F, t, P_F) for a fully connected layer f\nand demonstrate the great effect of an activation function on the probability\nmeasure P_F ; (ii) we prove the entire architecture of MLPs as a Gibbs\ndistribution P; and (iii) the back-propagation aims to optimize the sample\nspace Omega_F of all the fully connected layers of MLPs for learning an optimal\nGibbs distribution P* to express the statistical connection between the input\nand the label. Based on the probabilistic explanations for MLPs, we improve the\ninformation-theoretic interpretability of MLPs in three aspects: (i) the random\nvariable of f is discrete and the corresponding entropy is finite; (ii) the\ninformation bottleneck theory cannot correctly explain the information flow in\nMLPs if we take into account the back-propagation; and (iii) we propose novel\ninformation-theoretic explanations for the generalization of MLPs. Finally, we\ndemonstrate the proposed probabilistic representation and information-theoretic\nexplanations for MLPs in a synthetic dataset and benchmark datasets.', 'The history of deep learning has shown that human-designed problem-specific\nnetworks can greatly improve the classification performance of general neural\nmodels. In most practical cases, however, choosing the optimal architecture for\na given task remains a challenging problem. Recent architecture-search methods\nare able to automatically build neural models with strong performance but fail\nto fully appreciate the interaction between neural architecture and weights.\nThis work investigates the problem of disentangling the role of the neural\nstructure and its edge weights, by showing that well-trained architectures may\nnot need any link-specific fine-tuning of the weights. We compare the\nperformance of such weight-free networks (in our case these are binary networks\nwith {0, 1}-valued weights) with random, weight-agnostic, pruned and standard\nfully connected networks. To find the optimal weight-agnostic network, we use a\nnovel and computationally efficient method that translates the hard\narchitecture-search problem into a feasible optimization problem.More\nspecifically, we look at the optimal task-specific architectures as the optimal\nconfiguration of binary networks with {0, 1}-valued weights, which can be found\nthrough an approximate gradient descent strategy. Theoretical convergence\nguarantees of the proposed algorithm are obtained by bounding the error in the\ngradient approximation and its practical performance is evaluated on two\nreal-world data sets. For measuring the structural similarities between\ndifferent architectures, we use a novel spectral approach that allows us to\nunderline the intrinsic differences between real-valued networks and\nweight-free architectures.', 'The aim of this paper is to propose an application of mutual\ninformation-based ensemble methods to the analysis and classification of heart\nbeats associated with different types of Arrhythmia. Models of multilayer\nperceptrons, support vector machines, and radial basis function neural networks\nwere trained and tested using the MIT-BIH arrhythmia database. This research\nbrings a focus to an ensemble method that, to our knowledge, is a novel\napplication in the area of ECG Arrhythmia detection. The proposed classifier\nensemble method showed improved performance, relative to either majority voting\nclassifier integration or to individual classifier performance. The overall\nensemble accuracy was 98.25%.', ""{\\em Hypernetworks} are architectures that produce the weights of a\ntask-specific {\\em primary network}. A notable application of hypernetworks in\nthe recent literature involves learning to output functional representations.\nIn these scenarios, the hypernetwork learns a representation corresponding to\nthe weights of a shallow MLP, which typically encodes shape or image\ninformation. While such representations have seen considerable success in\npractice, they remain lacking in the theoretical guarantees in the wide regime\nof the standard architectures. In this work, we study wide over-parameterized\nhypernetworks. We show that unlike typical architectures, infinitely wide\nhypernetworks do not guarantee convergence to a global minima under gradient\ndescent. We further show that convexity can be achieved by increasing the\ndimensionality of the hypernetwork's output, to represent wide MLPs. In the\ndually infinite-width regime, we identify the functional priors of these\narchitectures by deriving their corresponding GP and NTK kernels, the latter of\nwhich we refer to as the {\\em hyperkernel}. As part of this study, we make a\nmathematical contribution by deriving tight bounds on high order Taylor\nexpansion terms of standard fully connected ReLU networks."", 'We propose an end-to-end model based on convolutional and recurrent neural\nnetworks for speech enhancement. Our model is purely data-driven and does not\nmake any assumptions about the type or the stationarity of the noise. In\ncontrast to existing methods that use multilayer perceptrons (MLPs), we employ\nboth convolutional and recurrent neural network architectures. Thus, our\napproach allows us to exploit local structures in both the frequency and\ntemporal domains. By incorporating prior knowledge of speech signals into the\ndesign of model structures, we build a model that is more data-efficient and\nachieves better generalization on both seen and unseen noise. Based on\nexperiments with synthetic data, we demonstrate that our model outperforms\nexisting methods, improving PESQ by up to 0.6 on seen noise and 0.64 on unseen\nnoise.', 'Embedding based models have been the state of the art in collaborative\nfiltering for over a decade. Traditionally, the dot product or higher order\nequivalents have been used to combine two or more embeddings, e.g., most\nnotably in matrix factorization. In recent years, it was suggested to replace\nthe dot product with a learned similarity e.g. using a multilayer perceptron\n(MLP). This approach is often referred to as neural collaborative filtering\n(NCF). In this work, we revisit the experiments of the NCF paper that\npopularized learned similarities using MLPs. First, we show that with a proper\nhyperparameter selection, a simple dot product substantially outperforms the\nproposed learned similarities. Second, while a MLP can in theory approximate\nany function, we show that it is non-trivial to learn a dot product with an\nMLP. Finally, we discuss practical issues that arise when applying MLP based\nsimilarities and show that MLPs are too costly to use for item recommendation\nin production environments while dot products allow to apply very efficient\nretrieval algorithms. We conclude that MLPs should be used with care as\nembedding combiner and that dot products might be a better default choice.', 'How to leverage cross-document interactions to improve ranking performance is\nan important topic in information retrieval (IR) research. However, this topic\nhas not been well-studied in the learning-to-rank setting and most of the\nexisting work still treats each document independently while scoring. The\nrecent development of deep learning shows strength in modeling complex\nrelationships across sequences and sets. It thus motivates us to study how to\nleverage cross-document interactions for learning-to-rank in the deep learning\nframework. In this paper, we formally define the permutation-equivariance\nrequirement for a scoring function that captures cross-document interactions.\nWe then propose a self-attention based document interaction network and show\nthat it satisfies the permutation-equivariant requirement, and can generate\nscores for document sets of varying sizes. Our proposed methods can\nautomatically learn to capture document interactions without any auxiliary\ninformation, and can scale across large document sets. We conduct experiments\non three ranking datasets: the benchmark Web30k, a Gmail search, and a Google\nDrive Quick Access dataset. Experimental results show that our proposed methods\nare both more effective and efficient than baselines.', 'Artificial Neural Networks(ANN) has been phenomenally successful on various\npattern recognition tasks. However, the design of neural networks rely heavily\non the experience and intuitions of individual developers. In this article, the\nauthor introduces a mathematical structure called MLP algebra on the set of all\nMultilayer Perceptron Neural Networks(MLP), which can serve as a guiding\nprinciple to build MLPs accommodating to the particular data sets, and to build\ncomplex MLPs from simpler ones.', ""We develop a mathematically rigorous framework for multilayer neural networks\nin the mean field regime. As the network's widths increase, the network's\nlearning trajectory is shown to be well captured by a meaningful and\ndynamically nonlinear limit (the \\textit{mean field} limit), which is\ncharacterized by a system of ODEs. Our framework applies to a broad range of\nnetwork architectures, learning dynamics and network initializations. Central\nto the framework is the new idea of a \\textit{neuronal embedding}, which\ncomprises of a non-evolving probability space that allows to embed neural\nnetworks of arbitrary widths.\n  Using our framework, we prove several properties of large-width multilayer\nneural networks. Firstly we show that independent and identically distributed\ninitializations cause strong degeneracy effects on the network's learning\ntrajectory when the network's depth is at least four. Secondly we obtain\nseveral global convergence guarantees for feedforward multilayer networks under\na number of different setups. These include two-layer and three-layer networks\nwith independent and identically distributed initializations, and multilayer\nnetworks of arbitrary depths with a special type of correlated initializations\nthat is motivated by the new concept of \\textit{bidirectional diversity}.\nUnlike previous works that rely on convexity, our results admit non-convex\nlosses and hinge on a certain universal approximation property, which is a\ndistinctive feature of infinite-width neural networks and is shown to hold\nthroughout the training process. Aside from being the first known results for\nglobal convergence of multilayer networks in the mean field regime, they\ndemonstrate flexibility of our framework and incorporate several new ideas and\ninsights that depart from the conventional convex optimization wisdom."", ""Successful sequential recommendation systems rely on accurately capturing the\nuser's short-term and long-term interest. Although Transformer-based models\nachieved state-of-the-art performance in the sequential recommendation task,\nthey generally require quadratic memory and time complexity to the sequence\nlength, making it difficult to extract the long-term interest of users. On the\nother hand, Multi-Layer Perceptrons (MLP)-based models, renowned for their\nlinear memory and time complexity, have recently shown competitive results\ncompared to Transformer in various tasks. Given the availability of a massive\namount of the user's behavior history, the linear memory and time complexity of\nMLP-based models make them a promising alternative to explore in the sequential\nrecommendation task. To this end, we adopted MLP-based models in sequential\nrecommendation but consistently observed that MLP-based methods obtain lower\nperformance than those of Transformer despite their computational benefits.\nFrom experiments, we observed that introducing explicit high-order interactions\nto MLP layers mitigates such performance gap. In response, we propose the\nMulti-Order Interaction (MOI) layer, which is capable of expressing an\narbitrary order of interactions within the inputs while maintaining the memory\nand time complexity of the MLP layer. By replacing the MLP layer with the MOI\nlayer, our model was able to achieve comparable performance with\nTransformer-based models while retaining the MLP-based models' computational\nbenefits."", 'We propose a Label Propagation based algorithm for weakly supervised text\nclassification. We construct a graph where each document is represented by a\nnode and edge weights represent similarities among the documents. Additionally,\nwe discover underlying topics using Latent Dirichlet Allocation (LDA) and\nenrich the document graph by including the topics in the form of additional\nnodes. The edge weights between a topic and a text document represent level of\n""affinity"" between them. Our approach does not require document level\nlabelling, instead it expects manual labels only for topic nodes. This\nsignificantly minimizes the level of supervision needed as only a few topics\nare observed to be enough for achieving sufficiently high accuracy. The Label\nPropagation Algorithm is employed on this enriched graph to propagate labels\namong the nodes. Our approach combines the advantages of Label Propagation\n(through document-document similarities) and Topic Modelling (for minimal but\nsmart supervision). We demonstrate the effectiveness of our approach on various\ndatasets and compare with state-of-the-art weakly supervised text\nclassification approaches.', 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'We study the task of semi-supervised learning on multilayer graphs by taking\ninto account both labeled and unlabeled observations together with the\ninformation encoded by each individual graph layer. We propose a regularizer\nbased on the generalized matrix mean, which is a one-parameter family of matrix\nmeans that includes the arithmetic, geometric and harmonic means as particular\ncases. We analyze it in expectation under a Multilayer Stochastic Block Model\nand verify numerically that it outperforms state of the art methods. Moreover,\nwe introduce a matrix-free numerical scheme based on contour integral\nquadratures and Krylov subspace solvers that scales to large sparse multilayer\ngraphs.', ""The question answering system can answer questions from various fields and\nforms with deep neural networks, but it still lacks effective ways when facing\nmultiple evidences. We introduce a new model called SRQA, which means Synthetic\nReader for Factoid Question Answering. This model enhances the question\nanswering system in the multi-document scenario from three aspects: model\nstructure, optimization goal, and training method, corresponding to Multilayer\nAttention (MA), Cross Evidence (CE), and Adversarial Training (AT)\nrespectively. First, we propose a multilayer attention network to obtain a\nbetter representation of the evidences. The multilayer attention mechanism\nconducts interaction between the question and the passage within each layer,\nmaking the token representation of evidences in each layer takes the\nrequirement of the question into account. Second, we design a cross evidence\nstrategy to choose the answer span within more evidences. We improve the\noptimization goal, considering all the answers' locations in multiple evidences\nas training targets, which leads the model to reason among multiple evidences.\nThird, adversarial training is employed to high-level variables besides the\nword embedding in our model. A new normalization method is also proposed for\nadversarial perturbations so that we can jointly add perturbations to several\ntarget variables. As an effective regularization method, adversarial training\nenhances the model's ability to process noisy data. Combining these three\nstrategies, we enhance the contextual representation and locating ability of\nour model, which could synthetically extract the answer span from several\nevidences. We perform SRQA on the WebQA dataset, and experiments show that our\nmodel outperforms the state-of-the-art models (the best fuzzy score of our\nmodel is up to 78.56%, with an improvement of about 2%)."", 'We propose a method for simultaneously detecting shared and unshared\ncommunities in heterogeneous multilayer weighted and undirected networks. The\nmultilayer network is assumed to follow a generative probabilistic model that\ntakes into account the similarities and dissimilarities between the\ncommunities. We make use of a variational Bayes approach for jointly inferring\nthe shared and unshared hidden communities from multilayer network\nobservations. We show that our approach outperforms state-of-the-art algorithms\nin detecting disparate (shared and private) communities on synthetic data as\nwell as on real genome-wide fibroblast proliferation dataset.', 'Multilayer graphs are commonly used for representing different relations\nbetween entities and handling heterogeneous data processing tasks. New\nchallenges arise in multilayer graph clustering for assigning clusters to a\ncommon multilayer node set and for combining information from each layer. This\npaper presents a theoretical framework for multilayer spectral graph clustering\nof the nodes via convex layer aggregation. Under a novel multilayer signal plus\nnoise model, we provide a phase transition analysis that establishes the\nexistence of a critical value on the noise level that permits reliable cluster\nseparation. The analysis also specifies analytical upper and lower bounds on\nthe critical value, where the bounds become exact when the clusters have\nidentical sizes. Numerical experiments on synthetic multilayer graphs are\nconducted to validate the phase transition analysis and study the effect of\nlayer weights and noise levels on clustering reliability.', 'We investigate the representation power of graph neural networks in the\nsemi-supervised node classification task under heterophily or low homophily,\ni.e., in networks where connected nodes may have different class labels and\ndissimilar features. Many popular GNNs fail to generalize to this setting, and\nare even outperformed by models that ignore the graph structure (e.g.,\nmultilayer perceptrons). Motivated by this limitation, we identify a set of key\ndesigns -- ego- and neighbor-embedding separation, higher-order neighborhoods,\nand combination of intermediate representations -- that boost learning from the\ngraph structure under heterophily. We combine them into a graph neural network,\nH2GCN, which we use as the base method to empirically evaluate the\neffectiveness of the identified designs. Going beyond the traditional\nbenchmarks with strong homophily, our empirical analysis shows that the\nidentified designs increase the accuracy of GNNs by up to 40% and 27% over\nmodels without them on synthetic and real networks with heterophily,\nrespectively, and yield competitive performance under homophily.', 'Like most learning algorithms, the multilayer perceptrons (MLP) is designed\nto learn a vector of parameters from data. However, in certain scenarios we are\ninterested in learning structured parameters (predictions) in the form of\nsymmetric positive definite matrices. Here, we introduce a variant of the MLP,\nreferred to as the matrix MLP, that is specialized at learning symmetric\npositive definite matrices. We also present an application of the model within\nthe context of the variational autoencoder (VAE). Our formulation of the VAE\nextends the vanilla formulation to the cases where the recognition and the\ngenerative networks can be from the parametric family of distributions with\ndense covariance matrices. Two specific examples are discussed in more detail:\nthe dense covariance Gaussian and its generalization, the power exponential\ndistribution. Our new developments are illustrated using both synthetic and\nreal data.', 'The main objective of eXplainable Artificial Intelligence (XAI) is to provide\neffective explanations for black-box classifiers. The existing literature lists\nmany desirable properties for explanations to be useful, but there is no\nconsensus on how to quantitatively evaluate explanations in practice. Moreover,\nexplanations are typically used only to inspect black-box models, and the\nproactive use of explanations as a decision support is generally overlooked.\nAmong the many approaches to XAI, a widely adopted paradigm is Local Linear\nExplanations - with LIME and SHAP emerging as state-of-the-art methods. We show\nthat these methods are plagued by many defects including unstable explanations,\ndivergence of actual implementations from the promised theoretical properties,\nand explanations for the wrong label. This highlights the need to have standard\nand unbiased evaluation procedures for Local Linear Explanations in the XAI\nfield. In this paper we address the problem of identifying a clear and\nunambiguous set of metrics for the evaluation of Local Linear Explanations.\nThis set includes both existing and novel metrics defined specifically for this\nclass of explanations. All metrics have been included in an open Python\nframework, named LEAF. The purpose of LEAF is to provide a reference for end\nusers to evaluate explanations in a standardised and unbiased way, and to guide\nresearchers towards developing improved explainable techniques.', 'Many recent developments in network analysis have focused on multilayer\nnetworks, which one can use to encode time-dependent interactions, multiple\ntypes of interactions, and other complications that arise in complex systems.\nLike their monolayer counterparts, multilayer networks in applications often\nhave mesoscale features, such as community structure. A prominent type of\nmethod for inferring such structures is the employment of multilayer stochastic\nblock models (SBMs). A common (but {potentially} inadequate) assumption of\nthese models is the sampling of edges in different layers independently,\nconditioned on the community labels of the nodes. In this paper, we relax this\nassumption of independence by incorporating edge correlations into an SBM-like\nmodel. We derive maximum-likelihood estimates of the key parameters of our\nmodel, and we propose a measure of layer correlation that reflects the\nsimilarity between connectivity patterns in different layers. Finally, we\nexplain how to use correlated models for edge ""prediction"" (i.e., inference) in\nmultilayer networks. By taking into account edge correlations, prediction\naccuracy improves both in synthetic networks and in a temporal network of\nshoppers who are connected to previously-purchased grocery products.', 'For the past ten years, CNN has reigned supreme in the world of computer\nvision, but recently, Transformer has been on the rise. However, the quadratic\ncomputational cost of self-attention has become a serious problem in practice\napplications. There has been much research on architectures without CNN and\nself-attention in this context. In particular, MLP-Mixer is a simple\narchitecture designed using MLPs and hit an accuracy comparable to the Vision\nTransformer. However, the only inductive bias in this architecture is the\nembedding of tokens. This leaves open the possibility of incorporating a\nnon-convolutional (or non-local) inductive bias into the architecture, so we\nused two simple ideas to incorporate inductive bias into the MLP-Mixer while\ntaking advantage of its ability to capture global correlations. A way is to\ndivide the token-mixing block vertically and horizontally. Another way is to\nmake spatial correlations denser among some channels of token-mixing. With this\napproach, we were able to improve the accuracy of the MLP-Mixer while reducing\nits parameters and computational complexity. The small model that is RaftMLP-S\nis comparable to the state-of-the-art global MLP-based model in terms of\nparameters and efficiency per calculation. In addition, we tackled the problem\nof fixed input image resolution for global MLP-based models by utilizing\nbicubic interpolation. We demonstrated that these models could be applied as\nthe backbone of architectures for downstream tasks such as object detection.\nHowever, it did not have significant performance and mentioned the need for\nMLP-specific architectures for downstream tasks for global MLP-based models.\nThe source code in PyTorch version is available at\n\\url{https://github.com/okojoalg/raft-mlp}.', 'High-level applications, such as machine learning, are evolving from simple\nmodels based on multilayer perceptrons for simple image recognition to much\ndeeper and more complex neural networks for self-driving vehicle control\nsystems.The rapid increase in the consumption of memory and computational\nresources by these models demands the use of multi-core parallel systems to\nscale the execution of the complex emerging applications that depend on them.\nHowever, parallel programs running on high-performance computers often suffer\nfrom data communication bottlenecks, limited memory bandwidth, and\nsynchronization overhead due to irregular critical sections. In this paper, we\npropose a framework to reduce the data communication and improve the\nscalability and performance of these applications in multi-core systems. We\ndesign a vertex cut framework for partitioning LLVM IR graphs into clusters\nwhile taking into consideration the data communication and workload balance\namong clusters. First, we construct LLVM graphs by compiling high-level\nprograms into LLVM IR, instrumenting code to obtain the execution order of\nbasic blocks and the execution time for each memory operation, and analyze data\ndependencies in dynamic LLVM traces. Next, we formulate the problem as Weight\nBalanced $p$-way Vertex Cut, and propose a generic and flexible framework,\nwherein four different greedy algorithms are proposed for solving this problem.\nLastly, we propose a memory-centric run-time mapping of the linear time\ncomplexity to map clusters generated from the vertex cut algorithms onto a\nmulti-core platform. We conclude that our best algorithm, WB-Libra, provides\nperformance improvements of 1.56x and 1.86x over existing state-of-the-art\napproaches for 8 and 1024 clusters running on a multi-core platform,\nrespectively.', ""This paper is about a machine learning approach based on the multilinear\nprojection of an unknown function (or probability distribution) to be estimated\ntowards a linear (or multilinear) dimensional space E'. The proposal transforms\nthe problem of predicting the target of an observation x into a problem of\ndetermining a consensus among the k nearest neighbors of x's image within the\ndimensional space E'. The algorithms that concretize it allow both regression\nand binary classification. Implementations carried out using Scala/Spark and\nassessed on a dozen LIBSVM datasets have demonstrated improvements in\nprediction accuracies in comparison with other prediction algorithms\nimplemented within Spark MLLib such as multilayer perceptrons, logistic\nregression classifiers and random forests."", 'Image denoising can be described as the problem of mapping from a noisy image\nto a noise-free image. The best currently available denoising methods\napproximate this mapping with cleverly engineered algorithms. In this work we\nattempt to learn this mapping directly with plain multi layer perceptrons (MLP)\napplied to image patches. We will show that by training on large image\ndatabases we are able to outperform the current state-of-the-art image\ndenoising methods. In addition, our method achieves results that are superior\nto one type of theoretical bound and goes a large way toward closing the gap\nwith a second type of theoretical bound. Our approach is easily adapted to less\nextensively studied types of noise, such as mixed Poisson-Gaussian noise, JPEG\nartifacts, salt-and-pepper noise and noise resembling stripes, for which we\nachieve excellent results as well. We will show that combining a block-matching\nprocedure with MLPs can further improve the results on certain images. In a\nsecond paper, we detail the training trade-offs and the inner mechanisms of our\nMLPs.', 'We present here a new model and algorithm which performs an efficient Natural\ngradient descent for Multilayer Perceptrons. Natural gradient descent was\noriginally proposed from a point of view of information geometry, and it\nperforms the steepest descent updates on manifolds in a Riemannian space. In\nparticular, we extend an approach taken by the ""Whitened neural networks""\nmodel. We make the whitening process not only in feed-forward direction as in\nthe original model, but also in the back-propagation phase. Its efficacy is\nshown by an application of this ""Bidirectional whitened neural networks"" model\nto a handwritten character recognition data (MNIST data).', ""We present a novel hierarchical approach to multi-class classification which\nis generic in that it can be applied to different classification models (e.g.,\nsupport vector machines, perceptrons), and makes no explicit assumptions about\nthe probabilistic structure of the problem as it is usually done in multi-class\nclassification. By adding a cascade of additional classifiers, each of which\nreceives the previous classifier's output in addition to regular input data,\nthe approach harnesses unused information that manifests itself in the form of,\ne.g., correlations between predicted classes. Using multilayer perceptrons as a\nclassification model, we demonstrate the validity of this approach by testing\nit on a complex ten-class 3D gesture recognition task."", ""We proposed a novel multilayer correlated topic model (MCTM) to analyze how\nthe main ideas inherit and vary between a document and its different segments,\nwhich helps understand an article's structure. The variational\nexpectation-maximization (EM) algorithm was derived to estimate the posterior\nand parameters in MCTM. We introduced two potential applications of MCTM,\nincluding the paragraph-level document analysis and market basket data\nanalysis. The effectiveness of MCTM in understanding the document structure has\nbeen verified by the great predictive performance on held-out documents and\nintuitive visualization. We also showed that MCTM could successfully capture\ncustomers' popular shopping patterns in the market basket analysis."", 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.', 'This paper shows that a long chain of perceptrons (that is, a multilayer\nperceptron, or MLP, with many hidden layers of width one) can be a universal\nclassifier. The classification procedure is not necessarily computationally\nefficient, but the technique throws some light on the kind of computations\npossible with narrow and deep MLPs.', 'The adoption of intelligent systems with Artificial Neural Networks (ANNs)\nembedded in hardware for real-time applications currently faces a growing\ndemand in fields like the Internet of Things (IoT) and Machine to Machine\n(M2M). However, the application of ANNs in this type of system poses a\nsignificant challenge due to the high computational power required to process\nits basic operations. This paper aims to show an implementation strategy of a\nMultilayer Perceptron (MLP) type neural network, in a microcontroller (a\nlow-cost, low-power platform). A modular matrix-based MLP with the full\nclassification process was implemented, and also the backpropagation training\nin the microcontroller. The testing and validation were performed through\nHardware in the Loop (HIL) of the Mean Squared Error (MSE) of the training\nprocess, classification result, and the processing time of each implementation\nmodule. The results revealed a linear relationship between the values of the\nhyperparameters and the processing time required for classification, also the\nprocessing time concurs with the required time for many applications on the\nfields mentioned above. These findings show that this implementation strategy\nand this platform can be applied successfully on real-time applications that\nrequire the capabilities of ANNs.', 'Artificial Neural Networks have emerged as an important tool for\nclassification and have been widely used to classify a non-linear separable\npattern. The most popular artificial neural networks model is a Multilayer\nPerceptron (MLP) as it is able to perform classification task with significant\nsuccess. However due to the complexity of MLP structure and also problems such\nas local minima trapping, over fitting and weight interference have made neural\nnetwork training difficult. Thus, the easy way to avoid these problems is to\nremove the hidden layers. This paper presents the ability of Functional Link\nNeural Network (FLNN) to overcome the complexity structure of MLP by using\nsingle layer architecture and propose an Artificial Bee Colony (ABC)\noptimization for training the FLNN. The proposed technique is expected to\nprovide better learning scheme for a classifier in order to get more accurate\nclassification result', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', ""Aiming at drastic speedup for point-feature embeddings at test time, we\npropose a new framework that uses a pair of multi-layer perceptrons (MLP) and a\nlookup table (LUT) to transform point-coordinate inputs into high-dimensional\nfeatures. When compared with PointNet's feature embedding part realized by MLP\nthat requires millions of dot products, the proposed framework at test time\nrequires no such layers of matrix-vector products but requires only looking up\nthe nearest entities from the tabulated MLP followed by interpolation, defined\nover discrete inputs on a 3D lattice that is substantially arranged\nirregularly. We call this framework LUTI-MLP: LUT Interpolation ML that\nprovides a way to train end-to-end irregularly tabulated MLP coupled to a LUT\nin a specific manner without the need for any approximation at test time.\nLUTI-MLP also provides significant speedup for Jacobian computation of the\nembedding function wrt global pose coordinate on Lie algebra $\\mathfrak{se}(3)$\nat test time, which could be used for point-set registration problems. After\nextensive evaluation using the ModelNet40, we confirmed that the LUTI-MLP even\nwith a small (e.g., $4^3$) lattice yields performance comparable to that of the\nMLP while achieving significant speedup: $100\\times$ for the embedding,\n$12\\times$ for the approximate Jacobian, and $860\\times$ for the canonical\nJacobian."", 'Perceptrons are neuronal devices capable of fully discriminating linearly\nseparable classes. Although straightforward to implement and train, their\napplicability is usually hindered by non-trivial requirements imposed by\nreal-world classification problems. Therefore, several approaches, such as\nkernel perceptrons, have been conceived to counteract such difficulties. In\nthis paper, we investigate an enhanced perceptron model based on the notion of\ncontrastive biclusters. From this perspective, a good discriminative bicluster\ncomprises a subset of data instances belonging to one class that show high\ncoherence across a subset of features and high differentiation from nearest\ninstances of the other class under the same features (referred to as its\ncontrastive bicluster). Upon each local subspace associated with a pair of\ncontrastive biclusters a perceptron is trained and the model with highest area\nunder the receiver operating characteristic curve (AUC) value is selected as\nthe final classifier. Experiments conducted on a range of data sets, including\nthose related to a difficult biosignal classification problem, show that the\nproposed variant can be indeed very useful, prevailing in most of the cases\nupon standard and kernel perceptrons in terms of accuracy and AUC measures.', 'In this paper, we focus on a typical two-phase phenomenon in the learning of\nmulti-layer perceptrons (MLPs), and we aim to explain the reason for the\ndecrease of feature diversity in the first phase. Specifically, people find\nthat, in the training of MLPs, the training loss does not decrease\nsignificantly until the second phase. To this end, we further explore the\nreason why the diversity of features over different samples keeps decreasing in\nthe first phase, which hurts the optimization of MLPs. We explain such a\nphenomenon in terms of the learning dynamics of MLPs. Furthermore, we\ntheoretically explain why four typical operations can alleviate the decrease of\nthe feature diversity.', 'Nowadays, neural networks are widely used in many applications as artificial\nintelligence models for learning tasks. Since typically neural networks process\na very large amount of data, it is convenient to formulate them within the\nmean-field and kinetic theory. In this work we focus on a particular class of\nneural networks, i.e. the residual neural networks, assuming that each layer is\ncharacterized by the same number of neurons $N$, which is fixed by the\ndimension of the data. This assumption allows to interpret the residual neural\nnetwork as a time-discretized ordinary differential equation, in analogy with\nneural differential equations. The mean-field description is then obtained in\nthe limit of infinitely many input data. This leads to a Vlasov-type partial\ndifferential equation which describes the evolution of the distribution of the\ninput data. We analyze steady states and sensitivity with respect to the\nparameters of the network, namely the weights and the bias. In the simple\nsetting of a linear activation function and one-dimensional input data, the\nstudy of the moments provides insights on the choice of the parameters of the\nnetwork. Furthermore, a modification of the microscopic dynamics, inspired by\nstochastic residual neural networks, leads to a Fokker-Planck formulation of\nthe network, in which the concept of network training is replaced by the task\nof fitting distributions. The performed analysis is validated by artificial\nnumerical simulations. In particular, results on classification and regression\nproblems are presented.', 'The ability to identify applications based on the network data they generate\ncould be a valuable tool for cyber defense. We report on a machine learning\ntechnique capable of using netflow-like features to predict the application\nthat generated the traffic. In our experiments, we used ground-truth labels\nobtained from host-based sensors deployed in a large enterprise environment; we\napplied random forests and multilayer perceptrons to the tasks of browser vs.\nnon-browser identification, browser fingerprinting, and process name\nprediction. For each of these tasks, we demonstrate how machine learning models\ncan achieve high classification accuracy using only netflow-like features as\nthe basis for classification.', 'Recently, visual Transformer (ViT) and its following works abandon the\nconvolution and exploit the self-attention operation, attaining a comparable or\neven higher accuracy than CNNs. More recently, MLP-Mixer abandons both the\nconvolution and the self-attention operation, proposing an architecture\ncontaining only MLP layers. To achieve cross-patch communications, it devises\nan additional token-mixing MLP besides the channel-mixing MLP. It achieves\npromising results when training on an extremely large-scale dataset. But it\ncannot achieve as outstanding performance as its CNN and ViT counterparts when\ntraining on medium-scale datasets such as ImageNet1K and ImageNet21K. The\nperformance drop of MLP-Mixer motivates us to rethink the token-mixing MLP. We\ndiscover that the token-mixing MLP is a variant of the depthwise convolution\nwith a global reception field and spatial-specific configuration. But the\nglobal reception field and the spatial-specific property make token-mixing MLP\nprone to over-fitting. In this paper, we propose a novel pure MLP architecture,\nspatial-shift MLP (S$^2$-MLP). Different from MLP-Mixer, our S$^2$-MLP only\ncontains channel-mixing MLP. We utilize a spatial-shift operation for\ncommunications between patches. It has a local reception field and is\nspatial-agnostic. It is parameter-free and efficient for computation. The\nproposed S$^2$-MLP attains higher recognition accuracy than MLP-Mixer when\ntraining on ImageNet-1K dataset. Meanwhile, S$^2$-MLP accomplishes as excellent\nperformance as ViT on ImageNet-1K dataset with considerably simpler\narchitecture and fewer FLOPs and parameters.', 'The light scattering of multilayer nanoparticles can be solved by Maxwell\nequations. However, it is difficult to solve the inverse design of multilayer\nnanoparticles by using the traditional trial-and-error method. Here, we present\na method for forward simulation and inverse design of multilayer nanoparticles.\nWe combine the global search ability of genetic algorithm with the local search\nability of neural network. First, the genetic algorithm is used to find a\nsuitable solution, and then the neural network is used to fine-tune it. Due to\nthe non-unique relationship between physical structures and optical responses,\nwe first train a forward neural network, and then it is applied to the inverse\ndesign of multilayer nanoparticles. Not only here, this method can easily be\nextended to predict and find the best design parameters for other optical\nstructures.', ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'Metric learning aims to learn a distance metric such that semantically\nsimilar instances are pulled together while dissimilar instances are pushed\naway. Many existing methods consider maximizing or at least constraining a\ndistance margin in the feature space that separates similar and dissimilar\npairs of instances to guarantee their generalization ability. In this paper, we\nadvocate imposing an adversarial margin in the input space so as to improve the\ngeneralization and robustness of metric learning algorithms. We first show\nthat, the adversarial margin, defined as the distance between training\ninstances and their closest adversarial examples in the input space, takes\naccount of both the distance margin in the feature space and the correlation\nbetween the metric and triplet constraints. Next, to enhance robustness to\ninstance perturbation, we propose to enlarge the adversarial margin through\nminimizing a derived novel loss function termed the perturbation loss. The\nproposed loss can be viewed as a data-dependent regularizer and easily plugged\ninto any existing metric learning methods. Finally, we show that the enlarged\nmargin is beneficial to the generalization ability by using the theoretical\ntechnique of algorithmic robustness. Experimental results on 16 datasets\ndemonstrate the superiority of the proposed method over existing\nstate-of-the-art methods in both discrimination accuracy and robustness against\npossible noise.', 'In this work we use Recurrent Neural Networks and Multilayer Perceptrons to\npredict NYSE, NASDAQ and AMEX stock prices from historical data. We experiment\nwith different architectures and compare data normalization techniques. Then,\nwe leverage those findings to question the efficient-market hypothesis through\na formal statistical test.', 'Modern entity linking systems rely on large collections of documents\nspecifically annotated for the task (e.g., AIDA CoNLL). In contrast, we propose\nan approach which exploits only naturally occurring information: unlabeled\ndocuments and Wikipedia. Our approach consists of two stages. First, we\nconstruct a high recall list of candidate entities for each mention in an\nunlabeled document. Second, we use the candidate lists as weak supervision to\nconstrain our document-level entity linking model. The model treats entities as\nlatent variables and, when estimated on a collection of unlabelled texts,\nlearns to choose entities relying both on local context of each mention and on\ncoherence with other entities in the document. The resulting approach rivals\nfully-supervised state-of-the-art systems on standard test sets. It also\napproaches their performance in the very challenging setting: when tested on a\ntest set sampled from the data used to estimate the supervised systems. By\ncomparing to Wikipedia-only training of our model, we demonstrate that modeling\nunlabeled documents is beneficial.', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', 'Multilayer graphs are commonly used for representing different relations\nbetween entities and handling heterogeneous data processing tasks. Non-standard\nmultilayer graph clustering methods are needed for assigning clusters to a\ncommon multilayer node set and for combining information from each layer. This\npaper presents a multilayer spectral graph clustering (SGC) framework that\nperforms convex layer aggregation. Under a multilayer signal plus noise model,\nwe provide a phase transition analysis of clustering reliability. Moreover, we\nuse the phase transition criterion to propose a multilayer iterative model\norder selection algorithm (MIMOSA) for multilayer SGC, which features automated\ncluster assignment and layer weight adaptation, and provides statistical\nclustering reliability guarantees. Numerical simulations on synthetic\nmultilayer graphs verify the phase transition analysis, and experiments on\nreal-world multilayer graphs show that MIMOSA is competitive or better than\nother clustering methods.', 'Multilayer perceptron (MLP) is a class of networks composed of multiple\nlayers of perceptrons, and it is essentially a mathematical function. Based on\nMLP, we develop a new numerical method to find the extrema of functionals. As\ndemonstrations, we present our solutions in three physic scenes. Ideally, the\nsame method is applicable to any cases where the objective curve/surface can be\nfitted by second-order differentiable functions. This method can also be\nextended to cases where there are a finite number of non-differentiable (but\ncontinuous) points/surfaces.', ""We are interested in the widespread problem of clustering documents and\nfinding topics in large collections of written documents in the presence of\nmetadata and hyperlinks. To tackle the challenge of accounting for these\ndifferent types of datasets, we propose a novel framework based on Multilayer\nNetworks and Stochastic Block Models. The main innovation of our approach over\nother techniques is that it applies the same non-parametric probabilistic\nframework to the different sources of datasets simultaneously. The key\ndifference to other multilayer complex networks is the strong unbalance between\nthe layers, with the average degree of different node types scaling differently\nwith system size. We show that the latter observation is due to generic\nproperties of text, such as Heaps' law, and strongly affects the inference of\ncommunities. We present and discuss the performance of our method in different\ndatasets (hundreds of Wikipedia documents, thousands of scientific papers, and\nthousands of E-mails) showing that taking into account multiple types of\ninformation provides a more nuanced view on topic- and document-clusters and\nincreases the ability to predict missing links."", ""Policy distillation in deep reinforcement learning provides an effective way\nto transfer control policies from a larger network to a smaller untrained\nnetwork without a significant degradation in performance. However, policy\ndistillation is underexplored in deep reinforcement learning, and existing\napproaches are computationally inefficient, resulting in a long distillation\ntime. In addition, the effectiveness of the distillation process is still\nlimited to the model capacity. We propose a new distillation mechanism, called\nreal-time policy distillation, in which training the teacher model and\ndistilling the policy to the student model occur simultaneously. Accordingly,\nthe teacher's latest policy is transferred to the student model in real time.\nThis reduces the distillation time to half the original time or even less and\nalso makes it possible for extremely small student models to learn skills at\nthe expert level. We evaluated the proposed algorithm in the Atari 2600 domain.\nThe results show that our approach can achieve full distillation in most games,\neven with compression ratios up to 1.7%."", 'We propose a new framework for estimating generative models via an\nadversarial process, in which we simultaneously train two models: a generative\nmodel G that captures the data distribution, and a discriminative model D that\nestimates the probability that a sample came from the training data rather than\nG. The training procedure for G is to maximize the probability of D making a\nmistake. This framework corresponds to a minimax two-player game. In the space\nof arbitrary functions G and D, a unique solution exists, with G recovering the\ntraining data distribution and D equal to 1/2 everywhere. In the case where G\nand D are defined by multilayer perceptrons, the entire system can be trained\nwith backpropagation. There is no need for any Markov chains or unrolled\napproximate inference networks during either training or generation of samples.\nExperiments demonstrate the potential of the framework through qualitative and\nquantitative evaluation of the generated samples.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Neural Networks (NNs) can be used to solve Ordinary and Partial Differential\nEquations (ODEs and PDEs) by redefining the question as an optimization\nproblem. The objective function to be optimized is the sum of the squares of\nthe PDE to be solved and of the initial/boundary conditions. A feed forward NN\nis trained to minimise this loss function evaluated on a set of collocation\npoints sampled from the domain where the problem is defined. A compact and\nsmooth solution, that only depends on the weights of the trained NN, is then\nobtained. This approach is often referred to as PINN, from Physics Informed\nNeural Network~\\cite{raissi2017physics_1, raissi2017physics_2}. Despite the\nsuccess of the PINN approach in solving various classes of PDEs, an\nimplementation of this idea that is capable of solving a large class of ODEs\nand PDEs with good accuracy and without the need to finely tune the\nhyperparameters of the network, is not available yet. In this paper, we\nintroduce a new implementation of this concept - called dNNsolve - that makes\nuse of dual Neural Networks to solve ODEs/PDEs. These include: i) sine and\nsigmoidal activation functions, that provide a more efficient basis to capture\nboth secular and periodic patterns in the solutions; ii) a newly designed\narchitecture, that makes it easy for the the NN to approximate the solution\nusing the basis functions mentioned above. We show that dNNsolve is capable of\nsolving a broad range of ODEs/PDEs in 1, 2 and 3 spacetime dimensions, without\nthe need of hyperparameter fine-tuning.', 'The significance of heating load (HL) accurate approximation is the primary\nmotivation of this research to distinguish the most efficient predictive model\namong several neural-metaheuristic models. The proposed models are through\nsynthesizing multi-layer perceptron network (MLP) with ant lion optimization\n(ALO), biogeography-based optimization (BBO), dragonfly algorithm (DA),\nevolutionary strategy (ES), invasive weed optimization (IWO), and league\nchampion optimization (LCA) hybrid algorithms. Each ensemble is optimized in\nterms of the operating population. Accordingly, the ALO-MLP, BBO-MLP, DA-MLP,\nES-MLP, IWO-MLP, and LCA-MLP presented their best performance for population\nsizes of 350, 400, 200, 500, 50, and 300, respectively. The comparison was\ncarried out by implementing a ranking system. Based on the obtained overall\nscores (OSs), the BBO (OS = 36) featured as the most capable optimization\ntechnique, followed by ALO (OS = 27) and ES (OS = 20). Due to the efficient\nperformance of these algorithms, the corresponding MLPs can be promising\nsubstitutes for traditional methods used for HL analysis.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am seeking to develop a robotic agent capable of comprehending user input and completing assigned tasks. I believe the agent's functionality can be separated into two segments: understanding commands and executing actions. First, I require a model that can interpret user commands, whether written or verbal, and translate them into a list of real-world actions. I believe a language model would be beneficial at this stage. However, I need advice on how to strike a balance between the size of the language model and the accuracy of its interpretation of the input. Additionally, I require a reinforcement learning framework for the action execution phase. However, I've noticed that users often provide detailed yet suboptimal instructions. Therefore, I am in search of a reward function that incorporates social psychology to enhance service quality. I anticipate an algorithm that empowers the agent to assess user input and carry out the most efficient actions while not straying too far from the original input.","['Robots operating alongside humans in diverse, stochastic environments must be\nable to accurately interpret natural language commands. These instructions\noften fall into one of two categories: those that specify a goal condition or\ntarget state, and those that specify explicit actions, or how to perform a\ngiven task. Recent approaches have used reward functions as a semantic\nrepresentation of goal-based commands, which allows for the use of a\nstate-of-the-art planner to find a policy for the given task. However, these\nreward functions cannot be directly used to represent action-oriented commands.\nWe introduce a new hybrid approach, the Deep Recurrent Action-Goal Grounding\nNetwork (DRAGGN), for task grounding and execution that handles natural\nlanguage from either category as input, and generalizes to unseen environments.\nOur robot-simulation results demonstrate that a system successfully\ninterpreting both goal-oriented and action-oriented task specifications brings\nus closer to robust natural language understanding for human-robot interaction.', ""We are increasingly surrounded by artificially intelligent technology that\ntakes decisions and executes actions on our behalf. This creates a pressing\nneed for general means to communicate with, instruct and guide artificial\nagents, with human language the most compelling means for such communication.\nTo achieve this in a scalable fashion, agents must be able to relate language\nto the world and to actions; that is, their understanding of language must be\ngrounded and embodied. However, learning grounded language is a notoriously\nchallenging problem in artificial intelligence research. Here we present an\nagent that learns to interpret language in a simulated 3D environment where it\nis rewarded for the successful execution of written instructions. Trained via a\ncombination of reinforcement and unsupervised learning, and beginning with\nminimal prior knowledge, the agent learns to relate linguistic symbols to\nemergent perceptual representations of its physical surroundings and to\npertinent sequences of actions. The agent's comprehension of language extends\nbeyond its prior experience, enabling it to apply familiar language to\nunfamiliar situations and to interpret entirely novel instructions. Moreover,\nthe speed with which this agent learns new words increases as its semantic\nknowledge grows. This facility for generalising and bootstrapping semantic\nknowledge indicates the potential of the present approach for reconciling\nambiguous natural language with the complexity of the physical world."", 'Recent work has described neural-network-based agents that are trained with\nreinforcement learning (RL) to execute language-like commands in simulated\nworlds, as a step towards an intelligent agent or robot that can be instructed\nby human users. However, the optimisation of multi-goal motor policies via deep\nRL from scratch requires many episodes of experience. Consequently,\ninstruction-following with deep RL typically involves language generated from\ntemplates (by an environment simulator), which does not reflect the varied or\nambiguous expressions of real users. Here, we propose a conceptually simple\nmethod for training instruction-following agents with deep RL that are robust\nto natural human instructions. By applying our method with a state-of-the-art\npre-trained text-based language model (BERT), on tasks requiring agents to\nidentify and position everyday objects relative to other objects in a\nnaturalistic 3D simulated room, we demonstrate substantially-above-chance\nzero-shot transfer from synthetic template commands to natural instructions\ngiven by humans. Our approach is a general recipe for training any deep\nRL-based system to interface with human users, and bridges the gap between two\nresearch directions of notable recent success: agent-centric motor behavior and\ntext-based representation learning.', 'It is highly desirable for robots that work alongside humans to be able to\nunderstand instructions in natural language. Existing language conditioned\nimitation learning models directly predict the actuator commands from the image\nobservation and the instruction text. Rather than directly predicting actuator\ncommands, we propose translating the natural language instruction to a Python\nfunction which queries the scene by accessing the output of the object detector\nand controls the robot to perform the specified task. This enables the use of\nnon-differentiable modules such as a constraint solver when computing commands\nto the robot. Moreover, the labels in this setup are significantly more\ninformative computer programs that capture the intent of the expert rather than\nteleoperated demonstrations. We show that the proposed method performs better\nthan training a neural network to directly predict the robot actions.', 'Interactive reinforcement learning agents use human feedback or instruction\nto help them learn in complex environments. Often, this feedback comes in the\nform of a discrete signal that is either positive or negative. While\ninformative, this information can be difficult to generalize on its own. In\nthis work, we explore how natural language advice can be used to provide a\nricher feedback signal to a reinforcement learning agent by extending policy\nshaping, a well-known Interactive reinforcement learning technique. Usually\npolicy shaping employs a human feedback policy to help an agent to learn more\nabout how to achieve its goal. In our case, we replace this human feedback\npolicy with policy generated based on natural language advice. We aim to\ninspect if the generated natural language reasoning provides support to a deep\nreinforcement learning agent to decide its actions successfully in any given\nenvironment. So, we design our model with three networks: first one is the\nexperience driven, next is the advice generator and third one is the advice\ndriven. While the experience driven reinforcement learning agent chooses its\nactions being influenced by the environmental reward, the advice driven neural\nnetwork with generated feedback by the advice generator for any new state\nselects its actions to assist the reinforcement learning agent to better policy\nshaping.', 'Our goal is to create an interactive natural language interface that\nefficiently and reliably learns from users to complete tasks in simulated\nrobotics settings. We introduce a neural semantic parsing system that learns\nnew high-level abstractions through decomposition: users interactively teach\nthe system by breaking down high-level utterances describing novel behavior\ninto low-level steps that it can understand. Unfortunately, existing methods\neither rely on grammars which parse sentences with limited flexibility, or\nneural sequence-to-sequence models that do not learn efficiently or reliably\nfrom individual examples. Our approach bridges this gap, demonstrating the\nflexibility of modern neural systems, as well as the one-shot reliable\ngeneralization of grammar-based methods. Our crowdsourced interactive\nexperiments suggest that over time, users complete complex tasks more\nefficiently while using our system by leveraging what they just taught. At the\nsame time, getting users to trust the system enough to be incentivized to teach\nhigh-level utterances is still an ongoing challenge. We end with a discussion\nof some of the obstacles we need to overcome to fully realize the potential of\nthe interactive paradigm.', 'The recently proposed ALFRED challenge task aims for a virtual robotic agent\nto complete complex multi-step everyday tasks in a virtual home environment\nfrom high-level natural language directives, such as ""put a hot piece of bread\non a plate"". Currently, the best-performing models are able to complete less\nthan 5% of these tasks successfully. In this work we focus on modeling the\ntranslation problem of converting natural language directives into detailed\nmulti-step sequences of actions that accomplish those goals in the virtual\nenvironment. We empirically demonstrate that it is possible to generate gold\nmulti-step plans from language directives alone without any visual input in 26%\nof unseen cases. When a small amount of visual information is incorporated,\nnamely the starting location in the virtual environment, our best-performing\nGPT-2 model successfully generates gold command sequences in 58% of cases. Our\nresults suggest that contextualized language models may provide strong visual\nsemantic planning modules for grounded virtual agents.', 'People are becoming increasingly comfortable using Digital Assistants (DAs)\nto interact with services or connected objects. However, for non-programming\nusers, the available possibilities for customizing their DA are limited and do\nnot include the possibility of teaching the assistant new tasks. To make the\nmost of the potential of DAs, users should be able to customize assistants by\ninstructing them through Natural Language (NL). To provide such\nfunctionalities, NL interpretation in traditional assistants should be\nimproved: (1) The intent identification system should be able to recognize new\nforms of known intents, and to acquire new intents as they are expressed by the\nuser. (2) In order to be adaptive to novel intents, the Natural Language\nUnderstanding module should be sample efficient, and should not rely on a\npretrained model. Rather, the system should continuously collect the training\ndata as it learns new intents from the user. In this work, we propose AidMe\n(Adaptive Intent Detection in Multi-Domain Environments), a user-in-the-loop\nadaptive intent detection framework that allows the assistant to adapt to its\nuser by learning his intents as their interaction progresses. AidMe builds its\nrepertoire of intents and collects data to train a model of semantic similarity\nevaluation that can discriminate between the learned intents and autonomously\ndiscover new forms of known intents. AidMe addresses two major issues - intent\nlearning and user adaptation - for instructable digital assistants. We\ndemonstrate the capabilities of AidMe as a standalone system by comparing it\nwith a one-shot learning system and a pretrained NLU module through simulations\nof interactions with a user. We also show how AidMe can smoothly integrate to\nan existing instructable digital assistant.', ""Wheelchair-mounted robotic arms (and other assistive robots) should help\ntheir users perform everyday tasks. One way robots can provide this assistance\nis shared autonomy. Within shared autonomy, both the human and robot maintain\ncontrol over the robot's motion: as the robot becomes confident it understands\nwhat the human wants, it increasingly intervenes to automate the task. But how\ndoes the robot know what tasks the human may want to perform in the first\nplace? Today's shared autonomy approaches often rely on prior knowledge: for\nexample, the robot must know the set of possible human goals a priori. In the\nlong-term, however, this prior knowledge will inevitably break down -- sooner\nor later the human will reach for a goal that the robot did not expect. In this\npaper we propose a learning approach to shared autonomy that takes advantage of\nrepeated interactions. Learning to assist humans would be impossible if they\nperformed completely different tasks at every interaction: but our insight is\nthat users living with physical disabilities repeat important tasks on a daily\nbasis (e.g., opening the fridge, making coffee, and having dinner). We\nintroduce an algorithm that exploits these repeated interactions to recognize\nthe human's task, replicate similar demonstrations, and return control when\nunsure. As the human repeatedly works with this robot, our approach continually\nlearns to assist tasks that were never specified beforehand: these tasks\ninclude both discrete goals (e.g., reaching a cup) and continuous skills (e.g.,\nopening a drawer). Across simulations and an in-person user study, we\ndemonstrate that robots leveraging our approach match existing shared autonomy\nmethods for known goals, and outperform imitation learning baselines on new\ntasks. See videos here: https://youtu.be/Plh4t3wQeIA"", 'We propose a developmental approach that allows a robot to interpret and\ndescribe the actions of human agents by reusing previous experience. The robot\nfirst learns the association between words and object affordances by\nmanipulating the objects in its environment. It then uses this information to\nlearn a mapping between its own actions and those performed by a human in a\nshared environment. It finally fuses the information from these two models to\ninterpret and describe human actions in light of its own experience. In our\nexperiments, we show that the model can be used flexibly to do inference on\ndifferent aspects of the scene. We can predict the effects of an action on the\nbasis of object properties. We can revise the belief that a certain action\noccurred, given the observed effects of the human action. In an early action\nrecognition fashion, we can anticipate the effects when the action has only\nbeen partially observed. By estimating the probability of words given the\nevidence and feeding them into a pre-defined grammar, we can generate relevant\ndescriptions of the scene. We believe that this is a step towards providing\nrobots with the fundamental skills to engage in social collaboration with\nhumans.', 'Behavioral skills or policies for autonomous agents are conventionally\nlearned from reward functions, via reinforcement learning, or from\ndemonstrations, via imitation learning. However, both modes of task\nspecification have their disadvantages: reward functions require manual\nengineering, while demonstrations require a human expert to be able to actually\nperform the task in order to generate the demonstration. Instruction following\nfrom natural language instructions provides an appealing alternative: in the\nsame way that we can specify goals to other humans simply by speaking or\nwriting, we would like to be able to specify tasks for our machines. However, a\nsingle instruction may be insufficient to fully communicate our intent or, even\nif it is, may be insufficient for an autonomous agent to actually understand\nhow to perform the desired task. In this work, we propose an interactive\nformulation of the task specification problem, where iterative language\ncorrections are provided to an autonomous agent, guiding it in acquiring the\ndesired skill. Our proposed language-guided policy learning algorithm can\nintegrate an instruction and a sequence of corrections to acquire new skills\nvery quickly. In our experiments, we show that this method can enable a policy\nto follow instructions and corrections for simulated navigation and\nmanipulation tasks, substantially outperforming direct, non-interactive\ninstruction following.', 'Humans generally use natural language to communicate task requirements to\neach other. Ideally, natural language should also be usable for communicating\ngoals to autonomous machines (e.g., robots) to minimize friction in task\nspecification. However, understanding and mapping natural language goals to\nsequences of states and actions is challenging. Specifically, existing work\nalong these lines has encountered difficulty in generalizing learned policies\nto new natural language goals and environments. In this paper, we propose a\nnovel adversarial inverse reinforcement learning algorithm to learn a\nlanguage-conditioned policy and reward function. To improve generalization of\nthe learned policy and reward function, we use a variational goal generator to\nrelabel trajectories and sample diverse goals during training. Our algorithm\noutperforms multiple baselines by a large margin on a vision-based natural\nlanguage instruction following dataset (Room-2-Room), demonstrating a promising\nadvance in enabling the use of natural language instructions in specifying\nagent goals.', 'Natural language understanding for robotics can require substantial domain-\nand platform-specific engineering. For example, for mobile robots to\npick-and-place objects in an environment to satisfy human commands, we can\nspecify the language humans use to issue such commands, and connect concept\nwords like red can to physical object properties. One way to alleviate this\nengineering for a new domain is to enable robots in human environments to adapt\ndynamically---continually learning new language constructions and perceptual\nconcepts. In this work, we present an end-to-end pipeline for translating\nnatural language commands to discrete robot actions, and use clarification\ndialogs to jointly improve language parsing and concept grounding. We train and\nevaluate this agent in a virtual setting on Amazon Mechanical Turk, and we\ntransfer the learned agent to a physical robot platform to demonstrate it in\nthe real world.', 'In this paper we present a framework for the teleoperation of pick-and-place\ntasks. We define a shared control policy that allows to blend between direct\nuser control and autonomous control based on user intent inference. One of the\nmain challenges in shared autonomy systems is to define the arbitration\nfunction, which decides when to let the autonomous agent take over. In this\nwork, we propose a model and training method to learn the arbitration function.\nOur model is based on a recurrent neural network that takes as input the state,\nintent prediction scores and user command to produce an arbitration between\nuser and robot commands. This work extends our previous work on differentiable\npolicies for shared autonomy. Differentiability of the policy is desirable to\nfurther train the shared autonomy system end-to-end. In this work we propose\ntraining of the arbitration function by using data from user performing the\ntask with shared control. We present initial results by teleoperating a gripper\nin a virtual environment using pre-trained motion generation and intent\nprediction. We compare our data aggregation training procedure to a handcrafted\narbitration function. Our preliminary results show the efficacy of the approach\nand shed light on limitations that we believe demonstrate the need for user\nadaptability in shared autonomy systems.', 'Recent successes combine reinforcement learning algorithms and deep neural\nnetworks, despite reinforcement learning not being widely applied to robotics\nand real world scenarios. This can be attributed to the fact that current\nstate-of-the-art, end-to-end reinforcement learning approaches still require\nthousands or millions of data samples to converge to a satisfactory policy and\nare subject to catastrophic failures during training. Conversely, in real world\nscenarios and after just a few data samples, humans are able to either provide\ndemonstrations of the task, intervene to prevent catastrophic actions, or\nsimply evaluate if the policy is performing correctly. This research\ninvestigates how to integrate these human interaction modalities to the\nreinforcement learning loop, increasing sample efficiency and enabling\nreal-time reinforcement learning in robotics and real world scenarios. This\nnovel theoretical foundation is called Cycle-of-Learning, a reference to how\ndifferent human interaction modalities, namely, task demonstration,\nintervention, and evaluation, are cycled and combined to reinforcement learning\nalgorithms. Results presented in this work show that the reward signal that is\nlearned based upon human interaction accelerates the rate of learning of\nreinforcement learning algorithms and that learning from a combination of human\ndemonstrations and interventions is faster and more sample efficient when\ncompared to traditional supervised learning algorithms. Finally,\nCycle-of-Learning develops an effective transition between policies learned\nusing human demonstrations and interventions to reinforcement learning. The\ntheoretical foundation developed by this research opens new research paths to\nhuman-agent teaming scenarios where autonomous agents are able to learn from\nhuman teammates and adapt to mission performance metrics in real-time and in\nreal world scenarios.', 'This article surveys reinforcement learning approaches in social robotics.\nReinforcement learning is a framework for decision-making problems in which an\nagent interacts through trial-and-error with its environment to discover an\noptimal behavior. Since interaction is a key component in both reinforcement\nlearning and social robotics, it can be a well-suited approach for real-world\ninteractions with physically embodied social robots. The scope of the paper is\nfocused particularly on studies that include social physical robots and\nreal-world human-robot interactions with users. We present a thorough analysis\nof reinforcement learning approaches in social robotics. In addition to a\nsurvey, we categorize existent reinforcement learning approaches based on the\nused method and the design of the reward mechanisms. Moreover, since\ncommunication capability is a prominent feature of social robots, we discuss\nand group the papers based on the communication medium used for reward\nformulation. Considering the importance of designing the reward function, we\nalso provide a categorization of the papers based on the nature of the reward.\nThis categorization includes three major themes: interactive reinforcement\nlearning, intrinsically motivated methods, and task performance-driven methods.\nThe benefits and challenges of reinforcement learning in social robotics,\nevaluation methods of the papers regarding whether or not they use subjective\nand algorithmic measures, a discussion in the view of real-world reinforcement\nlearning challenges and proposed solutions, the points that remain to be\nexplored, including the approaches that have thus far received less attention\nis also given in the paper. Thus, this paper aims to become a starting point\nfor researchers interested in using and applying reinforcement learning methods\nin this particular research field.', 'Robots navigating in human environments should use language to ask for\nassistance and be able to understand human responses. To study this challenge,\nwe introduce Cooperative Vision-and-Dialog Navigation, a dataset of over 2k\nembodied, human-human dialogs situated in simulated, photorealistic home\nenvironments. The Navigator asks questions to their partner, the Oracle, who\nhas privileged access to the best next steps the Navigator should take\naccording to a shortest path planner. To train agents that search an\nenvironment for a goal location, we define the Navigation from Dialog History\ntask. An agent, given a target object and a dialog history between humans\ncooperating to find that object, must infer navigation actions towards the goal\nin unexplored environments. We establish an initial, multi-modal\nsequence-to-sequence model and demonstrate that looking farther back in the\ndialog history improves performance. Sourcecode and a live interface demo can\nbe found at https://cvdn.dev/', ""Goal-oriented conversational agents are becoming prevalent in our daily\nlives. For these systems to engage users and achieve their goals, they need to\nexhibit appropriate social behavior as well as provide informative replies that\nguide users through tasks. The first component of the research in this paper\napplies statistical modeling techniques to understand conversations between\nusers and human agents for customer service. Analyses show that social language\nused by human agents is associated with greater users' responsiveness and task\ncompletion. The second component of the research is the construction of a\nconversational agent model capable of injecting social language into an agent's\nresponses while still preserving content. The model uses a sequence-to-sequence\ndeep learning architecture, extended with a social language understanding\nelement. Evaluation in terms of content preservation and social language level\nusing both human judgment and automatic linguistic measures shows that the\nmodel can generate responses that enable agents to address users' issues in a\nmore socially appropriate way."", 'Augmented-reality (AR) glasses that will have access to onboard sensors and\nan ability to display relevant information to the user present an opportunity\nto provide user assistance in quotidian tasks. Many such tasks can be\ncharacterized as object-rearrangement tasks. We introduce a novel framework for\ncomputing and displaying AR assistance that consists of (1) associating an\noptimal action sequence with the policy of an embodied agent and (2) presenting\nthis sequence to the user as suggestions in the AR system\'s heads-up display.\nThe embodied agent comprises a ""hybrid"" between the AR system and the user,\nwith the AR system\'s observation space (i.e., sensors) and the user\'s action\nspace (i.e., task-execution actions); its policy is learned by minimizing the\ntask-completion time. In this initial study, we assume that the AR system\'s\nobservations include the environment\'s map and localization of the objects and\nthe user. These choices allow us to formalize the problem of computing AR\nassistance for any object-rearrangement task as a planning problem,\nspecifically as a capacitated vehicle-routing problem. Further, we introduce a\nnovel AR simulator that can enable web-based evaluation of AR-like assistance\nand associated at-scale data collection via the Habitat simulator for embodied\nartificial intelligence. Finally, we perform a study that evaluates user\nresponse to the proposed form of AR assistance on a specific quotidian\nobject-rearrangement task, house cleaning, using our proposed AR simulator on\nmechanical turk. In particular, we study the effect of the proposed AR\nassistance on users\' task performance and sense of agency over a range of task\ndifficulties. Our results indicate that providing users with such assistance\nimproves their overall performance and while users report a negative impact to\ntheir agency, they may still prefer the proposed assistance to having no\nassistance at all.', 'Human collaborators coordinate effectively their actions through both verbal\nand non-verbal communication. We believe that the the same should hold for\nhuman-robot teams. We propose a formalism that enables a robot to decide\noptimally between doing a task and issuing an utterance. We focus on two types\nof utterances: verbal commands, where the robot expresses how it wants its\nhuman teammate to behave, and state-conveying actions, where the robot explains\nwhy it is behaving this way. Human subject experiments show that enabling the\nrobot to issue verbal commands is the most effective form of communicating\nobjectives, while retaining user trust in the robot. Communicating why\ninformation should be done judiciously, since many participants questioned the\ntruthfulness of the robot statements.', 'Modern approaches to text to speech require the entire input character\nsequence to be processed before any audio is synthesised. This latency limits\nthe suitability of such models for time-sensitive tasks like simultaneous\ninterpretation. Interleaving the action of reading a character with that of\nsynthesising audio reduces this latency. However, the order of this sequence of\ninterleaved actions varies across sentences, which raises the question of how\nthe actions should be chosen. We propose a reinforcement learning based\nframework to train an agent to make this decision. We compare our performance\nagainst that of deterministic, rule-based systems. Our results demonstrate that\nour agent successfully balances the trade-off between the latency of audio\ngeneration and the quality of synthesised audio. More broadly, we show that\nneural sequence-to-sequence models can be adapted to run in an incremental\nmanner.', ""We address the problem of bootstrapping language acquisition for an\nartificial system similarly to what is observed in experiments with human\ninfants. Our method works by associating meanings to words in manipulation\ntasks, as a robot interacts with objects and listens to verbal descriptions of\nthe interactions. The model is based on an affordance network, i.e., a mapping\nbetween robot actions, robot perceptions, and the perceived effects of these\nactions upon objects. We extend the affordance model to incorporate spoken\nwords, which allows us to ground the verbal symbols to the execution of actions\nand the perception of the environment. The model takes verbal descriptions of a\ntask as the input and uses temporal co-occurrence to create links between\nspeech utterances and the involved objects, actions, and effects. We show that\nthe robot is able form useful word-to-meaning associations, even without\nconsidering grammatical structure in the learning process and in the presence\nof recognition errors. These word-to-meaning associations are embedded in the\nrobot's own understanding of its actions. Thus, they can be directly used to\ninstruct the robot to perform tasks and also allow to incorporate context in\nthe speech recognition task. We believe that the encouraging results with our\napproach may afford robots with a capacity to acquire language descriptors in\ntheir operation's environment as well as to shed some light as to how this\nchallenging process develops with human infants."", 'Machine Learning models become increasingly proficient in complex tasks.\nHowever, even for experts in the field, it can be difficult to understand what\nthe model learned. This hampers trust and acceptance, and it obstructs the\npossibility to correct the model. There is therefore a need for transparency of\nmachine learning models. The development of transparent classification models\nhas received much attention, but there are few developments for achieving\ntransparent Reinforcement Learning (RL) models. In this study we propose a\nmethod that enables a RL agent to explain its behavior in terms of the expected\nconsequences of state transitions and outcomes. First, we define a translation\nof states and actions to a description that is easier to understand for human\nusers. Second, we developed a procedure that enables the agent to obtain the\nconsequences of a single action, as well as its entire policy. The method\ncalculates contrasts between the consequences of a policy derived from a user\nquery, and of the learned policy of the agent. Third, a format for generating\nexplanations was constructed. A pilot survey study was conducted to explore\npreferences of users for different explanation properties. Results indicate\nthat human users tend to favor explanations about policy rather than about\nsingle actions.', 'One of the long-term goals of artificial intelligence is to build an agent\nthat can communicate intelligently with human in natural language. Most\nexisting work on natural language learning relies heavily on training over a\npre-collected dataset with annotated labels, leading to an agent that\nessentially captures the statistics of the fixed external training data. As the\ntraining data is essentially a static snapshot representation of the knowledge\nfrom the annotator, the agent trained this way is limited in adaptiveness and\ngeneralization of its behavior. Moreover, this is very different from the\nlanguage learning process of humans, where language is acquired during\ncommunication by taking speaking action and learning from the consequences of\nspeaking action in an interactive manner. This paper presents an interactive\nsetting for grounded natural language learning, where an agent learns natural\nlanguage by interacting with a teacher and learning from feedback, thus\nlearning and improving language skills while taking part in the conversation.\nTo achieve this goal, we propose a model which incorporates both imitation and\nreinforcement by leveraging jointly sentence and reward feedbacks from the\nteacher. Experiments are conducted to validate the effectiveness of the\nproposed approach.', ""The combination of deep neural network models and reinforcement learning\nalgorithms can make it possible to learn policies for robotic behaviors that\ndirectly read in raw sensory inputs, such as camera images, effectively\nsubsuming both estimation and control into one model. However, real-world\napplications of reinforcement learning must specify the goal of the task by\nmeans of a manually programmed reward function, which in practice requires\neither designing the very same perception pipeline that end-to-end\nreinforcement learning promises to avoid, or else instrumenting the environment\nwith additional sensors to determine if the task has been performed\nsuccessfully. In this paper, we propose an approach for removing the need for\nmanual engineering of reward specifications by enabling a robot to learn from a\nmodest number of examples of successful outcomes, followed by actively\nsolicited queries, where the robot shows the user a state and asks for a label\nto determine whether that state represents successful completion of the task.\nWhile requesting labels for every single state would amount to asking the user\nto manually provide the reward signal, our method requires labels for only a\ntiny fraction of the states seen during training, making it an efficient and\npractical approach for learning skills without manually engineered rewards. We\nevaluate our method on real-world robotic manipulation tasks where the\nobservations consist of images viewed by the robot's camera. In our\nexperiments, our method effectively learns to arrange objects, place books, and\ndrape cloth, directly from images and without any manually specified reward\nfunctions, and with only 1-4 hours of interaction with the real world.""]","[""Reinforcement learning (RL) can be used to create a tactical decision-making\nagent for autonomous driving. However, previous approaches only output\ndecisions and do not provide information about the agent's confidence in the\nrecommended actions. This paper investigates how a Bayesian RL technique, based\non an ensemble of neural networks with additional randomized prior functions\n(RPF), can be used to estimate the uncertainty of decisions in autonomous\ndriving. A method for classifying whether or not an action should be considered\nsafe is also introduced. The performance of the ensemble RPF method is\nevaluated by training an agent on a highway driving scenario. It is shown that\nthe trained agent can estimate the uncertainty of its decisions and indicate an\nunacceptable level when the agent faces a situation that is far from the\ntraining distribution. Furthermore, within the training distribution, the\nensemble RPF agent outperforms a standard Deep Q-Network agent. In this study,\nthe estimated uncertainty is used to choose safe actions in unknown situations.\nHowever, the uncertainty information could also be used to identify situations\nthat should be added to the training process."", 'Obtaining policies that can generalise to new environments in reinforcement\nlearning is challenging. In this work, we demonstrate that language\nunderstanding via a reading policy learner is a promising vehicle for\ngeneralisation to new environments. We propose a grounded policy learning\nproblem, Read to Fight Monsters (RTFM), in which the agent must jointly reason\nover a language goal, relevant dynamics described in a document, and\nenvironment observations. We procedurally generate environment dynamics and\ncorresponding language descriptions of the dynamics, such that agents must read\nto understand new environment dynamics instead of memorising any particular\ninformation. In addition, we propose txt2$\\pi$, a model that captures three-way\ninteractions between the goal, document, and observations. On RTFM, txt2$\\pi$\ngeneralises to new environments with dynamics not seen during training via\nreading. Furthermore, our model outperforms baselines such as FiLM and\nlanguage-conditioned CNNs on RTFM. Through curriculum learning, txt2$\\pi$\nproduces policies that excel on complex RTFM tasks requiring several reasoning\nand coreference steps.', ""We study the problem of designing AI agents that can robustly cooperate with\npeople in human-machine partnerships. Our work is inspired by real-life\nscenarios in which an AI agent, e.g., a virtual assistant, has to cooperate\nwith new users after its deployment. We model this problem via a parametric MDP\nframework where the parameters correspond to a user's type and characterize her\nbehavior. In the test phase, the AI agent has to interact with a user of\nunknown type. Our approach to designing a robust AI agent relies on observing\nthe user's actions to make inferences about the user's type and adapting its\npolicy to facilitate efficient cooperation. We show that without being\nadaptive, an AI agent can end up performing arbitrarily bad in the test phase.\nWe develop two algorithms for computing policies that automatically adapt to\nthe user in the test phase. We demonstrate the effectiveness of our approach in\nsolving a two-agent collaborative task."", ""We tackle a task where an agent learns to navigate in a 2D maze-like\nenvironment called XWORLD. In each session, the agent perceives a sequence of\nraw-pixel frames, a natural language command issued by a teacher, and a set of\nrewards. The agent learns the teacher's language from scratch in a grounded and\ncompositional manner, such that after training it is able to correctly execute\nzero-shot commands: 1) the combination of words in the command never appeared\nbefore, and/or 2) the command contains new object concepts that are learned\nfrom another task but never learned from navigation. Our deep framework for the\nagent is trained end to end: it learns simultaneously the visual\nrepresentations of the environment, the syntax and semantics of the language,\nand the action module that outputs actions. The zero-shot learning capability\nof our framework results from its compositionality and modularity with\nparameter tying. We visualize the intermediate outputs of the framework,\ndemonstrating that the agent truly understands how to solve the problem. We\nbelieve that our results provide some preliminary insights on how to train an\nagent with similar abilities in a 3D environment."", 'Assistive robotic arms enable users with physical disabilities to perform\neveryday tasks without relying on a caregiver. Unfortunately, the very\ndexterity that makes these arms useful also makes them challenging to\nteleoperate: the robot has more degrees-of-freedom than the human can directly\ncoordinate with a handheld joystick. Our insight is that we can make assistive\nrobots easier for humans to control by leveraging latent actions. Latent\nactions provide a low-dimensional embedding of high-dimensional robot behavior:\nfor example, one latent dimension might guide the assistive arm along a pouring\nmotion. In this paper, we design a teleoperation algorithm for assistive robots\nthat learns latent actions from task demonstrations. We formulate the\ncontrollability, consistency, and scaling properties that user-friendly latent\nactions should have, and evaluate how different low-dimensional embeddings\ncapture these properties. Finally, we conduct two user studies on a robotic arm\nto compare our latent action approach to both state-of-the-art shared autonomy\nbaselines and a teleoperation strategy currently used by assistive arms.\nParticipants completed assistive eating and cooking tasks more efficiently when\nleveraging our latent actions, and also subjectively reported that latent\nactions made the task easier to perform. The video accompanying this paper can\nbe found at: https://youtu.be/wjnhrzugBj4.', 'Building intelligent agents that can communicate with and learn from humans\nin natural language is of great value. Supervised language learning is limited\nby the ability of capturing mainly the statistics of training data, and is\nhardly adaptive to new scenarios or flexible for acquiring new knowledge\nwithout inefficient retraining or catastrophic forgetting. We highlight the\nperspective that conversational interaction serves as a natural interface both\nfor language learning and for novel knowledge acquisition and propose a joint\nimitation and reinforcement approach for grounded language learning through an\ninteractive conversational game. The agent trained with this approach is able\nto actively acquire information by asking questions about novel objects and use\nthe just-learned knowledge in subsequent conversations in a one-shot fashion.\nResults compared with other methods verified the effectiveness of the proposed\napproach.', ""In the vision and language navigation task, the agent may encounter ambiguous\nsituations that are hard to interpret by just relying on visual information and\nnatural language instructions. We propose an interactive learning framework to\nendow the agent with the ability to ask for users' help in such situations. As\npart of this framework, we investigate multiple learning approaches for the\nagent with different levels of complexity. The simplest model-confusion-based\nmethod lets the agent ask questions based on its confusion, relying on the\npredefined confidence threshold of a next action prediction model. To build on\nthis confusion-based method, the agent is expected to demonstrate more\nsophisticated reasoning such that it discovers the timing and locations to\ninteract with a human. We achieve this goal using reinforcement learning (RL)\nwith a proposed reward shaping term, which enables the agent to ask questions\nonly when necessary. The success rate can be boosted by at least 15% with only\none question asked on average during the navigation. Furthermore, we show that\nthe RL agent is capable of adjusting dynamically to noisy human responses.\nFinally, we design a continual learning strategy, which can be viewed as a data\naugmentation method, for the agent to improve further utilizing its interaction\nhistory with a human. We demonstrate the proposed strategy is substantially\nmore realistic and data-efficient compared to previously proposed\npre-exploration techniques."", 'Modern robotics applications that involve human-robot interaction require\nrobots to be able to communicate with humans seamlessly and effectively.\nNatural language provides a flexible and efficient medium through which robots\ncan exchange information with their human partners. Significant advancements\nhave been made in developing robots capable of interpreting free-form\ninstructions, but less attention has been devoted to endowing robots with the\nability to generate natural language. We propose a navigational guide model\nthat enables robots to generate natural language instructions that allow humans\nto navigate a priori unknown environments. We first decide which information to\nshare with the user according to their preferences, using a policy trained from\nhuman demonstrations via inverse reinforcement learning. We then ""translate""\nthis information into a natural language instruction using a neural\nsequence-to-sequence model that learns to generate free-form instructions from\nnatural language corpora. We evaluate our method on a benchmark route\ninstruction dataset and achieve a BLEU score of 72.18% when compared to\nhuman-generated reference instructions. We additionally conduct navigation\nexperiments with human participants that demonstrate that our method generates\ninstructions that people follow as accurately and easily as those produced by\nhumans.', ""As environments involving both robots and humans become increasingly common,\nso does the need to account for people during planning. To plan effectively,\nrobots must be able to respond to and sometimes influence what humans do. This\nrequires a human model which predicts future human actions. A simple model may\nassume the human will continue what they did previously; a more complex one\nmight predict that the human will act optimally, disregarding the robot;\nwhereas an even more complex one might capture the robot's ability to influence\nthe human. These models make different trade-offs between computational time\nand performance of the resulting robot plan. Using only one model of the human\neither wastes computational resources or is unable to handle critical\nsituations. In this work, we give the robot access to a suite of human models\nand enable it to assess the performance-computation trade-off online. By\nestimating how an alternate model could improve human prediction and how that\nmay translate to performance gain, the robot can dynamically switch human\nmodels whenever the additional computation is justified. Our experiments in a\ndriving simulator showcase how the robot can achieve performance comparable to\nalways using the best human model, but with greatly reduced computation."", 'When deploying autonomous agents in the real world, we need effective ways of\ncommunicating objectives to them. Traditional skill learning has revolved\naround reinforcement and imitation learning, each with rigid constraints on the\nformat of information exchanged between the human and the agent. While scalar\nrewards carry little information, demonstrations require significant effort to\nprovide and may carry more information than is necessary. Furthermore, rewards\nand demonstrations are often defined and collected before training begins, when\nthe human is most uncertain about what information would help the agent. In\ncontrast, when humans communicate objectives with each other, they make use of\na large vocabulary of informative behaviors, including non-verbal\ncommunication, and often communicate throughout learning, responding to\nobserved behavior. In this way, humans communicate intent with minimal effort.\nIn this paper, we propose such interactive learning as an alternative to reward\nor demonstration-driven learning. To accomplish this, we introduce a\nmulti-agent training framework that enables an agent to learn from another\nagent who knows the current task. Through a series of experiments, we\ndemonstrate the emergence of a variety of interactive learning behaviors,\nincluding information-sharing, information-seeking, and question-answering.\nMost importantly, we find that our approach produces an agent that is capable\nof learning interactively from a human user, without a set of explicit\ndemonstrations or a reward function, and achieving significantly better\nperformance cooperatively with a human than a human performing the task alone.', 'We present a hierarchical planning and control framework that enables an\nagent to perform various tasks and adapt to a new task flexibly. Rather than\nlearning an individual policy for each particular task, the proposed framework,\nDISH, distills a hierarchical policy from a set of tasks by representation and\nreinforcement learning. The framework is based on the idea of latent variable\nmodels that represent high-dimensional observations using low-dimensional\nlatent variables. The resulting policy consists of two levels of hierarchy: (i)\na planning module that reasons a sequence of latent intentions that would lead\nto an optimistic future and (ii) a feedback control policy, shared across the\ntasks, that executes the inferred intention. Because the planning is performed\nin low-dimensional latent space, the learned policy can immediately be used to\nsolve or adapt to new tasks without additional training. We demonstrate the\nproposed framework can learn compact representations (3- and 1-dimensional\nlatent states and commands for a humanoid with 197- and 36-dimensional state\nfeatures and actions) while solving a small number of imitation tasks, and the\nresulting policy is directly applicable to other types of tasks, i.e.,\nnavigation in cluttered environments. Video: https://youtu.be/HQsQysUWOhg', 'This paper presents a research platform that supports spoken dialogue\ninteraction with multiple robots. The demonstration showcases our crafted\nMultiBot testing scenario in which users can verbally issue search, navigate,\nand follow instructions to two robotic teammates: a simulated ground robot and\nan aerial robot. This flexible language and robotic platform takes advantage of\nexisting tools for speech recognition and dialogue management that are\ncompatible with new domains, and implements an inter-agent communication\nprotocol (tactical behavior specification), where verbal instructions are\nencoded for tasks assigned to the appropriate robot.', 'We demonstrate a reinforcement learning agent which uses a compositional\nrecurrent neural network that takes as input an LTL formula and determines\nsatisfying actions. The input LTL formulas have never been seen before, yet the\nnetwork performs zero-shot generalization to satisfy them. This is a novel form\nof multi-task learning for RL agents where agents learn from one diverse set of\ntasks and generalize to a new set of diverse tasks. The formulation of the\nnetwork enables this capacity to generalize. We demonstrate this ability in two\ndomains. In a symbolic domain, the agent finds a sequence of letters that is\naccepted. In a Minecraft-like environment, the agent finds a sequence of\nactions that conform to the formula. While prior work could learn to execute\none formula reliably given examples of that formula, we demonstrate how to\nencode all formulas reliably. This could form the basis of new multitask agents\nthat discover sub-tasks and execute them without any additional training, as\nwell as the agents which follow more complex linguistic commands. The\nstructures required for this generalization are specific to LTL formulas, which\nopens up an interesting theoretical question: what structures are required in\nneural networks for zero-shot generalization to different logics?', ""We present a novel approach to shared control of human-machine systems. Our\nmethod assumes no a priori knowledge of the system dynamics. Instead, we learn\nboth the dynamics and information about the user's interaction from observation\nthrough the use of the Koopman operator. Using the learned model, we define an\noptimization problem to compute the optimal policy for a given task, and\ncompare the user input to the optimal input. We demonstrate the efficacy of our\napproach with a user study. We also analyze the individual nature of the\nlearned models by comparing the effectiveness of our approach when the\ndemonstration data comes from a user's own interactions, from the interactions\nof a group of users and from a domain expert. Positive results include\nstatistically significant improvements on task metrics when comparing a\nuser-only control paradigm with our shared control paradigm. Surprising results\ninclude findings that suggest that individualizing the model based on a user's\nown data does not effect the ability to learn a useful dynamic system. We\nexplore this tension as it relates to developing human-in-the-loop systems\nfurther in the discussion."", ""Inferring intent from observed behavior has been studied extensively within\nthe frameworks of Bayesian inverse planning and inverse reinforcement learning.\nThese methods infer a goal or reward function that best explains the actions of\nthe observed agent, typically a human demonstrator. Another agent can use this\ninferred intent to predict, imitate, or assist the human user. However, a\ncentral assumption in inverse reinforcement learning is that the demonstrator\nis close to optimal. While models of suboptimal behavior exist, they typically\nassume that suboptimal actions are the result of some type of random noise or a\nknown cognitive bias, like temporal inconsistency. In this paper, we take an\nalternative approach, and model suboptimal behavior as the result of internal\nmodel misspecification: the reason that user actions might deviate from\nnear-optimal actions is that the user has an incorrect set of beliefs about the\nrules -- the dynamics -- governing how actions affect the environment. Our\ninsight is that while demonstrated actions may be suboptimal in the real world,\nthey may actually be near-optimal with respect to the user's internal model of\nthe dynamics. By estimating these internal beliefs from observed behavior, we\narrive at a new method for inferring intent. We demonstrate in simulation and\nin a user study with 12 participants that this approach enables us to more\naccurately model human intent, and can be used in a variety of applications,\nincluding offering assistance in a shared autonomy framework and inferring\nhuman preferences."", 'Designing reward functions is difficult: the designer has to specify what to\ndo (what it means to complete the task) as well as what not to do (side effects\nthat should be avoided while completing the task). To alleviate the burden on\nthe reward designer, we propose an algorithm to automatically generate an\nauxiliary reward function that penalizes side effects. This auxiliary objective\nrewards the ability to complete possible future tasks, which decreases if the\nagent causes side effects during the current task. The future task reward can\nalso give the agent an incentive to interfere with events in the environment\nthat make future tasks less achievable, such as irreversible actions by other\nagents. To avoid this interference incentive, we introduce a baseline policy\nthat represents a default course of action (such as doing nothing), and use it\nto filter out future tasks that are not achievable by default. We formally\ndefine interference incentives and show that the future task approach with a\nbaseline policy avoids these incentives in the deterministic case. Using\ngridworld environments that test for side effects and interference, we show\nthat our method avoids interference and is more effective for avoiding side\neffects than the common approach of penalizing irreversible actions.', 'While reinforcement learning has achieved considerable successes in recent\nyears, state-of-the-art models are often still limited by the size of state and\naction spaces. Model-free reinforcement learning approaches use some form of\nstate representations and the latest work has explored embedding techniques for\nactions, both with the aim of achieving better generalization and\napplicability. However, these approaches consider only states or actions,\nignoring the interaction between them when generating embedded representations.\nIn this work, we establish the theoretical foundations for the validity of\ntraining a reinforcement learning agent using embedded states and actions. We\nthen propose a new approach for jointly learning embeddings for states and\nactions that combines aspects of model-free and model-based reinforcement\nlearning, which can be applied in both discrete and continuous domains.\nSpecifically, we use a model of the environment to obtain embeddings for states\nand actions and present a generic architecture that leverages these to learn a\npolicy. In this way, the embedded representations obtained via our approach\nenable better generalization over both states and actions by capturing\nsimilarities in the embedding spaces. Evaluations of our approach on several\ngaming, robotic control, and recommender systems show it significantly\noutperforms state-of-the-art models in both discrete/continuous domains with\nlarge state/action spaces, thus confirming its efficacy.', ""Maximizing the utility of human-robot teams in disaster response and search\nand rescue (SAR) missions remains to be a challenging problem. This is due to\nthe dynamic, uncertain nature of the environment and the variability in\ncognitive performance of the human operators. By having an autonomous agent\nshare control with the operator, we can achieve near-optimal performance by\naugmenting the operator's input and compensate for the factors resulting in\ndegraded performance. What this solution does not consider though is the human\ninput latency and errors caused by potential hardware failures that can occur\nduring task completion when operating in disaster response and SAR scenarios.\nIn this paper, we propose the use of blended shared control (BSC) architecture\nto address these issues and investigate the architecture's performance in\nconstrained, dynamic environments with a differential drive robot that has\ninput latency and erroneous odometry feedback. We conduct a validation study\n(n=12) for our control architecture and then a user study (n=14) in 2 different\nenvironments that are unknown to both the human operator and the autonomous\nagent. The results demonstrate that the BSC architecture can prevent collisions\nand enhance operator performance without the need of a complete transfer of\ncontrol between the human operator and autonomous agent."", ""One obstacle to applying reinforcement learning algorithms to real-world\nproblems is the lack of suitable reward functions. Designing such reward\nfunctions is difficult in part because the user only has an implicit\nunderstanding of the task objective. This gives rise to the agent alignment\nproblem: how do we create agents that behave in accordance with the user's\nintentions? We outline a high-level research direction to solve the agent\nalignment problem centered around reward modeling: learning a reward function\nfrom interaction with the user and optimizing the learned reward function with\nreinforcement learning. We discuss the key challenges we expect to face when\nscaling reward modeling to complex and general domains, concrete approaches to\nmitigate these challenges, and ways to establish trust in the resulting agents."", 'Learning predictive models from interaction with the world allows an agent,\nsuch as a robot, to learn about how the world works, and then use this learned\nmodel to plan coordinated sequences of actions to bring about desired outcomes.\nHowever, learning a model that captures the dynamics of complex skills\nrepresents a major challenge: if the agent needs a good model to perform these\nskills, it might never be able to collect the experience on its own that is\nrequired to learn these delicate and complex behaviors. Instead, we can imagine\naugmenting the training set with observational data of other agents, such as\nhumans. Such data is likely more plentiful, but represents a different\nembodiment. For example, videos of humans might show a robot how to use a tool,\nbut (i) are not annotated with suitable robot actions, and (ii) contain a\nsystematic distributional shift due to the embodiment differences between\nhumans and robots. We address the first challenge by formulating the\ncorresponding graphical model and treating the action as an observed variable\nfor the interaction data and an unobserved variable for the observation data,\nand the second challenge by using a domain-dependent prior. In addition to\ninteraction data, our method is able to leverage videos of passive observations\nin a driving dataset and a dataset of robotic manipulation videos. A robotic\nplanning agent equipped with our method can learn to use tools in a tabletop\nrobotic manipulation setting by observing humans without ever seeing a robotic\nvideo of tool use.', 'To interact with humans in collaborative environments, machines need to be\nable to predict (i.e., anticipate) future events, and execute actions in a\ntimely manner. However, the observation of the human limb movements may not be\nsufficient to anticipate their actions unambiguously. In this work, we consider\ntwo additional sources of information (i.e., context) over time, gaze, movement\nand object information, and study how these additional contextual cues improve\nthe action anticipation performance. We address action anticipation as a\nclassification task, where the model takes the available information as the\ninput and predicts the most likely action. We propose to use the uncertainty\nabout each prediction as an online decision-making criterion for action\nanticipation. Uncertainty is modeled as a stochastic process applied to a\ntime-based neural network architecture, which improves the conventional\nclass-likelihood (i.e., deterministic) criterion. The main contributions of\nthis paper are four-fold: (i) We propose a novel and effective decision-making\ncriterion that can be used to anticipate actions even in situations of high\nambiguity; (ii) we propose a deep architecture that outperforms previous\nresults in the action anticipation task when using the Acticipate collaborative\ndataset; (iii) we show that contextual information is important to disambiguate\nthe interpretation of similar actions; and (iv) we also provide a formal\ndescription of three existing performance metrics that can be easily used to\nevaluate action anticipation models.Our results on the Acticipate dataset\nshowed the importance of contextual information and the uncertainty criterion\nfor action anticipation. We achieve an average accuracy of 98.75% in the\nanticipation task using only an average of 25% of observations.', ""Contemporary coding education often presents students with the task of\ndeveloping programs that have user interaction and complex dynamic systems,\nsuch as mouse based games. While pedagogically compelling, there are no\ncontemporary autonomous methods for providing feedback. Notably, interactive\nprograms are impossible to grade by traditional unit tests. In this paper we\nformalize the challenge of providing feedback to interactive programs as a task\nof classifying Markov Decision Processes (MDPs). Each student's program fully\nspecifies an MDP where the agent needs to operate and decide, under reasonable\ngeneralization, if the dynamics and reward model of the input MDP should be\ncategorized as correct or broken. We demonstrate that by designing a\ncooperative objective between an agent and an autoregressive model, we can use\nthe agent to sample differential trajectories from the input MDP that allows a\nclassifier to determine membership: Play to Grade. Our method enables an\nautomatic feedback system for interactive code assignments. We release a\ndataset of 711,274 anonymized student submissions to a single assignment with\nhand-coded bug labels to support future research."", 'Autonomous driving is a complex task, which has been tackled since the first\nself-driving car ALVINN in 1989, with a supervised learning approach, or\nbehavioral cloning (BC). In BC, a neural network is trained with state-action\npairs that constitute the training set made by an expert, i.e., a human driver.\nHowever, this type of imitation learning does not take into account the\ntemporal dependencies that might exist between actions taken in different\nmoments of a navigation trajectory. These type of tasks are better handled by\nreinforcement learning (RL) algorithms, which need to define a reward function.\nOn the other hand, more recent approaches to imitation learning, such as\nGenerative Adversarial Imitation Learning (GAIL), can train policies without\nexplicitly requiring to define a reward function, allowing an agent to learn by\ntrial and error directly on a training set of expert trajectories. In this\nwork, we propose two variations of GAIL for autonomous navigation of a vehicle\nin the realistic CARLA simulation environment for urban scenarios. Both of them\nuse the same network architecture, which process high dimensional image input\nfrom three frontal cameras, and other nine continuous inputs representing the\nvelocity, the next point from the sparse trajectory and a high-level driving\ncommand. We show that both of them are capable of imitating the expert\ntrajectory from start to end after training ends, but the GAIL loss function\nthat is augmented with BC outperforms the former in terms of convergence time\nand training stability.', ""We analyze the language learned by an agent trained with reinforcement\nlearning as a component of the ActiveQA system [Buck et al., 2017]. In\nActiveQA, question answering is framed as a reinforcement learning task in\nwhich an agent sits between the user and a black box question-answering system.\nThe agent learns to reformulate the user's questions to elicit the optimal\nanswers. It probes the system with many versions of a question that are\ngenerated via a sequence-to-sequence question reformulation model, then\naggregates the returned evidence to find the best answer. This process is an\ninstance of \\emph{machine-machine} communication. The question reformulation\nmodel must adapt its language to increase the quality of the answers returned,\nmatching the language of the question answering system. We find that the agent\ndoes not learn transformations that align with semantic intuitions but\ndiscovers through learning classical information retrieval techniques such as\ntf-idf re-weighting and stemming."", 'Despite widespread interests in reinforcement-learning for task-oriented\ndialogue systems, several obstacles can frustrate research and development\nprogress. First, reinforcement learners typically require interaction with the\nenvironment, so conventional dialogue corpora cannot be used directly. Second,\neach task presents specific challenges, requiring separate corpus of\ntask-specific annotated data. Third, collecting and annotating human-machine or\nhuman-human conversations for task-oriented dialogues requires extensive domain\nknowledge. Because building an appropriate dataset can be both financially\ncostly and time-consuming, one popular approach is to build a user simulator\nbased upon a corpus of example dialogues. Then, one can train reinforcement\nlearning agents in an online fashion as they interact with the simulator.\nDialogue agents trained on these simulators can serve as an effective starting\npoint. Once agents master the simulator, they may be deployed in a real\nenvironment to interact with humans, and continue to be trained online. To ease\nempirical algorithmic comparisons in dialogues, this paper introduces a new,\npublicly available simulation framework, where our simulator, designed for the\nmovie-booking domain, leverages both rules and collected data. The simulator\nsupports two tasks: movie ticket booking and movie seeking. Finally, we\ndemonstrate several agents and detail the procedure to add and test your own\nagent in the proposed framework.', 'Deep learning often requires the manual collection and annotation of a\ntraining set. On robotic platforms, can we partially automate this task by\ntraining the robot to be curious, i.e., to seek out beneficial training\ninformation in the environment? In this work, we address the problem of\ncuriosity as it relates to online, real-time, human-in-the-loop training of an\nobject detection algorithm onboard a drone, where motion is constrained to two\ndimensions. We use a 3D simulation environment and deep reinforcement learning\nto train a curiosity agent to, in turn, train the object detection model. This\nagent could have one of two conflicting objectives: train as quickly as\npossible, or train with minimal human input. We outline a reward function that\nallows the curiosity agent to learn either of these objectives, while taking\ninto account some of the physical characteristics of the drone platform on\nwhich it is meant to run. In addition, We show that we can weigh the importance\nof achieving these objectives by adjusting a parameter in the reward function.', 'Multi-agent reinforcement learning involves multiple agents interacting with\neach other and a shared environment to complete tasks. When rewards provided by\nthe environment are sparse, agents may not receive immediate feedback on the\nquality of actions that they take, thereby affecting learning of policies. In\nthis paper, we propose a method called Shaping Advice in deep Multi-agent\nreinforcement learning (SAM) to augment the reward signal from the environment\nwith an additional reward termed shaping advice. The shaping advice is given by\na difference of potential functions at consecutive time-steps. Each potential\nfunction is a function of observations and actions of the agents. The shaping\nadvice needs to be specified only once at the start of training, and can be\neasily provided by non-experts. We show through theoretical analyses and\nexperimental validation that shaping advice provided by SAM does not distract\nagents from completing tasks specified by the environment reward.\nTheoretically, we prove that convergence of policy gradients and value\nfunctions when using SAM implies convergence of these quantities in the absence\nof SAM. Experimentally, we evaluate SAM on three tasks in the multi-agent\nParticle World environment that have sparse rewards. We observe that using SAM\nresults in agents learning policies to complete tasks faster, and obtain higher\nrewards than: i) using sparse rewards alone; ii) a state-of-the-art reward\nredistribution method.', 'Most model-free reinforcement learning methods leverage state representations\n(embeddings) for generalization, but either ignore structure in the space of\nactions or assume the structure is provided a priori. We show how a policy can\nbe decomposed into a component that acts in a low-dimensional space of action\nrepresentations and a component that transforms these representations into\nactual actions. These representations improve generalization over large, finite\naction sets by allowing the agent to infer the outcomes of actions similar to\nactions already taken. We provide an algorithm to both learn and use action\nrepresentations and provide conditions for its convergence. The efficacy of the\nproposed method is demonstrated on large-scale real-world problems.', ""The intuitive collaboration of humans and intelligent robots (embodied AI) in\nthe real-world is an essential objective for many desirable applications of\nrobotics. Whilst there is much research regarding explicit communication, we\nfocus on how humans and robots interact implicitly, on motor adaptation level.\nWe present a real-world setup of a human-robot collaborative maze game,\ndesigned to be non-trivial and only solvable through collaboration, by limiting\nthe actions to rotations of two orthogonal axes, and assigning each axes to one\nplayer. This results in neither the human nor the agent being able to solve the\ngame on their own. We use deep reinforcement learning for the control of the\nrobotic agent, and achieve results within 30 minutes of real-world play,\nwithout any type of pre-training. We then use this setup to perform systematic\nexperiments on human/agent behaviour and adaptation when co-learning a policy\nfor the collaborative game. We present results on how co-policy learning occurs\nover time between the human and the robotic agent resulting in each\nparticipant's agent serving as a representation of how they would play the\ngame. This allows us to relate a person's success when playing with different\nagents than their own, by comparing the policy of the agent with that of their\nown agent."", 'A common vision from science fiction is that robots will one day inhabit our\nphysical spaces, sense the world as we do, assist our physical labours, and\ncommunicate with us through natural language. Here we study how to design\nartificial agents that can interact naturally with humans using the\nsimplification of a virtual environment. This setting nevertheless integrates a\nnumber of the central challenges of artificial intelligence (AI) research:\ncomplex visual perception and goal-directed physical control, grounded language\ncomprehension and production, and multi-agent social interaction. To build\nagents that can robustly interact with humans, we would ideally train them\nwhile they interact with humans. However, this is presently impractical.\nTherefore, we approximate the role of the human with another learned agent, and\nuse ideas from inverse reinforcement learning to reduce the disparities between\nhuman-human and agent-agent interactive behaviour. Rigorously evaluating our\nagents poses a great challenge, so we develop a variety of behavioural tests,\nincluding evaluation by humans who watch videos of agents or interact directly\nwith them. These evaluations convincingly demonstrate that interactive training\nand auxiliary losses improve agent behaviour beyond what is achieved by\nsupervised learning of actions alone. Further, we demonstrate that agent\ncapabilities generalise beyond literal experiences in the dataset. Finally, we\ntrain evaluation models whose ratings of agents agree well with human\njudgement, thus permitting the evaluation of new agent models without\nadditional effort. Taken together, our results in this virtual environment\nprovide evidence that large-scale human behavioural imitation is a promising\ntool to create intelligent, interactive agents, and the challenge of reliably\nevaluating such agents is possible to surmount.', 'This paper develops a model that addresses sentence embedding, a hot topic in\ncurrent natural language processing research, using recurrent neural networks\nwith Long Short-Term Memory (LSTM) cells. Due to its ability to capture long\nterm memory, the LSTM-RNN accumulates increasingly richer information as it\ngoes through the sentence, and when it reaches the last word, the hidden layer\nof the network provides a semantic representation of the whole sentence. In\nthis paper, the LSTM-RNN is trained in a weakly supervised manner on user\nclick-through data logged by a commercial web search engine. Visualization and\nanalysis are performed to understand how the embedding process works. The model\nis found to automatically attenuate the unimportant words and detects the\nsalient keywords in the sentence. Furthermore, these detected keywords are\nfound to automatically activate different cells of the LSTM-RNN, where words\nbelonging to a similar topic activate the same cell. As a semantic\nrepresentation of the sentence, the embedding vector can be used in many\ndifferent applications. These automatic keyword detection and topic allocation\nabilities enabled by the LSTM-RNN allow the network to perform document\nretrieval, a difficult language processing task, where the similarity between\nthe query and documents can be measured by the distance between their\ncorresponding sentence embedding vectors computed by the LSTM-RNN. On a web\nsearch task, the LSTM-RNN embedding is shown to significantly outperform\nseveral existing state of the art methods. We emphasize that the proposed model\ngenerates sentence embedding vectors that are specially useful for web document\nretrieval tasks. A comparison with a well known general sentence embedding\nmethod, the Paragraph Vector, is performed. The results show that the proposed\nmethod in this paper significantly outperforms it for web document retrieval\ntask.', ""Writing is a complex non-linear process that begins with a mental model of\nintent, and progresses through an outline of ideas, to words on paper (and\ntheir subsequent refinement). Despite past research in understanding writing,\nWeb-scale consumer and enterprise collaborative digital writing environments\nare yet to greatly benefit from intelligent systems that understand the stages\nof document evolution, providing opportune assistance based on authors'\nsituated actions and context. In this paper, we present three studies that\nexplore temporal stages of document authoring. We first survey information\nworkers at a large technology company about their writing habits and\npreferences, concluding that writers do in fact conceptually progress through\nseveral distinct phases while authoring documents. We also explore,\nqualitatively, how writing stages are linked to document lifespan. We\nsupplement these qualitative findings with an analysis of the longitudinal user\ninteraction logs of a popular digital writing platform over several million\ndocuments. Finally, as a first step towards facilitating an intelligent digital\nwriting assistant, we conduct a preliminary investigation into the utility of\nuser interaction log data for predicting the temporal stage of a document. Our\nresults support the benefit of tools tailored to writing stages, identify\nprimary tasks associated with these stages, and show that it is possible to\npredict stages from anonymous interaction logs. Together, these results argue\nfor the benefit and feasibility of more tailored digital writing assistance."", 'While deep reinforcement learning techniques have led to agents that are\nsuccessfully able to learn to perform a number of tasks that had been\npreviously unlearnable, these techniques are still susceptible to the\nlongstanding problem of reward sparsity. This is especially true for tasks such\nas training an agent to play StarCraft II, a real-time strategy game where\nreward is only given at the end of a game which is usually very long. While\nthis problem can be addressed through reward shaping, such approaches typically\nrequire a human expert with specialized knowledge. Inspired by the vision of\nenabling reward shaping through the more-accessible paradigm of\nnatural-language narration, we develop a technique that can provide the\nbenefits of reward shaping using natural language commands. Our\nnarration-guided RL agent projects sequences of natural-language commands into\nthe same high-dimensional representation space as corresponding goal states. We\nshow that we can get improved performance with our method compared to\ntraditional reward-shaping approaches. Additionally, we demonstrate the ability\nof our method to generalize to unseen natural-language commands.', ""We seek to align agent behavior with a user's objectives in a reinforcement\nlearning setting with unknown dynamics, an unknown reward function, and unknown\nunsafe states. The user knows the rewards and unsafe states, but querying the\nuser is expensive. To address this challenge, we propose an algorithm that\nsafely and interactively learns a model of the user's reward function. We start\nwith a generative model of initial states and a forward dynamics model trained\non off-policy data. Our method uses these models to synthesize hypothetical\nbehaviors, asks the user to label the behaviors with rewards, and trains a\nneural network to predict the rewards. The key idea is to actively synthesize\nthe hypothetical behaviors from scratch by maximizing tractable proxies for the\nvalue of information, without interacting with the environment. We call this\nmethod reward query synthesis via trajectory optimization (ReQueST). We\nevaluate ReQueST with simulated users on a state-based 2D navigation task and\nthe image-based Car Racing video game. The results show that ReQueST\nsignificantly outperforms prior methods in learning reward models that transfer\nto new environments with different initial state distributions. Moreover,\nReQueST safely trains the reward model to detect unsafe states, and corrects\nreward hacking before deploying the agent."", 'We present an efficient document representation learning framework, Document\nVector through Corruption (Doc2VecC). Doc2VecC represents each document as a\nsimple average of word embeddings. It ensures a representation generated as\nsuch captures the semantic meanings of the document during learning. A\ncorruption model is included, which introduces a data-dependent regularization\nthat favors informative or rare words while forcing the embeddings of common\nand non-discriminative ones to be close to zero. Doc2VecC produces\nsignificantly better word embeddings than Word2Vec. We compare Doc2VecC with\nseveral state-of-the-art document representation learning algorithms. The\nsimple model architecture introduced by Doc2VecC matches or out-performs the\nstate-of-the-art in generating high-quality document representations for\nsentiment analysis, document classification as well as semantic relatedness\ntasks. The simplicity of the model enables training on billions of words per\nhour on a single machine. At the same time, the model is very efficient in\ngenerating representations of unseen documents at test time.', ""In this study, we investigate how a robot can generate novel and creative\nactions from its own experience of learning basic actions. Inspired by a\nmachine learning approach to computational creativity, we propose a dynamic\nneural network model that can learn and generate robot's actions. We conducted\na set of simulation experiments with a humanoid robot. The results showed that\nthe proposed model was able to learn the basic actions and also to generate\nnovel actions by modulating and combining those learned actions. The analysis\non the neural activities illustrated that the ability to generate creative\nactions emerged from the model's nonlinear memory structure self-organized\nduring training. The results also showed that the different way of learning the\nbasic actions induced the self-organization of the memory structure with the\ndifferent characteristics, resulting in the generation of different levels of\ncreative actions. Our approach can be utilized in human-robot interaction in\nwhich a user can interactively explore the robot's memory to control its\nbehavior and also discover other novel actions."", 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'This paper develops a stochastic programming framework for multi-agent\nsystems where task decomposition, assignment, and scheduling problems are\nsimultaneously optimized. The framework can be applied to heterogeneous mobile\nrobot teams with distributed sub-tasks. Examples include pandemic robotic\nservice coordination, explore and rescue, and delivery systems with\nheterogeneous vehicles. Due to their inherent flexibility and robustness,\nmulti-agent systems are applied in a growing range of real-world problems that\ninvolve heterogeneous tasks and uncertain information. Most previous works\nassume one fixed way to decompose a task into roles that can later be assigned\nto the agents. This assumption is not valid for a complex task where the roles\ncan vary and multiple decomposition structures exist. Meanwhile, it is unclear\nhow uncertainties in task requirements and agent capabilities can be\nsystematically quantified and optimized under a multi-agent system setting. A\nrepresentation for complex tasks is proposed: agent capabilities are\nrepresented as a vector of random distributions, and task requirements are\nverified by a generalizable binary function. The conditional value at risk\n(CVaR) is chosen as a metric in the objective function to generate robust\nplans. An efficient algorithm is described to solve the model, and the whole\nframework is evaluated in two different practical test cases: capture-the-flag\nand robotic service coordination during a pandemic (e.g., COVID-19). Results\ndemonstrate that the framework is generalizable, scalable up to 140 agents and\n40 tasks for the example test cases, and provides low-cost plans that ensure a\nhigh probability of success.', 'Reinforcement learning agents learn by encouraging behaviours which maximize\ntheir total reward, usually provided by the environment. In many environments,\nhowever, the reward is provided after a series of actions rather than each\nsingle action, leading the agent to experience ambiguity in terms of whether\nthose actions are effective, an issue known as the credit assignment problem.\nIn this paper, we propose two strategies inspired by behavioural psychology to\nenable the agent to intrinsically estimate more informative reward values for\nactions with no reward. The first strategy, called self-punishment (SP),\ndiscourages the agent from making mistakes that lead to undesirable terminal\nstates. The second strategy, called the rewards backfill (RB), backpropagates\nthe rewards between two rewarded actions. We prove that, under certain\nassumptions and regardless of the reinforcement learning algorithm used, these\ntwo strategies maintain the order of policies in the space of all possible\npolicies in terms of their total reward, and, by extension, maintain the\noptimal policy. Hence, our proposed strategies integrate with any reinforcement\nlearning algorithm that learns a value or action-value function through\nexperience. We incorporated these two strategies into three popular deep\nreinforcement learning approaches and evaluated the results on thirty Atari\ngames. After parameter tuning, our results indicate that the proposed\nstrategies improve the tested methods in over 65 percent of tested games by up\nto over 25 times performance improvement.', ""Intelligent assistants that follow commands or answer simple questions, such\nas Siri and Google search, are among the most economically important\napplications of AI. Future conversational AI assistants promise even greater\ncapabilities and a better user experience through a deeper understanding of the\ndomain, the user, or the user's purposes. But what domain and what methods are\nbest suited to researching and realizing this promise? In this article we argue\nfor the domain of voice document editing and for the methods of model-based\nreinforcement learning. The primary advantages of voice document editing are\nthat the domain is tightly scoped and that it provides something for the\nconversation to be about (the document) that is delimited and fully accessible\nto the intelligent assistant. The advantages of reinforcement learning in\ngeneral are that its methods are designed to learn from interaction without\nexplicit instruction and that it formalizes the purposes of the assistant.\nModel-based reinforcement learning is needed in order to genuinely understand\nthe domain of discourse and thereby work efficiently with the user to achieve\ntheir goals. Together, voice document editing and model-based reinforcement\nlearning comprise a promising research direction for achieving conversational\nAI."", 'Conversational search is an approach to information retrieval (IR), where\nusers engage in a dialogue with an agent in order to satisfy their information\nneeds. Previous conceptual work described properties and actions a good agent\nshould exhibit. Unlike them, we present a novel conceptual model defined in\nterms of conversational goals, which enables us to reason about current\nresearch practices in conversational search. Based on the literature, we elicit\nhow existing tasks and test collections from the fields of IR, natural language\nprocessing (NLP) and dialogue systems (DS) fit into this model. We describe a\nset of characteristics that an ideal conversational search dataset should have.\nLastly, we introduce MANtIS (the code and dataset are available at\nhttps://guzpenha.github.io/MANtIS/), a large-scale dataset containing\nmulti-domain and grounded information seeking dialogues that fulfill all of our\ndataset desiderata. We provide baseline results for the conversation response\nranking and user intent prediction tasks.', 'Reinforcement learning is a powerful technique to train an agent to perform a\ntask. However, an agent that is trained using reinforcement learning is only\ncapable of achieving the single task that is specified via its reward function.\nSuch an approach does not scale well to settings in which an agent needs to\nperform a diverse set of tasks, such as navigating to varying positions in a\nroom or moving objects to varying locations. Instead, we propose a method that\nallows an agent to automatically discover the range of tasks that it is capable\nof performing. We use a generator network to propose tasks for the agent to try\nto achieve, specified as goal states. The generator network is optimized using\nadversarial training to produce tasks that are always at the appropriate level\nof difficulty for the agent. Our method thus automatically produces a\ncurriculum of tasks for the agent to learn. We show that, by using this\nframework, an agent can efficiently and automatically learn to perform a wide\nset of tasks without requiring any prior knowledge of its environment. Our\nmethod can also learn to achieve tasks with sparse rewards, which traditionally\npose significant challenges.', ""In the Vision-and-Language Navigation (VLN) task an embodied agent navigates\na 3D environment, following natural language instructions. A challenge in this\ntask is how to handle 'off the path' scenarios where an agent veers from a\nreference path. Prior work supervises the agent with actions based on the\nshortest path from the agent's location to the goal, but such goal-oriented\nsupervision is often not in alignment with the instruction. Furthermore, the\nevaluation metrics employed by prior work do not measure how much of a language\ninstruction the agent is able to follow. In this work, we propose a simple and\neffective language-aligned supervision scheme, and a new metric that measures\nthe number of sub-instructions the agent has completed during navigation."", ""The role of conversational assistants has become more prevalent in helping\npeople increase their productivity. Document-centered assistance, for example\nto help an individual quickly review a document, has seen less significant\nprogress, even though it has the potential to tremendously increase a user's\nproductivity. This type of document-centered assistance is the focus of this\npaper. Our contributions are three-fold: (1) We first present a survey to\nunderstand the space of document-centered assistance and the capabilities\npeople expect in this scenario. (2) We investigate the types of queries that\nusers will pose while seeking assistance with documents, and show that\ndocument-centered questions form the majority of these queries. (3) We present\na set of initial machine learned models that show that (a) we can accurately\ndetect document-centered questions, and (b) we can build reasonably accurate\nmodels for answering such questions. These positive results are encouraging,\nand suggest that even greater results may be attained with continued study of\nthis interesting and novel problem space. Our findings have implications for\nthe design of intelligent systems to support task completion via natural\ninteractions with documents."", 'With the increase of information, document classification as one of the\nmethods of text mining, plays vital role in many management and organizing\ninformation. Document classification is the process of assigning a document to\none or more predefined category labels. Document classification includes\ndifferent parts such as text processing, term selection, term weighting and\nfinal classification. The accuracy of document classification is very\nimportant. Thus improvement in each part of classification should lead to\nbetter results and higher precision. Term weighting has a great impact on the\naccuracy of the classification. Most of the existing weighting methods exploit\nthe statistical information of terms in documents and do not consider semantic\nrelations between words. In this paper, an automated document classification\nsystem is presented that uses a novel term weighting method based on semantic\nrelations between terms. To evaluate the proposed method, three standard\nPersian corpuses are used. Experiment results show 2 to 4 percent improvement\nin classification accuracy compared with the best previous designed system for\nPersian documents.', 'How to leverage cross-document interactions to improve ranking performance is\nan important topic in information retrieval (IR) research. However, this topic\nhas not been well-studied in the learning-to-rank setting and most of the\nexisting work still treats each document independently while scoring. The\nrecent development of deep learning shows strength in modeling complex\nrelationships across sequences and sets. It thus motivates us to study how to\nleverage cross-document interactions for learning-to-rank in the deep learning\nframework. In this paper, we formally define the permutation-equivariance\nrequirement for a scoring function that captures cross-document interactions.\nWe then propose a self-attention based document interaction network and show\nthat it satisfies the permutation-equivariant requirement, and can generate\nscores for document sets of varying sizes. Our proposed methods can\nautomatically learn to capture document interactions without any auxiliary\ninformation, and can scale across large document sets. We conduct experiments\non three ranking datasets: the benchmark Web30k, a Gmail search, and a Google\nDrive Quick Access dataset. Experimental results show that our proposed methods\nare both more effective and efficient than baselines.', 'Document network embedding aims at learning representations for a structured\ntext corpus i.e. when documents are linked to each other. Recent algorithms\nextend network embedding approaches by incorporating the text content\nassociated with the nodes in their formulations. In most cases, it is hard to\ninterpret the learned representations. Moreover, little importance is given to\nthe generalization to new documents that are not observed within the network.\nIn this paper, we propose an interpretable and inductive document network\nembedding method. We introduce a novel mechanism, the Topic-Word Attention\n(TWA), that generates document representations based on the interplay between\nword and topic representations. We train these word and topic vectors through\nour general model, Inductive Document Network Embedding (IDNE), by leveraging\nthe connections in the document network. Quantitative evaluations show that our\napproach achieves state-of-the-art performance on various networks and we\nqualitatively show that our model produces meaningful and interpretable\nrepresentations of the words, topics and documents.', ""An accurate abstractive summary of a document should contain all its salient\ninformation and should be logically entailed by the input document. We improve\nthese important aspects of abstractive summarization via multi-task learning\nwith the auxiliary tasks of question generation and entailment generation,\nwhere the former teaches the summarization model how to look for salient\nquestioning-worthy details, and the latter teaches the model how to rewrite a\nsummary which is a directed-logical subset of the input document. We also\npropose novel multi-task architectures with high-level (semantic)\nlayer-specific sharing across multiple encoder and decoder layers of the three\ntasks, as well as soft-sharing mechanisms (and show performance ablations and\nanalysis examples of each contribution). Overall, we achieve statistically\nsignificant improvements over the state-of-the-art on both the CNN/DailyMail\nand Gigaword datasets, as well as on the DUC-2002 transfer setup. We also\npresent several quantitative and qualitative analysis studies of our model's\nlearned saliency and entailment skills."", 'We propose a Label Propagation based algorithm for weakly supervised text\nclassification. We construct a graph where each document is represented by a\nnode and edge weights represent similarities among the documents. Additionally,\nwe discover underlying topics using Latent Dirichlet Allocation (LDA) and\nenrich the document graph by including the topics in the form of additional\nnodes. The edge weights between a topic and a text document represent level of\n""affinity"" between them. Our approach does not require document level\nlabelling, instead it expects manual labels only for topic nodes. This\nsignificantly minimizes the level of supervision needed as only a few topics\nare observed to be enough for achieving sufficiently high accuracy. The Label\nPropagation Algorithm is employed on this enriched graph to propagate labels\namong the nodes. Our approach combines the advantages of Label Propagation\n(through document-document similarities) and Topic Modelling (for minimal but\nsmart supervision). We demonstrate the effectiveness of our approach on various\ndatasets and compare with state-of-the-art weakly supervised text\nclassification approaches.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'This paper develops an inverse reinforcement learning algorithm aimed at\nrecovering a reward function from the observed actions of an agent. We\nintroduce a strategy to flexibly handle different types of actions with two\napproximations of the Bellman Optimality Equation, and a Bellman Gradient\nIteration method to compute the gradient of the Q-value with respect to the\nreward function. These methods allow us to build a differentiable relation\nbetween the Q-value and the reward function and learn an approximately optimal\nreward function with gradient methods. We test the proposed method in two\nsimulated environments by evaluating the accuracy of different approximations\nand comparing the proposed method with existing solutions. The results show\nthat even with a linear reward function, the proposed method has a comparable\naccuracy with the state-of-the-art method adopting a non-linear reward\nfunction, and the proposed method is more flexible because it is defined on\nobserved actions instead of trajectories.', ""In many robotic applications, an autonomous agent must act within and explore\na partially observed environment that is unobserved by its human teammate. We\nconsider such a setting in which the agent can, while acting, transmit\ndeclarative information to the human that helps them understand aspects of this\nunseen environment. In this work, we address the algorithmic question of how\nthe agent should plan out what actions to take and what information to\ntransmit. Naturally, one would expect the human to have preferences, which we\nmodel information-theoretically by scoring transmitted information based on the\nchange it induces in weighted entropy of the human's belief state. We formulate\nthis setting as a belief MDP and give a tractable algorithm for solving it\napproximately. Then, we give an algorithm that allows the agent to learn the\nhuman's preferences online, through exploration. We validate our approach\nexperimentally in simulated discrete and continuous partially observed\nsearch-and-recover domains. Visit http://tinyurl.com/chitnis-corl-18 for a\nsupplementary video."", 'Feature based local attribution methods are amongst the most prevalent in\nexplainable artificial intelligence (XAI) literature. Going beyond standard\ncorrelation, recently, methods have been proposed that highlight what should be\nminimally sufficient to justify the classification of an input (viz. pertinent\npositives). While minimal sufficiency is an attractive property, the resulting\nexplanations are often too sparse for a human to understand and evaluate the\nlocal behavior of the model, thus making it difficult to judge its overall\nquality. To overcome these limitations, we propose a novel method called\nPath-Sufficient Explanations Method (PSEM) that outputs a sequence of\nsufficient explanations for a given input of strictly decreasing size (or\nvalue) -- from original input to a minimally sufficient explanation -- which\ncan be thought to trace the local boundary of the model in a smooth manner,\nthus providing better intuition about the local model behavior for the specific\ninput. We validate these claims, both qualitatively and quantitatively, with\nexperiments that show the benefit of PSEM across all three modalities (image,\ntabular and text). A user study depicts the strength of the method in\ncommunicating the local behavior, where (many) users are able to correctly\ndetermine the prediction made by a model.', 'When estimating the relevancy between a query and a document, ranking models\nlargely neglect the mutual information among documents. A common wisdom is that\nif two documents are similar in terms of the same query, they are more likely\nto have similar relevance score. To mitigate this problem, in this paper, we\npropose a multi-agent reinforced ranking model, named MarlRank. In particular,\nby considering each document as an agent, we formulate the ranking process as a\nmulti-agent Markov Decision Process (MDP), where the mutual interactions among\ndocuments are incorporated in the ranking process. To compute the ranking list,\neach document predicts its relevance to a query considering not only its own\nquery-document features but also its similar documents features and actions. By\ndefining reward as a function of NDCG, we can optimize our model directly on\nthe ranking performance measure. Our experimental results on two LETOR\nbenchmark datasets show that our model has significant performance gains over\nthe state-of-art baselines. We also find that the NDCG shows an overall\nincreasing trend along with the step of interactions, which demonstrates that\nthe mutual information among documents helps improve the ranking performance.', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', 'A major challenge of multiagent reinforcement learning (MARL) is the curse of\nmultiagents, where the size of the joint action space scales exponentially with\nthe number of agents. This remains to be a bottleneck for designing efficient\nMARL algorithms even in a basic scenario with finitely many states and actions.\nThis paper resolves this challenge for the model of episodic Markov games. We\ndesign a new class of fully decentralized algorithms -- V-learning, which\nprovably learns Nash equilibria (in the two-player zero-sum setting),\ncorrelated equilibria and coarse correlated equilibria (in the multiplayer\ngeneral-sum setting) in a number of samples that only scales with\n$\\max_{i\\in[m]} A_i$, where $A_i$ is the number of actions for the $i^{\\rm th}$\nplayer. This is in sharp contrast to the size of the joint action space which\nis $\\prod_{i=1}^m A_i$. V-learning (in its basic form) is a new class of\nsingle-agent RL algorithms that convert any adversarial bandit algorithm with\nsuitable regret guarantees into a RL algorithm. Similar to the classical\nQ-learning algorithm, it performs incremental updates to the value functions.\nDifferent from Q-learning, it only maintains the estimates of V-values instead\nof Q-values. This key difference allows V-learning to achieve the claimed\nguarantees in the MARL setting by simply letting all agents run V-learning\nindependently.', 'A recent trend in IR has been the usage of neural networks to learn retrieval\nmodels for text based adhoc search. While various approaches and architectures\nhave yielded significantly better performance than traditional retrieval models\nsuch as BM25, it is still difficult to understand exactly why a document is\nrelevant to a query. In the ML community several approaches for explaining\ndecisions made by deep neural networks have been proposed -- including DeepSHAP\nwhich modifies the DeepLift algorithm to estimate the relative importance\n(shapley values) of input features for a given decision by comparing the\nactivations in the network for a given image against the activations caused by\na reference input. In image classification, the reference input tends to be a\nplain black image. While DeepSHAP has been well studied for image\nclassification tasks, it remains to be seen how we can adapt it to explain the\noutput of Neural Retrieval Models (NRMs). In particular, what is a good ""black""\nimage in the context of IR? In this paper we explored various reference input\ndocument construction techniques. Additionally, we compared the explanations\ngenerated by DeepSHAP to LIME (a model agnostic approach) and found that the\nexplanations differ considerably. Our study raises concerns regarding the\nrobustness and accuracy of explanations produced for NRMs. With this paper we\naim to shed light on interesting problems surrounding interpretability in NRMs\nand highlight areas of future work.', ""BERT based ranking models have achieved superior performance on various\ninformation retrieval tasks. However, the large number of parameters and\ncomplex self-attention operation come at a significant latency overhead. To\nremedy this, recent works propose late-interaction architectures, which allow\npre-computation of intermediate document representations, thus reducing the\nruntime latency. Nonetheless, having solved the immediate latency issue, these\nmethods now introduce storage costs and network fetching latency, which limits\ntheir adoption in real-life production systems.\n  In this work, we propose the Succinct Document Representation (SDR) scheme\nthat computes highly compressed intermediate document representations,\nmitigating the storage/network issue. Our approach first reduces the dimension\nof token representations by encoding them using a novel autoencoder\narchitecture that uses the document's textual content in both the encoding and\ndecoding phases. After this token encoding step, we further reduce the size of\nentire document representations using a modern quantization technique.\n  Extensive evaluations on passage re-reranking on the MSMARCO dataset show\nthat compared to existing approaches using compressed document representations,\nour method is highly efficient, achieving 4x-11.6x better compression rates for\nthe same ranking quality."", ""Shared control in teleoperation leverages both human and robot's strengths\nand has demonstrated great advantages of reducing the difficulties in\nteleoperating a robot and increasing the task performance. One fundamental\nquestion in shared control is how to effectively allocate the control power to\nthe human and robot. Researchers have been subjectively defining the arbitrate\npolicies following conflicting principles, which resulted in great\ninconsistency in the policies. We attribute this inconsistency to the\ninconsiderateness of the multi-resource uncertainty in the human-robot system.\nTo fill the gap, we developed a multi-source uncertainty model that was\napplicable to various types of uncertainty in real world, and then a general\narbitration model was developed to comprehensively fuse the uncertainty and\nregulate the arbitration weight assigned to the robotic agent. Beside\ntraditional macro performance metrics, we introduced objective and quantitative\nmetrics of robotic helpfulness and friendliness that evaluated the assistive\nrobot's cooperation at micro and macro levels. Results from simulations and\nexperiments showed the new arbitration model was more effective and friendly\nover the existing policies and was robust to coping with multi-source\nuncertainty. With this new arbitration model, we expect the increased adoption\nof human-robot shared control in practical and complex teleoperation tasks."", 'Human interaction involves very sophisticated non-verbal communication skills\nlike understanding the goals and actions of others and coordinating our own\nactions accordingly. Neuroscience refers to this mechanism as motor resonance,\nin the sense that the perception of another person\'s actions and sensory\nexperiences activates the observer\'s brain as if (s)he would be performing the\nsame actions and having the same experiences. We analyze and model non-verbal\ncues (arm movements) exchanged between two humans that interact and execute\nhandover actions. The contributions of this paper are the following: (i)\ncomputational models, using recorded motion data, describing the motor\nbehaviour of each actor in action-in-interaction situations, (ii) a\ncomputational model that captures the behaviour if the ""giver"" and ""receiver""\nduring an object handover action, by coupling the arm motion of both actors,\nand (iii) embedded these models in the iCub robot for both action execution and\nrecognition. Our results show that: (i) the robot can interpret the human arm\nmotion and recognize handover actions; and (ii) behave in a ""human-like"" manner\nto receive the object of the recognized handover action.', 'Human recognition of the actions of other humans is very efficient and is\nbased on patterns of movements. Our theoretical starting point is that the\ndynamics of the joint movements is important to action categorization. On the\nbasis of this theory, we present a novel action recognition system that employs\na hierarchy of Self-Organizing Maps together with a custom supervised neural\nnetwork that learns to categorize actions. The system preprocesses the input\nfrom a Kinect like 3D camera to exploit the information not only about joint\npositions, but also their first and second order dynamics. We evaluate our\nsystem in two experiments with publicly available data sets, and compare its\nperformance to the performance with less sophisticated preprocessing of the\ninput. The results show that including the dynamics of the actions improves the\nperformance. We also apply an attention mechanism that focuses on the parts of\nthe body that are the most involved in performing the actions.', ""A fundamental question in neuroscience is how the brain creates an internal\nmodel of the world to guide actions using sequences of ambiguous sensory\ninformation. This is naturally formulated as a reinforcement learning problem\nunder partial observations, where an agent must estimate relevant latent\nvariables in the world from its evidence, anticipate possible future states,\nand choose actions that optimize total expected reward. This problem can be\nsolved by control theory, which allows us to find the optimal actions for a\ngiven system dynamics and objective function. However, animals often appear to\nbehave suboptimally. Why? We hypothesize that animals have their own flawed\ninternal model of the world, and choose actions with the highest expected\nsubjective reward according to that flawed model. We describe this behavior as\nrational but not optimal. The problem of Inverse Rational Control (IRC) aims to\nidentify which internal model would best explain an agent's actions. Our\ncontribution here generalizes past work on Inverse Rational Control which\nsolved this problem for discrete control in partially observable Markov\ndecision processes. Here we accommodate continuous nonlinear dynamics and\ncontinuous actions, and impute sensory observations corrupted by unknown noise\nthat is private to the animal. We first build an optimal Bayesian agent that\nlearns an optimal policy generalized over the entire model space of dynamics\nand subjective rewards using deep reinforcement learning. Crucially, this\nallows us to compute a likelihood over models for experimentally observable\naction trajectories acquired from a suboptimal agent. We then find the model\nparameters that maximize the likelihood using gradient ascent."", 'Humans have the fascinating capacity of processing non-verbal visual cues to\nunderstand and anticipate the actions of other humans. This ""intention reading""\nability is underpinned by shared motor-repertoires and action-models, which we\nuse to interpret the intentions of others as if they were our own.\n  We investigate how the different cues contribute to the legibility of human\nactions during interpersonal interactions. Our first contribution is a publicly\navailable dataset with recordings of human body-motion and eye-gaze, acquired\nin an experimental scenario with an actor interacting with three subjects. From\nthese data, we conducted a human study to analyse the importance of the\ndifferent non-verbal cues for action perception. As our second contribution, we\nused the motion/gaze recordings to build a computational model describing the\ninteraction between two persons. As a third contribution, we embedded this\nmodel in the controller of an iCub humanoid robot and conducted a second human\nstudy, in the same scenario with the robot as an actor, to validate the model\'s\n""intention reading"" capability.\n  Our results show that it is possible to model (non-verbal) signals exchanged\nby humans during interaction, and how to incorporate such a mechanism in\nrobotic systems with the twin goal of : (i) being able to ""read"" human action\nintentions, and (ii) acting in a way that is legible by humans.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'We propose a novel approach to learn goal-conditioned policies for locomotion\nin a batch RL setting. The batch data is collected by a policy that is not\ngoal-conditioned. For the locomotion task, this translates to data collection\nusing a policy learnt by the agent for walking straight in one direction, and\nusing that data to learn a goal-conditioned policy that enables the agent to\nwalk in any direction. The data collection policy used should be invariant to\nthe direction the agent is facing i.e. regardless of its initial orientation,\nthe agent should take the same actions to walk forward. We exploit this\nproperty to learn a goal-conditioned policy using two key ideas: (1) augmenting\ndata by generating trajectories with the same actions in different directions,\nand (2) learning an encoder that enforces invariance between these rotated\ntrajectories with a Siamese framework. We show that our approach outperforms\nexisting RL algorithms on 3-D locomotion agents like Ant, Humanoid and\nMinitaur.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'User preference learning is generally a hard problem. Individual preferences\nare typically unknown even to users themselves, while the space of choices is\ninfinite. Here we study user preference learning from information-theoretic\nperspective. We model preference learning as a system with two interacting\nsub-systems, one representing a user with his/her preferences and another one\nrepresenting an agent that has to learn these preferences. The user with\nhis/her behaviour is modeled by a parametric preference function. To\nefficiently learn the preferences and reduce search space quickly, we propose\nthe agent that interacts with the user to collect the most informative data for\nlearning. The agent presents two proposals to the user for evaluation, and the\nuser rates them based on his/her preference function. We show that the optimum\nagent strategy for data collection and preference learning is a result of\nmaximin optimization of the normalized weighted Kullback-Leibler (KL)\ndivergence between true and agent-assigned predictive user response\ndistributions. The resulting value of KL-divergence, which we also call\nremaining system uncertainty (RSU), provides an efficient performance metric in\nthe absence of the ground truth. This metric characterises how well the agent\ncan predict user and, thus, the quality of the underlying learned user\n(preference) model. Our proposed agent comprises sequential mechanisms for user\nmodel inference and proposal generation. To infer the user model (preference\nfunction), Bayesian approximate inference is used in the agent. The data\ncollection strategy is to generate proposals, responses to which help resolving\nuncertainty associated with prediction of the user responses the most. The\nefficiency of our approach is validated by numerical simulations.', 'Recent years have witnessed an increasing number of interpretation methods\nbeing developed for improving transparency of NLP models. Meanwhile,\nresearchers also try to answer the question that whether the obtained\ninterpretation is faithful in explaining mechanisms behind model prediction?\nSpecifically, (Jain and Wallace, 2019) proposes that ""attention is not\nexplanation"" by comparing attention interpretation with gradient alternatives.\nHowever, it raises a new question that can we safely pick one interpretation\nmethod as the ground-truth? If not, on what basis can we compare different\ninterpretation methods? In this work, we propose that it is crucial to have a\nconcrete definition of interpretation before we could evaluate faithfulness of\nan interpretation. The definition will affect both the algorithm to obtain\ninterpretation and, more importantly, the metric used in evaluation. Through\nboth theoretical and experimental analysis, we find that although\ninterpretation methods perform differently under a certain evaluation metric,\nsuch a difference may not result from interpretation quality or faithfulness,\nbut rather the inherent bias of the evaluation metric.', 'In this paper, we propose an AI-FML robotic agent for student learning\nbehavior ontology construction which can be applied in English speaking and\nlistening domain. The AI-FML robotic agent with the ontology contains the\nperception intelligence, computational intelligence, and cognition intelligence\nfor analyzing student learning behavior. In addition, there are three\nintelligent agents, including a perception agent, a computational agent, and a\ncognition agent in the AI-FML robotic agent. We deploy the perception agent and\nthe cognition agent on the robot Kebbi Air. Moreover, the computational agent\nwith the Deep Neural Network (DNN) model is performed in the cloud and can\ncommunicate with the perception agent and cognition agent via the Internet. The\nproposed AI-FML robotic agent is applied in Taiwan and tested in Japan. The\nexperimental results show that the agents can be utilized in the human and\nmachine co-learning model for the future education.', 'In this paper, we present a multi-lingual sentence encoder that can be used\nin search engines as a query and document encoder. This embedding enables a\nsemantic similarity score between queries and documents that can be an\nimportant feature in document ranking and relevancy. To train such a customized\nsentence encoder, it is beneficial to leverage users search data in the form of\nquery-document clicked pairs however, we must avoid relying too much on search\nclick data as it is biased and does not cover many unseen cases. The search\ndata is heavily skewed towards short queries and for long queries is small and\noften noisy. The goal is to design a universal multi-lingual encoder that works\nfor all cases and covers both short and long queries. We select a number of\npublic NLI datasets in different languages and translation data and together\nwith user search data we train a language model using a multi-task approach. A\nchallenge is that these datasets are not homogeneous in terms of content, size\nand the balance ratio. While the public NLI datasets are usually two-sentence\nbased with the same portion of positive and negative pairs, the user search\ndata can contain multi-sentence documents and only positive pairs. We show how\nmulti-task training enables us to leverage all these datasets and exploit\nknowledge sharing across these tasks.', 'Anomaly detection incurs certain sampling and sensing costs and therefore it\nis of great importance to strike a balance between the detection accuracy and\nthese costs. In this work, we study anomaly detection by considering the\ndetection of threshold crossings in a stochastic time series without the\nknowledge of its statistics. To reduce the sampling cost in this detection\nprocess, we propose the use of hierarchical generative adversarial networks\n(GANs) to perform nonuniform sampling. In order to improve the detection\naccuracy and reduce the delay in detection, we introduce a buffer zone in the\noperation of the proposed GAN-based detector. In the experiments, we analyze\nthe performance of the proposed hierarchical GAN detector considering the\nmetrics of detection delay, miss rates, average cost of error, and sampling\nratio. We identify the tradeoffs in the performance as the buffer zone sizes\nand the number of GAN levels in the hierarchy vary. We also compare the\nperformance with that of a sampling policy that approximately minimizes the sum\nof average costs of sampling and error given the parameters of the stochastic\nprocess. We demonstrate that the proposed GAN-based detector can have\nsignificant performance improvements in terms of detection delay and average\ncost of error with a larger buffer zone but at the cost of increased sampling\nrates.', ""In reinforcement learning, it is common to let an agent interact for a fixed\namount of time with its environment before resetting it and repeating the\nprocess in a series of episodes. The task that the agent has to learn can\neither be to maximize its performance over (i) that fixed period, or (ii) an\nindefinite period where time limits are only used during training to diversify\nexperience. In this paper, we provide a formal account for how time limits\ncould effectively be handled in each of the two cases and explain why not doing\nso can cause state aliasing and invalidation of experience replay, leading to\nsuboptimal policies and training instability. In case (i), we argue that the\nterminations due to time limits are in fact part of the environment, and thus a\nnotion of the remaining time should be included as part of the agent's input to\navoid violation of the Markov property. In case (ii), the time limits are not\npart of the environment and are only used to facilitate learning. We argue that\nthis insight should be incorporated by bootstrapping from the value of the\nstate at the end of each partial episode. For both cases, we illustrate\nempirically the significance of our considerations in improving the performance\nand stability of existing reinforcement learning algorithms, showing\nstate-of-the-art results on several control tasks."", ""Re-finding electronic documents from a personal computer is a frequent demand\nto users. In a simple re-finding task, people can use many methods to retrieve\na document, such as navigating directly to the document's folder, searching\nwith a desktop search engine, or checking the Recent Files List. However, when\nencountering a difficult re-finding task, people usually cannot remember the\nattributes used by conventional re-finding methods, such as file path, file\nname, keywords etc., the re-finding would fail. We propose a new method to\nsupport difficult re-finding tasks. When a user is reading a document, we\ncollect all kinds of possible memory pieces of the user about the document,\nsuch as number of pages, number of images, number of math formulas, cumulative\nreading time, reading frequency, printing experiences etc. If the user wants to\nre-find a document later, we use these collected attributes to filter out the\ntarget document. To alleviate the user's cognitive burden, we use a question\nand answer wizard interface and provide recommendations to the answers for the\nuser, the recommendations are generated by analyzing the collected attributes\nof each document and the user's experiences about them."", 'We introduce a novel type of text representation that preserves the 2D layout\nof a document. This is achieved by encoding each document page as a\ntwo-dimensional grid of characters. Based on this representation, we present a\ngeneric document understanding pipeline for structured documents. This pipeline\nmakes use of a fully convolutional encoder-decoder network that predicts a\nsegmentation mask and bounding boxes. We demonstrate its capabilities on an\ninformation extraction task from invoices and show that it significantly\noutperforms approaches based on sequential text or document images.', ""Inverse reinforcement learning (IRL) is the problem of finding a reward\nfunction that generates a given optimal policy for a given Markov Decision\nProcess. This paper looks at an algorithmic-independent geometric analysis of\nthe IRL problem with finite states and actions. A L1-regularized Support Vector\nMachine formulation of the IRL problem motivated by the geometric analysis is\nthen proposed with the basic objective of the inverse reinforcement problem in\nmind: to find a reward function that generates a specified optimal policy. The\npaper further analyzes the proposed formulation of inverse reinforcement\nlearning with $n$ states and $k$ actions, and shows a sample complexity of\n$O(n^2 \\log (nk))$ for recovering a reward function that generates a policy\nthat satisfies Bellman's optimality condition with respect to the true\ntransition probabilities."", 'Traditional Relational Topic Models provide a way to discover the hidden\ntopics from a document network. Many theoretical and practical tasks, such as\ndimensional reduction, document clustering, link prediction, benefit from this\nrevealed knowledge. However, existing relational topic models are based on an\nassumption that the number of hidden topics is known in advance, and this is\nimpractical in many real-world applications. Therefore, in order to relax this\nassumption, we propose a nonparametric relational topic model in this paper.\nInstead of using fixed-dimensional probability distributions in its generative\nmodel, we use stochastic processes. Specifically, a gamma process is assigned\nto each document, which represents the topic interest of this document.\nAlthough this method provides an elegant solution, it brings additional\nchallenges when mathematically modeling the inherent network structure of\ntypical document network, i.e., two spatially closer documents tend to have\nmore similar topics. Furthermore, we require that the topics are shared by all\nthe documents. In order to resolve these challenges, we use a subsampling\nstrategy to assign each document a different gamma process from the global\ngamma process, and the subsampling probabilities of documents are assigned with\na Markov Random Field constraint that inherits the document network structure.\nThrough the designed posterior inference algorithm, we can discover the hidden\ntopics and its number simultaneously. Experimental results on both synthetic\nand real-world network datasets demonstrate the capabilities of learning the\nhidden topics and, more importantly, the number of topics.', ""Imitation learning is an effective alternative approach to learn a policy\nwhen the reward function is sparse. In this paper, we consider a challenging\nsetting where an agent and an expert use different actions from each other. We\nassume that the agent has access to a sparse reward function and state-only\nexpert observations. We propose a method which gradually balances between the\nimitation learning cost and the reinforcement learning objective. In addition,\nthis method adapts the agent's policy based on either mimicking expert behavior\nor maximizing sparse reward. We show, through navigation scenarios, that (i) an\nagent is able to efficiently leverage sparse rewards to outperform standard\nstate-only imitation learning, (ii) it can learn a policy even when its actions\nare different from the expert, and (iii) the performance of the agent is not\nbounded by that of the expert, due to the optimized usage of sparse rewards."", 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Latent topic models have been successfully applied as an unsupervised topic\ndiscovery technique in large document collections. With the proliferation of\nhypertext document collection such as the Internet, there has also been great\ninterest in extending these approaches to hypertext [6, 9]. These approaches\ntypically model links in an analogous fashion to how they model words - the\ndocument-link co-occurrence matrix is modeled in the same way that the\ndocument-word co-occurrence matrix is modeled in standard topic models. In this\npaper we present a probabilistic generative model for hypertext document\ncollections that explicitly models the generation of links. Specifically, links\nfrom a word w to a document d depend directly on how frequent the topic of w is\nin d, in addition to the in-degree of d. We show how to perform EM learning on\nthis model efficiently. By not modeling links as analogous to words, we end up\nusing far fewer free parameters and obtain better link prediction results.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.', 'We present LDAExplore, a tool to visualize topic distributions in a given\ndocument corpus that are generated using Topic Modeling methods. Latent\nDirichlet Allocation (LDA) is one of the basic methods that is predominantly\nused to generate topics. One of the problems with methods like LDA is that\nusers who apply them may not understand the topics that are generated. Also,\nusers may find it difficult to search correlated topics and correlated\ndocuments. LDAExplore, tries to alleviate these problems by visualizing topic\nand word distributions generated from the document corpus and allowing the user\nto interact with them. The system is designed for users, who have minimal\nknowledge of LDA or Topic Modelling methods. To evaluate our design, we run a\npilot study which uses the abstracts of 322 Information Visualization papers,\nwhere every abstract is considered a document. The topics generated are then\nexplored by users. The results show that users are able to find correlated\ndocuments and group them based on topics that are similar.', 'Supernumerary Robotic Limbs (SRLs) exhibit inherently compliant behavior due\nto the elasticity present at the intersection of human tissue and the robot.\nThis compliance, can prominently influence the operation of some SRLs,\ndepending on the application. In order to control the residual vibrations of\nSRLs, we have used an input-shaping method which is a computationally\ninexpensive approach. The effectiveness of this method in controlling the\nresidual vibrations of a SRL has been proven using robustness analysis. User\nstudies show that reducing the vibrations using input shaping directly\nincreases the user satisfaction and comfort by at least 9%. It is also observed\nthat 36% of the users preferred unshaped commands. We hypothesize that the\nshaped commands put a higher cognitive load on the user compared to unshaped\ncommands. This shows that when dealing with human-robot interaction, user\nsatisfaction becomes an equally important parameter as traditional performance\ncriteria and should be taken into account while evaluating the success of any\nvibration-control method.', 'Opinion summarization is the task of automatically creating summaries that\nreflect subjective information expressed in multiple documents, such as product\nreviews. While the majority of previous work has focused on the extractive\nsetting, i.e., selecting fragments from input reviews to produce a summary, we\nlet the model generate novel sentences and hence produce abstractive summaries.\nRecent progress in summarization has seen the development of supervised models\nwhich rely on large quantities of document-summary pairs. Since such training\ndata is expensive to acquire, we instead consider the unsupervised setting, in\nother words, we do not use any summaries in training. We define a generative\nmodel for a review collection which capitalizes on the intuition that when\ngenerating a new review given a set of other reviews of a product, we should be\nable to control the ""amount of novelty"" going into the new review or,\nequivalently, vary the extent to which it deviates from the input. At test\ntime, when generating summaries, we force the novelty to be minimal, and\nproduce a text reflecting consensus opinions. We capture this intuition by\ndefining a hierarchical variational autoencoder model. Both individual reviews\nand the products they correspond to are associated with stochastic latent\ncodes, and the review generator (""decoder"") has direct access to the text of\ninput reviews through the pointer-generator mechanism. Experiments on Amazon\nand Yelp datasets, show that setting at test time the review\'s latent code to\nits mean, allows the model to produce fluent and coherent summaries reflecting\ncommon opinions.', ""We consider the problem of minimizing regret in an $N$ agent heterogeneous\nstochastic linear bandits framework, where the agents (users) are similar but\nnot all identical. We model user heterogeneity using two popularly used ideas\nin practice; (i) A clustering framework where users are partitioned into groups\nwith users in the same group being identical to each other, but different\nacross groups, and (ii) a personalization framework where no two users are\nnecessarily identical, but a user's parameters are close to that of the\npopulation average. In the clustered users' setup, we propose a novel\nalgorithm, based on successive refinement of cluster identities and regret\nminimization. We show that, for any agent, the regret scales as\n$\\mathcal{O}(\\sqrt{T/N})$, if the agent is in a `well separated' cluster, or\nscales as $\\mathcal{O}(T^{\\frac{1}{2} + \\varepsilon}/(N)^{\\frac{1}{2}\n-\\varepsilon})$ if its cluster is not well separated, where $\\varepsilon$ is\npositive and arbitrarily close to $0$. Our algorithm is adaptive to the cluster\nseparation, and is parameter free -- it does not need to know the number of\nclusters, separation and cluster size, yet the regret guarantee adapts to the\ninherent complexity. In the personalization framework, we introduce a natural\nalgorithm where, the personal bandit instances are initialized with the\nestimates of the global average model. We show that, an agent $i$ whose\nparameter deviates from the population average by $\\epsilon_i$, attains a\nregret scaling of $\\widetilde{O}(\\epsilon_i\\sqrt{T})$. This demonstrates that\nif the user representations are close (small $\\epsilon_i)$, the resulting\nregret is low, and vice-versa. The results are empirically validated and we\nobserve superior performance of our adaptive algorithms over non-adaptive\nbaselines."", ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'This paper describes the short-term competition on the Components\nSegmentation Task of Document Photos that was prepared in the context of the\n16th International Conference on Document Analysis and Recognition (ICDAR\n2021). This competition aims to bring together researchers working in the field\nof identification document image processing and provides them a suitable\nbenchmark to compare their techniques on the component segmentation task of\ndocument images. Three challenge tasks were proposed entailing different\nsegmentation assignments to be performed on a provided dataset. The collected\ndata are from several types of Brazilian ID documents, whose personal\ninformation was conveniently replaced. There were 16 participants whose results\nobtained for some or all the three tasks show different rates for the adopted\nmetrics, like Dice Similarity Coefficient ranging from 0.06 to 0.99. Different\nDeep Learning models were applied by the entrants with diverse strategies to\nachieve the best results in each of the tasks. Obtained results show that the\ncurrently applied methods for solving one of the proposed tasks (document\nboundary detection) are already well established. However, for the other two\nchallenge tasks (text zone and handwritten sign detection) research and\ndevelopment of more robust approaches are still required to achieve acceptable\nresults.', 'We show that Bayes\' rule provides an effective mechanism for creating\ndocument translation models that can be learned from only parallel sentences\nand monolingual documents---a compelling benefit as parallel documents are not\nalways available. In our formulation, the posterior probability of a candidate\ntranslation is the product of the unconditional (prior) probability of the\ncandidate output document and the ""reverse translation probability"" of\ntranslating the candidate output back into the source language. Our proposed\nmodel uses a powerful autoregressive language model as the prior on target\nlanguage documents, but it assumes that each sentence is translated\nindependently from the target to the source language. Crucially, at test time,\nwhen a source document is observed, the document language model prior induces\ndependencies between the translations of the source sentences in the posterior.\nThe model\'s independence assumption not only enables efficient use of available\ndata, but it additionally admits a practical left-to-right beam-search\nalgorithm for carrying out inference. Experiments show that our model benefits\nfrom using cross-sentence context in the language model, and it outperforms\nexisting document translation approaches.', 'Document clustering is a text mining technique used to provide better\ndocument search and browsing in digital libraries or online corpora. A lot of\nresearch has been done on biomedical document clustering that is based on using\nexisting ontology. But, associations and co-occurrences of the medical concepts\nare not well represented by using ontology. In this research, a vector\nrepresentation of concepts of diseases and similarity measurement between\nconcepts are proposed. They identify the closest concepts of diseases in the\ncontext of a corpus. Each document is represented by using the vector space\nmodel. A weight scheme is proposed to consider both local content and\nassociations between concepts. A Self-Organizing Map is used as document\nclustering algorithm. The vector projection and visualization features of SOM\nenable visualization and analysis of the clusters distributions and\nrelationships on the two dimensional space. The experimental results show that\nthe proposed document clustering framework generates meaningful clusters and\nfacilitate visualization of the clusters based on the concepts of diseases.', 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', ""A Bayesian view of data interpretation suggests that a visualization user\nshould update their existing beliefs about a parameter's value in accordance\nwith the amount of information about the parameter value captured by the new\nobservations. Extending recent work applying Bayesian models to understand and\nevaluate belief updating from visualizations, we show how the predictions of\nBayesian inference can be used to guide more rational belief updating. We\ndesign a Bayesian inference-assisted uncertainty analogy that numerically\nrelates uncertainty in observed data to the user's subjective uncertainty, and\na posterior visualization that prescribes how a user should update their\nbeliefs given their prior beliefs and the observed data. In a pre-registered\nexperiment on 4,800 people, we find that when a newly observed data sample is\nrelatively small (N=158), both techniques reliably improve people's Bayesian\nupdating on average compared to the current best practice of visualizing\nuncertainty in the observed data. For large data samples (N=5208), where\npeople's updated beliefs tend to deviate more strongly from the prescriptions\nof a Bayesian model, we find evidence that the effectiveness of the two forms\nof Bayesian assistance may depend on people's proclivity toward trusting the\nsource of the data. We discuss how our results provide insight into individual\nprocesses of belief updating and subjective uncertainty, and how understanding\nthese aspects of interpretation paves the way for more sophisticated\ninteractive visualizations for analysis and communication."", 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', 'When we view visualizations, we not only have a visual representation of the\ndata, but also a verbal one. Recent work has shown that these visual\nrepresentations of data can be biased, such that the position of a line in a\nchart will be consistently underestimated. But are the verbal representations\nof position encodings also biased in the same manner, or is this a purely\nvisual bias that can be mitigated with verbal context? We explored the bias in\nposition reproductions for simple uniform lines for both visual and verbal\nrepresentations. We find that the direction of the bias changed depending on\nthe response modality, with visual reproductions showing a position\nunderestimation while verbal responses showed overestimation. This finding\nindicates that, even for simple line charts, biases are still present for both\nvisual and verbal representations, although the directionality of this bias\ndepends on the modality.', 'We address the problem of solving mixed random linear equations. We have\nunlabeled observations coming from multiple linear regressions, and each\nobservation corresponds to exactly one of the regression models. The goal is to\nlearn the linear regressors from the observations. Classically, Alternating\nMinimization (AM) (which is a variant of Expectation Maximization (EM)) is used\nto solve this problem. AM iteratively alternates between the estimation of\nlabels and solving the regression problems with the estimated labels.\nEmpirically, it is observed that, for a large variety of non-convex problems\nincluding mixed linear regression, AM converges at a much faster rate compared\nto gradient based algorithms. However, the existing theory suggests similar\nrate of convergence for AM and gradient based methods, failing to capture this\nempirical behavior. In this paper, we close this gap between theory and\npractice for the special case of a mixture of $2$ linear regressions. We show\nthat, provided initialized properly, AM enjoys a \\emph{super-linear} rate of\nconvergence in certain parameter regimes. To the best of our knowledge, this is\nthe first work that theoretically establishes such rate for AM. Hence, if we\nwant to recover the unknown regressors upto an error (in $\\ell_2$ norm) of\n$\\epsilon$, AM only takes $\\mathcal{O}(\\log \\log (1/\\epsilon))$ iterations.\nFurthermore, we compare AM with a gradient based heuristic algorithm\nempirically and show that AM dominates in iteration complexity as well as\nwall-clock time.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am seeking to develop a method that yields a range of promising output candidates rather than a single solution, a concept known as conformal prediction, in scenarios where data availability for the target task is limited. I need a model capable of automatically analyzing task features and identifying similar tasks with sufficiently available training data. This model should then discern the differences between tasks and transfer knowledge from similar tasks to new ones. It should also retain the candidate that the user ultimately selects, storing this data under the new task for future reference. I am currently undecided on whether to use a transformer or a convolutional neural network (CNN) for this model's implementation, so I would appreciate a comparison of these two structures. Additionally, I would like examples demonstrating fields where conformal prediction outperforms the conventional best-one prediction.","['Existing research on continual learning of a sequence of tasks focused on\ndealing with catastrophic forgetting, where the tasks are assumed to be\ndissimilar and have little shared knowledge. Some work has also been done to\ntransfer previously learned knowledge to the new task when the tasks are\nsimilar and have shared knowledge. To the best of our knowledge, no technique\nhas been proposed to learn a sequence of mixed similar and dissimilar tasks\nthat can deal with forgetting and also transfer knowledge forward and backward.\nThis paper proposes such a technique to learn both types of tasks in the same\nnetwork. For dissimilar tasks, the algorithm focuses on dealing with\nforgetting, and for similar tasks, the algorithm focuses on selectively\ntransferring the knowledge learned from some similar previous tasks to improve\nthe new task learning. Additionally, the algorithm automatically detects\nwhether a new task is similar to any previous tasks. Empirical evaluation using\nsequences of mixed tasks demonstrates the effectiveness of the proposed model.', 'Most recently proposed methods for Neural Program Induction work under the\nassumption of having a large set of input/output (I/O) examples for learning\nany underlying input-output mapping. This paper aims to address the problem of\ndata and computation efficiency of program induction by leveraging information\nfrom related tasks. Specifically, we propose two approaches for cross-task\nknowledge transfer to improve program induction in limited-data scenarios. In\nour first proposal, portfolio adaptation, a set of induction models is\npretrained on a set of related tasks, and the best model is adapted towards the\nnew task using transfer learning. In our second approach, meta program\ninduction, a $k$-shot learning approach is used to make a model generalize to\nnew tasks without additional training. To test the efficacy of our methods, we\nconstructed a new benchmark of programs written in the Karel programming\nlanguage. Using an extensive experimental evaluation on the Karel benchmark, we\ndemonstrate that our proposals dramatically outperform the baseline induction\nmethod that does not use knowledge transfer. We also analyze the relative\nperformance of the two approaches and study conditions in which they perform\nbest. In particular, meta induction outperforms all existing approaches under\nextreme data sparsity (when a very small number of examples are available),\ni.e., fewer than ten. As the number of available I/O examples increase (i.e. a\nthousand or more), portfolio adapted program induction becomes the best\napproach. For intermediate data sizes, we demonstrate that the combined method\nof adapted meta program induction has the strongest performance.', 'Training deep learning models with limited labelled data is an attractive\nscenario for many NLP tasks, including document classification. While with the\nrecent emergence of BERT, deep learning language models can achieve reasonably\ngood performance in document classification with few labelled instances, there\nis a lack of evidence in the utility of applying BERT-like models on long\ndocument classification. This work introduces a long-text-specific model -- the\nHierarchical BERT Model (HBM) -- that learns sentence-level features of the\ntext and works well in scenarios with limited labelled data. Various evaluation\nexperiments have demonstrated that HBM can achieve higher performance in\ndocument classification than the previous state-of-the-art methods with only 50\nto 200 labelled instances, especially when documents are long. Also, as an\nextra benefit of HBM, the salient sentences identified by learned HBM are\nuseful as explanations for labelling documents based on a user study.', 'Given new tasks with very little data$-$such as new classes in a\nclassification problem or a domain shift in the input$-$performance of modern\nvision systems degrades remarkably quickly. In this work, we illustrate how the\nneural network representations which underpin modern vision systems are subject\nto supervision collapse, whereby they lose any information that is not\nnecessary for performing the training task, including information that may be\nnecessary for transfer to new tasks or domains. We then propose two methods to\nmitigate this problem. First, we employ self-supervised learning to encourage\ngeneral-purpose features that transfer better. Second, we propose a novel\nTransformer based neural network architecture called CrossTransformers, which\ncan take a small number of labeled images and an unlabeled query, find coarse\nspatial correspondence between the query and the labeled images, and then infer\nclass membership by computing distances between spatially-corresponding\nfeatures. The result is a classifier that is more robust to task and domain\nshift, which we demonstrate via state-of-the-art performance on Meta-Dataset, a\nrecent dataset for evaluating transfer from ImageNet to many other vision\ndatasets.', 'Imitation learning has been commonly applied to solve different tasks in\nisolation. This usually requires either careful feature engineering, or a\nsignificant number of samples. This is far from what we desire: ideally, robots\nshould be able to learn from very few demonstrations of any given task, and\ninstantly generalize to new situations of the same task, without requiring\ntask-specific engineering. In this paper, we propose a meta-learning framework\nfor achieving such capability, which we call one-shot imitation learning.\n  Specifically, we consider the setting where there is a very large set of\ntasks, and each task has many instantiations. For example, a task could be to\nstack all blocks on a table into a single tower, another task could be to place\nall blocks on a table into two-block towers, etc. In each case, different\ninstances of the task would consist of different sets of blocks with different\ninitial states. At training time, our algorithm is presented with pairs of\ndemonstrations for a subset of all tasks. A neural net is trained that takes as\ninput one demonstration and the current state (which initially is the initial\nstate of the other demonstration of the pair), and outputs an action with the\ngoal that the resulting sequence of states and actions matches as closely as\npossible with the second demonstration. At test time, a demonstration of a\nsingle instance of a new task is presented, and the neural net is expected to\nperform well on new instances of this new task. The use of soft attention\nallows the model to generalize to conditions and tasks unseen in the training\ndata. We anticipate that by training this model on a much greater variety of\ntasks and settings, we will obtain a general system that can turn any\ndemonstrations into robust policies that can accomplish an overwhelming variety\nof tasks.\n  Videos available at https://bit.ly/nips2017-oneshot .', 'The performance of deep neural networks improves with more annotated data.\nThe problem is that the budget for annotation is limited. One solution to this\nis active learning, where a model asks human to annotate data that it perceived\nas uncertain. A variety of recent methods have been proposed to apply active\nlearning to deep networks but most of them are either designed specific for\ntheir target tasks or computationally inefficient for large networks. In this\npaper, we propose a novel active learning method that is simple but\ntask-agnostic, and works efficiently with the deep networks. We attach a small\nparametric module, named ""loss prediction module,"" to a target network, and\nlearn it to predict target losses of unlabeled inputs. Then, this module can\nsuggest data that the target model is likely to produce a wrong prediction.\nThis method is task-agnostic as networks are learned from a single loss\nregardless of target tasks. We rigorously validate our method through image\nclassification, object detection, and human pose estimation, with the recent\nnetwork architectures. The results demonstrate that our method consistently\noutperforms the previous methods over the tasks.', 'Text alignment finds application in tasks such as citation recommendation and\nplagiarism detection. Existing alignment methods operate at a single,\npredefined level and cannot learn to align texts at, for example, sentence and\ndocument levels. We propose a new learning approach that equips previously\nestablished hierarchical attention encoders for representing documents with a\ncross-document attention component, enabling structural comparisons across\ndifferent levels (document-to-document and sentence-to-document). Our component\nis weakly supervised from document pairs and can align at multiple levels. Our\nevaluation on predicting document-to-document relationships and\nsentence-to-document relationships on the tasks of citation recommendation and\nplagiarism detection shows that our approach outperforms previously established\nhierarchical, attention encoders based on recurrent and transformer\ncontextualization that are unaware of structural correspondence between\ndocuments.', 'Most automation in machine learning focuses on model selection and hyper\nparameter tuning, and many overlook the challenge of automatically defining\npredictive tasks. We still heavily rely on human experts to define prediction\ntasks, and generate labels by aggregating raw data. In this paper, we tackle\nthe challenge of defining useful prediction problems on event-driven\ntime-series data. We introduce MLFriend to address this challenge. MLFriend\nfirst generates all possible prediction tasks under a predefined space, then\ninteracts with a data scientist to learn the context of the data and recommend\ngood prediction tasks from all the tasks in the space. We evaluate our system\non three different datasets and generate a total of 2885 prediction tasks and\nsolve them. Out of these 722 were deemed useful by expert data scientists. We\nalso show that an automatic prediction task discovery system is able to\nidentify top 10 tasks that a user may like within a batch of 100 tasks.', 'State-of-the-art systems in deep question answering proceed as follows: (1)\nan initial document retrieval selects relevant documents, which (2) are then\nprocessed by a neural network in order to extract the final answer. Yet the\nexact interplay between both components is poorly understood, especially\nconcerning the number of candidate documents that should be retrieved. We show\nthat choosing a static number of documents -- as used in prior research --\nsuffers from a noise-information trade-off and yields suboptimal results. As a\nremedy, we propose an adaptive document retrieval model. This learns the\noptimal candidate number for document retrieval, conditional on the size of the\ncorpus and the query. We report extensive experimental results showing that our\nadaptive approach outperforms state-of-the-art methods on multiple benchmark\ndatasets, as well as in the context of corpora with variable sizes.', 'Building upon recent Deep Neural Network architectures, current approaches\nlying in the intersection of computer vision and natural language processing\nhave achieved unprecedented breakthroughs in tasks like automatic captioning or\nimage retrieval. Most of these learning methods, though, rely on large training\nsets of images associated with human annotations that specifically describe the\nvisual content. In this paper we propose to go a step further and explore the\nmore complex cases where textual descriptions are loosely related to the\nimages. We focus on the particular domain of News articles in which the textual\ncontent often expresses connotative and ambiguous relations that are only\nsuggested but not directly inferred from images. We introduce new deep learning\nmethods that address source detection, popularity prediction, article\nillustration and geolocation of articles. An adaptive CNN architecture is\nproposed, that shares most of the structure for all the tasks, and is suitable\nfor multitask and transfer learning. Deep Canonical Correlation Analysis is\ndeployed for article illustration, and a new loss function based on Great\nCircle Distance is proposed for geolocation. Furthermore, we present\nBreakingNews, a novel dataset with approximately 100K news articles including\nimages, text and captions, and enriched with heterogeneous meta-data (such as\nGPS coordinates and popularity metrics). We show this dataset to be appropriate\nto explore all aforementioned problems, for which we provide a baseline\nperformance using various Deep Learning architectures, and different\nrepresentations of the textual and visual features. We report very promising\nresults and bring to light several limitations of current state-of-the-art in\nthis kind of domain, which we hope will help spur progress in the field.', 'Over the past decade, convolutional neural networks (CNN) have shown very\ncompetitive performance in medical image analysis tasks, such as disease\nclassification, tumor segmentation, and lesion detection. CNN has great\nadvantages in extracting local features of images. However, due to the locality\nof convolution operation, it can not deal with long-range relationships well.\nRecently, transformers have been applied to computer vision and achieved\nremarkable success in large-scale datasets. Compared with natural images,\nmulti-modal medical images have explicit and important long-range dependencies,\nand effective multi-modal fusion strategies can greatly improve the performance\nof deep models. This prompts us to study transformer-based structures and apply\nthem to multi-modal medical images. Existing transformer-based network\narchitectures require large-scale datasets to achieve better performance.\nHowever, medical imaging datasets are relatively small, which makes it\ndifficult to apply pure transformers to medical image analysis. Therefore, we\npropose TransMed for multi-modal medical image classification. TransMed\ncombines the advantages of CNN and transformer to efficiently extract low-level\nfeatures of images and establish long-range dependencies between modalities. We\nevaluated our model for the challenging problem of preoperative diagnosis of\nparotid gland tumors, and the experimental results show the advantages of our\nproposed method. We argue that the combination of CNN and transformer has\ntremendous potential in a large number of medical image analysis tasks. To our\nbest knowledge, this is the first work to apply transformers to medical image\nclassification.', 'This paper develops a model that addresses sentence embedding, a hot topic in\ncurrent natural language processing research, using recurrent neural networks\nwith Long Short-Term Memory (LSTM) cells. Due to its ability to capture long\nterm memory, the LSTM-RNN accumulates increasingly richer information as it\ngoes through the sentence, and when it reaches the last word, the hidden layer\nof the network provides a semantic representation of the whole sentence. In\nthis paper, the LSTM-RNN is trained in a weakly supervised manner on user\nclick-through data logged by a commercial web search engine. Visualization and\nanalysis are performed to understand how the embedding process works. The model\nis found to automatically attenuate the unimportant words and detects the\nsalient keywords in the sentence. Furthermore, these detected keywords are\nfound to automatically activate different cells of the LSTM-RNN, where words\nbelonging to a similar topic activate the same cell. As a semantic\nrepresentation of the sentence, the embedding vector can be used in many\ndifferent applications. These automatic keyword detection and topic allocation\nabilities enabled by the LSTM-RNN allow the network to perform document\nretrieval, a difficult language processing task, where the similarity between\nthe query and documents can be measured by the distance between their\ncorresponding sentence embedding vectors computed by the LSTM-RNN. On a web\nsearch task, the LSTM-RNN embedding is shown to significantly outperform\nseveral existing state of the art methods. We emphasize that the proposed model\ngenerates sentence embedding vectors that are specially useful for web document\nretrieval tasks. A comparison with a well known general sentence embedding\nmethod, the Paragraph Vector, is performed. The results show that the proposed\nmethod in this paper significantly outperforms it for web document retrieval\ntask.', ""This paper presents a meta-learning framework for few-shots One-Class\nClassification (OCC) at test-time, a setting where labeled examples are only\navailable for the positive class, and no supervision is given for the negative\nexample. We consider that we have a set of `one-class classification'\nobjective-tasks with only a small set of positive examples available for each\ntask, and a set of training tasks with full supervision (i.e. highly imbalanced\nclassification). We propose an approach using order-equivariant networks to\nlearn a 'meta' binary-classifier. The model will take as input an example to\nclassify from a given task, as well as the corresponding supervised set of\npositive examples for this OCC task. Thus, the output of the model will be\n'conditioned' on the available positive example of a given task, allowing to\npredict on new tasks and new examples without labeled negative examples. In\nthis paper, we are motivated by an astronomy application. Our goal is to\nidentify if stars belong to a specific stellar group (the 'one-class' for a\ngiven task), called \\textit{stellar streams}, where each stellar stream is a\ndifferent OCC-task. We show that our method transfers well on unseen (test)\nsynthetic streams, and outperforms the baselines even though it is not\nretrained and accesses a much smaller part of the data per task to predict\n(only positive supervision). We see however that it doesn't transfer as well on\nthe real stream GD-1. This could come from intrinsic differences from the\nsynthetic and real stream, highlighting the need for consistency in the\n'nature' of the task for this method. However, light fine-tuning improve\nperformances and outperform our baselines. Our experiments show encouraging\nresults to further explore meta-learning methods for OCC tasks."", 'Lifelong learning has recently attracted attention in building machine\nlearning systems that continually accumulate and transfer knowledge to help\nfuture learning. Unsupervised topic modeling has been popularly used to\ndiscover topics from document collections. However, the application of topic\nmodeling is challenging due to data sparsity, e.g., in a small collection of\n(short) documents and thus, generate incoherent topics and sub-optimal document\nrepresentations. To address the problem, we propose a lifelong learning\nframework for neural topic modeling that can continuously process streams of\ndocument collections, accumulate topics and guide future topic modeling tasks\nby knowledge transfer from several sources to better deal with the sparse data.\nIn the lifelong process, we particularly investigate jointly: (1) sharing\ngenerative homologies (latent topics) over lifetime to transfer prior\nknowledge, and (2) minimizing catastrophic forgetting to retain the past\nlearning via novel selective data augmentation, co-training and topic\nregularization approaches. Given a stream of document collections, we apply the\nproposed Lifelong Neural Topic Modeling (LNTM) framework in modeling three\nsparse document collections as future tasks and demonstrate improved\nperformance quantified by perplexity, topic coherence and information retrieval\ntask.', 'A general-purpose robot should be able to master a wide range of tasks and\nquickly learn a novel one by leveraging past experiences. One-shot imitation\nlearning (OSIL) approaches this goal by training an agent with (pairs of)\nexpert demonstrations, such that at test time, it can directly execute a new\ntask from just one demonstration. However, so far this framework has been\nlimited to training on many variations of one task, and testing on other unseen\nbut similar variations of the same task. In this work, we push for a higher\nlevel of generalization ability by investigating a more ambitious multi-task\nsetup. We introduce a diverse suite of vision-based robot manipulation tasks,\nconsisting of 7 tasks, a total of 61 variations, and a continuum of instances\nwithin each variation. For consistency and comparison purposes, we first train\nand evaluate single-task agents (as done in prior few-shot imitation work). We\nthen study the multi-task setting, where multi-task training is followed by (i)\none-shot imitation on variations within the training tasks, (ii) one-shot\nimitation on new tasks, and (iii) fine-tuning on new tasks. Prior\nstate-of-the-art, while performing well within some single tasks, struggles in\nthese harder multi-task settings. To address these limitations, we propose\nMOSAIC (Multi-task One-Shot Imitation with self-Attention and Contrastive\nlearning), which integrates a self-attention model architecture and a temporal\ncontrastive module to enable better task disambiguation and more robust\nrepresentation learning. Our experiments show that MOSAIC outperforms prior\nstate of the art in learning efficiency, final performance, and learns a\nmulti-task policy with promising generalization ability via fine-tuning on\nnovel tasks.', 'Transfer learning has been developed to improve the performances of different\nbut related tasks in machine learning. However, such processes become less\nefficient with the increase of the size of training data and the number of\ntasks. Moreover, privacy can be violated as some tasks may contain sensitive\nand private data, which are communicated between nodes and tasks. We propose a\nconsensus-based distributed transfer learning framework, where several tasks\naim to find the best linear support vector machine (SVM) classifiers in a\ndistributed network. With alternating direction method of multipliers, tasks\ncan achieve better classification accuracies more efficiently and privately, as\neach node and each task train with their own data, and only decision variables\nare transferred between different tasks and nodes. Numerical experiments on\nMNIST datasets show that the knowledge transferred from the source tasks can be\nused to decrease the risks of the target tasks that lack training data or have\nunbalanced training labels. We show that the risks of the target tasks in the\nnodes without the data of the source tasks can also be reduced using the\ninformation transferred from the nodes who contain the data of the source\ntasks. We also show that the target tasks can enter and leave in real-time\nwithout rerunning the whole algorithm.', 'Many structured prediction tasks in machine vision have a collection of\nacceptable answers, instead of one definitive ground truth answer. Segmentation\nof images, for example, is subject to human labeling bias. Similarly, there are\nmultiple possible pixel values that could plausibly complete occluded image\nregions. State-of-the art supervised learning methods are typically optimized\nto make a single test-time prediction for each query, failing to find other\nmodes in the output space. Existing methods that allow for sampling often\nsacrifice speed or accuracy.\n  We introduce a simple method for training a neural network, which enables\ndiverse structured predictions to be made for each test-time query. For a\nsingle input, we learn to predict a range of possible answers. We compare\nfavorably to methods that seek diversity through an ensemble of networks. Such\nstochastic multiple choice learning faces mode collapse, where one or more\nensemble members fail to receive any training signal. Our best performing\nsolution can be deployed for various tasks, and just involves small\nmodifications to the existing single-mode architecture, loss function, and\ntraining regime. We demonstrate that our method results in quantitative\nimprovements across three challenging tasks: 2D image completion, 3D volume\nestimation, and flow prediction.']","['Transferring knowledge from prior source tasks in solving a new target task\ncan be useful in several learning applications. The application of transfer\nposes two serious challenges which have not been adequately addressed. First,\nthe agent should be able to avoid negative transfer, which happens when the\ntransfer hampers or slows down the learning instead of helping it. Second, the\nagent should be able to selectively transfer, which is the ability to select\nand transfer from different and multiple source tasks for different parts of\nthe state space of the target task. We propose A2T (Attend, Adapt and\nTransfer), an attentive deep architecture which adapts and transfers from these\nsource tasks. Our model is generic enough to effect transfer of either policies\nor value functions. Empirical evaluations on different learning algorithms show\nthat A2T is an effective architecture for transfer by being able to avoid\nnegative transfer while transferring selectively from multiple source tasks in\nthe same domain.', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', 'Modern deep learning based classifiers show very high accuracy on test data\nbut this does not provide sufficient guarantees for safe deployment, especially\nin high-stake AI applications such as medical diagnosis. Usually, predictions\nare obtained without a reliable uncertainty estimate or a formal guarantee.\nConformal prediction (CP) addresses these issues by using the classifier\'s\npredictions, e.g., its probability estimates, to predict confidence sets\ncontaining the true class with a user-specified probability. However, using CP\nas a separate processing step after training prevents the underlying model from\nadapting to the prediction of confidence sets. Thus, this paper explores\nstrategies to differentiate through CP during training with the goal of\ntraining model with the conformal wrapper end-to-end. In our approach,\nconformal training (ConfTr), we specifically ""simulate"" conformalization on\nmini-batches during training. Compared to standard training, ConfTr reduces the\naverage confidence set size (inefficiency) of state-of-the-art CP methods\napplied after training. Moreover, it allows to ""shape"" the confidence sets\npredicted at test time, which is difficult for standard CP. On experiments with\nseveral datasets, we show ConfTr can influence how inefficiency is distributed\nacross classes, or guide the composition of confidence sets in terms of the\nincluded classes, while retaining the guarantees offered by CP.', 'Convolutional Neural Networks (CNNs) have recently emerged as the dominant\nmodel in computer vision. If provided with enough training data, they predict\nalmost any visual quantity. In a discrete setting, such as classification, CNNs\nare not only able to predict a label but often predict a confidence in the form\nof a probability distribution over the output space. In continuous regression\ntasks, such a probability estimate is often lacking. We present a regression\nframework which models the output distribution of neural networks. This output\ndistribution allows us to infer the most likely labeling following a set of\nphysical or modeling constraints. These constraints capture the intricate\ninterplay between different input and output variables, and complement the\noutput of a CNN. However, they may not hold everywhere. Our setup further\nallows to learn a confidence with which a constraint holds, in the form of a\ndistribution of the constrain satisfaction. We evaluate our approach on the\nproblem of intrinsic image decomposition, and show that constrained structured\nregression significantly increases the state-of-the-art.', 'Although recent multi-task learning methods have shown to be effective in\nimproving the generalization of deep neural networks, they should be used with\ncaution for safety-critical applications, such as clinical risk prediction.\nThis is because even if they achieve improved task-average performance, they\nmay still yield degraded performance on individual tasks, which may be critical\n(e.g., prediction of mortality risk). Existing asymmetric multi-task learning\nmethods tackle this negative transfer problem by performing knowledge transfer\nfrom tasks with low loss to tasks with high loss. However, using loss as a\nmeasure of reliability is risky since it could be a result of overfitting. In\nthe case of time-series prediction tasks, knowledge learned for one task (e.g.,\npredicting the sepsis onset) at a specific timestep may be useful for learning\nanother task (e.g., prediction of mortality) at a later timestep, but lack of\nloss at each timestep makes it difficult to measure the reliability at each\ntimestep. To capture such dynamically changing asymmetric relationships between\ntasks in time-series data, we propose a novel temporal asymmetric multi-task\nlearning model that performs knowledge transfer from certain tasks/timesteps to\nrelevant uncertain tasks, based on feature-level uncertainty. We validate our\nmodel on multiple clinical risk prediction tasks against various deep learning\nmodels for time-series prediction, which our model significantly outperforms,\nwithout any sign of negative transfer. Further qualitative analysis of learned\nknowledge graphs by clinicians shows that they are helpful in analyzing the\npredictions of the model. Our final code is available at\nhttps://github.com/anhtuan5696/TPAMTL.', 'Convolutional Neural Networks experience catastrophic forgetting when\noptimized on a sequence of learning problems: as they meet the objective of the\ncurrent training examples, their performance on previous tasks drops\ndrastically. In this work, we introduce a novel framework to tackle this\nproblem with conditional computation. We equip each convolutional layer with\ntask-specific gating modules, selecting which filters to apply on the given\ninput. This way, we achieve two appealing properties. Firstly, the execution\npatterns of the gates allow to identify and protect important filters, ensuring\nno loss in the performance of the model for previously learned tasks. Secondly,\nby using a sparsity objective, we can promote the selection of a limited set of\nkernels, allowing to retain sufficient model capacity to digest new\ntasks.Existing solutions require, at test time, awareness of the task to which\neach example belongs to. This knowledge, however, may not be available in many\npractical scenarios. Therefore, we additionally introduce a task classifier\nthat predicts the task label of each example, to deal with settings in which a\ntask oracle is not available. We validate our proposal on four continual\nlearning datasets. Results show that our model consistently outperforms\nexisting methods both in the presence and the absence of a task oracle.\nNotably, on Split SVHN and Imagenet-50 datasets, our model yields up to 23.98%\nand 17.42% improvement in accuracy w.r.t. competing methods.', 'There are relatively few works dealing with conformal prediction for\nmulti-task learning issues, and this is particularly true for multi-target\nregression. This paper focuses on the problem of providing valid (i.e.,\nfrequency calibrated) multi-variate predictions. To do so, we propose to use\ncopula functions applied to deep neural networks for inductive conformal\nprediction. We show that the proposed method ensures efficiency and validity\nfor multi-target regression problems on various data sets.', 'Multitask learning can be effective when features useful in one task are also\nuseful for other tasks, and the group lasso is a standard method for selecting\na common subset of features. In this paper, we are interested in a less\nrestrictive form of multitask learning, wherein (1) the available features can\nbe organized into subsets according to a notion of similarity and (2) features\nuseful in one task are similar, but not necessarily identical, to the features\nbest suited for other tasks. The main contribution of this paper is a new\nprocedure called Sparse Overlapping Sets (SOS) lasso, a convex optimization\nthat automatically selects similar features for related learning tasks. Error\nbounds are derived for SOSlasso and its consistency is established for squared\nerror loss. In particular, SOSlasso is motivated by multi- subject fMRI studies\nin which functional activity is classified using brain voxels as features.\nExperiments with real and synthetic data demonstrate the advantages of SOSlasso\ncompared to the lasso and group lasso.', 'Conformal predictors are machine learning algorithms that output prediction\nsets that have a guarantee of marginal validity for finite samples with minimal\ndistributional assumptions. This is a property that makes conformal predictors\nuseful for machine learning tasks where we require reliable predictions. It\nwould also be desirable to achieve conditional validity in the same setting, in\nthe sense that validity of the prediction intervals remains valid regardless of\nconditioning on any particular property of the object of the prediction.\nUnfortunately, it has been shown that such conditional validity is impossible\nto guarantee for non-trivial prediction problems for finite samples. In this\narticle, instead of trying to achieve a strong conditional validity result, the\nweaker goal of achieving an approximation to conditional validity is\nconsidered. A new algorithm is introduced to do this by iteratively adjusting a\nconformity measure to deviations from object conditional validity measured in\nthe training data. Along with some theoretical results, experimental results\nare provided for three data sets that demonstrate (1) in real world machine\nlearning tasks, lack of conditional validity is a measurable problem and (2)\nthat the proposed algorithm is effective at alleviating this problem.', 'In class-incremental learning, an agent with limited resources needs to learn\na sequence of classification tasks, forming an ever growing classification\nproblem, with the constraint of not being able to access data from previous\ntasks. The main difference with task-incremental learning, where a task-ID is\navailable at inference time, is that the learner also needs to perform\ncross-task discrimination, i.e. distinguish between classes that have not been\nseen together. Approaches to tackle this problem are numerous and mostly make\nuse of an external memory (buffer) of non-negligible size. In this paper, we\nablate the learning of cross-task features and study its influence on the\nperformance of basic replay strategies used for class-IL. We also define a new\nforgetting measure for class-incremental learning, and see that forgetting is\nnot the principal cause of low performance. Our experimental results show that\nfuture algorithms for class-incremental learning should not only prevent\nforgetting, but also aim to improve the quality of the cross-task features, and\nthe knowledge transfer between tasks. This is especially important when tasks\ncontain limited amount of data.', 'Convolutional Neural Networks (CNNs) have shown to be powerful medical image\nsegmentation models. In this study, we address some of the main unresolved\nissues regarding these models. Specifically, training of these models on small\nmedical image datasets is still challenging, with many studies promoting\ntechniques such as transfer learning. Moreover, these models are infamous for\nproducing over-confident predictions and for failing silently when presented\nwith out-of-distribution (OOD) data at test time. In this paper, we advocate\nfor multi-task learning, i.e., training a single model on several different\ndatasets, spanning several different organs of interest and different imaging\nmodalities. We show that not only a single CNN learns to automatically\nrecognize the context and accurately segment the organ of interest in each\ncontext, but also that such a joint model often has more accurate and\nbetter-calibrated predictions than dedicated models trained separately on each\ndataset. Our experiments show that multi-task learning can outperform transfer\nlearning in medical image segmentation tasks. For detecting OOD data, we\npropose a method based on spectral analysis of CNN feature maps. We show that\ndifferent datasets, representing different imaging modalities and/or different\norgans of interest, have distinct spectral signatures, which can be used to\nidentify whether or not a test image is similar to the images used to train a\nmodel. We show that this approach is far more accurate than OOD detection based\non prediction uncertainty. The methods proposed in this paper contribute\nsignificantly to improving the accuracy and reliability of CNN-based medical\nimage segmentation models.', ""The task of multi-label image classification is to recognize all the object\nlabels presented in an image. Though advancing for years, small objects,\nsimilar objects and objects with high conditional probability are still the\nmain bottlenecks of previous convolutional neural network(CNN) based models,\nlimited by convolutional kernels' representational capacity. Recent vision\ntransformer networks utilize the self-attention mechanism to extract the\nfeature of pixel granularity, which expresses richer local semantic\ninformation, while is insufficient for mining global spatial dependence. In\nthis paper, we point out the three crucial problems that CNN-based methods\nencounter and explore the possibility of conducting specific transformer\nmodules to settle them. We put forward a Multi-label Transformer\narchitecture(MlTr) constructed with windows partitioning, in-window pixel\nattention, cross-window attention, particularly improving the performance of\nmulti-label image classification tasks. The proposed MlTr shows\nstate-of-the-art results on various prevalent multi-label datasets such as\nMS-COCO, Pascal-VOC, and NUS-WIDE with 88.5%, 95.8%, and 65.5% respectively.\nThe code will be available soon at https://github.com/starmemda/MlTr/"", 'We consider the problem of transfer learning -- gaining knowledge from one\nsource task and applying it to a different but related target task. A\nfundamental question in transfer learning is whether combining the data of both\ntasks works better than using only the target task\'s data (equivalently,\nwhether a ""positive information transfer"" happens). We study this question\nformally in a linear regression setting where a two-layer linear neural network\nestimator combines both tasks\' data. The estimator uses a shared parameter\nvector for both tasks and exhibits positive or negative information transfer by\nvarying dataset characteristics.\n  We characterize the precise asymptotic limit of the prediction risk of the\nabove estimator when the sample sizes increase with the feature dimension\nproportionally at fixed ratios. We also show that the asymptotic limit is\nsufficiently accurate for finite dimensions. Then, we provide the exact\ncondition to determine positive (and negative) information transfer in a\nrandom-effect model, leading to several theoretical insights. For example, the\nrisk curve is non-monotone under model shift, thus motivating a transfer\nlearning procedure that progressively adds data from the source task. We\nvalidate this procedure\'s efficiency on text classification tasks with a neural\nnetwork that applies a shared feature space for both tasks, similar to the\nabove estimator. The main ingredient of the analysis is finding the\nhigh-dimensional asymptotic limits of various functions involving the sum of\ntwo independent sample covariance matrices with different population covariance\nmatrices, which may be of independent interest.', 'This paper presents an online transfer learning framework for improving\ntemperature predictions in residential buildings. In transfer learning,\nprediction models trained under a set of available data from a target domain\n(e.g., house with limited data) can be improved through the use of data\ngenerated from similar source domains (e.g., houses with rich data). Given also\nthe need for prediction models that can be trained online (e.g., as part of a\nmodel-predictive-control implementation), this paper introduces the generalized\nonline transfer learning algorithm (GOTL). It employs a weighted combination of\nthe available predictors (i.e., the target and source predictors) and\nguarantees convergence to the best weighted predictor. Furthermore, the use of\nTransfer Component Analysis (TCA) allows for using more than a single source\ndomains, since it may facilitate the fit of a single model on more than one\nsource domains (houses). This allows GOTL to transfer knowledge from more than\none source domains. We further validate our results through experiments in\nclimate control for residential buildings and show that GOTL may lead to\nnon-negligible energy savings for given comfort levels.', 'In this work, we focus on the task of generating natural language\ndescriptions from a structured table of facts containing fields (such as\nnationality, occupation, etc) and values (such as Indian, actor, director,\netc). One simple choice is to treat the table as a sequence of fields and\nvalues and then use a standard seq2seq model for this task. However, such a\nmodel is too generic and does not exploit task-specific characteristics. For\nexample, while generating descriptions from a table, a human would attend to\ninformation at two levels: (i) the fields (macro level) and (ii) the values\nwithin the field (micro level). Further, a human would continue attending to a\nfield for a few timesteps till all the information from that field has been\nrendered and then never return back to this field (because there is nothing\nleft to say about it). To capture this behavior we use (i) a fused bifocal\nattention mechanism which exploits and combines this micro and macro level\ninformation and (ii) a gated orthogonalization mechanism which tries to ensure\nthat a field is remembered for a few time steps and then forgotten. We\nexperiment with a recently released dataset which contains fact tables about\npeople and their corresponding one line biographical descriptions in English.\nIn addition, we also introduce two similar datasets for French and German. Our\nexperiments show that the proposed model gives 21% relative improvement over a\nrecently proposed state of the art method and 10% relative improvement over\nbasic seq2seq models. The code and the datasets developed as a part of this\nwork are publicly available.', 'Previous researchers have considered sentiment analysis as a document\nclassification task, in which input documents are classified into predefined\nsentiment classes. Although there are sentences in a document that support\nimportant evidences for sentiment analysis and sentences that do not, they have\ntreated the document as a bag of sentences. In other words, they have not\nconsidered the importance of each sentence in the document. To effectively\ndetermine polarity of a document, each sentence in the document should be dealt\nwith different degrees of importance. To address this problem, we propose a\ndocument-level sentence classification model based on deep neural networks, in\nwhich the importance degrees of sentences in documents are automatically\ndetermined through gate mechanisms. To verify our new sentiment analysis model,\nwe conducted experiments using the sentiment datasets in the four different\ndomains such as movie reviews, hotel reviews, restaurant reviews, and music\nreviews. In the experiments, the proposed model outperformed previous\nstate-of-the-art models that do not consider importance differences of\nsentences in a document. The experimental results show that the importance of\nsentences should be considered in a document-level sentiment classification\ntask.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', ""An accurate abstractive summary of a document should contain all its salient\ninformation and should be logically entailed by the input document. We improve\nthese important aspects of abstractive summarization via multi-task learning\nwith the auxiliary tasks of question generation and entailment generation,\nwhere the former teaches the summarization model how to look for salient\nquestioning-worthy details, and the latter teaches the model how to rewrite a\nsummary which is a directed-logical subset of the input document. We also\npropose novel multi-task architectures with high-level (semantic)\nlayer-specific sharing across multiple encoder and decoder layers of the three\ntasks, as well as soft-sharing mechanisms (and show performance ablations and\nanalysis examples of each contribution). Overall, we achieve statistically\nsignificant improvements over the state-of-the-art on both the CNN/DailyMail\nand Gigaword datasets, as well as on the DUC-2002 transfer setup. We also\npresent several quantitative and qualitative analysis studies of our model's\nlearned saliency and entailment skills."", 'Black-box machine learning models are now routinely used in high-risk\nsettings, like medical diagnostics, which demand uncertainty quantification to\navoid consequential model failures. Conformal prediction is a user-friendly\nparadigm for creating statistically rigorous uncertainty sets/intervals for the\npredictions of such models. Critically, the sets are valid in a\ndistribution-free sense: they possess explicit, non-asymptotic guarantees even\nwithout distributional assumptions or model assumptions. One can use conformal\nprediction with any pre-trained model, such as a neural network, to produce\nsets that are guaranteed to contain the ground truth with a user-specified\nprobability, such as 90%. It is easy-to-understand, easy-to-use, and general,\napplying naturally to problems arising in the fields of computer vision,\nnatural language processing, deep reinforcement learning, and so on.\n  This hands-on introduction is aimed to provide the reader a working\nunderstanding of conformal prediction and related distribution-free uncertainty\nquantification techniques with one self-contained document. We lead the reader\nthrough practical theory for and examples of conformal prediction and describe\nits extensions to complex machine learning tasks involving structured outputs,\ndistribution shift, time-series, outliers, models that abstain, and more.\nThroughout, there are many explanatory illustrations, examples, and code\nsamples in Python. With each code sample comes a Jupyter notebook implementing\nthe method on a real-data example; the notebooks can be accessed and easily run\nusing our codebase.', 'In many natural language processing (NLP) tasks, a document is commonly\nmodeled as a bag of words using the term frequency-inverse document frequency\n(TF-IDF) vector. One major shortcoming of the frequency-based TF-IDF feature\nvector is that it ignores word orders that carry syntactic and semantic\nrelationships among the words in a document, and they can be important in some\nNLP tasks such as genre classification. This paper proposes a novel distributed\nvector representation of a document: a simple recurrent-neural-network language\nmodel (RNN-LM) or a long short-term memory RNN language model (LSTM-LM) is\nfirst created from all documents in a task; some of the LM parameters are then\nadapted by each document, and the adapted parameters are vectorized to\nrepresent the document. The new document vectors are labeled as DV-RNN and\nDV-LSTM respectively. We believe that our new document vectors can capture some\nhigh-level sequential information in the documents, which other current\ndocument representations fail to capture. The new document vectors were\nevaluated in the genre classification of documents in three corpora: the Brown\nCorpus, the BNC Baby Corpus and an artificially created Penn Treebank dataset.\nTheir classification performances are compared with the performance of TF-IDF\nvector and the state-of-the-art distributed memory model of paragraph vector\n(PV-DM). The results show that DV-LSTM significantly outperforms TF-IDF and\nPV-DM in most cases, and combinations of the proposed document vectors with\nTF-IDF or PV-DM may further improve performance.', 'Different people have different habits of describing their intents in\nconversations. Some people tend to deliberate their intents in several\nsuccessive utterances, i.e., they use several consistent messages for\nreadability instead of a long sentence to express their question. This creates\na predicament faced by the application of dialogue systems, especially in\nreal-world industry scenarios, in which the dialogue system is unsure whether\nit should answer the query of user immediately or wait for further\nsupplementary input. Motivated by such an interesting predicament, we define a\nnovel Wait-or-Answer task for dialogue systems. We shed light on a new research\ntopic about how the dialogue system can be more intelligent to behave in this\nWait-or-Answer quandary. Further, we propose a predictive approach named\nPredict-then-Decide (PTD) to tackle this Wait-or-Answer task. More\nspecifically, we take advantage of a decision model to help the dialogue system\ndecide whether to wait or answer. The decision of decision model is made with\nthe assistance of two ancillary prediction models: a user prediction and an\nagent prediction. The user prediction model tries to predict what the user\nwould supplement and uses its prediction to persuade the decision model that\nthe user has some information to add, so the dialogue system should wait. The\nagent prediction model tries to predict the answer of the dialogue system and\nconvince the decision model that it is a superior choice to answer the query of\nuser immediately since the input of user has come to an end. We conduct our\nexperiments on two real-life scenarios and three public datasets. Experimental\nresults on five datasets show our proposed PTD approach significantly\noutperforms the existing models in solving this Wait-or-Answer problem.', 'The Query Focused Text Summarization (QFTS) task aims at building systems\nthat generate the summary of the text document(s) based on the given query. A\nkey challenge in addressing this task is the lack of large labeled data for\ntraining the summarization model. In this paper, we address this challenge by\nexploring a series of domain adaptation techniques. Given the recent success of\npre-trained transformer models in a wide range of natural language processing\ntasks, we utilize such models to generate abstractive summaries for the QFTS\ntask for both single-document and multi-document scenarios. For domain\nadaptation, we apply a variety of techniques using pre-trained\ntransformer-based summarization models including transfer learning, weakly\nsupervised learning, and distant supervision. Extensive experiments on six\ndatasets show that our proposed approach is very effective in generating\nabstractive summaries for the QFTS task while setting a new state-of-the-art\nresult in several datasets across a set of automatic and human evaluation\nmetrics.', 'Conformal predictors are an important class of algorithms that allow\npredictions to be made with a user-defined confidence level. They are able to\ndo this by outputting prediction sets, rather than simple point predictions.\nThe conformal predictor is valid in the sense that the accuracy of its\npredictions is guaranteed to meet the confidence level, only assuming\nexchangeability in the data. Since accuracy is guaranteed, the performance of a\nconformal predictor is measured through the efficiency of the prediction sets.\nTypically, a conformal predictor is built on an underlying machine learning\nalgorithm and hence its predictive power is inherited from this algorithm.\nHowever, since the underlying machine learning algorithm is not trained with\nthe objective of minimizing predictive efficiency it means that the resulting\nconformal predictor may be sub-optimal and not aligned sufficiently to this\nobjective. Hence, in this study we consider an approach to train the conformal\npredictor directly with maximum predictive efficiency as the optimization\nobjective, and we focus specifically on the inductive conformal predictor for\nclassification. To do this, the conformal predictor is approximated by a\ndifferentiable objective function and gradient descent used to optimize it. The\nresulting parameter estimates are then passed to a proper inductive conformal\npredictor to give valid prediction sets. We test the method on several real\nworld data sets and find that the method is promising and in most cases gives\nimproved predictive efficiency against a baseline conformal predictor.', ""Neural machine translation (NMT) has arguably achieved human level parity\nwhen trained and evaluated at the sentence-level. Document-level neural machine\ntranslation has received less attention and lags behind its sentence-level\ncounterpart. The majority of the proposed document-level approaches investigate\nways of conditioning the model on several source or target sentences to capture\ndocument context. These approaches require training a specialized NMT model\nfrom scratch on parallel document-level corpora. We propose an approach that\ndoesn't require training a specialized model on parallel document-level corpora\nand is applied to a trained sentence-level NMT model at decoding time. We\nprocess the document from left to right multiple times and self-train the\nsentence-level model on pairs of source sentences and generated translations.\nOur approach reinforces the choices made by the model, thus making it more\nlikely that the same choices will be made in other sentences in the document.\nWe evaluate our approach on three document-level datasets: NIST\nChinese-English, WMT'19 Chinese-English and OpenSubtitles English-Russian. We\ndemonstrate that our approach has higher BLEU score and higher human preference\nthan the baseline. Qualitative analysis of our approach shows that choices made\nby model are consistent across the document."", 'Conformal Prediction is a framework that produces prediction intervals based\non the output from a machine learning algorithm. In this paper we explore the\ncase when training data is made up of multiple parts available in different\nsources that cannot be pooled. We here consider the regression case and propose\na method where a conformal predictor is trained on each data source\nindependently, and where the prediction intervals are then combined into a\nsingle interval. We call the approach Non-Disclosed Conformal Prediction\n(NDCP), and we evaluate it on a regression dataset from the UCI machine\nlearning repository using support vector regression as the underlying machine\nlearning algorithm, with varying number of data sources and sizes. The results\nshow that the proposed method produces conservatively valid prediction\nintervals, and while we cannot retain the same efficiency as when all data is\nused, efficiency is improved through the proposed approach as compared to\npredicting using a single arbitrarily chosen source.', 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'We first exhibit a multimodal image registration task, for which a neural\nnetwork trained on a dataset with noisy labels reaches almost perfect accuracy,\nfar beyond noise variance. This surprising auto-denoising phenomenon can be\nexplained as a noise averaging effect over the labels of similar input\nexamples. This effect theoretically grows with the number of similar examples;\nthe question is then to define and estimate the similarity of examples.\n  We express a proper definition of similarity, from the neural network\nperspective, i.e. we quantify how undissociable two inputs $A$ and $B$ are,\ntaking a machine learning viewpoint: how much a parameter variation designed to\nchange the output for $A$ would impact the output for $B$ as well?\n  We study the mathematical properties of this similarity measure, and show how\nto use it on a trained network to estimate sample density, in low complexity,\nenabling new types of statistical analysis for neural networks. We analyze data\nby retrieving samples perceived as similar by the network, and are able to\nquantify the denoising effect without requiring true labels. We also propose,\nduring training, to enforce that examples known to be similar should also be\nseen as similar by the network, and notice speed-up training effects for\ncertain datasets.', 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', 'Novel contexts may often arise in complex querying scenarios such as in\nevidence-based medicine (EBM) involving biomedical literature, that may not\nexplicitly refer to entities or canonical concept forms occurring in any fact-\nor rule-based knowledge source such as an ontology like the UMLS. Moreover,\nhidden associations between candidate concepts meaningful in the current\ncontext, may not exist within a single document, but within the collection, via\nalternate lexical forms. Therefore, inspired by the recent success of\nsequence-to-sequence neural models in delivering the state-of-the-art in a wide\nrange of NLP tasks, we develop a novel sequence-to-set framework with neural\nattention for learning document representations that can effect term transfer\nwithin the corpus, for semantically tagging a large collection of documents. We\ndemonstrate that our proposed method can be effective in both a supervised\nmulti-label classification setup for text categorization, as well as in a\nunique unsupervised setting with no human-annotated document labels that uses\nno external knowledge resources and only corpus-derived term statistics to\ndrive the training. Further, we show that semi-supervised training using our\narchitecture on large amounts of unlabeled data can augment performance on the\ntext categorization task when limited labeled data is available. Our approach\nto generate document encodings employing our sequence-to-set models for\ninference of semantic tags, gives to the best of our knowledge, the\nstate-of-the-art for both, the unsupervised query expansion task for the TREC\nCDS 2016 challenge dataset when evaluated on an Okapi BM25--based document\nretrieval system; and also over the MLTM baseline (Soleimani et al, 2016), for\nboth supervised and semi-supervised multi-label prediction tasks on the\ndel.icio.us and Ohsumed datasets. We will make our code and data publicly\navailable.', 'While the use of deep learning in drug discovery is gaining increasing\nattention, the lack of methods to compute reliable errors in prediction for\nNeural Networks prevents their application to guide decision making in domains\nwhere identifying unreliable predictions is essential, e.g. precision medicine.\nHere, we present a framework to compute reliable errors in prediction for\nNeural Networks using Test-Time Dropout and Conformal Prediction. Specifically,\nthe algorithm consists of training a single Neural Network using dropout, and\nthen applying it N times to both the validation and test sets, also employing\ndropout in this step. Therefore, for each instance in the validation and test\nsets an ensemble of predictions were generated. The residuals and absolute\nerrors in prediction for the validation set were then used to compute\nprediction errors for test set instances using Conformal Prediction. We show\nusing 24 bioactivity data sets from ChEMBL 23 that dropout Conformal Predictors\nare valid (i.e., the fraction of instances whose true value lies within the\npredicted interval strongly correlates with the confidence level) and\nefficient, as the predicted confidence intervals span a narrower set of values\nthan those computed with Conformal Predictors generated using Random Forest\n(RF) models. Lastly, we show in retrospective virtual screening experiments\nthat dropout and RF-based Conformal Predictors lead to comparable retrieval\nrates of active compounds. Overall, we propose a computationally efficient\nframework (as only N extra forward passes are required in addition to training\na single network) to harness Test-Time Dropout and the Conformal Prediction\nframework, and to thereby generate reliable prediction errors for deep Neural\nNetworks.', 'Most classifiers operate by selecting the maximum of an estimate of the\nconditional distribution $p(y|x)$ where $x$ stands for the features of the\ninstance to be classified and $y$ denotes its label. This often results in a\n{\\em hubristic bias}: overconfidence in the assignment of a definite label.\nUsually, the observations are concentrated on a small volume but the classifier\nprovides definite predictions for the entire space. We propose constructing\nconformal prediction sets which contain a set of labels rather than a single\nlabel. These conformal prediction sets contain the true label with probability\n$1-\\alpha$. Our construction is based on $p(x|y)$ rather than $p(y|x)$ which\nresults in a classifier that is very cautious: it outputs the null set ---\nmeaning ""I don\'t know"" --- when the object does not resemble the training\nexamples. An important property of our approach is that adversarial attacks are\nlikely to be predicted as the null set or would also include the true label. We\ndemonstrate the performance on the ImageNet ILSVRC dataset and the CelebA and\nIMDB-Wiki facial datasets using high dimensional features obtained from state\nof the art convolutional neural networks.', 'Standard neural machine translation (NMT) is on the assumption of\ndocument-level context independent. Most existing document-level NMT methods\nare satisfied with a smattering sense of brief document-level information,\nwhile this work focuses on exploiting detailed document-level context in terms\nof multiple forms of document embeddings, which is capable of sufficiently\nmodeling deeper and richer document-level context. The proposed document-aware\nNMT is implemented to enhance the Transformer baseline by introducing both\nglobal and local document-level clues on the source end. Experiments show that\nthe proposed method significantly improves the translation performance over\nstrong baselines and other related studies.', 'Sharing information between multiple tasks enables algorithms to achieve good\ngeneralization performance even from small amounts of training data. However,\nin a realistic scenario of multi-task learning not all tasks are equally\nrelated to each other, hence it could be advantageous to transfer information\nonly between the most related tasks. In this work we propose an approach that\nprocesses multiple tasks in a sequence with sharing between subsequent tasks\ninstead of solving all tasks jointly. Subsequently, we address the question of\ncurriculum learning of tasks, i.e. finding the best order of tasks to be\nlearned. Our approach is based on a generalization bound criterion for choosing\nthe task order that optimizes the average expected classification performance\nover all tasks. Our experimental results show that learning multiple related\ntasks sequentially can be more effective than learning them jointly, the order\nin which tasks are being solved affects the overall performance, and that our\nmodel is able to automatically discover the favourable order of tasks.', 'We investigate whether and where multi-task learning (MTL) can improve\nperformance on NLP problems related to argumentation mining (AM), in particular\nargument component identification. Our results show that MTL performs\nparticularly well (and better than single-task learning) when little training\ndata is available for the main task, a common scenario in AM. Our findings\nchallenge previous assumptions that conceptualizations across AM datasets are\ndivergent and that MTL is difficult for semantic or higher-level tasks.', 'The machine learning literature contains several constructions for prediction\nintervals that are intuitively reasonable but ultimately ad-hoc in that they do\nnot come with provable performance guarantees. We present methods from the\nstatistics literature that can be used efficiently with neural networks under\nminimal assumptions with guaranteed performance. We propose a neural network\nthat outputs three values instead of a single point estimate and optimizes a\nloss function motivated by the standard quantile regression loss. We provide\ntwo prediction interval methods with finite sample coverage guarantees solely\nunder the assumption that the observations are independent and identically\ndistributed. The first method leverages the conformal inference framework and\nprovides average coverage. The second method provides a new, stronger guarantee\nby conditioning on the observed data. Lastly, our loss function does not\ncompromise the predictive accuracy of the network like other prediction\ninterval methods. We demonstrate the ease of use of our procedures as well as\nits improvements over other methods on both simulated and real data. As most\ndeep networks can easily be modified by our method to output predictions with\nvalid prediction intervals, its use should become standard practice, much like\nreporting standard errors along with mean estimates.', 'In the era of pre-trained language models, Transformers are the de facto\nchoice of model architectures. While recent research has shown promise in\nentirely convolutional, or CNN, architectures, they have not been explored\nusing the pre-train-fine-tune paradigm. In the context of language models, are\nconvolutional models competitive to Transformers when pre-trained? This paper\ninvestigates this research question and presents several interesting findings.\nAcross an extensive set of experiments on 8 datasets/tasks, we find that\nCNN-based pre-trained models are competitive and outperform their Transformer\ncounterpart in certain scenarios, albeit with caveats. Overall, the findings\noutlined in this paper suggest that conflating pre-training and architectural\nadvances is misguided and that both advances should be considered\nindependently. We believe our research paves the way for a healthy amount of\noptimism in alternative architectures.', 'Due to the increasing amount of data on the internet, finding a\nhighly-informative, low-dimensional representation for text is one of the main\nchallenges for efficient natural language processing tasks including text\nclassification. This representation should capture the semantic information of\nthe text while retaining their relevance level for document classification.\nThis approach maps the documents with similar topics to a similar space in\nvector space representation. To obtain representation for large text, we\npropose the utilization of deep Siamese neural networks. To embed document\nrelevance in topics in the distributed representation, we use a Siamese neural\nnetwork to jointly learn document representations. Our Siamese network consists\nof two sub-network of multi-layer perceptron. We examine our representation for\nthe text categorization task on BBC news dataset. The results show that the\nproposed representations outperform the conventional and state-of-the-art\nrepresentations in the text classification task on this dataset.', 'Training recurrent neural networks to model long term dependencies is\ndifficult. Hence, we propose to use external linguistic knowledge as an\nexplicit signal to inform the model which memories it should utilize.\nSpecifically, external knowledge is used to augment a sequence with typed edges\nbetween arbitrarily distant elements, and the resulting graph is decomposed\ninto directed acyclic subgraphs. We introduce a model that encodes such graphs\nas explicit memory in recurrent neural networks, and use it to model\ncoreference relations in text. We apply our model to several text comprehension\ntasks and achieve new state-of-the-art results on all considered benchmarks,\nincluding CNN, bAbi, and LAMBADA. On the bAbi QA tasks, our model solves 15 out\nof the 20 tasks with only 1000 training examples per task. Analysis of the\nlearned representations further demonstrates the ability of our model to encode\nfine-grained entity information across a document.', ""A reasonable approach for fact checking a claim involves retrieving\npotentially relevant documents from different sources (e.g., news websites,\nsocial media, etc.), determining the stance of each document with respect to\nthe claim, and finally making a prediction about the claim's factuality by\naggregating the strength of the stances, while taking the reliability of the\nsource into account. Moreover, a fact checking system should be able to explain\nits decision by providing relevant extracts (rationales) from the documents.\nYet, this setup is not directly supported by existing datasets, which treat\nfact checking, document retrieval, source credibility, stance detection and\nrationale extraction as independent tasks. In this paper, we support the\ninterdependencies between these tasks as annotations in the same corpus. We\nimplement this setup on an Arabic fact checking corpus, the first of its kind."", 'Explaining the predictions of AI models is paramount in safety-critical\napplications, such as in legal or medical domains. One form of explanation for\na prediction is an extractive rationale, i.e., a subset of features of an\ninstance that lead the model to give its prediction on the instance. Previous\nworks on generating extractive rationales usually employ a two-phase model: a\nselector that selects the most important features (i.e., the rationale)\nfollowed by a predictor that makes the prediction based exclusively on the\nselected features. One disadvantage of these works is that the main signal for\nlearning to select features comes from the comparison of the answers given by\nthe predictor and the ground-truth answers. In this work, we propose to squeeze\nmore information from the predictor via an information calibration method. More\nprecisely, we train two models jointly: one is a typical neural model that\nsolves the task at hand in an accurate but black-box manner, and the other is a\nselector-predictor model that additionally produces a rationale for its\nprediction. The first model is used as a guide to the second model. We use an\nadversarial-based technique to calibrate the information extracted by the two\nmodels such that the difference between them is an indicator of the missed or\nover-selected features. In addition, for natural language tasks, we propose to\nuse a language-model-based regularizer to encourage the extraction of fluent\nrationales. Experimental results on a sentiment analysis task as well as on\nthree tasks from the legal domain show the effectiveness of our approach to\nrationale extraction.', 'Hierarchical neural architectures are often used to capture long-distance\ndependencies and have been applied to many document-level tasks such as\nsummarization, document segmentation, and sentiment analysis. However,\neffective usage of such a large context can be difficult to learn, especially\nin the case where there is limited labeled data available. Building on the\nrecent success of language model pretraining methods for learning flat\nrepresentations of text, we propose algorithms for pre-training hierarchical\ndocument representations from unlabeled data. Unlike prior work, which has\nfocused on pre-training contextual token representations or context-independent\n{sentence/paragraph} representations, our hierarchical document representations\ninclude fixed-length sentence/paragraph representations which integrate\ncontextual information from the entire documents. Experiments on document\nsegmentation, document-level question answering, and extractive document\nsummarization demonstrate the effectiveness of the proposed pre-training\nalgorithms.', 'Modern machine learning algorithms are capable of providing remarkably\naccurate point-predictions; however, questions remain about their statistical\nreliability. Unlike conventional machine learning methods, conformal prediction\nalgorithms return confidence sets (i.e., set-valued predictions) that\ncorrespond to a given significance level. Moreover, these confidence sets are\nvalid in the sense that they guarantee finite sample control over type 1 error\nprobabilities, allowing the practitioner to choose an acceptable error rate. In\nour paper, we propose inductive conformal prediction (ICP) algorithms for the\ntasks of text infilling and part-of-speech (POS) prediction for natural\nlanguage data. We construct new conformal prediction-enhanced bidirectional\nencoder representations from transformers (BERT) and bidirectional long\nshort-term memory (BiLSTM) algorithms for POS tagging and a new conformal\nprediction-enhanced BERT algorithm for text infilling. We analyze the\nperformance of the algorithms in simulations using the Brown Corpus, which\ncontains over 57,000 sentences. Our results demonstrate that the ICP algorithms\nare able to produce valid set-valued predictions that are small enough to be\napplicable in real-world applications. We also provide a real data example for\nhow our proposed set-valued predictions can improve machine generated audio\ntranscriptions.', 'In this paper, we present a novel approach for conformal prediction (CP), in\nwhich we aim to identify a set of promising prediction candidates -- in place\nof a single prediction. This set is guaranteed to contain a correct answer with\nhigh probability, and is well-suited for many open-ended classification tasks.\nIn the standard CP paradigm, the predicted set can often be unusably large and\nalso costly to obtain. This is particularly pervasive in settings where the\ncorrect answer is not unique, and the number of total possible answers is high.\nWe first expand the CP correctness criterion to allow for additional, inferred\n""admissible"" answers, which can substantially reduce the size of the predicted\nset while still providing valid performance guarantees. Second, we amortize\ncosts by conformalizing prediction cascades, in which we aggressively prune\nimplausible labels early on by using progressively stronger classifiers --\nagain, while still providing valid performance guarantees. We demonstrate the\nempirical effectiveness of our approach for multiple applications in natural\nlanguage processing and computational chemistry for drug discovery.', 'The great success of Convolutional Neural Networks (CNN) for facial attribute\nprediction relies on a large amount of labeled images. Facial image datasets\nare usually annotated by some commonly used attributes (e.g., gender), while\nlabels for the other attributes (e.g., big nose) are limited which causes their\nprediction challenging. To address this problem, we use a new Multi-Task\nLearning (MTL) paradigm in which a facial attribute predictor uses the\nknowledge of other related attributes to obtain a better generalization\nperformance. Here, we leverage MLT paradigm in two problem settings. First, it\nis assumed that the structure of the tasks (e.g., grouping pattern of facial\nattributes) is known as a prior knowledge, and parameters of the tasks (i.e.,\npredictors) within the same group are represented by a linear combination of a\nlimited number of underlying basis tasks. Here, a sparsity constraint on the\ncoefficients of this linear combination is also considered such that each task\nis represented in a more structured and simpler manner. Second, it is assumed\nthat the structure of the tasks is unknown, and then structure and parameters\nof the tasks are learned jointly by using a Laplacian regularization framework.\nOur MTL methods are compared with competing methods for facial attribute\nprediction to show its effectiveness.', ""The goal of our research is to develop methods advancing automatic visual\nrecognition. In order to predict the unique or multiple labels associated to an\nimage, we study different kind of Deep Neural Networks architectures and\nmethods for supervised features learning. We first draw up a state-of-the-art\nreview of the Convolutional Neural Networks aiming to understand the history\nbehind this family of statistical models, the limit of modern architectures and\nthe novel techniques currently used to train deep CNNs. The originality of our\nwork lies in our approach focusing on tasks with a low amount of data. We\nintroduce different models and techniques to achieve the best accuracy on\nseveral kind of datasets, such as a medium dataset of food recipes (100k\nimages) for building a web API, or a small dataset of satellite images (6,000)\nfor the DSG online challenge that we've won. We also draw up the\nstate-of-the-art in Weakly Supervised Learning, introducing different kind of\nCNNs able to localize regions of interest. Our last contribution is a\nframework, build on top of Torch7, for training and testing deep models on any\nvisual recognition tasks and on datasets of any scale."", 'Recurrent Neural Networks were, until recently, one of the best ways to\ncapture the timely dependencies in sequences. However, with the introduction of\nthe Transformer, it has been proven that an architecture with only\nattention-mechanisms without any RNN can improve on the results in various\nsequence processing tasks (e.g. NLP). Multiple studies since then have shown\nthat similar approaches can be applied for images, point clouds, video, audio\nor time series forecasting. Furthermore, solutions such as the Perceiver or the\nInformer have been introduced to expand on the applicability of the\nTransformer. Our main objective is testing and evaluating the effectiveness of\napplying Transformer-like models on time series data, tackling susceptibility\nto anomalies, context awareness and space complexity by fine-tuning the\nhyperparameters, preprocessing the data, applying dimensionality reduction or\nconvolutional encodings, etc. We are also looking at the problem of next-frame\nprediction and exploring ways to modify existing solutions in order to achieve\nhigher performance and learn generalized knowledge.', 'Transformer, which originates from machine translation, is particularly\npowerful at modeling long-range dependencies. Currently, the transformer is\nmaking revolutionary progress in various vision tasks, leading to significant\nperformance improvements compared with the convolutional neural network (CNN)\nbased frameworks. In this paper, we conduct extensive research on exploiting\nthe contributions of transformers for accurate and reliable salient object\ndetection. For the former, we apply transformer to a deterministic model, and\nexplain that the effective structure modeling and global context modeling\nabilities lead to its superior performance compared with the CNN based\nframeworks. For the latter, we observe that both CNN and transformer based\nframeworks suffer greatly from the over-confidence issue, where the models tend\nto generate wrong predictions with high confidence. To estimate the reliability\ndegree of both CNN- and transformer-based frameworks, we further present a\nlatent variable model, namely inferential generative adversarial network\n(iGAN), based on the generative adversarial network (GAN). The stochastic\nattribute of the latent variable makes it convenient to estimate the predictive\nuncertainty, serving as an auxiliary output to evaluate the reliability of\nmodel prediction. Different from the conventional GAN, which defines the\ndistribution of the latent variable as fixed standard normal distribution\n$\\mathcal{N}(0,\\mathbf{I})$, the proposed iGAN infers the latent variable by\ngradient-based Markov Chain Monte Carlo (MCMC), namely Langevin dynamics,\nleading to an input-dependent latent variable model. We apply our proposed iGAN\nto both fully and weakly supervised salient object detection, and explain that\niGAN within the transformer framework leads to both accurate and reliable\nsalient object detection.', 'In medical applications, deep learning methods are built to automate\ndiagnostic tasks. However, a clinically relevant question that practitioners\nusually face, is how to predict the future trajectory of a disease (prognosis).\nCurrent methods for such a problem often require domain knowledge, and are\ncomplicated to apply. In this paper, we formulate the prognosis prediction\nproblem as a one-to-many forecasting problem from multimodal data. Inspired by\na clinical decision-making process with two agents -- a radiologist and a\ngeneral practitioner, we model a prognosis prediction problem with two\ntransformer-based components that share information between each other. The\nfirst block in this model aims to analyze the imaging data, and the second\nblock leverages the internal representations of the first one as inputs, also\nfusing them with auxiliary patient data. We show the effectiveness of our\nmethod in predicting the development of structural knee osteoarthritis changes\nover time. Our results show that the proposed method outperforms the\nstate-of-the-art baselines in terms of various performance metrics. In\naddition, we empirically show that the existence of the multi-agent\ntransformers with depths of 2 is sufficient to achieve good performances. Our\ncode is publicly available at \\url{https://github.com/MIPT-Oulu/CLIMAT}.', ""Deep Neural Networks achieve state-of-the-art results in many different\nproblem settings by exploiting vast amounts of training data. However,\ncollecting, storing and - in the case of supervised learning - labelling the\ndata is expensive and time-consuming. Additionally, assessing the networks'\ngeneralization abilities or predicting how the inferred output changes under\ninput transformations is complicated since the networks are usually treated as\na black box. Both of these problems can be mitigated by incorporating prior\nknowledge into the neural network. One promising approach, inspired by the\nsuccess of convolutional neural networks in computer vision tasks, is to\nincorporate knowledge about symmetric geometrical transformations of the\nproblem to solve that affect the output in a predictable way. This promises an\nincreased data efficiency and more interpretable network outputs. In this\nsurvey, we try to give a concise overview about different approaches that\nincorporate geometrical prior knowledge into neural networks. Additionally, we\nconnect those methods to 3D object detection for autonomous driving, where we\nexpect promising results when applying those methods."", 'The data distribution commonly evolves over time leading to problems such as\nconcept drift that often decrease classifier performance. Current techniques\nare not adequate for this problem because they either require detailed\nknowledge of the transformation or are not suited for anticipating unseen\ndomains but can only adapt to domains, where data samples are available. We\nseek to predict unseen data (and their labels) allowing us to tackle challenges\ns a non-constant data distribution in a proactive manner rather than detecting\nand reacting to already existing changes that might already have led to errors.\nTo this end, we learn a domain transformer in an unsupervised manner that\nallows generating data of unseen domains. Our approach first matches\nindependently learned latent representations of two given domains obtained from\nan auto-encoder using a Cycle-GAN. In turn, a transformation of the original\nsamples can be learned that can be applied iteratively to extrapolate to unseen\ndomains. Our evaluation of CNNs on image data confirms the usefulness of the\napproach. It also achieves very good results on the well-known problem of\nunsupervised domain adaption, where only labels but no samples have to be\npredicted. Code is available at https://github.com/JohnTailor/DoTra.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.', 'Deep reading models for question-answering have demonstrated promising\nperformance over the last couple of years. However current systems tend to\nlearn how to cleverly extract a span of the source document, based on its\nsimilarity with the question, instead of seeking for the appropriate answer.\nIndeed, a reading machine should be able to detect relevant passages in a\ndocument regarding a question, but more importantly, it should be able to\nreason over the important pieces of the document in order to produce an answer\nwhen it is required. To motivate this purpose, we present ReviewQA, a\nquestion-answering dataset based on hotel reviews. The questions of this\ndataset are linked to a set of relational understanding competencies that we\nexpect a model to master. Indeed, each question comes with an associated type\nthat characterizes the required competency. With this framework, it is possible\nto benchmark the main families of models and to get an overview of what are the\nstrengths and the weaknesses of a given model on the set of tasks evaluated in\nthis dataset. Our corpus contains more than 500.000 questions in natural\nlanguage over 100.000 hotel reviews. Our setup is projective, the answer of a\nquestion does not need to be extracted from a document, like in most of the\nrecent datasets, but selected among a set of candidates that contains all the\npossible answers to the questions of the dataset. Finally, we present several\nbaselines over this dataset.', 'Under special circumstances, summaries should conform to a particular style\nwith patterns, such as court judgments and abstracts in academic papers. To\nthis end, the prototype document-summary pairs can be utilized to generate\nbetter summaries. There are two main challenges in this task: (1) the model\nneeds to incorporate learned patterns from the prototype, but (2) should avoid\ncopying contents other than the patternized words---such as irrelevant\nfacts---into the generated summaries. To tackle these challenges, we design a\nmodel named Prototype Editing based Summary Generator (PESG). PESG first learns\nsummary patterns and prototype facts by analyzing the correlation between a\nprototype document and its summary. Prototype facts are then utilized to help\nextract facts from the input document. Next, an editing generator generates new\nsummary based on the summary pattern or extracted facts. Finally, to address\nthe second challenge, a fact checker is used to estimate mutual information\nbetween the input document and generated summary, providing an additional\nsignal for the generator. Extensive experiments conducted on a large-scale\nreal-world text summarization dataset show that PESG achieves the\nstate-of-the-art performance in terms of both automatic metrics and human\nevaluations.', 'Conformal prediction is a method of producing prediction sets that can be\napplied on top of a wide range of prediction algorithms. The method has a\nguaranteed coverage probability under the standard IID assumption regardless of\nwhether the assumptions (often considerably more restrictive) of the underlying\nalgorithm are satisfied. However, for the method to be really useful it is\ndesirable that in the case where the assumptions of the underlying algorithm\nare satisfied, the conformal predictor loses little in efficiency as compared\nwith the underlying algorithm (whereas being a conformal predictor, it has the\nstronger guarantee of validity). In this paper we explore the degree to which\nthis additional requirement of efficiency is satisfied in the case of Bayesian\nridge regression; we find that asymptotically conformal prediction sets differ\nlittle from ridge regression prediction intervals when the standard Bayesian\nassumptions are satisfied.', 'In the context of investigative journalism, we address the problem of\nautomatically identifying which claims in a given document are most worthy and\nshould be prioritized for fact-checking. Despite its importance, this is a\nrelatively understudied problem. Thus, we create a new dataset of political\ndebates, containing statements that have been fact-checked by nine reputable\nsources, and we train machine learning models to predict which claims should be\nprioritized for fact-checking, i.e., we model the problem as a ranking task.\nUnlike previous work, which has looked primarily at sentences in isolation, in\nthis paper we focus on a rich input representation modeling the context:\nrelationship between the target statement and the larger context of the debate,\ninteraction between the opponents, and reaction by the moderator and by the\npublic. Our experiments show state-of-the-art results, outperforming a strong\nrivaling system by a margin, while also confirming the importance of the\ncontextual information.', 'Text classification is an important and classical problem in natural language\nprocessing. There have been a number of studies that applied convolutional\nneural networks (convolution on regular grid, e.g., sequence) to\nclassification. However, only a limited number of studies have explored the\nmore flexible graph convolutional neural networks (convolution on non-grid,\ne.g., arbitrary graph) for the task. In this work, we propose to use graph\nconvolutional networks for text classification. We build a single text graph\nfor a corpus based on word co-occurrence and document word relations, then\nlearn a Text Graph Convolutional Network (Text GCN) for the corpus. Our Text\nGCN is initialized with one-hot representation for word and document, it then\njointly learns the embeddings for both words and documents, as supervised by\nthe known class labels for documents. Our experimental results on multiple\nbenchmark datasets demonstrate that a vanilla Text GCN without any external\nword embeddings or knowledge outperforms state-of-the-art methods for text\nclassification. On the other hand, Text GCN also learns predictive word and\ndocument embeddings. In addition, experimental results show that the\nimprovement of Text GCN over state-of-the-art comparison methods become more\nprominent as we lower the percentage of training data, suggesting the\nrobustness of Text GCN to less training data in text classification.', 'We explore using multilingual document embeddings for nearest neighbor mining\nof parallel data. Three document-level representations are investigated: (i)\ndocument embeddings generated by simply averaging multilingual sentence\nembeddings; (ii) a neural bag-of-words (BoW) document encoding model; (iii) a\nhierarchical multilingual document encoder (HiDE) that builds on our\nsentence-level model. The results show document embeddings derived from\nsentence-level averaging are surprisingly effective for clean datasets, but\nsuggest models trained hierarchically at the document-level are more effective\non noisy data. Analysis experiments demonstrate our hierarchical models are\nvery robust to variations in the underlying sentence embedding quality. Using\ndocument embeddings trained with HiDE achieves state-of-the-art performance on\nUnited Nations (UN) parallel document mining, 94.9% P@1 for en-fr and 97.3% P@1\nfor en-es.', 'With the increase of information, document classification as one of the\nmethods of text mining, plays vital role in many management and organizing\ninformation. Document classification is the process of assigning a document to\none or more predefined category labels. Document classification includes\ndifferent parts such as text processing, term selection, term weighting and\nfinal classification. The accuracy of document classification is very\nimportant. Thus improvement in each part of classification should lead to\nbetter results and higher precision. Term weighting has a great impact on the\naccuracy of the classification. Most of the existing weighting methods exploit\nthe statistical information of terms in documents and do not consider semantic\nrelations between words. In this paper, an automated document classification\nsystem is presented that uses a novel term weighting method based on semantic\nrelations between terms. To evaluate the proposed method, three standard\nPersian corpuses are used. Experiment results show 2 to 4 percent improvement\nin classification accuracy compared with the best previous designed system for\nPersian documents.', 'The property of conformal predictors to guarantee the required accuracy rate\nmakes this framework attractive in various practical applications. However,\nthis property is achieved at a price of reduction in precision. In the case of\nconformal classification, the systems can output multiple class labels instead\nof one. It is also known from the literature, that the choice of nonconformity\nfunction has a major impact on the efficiency of conformal classifiers.\nRecently, it was shown that different model-agnostic nonconformity functions\nresult in conformal classifiers with different characteristics. For a Neural\nNetwork-based conformal classifier, the inverse probability (or hinge loss)\nallows minimizing the average number of predicted labels, and margin results in\na larger fraction of singleton predictions. In this work, we aim to further\nextend this study. We perform an experimental evaluation using 8 different\nclassification algorithms and discuss when the previously observed relationship\nholds or not. Additionally, we propose a successful method to combine the\nproperties of these two nonconformity functions. The experimental evaluation is\ndone using 11 real and 5 synthetic datasets.', 'This paper summarizes our entries to both subtasks of the first DialDoc\nshared task which focuses on the agent response prediction task in\ngoal-oriented document-grounded dialogs. The task is split into two subtasks:\npredicting a span in a document that grounds an agent turn and generating an\nagent response based on a dialog and grounding document. In the first subtask,\nwe restrict the set of valid spans to the ones defined in the dataset, use a\nbiaffine classifier to model spans, and finally use an ensemble of different\nmodels. For the second subtask, we use a cascaded model which grounds the\nresponse prediction on the predicted span instead of the full document. With\nthese approaches, we obtain significant improvements in both subtasks compared\nto the baseline.', 'In the Query Focused Multi-Document Summarization (QF-MDS) task, a set of\ndocuments and a query are given where the goal is to generate a summary from\nthese documents based on the given query. However, one major challenge for this\ntask is the lack of availability of labeled training datasets. To overcome this\nissue, in this paper, we propose a novel weakly supervised learning approach\nvia utilizing distant supervision. In particular, we use datasets similar to\nthe target dataset as the training data where we leverage pre-trained sentence\nsimilarity models to generate the weak reference summary of each individual\ndocument in a document set from the multi-document gold reference summaries.\nThen, we iteratively train our summarization model on each single-document to\nalleviate the computational complexity issue that occurs while training neural\nsummarization models in multiple documents (i.e., long sequences) at once.\nExperimental results in Document Understanding Conferences (DUC) datasets show\nthat our proposed approach sets a new state-of-the-art result in terms of\nvarious evaluation metrics.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Most existing examples of full conformal predictive systems, split-conformal\npredictive systems, and cross-conformal predictive systems impose severe\nrestrictions on the adaptation of predictive distributions to the test object\nat hand. In this paper we develop split-conformal and cross-conformal\npredictive systems that are fully adaptive. Our method consists in calibrating\nexisting predictive systems; the input predictive system is not supposed to\nsatisfy any properties of validity, whereas the output predictive system is\nguaranteed to be calibrated in probability. It is interesting that the method\nmay also work without the IID assumption, standard in conformal prediction.', ""Neural conversation models are known to generate appropriate but\nnon-informative responses in general. A scenario where informativeness can be\nsignificantly enhanced is Conversing by Reading (CbR), where conversations take\nplace with respect to a given external document. In previous work, the external\ndocument is utilized by (1) creating a context-aware document memory that\nintegrates information from the document and the conversational context, and\nthen (2) generating responses referring to the memory. In this paper, we\npropose to create the document memory with some anticipated responses in mind.\nThis is achieved using a teacher-student framework. The teacher is given the\nexternal document, the context, and the ground-truth response, and learns how\nto build a response-aware document memory from three sources of information.\nThe student learns to construct a response-anticipated document memory from the\nfirst two sources, and the teacher's insight on memory creation. Empirical\nresults show that our model outperforms the previous state-of-the-art for the\nCbR task."", 'Conformal predictors are set predictors that are automatically valid in the\nsense of having coverage probability equal to or exceeding a given confidence\nlevel. Inductive conformal predictors are a computationally efficient version\nof conformal predictors satisfying the same property of validity. However,\ninductive conformal predictors have been only known to control unconditional\ncoverage probability. This paper explores various versions of conditional\nvalidity and various ways to achieve them using inductive conformal predictors\nand their modifications.', 'When one observes a sequence of variables $(x_1, y_1), \\ldots, (x_n, y_n)$,\nConformal Prediction (CP) is a methodology that allows to estimate a confidence\nset for $y_{n+1}$ given $x_{n+1}$ by merely assuming that the distribution of\nthe data is exchangeable. CP sets have guaranteed coverage for any finite\npopulation size $n$. While appealing, the computation of such a set turns out\nto be infeasible in general, e.g. when the unknown variable $y_{n+1}$ is\ncontinuous. The bottleneck is that it is based on a procedure that readjusts a\nprediction model on data where we replace the unknown target by all its\npossible values in order to select the most probable one. This requires\ncomputing an infinite number of models, which often makes it intractable. In\nthis paper, we combine CP techniques with classical algorithmic stability\nbounds to derive a prediction set computable with a single model fit. We\ndemonstrate that our proposed confidence set does not lose any coverage\nguarantees while avoiding the need for data splitting as currently done in the\nliterature. We provide some numerical experiments to illustrate the tightness\nof our estimation when the sample size is sufficiently large, on both synthetic\nand real datasets.', 'An important challenge facing modern machine learning is how to rigorously\nquantify the uncertainty of model predictions. Conveying uncertainty is\nespecially important when there are changes to the underlying data distribution\nthat might invalidate the predictive model. Yet, most existing uncertainty\nquantification algorithms break down in the presence of such shifts. We propose\na novel approach that addresses this challenge by constructing \\emph{probably\napproximately correct (PAC)} prediction sets in the presence of covariate\nshift. Our approach focuses on the setting where there is a covariate shift\nfrom the source distribution (where we have labeled training examples) to the\ntarget distribution (for which we want to quantify uncertainty). Our algorithm\nassumes given importance weights that encode how the probabilities of the\ntraining examples change under the covariate shift. In practice, importance\nweights typically need to be estimated; thus, we extend our algorithm to the\nsetting where we are given confidence intervals for the importance weights. We\ndemonstrate the effectiveness of our approach on covariate shifts based on\nDomainNet and ImageNet. Our algorithm satisfies the PAC constraint, and gives\nprediction sets with the smallest average normalized size among approaches that\nalways satisfy the PAC constraint.', 'Convolutional neural networks (CNN) are the dominant deep neural network\n(DNN) architecture for computer vision. Recently, Transformer and multi-layer\nperceptron (MLP)-based models, such as Vision Transformer and MLP-Mixer,\nstarted to lead new trends as they showed promising results in the ImageNet\nclassification task. In this paper, we conduct empirical studies on these DNN\nstructures and try to understand their respective pros and cons. To ensure a\nfair comparison, we first develop a unified framework called SPACH which adopts\nseparate modules for spatial and channel processing. Our experiments under the\nSPACH framework reveal that all structures can achieve competitive performance\nat a moderate scale. However, they demonstrate distinctive behaviors when the\nnetwork size scales up. Based on our findings, we propose two hybrid models\nusing convolution and Transformer modules. The resulting Hybrid-MS-S+ model\nachieves 83.9% top-1 accuracy with 63M parameters and 12.3G FLOPS. It is\nalready on par with the SOTA models with sophisticated designs. The code and\nmodels are publicly available at https://github.com/microsoft/SPACH.', 'Document categorization, which aims to assign a topic label to each document,\nplays a fundamental role in a wide variety of applications. Despite the success\nof existing studies in conventional supervised document classification, they\nare less concerned with two real problems: (1) the presence of metadata: in\nmany domains, text is accompanied by various additional information such as\nauthors and tags. Such metadata serve as compelling topic indicators and should\nbe leveraged into the categorization framework; (2) label scarcity: labeled\ntraining samples are expensive to obtain in some cases, where categorization\nneeds to be performed using only a small set of annotated data. In recognition\nof these two challenges, we propose MetaCat, a minimally supervised framework\nto categorize text with metadata. Specifically, we develop a generative process\ndescribing the relationships between words, documents, labels, and metadata.\nGuided by the generative model, we embed text and metadata into the same\nsemantic space to encode heterogeneous signals. Then, based on the same\ngenerative process, we synthesize training samples to address the bottleneck of\nlabel scarcity. We conduct a thorough evaluation on a wide range of datasets.\nExperimental results prove the effectiveness of MetaCat over many competitive\nbaselines.', 'Data representation is a fundamental task in machine learning. The\nrepresentation of data affects the performance of the whole machine learning\nsystem. In a long history, the representation of data is done by feature\nengineering, and researchers aim at designing better features for specific\ntasks. Recently, the rapid development of deep learning and representation\nlearning has brought new inspiration to various domains.\n  In natural language processing, the most widely used feature representation\nis the Bag-of-Words model. This model has the data sparsity problem and cannot\nkeep the word order information. Other features such as part-of-speech tagging\nor more complex syntax features can only fit for specific tasks in most cases.\nThis thesis focuses on word representation and document representation. We\ncompare the existing systems and present our new model.\n  First, for generating word embeddings, we make comprehensive comparisons\namong existing word embedding models. In terms of theory, we figure out the\nrelationship between the two most important models, i.e., Skip-gram and GloVe.\nIn our experiments, we analyze three key points in generating word embeddings,\nincluding the model construction, the training corpus and parameter design. We\nevaluate word embeddings with three types of tasks, and we argue that they\ncover the existing use of word embeddings. Through theory and practical\nexperiments, we present some guidelines for how to generate a good word\nembedding.\n  Second, in Chinese character or word representation. We introduce the joint\ntraining of Chinese character and word. ...\n  Third, for document representation, we analyze the existing document\nrepresentation models, including recursive NNs, recurrent NNs and convolutional\nNNs. We point out the drawbacks of these models and present our new model, the\nrecurrent convolutional neural networks. ...', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'Predicting the quality of a text document is a critical task when presented\nwith the problem of measuring the performance of a document before its release.\nIn this work, we evaluate various features including those extracted from the\ntext content (textual) and those describing higher-level characteristics of the\ntext (meta) features that are not directly available from the text, and show\nhow these features inform prediction of document quality in different ways.\nMoreover, we also compare our methods on both social user-generated data such\nas tweets, and scholarly user-generated data such as academic articles, showing\nhow the same features differently influence prediction of quality across these\ndisparate domains.', 'This paper applies conformal prediction techniques to compute simultaneous\nprediction bands and clustering trees for functional data. These tools can be\nused to detect outliers and clusters. Both our prediction bands and clustering\ntrees provide prediction sets for the underlying stochastic process with a\nguaranteed finite sample behavior, under no distributional assumptions. The\nprediction sets are also informative in that they correspond to the high\ndensity region of the underlying process. While ordinary conformal prediction\nhas high computational cost for functional data, we use the inductive conformal\npredictor, together with several novel choices of conformity scores, to\nsimplify the computation. Our methods are illustrated on some real data\nexamples.', 'Document grounded generation is the task of using the information provided in\na document to improve text generation. This work focuses on two different\ndocument grounded generation tasks: Wikipedia Update Generation task and\nDialogue response generation. Our work introduces two novel adaptations of\nlarge scale pre-trained encoder-decoder models focusing on building context\ndriven representation of the document and enabling specific attention to the\ninformation in the document. Additionally, we provide a stronger BART baseline\nfor these tasks. Our proposed techniques outperform existing methods on both\nautomated (at least 48% increase in BLEU-4 points) and human evaluation for\ncloseness to reference and relevance to the document. Furthermore, we perform\ncomprehensive manual inspection of the generated output and categorize errors\nto provide insights into future directions in modeling these tasks.', ""In real-world settings involving consequential decision-making, the\ndeployment of machine learning systems generally requires both reliable\nuncertainty quantification and protection of individuals' privacy. We present a\nframework that treats these two desiderata jointly. Our framework is based on\nconformal prediction, a methodology that augments predictive models to return\nprediction sets that provide uncertainty quantification -- they provably cover\nthe true response with a user-specified probability, such as 90%. One might\nhope that when used with privately-trained models, conformal prediction would\nyield privacy guarantees for the resulting prediction sets; unfortunately this\nis not the case. To remedy this key problem, we develop a method that takes any\npre-trained predictive model and outputs differentially private prediction\nsets. Our method follows the general approach of split conformal prediction; we\nuse holdout data to calibrate the size of the prediction sets but preserve\nprivacy by using a privatized quantile subroutine. This subroutine compensates\nfor the noise introduced to preserve privacy in order to guarantee correct\ncoverage. We evaluate the method on large-scale computer vision datasets."", ""When tracking user-specific online activities, each user's preference is\nrevealed in the form of choices and comparisons. For example, a user's purchase\nhistory is a record of her choices, i.e. which item was chosen among a subset\nof offerings. A user's preferences can be observed either explicitly as in\nmovie ratings or implicitly as in viewing times of news articles. Given such\nindividualized ordinal data in the form of comparisons and choices, we address\nthe problem of collaboratively learning representations of the users and the\nitems. The learned features can be used to predict a user's preference of an\nunseen item to be used in recommendation systems. This also allows one to\ncompute similarities among users and items to be used for categorization and\nsearch. Motivated by the empirical successes of the MultiNomial Logit (MNL)\nmodel in marketing and transportation, and also more recent successes in word\nembedding and crowdsourced image embedding, we pose this problem as learning\nthe MNL model parameters that best explain the data. We propose a convex\nrelaxation for learning the MNL model, and show that it is minimax optimal up\nto a logarithmic factor by comparing its performance to a fundamental lower\nbound. This characterizes the minimax sample complexity of the problem, and\nproves that the proposed estimator cannot be improved upon other than by a\nlogarithmic factor. Further, the analysis identifies how the accuracy depends\non the topology of sampling via the spectrum of the sampling graph. This\nprovides a guideline for designing surveys when one can choose which items are\nto be compared. This is accompanied by numerical simulations on synthetic and\nreal data sets, confirming our theoretical predictions."", 'Recent progress in pretrained Transformer-based language models has shown\ngreat success in learning contextual representation of text. However, due to\nthe quadratic self-attention complexity, most of the pretrained Transformers\nmodels can only handle relatively short text. It is still a challenge when it\ncomes to modeling very long documents. In this work, we propose to use a graph\nattention network on top of the available pretrained Transformers model to\nlearn document embeddings. This graph attention network allows us to leverage\nthe high-level semantic structure of the document. In addition, based on our\ngraph document model, we design a simple contrastive learning strategy to\npretrain our models on a large amount of unlabeled corpus. Empirically, we\ndemonstrate the effectiveness of our approaches in document classification and\ndocument retrieval tasks.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', 'Recent advances in attention-based networks have shown that Vision\nTransformers can achieve state-of-the-art or near state-of-the-art results on\nmany image classification tasks. This puts transformers in the unique position\nof being a promising alternative to traditional convolutional neural networks\n(CNNs). While CNNs have been carefully studied with respect to adversarial\nattacks, the same cannot be said of Vision Transformers. In this paper, we\nstudy the robustness of Vision Transformers to adversarial examples. Our\nanalyses of transformer security is divided into three parts. First, we test\nthe transformer under standard white-box and black-box attacks. Second, we\nstudy the transferability of adversarial examples between CNNs and\ntransformers. We show that adversarial examples do not readily transfer between\nCNNs and transformers. Based on this finding, we analyze the security of a\nsimple ensemble defense of CNNs and transformers. By creating a new attack, the\nself-attention blended gradient attack, we show that such an ensemble is not\nsecure under a white-box adversary. However, under a black-box adversary, we\nshow that an ensemble can achieve unprecedented robustness without sacrificing\nclean accuracy. Our analysis for this work is done using six types of white-box\nattacks and two types of black-box attacks. Our study encompasses multiple\nVision Transformers, Big Transfer Models and CNN architectures trained on\nCIFAR-10, CIFAR-100 and ImageNet.', 'We develop a novel approach for confidently accelerating inference in the\nlarge and expensive multilayer Transformers that are now ubiquitous in natural\nlanguage processing (NLP). Amortized or approximate computational methods\nincrease efficiency, but can come with unpredictable performance costs. In this\nwork, we present CATs -- Confident Adaptive Transformers -- in which we\nsimultaneously increase computational efficiency, while guaranteeing a\nspecifiable degree of consistency with the original model with high confidence.\nOur method trains additional prediction heads on top of intermediate layers,\nand dynamically decides when to stop allocating computational effort to each\ninput using a meta consistency classifier. To calibrate our early prediction\nstopping rule, we formulate a unique extension of conformal prediction. We\ndemonstrate the effectiveness of this approach on four classification and\nregression tasks.', 'Quantifying uncertainty in model predictions is a common goal for\npractitioners seeking more than just point predictions. One tool for\nuncertainty quantification that requires minimal assumptions is conformal\ninference, which can help create probabilistically valid prediction regions for\nblack box models. Classical conformal prediction only provides marginal\nvalidity, whereas in many situations locally valid prediction regions are\ndesirable. Deciding how best to partition the feature space X when applying\nlocalized conformal prediction is still an open question. We present MD-split+,\na practical local conformal approach that creates X partitions based on\nlocalized model performance of conditional density estimation models. Our\nmethod handles complex real-world data settings where such models may be\nmisspecified, and scales to high-dimensional inputs. We discuss how our local\npartitions philosophically align with expected behavior from an unattainable\nconditional conformal inference approach. We also empirically compare our\nmethod against other local conformal approaches.', 'Relating entities and events in text is a key component of natural language\nunderstanding. Cross-document coreference resolution, in particular, is\nimportant for the growing interest in multi-document analysis tasks. In this\nwork we propose a new model that extends the efficient sequential prediction\nparadigm for coreference resolution to cross-document settings and achieves\ncompetitive results for both entity and event coreference while provides strong\nevidence of the efficacy of both sequential models and higher-order inference\nin cross-document settings. Our model incrementally composes mentions into\ncluster representations and predicts links between a mention and the already\nconstructed clusters, approximating a higher-order model. In addition, we\nconduct extensive ablation studies that provide new insights into the\nimportance of various inputs and representation types in coreference.', 'Syntactic annotation of corpora in the form of part-of-speech (POS) tags is a\nkey requirement for both linguistic research and subsequent automated natural\nlanguage processing (NLP) tasks. This problem is commonly tackled using machine\nlearning methods, i.e., by training a POS tagger on a sufficiently large corpus\nof labeled data. While the problem of POS tagging can essentially be considered\nas solved for modern languages, historical corpora turn out to be much more\ndifficult, especially due to the lack of native speakers and sparsity of\ntraining data. Moreover, most texts have no sentences as we know them today,\nnor a common orthography. These irregularities render the task of automated POS\ntagging more difficult and error-prone. Under these circumstances, instead of\nforcing the POS tagger to predict and commit to a single tag, it should be\nenabled to express its uncertainty. In this paper, we consider POS tagging\nwithin the framework of set-valued prediction, which allows the POS tagger to\nexpress its uncertainty via predicting a set of candidate POS tags instead of\nguessing a single one. The goal is to guarantee a high confidence that the\ncorrect POS tag is included while keeping the number of candidates small. In\nour experimental study, we find that extending state-of-the-art POS taggers to\nset-valued prediction yields more precise and robust taggings, especially for\nunknown words, i.e., words not occurring in the training data.', 'We study distribution free, nonparametric prediction bands with a special\nfocus on their finite sample behavior. First we investigate and develop\ndifferent notions of finite sample coverage guarantees. Then we give a new\nprediction band estimator by combining the idea of ""conformal prediction"" (Vovk\net al. 2009) with nonparametric conditional density estimation. The proposed\nestimator, called COPS (Conformal Optimized Prediction Set), always has finite\nsample guarantee in a stronger sense than the original conformal prediction\nestimator. Under regularity conditions the estimator converges to an oracle\nband at a minimax optimal rate. A fast approximation algorithm and a data\ndriven method for selecting the bandwidth are developed. The method is\nillustrated first in simulated data. Then, an application shows that the\nproposed method gives desirable prediction intervals in an automatic way, as\ncompared to the classical linear regression modeling.', ""In recent developments in the field of Computer Vision, a rise is seen in the\nuse of transformer-based architectures. They are surpassing the\nstate-of-the-art set by CNN architectures in accuracy but on the other hand,\nthey are computationally very expensive to train from scratch. As these models\nare quite recent in the Computer Vision field, there is a need to study it's\ntransfer learning capabilities and compare it with CNNs so that we can\nunderstand which architecture is better when applied to real world problems\nwith small data. In this work, we follow a simple yet restrictive method for\nfine-tuning both CNN and Transformer models pretrained on ImageNet1K on\nCIFAR-10 and compare them with each other. We only unfreeze the last\ntransformer/encoder or last convolutional block of a model and freeze all the\nlayers before it while adding a simple MLP at the end for classification. This\nsimple modification lets us use the raw learned weights of both these neural\nnetworks. From our experiments, we find out that transformers-based\narchitectures not only achieve higher accuracy than CNNs but some transformers\neven achieve this feat with around 4 times lesser number of parameters."", 'The increasing amount of online content motivated the development of\nmulti-document summarization methods. In this work, we explore straightforward\napproaches to extend single-document summarization methods to multi-document\nsummarization. The proposed methods are based on the hierarchical combination\nof single-document summaries, and achieves state of the art results.', 'Measuring similarity between texts is an important task for several\napplications. Available approaches to measure document similarity are\ninadequate for document pairs that have non-comparable lengths, such as a long\ndocument and its summary. This is because of the lexical, contextual and the\nabstraction gaps between a long document of rich details and its concise\nsummary of abstract information. In this paper, we present a document matching\napproach to bridge this gap, by comparing the texts in a common space of hidden\ntopics. We evaluate the matching algorithm on two matching tasks and find that\nit consistently and widely outperforms strong baselines. We also highlight the\nbenefits of incorporating domain knowledge to text matching.', 'The vast volume of data generated by modern astronomical surveys offers test\nbeds for the application of machine-learning. It is important to evaluate\npotential existing tools and determine those that are optimal for extracting\nscientific knowledge from the available observations. We explore the\npossibility of using clustering algorithms to separate stellar populations with\ndistinct chemical patterns. Star clusters are likely the most chemically\nhomogeneous populations in the Galaxy, and therefore any practical approach to\nidentifying distinct stellar populations should at least be able to separate\nclusters from each other. We applied eight clustering algorithms combined with\nfour dimensionality reduction strategies to automatically distinguish stellar\nclusters using chemical abundances of 13 elements. Our sample includes 18\nstellar clusters with a total of 453 stars. We use statistical tests showing\nthat some pairs of clusters are indistinguishable from each other when chemical\nabundances from the Apache Point Galactic Evolution Experiment (APOGEE) are\nused. However, for most clusters we are able to automatically assign membership\nwith metric scores similar to previous works. The confusion level of the\nautomatically selected clusters is consistent with statistical tests that\ndemonstrate the impossibility of perfectly distinguishing all the clusters from\neach other. These statistical tests and confusion levels establish a limit for\nthe prospect of blindly identifying stars born in the same cluster based solely\non chemical abundances. We find that some of the algorithms we explored are\ncapable of blindly identify stellar populations with similar ages and chemical\ndistributions in the APOGEE data. Because some stellar clusters are chemically\nindistinguishable, our study supports the notion of extending weak chemical\ntagging that involves families of clusters instead of individual clusters', 'We present a contrasting learning approach with data augmentation techniques\nto learn document representations in an unsupervised manner. Inspired by recent\ncontrastive self-supervised learning algorithms used for image and NLP\npretraining, we hypothesize that high-quality document embedding should be\ninvariant to diverse paraphrases that preserve the semantics of the original\ndocument. With different backbones and contrastive learning frameworks, our\nstudy reveals the enormous benefits of contrastive augmentation for document\nrepresentation learning with two additional insights: 1) including data\naugmentation in a contrastive way can substantially improve the embedding\nquality in unsupervised document representation learning, and 2) in general,\nstochastic augmentations generated by simple word-level manipulation work much\nbetter than sentence-level and document-level ones. We plug our method into a\nclassifier and compare it with a broad range of baseline methods on six\nbenchmark datasets. Our method can decrease the classification error rate by up\nto 6.4% over the SOTA approaches on the document classification task, matching\nor even surpassing fully-supervised methods.', 'Conformal prediction constructs a confidence set for an unobserved response\nof a feature vector based on previous identically distributed and exchangeable\nobservations of responses and features. It has a coverage guarantee at any\nnominal level without additional assumptions on their distribution. Its\ncomputation deplorably requires a refitting procedure for all replacement\ncandidates of the target response. In regression settings, this corresponds to\nan infinite number of model fits. Apart from relatively simple estimators that\ncan be written as pieces of linear function of the response, efficiently\ncomputing such sets is difficult, and is still considered as an open problem.\nWe exploit the fact that, \\emph{often}, conformal prediction sets are intervals\nwhose boundaries can be efficiently approximated by classical root-finding\nalgorithms. We investigate how this approach can overcome many limitations of\nformerly used strategies; we discuss its complexity and drawbacks.', 'Conformal predictive systems are a recent modification of conformal\npredictors that output, in regression problems, probability distributions for\nlabels of test observations rather than set predictions. The extra information\nprovided by conformal predictive systems may be useful, e.g., in decision\nmaking problems. Conformal predictive systems inherit the relative\ncomputational inefficiency of conformal predictors. In this paper we discuss\ntwo computationally efficient versions of conformal predictive systems, which\nwe call split conformal predictive systems and cross-conformal predictive\nsystems. The main advantage of split conformal predictive systems is their\nguaranteed validity, whereas for cross-conformal predictive systems validity\nonly holds empirically and in the absence of excessive randomization. The main\nadvantage of cross-conformal predictive systems is their greater predictive\nefficiency.', 'We investigate the pertinence of methods from algebraic topology for text\ndata analysis. These methods enable the development of\nmathematically-principled isometric-invariant mappings from a set of vectors to\na document embedding, which is stable with respect to the geometry of the\ndocument in the selected metric space. In this work, we evaluate the utility of\nthese topology-based document representations in traditional NLP tasks,\nspecifically document clustering and sentiment classification. We find that the\nembeddings do not benefit text analysis. In fact, performance is worse than\nsimple techniques like $\\textit{tf-idf}$, indicating that the geometry of the\ndocument does not provide enough variability for classification on the basis of\ntopic or sentiment in the chosen datasets.', 'This note introduces the method of cross-conformal prediction, which is a\nhybrid of the methods of inductive conformal prediction and cross-validation,\nand studies its validity and predictive efficiency empirically.', 'This paper proposes a new method of probabilistic prediction, which is based\non conformal prediction. The method is applied to the standard USPS data set\nand gives encouraging results.', 'Text Mining is a field that aims at extracting information from textual data.\nOne of the challenges of such field of study comes from the pre-processing\nstage in which a vector (and structured) representation should be extracted\nfrom unstructured data. The common extraction creates large and sparse vectors\nrepresenting the importance of each term to a document. As such, this usually\nleads to the curse-of-dimensionality that plagues most machine learning\nalgorithms. To cope with this issue, in this paper we propose a new supervised\nfeature extraction and reduction algorithm, named DCDistance, that creates\nfeatures based on the distance between a document to a representative of each\nclass label. As such, the proposed technique can reduce the features set in\nmore than 99% of the original set. Additionally, this algorithm was also\ncapable of improving the classification accuracy over a set of benchmark\ndatasets when compared to traditional and state-of-the-art features selection\nalgorithms.', 'Document AI, or Document Intelligence, is a relatively new research topic\nthat refers to the techniques for automatically reading, understanding, and\nanalyzing business documents. It is an important research direction for natural\nlanguage processing and computer vision. In recent years, the popularity of\ndeep learning technology has greatly advanced the development of Document AI,\nsuch as document layout analysis, visual information extraction, document\nvisual question answering, document image classification, etc. This paper\nbriefly reviews some of the representative models, tasks, and benchmark\ndatasets. Furthermore, we also introduce early-stage heuristic rule-based\ndocument analysis, statistical machine learning algorithms, and deep learning\napproaches especially pre-training methods. Finally, we look into future\ndirections for Document AI research.', 'Recent studies emphasize the need of document context in human evaluation of\nmachine translations, but little research has been done on the impact of user\ninterfaces on annotator productivity and the reliability of assessments. In\nthis work, we compare human assessment data from the last two WMT evaluation\ncampaigns collected via two different methods for document-level evaluation.\nOur analysis shows that a document-centric approach to evaluation where the\nannotator is presented with the entire document context on a screen leads to\nhigher quality segment and document level assessments. It improves the\ncorrelation between segment and document scores and increases inter-annotator\nagreement for document scores but is considerably more time consuming for\nannotators.', ""Neural networks with deep architectures have demonstrated significant\nperformance improvements in computer vision, speech recognition, and natural\nlanguage processing. The challenges in information retrieval (IR), however, are\ndifferent from these other application areas. A common form of IR involves\nranking of documents--or short passages--in response to keyword-based queries.\nEffective IR systems must deal with query-document vocabulary mismatch problem,\nby modeling relationships between different query and document terms and how\nthey indicate relevance. Models should also consider lexical matches when the\nquery contains rare terms--such as a person's name or a product model\nnumber--not seen during training, and to avoid retrieving semantically related\nbut irrelevant results. In many real-life IR tasks, the retrieval involves\nextremely large collections--such as the document index of a commercial Web\nsearch engine--containing billions of documents. Efficient IR methods should\ntake advantage of specialized IR data structures, such as inverted index, to\nefficiently retrieve from large collections. Given an information need, the IR\nsystem also mediates how much exposure an information artifact receives by\ndeciding whether it should be displayed, and where it should be positioned,\namong other results. Exposure-aware IR systems may optimize for additional\nobjectives, besides relevance, such as parity of exposure for retrieved items\nand content publishers. In this thesis, we present novel neural architectures\nand methods motivated by the specific needs and challenges of IR tasks."", ""This note discusses a simple modification of cross-conformal prediction\ninspired by recent work on e-values. The precursor of conformal prediction\ndeveloped in the 1990s by Gammerman, Vapnik, and Vovk was also based on\ne-values and is called conformal e-prediction in this note. Replacing e-values\nby p-values led to conformal prediction, which has important advantages over\nconformal e-prediction without obvious disadvantages. The situation with\ncross-conformal prediction is, however, different: whereas for cross-conformal\nprediction validity is only an empirical fact (and can be broken with excessive\nrandomization), this note draws the reader's attention to the obvious fact that\ncross-conformal e-prediction enjoys a guaranteed property of validity."", 'We address the problem of solving mixed random linear equations. We have\nunlabeled observations coming from multiple linear regressions, and each\nobservation corresponds to exactly one of the regression models. The goal is to\nlearn the linear regressors from the observations. Classically, Alternating\nMinimization (AM) (which is a variant of Expectation Maximization (EM)) is used\nto solve this problem. AM iteratively alternates between the estimation of\nlabels and solving the regression problems with the estimated labels.\nEmpirically, it is observed that, for a large variety of non-convex problems\nincluding mixed linear regression, AM converges at a much faster rate compared\nto gradient based algorithms. However, the existing theory suggests similar\nrate of convergence for AM and gradient based methods, failing to capture this\nempirical behavior. In this paper, we close this gap between theory and\npractice for the special case of a mixture of $2$ linear regressions. We show\nthat, provided initialized properly, AM enjoys a \\emph{super-linear} rate of\nconvergence in certain parameter regimes. To the best of our knowledge, this is\nthe first work that theoretically establishes such rate for AM. Hence, if we\nwant to recover the unknown regressors upto an error (in $\\ell_2$ norm) of\n$\\epsilon$, AM only takes $\\mathcal{O}(\\log \\log (1/\\epsilon))$ iterations.\nFurthermore, we compare AM with a gradient based heuristic algorithm\nempirically and show that AM dominates in iteration complexity as well as\nwall-clock time.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am interested in understanding how modern computer programs can emulate human causal logic. To achieve this, I would like to learn about the latest computational models that can predict human judgments and compare their performance with that of humans. Additionally, I am seeking references where human participants are presented with causal structures generated by Bayesian-related models, along with their explanations or feedback based on these model-generated outputs. I am curious about the most common differences between human decision-making and machine-based decisions, as well as the history of the development of models for causal reasoning, particularly how these models draw inspiration from the neural structures of living creatures.","[""Research in Cognitive Science suggests that humans understand and represent\nknowledge of the world through causal relationships. In addition to\nobservations, they can rely on experimenting and counterfactual reasoning --\ni.e. referring to an alternative course of events -- to identify causal\nrelations and explain atypical situations. Different instances of control\nsystems, such as smart homes, would benefit from having a similar causal model,\nas it would help the user understand the logic of the system and better react\nwhen needed. However, while data-driven methods achieve high levels of\ncorrelation detection, they mainly fall short of finding causal relations,\nnotably being limited to observations only. Notably, they struggle to identify\nthe cause from the effect when detecting a correlation between two variables.\nThis paper introduces a new way to learn causal models from a mixture of\nexperiments on the environment and observational data. The core of our method\nis the use of selected interventions, especially our learning takes into\naccount the variables where it is impossible to intervene, unlike other\napproaches. The causal model we obtain is then used to generate Causal Bayesian\nNetworks, which can be later used to perform diagnostic and predictive\ninference. We use our method on a smart home simulation, a use case where\nknowing causal relations pave the way towards explainable systems. Our\nalgorithm succeeds in generating a Causal Bayesian Network close to the\nsimulation's ground truth causal interactions, showing encouraging prospects\nfor application in real-life systems."", 'A major challenge in research involving artificial intelligence (AI) is the\ndevelopment of algorithms that can find solutions to problems that can\ngeneralize to different environments and tasks. Unlike AI, humans are adept at\nfinding solutions that can transfer. We hypothesize this is because their\nsolutions are informed by causal models. We propose to use human-guided causal\nknowledge to help robots find solutions that can generalize to a new\nenvironment. We develop and test the feasibility of a language interface that\nna\\""ive participants can use to communicate these causal models to a planner.\nWe find preliminary evidence that participants are able to use our interface\nand generate causal models that achieve near-generalization. We outline an\nexperiment aimed at testing far-generalization using our interface and describe\nour longer terms goals for these causal models.', 'Discriminating between causality and correlation is a major problem in\nmachine learning, and theoretical tools for determining causality are still\nbeing developed. However, people commonly make causality judgments and are\noften correct, even in unfamiliar domains. What are humans doing to make these\njudgments? This paper examines differences in human experts\' and non-experts\'\nability to attribute causality by comparing their performances to those of\nmachine-learning algorithms. We collected human judgments by using Amazon\nMechanical Turk (MTurk) and then divided the human subjects into two groups:\nexperts and non-experts. We also prepared expert and non-expert machine\nalgorithms based on different training of convolutional neural network (CNN)\nmodels. The results showed that human experts\' judgments were similar to those\nmade by an ""expert"" CNN model trained on a large number of examples from the\ntarget domain. The human non-experts\' judgments resembled the prediction\noutputs of the CNN model that was trained on only the small number of examples\nused during the MTurk instruction. We also analyzed the differences between the\nexpert and non-expert machine algorithms based on their neural representations\nto evaluate the performances, providing insight into the human experts\' and\nnon-experts\' cognitive abilities.', 'Causal Models are increasingly suggested as a means to reason about the\nbehavior of cyber-physical systems in socio-technical contexts. They allow us\nto analyze courses of events and reason about possible alternatives. Until now,\nhowever, such reasoning is confined to the technical domain and limited to\nsingle systems or at most groups of systems. The humans that are an integral\npart of any such socio-technical system are usually ignored or dealt with by\n""expert judgment"". We show how a technical causal model can be extended with\nmodels of human behavior to cover the complexity and interplay between humans\nand technical systems. This integrated socio-technical causal model can then be\nused to reason not only about actions and decisions taken by the machine, but\nalso about those taken by humans interacting with the system. In this paper we\ndemonstrate the feasibility of merging causal models about machines with causal\nmodels about humans and illustrate the usefulness of this approach with a\nhighly automated vehicle example.', 'Humans demonstrate remarkable abilities to predict physical events in complex\nscenes. Two classes of models for physical scene understanding have recently\nbeen proposed: ""Intuitive Physics Engines"", or IPEs, which posit that people\nmake predictions by running approximate probabilistic simulations in causal\nmental models similar in nature to video-game physics engines, and memory-based\nmodels, which make judgments based on analogies to stored experiences of\npreviously encountered scenes and physical outcomes. Versions of the latter\nhave recently been instantiated in convolutional neural network (CNN)\narchitectures. Here we report four experiments that, to our knowledge, are the\nfirst rigorous comparisons of simulation-based and CNN-based models, where both\napproaches are concretely instantiated in algorithms that can run on raw image\ninputs and produce as outputs physical judgments such as whether a stack of\nblocks will fall. Both approaches can achieve super-human accuracy levels and\ncan quantitatively predict human judgments to a similar degree, but only the\nsimulation-based models generalize to novel situations in ways that people do,\nand are qualitatively consistent with systematic perceptual illusions and\njudgment asymmetries that people show.', 'There is a brief description of the probabilistic causal graph model for\nrepresenting, reasoning with, and learning causal structure using Bayesian\nnetworks. It is then argued that this model is closely related to how humans\nreason with and learn causal structure. It is shown that studies in psychology\non discounting (reasoning concerning how the presence of one cause of an effect\nmakes another cause less probable) support the hypothesis that humans reach the\nsame judgments as algorithms for doing inference in Bayesian networks. Next, it\nis shown how studies by Piaget indicate that humans learn causal structure by\nobserving the same independencies and dependencies as those used by certain\nalgorithms for learning the structure of a Bayesian network. Based on this\nindication, a subjective definition of causality is forwarded. Finally, methods\nfor further testing the accuracy of these claims are discussed.', ""Prevalent theories in cognitive science propose that humans understand and\nrepresent the knowledge of the world through causal relationships. In making\nsense of the world, we build causal models in our mind to encode cause-effect\nrelations of events and use these to explain why new events happen. In this\npaper, we use causal models to derive causal explanations of behaviour of\nreinforcement learning agents. We present an approach that learns a structural\ncausal model during reinforcement learning and encodes causal relationships\nbetween variables of interest. This model is then used to generate explanations\nof behaviour based on counterfactual analysis of the causal model. We report on\na study with 120 participants who observe agents playing a real-time strategy\ngame (Starcraft II) and then receive explanations of the agents' behaviour. We\ninvestigated: 1) participants' understanding gained by explanations through\ntask prediction; 2) explanation satisfaction and 3) trust. Our results show\nthat causal model explanations perform better on these measures compared to two\nother baseline explanation models."", ""Probabilistic machine learning models are often insufficient to help with\ndecisions on interventions because those models find correlations - not causal\nrelationships. If observational data is only available and experimentation are\ninfeasible, the correct approach to study the impact of an intervention is to\ninvoke Pearl's causality framework. Even that framework assumes that the\nunderlying causal graph is known, which is seldom the case in practice. When\nthe causal structure is not known, one may use out-of-the-box algorithms to\nfind causal dependencies from observational data. However, there exists no\nmethod that also accounts for the decision-maker's prior knowledge when\ndeveloping the causal structure either. The objective of this paper is to\ndevelop rational approaches for making decisions from observational data in the\npresence of causal graph uncertainty and prior knowledge from the\ndecision-maker. We use ensemble methods like Bayesian Model Averaging (BMA) to\ninfer set of causal graphs that can represent the data generation process. We\nprovide decisions by computing the expected value and risk of potential\ninterventions explicitly. We demonstrate our approach by applying them in\ndifferent example contexts."", ""Causal reasoning is the main learning and explanation tool used by humans. AI\nsystems should possess causal reasoning capabilities to be deployed in the real\nworld with trust and reliability. Introducing the ideas of causality to machine\nlearning helps in providing better learning and explainable models.\nExplainability, causal disentanglement are some important aspects of any\nmachine learning model. Causal explanations are required to believe in a\nmodel's decision and causal disentanglement learning is important for transfer\nlearning applications. We exploit the ideas of causality to be used in deep\nlearning models to achieve better and causally explainable models that are\nuseful in fairness, disentangled representation, etc."", 'State-of-the-art recommender systems have the ability to generate\nhigh-quality recommendations, but usually cannot provide intuitive explanations\nto humans due to the usage of black-box prediction models. The lack of\ntransparency has highlighted the critical importance of improving the\nexplainability of recommender systems. In this paper, we propose to extract\ncausal rules from the user interaction history as post-hoc explanations for the\nblack-box sequential recommendation mechanisms, whilst maintain the predictive\naccuracy of the recommendation model. Our approach firstly achieves\ncounterfactual examples with the aid of a perturbation model, and then extracts\npersonalized causal relationships for the recommendation model through a causal\nrule mining algorithm. Experiments are conducted on several state-of-the-art\nsequential recommendation models and real-world datasets to verify the\nperformance of our model on generating causal explanations. Meanwhile, We\nevaluate the discovered causal explanations in terms of quality and fidelity,\nwhich show that compared with conventional association rules, causal rules can\nprovide personalized and more effective explanations for the behavior of\nblack-box recommendation models.', ""Those designing autonomous systems that interact with humans will invariably\nface questions about how humans think and make decisions. Fortunately,\ncomputational cognitive science offers insight into human decision-making using\ntools that will be familiar to those with backgrounds in optimization and\ncontrol (e.g., probability theory, statistical machine learning, and\nreinforcement learning). Here, we review some of this work, focusing on how\ncognitive science can provide forward models of human decision-making and\ninverse models of how humans think about others' decision-making. We highlight\nrelevant recent developments, including approaches that synthesize blackbox and\ntheory-driven modeling, accounts that recast heuristics and biases as forms of\nbounded optimality, and models that characterize human theory of mind and\ncommunication in decision-theoretic terms. In doing so, we aim to provide\nreaders with a glimpse of the range of frameworks, methodologies, and\nactionable insights that lie at the intersection of cognitive science and\ncontrol research."", 'Complex machine learning models are deployed in several critical domains\nincluding healthcare and autonomous vehicles nowadays, albeit as functional\nblack boxes. Consequently, there has been a recent surge in interpreting\ndecisions of such complex models in order to explain their actions to humans.\nModels that correspond to human interpretation of a task are more desirable in\ncertain contexts and can help attribute liability, build trust, expose biases\nand in turn build better models. It is, therefore, crucial to understand how\nand which models conform to human understanding of tasks. In this paper, we\npresent a large-scale crowdsourcing study that reveals and quantifies the\ndissonance between human and machine understanding, through the lens of an\nimage classification task. In particular, we seek to answer the following\nquestions: Which (well-performing) complex ML models are closer to humans in\ntheir use of features to make accurate predictions? How does task difficulty\naffect the feature selection capability of machines in comparison to humans?\nAre humans consistently better at selecting features that make image\nrecognition more accurate? Our findings have important implications on\nhuman-machine collaboration, considering that a long term goal in the field of\nartificial intelligence is to make machines capable of learning and reasoning\nlike humans.']","['Decision-making under uncertainty and causal thinking are fundamental aspects\nof intelligent reasoning. Decision-making has been well studied when the\navailable information is considered at the associative (probabilistic) level.\nThe classical Theorems of von Neumann-Morgenstern and Savage provide a formal\ncriterion for rational choice using associative information: maximize expected\nutility. There is an ongoing debate around the origin of probabilities involved\nin such calculation. In this work, we will show how the probabilities for\ndecision-making can be grounded in causal models by considering decision\nproblems in which the available actions and consequences are causally\nconnected. In this setting, actions are regarded as an intervention over a\ncausal model. Then, we extend a previous causal decision-making result, which\nrelies on a known causal model, to the case in which the causal mechanism that\ncontrols some environment is unknown to a rational decision-maker. In this way,\naction-outcome probabilities can be grounded in causal models in known and\nunknown cases. Finally, as an application, we extend the well-known concept of\nNash Equilibrium to the case in which the players of a strategic game consider\ncausal information.', 'Machine learning has made major advances in categorizing objects in images,\nyet the best algorithms miss important aspects of how people learn and think\nabout categories. People can learn richer concepts from fewer examples,\nincluding causal models that explain how members of a category are formed.\nHere, we explore the limits of this human ability to infer causal ""programs"" --\nlatent generating processes with nontrivial algorithmic properties -- from one,\ntwo, or three visual examples. People were asked to extrapolate the programs in\nseveral ways, for both classifying and generating new examples. As a theory of\nthese inductive abilities, we present a Bayesian program learning model that\nsearches the space of programs for the best explanation of the observations.\nAlthough variable, people\'s judgments are broadly consistent with the model and\ninconsistent with several alternatives, including a pre-trained deep neural\nnetwork for object recognition, indicating that people can learn and reason\nwith rich algorithmic abstractions from sparse input data.', 'Discovering and exploiting the causal structure in the environment is a\ncrucial challenge for intelligent agents. Here we explore whether causal\nreasoning can emerge via meta-reinforcement learning. We train a recurrent\nnetwork with model-free reinforcement learning to solve a range of problems\nthat each contain causal structure. We find that the trained agent can perform\ncausal reasoning in novel situations in order to obtain rewards. The agent can\nselect informative interventions, draw causal inferences from observational\ndata, and make counterfactual predictions. Although established formal causal\nreasoning algorithms also exist, in this paper we show that such reasoning can\narise from model-free reinforcement learning, and suggest that causal reasoning\nin complex settings may benefit from the more end-to-end learning-based\napproaches presented here. This work also offers new strategies for structured\nexploration in reinforcement learning, by providing agents with the ability to\nperform -- and interpret -- experiments.', 'The ability to reason about temporal and causal events from videos lies at\nthe core of human intelligence. Most video reasoning benchmarks, however, focus\non pattern recognition from complex visual and language input, instead of on\ncausal structure. We study the complementary problem, exploring the temporal\nand causal structures behind videos of objects with simple visual appearance.\nTo this end, we introduce the CoLlision Events for Video REpresentation and\nReasoning (CLEVRER), a diagnostic video dataset for systematic evaluation of\ncomputational models on a wide range of reasoning tasks. Motivated by the\ntheory of human casual judgment, CLEVRER includes four types of questions:\ndescriptive (e.g., ""what color""), explanatory (""what is responsible for""),\npredictive (""what will happen next""), and counterfactual (""what if""). We\nevaluate various state-of-the-art models for visual reasoning on our benchmark.\nWhile these models thrive on the perception-based task (descriptive), they\nperform poorly on the causal tasks (explanatory, predictive and\ncounterfactual), suggesting that a principled approach for causal reasoning\nshould incorporate the capability of both perceiving complex visual and\nlanguage inputs, and understanding the underlying dynamics and causal\nrelations. We also study an oracle model that explicitly combines these\ncomponents via symbolic representations.', 'The study and understanding of human behaviour is relevant to computer\nscience, artificial intelligence, neural computation, cognitive science,\nphilosophy, psychology, and several other areas. Presupposing cognition as\nbasis of behaviour, among the most prominent tools in the modelling of\nbehaviour are computational-logic systems, connectionist models of cognition,\nand models of uncertainty. Recent studies in cognitive science, artificial\nintelligence, and psychology have produced a number of cognitive models of\nreasoning, learning, and language that are underpinned by computation. In\naddition, efforts in computer science research have led to the development of\ncognitive computational systems integrating machine learning and automated\nreasoning. Such systems have shown promise in a range of applications,\nincluding computational biology, fault diagnosis, training and assessment in\nsimulators, and software verification. This joint survey reviews the personal\nideas and views of several researchers on neural-symbolic learning and\nreasoning. The article is organised in three parts: Firstly, we frame the scope\nand goals of neural-symbolic computation and have a look at the theoretical\nfoundations. We then proceed to describe the realisations of neural-symbolic\ncomputation, systems, and applications. Finally we present the challenges\nfacing the area and avenues for further research.', ""Recent work in explanation generation for decision making agents has looked\nat how unexplained behavior of autonomous systems can be understood in terms of\ndifferences in the model of the system and the human's understanding of the\nsame, and how the explanation process as a result of this mismatch can be then\nseen as a process of reconciliation of these models. Existing algorithms in\nsuch settings, while having been built on contrastive, selective and social\nproperties of explanations as studied extensively in the psychology literature,\nhave not, to the best of our knowledge, been evaluated in settings with actual\nhumans in the loop. As such, the applicability of such explanations to human-AI\nand human-robot interactions remains suspect. In this paper, we set out to\nevaluate these explanation generation algorithms in a series of studies in a\nmock search and rescue scenario with an internal semi-autonomous robot and an\nexternal human commander. We demonstrate to what extent the properties of these\nalgorithms hold as they are evaluated by humans, and how the dynamics of trust\nbetween the human and the robot evolve during the process of these\ninteractions."", 'The recent success of Bayesian methods in neuroscience and artificial\nintelligence gives rise to the hypothesis that the brain is a Bayesian machine.\nSince logic and learning are both practices of the human brain, it leads to\nanother hypothesis that there is a Bayesian interpretation underlying both\nlogical reasoning and machine learning. In this paper, we introduce a\ngenerative model of logical consequence relations. It formalises the process of\nhow the truth value of a sentence is probabilistically generated from the\nprobability distribution over states of the world. We show that the generative\nmodel characterises a classical consequence relation, paraconsistent\nconsequence relation and nonmonotonic consequence relation. In particular, the\ngenerative model gives a new consequence relation that outperforms them in\nreasoning with inconsistent knowledge. We also show that the generative model\ngives a new classification algorithm that outperforms several representative\nalgorithms in predictive accuracy and complexity on the Kaggle Titanic dataset.', 'Humans are able to perceive, understand and reason about causal events.\nDeveloping models with similar physical and causal understanding capabilities\nis a long-standing goal of artificial intelligence. As a step towards this\ndirection, we introduce CRAFT, a new video question answering dataset that\nrequires causal reasoning about physical forces and object interactions. It\ncontains 58K video and question pairs that are generated from 10K videos from\n20 different virtual environments, containing various objects in motion that\ninteract with each other and the scene. Two question categories in CRAFT\ninclude previously studied descriptive and counterfactual questions.\nAdditionally, inspired by the Force Dynamics Theory in cognitive linguistics,\nwe introduce a new causal question category that involves understanding the\ncausal interactions between objects through notions like cause, enable, and\nprevent. Our results show that even though the questions in CRAFT are easy for\nhumans, the tested baseline models, including existing state-of-the-art\nmethods, do not yet deal with the challenges posed in our benchmark.', 'Causal induction, i.e., identifying unobservable mechanisms that lead to the\nobservable relations among variables, has played a pivotal role in modern\nscientific discovery, especially in scenarios with only sparse and limited\ndata. Humans, even young toddlers, can induce causal relationships surprisingly\nwell in various settings despite its notorious difficulty. However, in contrast\nto the commonplace trait of human cognition is the lack of a diagnostic\nbenchmark to measure causal induction for modern Artificial Intelligence (AI)\nsystems. Therefore, in this work, we introduce the Abstract Causal REasoning\n(ACRE) dataset for systematic evaluation of current vision systems in causal\ninduction. Motivated by the stream of research on causal discovery in Blicket\nexperiments, we query a visual reasoning system with the following four types\nof questions in either an independent scenario or an interventional scenario:\ndirect, indirect, screening-off, and backward-blocking, intentionally going\nbeyond the simple strategy of inducing causal relationships by covariation. By\nanalyzing visual reasoning architectures on this testbed, we notice that pure\nneural models tend towards an associative strategy under their chance-level\nperformance, whereas neuro-symbolic combinations struggle in backward-blocking\nreasoning. These deficiencies call for future research in models with a more\ncomprehensive capability of causal induction.', 'Human understanding of narrative is mainly driven by reasoning about causal\nrelations between events and thus recognizing them is a key capability for\ncomputational models of language understanding. Computational work in this area\nhas approached this via two different routes: by focusing on acquiring a\nknowledge base of common causal relations between events, or by attempting to\nunderstand a particular story or macro-event, along with its storyline. In this\nposition paper, we focus on knowledge acquisition approach and claim that\nnewswire is a relatively poor source for learning fine-grained causal relations\nbetween everyday events. We describe experiments using an unsupervised method\nto learn causal relations between events in the narrative genres of\nfirst-person narratives and film scene descriptions. We show that our method\nlearns fine-grained causal relations, judged by humans as likely to be causal\nover 80% of the time. We also demonstrate that the learned event pairs do not\nexist in publicly available event-pair datasets extracted from newswire.', 'We propose to explain the behavior of black-box prediction methods (e.g.,\ndeep neural networks trained on image pixel data) using causal graphical\nmodels. Specifically, we explore learning the structure of a causal graph where\nthe nodes represent prediction outcomes along with a set of macro-level\n""interpretable"" features, while allowing for arbitrary unmeasured confounding\namong these variables. The resulting graph may indicate which of the\ninterpretable features, if any, are possible causes of the prediction outcome\nand which may be merely associated with prediction outcomes due to confounding.\nThe approach is motivated by a counterfactual theory of causal explanation\nwherein good explanations point to factors that are ""difference-makers"" in an\ninterventionist sense. The resulting analysis may be useful in algorithm\nauditing and evaluation, by identifying features which make a causal difference\nto the algorithm\'s output.', 'We present a randomized controlled trial for a model-in-the-loop regression\ntask, with the goal of measuring the extent to which (1) good explanations of\nmodel predictions increase human accuracy, and (2) faulty explanations decrease\nhuman trust in the model. We study explanations based on visual saliency in an\nimage-based age prediction task for which humans and learned models are\nindividually capable but not highly proficient and frequently disagree. Our\nexperimental design separates model quality from explanation quality, and makes\nit possible to compare treatments involving a variety of explanations of\nvarying levels of quality. We find that presenting model predictions improves\nhuman accuracy. However, visual explanations of various kinds fail to\nsignificantly alter human accuracy or trust in the model - regardless of\nwhether explanations characterize an accurate model, an inaccurate one, or are\ngenerated randomly and independently of the input image. These findings suggest\nthe need for greater evaluation of explanations in downstream decision making\ntasks, better design-based tools for presenting explanations to users, and\nbetter approaches for generating explanations.', 'Recent advances in computational cognitive science (i.e., simulation-based\nprobabilistic programs) have paved the way for significant progress in formal,\nimplementable models of pragmatics. Rather than describing a pragmatic\nreasoning process in prose, these models formalize and implement one, deriving\nboth qualitative and quantitative predictions of human behavior -- predictions\nthat consistently prove correct, demonstrating the viability and value of the\nframework. The current paper provides a practical introduction to and critical\nassessment of the Bayesian Rational Speech Act modeling framework, unpacking\ntheoretical foundations, exploring technological innovations, and drawing\nconnections to issues beyond current applications.', 'Research in social psychology has shown that people\'s biased, subjective\njudgments about another\'s personality based solely on their appearance are not\npredictive of their actual personality traits. But researchers and companies\noften utilize computer vision models to predict similarly subjective\npersonality attributes such as ""employability."" We seek to determine whether\nstate-of-the-art, black box face processing technology can learn human-like\nappearance biases. With features extracted with FaceNet, a widely used face\nrecognition framework, we train a transfer learning model on human subjects\'\nfirst impressions of personality traits in other faces as measured by social\npsychologists. We find that features extracted with FaceNet can be used to\npredict human appearance bias scores for deliberately manipulated faces but not\nfor randomly generated faces scored by humans. Additionally, in contrast to\nwork with human biases in social psychology, the model does not find a\nsignificant signal correlating politicians\' vote shares with perceived\ncompetence bias. With Local Interpretable Model-Agnostic Explanations (LIME),\nwe provide several explanations for this discrepancy. Our results suggest that\nsome signals of appearance bias documented in social psychology are not\nembedded by the machine learning techniques we investigate. We shed light on\nthe ways in which appearance bias could be embedded in face processing\ntechnology and cast further doubt on the practice of predicting subjective\ntraits based on appearances.', 'Despite the surprising power of many modern AI systems that often learn their\nown representations, there is significant discontent about their inscrutability\nand the attendant problems in their ability to interact with humans. While\nalternatives such as neuro-symbolic approaches have been proposed, there is a\nlack of consensus on what they are about. There are often two independent\nmotivations (i) symbols as a lingua franca for human-AI interaction and (ii)\nsymbols as system-produced abstractions used by the AI system in its internal\nreasoning. The jury is still out on whether AI systems will need to use symbols\nin their internal reasoning to achieve general intelligence capabilities.\nWhatever the answer there is, the need for (human-understandable) symbols in\nhuman-AI interaction seems quite compelling. Symbols, like emotions, may well\nnot be sine qua non for intelligence per se, but they will be crucial for AI\nsystems to interact with us humans -- as we can neither turn off our emotions\nnor get by without our symbols. In particular, in many human-designed domains,\nhumans would be interested in providing explicit (symbolic) knowledge and\nadvice -- and expect machine explanations in kind. This alone requires AI\nsystems to to maintain a symbolic interface for interaction with humans. In\nthis blue sky paper, we argue this point of view, and discuss research\ndirections that need to be pursued to allow for this type of human-AI\ninteraction.', 'When AI systems interact with humans in the loop, they are often called on to\nprovide explanations for their plans and behavior. Past work on plan\nexplanations primarily involved the AI system explaining the correctness of its\nplan and the rationale for its decision in terms of its own model. Such\nsoliloquy is wholly inadequate in most realistic scenarios where the humans\nhave domain and task models that differ significantly from that used by the AI\nsystem. We posit that the explanations are best studied in light of these\ndiffering models. In particular, we show how explanation can be seen as a\n""model reconciliation problem"" (MRP), where the AI system in effect suggests\nchanges to the human\'s model, so as to make its plan be optimal with respect to\nthat changed human model. We will study the properties of such explanations,\npresent algorithms for automatically computing them, and evaluate the\nperformance of the algorithms.', ""Uncovering causal relationships in data is a major objective of data\nanalytics. Causal relationships are normally discovered with designed\nexperiments, e.g. randomised controlled trials, which, however are expensive or\ninfeasible to be conducted in many cases. Causal relationships can also be\nfound using some well designed observational studies, but they require domain\nexperts' knowledge and the process is normally time consuming. Hence there is a\nneed for scalable and automated methods for causal relationship exploration in\ndata. Classification methods are fast and they could be practical substitutes\nfor finding causal signals in data. However, classification methods are not\ndesigned for causal discovery and a classification method may find false causal\nsignals and miss the true ones. In this paper, we develop a causal decision\ntree where nodes have causal interpretations. Our method follows a well\nestablished causal inference framework and makes use of a classic statistical\ntest. The method is practical for finding causal signals in large data sets."", 'The burgeoning of AI has prompted recommendations that AI techniques should\nbe ""human-centered"". However, there is no clear definition of what is meant by\nHuman Centered Artificial Intelligence, or for short, HCAI. This paper aims to\nimprove this situation by addressing some foundational aspects of HCAI. To do\nso, we introduce the term HCAI agent to refer to any physical or software\ncomputational agent equipped with AI components and that interacts and/or\ncollaborates with humans. This article identifies five main conceptual\ncomponents that participate in an HCAI agent: Observations, Requirements,\nActions, Explanations and Models. We see the notion of HCAI agent, together\nwith its components and functions, as a way to bridge the technical and\nnon-technical discussions on human-centered AI. In this paper, we focus our\nanalysis on scenarios consisting of a single agent operating in dynamic\nenvironments in presence of humans.', 'Human decision-making underlies all economic behavior. For the past four\ndecades, human decision-making under uncertainty has continued to be explained\nby theoretical models based on prospect theory, a framework that was awarded\nthe Nobel Prize in Economic Sciences. However, theoretical models of this kind\nhave developed slowly, and robust, high-precision predictive models of human\ndecisions remain a challenge. While machine learning is a natural candidate for\nsolving these problems, it is currently unclear to what extent it can improve\npredictions obtained by current theories. We argue that this is mainly due to\ndata scarcity, since noisy human behavior requires massive sample sizes to be\naccurately captured by off-the-shelf machine learning methods. To solve this\nproblem, what is needed are machine learning models with appropriate inductive\nbiases for capturing human behavior, and larger datasets. We offer two\ncontributions towards this end: first, we construct ""cognitive model priors"" by\npretraining neural networks with synthetic data generated by cognitive models\n(i.e., theoretical models developed by cognitive psychologists). We find that\nfine-tuning these networks on small datasets of real human decisions results in\nunprecedented state-of-the-art improvements on two benchmark datasets. Second,\nwe present the first large-scale dataset for human decision-making, containing\nover 240,000 human judgments across over 13,000 decision problems. This dataset\nreveals the circumstances where cognitive model priors are useful, and provides\na new standard for benchmarking prediction of human decisions under\nuncertainty.', 'With the rise of machines to human-level performance in complex recognition\ntasks, a growing amount of work is directed towards comparing information\nprocessing in humans and machines. These studies are an exciting chance to\nlearn about one system by studying the other. Here, we propose ideas on how to\ndesign, conduct and interpret experiments such that they adequately support the\ninvestigation of mechanisms when comparing human and machine perception. We\ndemonstrate and apply these ideas through three case studies. The first case\nstudy shows how human bias can affect how we interpret results, and that\nseveral analytic tools can help to overcome this human reference point. In the\nsecond case study, we highlight the difference between necessary and sufficient\nmechanisms in visual reasoning tasks. Thereby, we show that contrary to\nprevious suggestions, feedback mechanisms might not be necessary for the tasks\nin question. The third case study highlights the importance of aligning\nexperimental conditions. We find that a previously-observed difference in\nobject recognition does not hold when adapting the experiment to make\nconditions more equitable between humans and machines. In presenting a\nchecklist for comparative studies of visual reasoning in humans and machines,\nwe hope to highlight how to overcome potential pitfalls in design or inference.', ""The problem of replicating the flexibility of human common-sense reasoning\nhas captured the imagination of computer scientists since the early days of\nAlan Turing's foundational work on computation and the philosophy of artificial\nintelligence. In the intervening years, the idea of cognition as computation\nhas emerged as a fundamental tenet of Artificial Intelligence (AI) and\ncognitive science. But what kind of computation is cognition?\n  We describe a computational formalism centered around a probabilistic Turing\nmachine called QUERY, which captures the operation of probabilistic\nconditioning via conditional simulation. Through several examples and analyses,\nwe demonstrate how the QUERY abstraction can be used to cast common-sense\nreasoning as probabilistic inference in a statistical model of our observations\nand the uncertain structure of the world that generated that experience. This\nformulation is a recent synthesis of several research programs in AI and\ncognitive science, but it also represents a surprising convergence of several\nof Turing's pioneering insights in AI, the foundations of computation, and\nstatistics."", ""As AI systems demonstrate increasingly strong predictive performance, their\nadoption has grown in numerous domains. However, in high-stakes domains such as\ncriminal justice and healthcare, full automation is often not desirable due to\nsafety, ethical, and legal concerns, yet fully manual approaches can be\ninaccurate and time consuming. As a result, there is growing interest in the\nresearch community to augment human decision making with AI assistance. Besides\ndeveloping AI technologies for this purpose, the emerging field of human-AI\ndecision making must embrace empirical approaches to form a foundational\nunderstanding of how humans interact and work with AI to make decisions. To\ninvite and help structure research efforts towards a science of understanding\nand improving human-AI decision making, we survey recent literature of\nempirical human-subject studies on this topic. We summarize the study design\nchoices made in over 100 papers in three important aspects: (1) decision tasks,\n(2) AI models and AI assistance elements, and (3) evaluation metrics. For each\naspect, we summarize current trends, discuss gaps in current practices of the\nfield, and make a list of recommendations for future research. Our survey\nhighlights the need to develop common frameworks to account for the design and\nresearch spaces of human-AI decision making, so that researchers can make\nrigorous choices in study design, and the research community can build on each\nother's work and produce generalizable scientific knowledge. We also hope this\nsurvey will serve as a bridge for HCI and AI communities to work together to\nmutually shape the empirical science and computational technologies for\nhuman-AI decision making."", ""Understanding the fundamentals of human reasoning is central to the\ndevelopment of any system built to closely interact with humans. Cognitive\nscience pursues the goal of modeling human-like intelligence from a\ntheory-driven perspective with a strong focus on explainability. Syllogistic\nreasoning as one of the core domains of human reasoning research has seen a\nsurge of computational models being developed over the last years. However,\nrecent analyses of models' predictive performances revealed a stagnation in\nimprovement. We believe that most of the problems encountered in cognitive\nscience are not due to the specific models that have been developed but can be\ntraced back to the peculiarities of behavioral data instead.\n  Therefore, we investigate potential data-related reasons for the problems in\nhuman reasoning research by comparing model performances on human and\nartificially generated datasets. In particular, we apply collaborative\nfiltering recommenders to investigate the adversarial effects of\ninconsistencies and noise in data and illustrate the potential for data-driven\nmethods in a field of research predominantly concerned with gaining high-level\ntheoretical insight into a domain.\n  Our work (i) provides insight into the levels of noise to be expected from\nhuman responses in reasoning data, (ii) uncovers evidence for an upper-bound of\nperformance that is close to being reached urging for an extension of the\nmodeling task, and (iii) introduces the tools and presents initial results to\npioneer a new paradigm for investigating and modeling reasoning focusing on\npredicting responses for individual human reasoners."", ""Human collaborators can effectively communicate with their partners to finish\na common task by inferring each other's mental states (e.g., goals, beliefs,\nand desires). Such mind-aware communication minimizes the discrepancy among\ncollaborators' mental states, and is crucial to the success in human ad-hoc\nteaming. We believe that robots collaborating with human users should\ndemonstrate similar pedagogic behavior. Thus, in this paper, we propose a novel\nexplainable AI (XAI) framework for achieving human-like communication in\nhuman-robot collaborations, where the robot builds a hierarchical mind model of\nthe human user and generates explanations of its own mind as a form of\ncommunications based on its online Bayesian inference of the user's mental\nstate. To evaluate our framework, we conduct a user study on a real-time\nhuman-robot cooking task. Experimental results show that the generated\nexplanations of our approach significantly improves the collaboration\nperformance and user perception of the robot. Code and video demos are\navailable on our project website: https://xfgao.github.io/xCookingWeb/."", 'When human cognition is modeled in Philosophy and Cognitive Science, there is\na pervasive idea that humans employ mental representations in order to navigate\nthe world and make predictions about outcomes of future actions. By\nunderstanding how these representational structures work, we not only\nunderstand more about human cognition but also gain a better understanding for\nhow humans rationalise and explain decisions. This has an influencing effect on\nexplainable AI, where the goal is to provide explanations of computer\ndecision-making for a human audience. We show that the Contextual Importance\nand Utility method for XAI share an overlap with the current new wave of\naction-oriented predictive representational structures, in ways that makes CIU\na reliable tool for creating explanations that humans can relate to and trust.', 'We discuss the changes in an attitude to decision making at the fire ground.\nThe changes are driven by the recent technological shift. The emerging new\napproaches in sensing and data processing (under common umbrella of\nCyber-Physical Systems) allow for leveling off the gap, between humans and\nmachines, in perception of the fire ground. Furthermore, results from\ndescriptive decision theory question the rationality of human choices. This\ncreates the need for searching and testing new approaches for decision making\nduring emergency. We propose the framework that addresses this need. The\nprimary feature of the framework are possibilities for incorporation of\nnormative and prescriptive approaches to decision making. The framework also\nallows for comparison of the performance of decisions, between human and\nmachine.', 'Bayesian networks can be used to extract explanations about the observed\nstate of a subset of variables. In this paper, we explicate the desiderata of\nan explanation and confront them with the concept of explanation proposed by\nexisting methods. The necessity of taking into account causal approaches when a\ncausal graph is available is discussed. We then introduce causal explanation\ntrees, based on the construction of explanation trees using the measure of\ncausal information ow (Ay and Polani, 2006). This approach is compared to\nseveral other methods on known networks.', ""Higher-level cognition depends on the ability to learn models of the world.\nWe can characterize this at the computational level as a structure-learning\nproblem with the goal of best identifying the prevailing causal relationships\namong a set of relata. However, the computational cost of performing exact\nBayesian inference over causal models grows rapidly as the number of relata\nincreases. This implies that the cognitive processes underlying causal learning\nmust be substantially approximate. A powerful class of approximations that\nfocuses on the sequential absorption of successive inputs is captured by the\nNeurath's ship metaphor in philosophy of science, where theory change is cast\nas a stochastic and gradual process shaped as much by people's limited\nwillingness to abandon their current theory when considering alternatives as by\nthe ground truth they hope to approach. Inspired by this metaphor and by\nalgorithms for approximating Bayesian inference in machine learning, we propose\nan algorithmic-level model of causal structure learning under which learners\nrepresent only a single global hypothesis that they update locally as they\ngather evidence. We propose a related scheme for understanding how, under these\nlimitations, learners choose informative interventions that manipulate the\ncausal system to help elucidate its workings. We find support for our approach\nin the analysis of four experiments."", 'A large body of work in behavioral fields attempts to develop models that\ndescribe the way people, as opposed to rational agents, make decisions. A\nrecent Choice Prediction Competition (2015) challenged researchers to suggest a\nmodel that captures 14 classic choice biases and can predict human decisions\nunder risk and ambiguity. The competition focused on simple decision problems,\nin which human subjects were asked to repeatedly choose between two gamble\noptions.\n  In this paper we present our approach for predicting human decision behavior:\nwe suggest to use machine learning algorithms with features that are based on\nwell-established behavioral theories. The basic idea is that these\npsychological features are essential for the representation of the data and are\nimportant for the success of the learning process. We implement a vanilla model\nin which we train SVM models using behavioral features that rely on the\npsychological properties underlying the competition baseline model. We show\nthat this basic model captures the 14 choice biases and outperforms all the\nother learning-based models in the competition. The preliminary results suggest\nthat such hybrid models can significantly improve the prediction of human\ndecision making, and are a promising direction for future research.', 'Causal inference is central to many areas of artificial intelligence,\nincluding complex reasoning, planning, knowledge-base construction, robotics,\nexplanation, and fairness. An active community of researchers develops and\nenhances algorithms that learn causal models from data, and this work has\nproduced a series of impressive technical advances. However, evaluation\ntechniques for causal modeling algorithms have remained somewhat primitive,\nlimiting what we can learn from experimental studies of algorithm performance,\nconstraining the types of algorithms and model representations that researchers\nconsider, and creating a gap between theory and practice. We argue for more\nfrequent use of evaluation techniques that examine interventional measures\nrather than structural or observational measures, and that evaluate those\nmeasures on empirical data rather than synthetic data. We survey the current\npractice in evaluation and show that the techniques we recommend are rarely\nused in practice. We show that such techniques are feasible and that data sets\nare available to conduct such evaluations. We also show that these techniques\nproduce substantially different results than using structural measures and\nsynthetic data.', 'Recent progress in artificial intelligence (AI) has renewed interest in\nbuilding systems that learn and think like people. Many advances have come from\nusing deep neural networks trained end-to-end in tasks such as object\nrecognition, video games, and board games, achieving performance that equals or\neven beats humans in some respects. Despite their biological inspiration and\nperformance achievements, these systems differ from human intelligence in\ncrucial ways. We review progress in cognitive science suggesting that truly\nhuman-like learning and thinking machines will have to reach beyond current\nengineering trends in both what they learn, and how they learn it.\nSpecifically, we argue that these machines should (a) build causal models of\nthe world that support explanation and understanding, rather than merely\nsolving pattern recognition problems; (b) ground learning in intuitive theories\nof physics and psychology, to support and enrich the knowledge that is learned;\nand (c) harness compositionality and learning-to-learn to rapidly acquire and\ngeneralize knowledge to new tasks and situations. We suggest concrete\nchallenges and promising routes towards these goals that can combine the\nstrengths of recent neural network advances with more structured cognitive\nmodels.', 'The ability to learn and act in novel situations is still a prerogative of\nanimate intelligence, as current machine learning methods mostly fail when\nmoving beyond the standard i.i.d. setting. What is the reason for this\ndiscrepancy? Most machine learning tasks are anti-causal, i.e., we infer causes\n(labels) from effects (observations). Typically, in supervised learning we\nbuild systems that try to directly invert causal mechanisms. Instead, in this\npaper we argue that strong generalization capabilities crucially hinge on\nsearching and validating meaningful hypotheses, requiring access to a causal\nmodel. In such a framework, we want to find a cause that leads to the observed\neffect. Anti-causal models are used to drive this search, but a causal model is\nrequired for validation. We investigate the fundamental differences between\ncausal and anti-causal tasks, discuss implications for topics ranging from\nadversarial attacks to disentangling factors of variation, and provide\nextensive evidence from the literature to substantiate our view. We advocate\nfor incorporating causal models in supervised learning to shift the paradigm\nfrom inference only, to search and validation.', 'Human beings learn causal models and constantly use them to transfer\nknowledge between similar environments. We use this intuition to design a\ntransfer-learning framework using object-oriented representations to learn the\ncausal relationships between objects. A learned causal dynamics model can be\nused to transfer between variants of an environment with exchangeable\nperceptual features among objects but with the same underlying causal dynamics.\nWe adapt continuous optimization for structure learning techniques to\nexplicitly learn the cause and effects of the actions in an interactive\nenvironment and transfer to the target domain by categorization of the objects\nbased on causal knowledge. We demonstrate the advantages of our approach in a\ngridworld setting by combining causal model-based approach with model-free\napproach in reinforcement learning.', 'We consider decision problems under uncertainty where the options available\nto a decision maker and the resulting outcome are related through a causal\nmechanism which is unknown to the decision maker. We ask how a decision maker\ncan learn about this causal mechanism through sequential decision making as\nwell as using current causal knowledge inside each round in order to make\nbetter choices had she not considered causal knowledge and propose a decision\nmaking procedure in which an agent holds \\textit{beliefs} about her environment\nwhich are used to make a choice and are updated using the observed outcome. As\nproof of concept, we present an implementation of this causal decision making\nmodel and apply it in a simple scenario. We show that the model achieves a\nperformance similar to the classic Q-learning while it also acquires a causal\nmodel of the environment.', ""Causal reasoning has been an indispensable capability for humans and other\nintelligent animals to interact with the physical world. In this work, we\npropose to endow an artificial agent with the capability of causal reasoning\nfor completing goal-directed tasks. We develop learning-based approaches to\ninducing causal knowledge in the form of directed acyclic graphs, which can be\nused to contextualize a learned goal-conditional policy to perform tasks in\nnovel environments with latent causal structures. We leverage attention\nmechanisms in our causal induction model and goal-conditional policy, enabling\nus to incrementally generate the causal graph from the agent's visual\nobservations and to selectively use the induced graph for determining actions.\nOur experiments show that our method effectively generalizes towards completing\nnew tasks in novel environments with previously unseen causal structures."", 'Linear structural causal models (SCMs) -- in which each observed variable is\ngenerated by a subset of the other observed variables as well as a subset of\nthe exogenous sources -- are pervasive in causal inference and casual\ndiscovery. However, for the task of causal discovery, existing work almost\nexclusively focus on the submodel where each observed variable is associated\nwith a distinct source with non-zero variance. This results in the restriction\nthat no observed variable can deterministically depend on other observed\nvariables or latent confounders. In this paper, we extend the results on\nstructure learning by focusing on a subclass of linear SCMs which do not have\nthis property, i.e., models in which observed variables can be causally\naffected by any subset of the sources, and are allowed to be a deterministic\nfunction of other observed variables or latent confounders. This allows for a\nmore realistic modeling of influence or information propagation in systems. We\nfocus on the task of causal discovery form observational data generated from a\nmember of this subclass. We derive a set of necessary and sufficient conditions\nfor unique identifiability of the causal structure. To the best of our\nknowledge, this is the first work that gives identifiability results for causal\ndiscovery under both latent confounding and deterministic relationships.\nFurther, we propose an algorithm for recovering the underlying causal structure\nwhen the aforementioned conditions are satisfied. We validate our theoretical\nresults both on synthetic and real datasets.', 'Understanding predictions made by deep neural networks is notoriously\ndifficult, but also crucial to their dissemination. As all machine learning\nbased methods, they are as good as their training data, and can also capture\nunwanted biases. While there are tools that can help understand whether such\nbiases exist, they do not distinguish between correlation and causation, and\nmight be ill-suited for text-based models and for reasoning about high level\nlanguage concepts. A key problem of estimating the causal effect of a concept\nof interest on a given model is that this estimation requires the generation of\ncounterfactual examples, which is challenging with existing generation\ntechnology. To bridge that gap, we propose CausaLM, a framework for producing\ncausal model explanations using counterfactual language representation models.\nOur approach is based on fine-tuning of deep contextualized embedding models\nwith auxiliary adversarial tasks derived from the causal graph of the problem.\nConcretely, we show that by carefully choosing auxiliary adversarial\npre-training tasks, language representation models such as BERT can effectively\nlearn a counterfactual representation for a given concept of interest, and be\nused to estimate its true causal effect on model performance. A byproduct of\nour method is a language representation model that is unaffected by the tested\nconcept, which can be useful in mitigating unwanted bias ingrained in the data.', 'Intelligence can be defined as a predominantly human ability to accomplish\ntasks that are generally hard for computers and animals. Artificial\nIntelligence [AI] is a field attempting to accomplish such tasks with\ncomputers. AI is becoming increasingly widespread, as are claims of its\nrelationship with Biological Intelligence. Often these claims are made to imply\nhigher chances of a given technology succeeding, working on the assumption that\nAI systems which mimic the mechanisms of Biological Intelligence should be more\nsuccessful.\n  In this article I will discuss the similarities and differences between AI\nand the extent of our knowledge about the mechanisms of intelligence in\nbiology, especially within humans. I will also explore the validity of the\nassumption that biomimicry in AI systems aids their advancement, and I will\nargue that existing similarity to biological systems in the way Artificial\nNeural Networks [ANNs] tackle tasks is due to design decisions, rather than\ninherent similarity of underlying mechanisms. This article is aimed at people\nwho understand the basics of AI (especially ANNs), and would like to be better\nable to evaluate the often wild claims about the value of biomimicry in AI.', 'Human social behavior is structured by relationships. We form teams, groups,\ntribes, and alliances at all scales of human life. These structures guide\nmulti-agent cooperation and competition, but when we observe others these\nunderlying relationships are typically unobservable and hence must be inferred.\nHumans make these inferences intuitively and flexibly, often making rapid\ngeneralizations about the latent relationships that underlie behavior from just\nsparse and noisy observations. Rapid and accurate inferences are important for\ndetermining who to cooperate with, who to compete with, and how to cooperate in\norder to compete. Towards the goal of building machine-learning algorithms with\nhuman-like social intelligence, we develop a generative model of multi-agent\naction understanding based on a novel representation for these latent\nrelationships called Composable Team Hierarchies (CTH). This representation is\ngrounded in the formalism of stochastic games and multi-agent reinforcement\nlearning. We use CTH as a target for Bayesian inference yielding a new\nalgorithm for understanding behavior in groups that can both infer hidden\nrelationships as well as predict future actions for multiple agents interacting\ntogether. Our algorithm rapidly recovers an underlying causal model of how\nagents relate in spatial stochastic games from just a few observations. The\npatterns of inference made by this algorithm closely correspond with human\njudgments and the algorithm makes the same rapid generalizations that people\ndo.', 'In this paper, we aim to develop a unified view of causal and non-causal\nfeature selection methods. The unified view will fill in the gap in the\nresearch of the relation between the two types of methods. Based on the\nBayesian network framework and information theory, we first show that causal\nand non-causal feature selection methods share the same objective. That is to\nfind the Markov blanket of a class attribute, the theoretically optimal feature\nset for classification. We then examine the assumptions made by causal and\nnon-causal feature selection methods when searching for the optimal feature\nset, and unify the assumptions by mapping them to the restrictions on the\nstructure of the Bayesian network model of the studied problem. We further\nanalyze in detail how the structural assumptions lead to the different levels\nof approximations employed by the methods in their search, which then result in\nthe approximations in the feature sets found by the methods with respect to the\noptimal feature set. With the unified view, we are able to interpret the output\nof non-causal methods from a causal perspective and derive the error bounds of\nboth types of methods. Finally, we present practical understanding of the\nrelation between causal and non-causal methods using extensive experiments with\nsynthetic data and various types of real-word data.', ""We propose CX-ToM, short for counterfactual explanations with theory-of mind,\na new explainable AI (XAI) framework for explaining decisions made by a deep\nconvolutional neural network (CNN). In contrast to the current methods in XAI\nthat generate explanations as a single shot response, we pose explanation as an\niterative communication process, i.e. dialog, between the machine and human\nuser. More concretely, our CX-ToM framework generates sequence of explanations\nin a dialog by mediating the differences between the minds of machine and human\nuser. To do this, we use Theory of Mind (ToM) which helps us in explicitly\nmodeling human's intention, machine's mind as inferred by the human as well as\nhuman's mind as inferred by the machine. Moreover, most state-of-the-art XAI\nframeworks provide attention (or heat map) based explanations. In our work, we\nshow that these attention based explanations are not sufficient for increasing\nhuman trust in the underlying CNN model. In CX-ToM, we instead use\ncounterfactual explanations called fault-lines which we define as follows:\ngiven an input image I for which a CNN classification model M predicts class\nc_pred, a fault-line identifies the minimal semantic-level features (e.g.,\nstripes on zebra, pointed ears of dog), referred to as explainable concepts,\nthat need to be added to or deleted from I in order to alter the classification\ncategory of I by M to another specified class c_alt. We argue that, due to the\niterative, conceptual and counterfactual nature of CX-ToM explanations, our\nframework is practical and more natural for both expert and non-expert users to\nunderstand the internal workings of complex deep learning models. Extensive\nquantitative and qualitative experiments verify our hypotheses, demonstrating\nthat our CX-ToM significantly outperforms the state-of-the-art explainable AI\nmodels."", 'Humans make complex inferences on faces, ranging from objective properties\n(gender, ethnicity, expression, age, identity, etc) to subjective judgments\n(facial attractiveness, trustworthiness, sociability, friendliness, etc). While\nthe objective aspects of face perception have been extensively studied,\nrelatively fewer computational models have been developed for the social\nimpressions of faces. Bridging this gap, we develop a method to predict human\nimpressions of faces in 40 subjective social dimensions, using deep\nrepresentations from state-of-the-art neural networks. We find that model\nperformance grows as the human consensus on a face trait increases, and that\nmodel predictions outperform human groups in correlation with human averages.\nThis illustrates the learnability of subjective social perception of faces,\nespecially when there is high human consensus. Our system can be used to decide\nwhich photographs from a personal collection will make the best impression. The\nresults are significant for the field of social robotics, demonstrating that\nrobots can learn the subjective judgments defining the underlying fabric of\nhuman interaction.', ""Neural networks have long been used to model human intelligence, capturing\nelements of behavior and cognition, and their neural basis. Recent advancements\nin deep learning have enabled neural network models to reach and even surpass\nhuman levels of intelligence in many respects, yet unlike humans, their ability\nto learn new tasks quickly remains a challenge. People can reason not only in\nfamiliar domains, but can also rapidly learn to reason through novel problems\nand situations, raising the question of how well modern neural network models\ncapture human intelligence and in which ways they diverge. In this work, we\nexplore this gap by investigating human adults' ability to learn an abstract\nreasoning task based on Sudoku from a brief instructional tutorial with\nexplanatory feedback for incorrect responses using a narrow range of training\nexamples. We find that participants who master the task do so within a small\nnumber of trials and generalize well to puzzles outside of the training range.\nWe also find that most of those who master the task can describe a valid\nsolution strategy, and such participants perform better on transfer puzzles\nthan those whose strategy descriptions are vague or incomplete. Interestingly,\nfewer than half of our human participants were successful in acquiring a valid\nsolution strategy, and this ability is associated with high school mathematics\neducation. We consider the challenges these findings pose for building\ncomputational models that capture all aspects of our findings and point toward\na possible role for learning to engage in explanation-based reasoning to\nsupport rapid learning and generalization."", ""As environments involving both robots and humans become increasingly common,\nso does the need to account for people during planning. To plan effectively,\nrobots must be able to respond to and sometimes influence what humans do. This\nrequires a human model which predicts future human actions. A simple model may\nassume the human will continue what they did previously; a more complex one\nmight predict that the human will act optimally, disregarding the robot;\nwhereas an even more complex one might capture the robot's ability to influence\nthe human. These models make different trade-offs between computational time\nand performance of the resulting robot plan. Using only one model of the human\neither wastes computational resources or is unable to handle critical\nsituations. In this work, we give the robot access to a suite of human models\nand enable it to assess the performance-computation trade-off online. By\nestimating how an alternate model could improve human prediction and how that\nmay translate to performance gain, the robot can dynamically switch human\nmodels whenever the additional computation is justified. Our experiments in a\ndriving simulator showcase how the robot can achieve performance comparable to\nalways using the best human model, but with greatly reduced computation."", ""This document is written with the intention to describe in detail a method\nand means by which a computer program can reason about the world and in so\ndoing, increase its analogue to a living system. As the literature is rife and\nit is apparent we, as scientists and engineers, have not found the solution,\nthis document will attempt the solution by grounding its intellectual arguments\nwithin tenets of human cognition in Western philosophy. The result will be a\ncharacteristic description of a method to describe an artificial system\nanalogous to that performed for a human. The approach was the substance of my\nMaster's thesis, explored more deeply during the course of my postdoc research.\nIt focuses primarily on context awareness and choice set within a boundary of\navailable epistemology, which serves to describe it. Expanded upon, such a\ndescription strives to discover agreement with Kant's critique of reason to\nunderstand how it could be applied to define the architecture of its design.\nThe intention has never been to mimic human or biological systems, rather, to\nunderstand the profoundly fundamental rules, when leveraged correctly, results\nin an artificial consciousness as noumenon while in keeping with the perception\nof it as phenomenon."", 'This paper proposes a research direction to advance AI which draws\ninspiration from cognitive theories of human decision making. The premise is\nthat if we gain insights about the causes of some human capabilities that are\nstill lacking in AI (for instance, adaptability, generalizability, common\nsense, and causal reasoning), we may obtain similar capabilities in an AI\nsystem by embedding these causal components. We hope that the high-level\ndescription of our vision included in this paper, as well as the several\nresearch questions that we propose to consider, can stimulate the AI research\ncommunity to define, try and evaluate new methodologies, frameworks, and\nevaluation metrics, in the spirit of achieving a better understanding of both\nhuman and machine intelligence.', 'Causal Learner is a toolbox for learning causal structure and Markov blanket\n(MB) from data. It integrates functions for generating simulated Bayesian\nnetwork data, a set of state-of-the-art global causal structure learning\nalgorithms, a set of state-of-the-art local causal structure learning\nalgorithms, a set of state-of-the-art MB learning algorithms, and functions for\nevaluating algorithms. The data generation part of Causal Learner is written in\nR, and the rest of Causal Learner is written in MATLAB. Causal Learner aims to\nprovide researchers and practitioners with an open-source platform for causal\nlearning from data and for the development and evaluation of new causal\nlearning algorithms. The Causal Learner project is available at\nhttp://bigdata.ahu.edu.cn/causal-learner.', 'We consider the incorporation of causal knowledge about the presence or\nabsence of (possibly indirect) causal relations into a causal model. Such\ncausal relations correspond to directed paths in a causal model. This type of\nknowledge naturally arises from experimental data, among others. Specifically,\nwe consider the formalisms of Causal Bayesian Networks and Maximal Ancestral\nGraphs and their Markov equivalence classes: Partially Directed Acyclic Graphs\nand Partially Oriented Ancestral Graphs. We introduce sound and complete\nprocedures which are able to incorporate causal prior knowledge in such models.\nIn simulated experiments, we show that often considering even a few causal\nfacts leads to a significant number of new inferences. In a case study, we also\nshow how to use real experimental data to infer causal knowledge and\nincorporate it into a real biological causal network. The code is available at\nmensxmachina.org.', 'Deep neural networks have become increasingly successful at solving classic\nperception problems such as object recognition, semantic segmentation, and\nscene understanding, often reaching or surpassing human-level accuracy. This\nsuccess is due in part to the ability of DNNs to learn useful representations\nof high-dimensional inputs, a problem that humans must also solve. We examine\nthe relationship between the representations learned by these networks and\nhuman psychological representations recovered from similarity judgments. We\nfind that deep features learned in service of object classification account for\na significant amount of the variance in human similarity judgments for a set of\nanimal images. However, these features do not capture some qualitative\ndistinctions that are a key part of human representations. To remedy this, we\ndevelop a method for adapting deep features to align with human similarity\njudgments, resulting in image representations that can potentially be used to\nextend the scope of psychological experiments.', 'In this paper we propose a causal analog to the purely observational Dynamic\nBayesian Networks, which we call Dynamic Causal Networks. We provide a sound\nand complete algorithm for identification of Dynamic Causal Net- works, namely,\nfor computing the effect of an intervention or experiment, based on passive\nobservations only, whenever possible. We note the existence of two types of\nconfounder variables that affect in substantially different ways the iden-\ntification procedures, a distinction with no analog in either Dynamic Bayesian\nNetworks or standard causal graphs. We further propose a procedure for the\ntransportability of causal effects in Dynamic Causal Network settings, where\nthe re- sult of causal experiments in a source domain may be used for the\nidentification of causal effects in a target domain.', 'Causal discovery, the task of automatically constructing a causal model from\ndata, is of major significance across the sciences. Evaluating the performance\nof causal discovery algorithms should ideally involve comparing the inferred\nmodels to ground-truth models available for benchmark datasets, which in turn\nrequires a notion of distance between causal models. While such distances have\nbeen proposed previously, they are limited by focusing on graphical properties\nof the causal models being compared. Here, we overcome this limitation by\ndefining distances derived from the causal distributions induced by the models,\nrather than exclusively from their graphical structure. Pearl and Mackenzie\n(2018) have arranged the properties of causal models in a hierarchy called the\n""ladder of causation"" spanning three rungs: observational, interventional, and\ncounterfactual. Following this organization, we introduce a hierarchy of three\ndistances, one for each rung of the ladder. Our definitions are intuitively\nappealing as well as efficient to compute approximately. We put our causal\ndistances to use by benchmarking standard causal discovery systems on both\nsynthetic and real-world datasets for which ground-truth causal models are\navailable. Finally, we highlight the usefulness of our causal distances by\nbriefly discussing further applications beyond the evaluation of causal\ndiscovery techniques.', 'Probabilistic graphical models such as Bayesian Networks are one of the most\npowerful structures known by the Computer Science community for deriving\nprobabilistic inferences. However, modern cognitive psychology has revealed\nthat human decisions could not follow the rules of classical probability\ntheory, because humans cannot process large amounts of data in order to make\njudgements. Consequently, the inferences performed are based on limited data\ncoupled with several heuristics, leading to violations of the law of total\nprobability. This means that probabilistic graphical models based on classical\nprobability theory are too limited to fully simulate and explain various\naspects of human decision making.\n  Quantum probability theory was developed in order to accommodate the\nparadoxical findings that the classical theory could not explain. Recent\nfindings in cognitive psychology revealed that quantum probability can fully\ndescribe human decisions in an elegant framework. Their findings suggest that,\nbefore taking a decision, human thoughts are seen as superposed waves that can\ninterfere with each other, influencing the final decision.\n  In this work, we propose a new Bayesian Network based on the psychological\nfindings of cognitive scientists. We made experiments with two very well known\nBayesian Networks from the literature. The results obtained revealed that the\nquantum like Bayesian Network can affect drastically the probabilistic\ninferences, specially when the levels of uncertainty of the network are very\nhigh (no pieces of evidence observed). When the levels of uncertainty are very\nlow, then the proposed quantum like network collapses to its classical\ncounterpart.', 'While feedback loops are known to play important roles in many complex\nsystems, their existence is ignored in a large part of the causal discovery\nliterature, as systems are typically assumed to be acyclic from the outset.\nWhen applying causal discovery algorithms designed for the acyclic setting on\ndata generated by a system that involves feedback, one would not expect to\nobtain correct results. In this work, we show that---surprisingly---the output\nof the Fast Causal Inference (FCI) algorithm is correct if it is applied to\nobservational data generated by a system that involves feedback. More\nspecifically, we prove that for observational data generated by a simple and\n$\\sigma$-faithful Structural Causal Model (SCM), FCI is sound and complete, and\ncan be used to consistently estimate (i) the presence and absence of causal\nrelations, (ii) the presence and absence of direct causal relations, (iii) the\nabsence of confounders, and (iv) the absence of specific cycles in the causal\ngraph of the SCM. We extend these results to constraint-based causal discovery\nalgorithms that exploit certain forms of background knowledge, including the\ncausally sufficient setting (e.g., the PC algorithm) and the Joint Causal\nInference setting (e.g., the FCI-JCI algorithm).', ""How can humans remain in control of artificial intelligence (AI)-based\nsystems designed to perform tasks autonomously? Such systems are increasingly\nubiquitous, creating benefits - but also undesirable situations where moral\nresponsibility for their actions cannot be properly attributed to any\nparticular person or group. The concept of meaningful human control has been\nproposed to address responsibility gaps and mitigate them by establishing\nconditions that enable a proper attribution of responsibility for humans;\nhowever, clear requirements for researchers, designers, and engineers are yet\ninexistent, making the development of AI-based systems that remain under\nmeaningful human control challenging. In this paper, we address the gap between\nphilosophical theory and engineering practice by identifying, through an\niterative process of abductive thinking, four actionable properties for\nAI-based systems under meaningful human control, which we discuss making use of\ntwo applications scenarios: automated vehicles and AI-based hiring. First, a\nsystem in which humans and AI algorithms interact should have an explicitly\ndefined domain of morally loaded situations within which the system ought to\noperate. Second, humans and AI agents within the system should have appropriate\nand mutually compatible representations. Third, responsibility attributed to a\nhuman should be commensurate with that human's ability and authority to control\nthe system. Fourth, there should be explicit links between the actions of the\nAI agents and actions of humans who are aware of their moral responsibility. We\nargue that these four properties will support practically-minded professionals\nto take concrete steps toward designing and engineering for AI systems that\nfacilitate meaningful human control."", 'Counterfactual inference is a powerful tool, capable of solving challenging\nproblems in high-profile sectors. To perform counterfactual inference, one\nrequires knowledge of the underlying causal mechanisms. However, causal\nmechanisms cannot be uniquely determined from observations and interventions\nalone. This raises the question of how to choose the causal mechanisms so that\nresulting counterfactual inference is trustworthy in a given domain. This\nquestion has been addressed in causal models with binary variables, but the\ncase of categorical variables remains unanswered. We address this challenge by\nintroducing for causal models with categorical variables the notion of\ncounterfactual ordering, a principle that posits desirable properties causal\nmechanisms should posses, and prove that it is equivalent to specific\nfunctional constraints on the causal mechanisms. To learn causal mechanisms\nsatisfying these constraints, and perform counterfactual inference with them,\nwe introduce deep twin networks. These are deep neural networks that, when\ntrained, are capable of twin network counterfactual inference -- an alternative\nto the abduction, action, & prediction method. We empirically test our approach\non diverse real-world and semi-synthetic data from medicine, epidemiology, and\nfinance, reporting accurate estimation of counterfactual probabilities while\ndemonstrating the issues that arise with counterfactual reasoning when\ncounterfactual ordering is not enforced.', 'Humans learn complex latent structures from their environments (e.g., natural\nlanguage, mathematics, music, social hierarchies). In cognitive science and\ncognitive neuroscience, models that infer higher-order structures from sensory\nor first-order representations have been proposed to account for the complexity\nand flexibility of human behavior. But how do the structures that these models\ninvoke arise in neural systems in the first place? To answer this question, we\nexplain how a system can learn latent representational structures (i.e.,\npredicates) from experience with wholly unstructured data. During the process\nof predicate learning, an artificial neural network exploits the naturally\noccurring dynamic properties of distributed computing across neuronal\nassemblies in order to learn predicates, but also to combine them\ncompositionally, two computational aspects which appear to be necessary for\nhuman behavior as per formal theories in multiple domains. We describe how\npredicates can be combined generatively using neural oscillations to achieve\nhuman-like extrapolation and compositionality in an artificial neural network.\nThe ability to learn predicates from experience, to represent structures\ncompositionally, and to extrapolate to unseen data offers an inroads to\nunderstanding and modeling the most complex human behaviors.', 'This paper describes a Bayesian method for combining an arbitrary mixture of\nobservational and experimental data in order to learn causal Bayesian networks.\nObservational data are passively observed. Experimental data, such as that\nproduced by randomized controlled trials, result from the experimenter\nmanipulating one or more variables (typically randomly) and observing the\nstates of other variables. The paper presents a Bayesian method for learning\nthe causal structure and parameters of the underlying causal process that is\ngenerating the data, given that (1) the data contains a mixture of\nobservational and experimental case records, and (2) the causal process is\nmodeled as a causal Bayesian network. This learning method was applied using as\ninput various mixtures of experimental and observational data that were\ngenerated from the ALARM causal Bayesian network. In these experiments, the\nabsolute and relative quantities of experimental and observational data were\nvaried systematically. For each of these training datasets, the learning method\nwas applied to predict the causal structure and to estimate the causal\nparameters that exist among randomly selected pairs of nodes in ALARM that are\nnot confounded. The paper reports how these structure predictions and parameter\nestimates compare with the true causal structures and parameters as given by\nthe ALARM network.', 'Experiential AI is proposed as a new research agenda in which artists and\nscientists come together to dispel the mystery of algorithms and make their\nmechanisms vividly apparent. It addresses the challenge of finding novel ways\nof opening up the field of artificial intelligence to greater transparency and\ncollaboration between human and machine. The hypothesis is that art can mediate\nbetween computer code and human comprehension to overcome the limitations of\nexplanations in and for AI systems. Artists can make the boundaries of systems\nvisible and offer novel ways to make the reasoning of AI transparent and\ndecipherable. Beyond this, artistic practice can explore new configurations of\nhumans and algorithms, mapping the terrain of inter-agencies between people and\nmachines. This helps to viscerally understand the complex causal chains in\nenvironments with AI components, including questions about what data to collect\nor who to collect it about, how the algorithms are chosen, commissioned and\nconfigured or how humans are conditioned by their participation in algorithmic\nprocesses.', 'In attempts to produce ML models less reliant on spurious patterns in NLP\ndatasets, researchers have recently proposed curating counterfactually\naugmented data (CAD) via a human-in-the-loop process in which given some\ndocuments and their (initial) labels, humans must revise the text to make a\ncounterfactual label applicable. Importantly, edits that are not necessary to\nflip the applicable label are prohibited. Models trained on the augmented data\nappear, empirically, to rely less on semantically irrelevant words and to\ngeneralize better out of domain. While this work draws loosely on causal\nthinking, the underlying causal model (even at an abstract level) and the\nprinciples underlying the observed out-of-domain improvements remain unclear.\nIn this paper, we introduce a toy analog based on linear Gaussian models,\nobserving interesting relationships between causal models, measurement noise,\nout-of-domain generalization, and reliance on spurious signals. Our analysis\nprovides some insights that help to explain the efficacy of CAD. Moreover, we\ndevelop the hypothesis that while adding noise to causal features should\ndegrade both in-domain and out-of-domain performance, adding noise to\nnon-causal features should lead to relative improvements in out-of-domain\nperformance. This idea inspires a speculative test for determining whether a\nfeature attribution technique has identified the causal spans. If adding noise\n(e.g., by random word flips) to the highlighted spans degrades both in-domain\nand out-of-domain performance on a battery of challenge datasets, but adding\nnoise to the complement gives improvements out-of-domain, it suggests we have\nidentified causal spans. We present a large-scale empirical study comparing\nspans edited to create CAD to those selected by attention and saliency maps.\nAcross numerous domains and models, we find that the hypothesized phenomenon is\npronounced for CAD.', ""This paper presents a sound and completecalculus for causal relevance, based\nonPearl's functional models semantics.The calculus consists of axioms and\nrulesof inference for reasoning about causalrelevance relationships.We extend\nthe set of known axioms for causalrelevance with three new axioms, andintroduce\ntwo new rules of inference forreasoning about specific subclasses\nofmodels.These subclasses give a more refinedcharacterization of causal models\nthan the one given in Halpern's axiomatizationof counterfactual\nreasoning.Finally, we show how the calculus for causalrelevance can be used in\nthe task ofidentifying causal structure from non-observational data."", 'Probability trees are one of the simplest models of causal generative\nprocesses. They possess clean semantics and -- unlike causal Bayesian networks\n-- they can represent context-specific causal dependencies, which are necessary\nfor e.g. causal induction. Yet, they have received little attention from the AI\nand ML community. Here we present concrete algorithms for causal reasoning in\ndiscrete probability trees that cover the entire causal hierarchy (association,\nintervention, and counterfactuals), and operate on arbitrary propositional and\ncausal events. Our work expands the domain of causal reasoning to a very\ngeneral class of discrete stochastic processes.', 'In this work we propose a multi-valued extension of logic programs under the\nstable models semantics where each true atom in a model is associated with a\nset of justifications. These justifications are expressed in terms of causal\ngraphs formed by rule labels and edges that represent their application\nordering. For positive programs, we show that the causal justifications\nobtained for a given atom have a direct correspon- dence to (relevant)\nsyntactic proofs of that atom using the program rules involved in the graphs.\nThe most interesting contribution is that this causal information is obtained\nin a purely semantic way, by algebraic op- erations (product, sum and\napplication) on a lattice of causal values whose ordering relation expresses\nwhen a justification is stronger than another. Finally, for programs with\nnegation, we define the concept of causal stable model by introducing an\nanalogous transformation to Gelfond and Lifschitz\'s program reduct. As a\nresult, default negation behaves as ""absence of proof"" and no justification is\nderived from negative liter', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', 'Automation of tasks can have critical consequences when humans lose agency\nover decision processes. Deep learning models are particularly susceptible\nsince current black-box approaches lack explainable reasoning. We argue that\nboth the visual interface and model structure of deep learning systems need to\ntake into account interaction design. We propose a framework of collaborative\nsemantic inference (CSI) for the co-design of interactions and models to enable\nvisual collaboration between humans and algorithms. The approach exposes the\nintermediate reasoning process of models which allows semantic interactions\nwith the visual metaphors of a problem, which means that a user can both\nunderstand and control parts of the model reasoning process. We demonstrate the\nfeasibility of CSI with a co-designed case study of a document summarization\nsystem.', 'A Random Graph is a random object which take its values in the space of\ngraphs. We take advantage of the expressibility of graphs in order to model the\nuncertainty about the existence of causal relationships within a given set of\nvariables. We adopt a Bayesian point of view in order to capture a causal\nstructure via interaction and learning with a causal environment. We test our\nmethod over two different scenarios, and the experiments mainly confirm that\nour technique can learn a causal structure. Furthermore, the experiments and\nresults presented for the first test scenario demonstrate the usefulness of our\nmethod to learn a causal structure as well as the optimal action. On the other\nhand the second experiment, shows that our proposal manages to learn the\nunderlying causal structure of several tasks with different sizes and different\ncausal structures.', 'This paper discusses causal independence models and a generalization of these\nmodels called causal interaction models. Causal interaction models are models\nthat have independent mechanisms where a mechanism can have several causes. In\naddition to introducing several particular types of causal interaction models,\nwe show how we can apply the Bayesian approach to learning causal interaction\nmodels obtaining approximate posterior distributions for the models and obtain\nMAP and ML estimates for the parameters. We illustrate the approach with a\nsimulation study of learning model posteriors.', ""Standard automatic metrics, e.g. BLEU, are not reliable for document-level MT\nevaluation. They can neither distinguish document-level improvements in\ntranslation quality from sentence-level ones, nor identify the discourse\nphenomena that cause context-agnostic translations. This paper introduces a\nnovel automatic metric BlonDe to widen the scope of automatic MT evaluation\nfrom sentence to document level. BlonDe takes discourse coherence into\nconsideration by categorizing discourse-related spans and calculating the\nsimilarity-based F1 measure of categorized spans. We conduct extensive\ncomparisons on a newly constructed dataset BWB. The experimental results show\nthat BlonDe possesses better selectivity and interpretability at the\ndocument-level, and is more sensitive to document-level nuances. In a\nlarge-scale human study, BlonDe also achieves significantly higher Pearson's r\ncorrelation with human judgments compared to previous metrics."", 'This paper proposes a causal inference relation and causal programming as\ngeneral frameworks for causal inference with structural causal models. A tuple,\n$\\langle M, I, Q, F \\rangle$, is an instance of the relation if a formula, $F$,\ncomputes a causal query, $Q$, as a function of known population probabilities,\n$I$, in every model entailed by a set of model assumptions, $M$. Many problems\nin causal inference can be viewed as the problem of enumerating instances of\nthe relation that satisfy given criteria. This unifies a number of previously\nstudied problems, including causal effect identification, causal discovery and\nrecovery from selection bias. In addition, the relation supports formalizing\nnew problems in causal inference with structural causal models, such as the\nproblem of research design. Causal programming is proposed as a further\ngeneralization of causal inference as the problem of finding optimal instances\nof the relation, with respect to a cost function.', ""The act of explaining across two parties is a feedback loop, where one\nprovides information on what needs to be explained and the other provides an\nexplanation relevant to this information. We apply a reinforcement learning\nframework which emulates this format by providing explanations based on the\nexplainee's current mental model. We conduct novel online human experiments\nwhere explanations generated by various explanation methods are selected and\npresented to participants, using policies which observe participants' mental\nmodels, in order to optimize an interpretability proxy. Our results suggest\nthat mental model-based policies (anchored in our proposed state\nrepresentation) may increase interpretability over multiple sequential\nexplanations, when compared to a random selection baseline. This work provides\ninsight into how to select explanations which increase relevant information for\nusers, and into conducting human-grounded experimentation to understand\ninterpretability."", 'We present novel automatic metrics for machine translation evaluation that\nuse discourse structure and convolution kernels to compare the discourse tree\nof an automatic translation with that of the human reference. We experiment\nwith five transformations and augmentations of a base discourse tree\nrepresentation based on the rhetorical structure theory, and we combine the\nkernel scores for each of them into a single score. Finally, we add other\nmetrics from the ASIYA MT evaluation toolkit, and we tune the weights of the\ncombination on actual human judgments. Experiments on the WMT12 and WMT13\nmetrics shared task datasets show correlation with human judgments that\noutperforms what the best systems that participated in these years achieved,\nboth at the segment and at the system level.', 'The ability of humans to quickly identify general concepts from a handful of\nimages has proven difficult to emulate with robots. Recently, a computer\narchitecture was developed that allows robots to mimic some aspects of this\nhuman ability by modeling concepts as cognitive programs using an instruction\nset of primitive cognitive functions. This allowed a robot to emulate human\nimagination by simulating candidate programs in a world model before\ngeneralizing to the physical world. However, this model used a naive search\nalgorithm that required 30 minutes to discover a single concept, and became\nintractable for programs with more than 20 instructions. To circumvent this\nbottleneck, we present an algorithm that emulates the human cognitive\nheuristics of object factorization and sub-goaling, allowing human-level\ninference speed, improving accuracy, and making the output more explainable.', 'This paper introduces Whittemore, a language for causal programming. Causal\nprogramming is based on the theory of structural causal models and consists of\ntwo primary operations: identification, which finds formulas that compute\ncausal queries, and estimation, which applies formulas to transform probability\ndistributions to other probability distribution. Causal programming provides\nabstractions to declare models, queries, and distributions with syntax similar\nto standard mathematical notation, and conducts rigorous causal inference,\nwithout requiring detailed knowledge of the underlying algorithms. Examples of\ncausal inference with real data are provided, along with discussion of the\nimplementation and possibilities for future extension.', ""Representation learning is a key element of state-of-the-art deep learning\napproaches. It enables to transform raw data into structured vector space\nembeddings. Such embeddings are able to capture the distributional semantics of\ntheir context, e.g. by word windows on natural language sentences, graph walks\non knowledge graphs or convolutions on images. So far, this context is manually\ndefined, resulting in heuristics which are solely optimized for computational\nperformance on certain tasks like link-prediction. However, such heuristic\nmodels of context are fundamentally different to how humans capture\ninformation. For instance, when reading a multi-modal webpage (i) humans do not\nperceive all parts of a document equally: Some words and parts of images are\nskipped, others are revisited several times which makes the perception trace\nhighly non-sequential; (ii) humans construct meaning from a document's content\nby shifting their attention between text and image, among other things, guided\nby layout and design elements. In this paper we empirically investigate the\ndifference between human perception and context heuristics of basic embedding\nmodels. We conduct eye tracking experiments to capture the underlying\ncharacteristics of human perception of media documents containing a mixture of\ntext and images. Based on that, we devise a prototypical computational\nperception-trace model, called CMPM. We evaluate empirically how CMPM can\nimprove a basic skip-gram embedding approach. Our results suggest, that even\nwith a basic human-inspired computational perception model, there is a huge\npotential for improving embeddings since such a model does inherently capture\nmultiple modalities, as well as layout and design elements."", ""We fine-tune GPT-3 to answer long-form questions using a text-based\nweb-browsing environment, which allows the model to search and navigate the\nweb. By setting up the task so that it can be performed by humans, we are able\nto train models on the task using imitation learning, and then optimize answer\nquality with human feedback. To make human evaluation of factual accuracy\neasier, models must collect references while browsing in support of their\nanswers. We train and evaluate our models on ELI5, a dataset of questions asked\nby Reddit users. Our best model is obtained by fine-tuning GPT-3 using behavior\ncloning, and then performing rejection sampling against a reward model trained\nto predict human preferences. This model's answers are preferred by humans 56%\nof the time to those of our human demonstrators, and 69% of the time to the\nhighest-voted answer from Reddit."", 'A structural causal model is made of endogenous (manifest) and exogenous\n(latent) variables. We show that endogenous observations induce linear\nconstraints on the probabilities of the exogenous variables. This allows to\nexactly map a causal model into a credal network. Causal inferences, such as\ninterventions and counterfactuals, can consequently be obtained by standard\nalgorithms for the updating of credal nets. These natively return sharp values\nin the identifiable case, while intervals corresponding to the exact bounds are\nproduced for unidentifiable queries. A characterization of the causal models\nthat allow the map above to be compactly derived is given, along with a\ndiscussion about the scalability for general models. This contribution should\nbe regarded as a systematic approach to represent structural causal models by\ncredal networks and hence to systematically compute causal inferences. A number\nof demonstrative examples is presented to clarify our methodology. Extensive\nexperiments show that approximate algorithms for credal networks can\nimmediately be used to do causal inference in real-size problems.', ""We define a Causal Decision Problem as a Decision Problem where the available\nactions, the family of uncertain events and the set of outcomes are related\nthrough the variables of a Causal Graphical Model $\\mathcal{G}$. A solution\ncriteria based on Pearl's Do-Calculus and the Expected Utility criteria for\nrational preferences is proposed. The implementation of this criteria leads to\nan on-line decision making procedure that has been shown to have similar\nperformance to classic Reinforcement Learning algorithms while allowing for a\ncausal model of an environment to be learned. Thus, we aim to provide the\ntheoretical guarantees of the usefulness and optimality of a decision making\nprocedure based on causal information."", 'We model the behavioral biases of human decision-making in securing\ninterdependent systems and show that such behavioral decision-making leads to a\nsuboptimal pattern of resource allocation compared to non-behavioral (rational)\ndecision-making. We provide empirical evidence for the existence of such\nbehavioral bias model through a controlled subject study with 145 participants.\nWe then propose three learning techniques for enhancing decision-making in\nmulti-round setups. We illustrate the benefits of our decision-making model\nthrough multiple interdependent real-world systems and quantify the level of\ngain compared to the case in which the defenders are behavioral. We also show\nthe benefit of our learning techniques against different attack models. We\nidentify the effects of different system parameters on the degree of\nsuboptimality of security outcomes due to behavioral decision-making.', 'Interpolation is an important property of classical and many non-classical\nlogics that has been shown to have interesting applications in computer science\nand AI. Here we study the Interpolation Property for the the non-monotonic\nsystem of equilibrium logic, establishing weaker or stronger forms of\ninterpolation depending on the precise interpretation of the inference\nrelation. These results also yield a form of interpolation for ground logic\nprograms under the answer sets semantics. For disjunctive logic programs we\nalso study the property of uniform interpolation that is closely related to the\nconcept of variable forgetting. The first-order version of equilibrium logic\nhas analogous Interpolation properties whenever the collection of equilibrium\nmodels is (first-order) definable. Since this is the case for so-called safe\nprograms and theories, it applies to the usual situations that arise in\npractical answer set programming.', 'The encoder-decoder framework achieves state-of-the-art results in keyphrase\ngeneration (KG) tasks by predicting both present keyphrases that appear in the\nsource document and absent keyphrases that do not. However, relying solely on\nthe source document can result in generating uncontrollable and inaccurate\nabsent keyphrases. To address these problems, we propose a novel graph-based\nmethod that can capture explicit knowledge from related references. Our model\nfirst retrieves some document-keyphrases pairs similar to the source document\nfrom a pre-defined index as references. Then a heterogeneous graph is\nconstructed to capture relationships of different granularities between the\nsource document and its references. To guide the decoding process, a\nhierarchical attention and copy mechanism is introduced, which directly copies\nappropriate words from both the source document and its references based on\ntheir relevance and significance. The experimental results on multiple KG\nbenchmarks show that the proposed model achieves significant improvements\nagainst other baseline models, especially with regard to the absent keyphrase\nprediction.', 'The humanity has been facing a plethora of challenges associated with\ninfectious diseases, which kill more than 6 million people a year. Although\ncontinuous efforts have been applied to relieve the potential damages from such\nmisfortunate events, it is unquestionable that there are many persisting\nchallenges yet to overcome. One related issue we particularly address here is\nthe assessment and prediction of such epidemics. In this field of study,\ntraditional and ad-hoc models frequently fail to provide proper predictive\nsituation awareness (PSAW), characterized by understanding the current\nsituations and predicting the future situations. Comprehensive PSAW for\ninfectious disease can support decision making and help to hinder disease\nspread. In this paper, we develop a computing system platform focusing on\ncollective intelligence causal modeling, in order to support PSAW in the domain\nof infectious disease. Analyses of global epidemics require integration of\nmultiple different data and models, which can be originated from multiple\nindependent researchers. These models should be integrated to accurately assess\nand predict the infectious disease in terms of holistic view. The system shall\nprovide three main functions: (1) collaborative causal modeling, (2) causal\nmodel integration, and (3) causal model reasoning. These functions are\nsupported by subject-matter expert and artificial intelligence (AI), with\nuncertainty treatment. Subject-matter experts, as collective intelligence,\ndevelop causal models and integrate them as one joint causal model. The\nintegrated causal model shall be used to reason about: (1) the past, regarding\nhow the causal factors have occurred; (2) the present, regarding how the spread\nis going now; and (3) the future, regarding how it will proceed. Finally, we\nintroduce one use case of predictive situation awareness for the Ebola virus\ndisease.', ""The causal structure of any system can be analyzed at a multitude of spatial\nand temporal scales. It has long been thought that while higher scale (macro)\ndescriptions of causal structure may be useful to observers, they are at best a\ncompressed description and at worse leave out critical information. However,\nrecent research applying information theory to causal analysis has shown that\nthe causal structure of some systems can actually come into focus (be more\ninformative) at a macroscale (Hoel et al. 2013). That is, a macro model of a\nsystem (a map) can be more informative than a fully detailed model of the\nsystem (the territory). This has been called causal emergence. While causal\nemergence may at first glance seem counterintuitive, this paper grounds the\nphenomenon in a classic concept from information theory: Shannon's discovery of\nthe channel capacity. I argue that systems have a particular causal capacity,\nand that different causal models of those systems take advantage of that\ncapacity to various degrees. For some systems, only macroscale causal models\nuse the full causal capacity. Such macroscale causal models can either be\ncoarse-grains, or may leave variables and states out of the model (exogenous)\nin various ways, which can improve the model's efficacy and its informativeness\nvia the same mathematical principles of how error-correcting codes take\nadvantage of an information channel's capacity. As model choice increase, the\ncausal capacity of a system approaches the channel capacity. Ultimately, this\nprovides a general framework for understanding how the causal structure of some\nsystems cannot be fully captured by even the most detailed microscopic model."", 'For robots to be accommodated in human spaces and in humans daily activities,\nrobots should be able to understand messages from the human conversation\npartner. In the same light, humans must also understand the messages that are\nbeing communicated by robots, including the non-verbal ones. We conducted a\nweb-based video study wherein participants gave interpretations on the iconic\ngestures and emblems that were produced by an anthropomorphic robot. Out of the\n15 gestures presented, we found 6 robotic gestures that can be accurately\nrecognized by the human observer. These were nodding, clapping, hugging,\nexpressing anger, walking, and flying. We reviewed these gestures for their\nmeaning from literatures in human and animal behavior. We conclude by\ndiscussing the possible implications of these gestures for the design of social\nrobots that are aimed to have engaging interactions with humans.', 'This paper assumes the hypothesis that human learning is perception based,\nand consequently, the learning process and perceptions should not be\nrepresented and investigated independently or modeled in different simulation\nspaces. In order to keep the analogy between the artificial and human learning,\nthe former is assumed here as being based on the artificial perception. Hence,\ninstead of choosing to apply or develop a Computational Theory of (human)\nPerceptions, we choose to mirror the human perceptions in a numeric\n(computational) space as artificial perceptions and to analyze the\ninterdependence between artificial learning and artificial perception in the\nsame numeric space, using one of the simplest tools of Artificial Intelligence\nand Soft Computing, namely the perceptrons. As practical applications, we\nchoose to work around two examples: Optical Character Recognition and Iris\nRecognition. In both cases a simple Turing test shows that artificial\nperceptions of the difference between two characters and between two irides are\nfuzzy, whereas the corresponding human perceptions are, in fact, crisp.', 'We extend the definition of the marginal causal effect to the continuous\ntreatment setting and develop a novel characterization of causal bias in the\nframework of structural causal models. We prove that our derived bias\nexpression is zero if, and only if, the causal effect is identifiable via\ncovariate adjustment. We show that under some restrictions on the structural\nequations, the causal bias can be estimated efficiently and allows for causal\nregularization of predictive probabilistic models. We demonstrate the\neffectiveness of our method for causal bias quantification in various settings\nwhere (not) controlling for certain covariates would introduce causal bias.', 'In recent years, G\\""odel\'s ontological proof and variations of it were\nformalized and analyzed with automated tools in various ways. We supplement\nthese analyses with a modeling in an automated environment based on first-order\nlogic extended by predicate quantification. Formula macros are used to\nstructure complex formulas and tasks. The analysis is presented as a generated\ntype-set document where informal explanations are interspersed with\npretty-printed formulas and outputs of reasoners for first-order theorem\nproving and second-order quantifier elimination. Previously unnoticed or\nobscured aspects and details of G\\""odel\'s proof become apparent. Practical\napplication possibilities of second-order quantifier elimination are shown and\nthe encountered elimination tasks may serve as benchmarks.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', ""Untile recently crowdsourcing has been primarily conceived as an online\nactivity to harness resources for problem solving. However the emergence of\nopportunistic networking (ON) has opened up crowdsourcing to the spatial\ndomain. In this paper we bring the ON model for potential crowdsourcing in the\nsmart city environment. We introduce cognitive features to the ON that allow\nusers' mobile devices to become aware of the surrounding physical environment.\nSpecifically, we exploit cognitive psychology studies on dynamic memory\nstructures and cognitive heuristics, i.e. mental models that describe how the\nhuman brain handle decision-making amongst complex and real-time stimuli.\nCombined with ON, these cognitive features allow devices to act as proxies in\nthe cyber-world of their users and exchange knowledge to deliver awareness of\nplaces in an urban environment. This is done through tags associated with\nlocations. They represent features that are perceived by humans about a place.\nWe consider the extent to which this knowledge becomes available to\nparticipants, using interactions with locations and other nodes. This is\nassessed taking into account a wide range of cognitive parameters. Outcomes are\nimportant because this functionality could support a new type of recommendation\nsystem that is independent of the traditional forms of networking."", 'We used a token-wise and document-wise sentiment analysis using both\nunsupervised and supervised models applied to both news and user reviews\ndataset. And our token-wise sentiment analysis found a statistically\nsignificant difference in sentiment between the two groups (both of which were\nvery large N), our document-wise supervised sentiment analysis found no\nsignificant difference in sentiment.', 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', ""Neural networks with deep architectures have demonstrated significant\nperformance improvements in computer vision, speech recognition, and natural\nlanguage processing. The challenges in information retrieval (IR), however, are\ndifferent from these other application areas. A common form of IR involves\nranking of documents--or short passages--in response to keyword-based queries.\nEffective IR systems must deal with query-document vocabulary mismatch problem,\nby modeling relationships between different query and document terms and how\nthey indicate relevance. Models should also consider lexical matches when the\nquery contains rare terms--such as a person's name or a product model\nnumber--not seen during training, and to avoid retrieving semantically related\nbut irrelevant results. In many real-life IR tasks, the retrieval involves\nextremely large collections--such as the document index of a commercial Web\nsearch engine--containing billions of documents. Efficient IR methods should\ntake advantage of specialized IR data structures, such as inverted index, to\nefficiently retrieve from large collections. Given an information need, the IR\nsystem also mediates how much exposure an information artifact receives by\ndeciding whether it should be displayed, and where it should be positioned,\namong other results. Exposure-aware IR systems may optimize for additional\nobjectives, besides relevance, such as parity of exposure for retrieved items\nand content publishers. In this thesis, we present novel neural architectures\nand methods motivated by the specific needs and challenges of IR tasks."", 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.', 'This paper introduces application of Reflexive Game Theory to the matter of\nmultistage decision making processes. The idea behind is that each decision\nmaking session has certain parameters like ""when the session is taking place"",\n""who are the group members to make decision"", ""how group members influence on\neach other"", etc. This study illustrates the consecutive or sequential decision\nmaking process, which consist of two stages. During the stage 1 decisions about\nthe parameters of the ultimate decision making are made. Then stage 2 is\nimplementation of Ultimate decision making itself. Since during stage 1 there\ncan be multiple decision sessions. In such a case it takes more than two\nsessions to make ultimate (final) decision. Therefore the overall process of\nultimate decision making becomes multistage decision making process consisting\nof consecutive decision making sessions.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : As an education researcher, my goal is to create a system that can generate explanations for preschool children as part of their early education. I require a model that can provide real-time explanations for the problems inputted. Given that the target users are preschool children, the model needs to be capable of accepting verbal inputs, integrating phrases into natural language, and predicting the user's question. One challenge I face is that adults rarely discuss common sense and trivial questions, so I require an additional dataset to address these types of questions. It is crucial that the content in the dataset is thoroughly vetted by authorities, as the model will be used by children. The model also needs the ability to discard any inappropriate knowledge it may have previously learned. If possible, I would also like guidance on how to strike a balance between the response time of the model and the size of the dataset.","[""There is a growing interest in designing autonomous agents that can work\nalongside humans. Such agents will undoubtedly be expected to explain their\nbehavior and decisions. While generating explanations is an actively researched\ntopic, most works tend to focus on methods that generate explanations that are\none size fits all. As in the specifics of the user-model are completely\nignored. The handful of works that look at tailoring their explanation to the\nuser's background rely on having specific models of the users (either analytic\nmodels or learned labeling models). The goal of this work is thus to propose an\nend-to-end adaptive explanation generation system that begins by learning the\ndifferent types of users that the agent could interact with. Then during the\ninteraction with the target user, it is tasked with identifying the type on the\nfly and adjust its explanations accordingly. The former is achieved by a\ndata-driven clustering approach while for the latter, we compile our\nexplanation generation problem into a POMDP. We demonstrate the usefulness of\nour system on two domains using state-of-the-art POMDP solvers. We also report\nthe results of a user study that investigates the benefits of providing\npersonalized explanations in a human-robot interaction setting."", ""Providing meaningful and actionable explanations to end-users is a\nfundamental prerequisite for implementing explainable intelligent systems in\nthe real world. Explainability is a situated interaction between a user and the\nAI system rather than being static design principles. The content of\nexplanations is context-dependent and must be defined by evidence about the\nuser and its context. This paper seeks to operationalize this concept by\nproposing a formative architecture that defines the explanation space from a\nuser-inspired perspective. The architecture comprises five intertwined\ncomponents to outline explanation requirements for a task: (1) the end-users\nmental models, (2) the end-users cognitive process, (3) the user interface, (4)\nthe human-explainer agent, and the (5) agent process. We first define each\ncomponent of the architecture. Then we present the Abstracted Explanation\nSpace, a modeling tool that aggregates the architecture's components to support\ndesigners in systematically aligning explanations with the end-users work\npractices, needs, and goals. It guides the specifications of what needs to be\nexplained (content - end-users mental model), why this explanation is necessary\n(context - end-users cognitive process), to delimit how to explain it (format -\nhuman-explainer agent and user interface), and when should the explanations be\ngiven. We then exemplify the tool's use in an ongoing case study in the\naircraft maintenance domain. Finally, we discuss possible contributions of the\ntool, known limitations/areas for improvement, and future work to be done.""]","[""Explainability has been a goal for Artificial Intelligence (AI) systems since\ntheir conception, with the need for explainability growing as more complex AI\nmodels are increasingly used in critical, high-stakes settings such as\nhealthcare. Explanations have often added to an AI system in a non-principled,\npost-hoc manner. With greater adoption of these systems and emphasis on\nuser-centric explainability, there is a need for a structured representation\nthat treats explainability as a primary consideration, mapping end user needs\nto specific explanation types and the system's AI capabilities. We design an\nexplanation ontology to model both the role of explanations, accounting for the\nsystem and user attributes in the process, and the range of different\nliterature-derived explanation types. We indicate how the ontology can support\nuser requirements for explanations in the domain of healthcare. We evaluate our\nontology with a set of competency questions geared towards a system designer\nwho might use our ontology to decide which explanation types to include, given\na combination of users' needs and a system's capabilities, both in system\ndesign settings and in real-time operations. Through the use of this ontology,\nsystem designers will be able to make informed choices on which explanations AI\nsystems can and should provide."", 'Many methods now exist for conditioning model outputs on task instructions,\nretrieved documents, and user-provided explanations and feedback. Rather than\nrelying solely on examples of task inputs and outputs, these approaches use\nvaluable additional data for improving model correctness and aligning learned\nmodels with human priors. Meanwhile, a growing body of evidence suggests that\nsome language models can (1) store a large amount of knowledge in their\nparameters, and (2) perform inference over tasks in textual inputs at test\ntime. These results raise the possibility that, for some tasks, humans cannot\nexplain to a model any more about the task than it already knows or could infer\non its own. In this paper, we study the circumstances under which explanations\nof individual data points can (or cannot) improve modeling performance. In\norder to carefully control important properties of the data and explanations,\nwe introduce a synthetic dataset for experiments, and we also make use of three\nexisting datasets with explanations: e-SNLI, TACRED, and SemEval. We first give\na formal framework for the available modeling approaches, in which explanation\ndata can be used as model inputs, as targets, or as a prior. After arguing that\nthe most promising role for explanation data is as model inputs, we propose to\nuse a retrieval-based method and show that it solves our synthetic task with\naccuracies upwards of 95%, while baselines without explanation data achieve\nbelow 65% accuracy. We then identify properties of datasets for which\nretrieval-based modeling fails. With the three existing datasets, we find no\nimprovements from explanation retrieval. Drawing on findings from our synthetic\ntask, we suggest that at least one of six preconditions for successful modeling\nfails to hold with these datasets. Our code is publicly available at\nhttps://github.com/peterbhase/ExplanationRoles', 'Neural predictive models have achieved remarkable performance improvements in\nvarious natural language processing tasks. However, most neural predictive\nmodels suffer from the lack of explainability of predictions, limiting their\npractical utility. This paper proposes a neural predictive approach to make a\nprediction and generate its corresponding explanation simultaneously. It\nleverages the knowledge entailed in explanations as an additional distillation\nsignal for more efficient learning. We conduct a preliminary study on Chinese\nmedical multiple-choice question answering, English natural language inference,\nand commonsense question answering tasks. The experimental results show that\nthe proposed approach can generate reasonable explanations for its predictions\neven with a small-scale training corpus. The proposed method also achieves\nimproved prediction accuracy on three datasets, which indicates that making\npredictions can benefit from generating the explanation in the decision\nprocess.', 'The adoption of machine learning in high-stakes applications such as\nhealthcare and law has lagged in part because predictions are not accompanied\nby explanations comprehensible to the domain user, who often holds the ultimate\nresponsibility for decisions and outcomes. In this paper, we propose an\napproach to generate such explanations in which training data is augmented to\ninclude, in addition to features and labels, explanations elicited from domain\nusers. A joint model is then learned to produce both labels and explanations\nfrom the input features. This simple idea ensures that explanations are\ntailored to the complexity expectations and domain knowledge of the consumer.\nEvaluation spans multiple modeling techniques on a game dataset, a (visual)\naesthetics dataset, a chemical odor dataset and a Melanoma dataset showing that\nour approach is generalizable across domains and algorithms. Results\ndemonstrate that meaningful explanations can be reliably taught to machine\nlearning algorithms, and in some cases, also improve modeling accuracy.', ""Explainable machine learning and artificial intelligence models have been\nused to justify a model's decision-making process. This added transparency aims\nto help improve user performance and understanding of the underlying model.\nHowever, in practice, explainable systems face many open questions and\nchallenges. Specifically, designers might reduce the complexity of deep\nlearning models in order to provide interpretability. The explanations\ngenerated by these simplified models, however, might not accurately justify and\nbe truthful to the model. This can further add confusion to the users as they\nmight not find the explanations meaningful with respect to the model\npredictions. Understanding how these explanations affect user behavior is an\nongoing challenge. In this paper, we explore how explanation veracity affects\nuser performance and agreement in intelligent systems. Through a controlled\nuser study with an explainable activity recognition system, we compare\nvariations in explanation veracity for a video review and querying task. The\nresults suggest that low veracity explanations significantly decrease user\nperformance and agreement compared to both accurate explanations and a system\nwithout explanations. These findings demonstrate the importance of accurate and\nunderstandable explanations and caution that poor explanations can sometimes be\nworse than no explanations with respect to their effect on user performance and\nreliance on an AI system."", ""With the increasing complexity of CPSs, their behavior and decisions become\nincreasingly difficult to understand and comprehend for users and other\nstakeholders. Our vision is to build self-explainable systems that can, at\nrun-time, answer questions about the system's past, current, and future\nbehavior. As hitherto no design methodology or reference framework exists for\nbuilding such systems, we propose the MAB-EX framework for building\nself-explainable systems that leverage requirements- and explainability models\nat run-time. The basic idea of MAB-EX is to first Monitor and Analyze a certain\nbehavior of a system, then Build an explanation from explanation models and\nconvey this EXplanation in a suitable way to a stakeholder. We also take into\naccount that new explanations can be learned, by updating the explanation\nmodels, should new and yet un-explainable behavior be detected by the system."", 'We propose a new method for generating explanations with AI and a tool to\ntest its expressive power within a user interface. In order to bridge the gap\nbetween philosophy and human-computer interfaces, we show a new approach for\nthe generation of interactive explanations based on a sophisticated pipeline of\nAI algorithms for structuring natural language documents into knowledge graphs,\nanswering questions effectively and satisfactorily. With this work we aim to\nprove that the philosophical theory of explanations presented by Achinstein can\nbe actually adapted for being implemented into a concrete software application,\nas an interactive and illocutionary process of answering questions.\nSpecifically, our contribution is an approach to frame illocution in a\ncomputer-friendly way, to achieve user-centrality with statistical question\nanswering. Indeed, we frame the illocution of an explanatory process as that\nmechanism responsible for anticipating the needs of the explainee in the form\nof unposed, implicit, archetypal questions, hence improving the user-centrality\nof the underlying explanatory process. Therefore, we hypothesise that if an\nexplanatory process is an illocutionary act of providing content-giving answers\nto questions, and illocution is as we defined it, the more explicit and\nimplicit questions can be answered by an explanatory tool, the more usable its\nexplanations. We tested our hypothesis with a user-study involving more than 60\nparticipants, on two XAI-based systems, one for credit approval and one for\nheart disease prediction. The results showed that increasing the illocutionary\npower of an explanatory tool can produce statistically significant improvements\non effectiveness. This, combined with a visible alignment between the\nincrements in effectiveness and satisfaction, suggests that our understanding\nof illocution can be correct, giving evidence in favour of our theory.', ""As Graph Neural Networks (GNNs) are widely adopted in digital pathology,\nthere is increasing attention to developing explanation models (explainers) of\nGNNs for improved transparency in clinical decisions.\n  Existing explainers discover an explanatory subgraph relevant to the\nprediction.\n  However, such a subgraph is insufficient to reveal all the critical\nbiological substructures for the prediction because the prediction will remain\nunchanged after removing that subgraph.\n  Hence, an explanatory subgraph should be not only necessary for prediction,\nbut also sufficient to uncover the most predictive regions for the explanation.\n  Such explanation requires a measurement of information transferred from\ndifferent input subgraphs to the predictive output, which we define as\ninformation flow.\n  In this work, we address these key challenges and propose IFEXPLAINER, which\ngenerates a necessary and sufficient explanation for GNNs.\n  To evaluate the information flow within GNN's prediction, we first propose a\nnovel notion of predictiveness, named $f$-information, which is directional and\nincorporates the realistic capacity of the GNN model.\n  Based on it, IFEXPLAINER generates the explanatory subgraph with maximal\ninformation flow to the prediction.\n  Meanwhile, it minimizes the information flow from the input to the predictive\nresult after removing the explanation.\n  Thus, the produced explanation is necessarily important to the prediction and\nsufficient to reveal the most crucial substructures.\n  We evaluate IFEXPLAINER to interpret GNN's predictions on breast cancer\nsubtyping.\n  Experimental results on the BRACS dataset show the superior performance of\nthe proposed method."", 'Counterfactual explanation methods interpret the outputs of a machine\nlearning model in the form of ""what-if scenarios"" without compromising the\nfidelity-interpretability trade-off. They explain how to obtain a desired\nprediction from the model by recommending small changes to the input features,\naka recourse. We believe an actionable recourse should be created based on\nsound counterfactual explanations originating from the distribution of the\nground-truth data and linked to the domain knowledge. Moreover, it needs to\npreserve the coherency between changed/unchanged features while satisfying\nuser/domain-specified constraints. This paper introduces CARE, a modular\nexplanation framework that addresses the model- and user-level desiderata in a\nconsecutive and structured manner. We tackle the existing requirements by\nproposing novel and efficient solutions that are formulated in a\nmulti-objective optimization framework. The designed framework enables\nincluding arbitrary requirements and generating counterfactual explanations and\nactionable recourse by choice. As a model-agnostic approach, CARE generates\nmultiple, diverse explanations for any black-box model in tabular\nclassification and regression settings. Several experiments on standard data\nsets and black-box models demonstrate the effectiveness of our modular\nframework and its superior performance compared to the baselines.', ""An accurate abstractive summary of a document should contain all its salient\ninformation and should be logically entailed by the input document. We improve\nthese important aspects of abstractive summarization via multi-task learning\nwith the auxiliary tasks of question generation and entailment generation,\nwhere the former teaches the summarization model how to look for salient\nquestioning-worthy details, and the latter teaches the model how to rewrite a\nsummary which is a directed-logical subset of the input document. We also\npropose novel multi-task architectures with high-level (semantic)\nlayer-specific sharing across multiple encoder and decoder layers of the three\ntasks, as well as soft-sharing mechanisms (and show performance ablations and\nanalysis examples of each contribution). Overall, we achieve statistically\nsignificant improvements over the state-of-the-art on both the CNN/DailyMail\nand Gigaword datasets, as well as on the DUC-2002 transfer setup. We also\npresent several quantitative and qualitative analysis studies of our model's\nlearned saliency and entailment skills."", 'A central issue addressed by the rapidly growing research area of eXplainable\nArtificial Intelligence (XAI) is to provide methods to give explanations for\nthe behaviours of Machine Learning (ML) non-interpretable models after the\ntraining. Recently, it is becoming more and more evident that new directions to\ncreate better explanations should take into account what a good explanation is\nto a human user. This paper suggests taking advantage of developing an XAI\nframework that allows producing multiple explanations for the response of image\na classification system in terms of potentially different middle-level input\nfeatures. To this end, we propose an XAI framework able to construct\nexplanations in terms of input features extracted by auto-encoders. We start\nfrom the hypothesis that some autoencoders, relying on standard data\nrepresentation approaches, could extract more salient and understandable input\nproperties, which we call here \\textit{Middle-Level input Features} (MLFs), for\na user with respect to raw low-level features. Furthermore, extracting\ndifferent types of MLFs through different type of autoencoders, different types\nof explanations for the same ML system behaviour can be returned. We\nexperimentally tested our method on two different image datasets and using\nthree different types of MLFs. The results are encouraging. Although our novel\napproach was tested in the context of image classification, it can potentially\nbe used on other data types to the extent that auto-encoders to extract humanly\nunderstandable representations can be applied.', 'Developing methods of automated inference that are able to provide users with\ncompelling human-readable justifications for why the answer to a question is\ncorrect is critical for domains such as science and medicine, where user trust\nand detecting costly errors are limiting factors to adoption. One of the\ncentral barriers to training question answering models on explainable inference\ntasks is the lack of gold explanations to serve as training data. In this paper\nwe present a corpus of explanations for standardized science exams, a recent\nchallenge task for question answering. We manually construct a corpus of\ndetailed explanations for nearly all publicly available standardized elementary\nscience question (approximately 1,680 3rd through 5th grade questions) and\nrepresent these as ""explanation graphs"" -- sets of lexically overlapping\nsentences that describe how to arrive at the correct answer to a question\nthrough a combination of domain and world knowledge. We also provide an\nexplanation-centered tablestore, a collection of semi-structured tables that\ncontain the knowledge to construct these elementary science explanations.\nTogether, these two knowledge resources map out a substantial portion of the\nknowledge required for answering and explaining elementary science exams, and\nprovide both structured and free-text training data for the explainable\ninference task.', 'Recently we see a rising number of methods in the field of eXplainable\nArtificial Intelligence. To our surprise, their development is driven by model\ndevelopers rather than a study of needs for human end users. The analysis of\nneeds, if done, takes the form of an A/B test rather than a study of open\nquestions. To answer the question ""What would a human operator like to ask the\nML model?"" we propose a conversational system explaining decisions of the\npredictive model. In this experiment, we developed a chatbot called dr_ant to\ntalk about machine learning model trained to predict survival odds on Titanic.\nPeople can talk with dr_ant about different aspects of the model to understand\nthe rationale behind its predictions. Having collected a corpus of 1000+\ndialogues, we analyse the most common types of questions that users would like\nto ask. To our knowledge, it is the first study which uses a conversational\nsystem to collect the needs of human operators from the interactive and\niterative dialogue explorations of a predictive model.', 'A desirable property of learning systems is to be both effective and\ninterpretable. Towards this goal, recent models have been proposed that first\ngenerate an extractive explanation from the input text and then generate a\nprediction on just the explanation called explain-then-predict models. These\nmodels primarily consider the task input as a supervision signal in learning an\nextractive explanation and do not effectively integrate rationales data as an\nadditional inductive bias to improve task performance. We propose a novel yet\nsimple approach ExPred, that uses multi-task learning in the explanation\ngeneration phase effectively trading-off explanation and prediction losses. And\nthen we use another prediction network on just the extracted explanations for\noptimizing the task performance. We conduct an extensive evaluation of our\napproach on three diverse language datasets -- fact verification, sentiment\nclassification, and QA -- and find that we substantially outperform existing\napproaches.', 'Deep neural networks are becoming more and more popular due to their\nrevolutionary success in diverse areas, such as computer vision, natural\nlanguage processing, and speech recognition. However, the decision-making\nprocesses of these models are generally not interpretable to users. In various\ndomains, such as healthcare, finance, or law, it is critical to know the\nreasons behind a decision made by an artificial intelligence system. Therefore,\nseveral directions for explaining neural models have recently been explored. In\nthis thesis, I investigate two major directions for explaining deep neural\nnetworks. The first direction consists of feature-based post-hoc explanatory\nmethods, that is, methods that aim to explain an already trained and fixed\nmodel (post-hoc), and that provide explanations in terms of input features,\nsuch as tokens for text and superpixels for images (feature-based). The second\ndirection consists of self-explanatory neural models that generate natural\nlanguage explanations, that is, models that have a built-in module that\ngenerates explanations for the predictions of the model.', 'We explore the problem of step-wise explaining how to solve constraint\nsatisfaction problems, with a use case on logic grid puzzles. More\nspecifically, we study the problem of explaining the inference steps that one\ncan take during propagation, in a way that is easy to interpret for a person.\nThereby, we aim to give the constraint solver explainable agency, which can\nhelp in building trust in the solver by being able to understand and even learn\nfrom the explanations. The main challenge is that of finding a sequence of\nsimple explanations, where each explanation should aim to be as cognitively\neasy as possible for a human to verify and understand. This contrasts with the\narbitrary combination of facts and constraints that the solver may use when\npropagating. We propose the use of a cost function to quantify how simple an\nindividual explanation of an inference step is, and identify the\nexplanation-production problem of finding the best sequence of explanations of\na CSP. Our approach is agnostic of the underlying constraint propagation\nmechanisms, and can provide explanations even for inference steps resulting\nfrom combinations of constraints. In case multiple constraints are involved, we\nalso develop a mechanism that allows to break the most difficult steps up and\nthus gives the user the ability to zoom in on specific parts of the\nexplanation. Our proposed algorithm iteratively constructs the explanation\nsequence by using an optimistic estimate of the cost function to guide the\nsearch for the best explanation at each step. Our experiments on logic grid\npuzzles show the feasibility of the approach in terms of the quality of the\nindividual explanations and the resulting explanation sequences obtained.', 'As machine learning models become more accurate, they typically become more\ncomplex and uninterpretable by humans. The black-box character of these models\nholds back its acceptance in practice, especially in high-risk domains where\nthe consequences of failure could be catastrophic such as health-care or\ndefense. Providing understandable and useful explanations behind ML models or\npredictions can increase the trust of the user. Example-based reasoning, which\nentails leveraging previous experience with analogous tasks to make a decision,\nis a well known strategy for problem solving and justification. This work\npresents a new explanation extraction method called LEAFAGE, for a prediction\nmade by any black-box ML model. The explanation consists of the visualization\nof similar examples from the training set and the importance of each feature.\nMoreover, these explanations are contrastive which aims to take the\nexpectations of the user into account. LEAFAGE is evaluated in terms of\nfidelity to the underlying black-box model and usefulness to the user. The\nresults showed that LEAFAGE performs overall better than the current\nstate-of-the-art method LIME in terms of fidelity, on ML models with non-linear\ndecision boundary. A user-study was conducted which focused on revealing the\ndifferences between example-based and feature importance-based explanations. It\nshowed that example-based explanations performed significantly better than\nfeature importance-based explanation, in terms of perceived transparency,\ninformation sufficiency, competence and confidence. Counter-intuitively, when\nthe gained knowledge of the participants was tested, it showed that they\nlearned less about the black-box model after seeing a feature importance-based\nexplanation than seeing no explanation at all. The participants found feature\nimportance-based explanation vague and hard to generalize it to other\ninstances.', ""Prior work on generating explanations in a planning and decision-making\ncontext has focused on providing the rationale behind an AI agent's decision\nmaking. While these methods provide the right explanations from the explainer's\nperspective, they fail to heed the cognitive requirement of understanding an\nexplanation from the explainee's (the human's) perspective. In this work, we\nset out to address this issue by first considering the influence of information\norder in an explanation, or the progressiveness of explanations. Intuitively,\nprogression builds later concepts on previous ones and is known to contribute\nto better learning. In this work, we aim to investigate similar effects during\nexplanation generation when an explanation is broken into multiple parts that\nare communicated sequentially. The challenge here lies in modeling the humans'\npreferences for information order in receiving such explanations to assist\nunderstanding. Given this sequential process, a formulation based on goal-based\nMDP for generating progressive explanations is presented. The reward function\nof this MDP is learned via inverse reinforcement learning based on explanations\nthat are retrieved via human subject studies. We first evaluated our approach\non a scavenger-hunt domain to demonstrate its effectively in capturing the\nhumans' preferences. Upon analyzing the results, it revealed something more\nfundamental: the preferences arise strongly from both domain dependent and\nindependence features. The correlation with domain independent features pushed\nus to verify this result further in an escape room domain. Results confirmed\nour hypothesis that the process of understanding an explanation was a dynamic\nprocess. The human preference that reflected this aspect corresponded exactly\nto the progression for knowledge assimilation hidden deeper in our cognitive\nprocess."", 'Recently, there has been increasing interest in transparency and\ninterpretability in Deep Reinforcement Learning (DRL) systems. Verbal\nexplanations, as the most natural way of communication in our daily life,\ndeserve more attention, since they allow users to gain a better understanding\nof the system which ultimately could lead to a high level of trust and smooth\ncollaboration. This paper reports a novel work in generating verbal\nexplanations for DRL behaviors agent. A rule-based model is designed to\nconstruct explanations using a series of rules which are predefined with prior\nknowledge. A learning model is then proposed to expand the implicit logic of\ngenerating verbal explanation to general situations by employing rule-based\nexplanations as training data. The learning model is shown to have better\nflexibility and generalizability than the static rule-based model. The\nperformance of both models is evaluated quantitatively through objective\nmetrics. The results show that verbal explanation generated by both models\nimprove subjective satisfaction of users towards the interpretability of DRL\nsystems. Additionally, seven variants of the learning model are designed to\nillustrate the contribution of input channels, attention mechanism, and\nproposed encoder in improving the quality of verbal explanation.', ""Local Interpretable Model-Agnostic Explanations (LIME) is a popular method to\nperform interpretability of any kind of Machine Learning (ML) model. It\nexplains one ML prediction at a time, by learning a simple linear model around\nthe prediction. The model is trained on randomly generated data points, sampled\nfrom the training dataset distribution and weighted according to the distance\nfrom the reference point - the one being explained by LIME. Feature selection\nis applied to keep only the most important variables. LIME is widespread across\ndifferent domains, although its instability - a single prediction may obtain\ndifferent explanations - is one of the major shortcomings. This is due to the\nrandomness in the sampling step, as well as to the flexibility in tuning the\nweights and determines a lack of reliability in the retrieved explanations,\nmaking LIME adoption problematic. In Medicine especially, clinical\nprofessionals trust is mandatory to determine the acceptance of an explainable\nalgorithm, considering the importance of the decisions at stake and the related\nlegal issues. In this paper, we highlight a trade-off between explanation's\nstability and adherence, namely how much it resembles the ML model. Exploiting\nour innovative discovery, we propose a framework to maximise stability, while\nretaining a predefined level of adherence. OptiLIME provides freedom to choose\nthe best adherence-stability trade-off level and more importantly, it clearly\nhighlights the mathematical properties of the retrieved explanation. As a\nresult, the practitioner is provided with tools to decide whether the\nexplanation is reliable, according to the problem at hand. We extensively test\nOptiLIME on a toy dataset - to present visually the geometrical findings - and\na medical dataset. In the latter, we show how the method comes up with\nmeaningful explanations both from a medical and mathematical standpoint."", 'The need for transparency of predictive systems based on Machine Learning\nalgorithms arises as a consequence of their ever-increasing proliferation in\nthe industry. Whenever black-box algorithmic predictions influence human\naffairs, the inner workings of these algorithms should be scrutinised and their\ndecisions explained to the relevant stakeholders, including the system\nengineers, the system\'s operators and the individuals whose case is being\ndecided. While a variety of interpretability and explainability methods is\navailable, none of them is a panacea that can satisfy all diverse expectations\nand competing objectives that might be required by the parties involved. We\naddress this challenge in this paper by discussing the promises of Interactive\nMachine Learning for improved transparency of black-box systems using the\nexample of contrastive explanations -- a state-of-the-art approach to\nInterpretable Machine Learning.\n  Specifically, we show how to personalise counterfactual explanations by\ninteractively adjusting their conditional statements and extract additional\nexplanations by asking follow-up ""What if?"" questions. Our experience in\nbuilding, deploying and presenting this type of system allowed us to list\ndesired properties as well as potential limitations, which can be used to guide\nthe development of interactive explainers. While customising the medium of\ninteraction, i.e., the user interface comprising of various communication\nchannels, may give an impression of personalisation, we argue that adjusting\nthe explanation itself and its content is more important. To this end,\nproperties such as breadth, scope, context, purpose and target of the\nexplanation have to be considered, in addition to explicitly informing the\nexplainee about its limitations and caveats...', ""The paper presents an approach to build a question and answer system that is\ncapable of processing the information in a large dataset and allows the user to\ngain knowledge from this dataset by asking questions in natural language form.\nKey content of this research covers four dimensions which are; Corpus\nPreprocessing, Question Preprocessing, Deep Neural Network for Answer\nExtraction and Answer Generation. The system is capable of understanding the\nquestion, responds to the user's query in natural language form as well. The\ngoal is to make the user feel as if they were interacting with a person than a\nmachine."", 'Feature based local attribution methods are amongst the most prevalent in\nexplainable artificial intelligence (XAI) literature. Going beyond standard\ncorrelation, recently, methods have been proposed that highlight what should be\nminimally sufficient to justify the classification of an input (viz. pertinent\npositives). While minimal sufficiency is an attractive property, the resulting\nexplanations are often too sparse for a human to understand and evaluate the\nlocal behavior of the model, thus making it difficult to judge its overall\nquality. To overcome these limitations, we propose a novel method called\nPath-Sufficient Explanations Method (PSEM) that outputs a sequence of\nsufficient explanations for a given input of strictly decreasing size (or\nvalue) -- from original input to a minimally sufficient explanation -- which\ncan be thought to trace the local boundary of the model in a smooth manner,\nthus providing better intuition about the local model behavior for the specific\ninput. We validate these claims, both qualitatively and quantitatively, with\nexperiments that show the benefit of PSEM across all three modalities (image,\ntabular and text). A user study depicts the strength of the method in\ncommunicating the local behavior, where (many) users are able to correctly\ndetermine the prediction made by a model.', 'When answering a question, people often draw upon their rich world knowledge\nin addition to the particular context. Recent work has focused primarily on\nanswering questions given some relevant document or context, and required very\nlittle general background. To investigate question answering with prior\nknowledge, we present CommonsenseQA: a challenging new dataset for commonsense\nquestion answering. To capture common sense beyond associations, we extract\nfrom ConceptNet (Speer et al., 2017) multiple target concepts that have the\nsame semantic relation to a single source concept. Crowd-workers are asked to\nauthor multiple-choice questions that mention the source concept and\ndiscriminate in turn between each of the target concepts. This encourages\nworkers to create questions with complex semantics that often require prior\nknowledge. We create 12,247 questions through this procedure and demonstrate\nthe difficulty of our task with a large number of strong baselines. Our best\nbaseline is based on BERT-large (Devlin et al., 2018) and obtains 56% accuracy,\nwell below human performance, which is 89%.', ""In consequential domains such as recidivism prediction, facility inspection,\nand benefit assignment, it's important for individuals to know the\ndecision-relevant information for the model's prediction. In addition,\npredictions should be fair both in terms of the outcome and the justification\nof the outcome. In other words, decision-relevant features should provide\nsufficient information for the predicted outcome and should be independent of\nthe membership of individuals in protected groups such as race and gender. In\nthis work, we focus on the problem of (un)fairness in the justification of the\ntext-based neural models. We tie the explanatory power of the model to fairness\nin the outcome and propose a fairness-aware summarization mechanism to detect\nand counteract the bias in such models. Given a potentially biased natural\nlanguage explanation for a decision, we use a multi-task neural model and an\nattribution mechanism based on integrated gradients to extract high-utility and\nlow-bias justifications in form of a summary. The extracted summary is then\nused for training a model to make decisions for individuals. Results on several\nreal world datasets suggest that our method drastically limits the demographic\nleakage in the input (fairness in justification) while moderately enhancing the\nfairness in the outcome. Our model is also effective in detecting and\ncounteracting several types of data poisoning attacks that synthesize\nrace-coded reasoning or irrelevant justifications."", 'The use of deep neural networks to make high risk decisions creates a need\nfor global and local explanations so that users and experts have confidence in\nthe modeling algorithms. We introduce a novel technique to find global and\nlocal explanations for time series data used in binary classification machine\nlearning systems. We identify the most salient of the original features used by\na black box model to distinguish between classes. The explanation can be made\non categorical, continuous, and time series data and can be generalized to any\nbinary classification model. The analysis is conducted on time series data to\ntrain a long short-term memory deep neural network and uses the time dependent\nstructure of the underlying features in the explanation. The proposed technique\nattributes weights to features to explain an observations risk of belonging to\na class as a multiplicative factor of a base hazard rate. We use a variation of\nthe Cox Proportional Hazards regression, a Generalized Additive Model, to\nexplain the effect of variables upon the probability of an in-class response\nfor a score output from the black box model. The covariates incorporate time\ndependence structure in the features so the explanation is inclusive of the\nunderlying time series data structure.', 'As AI becomes an integral part of our lives, the development of explainable\nAI, embodied in the decision-making process of an AI or robotic agent, becomes\nimperative. For a robotic teammate, the ability to generate explanations to\njustify its behavior is one of the key requirements of explainable agency.\nPrior work on explanation generation has been focused on supporting the\nrationale behind the robot\'s decision or behavior. These approaches, however,\nfail to consider the mental demand for understanding the received explanation.\nIn other words, the human teammate is expected to understand an explanation no\nmatter how much information is presented. In this work, we argue that\nexplanations, especially those of a complex nature, should be made in an online\nfashion during the execution, which helps spread out the information to be\nexplained and thus reduce the mental workload of humans in highly cognitive\ndemanding tasks. However, a challenge here is that the different parts of an\nexplanation may be dependent on each other, which must be taken into account\nwhen generating online explanations. To this end, a general formulation of\nonline explanation generation is presented with three variations satisfying\ndifferent ""online"" properties. The new explanation generation methods are based\non a model reconciliation setting introduced in our prior work. We evaluated\nour methods both with human subjects in a simulated rover domain, using NASA\nTask Load Index (TLX), and synthetically with ten different problems across two\nstandard IPC domains. Results strongly suggest that our methods generate\nexplanations that are perceived as less cognitively demanding and much\npreferred over the baselines and are computationally efficient.', 'Machine learning plays an increasing role in intelligent tutoring systems as\nboth the amount of data available and specialization among students grow.\nNowadays, these systems are frequently deployed on mobile applications. Users\non such mobile education platforms are dynamic, frequently being added,\naccessing the application with varying levels of focus, and changing while\nusing the service. The education material itself, on the other hand, is often\nstatic and is an exhaustible resource whose use in tasks such as problem\nrecommendation must be optimized. The ability to update user models with\nrespect to educational material in real-time is thus essential; however,\nexisting approaches require time-consuming re-training of user features\nwhenever new data is added. In this paper, we introduce a neural pedagogical\nagent for real-time user modeling in the task of predicting user response\ncorrectness, a central task for mobile education applications. Our model,\ninspired by work in natural language processing on sequence modeling and\nmachine translation, updates user features in real-time via bidirectional\nrecurrent neural networks with an attention mechanism over embedded\nquestion-response pairs. We experiment on the mobile education application\nSantaTOEIC, which has 559k users, 66M response data points as well as a set of\n10k study problems each expert-annotated with topic tags and gathered since\n2016. Our model outperforms existing approaches over several metrics in\npredicting user response correctness, notably out-performing other methods on\nnew users without large question-response histories. Additionally, our\nattention mechanism and annotated tag set allow us to create an interpretable\neducation platform, with a smart review system that addresses the\naforementioned issue of varied user attention and problem exhaustion.', 'Building a dialogue system that can communicate naturally with humans is a\nchallenging yet interesting problem of agent-based computing. The rapid growth\nin this area is usually hindered by the long-standing problem of data scarcity\nas these systems are expected to learn syntax, grammar, decision making, and\nreasoning from insufficient amounts of task-specific dataset. The recently\nintroduced pre-trained language models have the potential to address the issue\nof data scarcity and bring considerable advantages by generating contextualized\nword embeddings. These models are considered counterpart of ImageNet in NLP and\nhave demonstrated to capture different facets of language such as hierarchical\nrelations, long-term dependency, and sentiment. In this short survey paper, we\ndiscuss the recent progress made in the field of pre-trained language models.\nWe also deliberate that how the strengths of these language models can be\nleveraged in designing more engaging and more eloquent conversational agents.\nThis paper, therefore, intends to establish whether these pre-trained models\ncan overcome the challenges pertinent to dialogue systems, and how their\narchitecture could be exploited in order to overcome these challenges. Open\nchallenges in the field of dialogue systems have also been deliberated.', 'Explainable question answering systems predict an answer together with an\nexplanation showing why the answer has been selected. The goal is to enable\nusers to assess the correctness of the system and understand its reasoning\nprocess. However, we show that current models and evaluation settings have\nshortcomings regarding the coupling of answer and explanation which might cause\nserious issues in user experience. As a remedy, we propose a hierarchical model\nand a new regularization term to strengthen the answer-explanation coupling as\nwell as two evaluation scores to quantify the coupling. We conduct experiments\non the HOTPOTQA benchmark data set and perform a user study. The user study\nshows that our models increase the ability of the users to judge the\ncorrectness of the system and that scores like F1 are not enough to estimate\nthe usefulness of a model in a practical setting with human users. Our scores\nare better aligned with user experience, making them promising candidates for\nmodel selection.', 'The study of adverse childhood experiences and their consequences has emerged\nover the past 20 years. In this study, we aimed to leverage explainable\nartificial intelligence, and propose a proof-of-concept prototype for a\nknowledge-driven evidence-based recommendation system to improve surveillance\nof adverse childhood experiences. We used concepts from an ontology that we\nhave developed to build and train a question-answering agent using the Google\nDialogFlow engine. In addition to the question-answering agent, the initial\nprototype includes knowledge graph generation and recommendation components\nthat leverage third-party graph technology. To showcase the framework\nfunctionalities, we here present a prototype design and demonstrate the main\nfeatures through four use case scenarios motivated by an initiative currently\nimplemented at a children hospital in Memphis, Tennessee. Ongoing development\nof the prototype requires implementing an optimization algorithm of the\nrecommendations, incorporating a privacy layer through a personal health\nlibrary, and conducting a clinical trial to assess both usability and\nusefulness of the implementation. This semantic-driven explainable artificial\nintelligence prototype can enhance health care practitioners ability to provide\nexplanations for the decisions they make.', 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'In this paper we propose a novel method that provides contrastive\nexplanations justifying the classification of an input by a black box\nclassifier such as a deep neural network. Given an input we find what should be\n%necessarily and minimally and sufficiently present (viz. important object\npixels in an image) to justify its classification and analogously what should\nbe minimally and necessarily \\emph{absent} (viz. certain background pixels). We\nargue that such explanations are natural for humans and are used commonly in\ndomains such as health care and criminology. What is minimally but critically\n\\emph{absent} is an important part of an explanation, which to the best of our\nknowledge, has not been explicitly identified by current explanation methods\nthat explain predictions of neural networks. We validate our approach on three\nreal datasets obtained from diverse domains; namely, a handwritten digits\ndataset MNIST, a large procurement fraud dataset and a brain activity strength\ndataset. In all three cases, we witness the power of our approach in generating\nprecise explanations that are also easy for human experts to understand and\nevaluate.', ""Providing explanations is considered an imperative ability for an AI agent in\na human-robot teaming framework. The right explanation provides the rationale\nbehind an AI agent's decision-making. However, to maintain the human teammate's\ncognitive demand to comprehend the provided explanations, prior works have\nfocused on providing explanations in a specific order or intertwining the\nexplanation generation with plan execution. Moreover, these approaches do not\nconsider the degree of details required to share throughout the provided\nexplanations. In this work, we argue that the agent-generated explanations,\nespecially the complex ones, should be abstracted to be aligned with the level\nof details the human teammate desires to maintain the recipient's cognitive\nload. Therefore, learning a hierarchical explanations model is a challenging\ntask. Moreover, the agent needs to follow a consistent high-level policy to\ntransfer the learned teammate preferences to a new scenario while lower-level\ndetailed plans are different. Our evaluation confirmed the process of\nunderstanding an explanation, especially a complex and detailed explanation, is\nhierarchical. The human preference that reflected this aspect corresponded\nexactly to creating and employing abstraction for knowledge assimilation hidden\ndeeper in our cognitive process. We showed that hierarchical explanations\nachieved better task performance and behavior interpretability while reduced\ncognitive load. These results shed light on designing explainable agents\nutilizing reinforcement learning and planning across various domains."", 'To generate trust with their users, Explainable Artificial Intelligence (XAI)\nsystems need to include an explanation model that can communicate the internal\ndecisions, behaviours and actions to the interacting humans. Successful\nexplanation involves both cognitive and social processes. In this paper we\nfocus on the challenge of meaningful interaction between an explainer and an\nexplainee and investigate the structural aspects of an explanation in order to\npropose a human explanation dialog model. We follow a bottom-up approach to\nderive the model by analysing transcripts of 398 different explanation dialog\ntypes. We use grounded theory to code and identify key components of which an\nexplanation dialog consists. We carry out further analysis to identify the\nrelationships between components and sequences and cycles that occur in a\ndialog. We present a generalized state model obtained by the analysis and\ncompare it with an existing conceptual dialog model of explanation.', 'Our goal is to create an interactive natural language interface that\nefficiently and reliably learns from users to complete tasks in simulated\nrobotics settings. We introduce a neural semantic parsing system that learns\nnew high-level abstractions through decomposition: users interactively teach\nthe system by breaking down high-level utterances describing novel behavior\ninto low-level steps that it can understand. Unfortunately, existing methods\neither rely on grammars which parse sentences with limited flexibility, or\nneural sequence-to-sequence models that do not learn efficiently or reliably\nfrom individual examples. Our approach bridges this gap, demonstrating the\nflexibility of modern neural systems, as well as the one-shot reliable\ngeneralization of grammar-based methods. Our crowdsourced interactive\nexperiments suggest that over time, users complete complex tasks more\nefficiently while using our system by leveraging what they just taught. At the\nsame time, getting users to trust the system enough to be incentivized to teach\nhigh-level utterances is still an ongoing challenge. We end with a discussion\nof some of the obstacles we need to overcome to fully realize the potential of\nthe interactive paradigm.', ""A long-term goal of machine learning research is to build an intelligent\ndialog agent. Most research in natural language understanding has focused on\nlearning from fixed training sets of labeled data, with supervision either at\nthe word level (tagging, parsing tasks) or sentence level (question answering,\nmachine translation). This kind of supervision is not realistic of how humans\nlearn, where language is both learned by, and used for, communication. In this\nwork, we study dialog-based language learning, where supervision is given\nnaturally and implicitly in the response of the dialog partner during the\nconversation. We study this setup in two domains: the bAbI dataset of (Weston\net al., 2015) and large-scale question answering from (Dodge et al., 2015). We\nevaluate a set of baseline learning strategies on these tasks, and show that a\nnovel model incorporating predictive lookahead is a promising approach for\nlearning from a teacher's response. In particular, a surprising result is that\nit can learn to answer questions correctly without any reward-based supervision\nat all."", 'Commonsense inference to understand and explain human language is a\nfundamental research problem in natural language processing. Explaining human\nconversations poses a great challenge as it requires contextual understanding,\nplanning, inference, and several aspects of reasoning including causal,\ntemporal, and commonsense reasoning. In this work, we introduce CIDER -- a\nmanually curated dataset that contains dyadic dialogue explanations in the form\nof implicit and explicit knowledge triplets inferred using contextual\ncommonsense inference. Extracting such rich explanations from conversations can\nbe conducive to improving several downstream applications. The annotated\ntriplets are categorized by the type of commonsense knowledge present (e.g.,\ncausal, conditional, temporal). We set up three different tasks conditioned on\nthe annotated dataset: Dialogue-level Natural Language Inference, Span\nExtraction, and Multi-choice Span Selection. Baseline results obtained with\ntransformer-based models reveal that the tasks are difficult, paving the way\nfor promising future research. The dataset and the baseline implementations are\npublicly available at https://cider-task.github.io/cider/.', 'Current conversational AI systems aim to understand a set of pre-designed\nrequests and execute related actions, which limits them to evolve naturally and\nadapt based on human interactions. Motivated by how children learn their first\nlanguage interacting with adults, this paper describes a new Teachable AI\nsystem that is capable of learning new language nuggets called concepts,\ndirectly from end users using live interactive teaching sessions. The proposed\nsetup uses three models to: a) Identify gaps in understanding automatically\nduring live conversational interactions, b) Learn the respective\ninterpretations of such unknown concepts from live interactions with users, and\nc) Manage a classroom sub-dialogue specifically tailored for interactive\nteaching sessions. We propose state-of-the-art transformer based neural\narchitectures of models, fine-tuned on top of pre-trained models, and show\naccuracy improvements on the respective components. We demonstrate that this\nmethod is very promising in leading way to build more adaptive and personalized\nlanguage understanding models.', 'When experiencing an information need, users want to engage with a domain\nexpert, but often turn to an information retrieval system, such as a search\nengine, instead. Classical information retrieval systems do not answer\ninformation needs directly, but instead provide references to (hopefully\nauthoritative) answers. Successful question answering systems offer a limited\ncorpus created on-demand by human experts, which is neither timely nor\nscalable. Pre-trained language models, by contrast, are capable of directly\ngenerating prose that may be responsive to an information need, but at present\nthey are dilettantes rather than domain experts -- they do not have a true\nunderstanding of the world, they are prone to hallucinating, and crucially they\nare incapable of justifying their utterances by referring to supporting\ndocuments in the corpus they were trained over. This paper examines how ideas\nfrom classical information retrieval and pre-trained language models can be\nsynthesized and evolved into systems that truly deliver on the promise of\ndomain expert advice.', 'The encoder-decoder framework achieves state-of-the-art results in keyphrase\ngeneration (KG) tasks by predicting both present keyphrases that appear in the\nsource document and absent keyphrases that do not. However, relying solely on\nthe source document can result in generating uncontrollable and inaccurate\nabsent keyphrases. To address these problems, we propose a novel graph-based\nmethod that can capture explicit knowledge from related references. Our model\nfirst retrieves some document-keyphrases pairs similar to the source document\nfrom a pre-defined index as references. Then a heterogeneous graph is\nconstructed to capture relationships of different granularities between the\nsource document and its references. To guide the decoding process, a\nhierarchical attention and copy mechanism is introduced, which directly copies\nappropriate words from both the source document and its references based on\ntheir relevance and significance. The experimental results on multiple KG\nbenchmarks show that the proposed model achieves significant improvements\nagainst other baseline models, especially with regard to the absent keyphrase\nprediction.', 'Visual reasoning with compositional natural language instructions, e.g.,\nbased on the newly-released Cornell Natural Language Visual Reasoning (NLVR)\ndataset, is a challenging task, where the model needs to have the ability to\ncreate an accurate mapping between the diverse phrases and the several objects\nplaced in complex arrangements in the image. Further, this mapping needs to be\nprocessed to answer the question in the statement given the ordering and\nrelationship of the objects across three similar images. In this paper, we\npropose a novel end-to-end neural model for the NLVR task, where we first use\njoint bidirectional attention to build a two-way conditioning between the\nvisual information and the language phrases. Next, we use an RL-based pointer\nnetwork to sort and process the varying number of unordered objects (so as to\nmatch the order of the statement phrases) in each of the three images and then\npool over the three decisions. Our model achieves strong improvements (of 4-6%\nabsolute) over the state-of-the-art on both the structured representation and\nraw image versions of the dataset.', ""With the ever-increasing popularity and applications of graph neural\nnetworks, several proposals have been made to explain and understand the\ndecisions of a graph neural network. Explanations for graph neural networks\ndiffer in principle from other input settings. It is important to attribute the\ndecision to input features and other related instances connected by the graph\nstructure. We find that the previous explanation generation approaches that\nmaximize the mutual information between the label distribution produced by the\nmodel and the explanation to be restrictive. Specifically, existing approaches\ndo not enforce explanations to be valid, sparse, or robust to input\nperturbations. In this paper, we lay down some of the fundamental principles\nthat an explanation method for graph neural networks should follow and\nintroduce a metric RDT-Fidelity as a measure of the explanation's\neffectiveness. We propose a novel approach Zorro based on the principles from\nrate-distortion theory that uses a simple combinatorial procedure to optimize\nfor RDT-Fidelity. Extensive experiments on real and synthetic datasets reveal\nthat Zorro produces sparser, stable, and more faithful explanations than\nexisting graph neural network explanation approaches."", 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', ""Major advances have recently been made in merging language and vision\nrepresentations. But most tasks considered so far have confined themselves to\nthe processing of objects and lexicalised relations amongst objects (content\nwords). We know, however, that humans (even pre-school children) can abstract\nover raw data to perform certain types of higher-level reasoning, expressed in\nnatural language by function words. A case in point is given by their ability\nto learn quantifiers, i.e. expressions like 'few', 'some' and 'all'. From\nformal semantics and cognitive linguistics, we know that quantifiers are\nrelations over sets which, as a simplification, we can see as proportions. For\ninstance, in 'most fish are red', most encodes the proportion of fish which are\nred fish. In this paper, we study how well current language and vision\nstrategies model such relations. We show that state-of-the-art attention\nmechanisms coupled with a traditional linguistic formalisation of quantifiers\ngives best performance on the task. Additionally, we provide insights on the\nrole of 'gist' representations in quantification. A 'logical' strategy to\ntackle the task would be to first obtain a numerosity estimation for the two\ninvolved sets and then compare their cardinalities. We however argue that\nprecisely identifying the composition of the sets is not only beyond current\nstate-of-the-art models but perhaps even detrimental to a task that is most\nefficiently performed by refining the approximate numerosity estimator of the\nsystem."", 'The Visual Dialogue task requires an agent to engage in a conversation about\nan image with a human. It represents an extension of the Visual Question\nAnswering task in that the agent needs to answer a question about an image, but\nit needs to do so in light of the previous dialogue that has taken place. The\nkey challenge in Visual Dialogue is thus maintaining a consistent, and natural\ndialogue while continuing to answer questions correctly. We present a novel\napproach that combines Reinforcement Learning and Generative Adversarial\nNetworks (GANs) to generate more human-like responses to questions. The GAN\nhelps overcome the relative paucity of training data, and the tendency of the\ntypical MLE-based approach to generate overly terse answers. Critically, the\nGAN is tightly integrated into the attention mechanism that generates\nhuman-interpretable reasons for each answer. This means that the discriminative\nmodel of the GAN has the task of assessing whether a candidate answer is\ngenerated by a human or not, given the provided reason. This is significant\nbecause it drives the generative model to produce high quality answers that are\nwell supported by the associated reasoning. The method also generates the\nstate-of-the-art results on the primary benchmark.', ""The role of conversational assistants has become more prevalent in helping\npeople increase their productivity. Document-centered assistance, for example\nto help an individual quickly review a document, has seen less significant\nprogress, even though it has the potential to tremendously increase a user's\nproductivity. This type of document-centered assistance is the focus of this\npaper. Our contributions are three-fold: (1) We first present a survey to\nunderstand the space of document-centered assistance and the capabilities\npeople expect in this scenario. (2) We investigate the types of queries that\nusers will pose while seeking assistance with documents, and show that\ndocument-centered questions form the majority of these queries. (3) We present\na set of initial machine learned models that show that (a) we can accurately\ndetect document-centered questions, and (b) we can build reasonably accurate\nmodels for answering such questions. These positive results are encouraging,\nand suggest that even greater results may be attained with continued study of\nthis interesting and novel problem space. Our findings have implications for\nthe design of intelligent systems to support task completion via natural\ninteractions with documents."", ""In multiple-choice exams, students select one answer from among typically\nfour choices and can explain why they made that particular choice. Students are\ngood at understanding natural language questions and based on their domain\nknowledge can easily infer the question's answer by 'connecting the dots'\nacross various pertinent facts.\n  Considering automated reasoning for elementary science question answering, we\naddress the novel task of generating explanations for answers from\nhuman-authored facts. For this, we examine the practically scalable framework\nof feature-rich support vector machines leveraging domain-targeted,\nhand-crafted features. Explanations are created from a human-annotated set of\nnearly 5,000 candidate facts in the WorldTree corpus. Our aim is to obtain\nbetter matches for valid facts of an explanation for the correct answer of a\nquestion over the available fact candidates. To this end, our features offer a\ncomprehensive linguistic and semantic unification paradigm. The machine\nlearning problem is the preference ordering of facts, for which we test\npointwise regression versus pairwise learning-to-rank.\n  Our contributions are: (1) a case study in which two preference ordering\napproaches are systematically compared; (2) it is a practically competent\napproach that can outperform some variants of BERT-based reranking models; and\n(3) the human-engineered features make it an interpretable machine learning\nmodel for the task."", ""Building a machine learning driven spoken dialog system for goal-oriented\ninteractions involves careful design of intents and data collection along with\ndevelopment of intent recognition models and dialog policy learning algorithms.\nThe models should be robust enough to handle various user distractions during\nthe interaction flow and should steer the user back into an engaging\ninteraction for successful completion of the interaction. In this work, we have\ndesigned a goal-oriented interaction system where children can engage with\nagents for a series of interactions involving `Meet \\& Greet' and `Simon Says'\ngame play. We have explored various feature extractors and models for improved\nintent recognition and looked at leveraging previous user and system\ninteractions in novel ways with attention models. We have also looked at dialog\nadaptation methods for entrained response selection. Our bootstrapped models\nfrom limited training data perform better than many baseline approaches we have\nlooked at for intent recognition and dialog action prediction."", 'Neural network-based systems can now learn to locate the referents of words\nand phrases in images, answer questions about visual scenes, and execute\nsymbolic instructions as first-person actors in partially-observable worlds. To\nachieve this so-called grounded language learning, models must overcome\nchallenges that infants face when learning their first words. While it is\nnotable that models with no meaningful prior knowledge overcome these\nobstacles, researchers currently lack a clear understanding of how they do so,\na problem that we attempt to address in this paper. For maximum control and\ngenerality, we focus on a simple neural network-based language learning agent,\ntrained via policy-gradient methods, which can interpret single-word\ninstructions in a simulated 3D world. Whilst the goal is not to explicitly\nmodel infant word learning, we take inspiration from experimental paradigms in\ndevelopmental psychology and apply some of these to the artificial agent,\nexploring the conditions under which established human biases and learning\neffects emerge. We further propose a novel method for visualising semantic\nrepresentations in the agent.', 'Context-aware neural machine translation (NMT) is a promising direction to\nimprove the translation quality by making use of the additional context, e.g.,\ndocument-level translation, or having meta-information. Although there exist\nvarious architectures and analyses, the effectiveness of different\ncontext-aware NMT models is not well explored yet. This paper analyzes the\nperformance of document-level NMT models on four diverse domains with a varied\namount of parallel document-level bilingual data. We conduct a comprehensive\nset of experiments to investigate the impact of document-level NMT. We find\nthat there is no single best approach to document-level NMT, but rather that\ndifferent architectures come out on top on different tasks. Looking at\ntask-specific problems, such as pronoun resolution or headline translation, we\nfind improvements in the context-aware systems, even in cases where the\ncorpus-level metrics like BLEU show no significant improvement. We also show\nthat document-level back-translation significantly helps to compensate for the\nlack of document-level bi-texts.', ""The increasing incorporation of Artificial Intelligence in the form of\nautomated systems into decision-making procedures highlights not only the\nimportance of decision theory for automated systems but also the need for these\ndecision procedures to be explainable to the people involved in them.\nTraditional realist accounts of explanation, wherein explanation is a relation\nthat holds (or does not hold) eternally between an explanans and an\nexplanandum, are not adequate to account for the notion of explanation required\nfor artificial decision procedures. We offer an alternative account of\nexplanation as used in the context of automated decision-making that makes\nexplanation an epistemic phenomenon, and one that is dependent on context. This\naccount of explanation better accounts for the way that we talk about, and use,\nexplanations and derived concepts, such as `explanatory power', and also allows\nus to differentiate between reasons or causes on the one hand, which do not\nneed to have an epistemic aspect, and explanations on the other, which do have\nsuch an aspect. Against this theoretical backdrop we then review existing\napproaches to explanation in Artificial Intelligence and Machine Learning, and\nsuggest desiderata which truly explainable decision systems should fulfill."", 'Transportation services play a crucial part in the development of modern\nsmart cities. In particular, on-demand ridesharing services, which group\ntogether passengers with similar itineraries, are already operating in several\nmetropolitan areas. These services can be of significant social and\nenvironmental benefit, by reducing travel costs, road congestion and CO2\nemissions.\n  Unfortunately, despite their advantages, not many people opt to use these\nridesharing services. We believe that increasing the user satisfaction from the\nservice will cause more people to utilize it, which, in turn, will improve the\nquality of the service, such as the waiting time, cost, travel time, and\nservice availability. One possible way for increasing user satisfaction is by\nproviding appropriate explanations comparing the alternative modes of\ntransportation, such as a private taxi ride and public transportation. For\nexample, a passenger may be more satisfied from a shared-ride if she is told\nthat a private taxi ride would have cost her 50% more. Therefore, the problem\nis to develop an agent that provides explanations that will increase the user\nsatisfaction.\n  We model our environment as a signaling game and show that a rational agent,\nwhich follows the perfect Bayesian equilibrium, must reveal all of the\ninformation regarding the possible alternatives to the passenger. In addition,\nwe develop a machine learning based agent that, when given a shared-ride along\nwith its possible alternatives, selects the explanations that are most likely\nto increase user satisfaction. Using feedback from humans we show that our\nmachine learning based agent outperforms the rational agent and an agent that\nrandomly chooses explanations, in terms of user satisfaction.', ""Explainable machine learning attracts increasing attention as it improves\ntransparency of models, which is helpful for machine learning to be trusted in\nreal applications. However, explanation methods have recently been demonstrated\nto be vulnerable to manipulation, where we can easily change a model's\nexplanation while keeping its prediction constant. To tackle this problem, some\nefforts have been paid to use more stable explanation methods or to change\nmodel configurations. In this work, we tackle the problem from the training\nperspective, and propose a new training scheme called Adversarial Training on\nEXplanations (ATEX) to improve the internal explanation stability of a model\nregardless of the specific explanation method being applied. Instead of\ndirectly specifying explanation values over data instances, ATEX only puts\nrequirement on model predictions which avoids involving second-order\nderivatives in optimization. As a further discussion, we also find that\nexplanation stability is closely related to another property of the model,\ni.e., the risk of being exposed to adversarial attack. Through experiments,\nbesides showing that ATEX improves model robustness against manipulation\ntargeting explanation, it also brings additional benefits including smoothing\nexplanations and improving the efficacy of adversarial training if applied to\nthe model."", 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', 'Future coalition operations can be substantially augmented through agile\nteaming between human and machine agents, but in a coalition context these\nagents may be unfamiliar to the human users and expected to operate in a broad\nset of scenarios rather than being narrowly defined for particular purposes. In\nsuch a setting it is essential that the human agents can rapidly build trust in\nthe machine agents through appropriate transparency of their behaviour, e.g.,\nthrough explanations. The human agents are also able to bring their local\nknowledge to the team, observing the situation unfolding and deciding which key\ninformation should be communicated to the machine agents to enable them to\nbetter account for the particular environment. In this paper we describe the\ninitial steps towards this human-agent knowledge fusion (HAKF) environment\nthrough a recap of the key requirements, and an explanation of how these can be\nfulfilled for an example situation. We show how HAKF has the potential to bring\nvalue to both human and machine agents working as part of a distributed\ncoalition team in a complex event processing setting with uncertain sources.', 'Issues regarding explainable AI involve four components: users, laws &\nregulations, explanations and algorithms. Together these components provide a\ncontext in which explanation methods can be evaluated regarding their adequacy.\nThe goal of this chapter is to bridge the gap between expert users and lay\nusers. Different kinds of users are identified and their concerns revealed,\nrelevant statements from the General Data Protection Regulation are analyzed in\nthe context of Deep Neural Networks (DNNs), a taxonomy for the classification\nof existing explanation methods is introduced, and finally, the various classes\nof explanation methods are analyzed to verify if user concerns are justified.\nOverall, it is clear that (visual) explanations can be given about various\naspects of the influence of the input on the output. However, it is noted that\nexplanation methods or interfaces for lay users are missing and we speculate\nwhich criteria these methods / interfaces should satisfy. Finally it is noted\nthat two important concerns are difficult to address with explanation methods:\nthe concern about bias in datasets that leads to biased DNNs, as well as the\nsuspicion about unfair outcomes.', 'The complexity of state-of-the-art modeling techniques for image\nclassification impedes the ability to explain model predictions in an\ninterpretable way. Existing explanation methods generally create importance\nrankings in terms of pixels or pixel groups. However, the resulting\nexplanations lack an optimal size, do not consider feature dependence and are\nonly related to one class. Counterfactual explanation methods are considered\npromising to explain complex model decisions, since they are associated with a\nhigh degree of human interpretability. In this paper, SEDC is introduced as a\nmodel-agnostic instance-level explanation method for image classification to\nobtain visual counterfactual explanations. For a given image, SEDC searches a\nsmall set of segments that, in case of removal, alters the classification. As\nimage classification tasks are typically multiclass problems, SEDC-T is\nproposed as an alternative method that allows specifying a target\ncounterfactual class. We compare SEDC(-T) with popular feature importance\nmethods such as LRP, LIME and SHAP, and we describe how the mentioned\nimportance ranking issues are addressed. Moreover, concrete examples and\nexperiments illustrate the potential of our approach (1) to obtain trust and\ninsight, and (2) to obtain input for model improvement by explaining\nmisclassifications.', 'In this document I present an approach to answer validation and reranking for\nquestion answering (QA) systems. A cased-based reasoning (CBR) system judges\nanswer candidates for questions from annotated answer candidates for earlier\nquestions. The promise of this approach is that user feedback will result in\nimproved answers of the QA system, due to the growing case base. In the paper,\nI present the adequate structuring of the case base and the appropriate\nselection of relevant similarity measures, in order to solve the answer\nvalidation problem. The structural case base is built from annotated MultiNet\ngraphs, which provide representations for natural language expressions, and\ncorresponding graph similarity measures. I cover a priori relations to\nexperienced answer candidates for former questions. I compare the CBR System\nresults to current approaches in an experiment integrating CBR into an existing\nframework for answer validation and reranking. This integration is achieved by\nadding CBR-related features to the input of a learned ranking model that\ndetermines the final answer ranking. In the experiments based on QA@CLEF\nquestions, the best learned models make heavy use of CBR features. Observing\nthe results with a continually growing case base, I present a positive effect\nof the size of the case base on the accuracy of the CBR subsystem.', 'With the ever-increasing use of complex machine learning models in critical\napplications within the finance domain, explaining the decisions of the model\nhas become a necessity. With applications spanning from credit scoring to\ncredit marketing, the impact of these models is undeniable. Among the multiple\nways in which one can explain the decisions of these complicated models, local\npost hoc model agnostic explanations have gained massive adoption. These\nmethods allow one to explain each prediction independent of the modelling\ntechnique that was used while training. As explanations, they either give\nindividual feature attributions or provide sufficient rules that represent\nconditions for a prediction to be made. The current state of the art methods\nuse rudimentary methods to generate synthetic data around the point to be\nexplained. This is followed by fitting simple linear models as surrogates to\nobtain a local interpretation of the prediction. In this paper, we seek to\nsignificantly improve on both, the method used to generate the explanations and\nthe nature of explanations produced. We use a Generative Adversarial Network\nfor synthetic data generation and train a piecewise linear model in the form of\nLinear Model Trees to be used as the surrogate model.In addition to individual\nfeature attributions, we also provide an accompanying context to our\nexplanations by leveraging the structure and property of our surrogate model.', ""Generating explanation to explain its behavior is an essential capability for\na robotic teammate. Explanations help human partners better understand the\nsituation and maintain trust of their teammates. Prior work on robot generating\nexplanations focuses on providing the reasoning behind its decision making.\nThese approaches, however, fail to heed the cognitive requirement of\nunderstanding an explanation. In other words, while they provide the right\nexplanations from the explainer's perspective, the explainee part of the\nequation is ignored. In this work, we address an important aspect along this\ndirection that contributes to a better understanding of a given explanation,\nwhich we refer to as the progressiveness of explanations. A progressive\nexplanation improves understanding by limiting the cognitive effort required at\neach step of making the explanation. As a result, such explanations are\nexpected to be smoother and hence easier to understand. A general formulation\nof progressive explanation is presented. Algorithms are provided based on\nseveral alternative quantifications of cognitive effort as an explanation is\nbeing made, which are evaluated in a standard planning competition domain."", 'Problems at the intersection of vision and language are of significant\nimportance both as challenging research questions and for the rich set of\napplications they enable. However, inherent structure in our world and bias in\nour language tend to be a simpler signal for learning than visual modalities,\nresulting in models that ignore visual information, leading to an inflated\nsense of their capability.\n  We propose to counter these language priors for the task of Visual Question\nAnswering (VQA) and make vision (the V in VQA) matter! Specifically, we balance\nthe popular VQA dataset by collecting complementary images such that every\nquestion in our balanced dataset is associated with not just a single image,\nbut rather a pair of similar images that result in two different answers to the\nquestion. Our dataset is by construction more balanced than the original VQA\ndataset and has approximately twice the number of image-question pairs. Our\ncomplete balanced dataset is publicly available at www.visualqa.org as part of\nthe 2nd iteration of the Visual Question Answering Dataset and Challenge (VQA\nv2.0).\n  We further benchmark a number of state-of-art VQA models on our balanced\ndataset. All models perform significantly worse on our balanced dataset,\nsuggesting that these models have indeed learned to exploit language priors.\nThis finding provides the first concrete empirical evidence for what seems to\nbe a qualitative sense among practitioners.\n  Finally, our data collection protocol for identifying complementary images\nenables us to develop a novel interpretable model, which in addition to\nproviding an answer to the given (image, question) pair, also provides a\ncounter-example based explanation. Specifically, it identifies an image that is\nsimilar to the original image, but it believes has a different answer to the\nsame question. This can help in building trust for machines among their users.', 'Recent years have seen the introduction of a range of methods for post-hoc\nexplainability of image classifier predictions. However, these post-hoc\nexplanations may not always be faithful to classifier predictions, which poses\na significant challenge when attempting to debug models based on such\nexplanations. To this end, we seek a methodology that can improve the\nfaithfulness of an explanation method with respect to model predictions which\ndoes not require ground truth explanations. We achieve this through a novel\nexplanation-driven data augmentation (EDDA) technique that augments the\ntraining data with occlusions inferred from model explanations; this is based\non the simple motivating principle that \\emph{if} the explainer is faithful to\nthe model \\emph{then} occluding salient regions for the model prediction should\ndecrease the model confidence in the prediction, while occluding non-salient\nregions should not change the prediction. To verify that the proposed\naugmentation method has the potential to improve faithfulness, we evaluate EDDA\nusing a variety of datasets and classification models. We demonstrate\nempirically that our approach leads to a significant increase of faithfulness,\nwhich can facilitate better debugging and successful deployment of image\nclassification models in real-world applications.', ""Our research is a step toward ascertaining the need for personalization, in\nXAI, and we do so in the context of investigating the value of explanations of\nAI-driven hints and feedback are useful in Intelligent Tutoring Systems (ITS).\nWe added an explanation functionality for the adaptive hints provided by the\nAdaptive CSP (ACSP) applet, an interactive simulation that helps students learn\nan algorithm for constraint satisfaction problems by providing AI-driven hints\nadapted to their predicted level of learning. We present the design of the\nexplanation functionality and the results of a controlled study to evaluate its\nimpact on students' learning and perception of the ACPS hints. The study\nincludes an analysis of how these outcomes are modulated by several user\ncharacteristics such as personality traits and cognitive abilities, to asses if\nexplanations should be personalized to these characteristics. Our results\nindicate that providing explanations increase students' trust in the ACPS\nhints, perceived usefulness of the hints, and intention to use them again. In\naddition, we show that students' access of the explanation and learning gains\nare modulated by user characteristics, providing insights toward designing\npersonalized Explainable AI (XAI) for ITS."", 'In the last years many accurate decision support systems have been\nconstructed as black boxes, that is as systems that hide their internal logic\nto the user. This lack of explanation constitutes both a practical and an\nethical issue. The literature reports many approaches aimed at overcoming this\ncrucial weakness sometimes at the cost of scarifying accuracy for\ninterpretability. The applications in which black box decision systems can be\nused are various, and each approach is typically developed to provide a\nsolution for a specific problem and, as a consequence, delineating explicitly\nor implicitly its own definition of interpretability and explanation. The aim\nof this paper is to provide a classification of the main problems addressed in\nthe literature with respect to the notion of explanation and the type of black\nbox system. Given a problem definition, a black box type, and a desired\nexplanation this survey should help the researcher to find the proposals more\nuseful for his own work. The proposed classification of approaches to open\nblack box models should also be useful for putting the many research open\nquestions in perspective.', 'We present a randomized controlled trial for a model-in-the-loop regression\ntask, with the goal of measuring the extent to which (1) good explanations of\nmodel predictions increase human accuracy, and (2) faulty explanations decrease\nhuman trust in the model. We study explanations based on visual saliency in an\nimage-based age prediction task for which humans and learned models are\nindividually capable but not highly proficient and frequently disagree. Our\nexperimental design separates model quality from explanation quality, and makes\nit possible to compare treatments involving a variety of explanations of\nvarying levels of quality. We find that presenting model predictions improves\nhuman accuracy. However, visual explanations of various kinds fail to\nsignificantly alter human accuracy or trust in the model - regardless of\nwhether explanations characterize an accurate model, an inaccurate one, or are\ngenerated randomly and independently of the input image. These findings suggest\nthe need for greater evaluation of explanations in downstream decision making\ntasks, better design-based tools for presenting explanations to users, and\nbetter approaches for generating explanations.', 'In automated planning, the need for explanations arises when there is a\nmismatch between a proposed plan and the user\'s expectation. We frame\nExplainable AI Planning in the context of the plan negotiation problem, in\nwhich a succession of hypothetical planning problems are generated and solved.\nThe object of the negotiation is for the user to understand and ultimately\narrive at a satisfactory plan. We present the results of a user study that\ndemonstrates that when users ask questions about plans, those questions are\ncontrastive, i.e. ""why A rather than B?"". We use the data from this study to\nconstruct a taxonomy of user questions that often arise during plan\nnegotiation. We formally define our approach to plan negotiation through model\nrestriction as an iterative process. This approach generates hypothetical\nproblems and contrastive plans by restricting the model through constraints\nimplied by user questions. We formally define model-based compilations in\nPDDL2.1 of each constraint derived from a user question in the taxonomy, and\nempirically evaluate the compilations in terms of computational complexity. The\ncompilations were implemented as part of an explanation framework that employs\niterative model restriction. We demonstrate its benefits in a second user\nstudy.', ""A feature-based model explanation denotes how much each input feature\ncontributes to a model's output for a given data point. As the number of\nproposed explanation functions grows, we lack quantitative evaluation criteria\nto help practitioners know when to use which explanation function. This paper\nproposes quantitative evaluation criteria for feature-based explanations: low\nsensitivity, high faithfulness, and low complexity. We devise a framework for\naggregating explanation functions. We develop a procedure for learning an\naggregate explanation function with lower complexity and then derive a new\naggregate Shapley value explanation function that minimizes sensitivity."", ""Theoretically as well as experimentally it is investigated how people\nrepresent their knowledge in order to make decisions or to share their\nknowledge with others. Experiment 1 probes into the ways how people 6ather\ninformation about the frequencies of events and how the requested response\nmode, that is, numerical vs. verbal estimates interferes with this knowledge.\nThe least interference occurs if the subjects are allowed to give verbal\nresponses. From this it is concluded that processing knowledge about\nuncertainty categorically, that is, by means of verbal expressions, imposes\nless mental work load on the decision matter than numerical processing.\nPossibility theory is used as a framework for modeling the individual usage of\nverbal categories for grades of uncertainty. The 'elastic' constraints on the\nverbal expressions for every sing1e subject are determined in Experiment 2 by\nmeans of sequential calibration. In further experiments it is shown that the\nsuperiority of the verbal processing of knowledge about uncertainty guise\ngenerally reduces persistent biases reported in the literature: conservatism\n(Experiment 3) and neg1igence of regression (Experiment 4). The reanalysis of\nHormann's data reveal that in verbal Judgments people exhibit sensitivity for\nbase rates and are not prone to the conjunction fallacy. In a final experiment\n(5) about predictions in a real-life situation it turns out that in a numerical\nforecasting task subjects restricted themselves to those parts of their\nknowledge which are numerical. On the other hand subjects in a verbal\nforecasting task accessed verbally as well as numerically stated knowledge.\nForecasting is structurally related to the estimation of probabilities for rare\nevents insofar as supporting and contradicting arguments have to be evaluated\nand the choice of the final Judgment has to be Justified according to the\nevidence brought forward. In order to assist people in such choice situations a\nformal model for the interactive checking of arguments has been developed. The\nmodel transforms the normal-language quantifiers used in the arguments into\nfuzzy numbers and evaluates the given train of arguments by means of fuzzy\nnumerica1 operations. Ambiguities in the meanings of quantifiers are resolved\ninteractively."", 'To increase trust in artificial intelligence systems, a promising research\ndirection consists of designing neural models capable of generating natural\nlanguage explanations for their predictions. In this work, we show that such\nmodels are nonetheless prone to generating mutually inconsistent explanations,\nsuch as ""Because there is a dog in the image"" and ""Because there is no dog in\nthe [same] image"", exposing flaws in either the decision-making process of the\nmodel or in the generation of the explanations. We introduce a simple yet\neffective adversarial framework for sanity checking models against the\ngeneration of inconsistent natural language explanations. Moreover, as part of\nthe framework, we address the problem of adversarial attacks with full target\nsequences, a scenario that was not previously addressed in sequence-to-sequence\nattacks. Finally, we apply our framework on a state-of-the-art neural natural\nlanguage inference model that provides natural language explanations for its\npredictions. Our framework shows that this model is capable of generating a\nsignificant number of inconsistent explanations.', 'The main objective of eXplainable Artificial Intelligence (XAI) is to provide\neffective explanations for black-box classifiers. The existing literature lists\nmany desirable properties for explanations to be useful, but there is no\nconsensus on how to quantitatively evaluate explanations in practice. Moreover,\nexplanations are typically used only to inspect black-box models, and the\nproactive use of explanations as a decision support is generally overlooked.\nAmong the many approaches to XAI, a widely adopted paradigm is Local Linear\nExplanations - with LIME and SHAP emerging as state-of-the-art methods. We show\nthat these methods are plagued by many defects including unstable explanations,\ndivergence of actual implementations from the promised theoretical properties,\nand explanations for the wrong label. This highlights the need to have standard\nand unbiased evaluation procedures for Local Linear Explanations in the XAI\nfield. In this paper we address the problem of identifying a clear and\nunambiguous set of metrics for the evaluation of Local Linear Explanations.\nThis set includes both existing and novel metrics defined specifically for this\nclass of explanations. All metrics have been included in an open Python\nframework, named LEAF. The purpose of LEAF is to provide a reference for end\nusers to evaluate explanations in a standardised and unbiased way, and to guide\nresearchers towards developing improved explainable techniques.', 'Autonomous agents situated in real-world environments must be able to master\nlarge repertoires of skills. While a single short skill can be learned quickly,\nit would be impractical to learn every task independently. Instead, the agent\nshould share knowledge across behaviors such that each task can be learned\nefficiently, and such that the resulting model can generalize to new tasks,\nespecially ones that are compositions or subsets of tasks seen previously. A\npolicy conditioned on a goal or demonstration has the potential to share\nknowledge between tasks if it sees enough diversity of inputs. However, these\nmethods may not generalize to a more complex task at test time. We introduce\ncompositional plan vectors (CPVs) to enable a policy to perform compositions of\ntasks without additional supervision. CPVs represent trajectories as the sum of\nthe subtasks within them. We show that CPVs can be learned within a one-shot\nimitation learning framework without any additional supervision or information\nabout task hierarchy, and enable a demonstration-conditioned policy to\ngeneralize to tasks that sequence twice as many skills as the tasks seen during\ntraining.\n  Analogously to embeddings such as word2vec in NLP, CPVs can also support\nsimple arithmetic operations -- for example, we can add the CPVs for two\ndifferent tasks to command an agent to compose both tasks, without any\nadditional training.', ""Local explanation methods such as LIME have become popular in MIR as tools\nfor generating post-hoc, model-agnostic explanations of a model's\nclassification decisions. The basic idea is to identify a small set of\nhuman-understandable features of the classified example that are most\ninfluential on the classifier's prediction. These are then presented as an\nexplanation. Evaluation of such explanations in publications often resorts to\naccepting what matches the expectation of a human without actually being able\nto verify if what the explanation shows is what really caused the model's\nprediction. This paper reports on targeted investigations where we try to get\nmore insight into the actual veracity of LIME's explanations in an audio\nclassification task. We deliberately design adversarial examples for the\nclassifier, in a way that gives us knowledge about which parts of the input are\npotentially responsible for the model's (wrong) prediction. Asking LIME to\nexplain the predictions for these adversaries permits us to study whether local\nexplanations do indeed detect these regions of interest. We also look at\nwhether LIME is more successful in finding perturbations that are more\nprominent and easily noticeable for a human. Our results suggest that LIME does\nnot necessarily manage to identify the most relevant input features and hence\nit remains unclear whether explanations are useful or even misleading."", 'The widespread adoption of black-box models in Artificial Intelligence has\nenhanced the need for explanation methods to reveal how these obscure models\nreach specific decisions. Retrieving explanations is fundamental to unveil\npossible biases and to resolve practical or ethical issues. Nowadays, the\nliterature is full of methods with different explanations. We provide a\ncategorization of explanation methods based on the type of explanation\nreturned. We present the most recent and widely used explainers, and we show a\nvisual comparison among explanations and a quantitative benchmarking.', 'eXplainable AI focuses on generating explanations for the output of an AI\nalgorithm to a user, usually a decision-maker. Such user needs to interpret the\nAI system in order to decide whether to trust the machine outcome. When\naddressing this challenge, therefore, proper attention should be given to\nproduce explanations that are interpretable by the target community of users.\nIn this chapter, we claim for the need to better investigate what constitutes a\nhuman explanation, i.e. a justification of the machine behaviour that is\ninterpretable and actionable by the human decision makers. In particular, we\nfocus on the contributions that Human Intelligence can bring to eXplainable AI,\nespecially in conjunction with the exploitation of Knowledge Graphs. Indeed, we\ncall for a better interplay between Knowledge Representation and Reasoning,\nSocial Sciences, Human Computation and Human-Machine Cooperation research -- as\nalready explored in other AI branches -- in order to support the goal of\neXplainable AI with the adoption of a Human-in-the-Loop approach.', 'Owing to the rapidly growing multimedia content available on the Internet,\nextractive spoken document summarization, with the purpose of automatically\nselecting a set of representative sentences from a spoken document to concisely\nexpress the most important theme of the document, has been an active area of\nresearch and experimentation. On the other hand, word embedding has emerged as\na newly favorite research subject because of its excellent performance in many\nnatural language processing (NLP)-related tasks. However, as far as we are\naware, there are relatively few studies investigating its use in extractive\ntext or speech summarization. A common thread of leveraging word embeddings in\nthe summarization process is to represent the document (or sentence) by\naveraging the word embeddings of the words occurring in the document (or\nsentence). Then, intuitively, the cosine similarity measure can be employed to\ndetermine the relevance degree between a pair of representations. Beyond the\ncontinued efforts made to improve the representation of words, this paper\nfocuses on building novel and efficient ranking models based on the general\nword embedding methods for extractive speech summarization. Experimental\nresults demonstrate the effectiveness of our proposed methods, compared to\nexisting state-of-the-art methods.', 'We administered the Verbal IQ (VIQ) part of the Wechsler Preschool and\nPrimary Scale of Intelligence (WPPSI-III) to the ConceptNet 4 AI system. The\ntest questions (e.g., ""Why do we shake hands?"") were translated into ConceptNet\n4 inputs using a combination of the simple natural language processing tools\nthat come with ConceptNet together with short Python programs that we wrote.\nThe question answering used a version of ConceptNet based on spectral methods.\nThe ConceptNet system scored a WPPSI-III VIQ that is average for a\nfour-year-old child, but below average for 5 to 7 year-olds. Large variations\namong subtests indicate potential areas of improvement. In particular, results\nwere strongest for the Vocabulary and Similarities subtests, intermediate for\nthe Information subtest, and lowest for the Comprehension and Word Reasoning\nsubtests. Comprehension is the subtest most strongly associated with common\nsense. The large variations among subtests and ordinary common sense strongly\nsuggest that the WPPSI-III VIQ results do not show that ""ConceptNet has the\nverbal abilities a four-year-old."" Rather, children\'s IQ tests offer one\nobjective metric for the evaluation and comparison of AI systems. Also, this\nwork continues previous research on Psychometric AI.', 'After learning a concept, humans are also able to continually generalize\ntheir learned concepts to new domains by observing only a few labeled instances\nwithout any interference with the past learned knowledge. In contrast, learning\nconcepts efficiently in a continual learning setting remains an open challenge\nfor current Artificial Intelligence algorithms as persistent model retraining\nis necessary. Inspired by the Parallel Distributed Processing learning and the\nComplementary Learning Systems theories, we develop a computational model that\nis able to expand its previously learned concepts efficiently to new domains\nusing a few labeled samples. We couple the new form of a concept to its past\nlearned forms in an embedding space for effective continual learning. Doing so,\na generative distribution is learned such that it is shared across the tasks in\nthe embedding space and models the abstract concepts. This procedure enables\nthe model to generate pseudo-data points to replay the past experience to\ntackle catastrophic forgetting.', 'Given questions regarding some prototypical situation such as Name something\nthat people usually do before they leave the house for work? a human can easily\nanswer them via acquired experiences. There can be multiple right answers for\nsuch questions, with some more common for a situation than others. This paper\nintroduces a new question answering dataset for training and evaluating common\nsense reasoning capabilities of artificial intelligence systems in such\nprototypical situations. The training set is gathered from an existing set of\nquestions played in a long-running international game show FAMILY- FEUD. The\nhidden evaluation set is created by gathering answers for each question from\n100 crowd-workers. We also propose a generative evaluation task where a model\nhas to output a ranked list of answers, ideally covering all prototypical\nanswers for a question. After presenting multiple competitive baseline models,\nwe find that human performance still exceeds model scores on all evaluation\nmetrics with a meaningful gap, supporting the challenging nature of the task.', ""Common sense has always been of interest in Artificial Intelligence, but has\nrarely taken center stage. Despite its mention in one of John McCarthy's\nearliest papers and years of work by dedicated researchers, arguably no AI\nsystem with a serious amount of general common sense has ever emerged. Why is\nthat? What's missing? Examples of AI systems' failures of common sense abound,\nand they point to AI's frequent focus on expertise as the cause. Those\nattempting to break the resulting brittleness barrier, even in the context of\nmodern deep learning, have tended to invest their energy in large numbers of\nsmall bits of commonsense knowledge. While important, all the commonsense\nknowledge fragments in the world don't add up to a system that actually\ndemonstrates common sense in a human-like way. We advocate examining common\nsense from a broader perspective than in the past. Common sense should be\nconsidered in the context of a full cognitive system with history, goals,\ndesires, and drives, not just in isolated circumscribed examples. A fresh look\nis needed: common sense is worthy of its own dedicated scientific exploration."", 'Cross-lingual document representations enable language understanding in\nmultilingual contexts and allow transfer learning from high-resource to\nlow-resource languages at the document level. Recently large pre-trained\nlanguage models such as BERT, XLM and XLM-RoBERTa have achieved great success\nwhen fine-tuned on sentence-level downstream tasks. It is tempting to apply\nthese cross-lingual models to document representation learning. However, there\nare two challenges: (1) these models impose high costs on long document\nprocessing and thus many of them have strict length limit; (2) model\nfine-tuning requires extra data and computational resources, which is not\npractical in resource-limited settings. In this work, we address these\nchallenges by proposing unsupervised Language-Agnostic Weighted Document\nRepresentations (LAWDR). We study the geometry of pre-trained sentence\nembeddings and leverage it to derive document representations without\nfine-tuning. Evaluated on cross-lingual document alignment, LAWDR demonstrates\ncomparable performance to state-of-the-art models on benchmark datasets.', 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', ""Neural networks with deep architectures have demonstrated significant\nperformance improvements in computer vision, speech recognition, and natural\nlanguage processing. The challenges in information retrieval (IR), however, are\ndifferent from these other application areas. A common form of IR involves\nranking of documents--or short passages--in response to keyword-based queries.\nEffective IR systems must deal with query-document vocabulary mismatch problem,\nby modeling relationships between different query and document terms and how\nthey indicate relevance. Models should also consider lexical matches when the\nquery contains rare terms--such as a person's name or a product model\nnumber--not seen during training, and to avoid retrieving semantically related\nbut irrelevant results. In many real-life IR tasks, the retrieval involves\nextremely large collections--such as the document index of a commercial Web\nsearch engine--containing billions of documents. Efficient IR methods should\ntake advantage of specialized IR data structures, such as inverted index, to\nefficiently retrieve from large collections. Given an information need, the IR\nsystem also mediates how much exposure an information artifact receives by\ndeciding whether it should be displayed, and where it should be positioned,\namong other results. Exposure-aware IR systems may optimize for additional\nobjectives, besides relevance, such as parity of exposure for retrieved items\nand content publishers. In this thesis, we present novel neural architectures\nand methods motivated by the specific needs and challenges of IR tasks."", 'There has been much discussion of the right to explanation in the EU General\nData Protection Regulation, and its existence, merits, and disadvantages.\nImplementing a right to explanation that opens the black box of algorithmic\ndecision-making faces major legal and technical barriers. Explaining the\nfunctionality of complex algorithmic decision-making systems and their\nrationale in specific cases is a technically challenging problem. Some\nexplanations may offer little meaningful information to data subjects, raising\nquestions around their value. Explanations of automated decisions need not\nhinge on the general public understanding how algorithmic systems function.\nEven though such interpretability is of great importance and should be pursued,\nexplanations can, in principle, be offered without opening the black box.\nLooking at explanations as a means to help a data subject act rather than\nmerely understand, one could gauge the scope and content of explanations\naccording to the specific goal or action they are intended to support. From the\nperspective of individuals affected by automated decision-making, we propose\nthree aims for explanations: (1) to inform and help the individual understand\nwhy a particular decision was reached, (2) to provide grounds to contest the\ndecision if the outcome is undesired, and (3) to understand what would need to\nchange in order to receive a desired result in the future, based on the current\ndecision-making model. We assess how each of these goals finds support in the\nGDPR. We suggest data controllers should offer a particular type of\nexplanation, unconditional counterfactual explanations, to support these three\naims. These counterfactual explanations describe the smallest change to the\nworld that can be made to obtain a desirable outcome, or to arrive at the\nclosest possible world, without needing to explain the internal logic of the\nsystem.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'Recently, many AI researchers and practitioners have embarked on research\nvisions that involve doing AI for ""Good"". This is part of a general drive\ntowards infusing AI research and practice with ethical thinking. One frequent\ntheme in current ethical guidelines is the requirement that AI be good for all,\nor: contribute to the Common Good. But what is the Common Good, and is it\nenough to want to be good? Via four lead questions, I will illustrate\nchallenges and pitfalls when determining, from an AI point of view, what the\nCommon Good is and how it can be enhanced by AI. The questions are: What is the\nproblem / What is a problem?, Who defines the problem?, What is the role of\nknowledge?, and What are important side effects and dynamics? The illustration\nwill use an example from the domain of ""AI for Social Good"", more specifically\n""Data Science for Social Good"". Even if the importance of these questions may\nbe known at an abstract level, they do not get asked sufficiently in practice,\nas shown by an exploratory study of 99 contributions to recent conferences in\nthe field. Turning these challenges and pitfalls into a positive\nrecommendation, as a conclusion I will draw on another characteristic of\ncomputer-science thinking and practice to make these impediments visible and\nattenuate them: ""attacks"" as a method for improving design. This results in the\nproposal of ethics pen-testing as a method for helping AI designs to better\ncontribute to the Common Good.', 'Document clustering is a text mining technique used to provide better\ndocument search and browsing in digital libraries or online corpora. A lot of\nresearch has been done on biomedical document clustering that is based on using\nexisting ontology. But, associations and co-occurrences of the medical concepts\nare not well represented by using ontology. In this research, a vector\nrepresentation of concepts of diseases and similarity measurement between\nconcepts are proposed. They identify the closest concepts of diseases in the\ncontext of a corpus. Each document is represented by using the vector space\nmodel. A weight scheme is proposed to consider both local content and\nassociations between concepts. A Self-Organizing Map is used as document\nclustering algorithm. The vector projection and visualization features of SOM\nenable visualization and analysis of the clusters distributions and\nrelationships on the two dimensional space. The experimental results show that\nthe proposed document clustering framework generates meaningful clusters and\nfacilitate visualization of the clusters based on the concepts of diseases.', 'We evaluate four computational models of explanation in Bayesian networks by\ncomparing model predictions to human judgments. In two experiments, we present\nhuman participants with causal structures for which the models make divergent\npredictions and either solicit the best explanation for an observed event\n(Experiment 1) or have participants rate provided explanations for an observed\nevent (Experiment 2). Across two versions of two causal structures and across\nboth experiments, we find that the Causal Explanation Tree and Most Relevant\nExplanation models provide better fits to human data than either Most Probable\nExplanation or Explanation Tree models. We identify strengths and shortcomings\nof these models and what they can reveal about human explanation. We conclude\nby suggesting the value of pursuing computational and psychological\ninvestigations of explanation in parallel.', 'Relation extraction (RE) aims to identify the semantic relations between\nnamed entities in text. Recent years have witnessed it raised to the document\nlevel, which requires complex reasoning with entities and mentions throughout\nan entire document. In this paper, we propose a novel model to document-level\nRE, by encoding the document information in terms of entity global and local\nrepresentations as well as context relation representations. Entity global\nrepresentations model the semantic information of all entities in the document,\nentity local representations aggregate the contextual information of multiple\nmentions of specific entities, and context relation representations encode the\ntopic information of other relations. Experimental results demonstrate that our\nmodel achieves superior performance on two public datasets for document-level\nRE. It is particularly effective in extracting relations between entities of\nlong distance and having multiple mentions.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', ""Children learning their first language face multiple problems of induction:\nhow to learn the meanings of words, and how to build meaningful phrases from\nthose words according to syntactic rules. We consider how children might solve\nthese problems efficiently by solving them jointly, via a computational model\nthat learns the syntax and semantics of multi-word utterances in a grounded\nreference game. We select a well-studied empirical case in which children are\naware of patterns linking the syntactic and semantic properties of words ---\nthat the properties picked out by base nouns tend to be related to shape, while\nprenominal adjectives tend to refer to other properties such as color. We show\nthat children applying such inductive biases are accurately reflecting the\nstatistics of child-directed speech, and that inducing similar biases in our\ncomputational model captures children's behavior in a classic adjective\nlearning experiment. Our model incorporating such biases also demonstrates a\nclear data efficiency in learning, relative to a baseline model that learns\nwithout forming syntax-sensitive overhypotheses of word meaning. Thus solving a\nmore complex joint inference problem may make the full problem of language\nacquisition easier, not harder."", 'Document-level relation extraction aims to discover relations between\nentities across a whole document. How to build the dependency of entities from\ndifferent sentences in a document remains to be a great challenge. Current\napproaches either leverage syntactic trees to construct document-level graphs\nor aggregate inference information from different sentences. In this paper, we\nbuild cross-sentence dependencies by inferring compositional relations between\ninter-sentence mentions. Adopting aggressive linking strategy, intermediate\nrelations are reasoned on the document-level graphs by mention convolution. We\nfurther notice the generalization problem of NA instances, which is caused by\nincomplete annotation and worsened by fully-connected mention pairs. An\nimproved ranking loss is proposed to attend this problem. Experiments show the\nconnections between different mentions are crucial to document-level relation\nextraction, which enables the model to extract more meaningful higher-level\ncompositional relations.', 'The concept of innateness is rarely discussed in the context of artificial\nintelligence. When it is discussed, or hinted at, it is often the context of\ntrying to reduce the amount of innate machinery in a given system. In this\npaper, I consider as a test case a recent series of papers by Silver et al\n(Silver et al., 2017a) on AlphaGo and its successors that have been presented\nas an argument that a ""even in the most challenging of domains: it is possible\nto train to superhuman level, without human examples or guidance"", ""starting\ntabula rasa.""\n  I argue that these claims are overstated, for multiple reasons. I close by\narguing that artificial intelligence needs greater attention to innateness, and\nI point to some proposals about what that innateness might look like.', ""Probability theory, epistemically interpreted, provides an excellent, if not\nthe best available account of inductive reasoning. This is so because there are\ngeneral and definite rules for the change of subjective probabilities through\ninformation or experience; induction and belief change are one and same topic,\nafter all. The most basic of these rules is simply to conditionalize with\nrespect to the information received; and there are similar and more general\nrules. 1 Hence, a fundamental reason for the epistemological success of\nprobability theory is that there at all exists a well-behaved concept of\nconditional probability. Still, people have, and have reasons for, various\nconcerns over probability theory. One of these is my starting point:\nIntuitively, we have the notion of plain belief; we believe propositions2 to be\ntrue (or to be false or neither). Probability theory, however, offers no formal\ncounterpart to this notion. Believing A is not the same as having probability 1\nfor A, because probability 1 is incorrigible3; but plain belief is clearly\ncorrigible. And believing A is not the same as giving A a probability larger\nthan some 1 - c, because believing A and believing B is usually taken to be\nequivalent to believing A & B.4 Thus, it seems that the formal representation\nof plain belief has to take a non-probabilistic route. Indeed, representing\nplain belief seems easy enough: simply represent an epistemic state by the set\nof all propositions believed true in it or, since I make the common assumption\nthat plain belief is deductively closed, by the conjunction of all propositions\nbelieved true in it. But this does not yet provide a theory of induction, i.e.\nan answer to the question how epistemic states so represented are changed\ntbrough information or experience. There is a convincing partial answer: if the\nnew information is compatible with the old epistemic state, then the new\nepistemic state is simply represented by the conjunction of the new information\nand the old beliefs. This answer is partial because it does not cover the quite\ncommon case where the new information is incompatible with the old beliefs. It\nis, however, important to complete the answer and to cover this case, too;\notherwise, we would not represent plain belief as conigible. The crucial\nproblem is that there is no good completion. When epistemic states are\nrepresented simply by the conjunction of all propositions believed true in it,\nthe answer cannot be completed; and though there is a lot of fruitful work, no\nother representation of epistemic states has been proposed, as far as I know,\nwhich provides a complete solution to this problem. In this paper, I want to\nsuggest such a solution. In [4], I have more fully argued that this is the only\nsolution, if certain plausible desiderata are to be satisfied. Here, in section\n2, I will be content with formally defining and intuitively explaining my\nproposal. I will compare my proposal with probability theory in section 3. It\nwill turn out that the theory I am proposing is structurally homomorphic to\nprobability theory in important respects and that it is thus equally easily\nimplementable, but moreover computationally simpler. Section 4 contains a very\nbrief comparison with various kinds of logics, in particular conditional logic,\nwith Shackle's functions of potential surprise and related theories, and with\nthe Dempster - Shafer theory of belief functions."", 'The rise of artificial intelligence (A.I.) based systems is already offering\nsubstantial benefits to the society as a whole. However, these systems may also\nenclose potential conflicts and unintended consequences. Notably, people will\ntend to adopt an A.I. system if it confers them an advantage, at which point\nnon-adopters might push for a strong regulation if that advantage for adopters\nis at a cost for them. Here we propose an agent-based game-theoretical model\nfor these conflicts, where agents may decide to resort to A.I. to use and\nacquire additional information on the payoffs of a stochastic game, striving to\nbring insights from simulation to what has been, hitherto, a mostly\nphilosophical discussion. We frame our results under the current discussion on\nethical A.I. and the conflict between individual and societal gains: the\nsocietal value alignment problem. We test the arising equilibria in the\nadoption of A.I. technology under different norms followed by artificial\nagents, their ensuing benefits, and the emergent levels of wealth inequality.\nWe show that without any regulation, purely selfish A.I. systems will have the\nstrongest advantage, even when a utilitarian A.I. provides significant benefits\nfor the individual and the society. Nevertheless, we show that it is possible\nto develop A.I. systems following human conscious policies that, when\nintroduced in society, lead to an equilibrium where the gains for the adopters\nare not at a cost for non-adopters, thus increasing the overall wealth of the\npopulation and lowering inequality. However, as shown, a self-organised\nadoption of such policies would require external regulation.', 'Question answering and conversational systems are often baffled and need help\nclarifying certain ambiguities. However, limitations of existing datasets\nhinder the development of large-scale models capable of generating and\nutilising clarification questions. In order to overcome these limitations, we\ndevise a novel bootstrapping framework (based on self-supervision) that assists\nin the creation of a diverse, large-scale dataset of clarification questions\nbased on post-comment tuples extracted from stackexchange. The framework\nutilises a neural network based architecture for classifying clarification\nquestions. It is a two-step method where the first aims to increase the\nprecision of the classifier and second aims to increase its recall. We\nquantitatively demonstrate the utility of the newly created dataset by applying\nit to the downstream task of question-answering. The final dataset, ClarQ,\nconsists of ~2M examples distributed across 173 domains of stackexchange. We\nrelease this dataset in order to foster research into the field of\nclarification question generation with the larger goal of enhancing dialog and\nquestion answering systems.', 'We propose a method for online news stream clustering that is a variant of\nthe non-parametric streaming K-means algorithm. Our model uses a combination of\nsparse and dense document representations, aggregates document-cluster\nsimilarity along these multiple representations and makes the clustering\ndecision using a neural classifier. The weighted document-cluster similarity\nmodel is learned using a novel adaptation of the triplet loss into a linear\nclassification objective. We show that the use of a suitable fine-tuning\nobjective and external knowledge in pre-trained transformer models yields\nsignificant improvements in the effectiveness of contextual embeddings for\nclustering. Our model achieves a new state-of-the-art on a standard stream\nclustering dataset of English documents.', 'Transferring as fast as possible the functioning of our brain to artificial\nintelligence is an ambitious goal that would help advance the state of the art\nin AI and robotics. It is in this perspective that we propose to start from\nhypotheses derived from an empirical study in a human-robot interaction and to\nverify if they are validated in the same way for children as for a basic\nreinforcement learning algorithm. Thus, we check whether receiving help from an\nexpert when solving a simple close-ended task (the Towers of Hano\\""i) allows to\naccelerate or not the learning of this task, depending on whether the\nintervention is canonical or requested by the player. Our experiences have\nallowed us to conclude that, whether requested or not, a Q-learning algorithm\nbenefits in the same way from expert help as children do.', 'What is the nature of curiosity? Is there any scientific way to understand\nthe origin of this mysterious force that drives the behavior of even the\nstupidest naturally intelligent systems and is completely absent in their\nsmartest artificial analogs? Can we build AI systems that could be curious\nabout something, systems that would have an intrinsic motivation to learn? Is\nsuch a motivation quantifiable? Is it implementable? I will discuss this\nproblem from the standpoint of physics. The relationship between physics and\nintelligence is a consequence of the fact that correctly predicted information\nis nothing but an energy resource, and the process of thinking can be viewed as\na process of accumulating and spending this resource through the acts of\nperception and, respectively, decision making. The natural motivation of any\nautonomous system to keep this accumulation/spending balance as high as\npossible allows one to treat the problem of describing the dynamics of thinking\nprocesses as a resource optimization problem. Here I will propose and discuss a\nsimple theoretical model of such an autonomous system which I call the\nAutonomous Turing Machine (ATM). The potential attractiveness of ATM lies in\nthe fact that it is the model of a self-propelled AI for which the only\navailable energy resource is the information itself. For ATM, the problem of\noptimal thinking, learning, and decision-making becomes conceptually simple and\nmathematically well tractable. This circumstance makes the ATM an ideal\nplayground for studying the dynamics of intelligent behavior and allows one to\nquantify many seemingly unquantifiable features of genuine intelligence.', 'This paper summarizes our entries to both subtasks of the first DialDoc\nshared task which focuses on the agent response prediction task in\ngoal-oriented document-grounded dialogs. The task is split into two subtasks:\npredicting a span in a document that grounds an agent turn and generating an\nagent response based on a dialog and grounding document. In the first subtask,\nwe restrict the set of valid spans to the ones defined in the dataset, use a\nbiaffine classifier to model spans, and finally use an ensemble of different\nmodels. For the second subtask, we use a cascaded model which grounds the\nresponse prediction on the predicted span instead of the full document. With\nthese approaches, we obtain significant improvements in both subtasks compared\nto the baseline.', 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am seeking a generative modeling approach capable of creating new levels and potentially game settings/environments for a video game with multiple existing levels of difficulty. Specifically, I am interested in exploring how Generative Adversarial Networks (GANs) and other generative methods could generate entirely new levels by emulating the style of previous ones. It is crucial that the newly generated levels are not merely derivative and that my generative model can optimize specific properties, such as the intensity or graphic nature of the game. Given that these properties are non-differentiable, I need a method to either render them differentiable or employ a reinforcement learning-centric approach to optimize these rewards. After generating a variety of levels, I require a method to select some of the best ones. One potential solution could be to evaluate the generated levels using an automatic metric, such as the performance of an AI agent playing the level. Alternatively, I am considering designing a derivative-free stochastic optimization algorithm to guide the search across the space of all synthetically generated levels, steering towards those that meet specific objectives.","['Generative adversarial networks (GANs) are quickly becoming a ubiquitous\napproach to procedurally generating video game levels. While GAN generated\nlevels are stylistically similar to human-authored examples, human designers\noften want to explore the generative design space of GANs to extract\ninteresting levels. However, human designers find latent vectors opaque and\nwould rather explore along dimensions the designer specifies, such as number of\nenemies or obstacles. We propose using state-of-the-art quality diversity\nalgorithms designed to optimize continuous spaces, i.e. MAP-Elites with a\ndirectional variation operator and Covariance Matrix Adaptation MAP-Elites, to\nefficiently explore the latent space of a GAN to extract levels that vary\nacross a set of specified gameplay measures. In the benchmark domain of Super\nMario Bros, we demonstrate how designers may specify gameplay measures to our\nsystem and extract high-quality (playable) levels with a diverse range of level\nmechanics, while still maintaining stylistic similarity to human authored\nexamples. An online user study shows how the different mechanics of the\nautomatically generated levels affect subjective ratings of their perceived\ndifficulty and appearance.', 'Several works have demonstrated the use of variational autoencoders (VAEs)\nfor generating levels in the style of existing games and blending levels across\ndifferent games. Further, quality-diversity (QD) algorithms have also become\npopular for generating varied game content by using evolution to explore a\nsearch space while focusing on both variety and quality. To reap the benefits\nof both these approaches, we present a level generation and game blending\napproach that combines the use of VAEs and QD algorithms. Specifically, we\ntrain VAEs on game levels and run the MAP-Elites QD algorithm using the learned\nlatent space of the VAE as the search space. The latent space captures the\nproperties of the games whose levels we want to generate and blend, while\nMAP-Elites searches this latent space to find a diverse set of levels\noptimizing a given objective such as playability. We test our method using\nmodels for 5 different platformer games as well as a blended domain spanning 3\nof these games. We refer to using MAP-Elites for blending as Blend-Elites. Our\nresults show that MAP-Elites in conjunction with VAEs enables the generation of\na diverse set of playable levels not just for each individual game but also for\nthe blended domain while illuminating game-specific regions of the blended\nlatent space.', 'Recent advancements in procedural content generation via machine learning\nenable the generation of video-game levels that are aesthetically similar to\nhuman-authored examples. However, the generated levels are often unplayable\nwithout additional editing. We propose a generate-then-repair framework for\nautomatic generation of playable levels adhering to specific styles. The\nframework constructs levels using a generative adversarial network (GAN)\ntrained with human-authored examples and repairs them using a mixed-integer\nlinear program (MIP) with playability constraints. A key component of the\nframework is computing minimum cost edits between the GAN generated level and\nthe solution of the MIP solver, which we cast as a minimum cost network flow\nproblem. Results show that the proposed framework generates a diverse range of\nplayable levels, that capture the spatial relationships between objects\nexhibited in the human-authored levels.', 'Methods for dynamic difficulty adjustment allow games to be tailored to\nparticular players to maximize their engagement. However, current methods often\nonly modify a limited set of game features such as the difficulty of the\nopponents, or the availability of resources. Other approaches, such as\nexperience-driven Procedural Content Generation (PCG), can generate complete\nlevels with desired properties such as levels that are neither too hard nor too\neasy, but require many iterations. This paper presents a method that can\ngenerate and search for complete levels with a specific target difficulty in\nonly a few trials. This advance is enabled by through an Intelligent\nTrial-and-Error algorithm, originally developed to allow robots to adapt\nquickly. Our algorithm first creates a large variety of different levels that\nvary across predefined dimensions such as leniency or map coverage. The\nperformance of an AI playing agent on these maps gives a proxy for how\ndifficult the level would be for another AI agent (e.g. one that employs Monte\nCarlo Tree Search instead of Greedy Tree Search); using this information, a\nBayesian Optimization procedure is deployed, updating the difficulty of the\nprior map to reflect the ability of the agent. The approach can reliably find\nlevels with a specific target difficulty for a variety of planning agents in\nonly a few trials, while maintaining an understanding of their skill landscape.', ""Deep reinforcement learning (RL) has shown impressive results in a variety of\ndomains, learning directly from high-dimensional sensory streams. However, when\nneural networks are trained in a fixed environment, such as a single level in a\nvideo game, they will usually overfit and fail to generalize to new levels.\nWhen RL models overfit, even slight modifications to the environment can result\nin poor agent performance. This paper explores how procedurally generated\nlevels during training can increase generality. We show that for some games\nprocedural level generation enables generalization to new levels within the\nsame distribution. Additionally, it is possible to achieve better performance\nwith less data by manipulating the difficulty of the levels in response to the\nperformance of the agent. The generality of the learned behaviors is also\nevaluated on a set of human-designed levels. The results suggest that the\nability to generalize to human-designed levels highly depends on the design of\nthe level generators. We apply dimensionality reduction and clustering\ntechniques to visualize the generators' distributions of levels and analyze to\nwhat degree they can produce levels similar to those designed by a human."", ""Generative Adversarial Networks (GANs) have demonstrated their ability to\nlearn patterns in data and produce new exemplars similar to, but different\nfrom, their training set in several domains, including video games. However,\nGANs have a fixed output size, so creating levels of arbitrary size for a\ndungeon crawling game is difficult. GANs also have trouble encoding semantic\nrequirements that make levels interesting and playable. This paper combines a\nGAN approach to generating individual rooms with a graph grammar approach to\ncombining rooms into a dungeon. The GAN captures design principles of\nindividual rooms, but the graph grammar organizes rooms into a global layout\nwith a sequence of obstacles determined by a designer. Room data from The\nLegend of Zelda is used to train the GAN. This approach is validated by a user\nstudy, showing that GAN dungeons are as enjoyable to play as a level from the\noriginal game, and levels generated with a graph grammar alone. However, GAN\ndungeons have rooms considered more complex, and plain graph grammar's dungeons\nare considered least complex and challenging. Only the GAN approach creates an\nextensive supply of both layouts and rooms, where rooms span across the\nspectrum of those seen in the training set to new creations merging design\nprinciples from multiple rooms."", 'Generative Adversarial Networks (GANs) are proving to be a powerful indirect\ngenotype-to-phenotype mapping for evolutionary search, but they have\nlimitations. In particular, GAN output does not scale to arbitrary dimensions,\nand there is no obvious way of combining multiple GAN outputs into a cohesive\nwhole, which would be useful in many areas, such as the generation of video\ngame levels. Game levels often consist of several segments, sometimes repeated\ndirectly or with variation, organized into an engaging pattern. Such patterns\ncan be produced with Compositional Pattern Producing Networks (CPPNs).\nSpecifically, a CPPN can define latent vector GAN inputs as a function of\ngeometry, which provides a way to organize level segments output by a GAN into\na complete level. This new CPPN2GAN approach is validated in both Super Mario\nBros. and The Legend of Zelda. Specifically, divergent search via MAP-Elites\ndemonstrates that CPPN2GAN can better cover the space of possible levels. The\nlayouts of the resulting levels are also more cohesive and aesthetically\nconsistent.', 'The procedural generation of levels and content in video games is a\nchallenging AI problem. Often such generation relies on an intelligent way of\nevaluating the content being generated so that constraints are satisfied and/or\nobjectives maximized. In this work, we address the problem of creating levels\nthat are not only playable but also revolve around specific mechanics in the\ngame. We use constrained evolutionary algorithms and quality-diversity\nalgorithms to generate small sections of Super Mario Bros levels called scenes,\nusing three different simulation approaches: Limited Agents, Punishing Model,\nand Mechanics Dimensions. All three approaches are able to create scenes that\ngive opportunity for a player to encounter or use targeted mechanics with\ndifferent properties. We conclude by discussing the advantages and\ndisadvantages of each approach and compare them to each other.', 'Previous work explored blending levels from existing games to create levels\nfor a new game that mixes properties of the original games. In this paper, we\nuse Variational Autoencoders (VAEs) for improving upon such techniques. VAEs\nare artificial neural networks that learn and use latent representations of\ndatasets to generate novel outputs. We train a VAE on level data from Super\nMario Bros. and Kid Icarus, enabling it to capture the latent space spanning\nboth games. We then use this space to generate level segments that combine\nproperties of levels from both games. Moreover, by applying evolutionary search\nin the latent space, we evolve level segments satisfying specific constraints.\nWe argue that these affordances make the VAE-based approach especially suitable\nfor co-creative level design and compare its performance with similar\ngenerative models like the GAN and the VAE-GAN.', 'Generative Adversarial Networks (GANs) have shown im-pressive results for\nimage generation. However, GANs facechallenges in generating contents with\ncertain types of con-straints, such as game levels. Specifically, it is\ndifficult togenerate levels that have aesthetic appeal and are playable atthe\nsame time. Additionally, because training data usually islimited, it is\nchallenging to generate unique levels with cur-rent GANs. In this paper, we\npropose a new GAN architec-ture namedConditional Embedding Self-Attention\nGenera-tive Adversarial Network(CESAGAN) and a new bootstrap-ping training\nprocedure. The CESAGAN is a modification ofthe self-attention GAN that\nincorporates an embedding fea-ture vector input to condition the training of\nthe discriminatorand generator. This allows the network to model\nnon-localdependency between game objects, and to count objects. Ad-ditionally,\nto reduce the number of levels necessary to trainthe GAN, we propose a\nbootstrapping mechanism in whichplayable generated levels are added to the\ntraining set. Theresults demonstrate that the new approach does not only\ngen-erate a larger number of levels that are playable but also gen-erates fewer\nduplicate levels compared to a standard GAN.', 'It has recently been shown that reinforcement learning can be used to train\ngenerators capable of producing high-quality game levels, with quality defined\nin terms of some user-specified heuristic. To ensure that these generators\'\noutput is sufficiently diverse (that is, not amounting to the reproduction of a\nsingle optimal level configuration), the generation process is constrained such\nthat the initial seed results in some variance in the generator\'s output.\nHowever, this results in a loss of control over the generated content for the\nhuman user. We propose to train generators capable of producing controllably\ndiverse output, by making them ""goal-aware."" To this end, we add conditional\ninputs representing how close a generator is to some heuristic, and also modify\nthe reward mechanism to incorporate that value. Testing on multiple domains, we\nshow that the resulting level generators are capable of exploring the space of\npossible levels in a targeted, controllable manner, producing levels of\ncomparable quality as their goal-unaware counterparts, that are diverse along\ndesigner-specified dimensions.', 'We investigate how reinforcement learning can be used to train\nlevel-designing agents. This represents a new approach to procedural content\ngeneration in games, where level design is framed as a game, and the content\ngenerator itself is learned. By seeing the design problem as a sequential task,\nwe can use reinforcement learning to learn how to take the next action so that\nthe expected final level quality is maximized. This approach can be used when\nfew or no examples exist to train from, and the trained generator is very fast.\nWe investigate three different ways of transforming two-dimensional level\ndesign problems into Markov decision processes and apply these to three game\nenvironments.']","[""This paper introduces a surrogate model of gameplay that learns the mapping\nbetween different game facets, and applies it to a generative system which\ndesigns new content in one of these facets. Focusing on the shooter game genre,\nthe paper explores how deep learning can help build a model which combines the\ngame level structure and the game's character class parameters as input and the\ngameplay outcomes as output. The model is trained on a large corpus of game\ndata from simulations with artificial agents in random sets of levels and class\nparameters. The model is then used to generate classes for specific levels and\nfor a desired game outcome, such as balanced matches of short duration.\nFindings in this paper show that the system can be expressive and can generate\nclasses for both computer generated and human authored levels."", ""This paper describes a new evolutionary algorithm that is especially well\nsuited to AI-Assisted Game Design. The approach adopted in this paper is to use\nobservations of AI agents playing the game to estimate the game's quality. Some\nof best agents for this purpose are General Video Game AI agents, since they\ncan be deployed directly on a new game without game-specific tuning; these\nagents tend to be based on stochastic algorithms which give robust but noisy\nresults and tend to be expensive to run. This motivates the main contribution\nof the paper: the development of the novel N-Tuple Bandit Evolutionary\nAlgorithm, where a model is used to estimate the fitness of unsampled points\nand a bandit approach is used to balance exploration and exploitation of the\nsearch space. Initial results on optimising a Space Battle game variant suggest\nthat the algorithm offers far more robust results than the Random Mutation Hill\nClimber and a Biased Mutation variant, which are themselves known to offer\ncompetitive performance across a range of problems. Subjective observations are\nalso given by human players on the nature of the evolved games, which indicate\na preference towards games generated by the N-Tuple algorithm."", 'In this work, we present TOAD-GAN (Token-based One-shot Arbitrary Dimension\nGenerative Adversarial Network), a novel Procedural Content Generation (PCG)\nalgorithm that generates token-based video game levels. TOAD-GAN follows the\nSinGAN architecture and can be trained using only one example. We demonstrate\nits application for Super Mario Bros. levels and are able to generate new\nlevels of similar style in arbitrary sizes. We achieve state-of-the-art results\nin modeling the patterns of the training level and provide a comparison with\ndifferent baselines under several metrics. Additionally, we present an\nextension of the method that allows the user to control the generation process\nof certain token structures to ensure a coherent global level layout. We\nprovide this tool to the community to spur further research by publishing our\nsource code.', 'Open-endedness, primarily studied in the context of artificial life, is the\nability of systems to generate potentially unbounded ontologies of increasing\nnovelty and complexity. Engineering generative systems displaying at least some\ndegree of this ability is a goal with clear applications to procedural content\ngeneration in games. The Paired Open-Ended Trailblazer (POET) algorithm,\nheretofore explored only in a biped walking domain, is a coevolutionary system\nthat simultaneously generates environments and agents that can solve them. This\npaper introduces a POET-Inspired Neuroevolutionary System for KreativitY\n(PINSKY) in games, which co-generates levels for multiple video games and\nagents that play them. This system leverages the General Video Game Artificial\nIntelligence (GVGAI) framework to enable co-generation of levels and agents for\nthe 2D Atari-style games Zelda and Solar Fox. Results demonstrate the ability\nof PINSKY to generate curricula of game levels, opening up a promising new\navenue for research at the intersection of procedural content generation and\nartificial life. At the same time, results in these challenging game domains\nhighlight the limitations of the current algorithm and opportunities for\nimprovement.', ""We introduce a procedural content generation (PCG) framework at the\nintersections of experience-driven PCG and PCG via reinforcement learning,\nnamed ED(PCG)RL, EDRL in short. EDRL is able to teach RL designers to generate\nendless playable levels in an online manner while respecting particular\nexperiences for the player as designed in the form of reward functions. The\nframework is tested initially in the Super Mario Bros game. In particular, the\nRL designers of Super Mario Bros generate and concatenate level segments while\nconsidering the diversity among the segments. The correctness of the generation\nis ensured by a neural net-assisted evolutionary level repairer and the\nplayability of the whole level is determined through AI-based testing. Our\nagents in this EDRL implementation learn to maximise a quantification of\nKoster's principle of fun by moderating the degree of diversity across level\nsegments. Moreover, we test their ability to design fun levels that are diverse\nover time and playable. Our proposed framework is capable of generating\nendless, playable Super Mario Bros levels with varying degrees of fun,\ndeviation from earlier segments, and playability. EDRL can be generalised to\nany game that is built as a segment-based sequential process and features a\nbuilt-in compressed representation of its game content."", ""Reinforcement learning algorithms have performed well in playing challenging\nboard and video games. More and more studies focus on improving the\ngeneralisation ability of reinforcement learning algorithms. The General Video\nGame AI Learning Competition aims to develop agents capable of learning to play\ndifferent game levels that were unseen during training. This paper summarises\nthe five years' General Video Game AI Learning Competition editions. At each\nedition, three new games were designed. The training and test levels were\ndesigned separately in the first three editions. Since 2020, three test levels\nof each game were generated by perturbing or combining two training levels.\nThen, we present a novel reinforcement learning technique with dual-observation\nfor general video game playing, assuming that it is more likely to observe\nsimilar local information in different levels rather than global information.\nInstead of directly inputting a single, raw pixel-based screenshot of the\ncurrent game screen, our proposed general technique takes the encoded,\ntransformed global and local observations of the game screen as two\nsimultaneous inputs, aiming at learning local information for playing new\nlevels. Our proposed technique is implemented with three state-of-the-art\nreinforcement learning algorithms and tested on the game set of the 2020\nGeneral Video Game AI Learning Competition. Ablation studies show the\noutstanding performance of using encoded, transformed global and local\nobservations as input."", ""Machine learning for procedural content generation has recently become an\nactive area of research. Levels vary in both form and function and are mostly\nunrelated to each other across games. This has made it difficult to assemble\nsuitably large datasets to bring machine learning to level design in the same\nway as it's been used for image generation. Here we propose Generative Playing\nNetworks which design levels for itself to play. The algorithm is built in two\nparts; an agent that learns to play game levels, and a generator that learns\nthe distribution of playable levels. As the agent learns and improves its\nability, the space of playable levels, as defined by the agent, grows. The\ngenerator targets the agent's playability estimates to then update its\nunderstanding of what constitutes a playable level. We call this process of\nlearning the distribution of data found through self-discovery with an\nenvironment, self-supervised inductive learning. Unlike previous approaches to\nprocedural content generation, Generative Playing Networks are end-to-end\ndifferentiable and do not require human-designed examples or domain knowledge.\nWe demonstrate the capability of this framework by training an agent and level\ngenerator for a 2D dungeon crawler game."", 'Algorithms that generate computer game content require game design knowledge.\nWe present an approach to automatically learn game design knowledge for level\ndesign from gameplay videos. We further demonstrate how the acquired design\nknowledge can be used to generate sections of game levels. Our approach\ninvolves parsing video of people playing a game to detect the appearance of\npatterns of sprites and utilizing machine learning to build a probabilistic\nmodel of sprite placement. We show how rich game design information can be\nautomatically parsed from gameplay videos and represented as a set of\ngenerative probabilistic models. We use Super Mario Bros. as a proof of\nconcept. We evaluate our approach on a measure of playability and stylistic\nsimilarity to the original levels as represented in the gameplay videos.', 'This survey explores Procedural Content Generation via Machine Learning\n(PCGML), defined as the generation of game content using machine learning\nmodels trained on existing content. As the importance of PCG for game\ndevelopment increases, researchers explore new avenues for generating\nhigh-quality content with or without human involvement; this paper addresses\nthe relatively new paradigm of using machine learning (in contrast with\nsearch-based, solver-based, and constructive methods). We focus on what is most\noften considered functional game content such as platformer levels, game maps,\ninteractive fiction stories, and cards in collectible card games, as opposed to\ncosmetic content such as sprites and sound effects. In addition to using PCG\nfor autonomous generation, co-creativity, mixed-initiative design, and\ncompression, PCGML is suited for repair, critique, and content analysis because\nof its focus on modeling existing content. We discuss various data sources and\nrepresentations that affect the resulting generated content. Multiple PCGML\nmethods are covered, including neural networks, long short-term memory (LSTM)\nnetworks, autoencoders, and deep convolutional networks; Markov models,\n$n$-grams, and multi-dimensional Markov chains; clustering; and matrix\nfactorization. Finally, we discuss open problems in the application of PCGML,\nincluding learning from small datasets, lack of training data, multi-layered\nlearning, style-transfer, parameter tuning, and PCG as a game mechanic.', 'Automatically adapting game content to players opens new doors for game\ndevelopment. In this paper we propose an architecture using persona agents and\nexperience metrics, which enables evolving procedurally generated levels\ntailored for particular player personas. Using our game, ""Grave Rave"", we\ndemonstrate that this approach successfully adapts to four rule-based persona\nagents over three different experience metrics. Furthermore, the adaptation is\nshown to be specific in nature, meaning that the levels are persona-conscious,\nand not just general optimizations with regard to the selected metric.', 'Generative Adversarial Networks (GANs) are a powerful indirect\ngenotype-to-phenotype mapping for evolutionary search. Much previous work\napplying GANs to level generation focuses on fixed-size segments combined into\na whole level, but individual segments may not fit together cohesively. In\ncontrast, segments in human designed levels are often repeated, directly or\nwith variation, and organized into patterns (the symmetric eagle in Level 1 of\nThe Legend of Zelda, or repeated pipe motifs in Super Mario Bros). Such\npatterns can be produced with Compositional Pattern Producing Networks (CPPNs).\nCPPNs define latent vector GAN inputs as a function of geometry, organizing\nsegments output by a GAN into complete levels. However, collections of latent\nvectors can also be evolved directly, producing more chaotic levels. We propose\na hybrid approach that evolves CPPNs first, but allows latent vectors to evolve\nlater, combining the benefits of both approaches. These approaches are\nevaluated in Super Mario Bros. and The Legend of Zelda. We previously\ndemonstrated via a Quality-Diversity algorithm that CPPNs better cover the\nspace of possible levels than directly evolved levels. Here, we show that the\nhybrid approach (1) covers areas that neither of the other methods can, and (2)\nachieves comparable or superior QD scores.', 'The General Video Game Artificial Intelligence (GVGAI) competition has been\nrunning for several years with various tracks. This paper focuses on the\nchallenge of the GVGAI learning track in which 3 games are selected and 2\nlevels are given for training, while 3 hidden levels are left for evaluation.\nThis setup poses a difficult challenge for current Reinforcement Learning (RL)\nalgorithms, as they typically require much more data. This work investigates 3\nversions of the Advantage Actor-Critic (A2C) algorithm trained on a maximum of\n2 levels from the available 5 from the GVGAI framework and compares their\nperformance on all levels. The selected sub-set of games have different\ncharacteristics, like stochasticity, reward distribution and objectives. We\nfound that stochasticity improves the generalisation, but too much can cause\nthe algorithms to fail to learn the training levels. The quality of the\ntraining levels also matters, different sets of training levels can boost\ngeneralisation over all levels. In the GVGAI competition agents are scored\nbased on their win rates and then their scores achieved in the games. We found\nthat solely using the rewards provided by the game might not encourage winning.', ""Video game level generation based on machine learning (ML), in particular,\ndeep generative models, has attracted attention as a technique to automate\nlevel generation. However, applications of existing ML-based level generations\nare mostly limited to tile-based level representation. When ML techniques are\napplied to game domains with non-tile-based level representation, such as Angry\nBirds, where objects in a level are specified by real-valued parameters, ML\noften fails to generate playable levels. In this study, we develop a\ndeep-generative-model-based level generation for the game domain of Angry\nBirds. To overcome these drawbacks, we propose a sequential encoding of a level\nand process it as text data, whereas existing approaches employ a tile-based\nencoding and process it as an image. Experiments show that the proposed level\ngenerator drastically improves the stability and diversity of generated levels\ncompared with existing approaches. We apply latent variable evolution with the\nproposed generator to control the feature of a generated level computed through\nan AI agent's play, while keeping the level stable and natural."", 'Generative Adversarial Networks (GANs) are an emerging form of indirect\nencoding. The GAN is trained to induce a latent space on training data, and a\nreal-valued evolutionary algorithm can search that latent space. Such Latent\nVariable Evolution (LVE) has recently been applied to game levels. However, it\nis hard for objective scores to capture level features that are appealing to\nplayers. Therefore, this paper introduces a tool for interactive LVE of\ntile-based levels for games. The tool also allows for direct exploration of the\nlatent dimensions, and allows users to play discovered levels. The tool works\nfor a variety of GAN models trained for both Super Mario Bros. and The Legend\nof Zelda, and is easily generalizable to other games. A user study shows that\nboth the evolution and latent space exploration features are appreciated, with\na slight preference for direct exploration, but combining these features allows\nusers to discover even better levels. User feedback also indicates how this\nsystem could eventually grow into a commercial design tool, with the addition\nof a few enhancements.', 'Generative Adversarial Networks (GANs) are capable of generating convincing\nimitations of elements from a training set, but the distribution of elements in\nthe training set affects to difficulty of properly training the GAN and the\nquality of the outputs it produces. This paper looks at six different GANs\ntrained on different subsets of data from the game Lode Runner. The quality\ndiversity algorithm MAP-Elites was used to explore the set of quality levels\nthat could be produced by each GAN, where quality was defined as being beatable\nand having the longest solution path possible. Interestingly, a GAN trained on\nonly 20 levels generated the largest set of diverse beatable levels while a GAN\ntrained on 150 levels generated the smallest set of diverse beatable levels,\nthus challenging the notion that more is always better when training GANs.', 'Applying latent variable evolution to game level design has become more and\nmore popular as little human expert knowledge is required. However, defective\nlevels with illegal patterns may be generated due to the violation of\nconstraints for level design. A traditional way of repairing the defective\nlevels is programming specific rule-based repairers to patch the flaw. However,\nprogramming these constraints is sometimes complex and not straightforward. An\nautonomous level repairer which is capable of learning the constraints is\nneeded. In this paper, we propose a novel approach, CNet, to learn the\nprobability distribution of tiles giving its surrounding tiles on a set of real\nlevels, and then detect the illegal tiles in generated new levels. Then, an\nevolutionary repairer is designed to search for optimal replacement schemes\nequipped with a novel search space being constructed with the help of CNet and\na novel heuristic function. The proposed approaches are proved to be effective\nin our case study of repairing GAN-generated and artificially destroyed levels\nof Super Mario Bros. game. Our CNet-assisted evolutionary repairer can also be\neasily applied to other games of which the levels can be represented by a\nmatrix of objects or tiles.', ""In games, as well as many user-facing systems, adapting content to users'\npreferences and experience is an important challenge. This paper explores a\nnovel method to realize this goal in the context of dynamic difficulty\nadjustment (DDA). Here the aim is to constantly adapt the content of a game to\nthe skill level of the player, keeping them engaged by avoiding states that are\neither too difficult or too easy. Current systems for DDA rely on expensive\ndata mining, or on hand-crafted rules designed for particular domains, and\nusually adapts to keep players in the flow, leaving no room for the designer to\npresent content that is purposefully easy or difficult. This paper presents\nFast Bayesian Content Adaption (FBCA), a system for DDA that is agnostic to the\ndomain and that can target particular difficulties. We deploy this framework in\ntwo different domains: the puzzle game Sudoku, and a simple Roguelike game. By\nmodifying the acquisition function's optimization, we are reliably able to\npresent a content with a bespoke difficulty for players with different skill\nlevels in less than five iterations for Sudoku and fifteen iterations for the\nsimple Roguelike. Our method significantly outperforms simpler DDA heuristics\nwith the added benefit of maintaining a model of the user. These results point\ntowards a promising alternative for content adaption in a variety of different\ndomains."", ""We present practical approaches of using deep learning to create and enhance\nlevel maps and textures for video games -- desktop, mobile, and web. We aim to\npresent new possibilities for game developers and level artists. The task of\ndesigning levels and filling them with details is challenging. It is both\ntime-consuming and takes effort to make levels rich, complex, and with a\nfeeling of being natural. Fortunately, recent progress in deep learning\nprovides new tools to accompany level designers and visual artists. Moreover,\nthey offer a way to generate infinite worlds for game replayability and adjust\neducational games to players' needs. We present seven approaches to create\nlevel maps, each using statistical methods, machine learning, or deep learning.\nIn particular, we include:\n  - Generative Adversarial Networks for creating new images from existing\nexamples (e.g. ProGAN).\n  - Super-resolution techniques for upscaling images while preserving crisp\ndetail (e.g. ESRGAN).\n  - Neural style transfer for changing visual themes.\n  - Image translation - turning semantic maps into images (e.g. GauGAN).\n  - Semantic segmentation for turning images into semantic masks (e.g. U-Net).\n  - Unsupervised semantic segmentation for extracting semantic features (e.g.\nTile2Vec).\n  - Texture synthesis - creating large patterns based on a smaller sample (e.g.\nInGAN)."", ""This paper presents an adaptive level generation algorithm for the\nphysics-based puzzle game Angry Birds. The proposed algorithm is based on a\npre-existing level generator for this game, but where the difficulty of the\ngenerated levels can be adjusted based on the player's performance. This allows\nfor the creation of personalised levels tailored specifically to the player's\nown abilities. The effectiveness of our proposed method is evaluated using\nseveral agents with differing strategies and AI techniques. By using these\nagents as models / representations of real human player's characteristics, we\ncan optimise level properties efficiently over a large number of generations.\nAs a secondary investigation, we also demonstrate that by combining the\nperformance of several agents together it is possible to generate levels that\nare especially challenging for certain players but not others."", ""Machine playtesting tools and game moment search engines require exposure to\nthe diversity of a game's state space if they are to report on or index the\nmost interesting moments of possible play. Meanwhile, mobile app distribution\nservices would like to quickly determine if a freshly-uploaded game is fit to\nbe published. Having access to a semantic map of reachable states in the game\nwould enable efficient inference in these applications. However, human gameplay\ndata is expensive to acquire relative to the coverage of a game that it\nprovides. We show that off-the-shelf automatic exploration strategies can\nexplore with an effectiveness comparable to human gameplay on the same\ntimescale. We contribute generic methods for quantifying exploration quality as\na function of time and demonstrate our metric on several elementary techniques\nand human players on a collection of commercial games sampled from multiple\ngame platforms (from Atari 2600 to Nintendo 64). Emphasizing the diversity of\nstates reached and the semantic map extracted, this work makes productive\ncontrast with the focus on finding a behavior policy or optimizing game score\nused in most automatic game playing research."", 'This paper provides a detailed investigation of using the Kullback-Leibler\n(KL) Divergence as a way to compare and analyse game-levels, and hence to use\nthe measure as the objective function of an evolutionary algorithm to evolve\nnew levels. We describe the benefits of its asymmetry for level analysis and\ndemonstrate how (not surprisingly) the quality of the results depends on the\nfeatures used. Here we use tile-patterns of various sizes as features.\n  When using the measure for evolution-based level generation, we demonstrate\nthat the choice of variation operator is critical in order to provide an\nefficient search process, and introduce a novel convolutional mutation operator\nto facilitate this. We compare the results with alternative generators,\nincluding evolving in the latent space of generative adversarial networks, and\nWave Function Collapse. The results clearly show the proposed method to provide\ncompetitive performance, providing reasonable quality results with very fast\ntraining and reasonably fast generation.', ""The Angry Birds AI competition has been held over many years to encourage the\ndevelopment of AI agents that can play Angry Birds game levels better than\nhuman players. Many different agents with various approaches have been employed\nover the competition's lifetime to solve this task. Even though the performance\nof these agents has increased significantly over the past few years, they still\nshow major drawbacks in playing deceptive levels. This is because most of the\ncurrent agents try to identify the best next shot rather than planning an\neffective sequence of shots. In order to encourage advancements in such agents,\nwe present an automated methodology to generate deceptive game levels for Angry\nBirds. Even though there are many existing content generators for Angry Birds,\nthey do not focus on generating deceptive levels. In this paper, we propose a\nprocedure to generate deceptive levels for six deception categories that can\nfool the state-of-the-art Angry Birds playing AI agents. Our results show that\ngenerated deceptive levels exhibit similar characteristics of human-created\ndeceptive levels. Additionally, we define metrics to measure the stability,\nsolvability, and degree of deception of the generated levels."", 'Generating novel graph structures that optimize given objectives while\nobeying some given underlying rules is fundamental for chemistry, biology and\nsocial science research. This is especially important in the task of molecular\ngraph generation, whose goal is to discover novel molecules with desired\nproperties such as drug-likeness and synthetic accessibility, while obeying\nphysical laws such as chemical valency. However, designing models to find\nmolecules that optimize desired properties while incorporating highly complex\nand non-differentiable rules remains to be a challenging task. Here we propose\nGraph Convolutional Policy Network (GCPN), a general graph convolutional\nnetwork based model for goal-directed graph generation through reinforcement\nlearning. The model is trained to optimize domain-specific rewards and\nadversarial loss through policy gradient, and acts in an environment that\nincorporates domain-specific rules. Experimental results show that GCPN can\nachieve 61% improvement on chemical property optimization over state-of-the-art\nbaselines while resembling known molecules, and achieve 184% improvement on the\nconstrained property optimization task.', 'Recent procedural content generation via machine learning (PCGML) methods\nallow learning from existing content to produce similar content automatically.\nWhile these approaches are able to generate content for different games (e.g.\nSuper Mario Bros., DOOM, Zelda, and Kid Icarus), it is an open questions how\nwell these approaches can capture large-scale visual patterns such as symmetry.\nIn this paper, we propose match-three games as a domain to test PCGML\nalgorithms regarding their ability to generate suitable patterns. We\ndemonstrate that popular algorithm such as Generative Adversarial Networks\nstruggle in this domain and propose adaptations to improve their performance.\nIn particular we augment the neighborhood of a Markov Random Fields approach to\nnot only take local but also symmetric positional information into account. We\nconduct several empirical tests including a user study that show the\nimprovements achieved by the proposed modifications, and obtain promising\nresults.', 'We propose the problem of tutorial generation for games, i.e. to generate\ntutorials which can teach players to play games, as an AI problem. This problem\ncan be approached in several ways, including generating natural language\ndescriptions of game rules, generating instructive game levels, and generating\ndemonstrations of how to play a game using agents that play in a human-like\nmanner. We further argue that the General Video Game AI framework provides a\nuseful testbed for addressing this problem.', ""Generative Adversarial Networks (GANs) can generate levels for a variety of\ngames. This paper focuses on combining GAN-generated segments in a snaking\npattern to create levels for Mega Man. Adjacent segments in such levels can be\northogonally adjacent in any direction, meaning that an otherwise fine segment\nmight impose a barrier between its neighbor depending on what sorts of segments\nin the training set are being most closely emulated: horizontal, vertical, or\ncorner segments. To pick appropriate segments, multiple GANs were trained on\ndifferent types of segments to ensure better flow between segments. Flow was\nfurther improved by evolving the latent vectors for the segments being joined\nin the level to maximize the length of the level's solution path. Using\nmultiple GANs to represent different types of segments results in significantly\nlonger solution paths than using one GAN for all segment types, and a human\nsubject study verifies that these levels are more fun and have more human-like\ndesign than levels produced by one GAN."", 'Successfully navigating a complex environment to obtain a desired outcome is\na difficult task, that up to recently was believed to be capable only by\nhumans. This perception has been broken down over time, especially with the\nintroduction of deep reinforcement learning, which has greatly increased the\ndifficulty of tasks that can be automated. However, for traditional\nreinforcement learning agents this requires an environment to be able to\nprovide frequent extrinsic rewards, which are not known or accessible for many\nreal-world environments. This project aims to explore and contrast existing\nreinforcement learning solutions that circumnavigate the difficulties of an\nenvironment that provide sparse rewards. Different reinforcement solutions will\nbe implemented over a several video game environments with varying difficulty\nand varying frequency of rewards, as to properly investigate the applicability\nof these solutions. This project introduces a novel reinforcement learning\nsolution by combining aspects of two existing state of the art sparse reward\nsolutions, curiosity driven exploration and unsupervised auxiliary tasks.', 'Levels are a key component of many different video games, and a large body of\nwork has been produced on how to procedurally generate game levels. Recently,\nMachine Learning techniques have been applied to video game level generation\ntowards the purpose of automatically generating levels that have the properties\nof the training corpus. Towards that end we have made available a corpora of\nvideo game levels in an easy to parse format ideal for different machine\nlearning and other game AI research purposes.', 'We describe a search-based approach to generating new levels for bullet hell\ngames, which are action games characterized by and requiring avoidance of a\nvery large amount of projectiles. Levels are represented using a\ndomain-specific description language, and search in the space defined by this\nlanguage is performed by a novel variant of the Map-Elites algorithm which\nincorporates a feasible- infeasible approach to constraint satisfaction.\nSimulation-based evaluation is used to gauge the fitness of levels, using an\nagent based on best-first search. The performance of the agent can be tuned\naccording to the two dimensions of strategy and dexterity, making it possible\nto search for level configurations that require a specific combination of both.\nAs far as we know, this paper describes the first generator for this game\ngenre, and includes several algorithmic innovations.', 'This paper reviews the field of Game AI, which not only deals with creating\nagents that can play a certain game, but also with areas as diverse as creating\ngame content automatically, game analytics, or player modelling. While Game AI\nwas for a long time not very well recognized by the larger scientific\ncommunity, it has established itself as a research area for developing and\ntesting the most advanced forms of AI algorithms and articles covering advances\nin mastering video games such as StarCraft 2 and Quake III appear in the most\nprestigious journals. Because of the growth of the field, a single review\ncannot cover it completely. Therefore, we put a focus on important recent\ndevelopments, including that advances in Game AI are starting to be extended to\nareas outside of games, such as robotics or the synthesis of chemicals. In this\narticle, we review the algorithms and methods that have paved the way for these\nbreakthroughs, report on the other important areas of Game AI research, and\nalso point out exciting directions for the future of Game AI.', 'We introduce the General Video Game Rule Generation problem, and the\neponymous software framework which will be used in a new track of the General\nVideo Game AI (GVGAI) competition. The problem is, given a game level as input,\nto generate the rules of a game that fits that level. This can be seen as the\ninverse of the General Video Game Level Generation problem. Conceptualizing\nthese two problems as separate helps breaking the very hard problem of\ngenerating complete games into smaller, more manageable subproblems. The\nproposed framework builds on the GVGAI software and thus asks the rule\ngenerator for rules defined in the Video Game Description Language. We describe\nthe API, and three different rule generators: a random, a constructive and a\nsearch-based generator. Early results indicate that the constructive generator\ngenerates playable and somewhat interesting game rules but has a limited\nexpressive range, whereas the search-based generator generates remarkably\ndiverse rulesets, but with an uneven quality.', 'Automated game design is the problem of automatically producing games through\ncomputational processes. Traditionally, these methods have relied on the\nauthoring of search spaces by a designer, defining the space of all possible\ngames for the system to author. In this paper, we instead learn representations\nof existing games from gameplay video and use these to approximate a search\nspace of novel games. In a human subject study we demonstrate that these novel\ngames are indistinguishable from human games in terms of challenge, and that\none of the novel games was equivalent to one of the human games in terms of\nfun, frustration, and likeability.', ""Modern machine learning techniques, such as deep neural networks, are\ntransforming many disciplines ranging from image recognition to language\nunderstanding, by uncovering patterns in big data and making accurate\npredictions. They have also shown promising results for synthesizing new\ndesigns, which is crucial for creating products and enabling innovation.\nGenerative models, including generative adversarial networks (GANs), have\nproven to be effective for design synthesis with applications ranging from\nproduct design to metamaterial design. These automated computational design\nmethods can support human designers, who typically create designs by a\ntime-consuming process of iteratively exploring ideas using experience and\nheuristics. However, there are still challenges remaining in automatically\nsynthesizing `creative' designs. GAN models, however, are not capable of\ngenerating unique designs, a key to innovation and a major gap in AI-based\ndesign automation applications. This paper proposes an automated method, named\nCreativeGAN, for generating novel designs. It does so by identifying components\nthat make a design unique and modifying a GAN model such that it becomes more\nlikely to generate designs with identified unique components. The method\ncombines state-of-art novelty detection, segmentation, novelty localization,\nrewriting, and generative models for creative design synthesis. Using a dataset\nof bicycle designs, we demonstrate that the method can create new bicycle\ndesigns with unique frames and handles, and generalize rare novelties to a\nbroad set of designs. Our automated method requires no human intervention and\ndemonstrates a way to rethink creative design synthesis and exploration."", 'Bayesian optimization has been successful at global optimization of\nexpensive-to-evaluate multimodal objective functions. However, unlike most\noptimization methods, Bayesian optimization typically does not use derivative\ninformation. In this paper we show how Bayesian optimization can exploit\nderivative information to decrease the number of objective function evaluations\nrequired for good performance. In particular, we develop a novel Bayesian\noptimization algorithm, the derivative-enabled knowledge-gradient (dKG), for\nwhich we show one-step Bayes-optimality, asymptotic consistency, and greater\none-step value of information than is possible in the derivative-free setting.\nOur procedure accommodates noisy and incomplete derivative information, comes\nin both sequential and batch forms, and can optionally reduce the computational\ncost of inference through automatically selected retention of a single\ndirectional derivative. We also compute the d-KG acquisition function and its\ngradient using a novel fast discretization-free technique. We show d-KG\nprovides state-of-the-art performance compared to a wide range of optimization\nprocedures with and without gradients, on benchmarks including logistic\nregression, deep learning, kernel learning, and k-nearest neighbors.', 'The procedural generation of video game levels has existed for at least 30\nyears, but only recently have machine learning approaches been used to generate\nlevels without specifying the rules for generation. A number of these have\nlooked at platformer levels as a sequence of characters and performed\ngeneration using Markov chains. In this paper we examine the use of Long\nShort-Term Memory recurrent neural networks (LSTMs) for the purpose of\ngenerating levels trained from a corpus of Super Mario Brothers levels. We\nanalyze a number of different data representations and how the generated levels\nfit into the space of human authored Super Mario Brothers levels.', 'Multiple Artificial Intelligence (AI) methods have been proposed over recent\nyears to create controllers to play multiple video games of different nature\nand complexity without revealing the specific mechanics of each of these games\nto the AI methods. In recent years, Evolutionary Algorithms (EAs) employing\nrolling horizon mechanisms have achieved extraordinary results in these type of\nproblems. However, some limitations are present in Rolling Horizon EAs making\nit a grand challenge of AI. These limitations include the wasteful mechanism of\ncreating a population and evolving it over a fraction of a second to propose an\naction to be executed by the game agent. Another limitation is to use a scalar\nvalue (fitness value) to direct evolutionary search instead of accounting for a\nmechanism that informs us how a particular agent behaves during the rolling\nhorizon simulation. In this work, we address both of these issues. We introduce\nthe use of a statistical tree that tackles the latter limitation. Furthermore,\nwe tackle the former limitation by employing a mechanism that allows us to seed\npart of the population using Monte Carlo Tree Search, a method that has\ndominated multiple General Video Game AI competitions. We show how the proposed\nnovel mechanism, called Statistical Tree-based Population Seeding, achieves\nbetter results compared to vanilla Rolling Horizon EAs in a set of 20 games,\nincluding 10 stochastic and 10 deterministic games.', 'Procedural content generation via machine learning (PCGML) has shown success\nat producing new video game content with machine learning. However, the\nmajority of the work has focused on the production of static game content,\nincluding game levels and visual elements. There has been much less work on\ndynamic game content, such as game mechanics. One reason for this is the lack\nof a consistent representation for dynamic game content, which is key for a\nnumber of statistical machine learning approaches. We present an autoencoder\nfor deriving what we call ""entity embeddings"", a consistent way to represent\ndifferent dynamic entities across multiple games in the same representation. In\nthis paper we introduce the learned representation, along with some evidence\ntowards its quality and future utility.', 'This work introduces World-GAN, the first method to perform data-driven\nProcedural Content Generation via Machine Learning in Minecraft from a single\nexample. Based on a 3D Generative Adversarial Network (GAN) architecture, we\nare able to create arbitrarily sized world snippets from a given sample. We\nevaluate our approach on creations from the community as well as structures\ngenerated with the Minecraft World Generator. Our method is motivated by the\ndense representations used in Natural Language Processing (NLP) introduced with\nword2vec [1]. The proposed block2vec representations make World-GAN independent\nfrom the number of different blocks, which can vary a lot in Minecraft, and\nenable the generation of larger levels. Finally, we demonstrate that changing\nthis new representation space allows us to change the generated style of an\nalready trained generator. World-GAN enables its users to generate Minecraft\nworlds based on parts of their creations.', 'Access to a fast and easily copied forward model of a game is essential for\nmodel-based reinforcement learning and for algorithms such as Monte Carlo tree\nsearch, and is also beneficial as a source of unlimited experience data for\nmodel-free algorithms. Learning forward models is an interesting and important\nchallenge in order to address problems where a model is not available. Building\nupon previous work on the Neural GPU, this paper introduces the Neural Game\nEngine, as a way to learn models directly from pixels. The learned models are\nable to generalise to different size game levels to the ones they were trained\non without loss of accuracy. Results on 10 deterministic General Video Game AI\ngames demonstrate competitive performance, with many of the games models being\nlearned perfectly both in terms of pixel predictions and reward predictions.\nThe pre-trained models are available through the OpenAI Gym interface and are\navailable publicly for future research here:\n\\url{https://github.com/Bam4d/Neural-Game-Engine}', 'In this paper we present a novel approach to optimise tactical and strategic\ndecision making in football (soccer). We model the game of football as a\nmulti-stage game which is made up from a Bayesian game to model the pre-match\ndecisions and a stochastic game to model the in-match state transitions and\ndecisions. Using this formulation, we propose a method to predict the\nprobability of game outcomes and the payoffs of team actions. Building upon\nthis, we develop algorithms to optimise team formation and in-game tactics with\ndifferent objectives. Empirical evaluation of our approach on real-world\ndatasets from 760 matches shows that by using optimised tactics from our\nBayesian and stochastic games, we can increase a team chances of winning by up\nto 16.1\\% and 3.4\\% respectively.', 'The design of video game levels is a complex and critical task. Levels need\nto elicit fun and challenge while avoiding frustration at all costs. In this\npaper, we present a framework to assist designers in the creation of levels for\n2D platformers. Our framework provides designers with a toolbox (i) to create\n2D platformer levels, (ii) to estimate the difficulty and probability of\nsuccess of single jump actions (the main mechanics of platformer games), and\n(iii) a set of metrics to evaluate the difficulty and probability of completion\nof entire levels. At the end, we present the results of a set of experiments we\ncarried out with human players to validate the metrics included in our\nframework.', ""Environments with procedurally generated content serve as important\nbenchmarks for testing systematic generalization in deep reinforcement\nlearning. In this setting, each level is an algorithmically created environment\ninstance with a unique configuration of its factors of variation. Training on a\nprespecified subset of levels allows for testing generalization to unseen\nlevels. What can be learned from a level depends on the current policy, yet\nprior work defaults to uniform sampling of training levels independently of the\npolicy. We introduce Prioritized Level Replay (PLR), a general framework for\nselectively sampling the next training level by prioritizing those with higher\nestimated learning potential when revisited in the future. We show TD-errors\neffectively estimate a level's future learning potential and, when used to\nguide the sampling procedure, induce an emergent curriculum of increasingly\ndifficult levels. By adapting the sampling of training levels, PLR\nsignificantly improves sample efficiency and generalization on Procgen\nBenchmark--matching the previous state-of-the-art in test return--and readily\ncombines with other methods. Combined with the previous leading method, PLR\nraises the state-of-the-art to over 76% improvement in test return relative to\nstandard RL baselines."", ""Reinforcement learning (RL) studies how an agent comes to achieve reward in\nan environment through interactions over time. Recent advances in machine RL\nhave surpassed human expertise at the world's oldest board games and many\nclassic video games, but they require vast quantities of experience to learn\nsuccessfully -- none of today's algorithms account for the human ability to\nlearn so many different tasks, so quickly. Here we propose a new approach to\nthis challenge based on a particularly strong form of model-based RL which we\ncall Theory-Based Reinforcement Learning, because it uses human-like intuitive\ntheories -- rich, abstract, causal models of physical objects, intentional\nagents, and their interactions -- to explore and model an environment, and plan\neffectively to achieve task goals. We instantiate the approach in a video game\nplaying agent called EMPA (the Exploring, Modeling, and Planning Agent), which\nperforms Bayesian inference to learn probabilistic generative models expressed\nas programs for a game-engine simulator, and runs internal simulations over\nthese models to support efficient object-based, relational exploration and\nheuristic planning. EMPA closely matches human learning efficiency on a suite\nof 90 challenging Atari-style video games, learning new games in just minutes\nof game play and generalizing robustly to new game situations and new levels.\nThe model also captures fine-grained structure in people's exploration\ntrajectories and learning dynamics. Its design and behavior suggest a way\nforward for building more general human-like AI systems."", 'General Video Game Playing (GVGP) aims at designing an agent that is capable\nof playing multiple video games with no human intervention. In 2014, The\nGeneral Video Game AI (GVGAI) competition framework was created and released\nwith the purpose of providing researchers a common open-source and easy to use\nplatform for testing their AI methods with potentially infinity of games\ncreated using Video Game Description Language (VGDL). The framework has been\nexpanded into several tracks during the last few years to meet the demand of\ndifferent research directions. The agents are required either to play multiple\nunknown games with or without access to game simulations, or to design new game\nlevels or rules. This survey paper presents the VGDL, the GVGAI framework,\nexisting tracks, and reviews the wide use of GVGAI framework in research,\neducation and competitions five years after its birth. A future plan of\nframework improvements is also described.', ""Generative Adversarial Networks (GANs) have recently attracted considerable\nattention in the AI community due to its ability to generate high-quality data\nof significant statistical resemblance to real data. Fundamentally, GAN is a\ngame between two neural networks trained in an adversarial manner to reach a\nzero-sum Nash equilibrium profile. Despite the improvement accomplished in GANs\nin the last few years, several issues remain to be solved. This paper reviews\nthe literature on the game theoretic aspects of GANs and addresses how game\ntheory models can address specific challenges of generative model and improve\nthe GAN's performance. We first present some preliminaries, including the basic\nGAN model and some game theory background. We then present taxonomy to classify\nstate-of-the-art solutions into three main categories: modified game models,\nmodified architectures, and modified learning methods. The classification is\nbased on modifications made to the basic GAN model by proposed game-theoretic\napproaches in the literature. We then explore the objectives of each category\nand discuss recent works in each category. Finally, we discuss the remaining\nchallenges in this field and present future research directions."", ""Procedural Content Generation via Machine Learning (PCGML) refers to a group\nof methods for creating game content (e.g. platformer levels, game maps, etc.)\nusing machine learning models. PCGML approaches rely on black box models, which\ncan be difficult to understand and debug by human designers who do not have\nexpert knowledge about machine learning. This can be even more tricky in\nco-creative systems where human designers must interact with AI agents to\ngenerate game content. In this paper we present an approach to explainable\nartificial intelligence in which certain training instances are offered to\nhuman users as an explanation for the AI agent's actions during a co-creation\nprocess. We evaluate this approach by approximating its ability to provide\nhuman users with the explanations of AI agent's actions and helping them to\nmore efficiently cooperate with the AI agent."", 'Generative models are undoubtedly a hot topic in Artificial Intelligence,\namong which the most common type is Generative Adversarial Networks (GANs).\nThese architectures let one synthesise artificial datasets by implicitly\nmodelling the underlying probability distribution of a real-world training\ndataset. With the introduction of Conditional GANs and their variants, these\nmethods were extended to generating samples conditioned on ancillary\ninformation available for each sample within the dataset. From a practical\nstandpoint, however, one might desire to generate data conditioned on partial\ninformation. That is, only a subset of the ancillary conditioning variables\nmight be of interest when synthesising data. In this work, we argue that\nstandard Conditional GANs are not suitable for such a task and propose a new\nAdversarial Network architecture and training strategy to deal with the ensuing\nproblems. Experiments illustrating the value of the proposed approach in digit\nand face image synthesis under partial conditioning information are presented,\nshowing that the proposed method can effectively outperform the standard\napproach under these circumstances.', ""The task of image generation started to receive some attention from artists\nand designers to inspire them in new creations. However, exploiting the results\nof deep generative models such as Generative Adversarial Networks can be long\nand tedious given the lack of existing tools. In this work, we propose a simple\nstrategy to inspire creators with new generations learned from a dataset of\ntheir choice, while providing some control on them. We design a simple\noptimization method to find the optimal latent parameters corresponding to the\nclosest generation to any input inspirational image. Specifically, we allow the\ngeneration given an inspirational image of the user choice by performing\nseveral optimization steps to recover optimal parameters from the model's\nlatent space. We tested several exploration methods starting with classic\ngradient descents to gradient-free optimizers. Many gradient-free optimizers\njust need comparisons (better/worse than another image), so that they can even\nbe used without numerical criterion, without inspirational image, but with only\nwith human preference. Thus, by iterating on one's preferences we could make\nrobust Facial Composite or Fashion Generation algorithms. High resolution of\nthe produced design generations are obtained using progressive growing of GANs.\nOur results on four datasets of faces, fashion images, and textures show that\nsatisfactory images are effectively retrieved in most cases."", 'As a new way of training generative models, Generative Adversarial Nets (GAN)\nthat uses a discriminative model to guide the training of the generative model\nhas enjoyed considerable success in generating real-valued data. However, it\nhas limitations when the goal is for generating sequences of discrete tokens. A\nmajor reason lies in that the discrete outputs from the generative model make\nit difficult to pass the gradient update from the discriminative model to the\ngenerative model. Also, the discriminative model can only assess a complete\nsequence, while for a partially generated sequence, it is non-trivial to\nbalance its current score and the future one once the entire sequence has been\ngenerated. In this paper, we propose a sequence generation framework, called\nSeqGAN, to solve the problems. Modeling the data generator as a stochastic\npolicy in reinforcement learning (RL), SeqGAN bypasses the generator\ndifferentiation problem by directly performing gradient policy update. The RL\nreward signal comes from the GAN discriminator judged on a complete sequence,\nand is passed back to the intermediate state-action steps using Monte Carlo\nsearch. Extensive experiments on synthetic data and real-world tasks\ndemonstrate significant improvements over strong baselines.', 'Despite recent advances, goal-directed generation of structured discrete data\nremains challenging. For problems such as program synthesis (generating source\ncode) and materials design (generating molecules), finding examples which\nsatisfy desired constraints or exhibit desired properties is difficult. In\npractice, expensive heuristic search or reinforcement learning algorithms are\noften employed. In this paper we investigate the use of conditional generative\nmodels which directly attack this inverse problem, by modeling the distribution\nof discrete structures given properties of interest. Unfortunately, maximum\nlikelihood training of such models often fails with the samples from the\ngenerative model inadequately respecting the input properties. To address this,\nwe introduce a novel approach to directly optimize a reinforcement learning\nobjective, maximizing an expected reward. We avoid high-variance score-function\nestimators that would otherwise be required by sampling from an approximation\nto the normalized rewards, allowing simple Monte Carlo estimation of model\ngradients. We test our methodology on two tasks: generating molecules with\nuser-defined properties and identifying short python expressions which evaluate\nto a given target value. In both cases, we find improvements over maximum\nlikelihood estimation and other baselines.', 'Angry Birds is a popular video game in which the player is provided with a\nsequence of birds to shoot from a slingshot. The task of the game is to destroy\nall green pigs with maximum possible score. Angry Birds appears to be a\ndifficult task to solve for artificially intelligent agents due to the\nsequential decision-making, non-deterministic game environment, enormous state\nand action spaces and requirement to differentiate between multiple birds,\ntheir abilities and optimum tapping times. We describe the application of Deep\nReinforcement learning by implementing Double Dueling Deep Q-network to play\nAngry Birds game. One of our main goals was to build an agent that is able to\ncompete with previous participants and humans on the first 21 levels. In order\nto do so, we have collected a dataset of game frames that we used to train our\nagent on. We present different approaches and settings for DQN agent. We\nevaluate our agent using results of the previous participants of AIBirds\ncompetition, results of volunteer human players and present the results of\nAIBirds 2018 competition.', 'Multiplayer games have long been used as testbeds in artificial intelligence\nresearch, aptly referred to as the Drosophila of artificial intelligence.\nTraditionally, researchers have focused on using well-known games to build\nstrong agents. This progress, however, can be better informed by characterizing\ngames and their topological landscape. Tackling this latter question can\nfacilitate understanding of agents and help determine what game an agent should\ntarget next as part of its training. Here, we show how network measures applied\nto response graphs of large-scale games enable the creation of a landscape of\ngames, quantifying relationships between games of varying sizes and\ncharacteristics. We illustrate our findings in domains ranging from canonical\ngames to complex empirical games capturing the performance of trained agents\npitted against one another. Our results culminate in a demonstration leveraging\nthis information to generate new and interesting games, including mixtures of\nempirical games synthesized from real world games.', 'Domain knowledge is crucial for effective performance in autonomous control\nsystems. Typically, human effort is required to encode this knowledge into a\ncontrol algorithm. In this paper, we present an approach to language grounding\nwhich automatically interprets text in the context of a complex control\napplication, such as a game, and uses domain knowledge extracted from the text\nto improve control performance. Both text analysis and control strategies are\nlearned jointly using only a feedback signal inherent to the application. To\neffectively leverage textual information, our method automatically extracts the\ntext segment most relevant to the current game state, and labels it with a\ntask-centric predicate structure. This labeled text is then used to bias an\naction selection policy for the game, guiding it towards promising regions of\nthe action space. We encode our model for text analysis and game playing in a\nmulti-layer neural network, representing linguistic decisions via latent\nvariables in the hidden layers, and game action quality via the output layer.\nOperating within the Monte-Carlo Search framework, we estimate model parameters\nusing feedback from simulated games. We apply our approach to the complex\nstrategy game Civilization II using the official game manual as the text guide.\nOur results show that a linguistically-informed game-playing agent\nsignificantly outperforms its language-unaware counterpart, yielding a 34%\nabsolute improvement and winning over 65% of games when playing against the\nbuilt-in AI of Civilization.', 'This paper introduces a fully automatic method for generating video game\ntutorials. The AtDELFI system (AuTomatically DEsigning Legible, Full\nInstructions for games) was created to investigate procedural generation of\ninstructions that teach players how to play video games. We present a\nrepresentation of game rules and mechanics using a graph system as well as a\ntutorial generation method that uses said graph representation. We demonstrate\nthe concept by testing it on games within the General Video Game Artificial\nIntelligence (GVG-AI) framework; the paper discusses tutorials generated for\neight different games. Our findings suggest that a graph representation scheme\nworks well for simple arcade style games such as Space Invaders and Pacman, but\nit appears that tutorials for more complex games might require higher-level\nunderstanding of the game than just single mechanics.', 'Policy gradients-based reinforcement learning has proven to be a promising\napproach for directly optimizing non-differentiable evaluation metrics for\nlanguage generation tasks. However, optimizing for a specific metric reward\nleads to improvements in mostly that metric only, suggesting that the model is\ngaming the formulation of that metric in a particular way without often\nachieving real qualitative improvements. Hence, it is more beneficial to make\nthe model optimize multiple diverse metric rewards jointly. While appealing,\nthis is challenging because one needs to manually decide the importance and\nscaling weights of these metric rewards. Further, it is important to consider\nusing a dynamic combination and curriculum of metric rewards that flexibly\nchanges over time. Considering the above aspects, in our work, we automate the\noptimization of multiple metric rewards simultaneously via a multi-armed bandit\napproach (DORB), where at each round, the bandit chooses which metric reward to\noptimize next, based on expected arm gains. We use the Exp3 algorithm for\nbandits and formulate two approaches for bandit rewards: (1) Single\nMulti-reward Bandit (SM-Bandit); (2) Hierarchical Multi-reward Bandit\n(HM-Bandit). We empirically show the effectiveness of our approaches via\nvarious automatic metrics and human evaluation on two important NLG tasks:\nquestion generation and data-to-text generation, including on an unseen-test\ntransfer setup. Finally, we present interpretable analyses of the learned\nbandit curriculum over the optimized rewards.', 'Behavioural cloning, where a computer is taught to perform a task based on\ndemonstrations, has been successfully applied to various video games and\nrobotics tasks, with and without reinforcement learning. This also includes\nend-to-end approaches, where a computer plays a video game like humans do: by\nlooking at the image displayed on the screen, and sending keystrokes to the\ngame. As a general approach to playing video games, this has many inviting\nproperties: no need for specialized modifications to the game, no lengthy\ntraining sessions and the ability to re-use the same tools across different\ngames. However, related work includes game-specific engineering to achieve the\nresults. We take a step towards a general approach and study the general\napplicability of behavioural cloning on twelve video games, including six\nmodern video games (published after 2010), by using human demonstrations as\ntraining data. Our results show that these agents cannot match humans in raw\nperformance but do learn basic dynamics and rules. We also demonstrate how the\nquality of the data matters, and how recording data from humans is subject to a\nstate-action mismatch, due to human reflexes.', 'General Video Game Artificial Intelligence is a general game playing\nframework for Artificial General Intelligence research in the video-games\ndomain. In this paper, we propose for the first time a screen capture learning\nagent for General Video Game AI framework. A Deep Q-Network algorithm was\napplied and improved to develop an agent capable of learning to play different\ngames in the framework. After testing this algorithm using various games of\ndifferent categories and difficulty levels, the results suggest that our\nproposed screen capture learning agent has the potential to learn many\ndifferent games using only a single learning algorithm.', 'Generative Adversarial Networks (GANs) are formulated as minimax game\nproblems, whereby generators attempt to approach real data distributions by\nvirtue of adversarial learning against discriminators. The intrinsic problem\ncomplexity poses the challenge to enhance the performance of generative\nnetworks. In this work, we aim to boost model learning from the perspective of\nnetwork architectures, by incorporating recent progress on automated\narchitecture search into GANs. To this end, we propose a fully differentiable\nsearch framework for generative adversarial networks, dubbed alphaGAN. The\nsearching process is formalized as solving a bi-level minimax optimization\nproblem, in which the outer-level objective aims for seeking a suitable network\narchitecture towards pure Nash Equilibrium conditioned on the generator and the\ndiscriminator network parameters optimized with a traditional GAN loss in the\ninner level. The entire optimization performs a first-order method by\nalternately minimizing the two-level objective in a fully differentiable\nmanner, enabling architecture search to be completed in an enormous search\nspace. Extensive experiments on CIFAR-10 and STL-10 datasets show that our\nalgorithm can obtain high-performing architectures only with 3-GPU hours on a\nsingle GPU in the search space comprised of approximate 2 ? 1011 possible\nconfigurations. We also provide a comprehensive analysis on the behavior of the\nsearching process and the properties of searched architectures, which would\nbenefit further research on architectures for generative models. Pretrained\nmodels and codes are available at https://github.com/yuesongtian/AlphaGAN.', 'Applying generative adversarial networks (GANs) to text-related tasks is\nchallenging due to the discrete nature of language. One line of research\nresolves this issue by employing reinforcement learning (RL) and optimizing the\nnext-word sampling policy directly in a discrete action space. Such methods\ncompute the rewards from complete sentences and avoid error accumulation due to\nexposure bias. Other approaches employ approximation techniques that map the\ntext to continuous representation in order to circumvent the non-differentiable\ndiscrete process. Particularly, autoencoder-based methods effectively produce\nrobust representations that can model complex discrete structures. In this\npaper, we propose a novel text GAN architecture that promotes the collaborative\ntraining of the continuous-space and discrete-space methods. Our method employs\nan autoencoder to learn an implicit data manifold, providing a learning\nobjective for adversarial training in a continuous space. Furthermore, the\ncomplete textual output is directly evaluated and updated via RL in a discrete\nspace. The collaborative interplay between the two adversarial trainings\neffectively regularize the text representations in different spaces. The\nexperimental results on three standard benchmark datasets show that our model\nsubstantially outperforms state-of-the-art text GANs with respect to quality,\ndiversity, and global consistency.', 'Open-domain dialog generation is a challenging problem; maximum likelihood\ntraining can lead to repetitive outputs, models have difficulty tracking\nlong-term conversational goals, and training on standard movie or online\ndatasets may lead to the generation of inappropriate, biased, or offensive\ntext. Reinforcement Learning (RL) is a powerful framework that could\npotentially address these issues, for example by allowing a dialog model to\noptimize for reducing toxicity and repetitiveness. However, previous approaches\nwhich apply RL to open-domain dialog generation do so at the word level, making\nit difficult for the model to learn proper credit assignment for long-term\nconversational rewards. In this paper, we propose a novel approach to\nhierarchical reinforcement learning, VHRL, which uses policy gradients to tune\nthe utterance-level embedding of a variational sequence model. This\nhierarchical approach provides greater flexibility for learning long-term,\nconversational rewards. We use self-play and RL to optimize for a set of\nhuman-centered conversation metrics, and show that our approach provides\nsignificant improvements -- in terms of both human evaluation and automatic\nmetrics -- over state-of-the-art dialog models, including Transformers.', 'Representing games through their pixels offers a promising approach for\nbuilding general-purpose and versatile game models. While games are not merely\nimages, neural network models trained on game pixels often capture differences\nof the visual style of the image rather than the content of the game. As a\nresult, such models cannot generalize well even within similar games of the\nsame genre. In this paper we build on recent advances in contrastive learning\nand showcase its benefits for representation learning in games. Learning to\ncontrast images of games not only classifies games in a more efficient manner;\nit also yields models that separate games in a more meaningful fashion by\nignoring the visual style and focusing, instead, on their content. Our results\nin a large dataset of sports video games containing 100k images across 175\ngames and 10 game genres suggest that contrastive learning is better suited for\nlearning generalized game representations compared to conventional supervised\nlearning. The findings of this study bring us closer to universal visual\nencoders for games that can be reused across previously unseen games without\nrequiring retraining or fine-tuning.', 'Fuzzing is the process of finding security vulnerabilities in\ninput-processing code by repeatedly testing the code with modified inputs. In\nthis paper, we formalize fuzzing as a reinforcement learning problem using the\nconcept of Markov decision processes. This in turn allows us to apply\nstate-of-the-art deep Q-learning algorithms that optimize rewards, which we\ndefine from runtime properties of the program under test. By observing the\nrewards caused by mutating with a specific set of actions performed on an\ninitial program input, the fuzzing agent learns a policy that can next generate\nnew higher-reward inputs. We have implemented this new approach, and\npreliminary empirical evidence shows that reinforcement fuzzing can outperform\nbaseline random fuzzing.', 'The process of playtesting a game is subjective, expensive and incomplete. In\nthis paper, we present a playtesting approach that explores the game space with\nautomated agents and collects data to answer questions posed by the designers.\nRather than have agents interacting with an actual game client, this approach\nrecreates the bare bone mechanics of the game as a separate system. Our agent\nis able to play in minutes what would take testers days of organic gameplay.\nThe analysis of thousands of game simulations exposed imbalances in game\nactions, identified inconsequential rewards and evaluated the effectiveness of\noptional strategic choices. Our test case game, The Sims Mobile, was recently\nreleased and the findings shown here influenced design changes that resulted in\nimproved player experience.', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', 'The automatic generation of game tutorials is a challenging AI problem. While\nit is possible to generate annotations and instructions that explain to the\nplayer how the game is played, this paper focuses on generating a gameplay\nexperience that introduces the player to a game mechanic. It evolves small\nlevels for the Mario AI Framework that can only be beaten by an agent that\nknows how to perform specific actions in the game. It uses variations of a\nperfect A* agent that are limited in various ways, such as not being able to\njump high or see enemies, to test how failing to do certain actions can stop\nthe player from beating the level.', 'In this article, we review recent Deep Learning advances in the context of\nhow they have been applied to play different types of video games such as\nfirst-person shooters, arcade games, and real-time strategy games. We analyze\nthe unique requirements that different game genres pose to a deep learning\nsystem and highlight important open challenges in the context of applying these\nmachine learning methods to video games, such as general game playing, dealing\nwith extremely large decision spaces and sparse rewards.', ""While Generative Adversarial Networks (GANs) show increasing performance and\nthe level of realism is becoming indistinguishable from natural images, this\nalso comes with high demands on data and computation. We show that\nstate-of-the-art GAN models -- such as they are being publicly released by\nresearchers and industry -- can be used for a range of applications beyond\nunconditional image generation. We achieve this by an iterative scheme that\nalso allows gaining control over the image generation process despite the\nhighly non-linear latent spaces of the latest GAN models. We demonstrate that\nthis opens up the possibility to re-use state-of-the-art, difficult to train,\npre-trained GANs with a high level of control even if only black-box access is\ngranted. Our work also raises concerns and awareness that the use cases of a\npublished GAN model may well reach beyond the creators' intention, which needs\nto be taken into account before a full public release. Code is available at\nhttps://github.com/a514514772/hijackgan."", 'Generative Adversarial Networks (GANs) have been extremely successful in\nvarious application domains such as computer vision, medicine, and natural\nlanguage processing. Moreover, transforming an object or person to a desired\nshape become a well-studied research in the GANs. GANs are powerful models for\nlearning complex distributions to synthesize semantically meaningful samples.\nHowever, there is a lack of comprehensive review in this field, especially lack\nof a collection of GANs loss-variant, evaluation metrics, remedies for diverse\nimage generation, and stable training. Given the current fast GANs development,\nin this survey, we provide a comprehensive review of adversarial models for\nimage synthesis. We summarize the synthetic image generation methods, and\ndiscuss the categories including image-to-image translation, fusion image\ngeneration, label-to-image mapping, and text-to-image translation. We organize\nthe literature based on their base models, developed ideas related to\narchitectures, constraints, loss functions, evaluation metrics, and training\ndatasets. We present milestones of adversarial models, review an extensive\nselection of previous works in various categories, and present insights on the\ndevelopment route from the model-based to data-driven methods. Further, we\nhighlight a range of potential future research directions. One of the unique\nfeatures of this review is that all software implementations of these GAN\nmethods and datasets have been collected and made available in one place at\nhttps://github.com/pshams55/GAN-Case-Study.', 'We present an algorithm to directly solve numerous image restoration problems\n(e.g., image deblurring, image dehazing, image deraining, etc.). These problems\nare highly ill-posed, and the common assumptions for existing methods are\nusually based on heuristic image priors. In this paper, we find that these\nproblems can be solved by generative models with adversarial learning. However,\nthe basic formulation of generative adversarial networks (GANs) does not\ngenerate realistic images, and some structures of the estimated images are\nusually not preserved well. Motivated by an interesting observation that the\nestimated results should be consistent with the observed inputs under the\nphysics models, we propose a physics model constrained learning algorithm so\nthat it can guide the estimation of the specific task in the conventional GAN\nframework. The proposed algorithm is trained in an end-to-end fashion and can\nbe applied to a variety of image restoration and related low-level vision\nproblems. Extensive experiments demonstrate that our method performs favorably\nagainst the state-of-the-art algorithms.', ""As we develop more assistive and automated game design systems, the question\nof how these systems should be integrated into game development workflows, and\nhow much adaptation may be required, becomes increasingly important. In this\npaper we explore the impact of software engineering decisions on the ability of\nan automated game design system to understand a game's codebase, generate new\ngame code, and evaluate its work. We argue that a new approach to software\nengineering may be required in order for game developers to fully benefit from\nautomated game designers."", 'A deep neural network (DNN)-based speech enhancement (SE) aiming to maximize\nthe performance of an automatic speech recognition (ASR) system is proposed in\nthis paper. In order to optimize the DNN-based SE model in terms of the\ncharacter error rate (CER), which is one of the metric to evaluate the ASR\nsystem and generally non-differentiable, our method uses two DNNs: one for\nspeech processing and one for mimicking the output CERs derived through an\nacoustic model (AM). Then both of DNNs are alternately optimized in the\ntraining phase. Even if the AM is a black-box, e.g., like one provided by a\nthird-party, the proposed method enables the DNN-based SE model to be optimized\nin terms of the CER since the DNN mimicking the AM is differentiable.\nConsequently, it becomes feasible to build CER-centric SE model that has no\nnegative effect, e.g., additional calculation cost and changing network\narchitecture, on the inference phase since our method is merely a training\nscheme for the existing DNN-based methods. Experimental results show that our\nmethod improved CER by 8.8% relative derived through a black-box AM although\ncertain noise levels are kept.', 'Our paper explores the game theoretic value of the 7-in-a-row game. We reduce\nthe problem to solving a finite board game, which we target using Proof Number\nSearch. We present a number of heuristic improvements to Proof Number Search\nand examine their effect within the context of this particular game. Although\nour paper does not solve the 7-in-a-row game, our experiments indicate that we\nhave made significant progress towards it.', 'This paper describes a new implementation of Planet Wars, designed from the\noutset for Game AI research. The skill-depth of the game makes it a challenge\nfor game-playing agents, and the speed of more than 1 million game ticks per\nsecond enables rapid experimentation and prototyping. The parameterised nature\nof the game together with an interchangeable actuator model make it well suited\nto automated game tuning. The game is designed to be fun to play for humans,\nand is directly playable by General Video Game AI agents.', 'Generative adversarial networks (GANs) has gained tremendous popularity\nlately due to an ability to reinforce quality of its predictive model with\ngenerated objects and the quality of the generative model with and supervised\nfeedback. GANs allow to synthesize images with a high degree of realism.\nHowever, the learning process of such models is a very complicated optimization\nproblem and certain limitation for such models were found. It affects the\nchoice of certain layers and nonlinearities when designing architectures. In\nparticular, it does not allow to train convolutional GAN models with\nfully-connected hidden layers. In our work, we propose a modification of the\npreviously described set of rules, as well as new approaches to designing\narchitectures that will allow us to train more powerful GAN models. We show the\neffectiveness of our methods on the problem of synthesizing projections of 3D\nobjects with the possibility of interpolation by class and view point.', 'One of the most significant challenges in statistical signal processing and\nmachine learning is how to obtain a generative model that can produce samples\nof large-scale data distribution, such as images and speeches. Generative\nAdversarial Network (GAN) is an effective method to address this problem. The\nGANs provide an appropriate way to learn deep representations without\nwidespread use of labeled training data. This approach has attracted the\nattention of many researchers in computer vision since it can generate a large\namount of data without precise modeling of the probability density function\n(PDF). In GANs, the generative model is estimated via a competitive process\nwhere the generator and discriminator networks are trained simultaneously. The\ngenerator learns to generate plausible data, and the discriminator learns to\ndistinguish fake data created by the generator from real data samples. Given\nthe rapid growth of GANs over the last few years and their application in\nvarious fields, it is necessary to investigate these networks accurately. In\nthis paper, after introducing the main concepts and the theory of GAN, two new\ndeep generative models are compared, the evaluation metrics utilized in the\nliterature and challenges of GANs are also explained. Moreover, the most\nremarkable GAN architectures are categorized and discussed. Finally, the\nessential applications in computer vision are examined.', 'Differentiable rendering has paved the way to training neural networks to\nperform ""inverse graphics"" tasks such as predicting 3D geometry from monocular\nphotographs. To train high performing models, most of the current approaches\nrely on multi-view imagery which are not readily available in practice. Recent\nGenerative Adversarial Networks (GANs) that synthesize images, in contrast,\nseem to acquire 3D knowledge implicitly during training: object viewpoints can\nbe manipulated by simply manipulating the latent codes. However, these latent\ncodes often lack further physical interpretation and thus GANs cannot easily be\ninverted to perform explicit 3D reasoning. In this paper, we aim to extract and\ndisentangle 3D knowledge learned by generative models by utilizing\ndifferentiable renderers. Key to our approach is to exploit GANs as a\nmulti-view data generator to train an inverse graphics network using an\noff-the-shelf differentiable renderer, and the trained inverse graphics network\nas a teacher to disentangle the GAN\'s latent code into interpretable 3D\nproperties. The entire architecture is trained iteratively using cycle\nconsistency losses. We show that our approach significantly outperforms\nstate-of-the-art inverse graphics networks trained on existing datasets, both\nquantitatively and via user studies. We further showcase the disentangled GAN\nas a controllable 3D ""neural renderer"", complementing traditional graphics\nrenderers.', 'Generative Adversarial Networks (GANs) are commonly used for modeling complex\ndistributions of data. Both the generators and discriminators of GANs are often\nmodeled by neural networks, posing a non-transparent optimization problem which\nis non-convex and non-concave over the generator and discriminator,\nrespectively. Such networks are often heuristically optimized with gradient\ndescent-ascent (GDA), but it is unclear whether the optimization problem\ncontains any saddle points, or whether heuristic methods can find them in\npractice. In this work, we analyze the training of Wasserstein GANs with\ntwo-layer neural network discriminators through the lens of convex duality, and\nfor a variety of generators expose the conditions under which Wasserstein GANs\ncan be solved exactly with convex optimization approaches, or can be\nrepresented as convex-concave games. Using this convex duality interpretation,\nwe further demonstrate the impact of different activation functions of the\ndiscriminator. Our observations are verified with numerical results\ndemonstrating the power of the convex interpretation, with applications in\nprogressive training of convex architectures corresponding to linear generators\nand quadratic-activation discriminators for CelebA image generation. The code\nfor our experiments is available at https://github.com/ardasahiner/ProCoGAN.', ""The Generative Models have gained considerable attention in the field of\nunsupervised learning via a new and practical framework called Generative\nAdversarial Networks (GAN) due to its outstanding data generation capability.\nMany models of GAN have proposed, and several practical applications emerged in\nvarious domains of computer vision and machine learning. Despite GAN's\nexcellent success, there are still obstacles to stable training. The problems\nare due to Nash-equilibrium, internal covariate shift, mode collapse, vanishing\ngradient, and lack of proper evaluation metrics. Therefore, stable training is\na crucial issue in different applications for the success of GAN. Herein, we\nsurvey several training solutions proposed by different researchers to\nstabilize GAN training. We survey, (I) the original GAN model and its modified\nclassical versions, (II) detail analysis of various GAN applications in\ndifferent domains, (III) detail study about the various GAN training obstacles\nas well as training solutions. Finally, we discuss several new issues as well\nas research outlines to the topic."", 'Rehabilitation robotics combined with video game technology provides a means\nof assisting in the rehabilitation of patients with neuromuscular disorders by\nperforming various facilitation movements. The current work presents ReHabGame,\na serious game using a fusion of implemented technologies that can be easily\nused by patients and therapists to assess and enhance sensorimotor performance\nand also increase the activities in the daily lives of patients. The game\nallows a player to control avatar movements through a Kinect Xbox, Myo armband\nand rudder foot pedal, and involves a series of reach-grasp-collect tasks whose\ndifficulty levels are learnt by a fuzzy interface. The orientation, angular\nvelocity, head and spine tilts and other data generated by the player are\nmonitored and saved, whilst the task completion is calculated by solving an\ninverse kinematics algorithm which orientates the upper limb joints of the\navatar. The different values in upper body quantities of movement provide fuzzy\ninput from which crisp output is determined and used to generate an appropriate\nsubsequent rehabilitation game level. The system can thus provide personalised,\nautonomously-learnt rehabilitation programmes for patients with neuromuscular\ndisorders with superior predictions to guide the development of improved\nclinical protocols compared to traditional theraputic activities.', 'In this paper we introduce DeepCrawl, a fully-playable Roguelike prototype\nfor iOS and Android in which all agents are controlled by policy networks\ntrained using Deep Reinforcement Learning (DRL). Our aim is to understand\nwhether recent advances in DRL can be used to develop convincing behavioral\nmodels for non-player characters in videogames. We begin with an analysis of\nrequirements that such an AI system should satisfy in order to be practically\napplicable in video game development, and identify the elements of the DRL\nmodel used in the DeepCrawl prototype. The successes and limitations of\nDeepCrawl are documented through a series of playability tests performed on the\nfinal game. We believe that the techniques we propose offer insight into\ninnovative new avenues for the development of behaviors for non-player\ncharacters in video games, as they offer the potential to overcome critical\nissues with', 'Many games often share common ideas or aspects between them, such as their\nrules, controls, or playing area. However, in the context of General Game\nPlaying (GGP) for board games, this area remains under-explored. We propose to\nformalise the notion of ""game concept"", inspired by terms generally used by\ngame players and designers. Through the Ludii General Game System, we describe\nconcepts for several levels of abstraction, such as the game itself, the moves\nplayed, or the states reached. This new GGP feature associated with the ludeme\nrepresentation of games opens many new lines of research. The creation of a\nhyper-agent selector, the transfer of AI learning between games, or explaining\nAI techniques using game terms, can all be facilitated by the use of game\nconcepts. Other applications which can benefit from game concepts are also\ndiscussed, such as the generation of plausible reconstructed rules for\nincomplete ancient games, or the implementation of a board game recommender\nsystem.', 'Coreference evaluation metrics are hard to optimize directly as they are\nnon-differentiable functions, not easily decomposable into elementary\ndecisions. Consequently, most approaches optimize objectives only indirectly\nrelated to the end goal, resulting in suboptimal performance. Instead, we\npropose a differentiable relaxation that lends itself to gradient-based\noptimisation, thus bypassing the need for reinforcement learning or heuristic\nmodification of cross-entropy. We show that by modifying the training objective\nof a competitive neural coreference system, we obtain a substantial gain in\nperformance. This suggests that our approach can be regarded as a viable\nalternative to using reinforcement learning or more computationally expensive\nimitation learning.', 'If you are an artificial intelligence researcher, you should look to video\ngames as ideal testbeds for the work you do. If you are a video game developer,\nyou should look to AI for the technology that makes completely new types of\ngames possible. This chapter lays out the case for both of these propositions.\nIt asks the question ""what can video games do for AI"", and discusses how in\nparticular general video game playing is the ideal testbed for artificial\ngeneral intelligence research. It then asks the question ""what can AI do for\nvideo games"", and lays out a vision for what video games might look like if we\nhad significantly more advanced AI at our disposal. The chapter is based on my\nkeynote at IJCCI 2015, and is written in an attempt to be accessible to a broad\naudience.', 'This project proposes a methodology for the automatic generation of action\nmodels from video game dynamics descriptions, as well as its integration with a\nplanning agent for the execution and monitoring of the plans. Planners use\nthese action models to get the deliberative behaviour for an agent in many\ndifferent video games and, combined with a reactive module, solve deterministic\nand no-deterministic levels. Experimental results validate the methodology and\nprove that the effort put by a knowledge engineer can be greatly reduced in the\ndefinition of such complex domains. Furthermore, benchmarks of the domains has\nbeen produced that can be of interest to the international planning community\nto evaluate planners in international planning competitions.', 'Over the past few years the Angry Birds AI competition has been held in an\nattempt to develop intelligent agents that can successfully and efficiently\nsolve levels for the video game Angry Birds. Many different agents and\nstrategies have been developed to solve the complex and challenging physical\nreasoning problems associated with such a game. However none of these agents\nattempt one of the key strategies which humans employ to solve Angry Birds\nlevels, which is restarting levels. Restarting is important in Angry Birds\nbecause sometimes the level is no longer solvable or some given shot made has\nlittle to no benefit towards the ultimate goal of the game. This paper proposes\na framework and experimental evaluation for when to restart levels in Angry\nBirds. We demonstrate that restarting is a viable strategy to improve agent\nperformance in many cases.', 'We propose the study of mathematical ludology, which aims to formally\ninterrogate questions of interest to game studies and game design in\nparticular. The goal is to extend our mathematical understanding of complex\ngames beyond decision-making---the typical focus of game theory and artificial\nintelligence efforts---to explore other aspects such as game mechanics,\nstructure, relationships between games, and connections between game rules and\nuser-interfaces, as well as exploring related gameplay phenomena and typical\nplayer behavior. In this paper, we build a basic foundation for this line of\nstudy by developing a hierarchy of game descriptions, mathematical formalism to\ncompactly describe complex discrete games, and equivalence relations on the\nspace of game systems.', 'Game-playing AI research has focused for a long time on learning to play\nvideo games from visual input or symbolic information. However, humans benefit\nfrom a wider array of sensors which we utilise in order to navigate the world\naround us. In particular, sounds and music are key to how many of us perceive\nthe world and influence the decisions we make. In this paper, we present\ninitial experiments on game-playing agents learning to play video games solely\nfrom audio cues. We expand the Video Game Description Language to allow for\naudio specification, and the General Video Game AI framework to provide new\naudio games and an API for learning agents to make use of audio observations.\nWe analyse the games and the audio game design process, include initial results\nwith simple Q~Learning agents, and encourage further research in this area.', ""Class labels have been empirically shown useful in improving the sample\nquality of generative adversarial nets (GANs). In this paper, we mathematically\nstudy the properties of the current variants of GANs that make use of class\nlabel information. With class aware gradient and cross-entropy decomposition,\nwe reveal how class labels and associated losses influence GAN's training.\nBased on that, we propose Activation Maximization Generative Adversarial\nNetworks (AM-GAN) as an advanced solution. Comprehensive experiments have been\nconducted to validate our analysis and evaluate the effectiveness of our\nsolution, where AM-GAN outperforms other strong baselines and achieves\nstate-of-the-art Inception Score (8.91) on CIFAR-10. In addition, we\ndemonstrate that, with the Inception ImageNet classifier, Inception Score\nmainly tracks the diversity of the generator, and there is, however, no\nreliable evidence that it can reflect the true sample quality. We thus propose\na new metric, called AM Score, to provide a more accurate estimation of the\nsample quality. Our proposed model also outperforms the baseline methods in the\nnew metric."", 'Generative Adversarial Networks (GANs) is a novel class of deep generative\nmodels which has recently gained significant attention. GANs learns complex and\nhigh-dimensional distributions implicitly over images, audio, and data.\nHowever, there exists major challenges in training of GANs, i.e., mode\ncollapse, non-convergence and instability, due to inappropriate design of\nnetwork architecture, use of objective function and selection of optimization\nalgorithm. Recently, to address these challenges, several solutions for better\ndesign and optimization of GANs have been investigated based on techniques of\nre-engineered network architectures, new objective functions and alternative\noptimization algorithms. To the best of our knowledge, there is no existing\nsurvey that has particularly focused on broad and systematic developments of\nthese solutions. In this study, we perform a comprehensive survey of the\nadvancements in GANs design and optimization solutions proposed to handle GANs\nchallenges. We first identify key research issues within each design and\noptimization technique and then propose a new taxonomy to structure solutions\nby key research issues. In accordance with the taxonomy, we provide a detailed\ndiscussion on different GANs variants proposed within each solution and their\nrelationships. Finally, based on the insights gained, we present the promising\nresearch directions in this rapidly growing field.', ""Generative adversarial network (GAN) is among the most popular deep learning\nmodels for learning complex data distributions. However, training a GAN is\nknown to be a challenging task. This is often attributed to the lack of\ncorrelation between the training progress and the trajectory of the generator\nand discriminator losses and the need for the GAN's subjective evaluation. A\nrecently proposed measure inspired by game theory - the duality gap, aims to\nbridge this gap. However, as we demonstrate, the duality gap's capability\nremains constrained due to limitations posed by its estimation process. This\npaper presents a theoretical understanding of this limitation and proposes a\nmore dependable estimation process for the duality gap. At the crux of our\napproach is the idea that local perturbations can help agents in a zero-sum\ngame escape non-Nash saddle points efficiently. Through exhaustive\nexperimentation across GAN models and datasets, we establish the efficacy of\nour approach in capturing the GAN training progress with minimal increase to\nthe computational complexity. Further, we show that our estimate, with its\nability to identify model convergence/divergence, is a potential performance\nmeasure that can be used to tune the hyperparameters of a GAN."", ""Generative Adversarial Networks (GANs) have become a dominant class of\ngenerative models. In recent years, GAN variants have yielded especially\nimpressive results in the synthesis of a variety of forms of data. Examples\ninclude compelling natural and artistic images, textures, musical sequences,\nand 3D object files. However, one obvious synthesis candidate is missing. In\nthis work, we answer one of deep learning's most pressing questions: GAN you do\nthe GAN GAN? That is, is it possible to train a GAN to model a distribution of\nGANs? We release the full source code for this project under the MIT license."", 'Generative Adversarial Networks (GAN) have received wide attention in the\nmachine learning field for their potential to learn high-dimensional, complex\nreal data distribution. Specifically, they do not rely on any assumptions about\nthe distribution and can generate real-like samples from latent space in a\nsimple manner. This powerful property leads GAN to be applied to various\napplications such as image synthesis, image attribute editing, image\ntranslation, domain adaptation and other academic fields. In this paper, we aim\nto discuss the details of GAN for those readers who are familiar with, but do\nnot comprehend GAN deeply or who wish to view GAN from various perspectives. In\naddition, we explain how GAN operates and the fundamental meaning of various\nobjective functions that have been suggested recently. We then focus on how the\nGAN can be combined with an autoencoder framework. Finally, we enumerate the\nGAN variants that are applied to various tasks and other fields for those who\nare interested in exploiting GAN for their research.', 'This paper describes a new algorithm for decision making in two-player\nreal-time video games. As with Monte Carlo Tree Search, the algorithm can be\nused without heuristics and has been developed for use in general video game\nAI. The approach is to extend recent work on rolling horizon evolutionary\nplanning, which has been shown to work well for single-player games, to two (or\nin principle many) player games. To select an action the algorithm co-evolves\ntwo (or in the general case N) populations, one for each player, where each\nindividual is a sequence of actions for the respective player. The fitness of\neach individual is evaluated by playing it against a selection of\naction-sequences from the opposing population. When choosing an action to take\nin the game, the first action is chosen from the fittest member of the\npopulation for that player. The new algorithm is compared with a number of\ngeneral video game AI algorithms on three variations of a two-player space\nbattle game, with promising results.', 'An approach for game bot detection in MMORPGs is proposed based on the\nanalysis of game playing behavior. Since MMORPGs are large scale games, users\ncan play in various ways. This variety in playing behavior makes it hard to\ndetect game bots based on play behaviors. In order to cope with this problem,\nthe proposed approach observes game playing behaviors of users and groups them\nby their behavioral similarities. Then, it develops a local bot detection model\nfor each player group. Since the locally optimized models can more accurately\ndetect game bots within each player group, the combination of those models\nbrings about overall improvement. For a practical purpose of reducing the\nworkloads of the game servers in service, the game data is collected at a low\nresolution in time. Behavioral features are selected and developed to\naccurately detect game bots with the low resolution data, considering common\naspects of MMORPG playing. Through the experiment with the real data from a\ngame currently in service, it is shown that the proposed local model approach\nyields more accurate results.', ""The game industry has long been troubled by malicious activities utilizing\ngame bots. The game bots disturb other game players and destroy the\nenvironmental system of the games. For these reasons, the game industry put\ntheir best efforts to detect the game bots among players' characters using the\nlearning-based detections. However, one problem with the detection\nmethodologies is that they do not provide rational explanations about their\ndecisions. To resolve this problem, in this work, we investigate the\nexplainabilities of the game bot detection. We develop the XAI model using a\ndataset from the Korean MMORPG, AION, which includes game logs of human players\nand game bots. More than one classification model has been applied to the\ndataset to be analyzed by applying interpretable models. This provides us\nexplanations about the game bots' behavior, and the truthfulness of the\nexplanations has been evaluated. Besides, interpretability contributes to\nminimizing false detection, which imposes unfair restrictions on human players."", ""This paper presents TotalBotWar, a new pseudo real-time multi-action\nchallenge for game AI, as well as some initial experiments that benchmark the\nframework with different agents. The game is based on the real-time battles of\nthe popular TotalWar games series where players manage an army to defeat the\nopponent's one. In the proposed game, a turn consists of a set of orders to\ncontrol the units. The number and specific orders that can be performed in a\nturn vary during the progression of the game. One interesting feature of the\ngame is that if a particular unit does not receive an order in a turn, it will\ncontinue performing the action specified in a previous turn. The turn-wise\nbranching factor becomes overwhelming for traditional algorithms and the\npartial observability of the game state makes the proposed game an interesting\nplatform to test modern AI algorithms."", 'Modern board games are a rich source of interesting and new challenges for\ncombinatorial problems. The game Nmbr9 is a solitaire style puzzle game using\npolyominoes. The rules of the game are simple to explain, but modelling the\ngame effectively using constraint programming is hard. This abstract presents\nthe game, contributes new generalized variants of the game suitable for\nbenchmarking and testing, and describes a model for the presented variants. The\nquestion of the top possible score in the standard game is an open challenge.', 'Cross-modal retrieval between videos and texts has attracted growing\nattentions due to the rapid emergence of videos on the web. The current\ndominant approach for this problem is to learn a joint embedding space to\nmeasure cross-modal similarities. However, simple joint embeddings are\ninsufficient to represent complicated visual and textual details, such as\nscenes, objects, actions and their compositions. To improve fine-grained\nvideo-text retrieval, we propose a Hierarchical Graph Reasoning (HGR) model,\nwhich decomposes video-text matching into global-to-local levels. To be\nspecific, the model disentangles texts into hierarchical semantic graph\nincluding three levels of events, actions, entities and relationships across\nlevels. Attention-based graph reasoning is utilized to generate hierarchical\ntextual embeddings, which can guide the learning of diverse and hierarchical\nvideo representations. The HGR model aggregates matchings from different\nvideo-text levels to capture both global and local details. Experimental\nresults on three video-text datasets demonstrate the advantages of our model.\nSuch hierarchical decomposition also enables better generalization across\ndatasets and improves the ability to distinguish fine-grained semantic\ndifferences.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : As a healthcare organization, our goal is to create a system that can effectively extract crucial information from Electronic Health Records (EHRs) to enhance patient care and medical research. Given the challenge of limited labeled data in EHRs, we intend to utilize transfer learning and cross-lingual methods to train a model for this task. Consequently, we are interested in identifying tasks that bear the highest similarity to information extraction from EHRs. Due to privacy concerns, we are unable to test the model on actual EHRs. Therefore, we require methods to generate datasets that closely resemble our format, using fictitious patient information, to test our model. Furthermore, since some EHRs may not convert doctors’ notes into text, the model should be capable of interpreting the doctor's handwriting and accurately extracting information from these notes. We have a preference for pre-trained models that have been benchmarked or models that come with training codes.","[""Electronic Health Records (EHRs) are commonly used by the machine learning\ncommunity for research on problems specifically related to health care and\nmedicine. EHRs have the advantages that they can be easily distributed and\ncontain many features useful for e.g. classification problems. What makes EHR\ndata sets different from typical machine learning data sets is that they are\noften very sparse, due to their high dimensionality, and often contain\nheterogeneous (mixed) data types. Furthermore, the data sets deal with\nsensitive information, which limits the distribution of any models learned\nusing them, due to privacy concerns. For these reasons, using EHR data in\npractice presents a real challenge. In this work, we explore using Generative\nAdversarial Networks to generate synthetic, heterogeneous EHRs with the goal of\nusing these synthetic records in place of existing data sets for downstream\nclassification tasks. We will further explore applying differential privacy\n(DP) preserving optimization in order to produce DP synthetic EHR data sets,\nwhich provide rigorous privacy guarantees, and are therefore shareable and\nusable in the real world. The performance (measured by AUROC, AUPRC and\naccuracy) of our model's synthetic, heterogeneous data is very close to the\noriginal data set (within 3 - 5% of the baseline) for the non-DP model when\ntested in a binary classification task. Using strong $(1, 10^{-5})$ DP, our\nmodel still produces data useful for machine learning tasks, albeit incurring a\nroughly 17% performance penalty in our tested classification task. We\nadditionally perform a sub-population analysis and find that our model does not\nintroduce any bias into the synthetic EHR data compared to the baseline in\neither male/female populations, or the 0-18, 19-50 and 51+ age groups in terms\nof classification performance for either the non-DP or DP variant."", 'Extractive summarization is very useful for physicians to better manage and\ndigest Electronic Health Records (EHRs). However, the training of a supervised\nmodel requires disease-specific medical background and is thus very expensive.\nWe studied how to utilize the intrinsic correlation between multiple EHRs to\ngenerate pseudo-labels and train a supervised model with no external\nannotation. Experiments on real-patient data validate that our model is\neffective in summarizing crucial disease-specific information for patients.', 'Electronic health records (EHRs), digital collections of patient healthcare\nevents and observations, are ubiquitous in medicine and critical to healthcare\ndelivery, operations, and research. Despite this central role, EHRs are\nnotoriously difficult to process automatically. Well over half of the\ninformation stored within EHRs is in the form of unstructured text (e.g.\nprovider notes, operation reports) and remains largely untapped for secondary\nuse. Recently, however, newer neural network and deep learning approaches to\nNatural Language Processing (NLP) have made considerable advances,\noutperforming traditional statistical and rule-based systems on a variety of\ntasks. In this survey paper, we summarize current neural NLP methods for EHR\napplications. We focus on a broad scope of tasks, namely, classification and\nprediction, word embeddings, extraction, generation, and other topics such as\nquestion answering, phenotyping, knowledge graphs, medical dialogue,\nmultilinguality, interpretability, etc.', 'Researchers require timely access to real-world longitudinal electronic\nhealth records (EHR) to develop, test, validate, and implement machine learning\nsolutions that improve the quality and efficiency of healthcare. In contrast,\nhealth systems value deeply patient privacy and data security. De-identified\nEHRs do not adequately address the needs of health systems, as de-identified\ndata are susceptible to re-identification and its volume is also limited.\nSynthetic EHRs offer a potential solution. In this paper, we propose EHR\nVariational Autoencoder (EVA) for synthesizing sequences of discrete EHR\nencounters (e.g., clinical visits) and encounter features (e.g., diagnoses,\nmedications, procedures). We illustrate that EVA can produce realistic EHR\nsequences, account for individual differences among patients, and can be\nconditioned on specific disease conditions, thus enabling disease-specific\nstudies. We design efficient, accurate inference algorithms by combining\nstochastic gradient Markov Chain Monte Carlo with amortized variational\ninference. We assess the utility of the methods on large real-world EHR\nrepositories containing over 250, 000 patients. Our experiments, which include\nuser studies with knowledgeable clinicians, indicate the generated EHR\nsequences are realistic. We confirmed the performance of predictive models\ntrained on the synthetic data are similar with those trained on real EHRs.\nAdditionally, our findings indicate that augmenting real data with synthetic\nEHRs results in the best predictive performance - improving the best baseline\nby as much as 8% in top-20 recall.', ""Electronic Health Records (EHRs) provide vital contextual information to\nradiologists and other physicians when making a diagnosis. Unfortunately,\nbecause a given patient's record may contain hundreds of notes and reports,\nidentifying relevant information within these in the short time typically\nallotted to a case is very difficult. We propose and evaluate models that\nextract relevant text snippets from patient records to provide a rough case\nsummary intended to aid physicians considering one or more diagnoses. This is\nhard because direct supervision (i.e., physician annotations of snippets\nrelevant to specific diagnoses in medical records) is prohibitively expensive\nto collect at scale. We propose a distantly supervised strategy in which we use\ngroups of International Classification of Diseases (ICD) codes observed in\n'future' records as noisy proxies for 'downstream' diagnoses. Using this we\ntrain a transformer-based neural model to perform extractive summarization\nconditioned on potential diagnoses. This model defines an attention mechanism\nthat is conditioned on potential diagnoses (queries) provided by the diagnosing\nphysician. We train (via distant supervision) and evaluate variants of this\nmodel on EHR data from Brigham and Women's Hospital in Boston and MIMIC-III\n(the latter to facilitate reproducibility). Evaluations performed by\nradiologists demonstrate that these distantly supervised models yield better\nextractive summaries than do unsupervised approaches. Such models may aid\ndiagnosis by identifying sentences in past patient reports that are clinically\nrelevant to a potential diagnosis."", ""The field of clinical natural language processing has been advanced\nsignificantly since the introduction of deep learning models. The\nself-supervised representation learning and the transfer learning paradigm\nbecame the methods of choice in many natural language processing application,\nin particular in the settings with the dearth of high quality manually\nannotated data. Electronic health record systems are ubiquitous and the\nmajority of patients' data are now being collected electronically and in\nparticular in the form of free text. Identification of medical concepts and\ninformation extraction is a challenging task, yet important ingredient for\nparsing unstructured data into structured and tabulated format for downstream\nanalytical tasks. In this work we introduced a named-entity recognition model\nfor clinical natural language processing. The model is trained to recognise\nseven categories: drug names, route, frequency, dosage, strength, form,\nduration. The model was first self-supervisedly pre-trained by predicting the\nnext word, using a collection of 2 million free-text patients' records from\nMIMIC-III corpora and then fine-tuned on the named-entity recognition task. The\nmodel achieved a lenient (strict) micro-averaged F1 score of 0.957 (0.893)\nacross all seven categories. Additionally, we evaluated the transferability of\nthe developed model using the data from the Intensive Care Unit in the US to\nsecondary care mental health records (CRIS) in the UK. A direct application of\nthe trained NER model to CRIS data resulted in reduced performance of F1=0.762,\nhowever after fine-tuning on a small sample from CRIS, the model achieved a\nreasonable performance of F1=0.944. This demonstrated that despite a close\nsimilarity between the data sets and the NER tasks, it is essential to\nfine-tune on the target domain data in order to achieve more accurate results."", 'In this paper, we present our approach to extracting structured information\nfrom unstructured Electronic Health Records (EHR) [2] which can be used to, for\nexample, study adverse drug reactions in patients due to chemicals in their\nproducts. Our solution uses a combination of Natural Language Processing (NLP)\ntechniques and a web-based annotation tool to optimize the performance of a\ncustom Named Entity Recognition (NER) [1] model trained on a limited amount of\nEHR training data. This work was presented at the first Health Search and Data\nMining Workshop (HSDM 2020) [26]. We showcase a combination of tools and\ntechniques leveraging the recent advancements in NLP aimed at targeting domain\nshifts by applying transfer learning and language model pre-training techniques\n[3]. We present a comparison of our technique to the current popular approaches\nand show the effective increase in performance of the NER model and the\nreduction in time to annotate data.A key observation of the results presented\nis that the F1 score of model (0.734) trained with our approach with just 50%\nof available training data outperforms the F1 score of the blank spaCy model\nwithout language model component (0.704) trained with 100% of the available\ntraining data. We also demonstrate an annotation tool to minimize domain expert\ntime and the manual effort required to generate such a training dataset.\nFurther, we plan to release the annotated dataset as well as the pre-trained\nmodel to the community to further research in medical health records.', 'Deep learning models have shown tremendous potential in learning\nrepresentations, which are able to capture some key properties of the data.\nThis makes them great candidates for transfer learning: Exploiting\ncommonalities between different learning tasks to transfer knowledge from one\ntask to another. Electronic health records (EHR) research is one of the domains\nthat has witnessed a growing number of deep learning techniques employed for\nlearning clinically-meaningful representations of medical concepts (such as\ndiseases and medications). Despite this growth, the approaches to benchmark and\nassess such learned representations (or, embeddings) is under-investigated;\nthis can be a big issue when such embeddings are shared to facilitate transfer\nlearning. In this study, we aim to (1) train some of the most prominent disease\nembedding techniques on a comprehensive EHR data from 3.1 million patients, (2)\nemploy qualitative and quantitative evaluation techniques to assess these\nembeddings, and (3) provide pre-trained disease embeddings for transfer\nlearning. This study can be the first comprehensive approach for clinical\nconcept embedding evaluation and can be applied to any embedding techniques and\nfor any EHR concept.', 'Electronic Health Record (EHR) data has been of tremendous utility in\nArtificial Intelligence (AI) for healthcare such as predicting future clinical\nevents. These tasks, however, often come with many challenges when using\nclassical machine learning models due to a myriad of factors including class\nimbalance and data heterogeneity (i.e., the complex intra-class variances). To\naddress some of these research gaps, this paper leverages the exciting\ncontrastive learning framework and proposes a novel contrastive regularized\nclinical classification model. The contrastive loss is found to substantially\naugment EHR-based prediction: it effectively characterizes the\nsimilar/dissimilar patterns (by its ""push-and-pull"" form), meanwhile mitigating\nthe highly skewed class distribution by learning more balanced feature spaces\n(as also echoed by recent findings). In particular, when naively exporting the\ncontrastive learning to the EHR data, one hurdle is in generating positive\nsamples, since EHR data is not as amendable to data augmentation as image data.\nTo this end, we have introduced two unique positive sampling strategies\nspecifically tailored for EHR data: a feature-based positive sampling that\nexploits the feature space neighborhood structure to reinforce the feature\nlearning; and an attribute-based positive sampling that incorporates\npre-generated patient similarity metrics to define the sample proximity. Both\nsampling approaches are designed with an awareness of unique high intra-class\nvariance in EHR data. Our overall framework yields highly competitive\nexperimental results in predicting the mortality risk on real-world COVID-19\nEHR data with a total of 5,712 patients admitted to a large, urban health\nsystem. Specifically, our method reaches a high AUROC prediction score of\n0.959, which outperforms other baselines and alternatives: cross-entropy(0.873)\nand focal loss(0.931).', 'The recent adoption of Electronic Health Records (EHRs) by health care\nproviders has introduced an important source of data that provides detailed and\nhighly specific insights into patient phenotypes over large cohorts. These\ndatasets, in combination with machine learning and statistical approaches,\ngenerate new opportunities for research and clinical care. However, many\nmethods require the patient representations to be in structured formats, while\nthe information in the EHR is often locked in unstructured texts designed for\nhuman readability. In this work, we develop the methodology to automatically\nextract clinical features from clinical narratives from large EHR corpora\nwithout the need for prior knowledge. We consider medical terms and sentences\nappearing in clinical narratives as atomic information units. We propose an\nefficient clustering strategy suitable for the analysis of large text corpora\nand to utilize the clusters to represent information about the patient\ncompactly. To demonstrate the utility of our approach, we perform an\nassociation study of clinical features with somatic mutation profiles from\n4,007 cancer patients and their tumors. We apply the proposed algorithm to a\ndataset consisting of about 65 thousand documents with a total of about 3.2\nmillion sentences. We identify 341 significant statistical associations between\nthe presence of somatic mutations and clinical features. We annotated these\nassociations according to their novelty, and report several known associations.\nWe also propose 32 testable hypotheses where the underlying biological\nmechanism does not appear to be known but plausible. These results illustrate\nthat the automated discovery of clinical features is possible and the joint\nanalysis of clinical and genetic datasets can generate appealing new\nhypotheses.']","['Automatically extracting useful information from electronic medical records\nalong with conducting disease diagnoses is a promising task for both clinical\ndecision support(CDS) and neural language processing(NLP). Most of the existing\nsystems are based on artificially constructed knowledge bases, and then\nauxiliary diagnosis is done by rule matching. In this study, we present a\nclinical intelligent decision approach based on Convolutional Neural\nNetworks(CNN), which can automatically extract high-level semantic information\nof electronic medical records and then perform automatic diagnosis without\nartificial construction of rules or knowledge bases. We use collected 18,590\ncopies of the real-world clinical electronic medical records to train and test\nthe proposed model. Experimental results show that the proposed model can\nachieve 98.67\\% accuracy and 96.02\\% recall, which strongly supports that using\nconvolutional neural network to automatically learn high-level semantic\nfeatures of electronic medical records and then conduct assist diagnosis is\nfeasible and effective.', 'Sharing electronic health records (EHRs) on a large scale may lead to privacy\nintrusions. Recent research has shown that risks may be mitigated by simulating\nEHRs through generative adversarial network (GAN) frameworks. Yet the methods\ndeveloped to date are limited because they 1) focus on generating data of a\nsingle type (e.g., diagnosis codes), neglecting other data types (e.g.,\ndemographics, procedures or vital signs) and 2) do not represent constraints\nbetween features. In this paper, we introduce a method to simulate EHRs\ncomposed of multiple data types by 1) refining the GAN model, 2) accounting for\nfeature constraints, and 3) incorporating key utility measures for such\ngeneration tasks. Our analysis with over $770,000$ EHRs from Vanderbilt\nUniversity Medical Center demonstrates that the new model achieves higher\nperformance in terms of retaining basic statistics, cross-feature correlations,\nlatent structural properties, feature constraints and associated patterns from\nreal data, without sacrificing privacy.', 'The wide adoption of Electronic Health Records (EHR) has resulted in large\namounts of clinical data becoming available, which promises to support service\ndelivery and advance clinical and informatics research. Deep learning\ntechniques have demonstrated performance in predictive analytic tasks using\nEHRs yet they typically lack model result transparency or explainability\nfunctionalities and require cumbersome pre-processing tasks. Moreover, EHRs\ncontain heterogeneous and multi-modal data points such as text, numbers and\ntime series which further hinder visualisation and interpretability. This paper\nproposes a deep learning framework to: 1) encode patient pathways from EHRs\ninto images, 2) highlight important events within pathway images, and 3) enable\nmore complex predictions with additional intelligibility. The proposed method\nrelies on a deep attention mechanism for visualisation of the predictions and\nallows predicting multiple sequential outcomes.', 'Evaluating the clinical similarities between pairwise patients is a\nfundamental problem in healthcare informatics. A proper patient similarity\nmeasure enables various downstream applications, such as cohort study and\ntreatment comparative effectiveness research. One major carrier for conducting\npatient similarity research is Electronic Health Records(EHRs), which are\nusually heterogeneous, longitudinal, and sparse. Though existing studies on\nlearning patient similarity from EHRs have shown being useful in solving real\nclinical problems, their applicability is limited due to the lack of medical\ninterpretations. Moreover, most previous methods assume a vector-based\nrepresentation for patients, which typically requires aggregation of medical\nevents over a certain time period. As a consequence, temporal information will\nbe lost. In this paper, we propose a patient similarity evaluation framework\nbased on the temporal matching of longitudinal patient EHRs. Two efficient\nmethods are presented, unsupervised and supervised, both of which preserve the\ntemporal properties in EHRs. The supervised scheme takes a convolutional neural\nnetwork architecture and learns an optimal representation of patient clinical\nrecords with medical concept embedding. The empirical results on real-world\nclinical data demonstrate substantial improvement over the baselines. We make\nour code and sample data available for further study.', 'The widespread availability of electronic health records (EHRs) promises to\nusher in the era of personalized medicine. However, the problem of extracting\nuseful clinical representations from longitudinal EHR data remains challenging.\nIn this paper, we explore deep neural network models with learned medical\nfeature embedding to deal with the problems of high dimensionality and\ntemporality. Specifically, we use a multi-layer convolutional neural network\n(CNN) to parameterize the model and is thus able to capture complex non-linear\nlongitudinal evolution of EHRs. Our model can effectively capture local/short\ntemporal dependency in EHRs, which is beneficial for risk prediction. To\naccount for high dimensionality, we use the embedding medical features in the\nCNN model which hold the natural medical concepts. Our initial experiments\nproduce promising results and demonstrate the effectiveness of both the medical\nfeature embedding and the proposed convolutional neural network in risk\nprediction on cohorts of congestive heart failure and diabetes patients\ncompared with several strong baselines.', 'Healthcare is becoming a more and more important research topic recently.\nWith the growing data in the healthcare domain, it offers a great opportunity\nfor deep learning to improve the quality of medical service. However, the\ncomplexity of electronic health records (EHR) data is a challenge for the\napplication of deep learning. Specifically, the data produced in the hospital\nadmissions are monitored by the EHR system, which includes structured data like\ndaily body temperature, and unstructured data like free text and laboratory\nmeasurements. Although there are some preprocessing frameworks proposed for\nspecific EHR data, the clinical notes that contain significant clinical value\nare beyond the realm of their consideration. Besides, whether these different\ndata from various views are all beneficial to the medical tasks and how to best\nutilize these data remain unclear. Therefore, in this paper, we first extract\nthe accompanying clinical notes from EHR and propose a method to integrate\nthese data, we also comprehensively study the different models and the data\nleverage methods for better medical task prediction. The results on two medical\nprediction tasks show that our fused model with different data outperforms the\nstate-of-the-art method that without clinical notes, which illustrates the\nimportance of our fusion method and the value of clinical note features. Our\ncode is available at https: //github.com/emnlp-mimic/mimic.', ""Electronic health records (EHR) systems contain vast amounts of medical\ninformation about patients. These data can be used to train machine learning\nmodels that can predict health status, as well as to help prevent future\ndiseases or disabilities. However, getting patients' medical data to obtain\nwell-trained machine learning models is a challenging task. This is because\nsharing the patients' medical records is prohibited by law in most countries\ndue to patients privacy concerns. In this paper, we tackle this problem by\nsharing the models instead of the original sensitive data by using the mimic\nlearning approach. The idea is first to train a model on the original sensitive\ndata, called the teacher model. Then, using this model, we can transfer its\nknowledge to another model, called the student model, without the need to learn\nthe original data used in training the teacher model. The student model is then\nshared to the public and can be used to make accurate predictions. To assess\nthe mimic learning approach, we have evaluated our scheme using different\nmedical datasets. The results indicate that the student model mimics the\nteacher model performance in terms of prediction accuracy without the need to\naccess to the patients' original data records."", ""Electronic Health Records (EHR) have been heavily used in modern healthcare\nsystems for recording patients' admission information to hospitals. Many\ndata-driven approaches employ temporal features in EHR for predicting specific\ndiseases, readmission times, or diagnoses of patients. However, most existing\npredictive models cannot fully utilize EHR data, due to an inherent lack of\nlabels in supervised training for some temporal events. Moreover, it is hard\nfor existing works to simultaneously provide generic and personalized\ninterpretability. To address these challenges, we first propose a hyperbolic\nembedding method with information flow to pre-train medical code\nrepresentations in a hierarchical structure. We incorporate these pre-trained\nrepresentations into a graph neural network to detect disease complications,\nand design a multi-level attention method to compute the contributions of\nparticular diseases and admissions, thus enhancing personalized\ninterpretability. We present a new hierarchy-enhanced historical prediction\nproxy task in our self-supervised learning framework to fully utilize EHR data\nand exploit medical domain knowledge. We conduct a comprehensive set of\nexperiments and case studies on widely used publicly available EHR datasets to\nverify the effectiveness of our model. The results demonstrate our model's\nstrengths in both predictive tasks and interpretable abilities."", 'Early detection of preventable diseases is important for better disease\nmanagement, improved inter-ventions, and more efficient health-care resource\nallocation. Various machine learning approacheshave been developed to utilize\ninformation in Electronic Health Record (EHR) for this task. Majorityof\nprevious attempts, however, focus on structured fields and lose the vast amount\nof information inthe unstructured notes. In this work we propose a general\nmulti-task framework for disease onsetprediction that combines both free-text\nmedical notes and structured information. We compareperformance of different\ndeep learning architectures including CNN, LSTM and hierarchical models.In\ncontrast to traditional text-based prediction models, our approach does not\nrequire disease specificfeature engineering, and can handle negations and\nnumerical values that exist in the text. Ourresults on a cohort of about 1\nmillion patients show that models using text outperform modelsusing just\nstructured data, and that models capable of using numerical values and\nnegations in thetext, in addition to the raw text, further improve performance.\nAdditionally, we compare differentvisualization methods for medical\nprofessionals to interpret model predictions.', 'The development of neural networks for clinical artificial intelligence (AI)\nis reliant on interpretability, transparency, and performance. The need to\ndelve into the black-box neural network and derive interpretable explanations\nof model output is paramount. A task of high clinical importance is predicting\nthe likelihood of a patient being readmitted to hospital in the near future to\nenable efficient triage. With the increasing adoption of electronic health\nrecords (EHRs), there is great interest in applications of natural language\nprocessing (NLP) to clinical free-text contained within EHRs. In this work, we\napply InfoCal, the current state-of-the-art model that produces extractive\nrationales for its predictions, to the task of predicting hospital readmission\nusing hospital discharge notes. We compare extractive rationales produced by\nInfoCal to competitive transformer-based models pretrained on clinical text\ndata and for which the attention mechanism can be used for interpretation. We\nfind each presented model with selected interpretability or feature importance\nmethods yield varying results, with clinical language domain expertise and\npretraining critical to performance and subsequent interpretability.', 'Electronic health records (EHRs) contain structured and unstructured data of\nsignificant clinical and research value. Various machine learning approaches\nhave been developed to employ information in EHRs for risk prediction. The\nmajority of these attempts, however, focus on structured EHR fields and lose\nthe vast amount of information in the unstructured texts. To exploit the\npotential information captured in EHRs, in this study we propose a multimodal\nrecurrent neural network model for cardiovascular risk prediction that\nintegrates both medical texts and structured clinical information. The proposed\nmultimodal bidirectional long short-term memory (BiLSTM) model concatenates\nword embeddings to classical clinical predictors before applying them to a\nfinal fully connected neural network. In the experiments, we compare\nperformance of different deep neural network (DNN) architectures including\nconvolutional neural network and long short-term memory in scenarios of using\nclinical variables and chest X-ray radiology reports. Evaluated on a data set\nof real world patients with manifest vascular disease or at high-risk for\ncardiovascular disease, the proposed BiLSTM model demonstrates state-of-the-art\nperformance and outperforms other DNN baseline architectures.', 'Widespread adoption of electronic health records (EHRs) has fueled the\ndevelopment of using machine learning to build prediction models for various\nclinical outcomes. This process is often constrained by having a relatively\nsmall number of patient records for training the model. We demonstrate that\nusing patient representation schemes inspired from techniques in natural\nlanguage processing can increase the accuracy of clinical prediction models by\ntransferring information learned from the entire patient population to the task\nof training a specific model, where only a subset of the population is\nrelevant. Such patient representation schemes enable a 3.5% mean improvement in\nAUROC on five prediction tasks compared to standard baselines, with the average\nimprovement rising to 19% when only a small number of patient records are\navailable for training the clinical prediction model.', 'The shift to electronic medical records (EMRs) has engendered research into\nmachine learning and natural language technologies to analyze patient records,\nand to predict from these clinical outcomes of interest. Two observations\nmotivate our aims here. First, unstructured notes contained within EMR often\ncontain key information, and hence should be exploited by models. Second, while\nstrong predictive performance is important, interpretability of models is\nperhaps equally so for applications in this domain. Together, these points\nsuggest that neural models for EMR may benefit from incorporation of attention\nover notes, which one may hope will both yield performance gains and afford\ntransparency in predictions. In this work we perform experiments to explore\nthis question using two EMR corpora and four different predictive tasks, that:\n(i) inclusion of attention mechanisms is critical for neural encoder modules\nthat operate over notes fields in order to yield competitive performance, but,\n(ii) unfortunately, while these boost predictive performance, it is decidedly\nless clear whether they provide meaningful support for predictions.', 'Deep learning models in healthcare may fail to generalize on data from unseen\ncorpora. Additionally, no quantitative metric exists to tell how existing\nmodels will perform on new data. Previous studies demonstrated that NLP models\nof medical notes generalize variably between institutions, but ignored other\nlevels of healthcare organization. We measured SciBERT diagnosis sentiment\nclassifier generalizability between medical specialties using EHR sentences\nfrom MIMIC-III. Models trained on one specialty performed better on internal\ntest sets than mixed or external test sets (mean AUCs 0.92, 0.87, and 0.83,\nrespectively; p = 0.016). When models are trained on more specialties, they\nhave better test performances (p < 1e-4). Model performance on new corpora is\ndirectly correlated to the similarity between train and test sentence content\n(p < 1e-4). Future studies should assess additional axes of generalization to\nensure deep learning models fulfil their intended purpose across institutions,\nspecialties, and practices.', 'The rapid growth of Electronic Health Records (EHRs), as well as the\naccompanied opportunities in Data-Driven Healthcare (DDH), has been attracting\nwidespread interests and attentions. Recent progress in the design and\napplications of deep learning methods has shown promising results and is\nforcing massive changes in healthcare academia and industry, but most of these\nmethods rely on massive labeled data. In this work, we propose a general deep\nlearning framework which is able to boost risk prediction performance with\nlimited EHR data. Our model takes a modified generative adversarial network\nnamely ehrGAN, which can provide plausible labeled EHR data by mimicking real\npatient records, to augment the training dataset in a semi-supervised learning\nmanner. We use this generative model together with a convolutional neural\nnetwork (CNN) based prediction model to improve the onset prediction\nperformance. Experiments on two real healthcare datasets demonstrate that our\nproposed framework produces realistic data samples and achieves significant\nimprovements on classification tasks with the generated data over several\nstat-of-the-art baselines.', ""A crucial step within secondary analysis of electronic health records (EHRs)\nis to identify the patient cohort under investigation. While EHRs contain\nmedical billing codes that aim to represent the conditions and treatments\npatients may have, much of the information is only present in the patient\nnotes. Therefore, it is critical to develop robust algorithms to infer\npatients' conditions and treatments from their written notes. In this paper, we\nintroduce a dataset for patient phenotyping, a task that is defined as the\nidentification of whether a patient has a given medical condition (also\nreferred to as clinical indication or phenotype) based on their patient note.\nNursing Progress Notes and Discharge Summaries from the Intensive Care Unit of\na large tertiary care hospital were manually annotated for the presence of\nseveral high-context phenotypes relevant to treatment and risk of\nre-hospitalization. This dataset contains 1102 Discharge Summaries and 1000\nNursing Progress Notes. Each Discharge Summary and Progress Note has been\nannotated by at least two expert human annotators (one clinical researcher and\none resident physician). Annotated phenotypes include treatment non-adherence,\nchronic pain, advanced/metastatic cancer, as well as 10 other phenotypes. This\ndataset can be utilized for academic and industrial research in medicine and\ncomputer science, particularly within the field of medical natural language\nprocessing."", 'Biomedical documents such as Electronic Health Records (EHRs) contain a large\namount of information in an unstructured format. The data in EHRs is a hugely\nvaluable resource documenting clinical narratives and decisions, but whilst the\ntext can be easily understood by human doctors it is challenging to use in\nresearch and clinical applications. To uncover the potential of biomedical\ndocuments we need to extract and structure the information they contain. The\ntask at hand is Named Entity Recognition and Linking (NER+L). The number of\nentities, ambiguity of words, overlapping and nesting make the biomedical area\nsignificantly more difficult than many others. To overcome these difficulties,\nwe have developed the Medical Concept Annotation Tool (MedCAT), an open-source\nunsupervised approach to NER+L. MedCAT uses unsupervised machine learning to\ndisambiguate entities. It was validated on MIMIC-III (a freely accessible\ncritical care database) and MedMentions (Biomedical papers annotated with\nmentions from the Unified Medical Language System). In case of NER+L, the\ncomparison with existing tools shows that MedCAT improves the previous best\nwith only unsupervised learning (F1=0.848 vs 0.691 for disease detection;\nF1=0.710 vs. 0.222 for general concept detection). A qualitative analysis of\nthe vector embeddings learnt by MedCAT shows that it captures latent medical\nknowledge available in EHRs (MIMIC-III). Unsupervised learning can improve the\nperformance of large scale entity extraction, but it has some limitations when\nworking with only a couple of entities and a small dataset. In that case\noptions are supervised learning or active learning, both of which are supported\nin MedCAT via the MedCATtrainer extension. Our approach can detect and link\nmillions of different biomedical concepts with state-of-the-art performance,\nwhilst being lightweight, fast and easy to use.', 'Extracting information from electronic health records (EHR) is a challenging\ntask since it requires prior knowledge of the reports and some natural language\nprocessing algorithm (NLP). With the growing number of EHR implementations,\nsuch knowledge is increasingly challenging to obtain in an efficient manner. We\naddress this challenge by proposing a novel methodology to analyze large sets\nof EHRs using a modified Count Sketch data streaming algorithm termed DreamNLP.\nBy using DreamNLP, we generate a dictionary of frequently occurring terms or\nheavy hitters in the EHRs using low computational memory compared to\nconventional counting approach other NLP programs use. We demonstrate the\nextraction of the most important breast diagnosis features from the EHRs in a\nset of patients that underwent breast imaging. Based on the analysis,\nextraction of these terms would be useful for defining important features for\ndownstream tasks such as machine learning for precision medicine.', ""Understanding patients' journeys in healthcare system is a fundamental\nprepositive task for a broad range of AI-based healthcare applications. This\ntask aims to learn an informative representation that can comprehensively\nencode hidden dependencies among medical events and its inner entities, and\nthen the use of encoding outputs can greatly benefit the downstream\napplication-driven tasks. A patient journey is a sequence of electronic health\nrecords (EHRs) over time that is organized at multiple levels: patient, visits\nand medical codes. The key challenge of patient journey understanding is to\ndesign an effective encoding mechanism which can properly tackle the\naforementioned multi-level structured patient journey data with temporal\nsequential visits and a set of medical codes. This paper proposes a novel\nself-attention mechanism that can simultaneously capture the contextual and\ntemporal relationships hidden in patient journeys. A multi-level self-attention\nnetwork (MusaNet) is specifically designed to learn the representations of\npatient journeys that is used to be a long sequence of activities. The MusaNet\nis trained in end-to-end manner using the training data derived from EHRs. We\nevaluated the efficacy of our method on two medical application tasks with\nreal-world benchmark datasets. The results have demonstrated the proposed\nMusaNet produces higher-quality representations than state-of-the-art baseline\nmethods. The source code is available in https://github.com/xueping/MusaNet."", 'In longitudinal electronic health records (EHRs), the event records of a\npatient are distributed over a long period of time and the temporal relations\nbetween the events reflect sufficient domain knowledge to benefit prediction\ntasks such as the rate of inpatient mortality. Medical concept embedding as a\nfeature extraction method that transforms a set of medical concepts with a\nspecific time stamp into a vector, which will be fed into a supervised learning\nalgorithm. The quality of the embedding significantly determines the learning\nperformance over the medical data. In this paper, we propose a medical concept\nembedding method based on applying a self-attention mechanism to represent each\nmedical concept. We propose a novel attention mechanism which captures the\ncontextual information and temporal relationships between medical concepts. A\nlight-weight neural net, ""Temporal Self-Attention Network (TeSAN)"", is then\nproposed to learn medical concept embedding based solely on the proposed\nattention mechanism. To test the effectiveness of our proposed methods, we have\nconducted clustering and prediction tasks on two public EHRs datasets comparing\nTeSAN against five state-of-the-art embedding methods. The experimental results\ndemonstrate that the proposed TeSAN model is superior to all the compared\nmethods. To the best of our knowledge, this work is the first to exploit\ntemporal self-attentive relations between medical events.', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', 'Recent grid-based document representations like BERTgrid allow the\nsimultaneous encoding of the textual and layout information of a document in a\n2D feature map so that state-of-the-art image segmentation and/or object\ndetection models can be straightforwardly leveraged to extract key information\nfrom documents. However, such methods have not achieved comparable performance\nto state-of-the-art sequence- and graph-based methods such as LayoutLM and PICK\nyet. In this paper, we propose a new multi-modal backbone network by\nconcatenating a BERTgrid to an intermediate layer of a CNN model, where the\ninput of CNN is a document image and the BERTgrid is a grid of word embeddings,\nto generate a more powerful grid-based document representation, named\nViBERTgrid. Unlike BERTgrid, the parameters of BERT and CNN in our multimodal\nbackbone network are trained jointly. Our experimental results demonstrate that\nthis joint training strategy improves significantly the representation ability\nof ViBERTgrid. Consequently, our ViBERTgrid-based key information extraction\napproach has achieved state-of-the-art performance on real-world datasets.', 'In these notes, we wish to provide an introduction to the signature method,\nfocusing on its basic theoretical properties and recent numerical applications.\n  The notes are split into two parts. The first part focuses on the definition\nand fundamental properties of the signature of a path, or the path signature.\nWe have aimed for a minimalistic approach, assuming only familiarity with\nclassical real analysis and integration theory, and supplementing theory with\nstraightforward examples. We have chosen to focus in detail on the principle\nproperties of the signature which we believe are fundamental to understanding\nits role in applications. We also present an informal discussion on some of its\ndeeper properties and briefly mention the role of the signature in rough paths\ntheory, which we hope could serve as a light introduction to rough paths for\nthe interested reader.\n  The second part of these notes discusses practical applications of the path\nsignature to the area of machine learning. The signature approach represents a\nnon-parametric way for extraction of characteristic features from data. The\ndata are converted into a multi-dimensional path by means of various embedding\nalgorithms and then processed for computation of individual terms of the\nsignature which summarise certain information contained in the data. The\nsignature thus transforms raw data into a set of features which are used in\nmachine learning tasks. We will review current progress in applications of\nsignatures to machine learning problems.', ""In this work we addressed the problem of capturing sequential information\ncontained in longitudinal electronic health records (EHRs). Clinical notes,\nwhich is a particular type of EHR data, are a rich source of information and\npractitioners often develop clever solutions how to maximise the sequential\ninformation contained in free-texts. We proposed a systematic methodology for\nlearning from chronological events available in clinical notes. The proposed\nmethodological {\\it path signature} framework creates a non-parametric\nhierarchical representation of sequential events of any type and can be used as\nfeatures for downstream statistical learning tasks. The methodology was\ndeveloped and externally validated using the largest in the UK secondary care\nmental health EHR data on a specific task of predicting survival risk of\npatients diagnosed with Alzheimer's disease. The signature-based model was\ncompared to a common survival random forest model. Our results showed a\n15.4$\\%$ increase of risk prediction AUC at the time point of 20 months after\nthe first admission to a specialist memory clinic and the signature method\noutperformed the baseline mixed-effects model by 13.2 $\\%$."", ""Clinicians spend a significant amount of time inputting free-form textual\nnotes into Electronic Health Records (EHR) systems. Much of this documentation\nwork is seen as a burden, reducing time spent with patients and contributing to\nclinician burnout. With the aspiration of AI-assisted note-writing, we propose\na new language modeling task predicting the content of notes conditioned on\npast data from a patient's medical record, including patient demographics,\nlabs, medications, and past notes. We train generative models using the public,\nde-identified MIMIC-III dataset and compare generated notes with those in the\ndataset on multiple measures. We find that much of the content can be\npredicted, and that many common templates found in notes can be learned. We\ndiscuss how such models can be useful in supporting assistive note-writing\nfeatures such as error-detection and auto-complete."", 'Information in electronic health records (EHR), such as clinical narratives,\nexamination reports, lab measurements, demographics, and other patient\nencounter entries, can be transformed into appropriate data representations\nthat can be used for downstream clinical machine learning tasks using\nrepresentation learning. Learning better representations is critical to improve\nthe performance of downstream tasks. Due to the advances in machine learning,\nwe now can learn better and meaningful representations from EHR through\ndisentangling the underlying factors inside data and distilling large amounts\nof information and knowledge from heterogeneous EHR sources. In this chapter,\nwe first introduce the background of learning representations and reasons why\nwe need good EHR representations in machine learning for medicine and\nhealthcare in Section 1. Next, we explain the commonly-used machine learning\nand evaluation methods for representation learning using a deep learning\napproach in Section 2. Following that, we review recent related studies of\nlearning patient state representation from EHR for clinical machine learning\ntasks in Section 3. Finally, in Section 4 we discuss more techniques, studies,\nand challenges for learning natural language representations when free texts,\nsuch as clinical notes, examination reports, or biomedical literature are used.\nWe also discuss challenges and opportunities in these rapidly growing research\nfields.', 'Crucial information about the practice of healthcare is recorded only in\nfree-form text, which creates an enormous opportunity for high-impact NLP.\nHowever, annotated healthcare datasets tend to be small and expensive to\nobtain, which raises the question of how to make maximally efficient uses of\nthe available data. To this end, we develop an LSTM-CRF model for combining\nunsupervised word representations and hand-built feature representations\nderived from publicly available healthcare ontologies. We show that this\ncombined model yields superior performance on five datasets of diverse kinds of\nhealthcare text (clinical, social, scientific, commercial). Each involves the\nlabeling of complex, multi-word spans that pick out different healthcare\nconcepts. We also introduce a new labeled dataset for identifying the treatment\nrelations between drugs and diseases.', 'The past decade has seen an explosion in the amount of digital information\nstored in electronic health records (EHR). While primarily designed for\narchiving patient clinical information and administrative healthcare tasks,\nmany researchers have found secondary use of these records for various clinical\ninformatics tasks. Over the same period, the machine learning community has\nseen widespread advances in deep learning techniques, which also have been\nsuccessfully applied to the vast amount of EHR data. In this paper, we review\nthese deep EHR systems, examining architectures, technical aspects, and\nclinical applications. We also identify shortcomings of current techniques and\ndiscuss avenues of future research for EHR-based deep learning.', ""Oral drugs are becoming increasingly common in oncology care. In contrast to\nintravenous chemotherapy, which is administered in the clinic and carefully\ntracked via structure electronic health records (EHRs), oral drug treatment is\nself-administered and therefore not tracked as well. Often, the details of oral\ncancer treatment occur only in unstructured clinic notes. Extracting this\ninformation is critical to understanding a patient's treatment history. Yet,\nthis a challenging task because treatment intervals must be inferred\nlongitudinally from both explicit mentions in the text as well as from document\ntimestamps. In this work, we present TIFTI (Temporally Integrated Framework for\nTreatment Intervals), a robust framework for extracting oral drug treatment\nintervals from a patient's unstructured notes. TIFTI leverages distinct sources\nof temporal information by breaking the problem down into two separate\nsubtasks: document-level sequence labeling and date extraction. On a labeled\ndataset of metastatic renal-cell carcinoma (RCC) patients, it exactly matched\nthe labeled start date in 46% of the examples (86% of the examples within 30\ndays), and it exactly matched the labeled end date in 52% of the examples (78%\nof the examples within 30 days). Without retraining, the model achieved a\nsimilar level of performance on a labeled dataset of advanced non-small-cell\nlung cancer (NSCLC) patients."", 'In recent years, many methods have been developed to identify important\nportions of text documents. Summarization tools can utilize these methods to\nextract summaries from large volumes of textual information. However, to\nidentify concepts representing central ideas within a text document and to\nextract the most informative sentences that best convey those concepts still\nremain two crucial tasks in summarization methods. In this paper, we introduce\na graph-based method to address these two challenges in the context of\nbiomedical text summarization. We show that how a summarizer can discover\nmeaningful concepts within a biomedical text document using the Helmholtz\nprinciple. The summarizer considers the meaningful concepts as the main topics\nand constructs a graph based on the topics that the sentences share. The\nsummarizer can produce an informative summary by extracting those sentences\nhaving higher values of the degree. We assess the performance of our method for\nsummarization of biomedical articles using the Recall-Oriented Understudy for\nGisting Evaluation (ROUGE) toolkit. The results show that the degree can be a\nuseful centrality measure to identify important sentences in this type of\ngraph-based modelling. Our method can improve the performance of biomedical\ntext summarization compared to some state-of-the-art and publicly available\nsummarizers. Combining a concept-based modelling strategy and a graph-based\napproach to sentence extraction, our summarizer can produce summaries with the\nhighest scores of informativeness among the comparison methods. This research\nwork can be regarded as a start point to the study of small-world networks in\nsummarization of biomedical texts.', ""The data available in Electronic Health Records (EHRs) provides the\nopportunity to transform care, and the best way to provide better care for one\npatient is through learning from the data available on all other patients.\nTemporal modelling of a patient's medical history, which takes into account the\nsequence of past events, can be used to predict future events such as a\ndiagnosis of a new disorder or complication of a previous or existing disorder.\nWhile most prediction approaches use mostly the structured data in EHRs or a\nsubset of single-domain predictions and outcomes, we present MedGPT a novel\ntransformer-based pipeline that uses Named Entity Recognition and Linking tools\n(i.e. MedCAT) to structure and organize the free text portion of EHRs and\nanticipate a range of future medical events (initially disorders). Since a\nlarge portion of EHR data is in text form, such an approach benefits from a\ngranular and detailed view of a patient while introducing modest additional\nnoise. MedGPT effectively deals with the noise and the added granularity, and\nachieves a precision of 0.344, 0.552 and 0.640 (vs LSTM 0.329, 0.538 and 0.633)\nwhen predicting the top 1, 3 and 5 candidate future disorders on real world\nhospital data from King's College Hospital, London, UK (\\textasciitilde600k\npatients). We also show that our model captures medical knowledge by testing it\non an experimental medical multiple choice question answering task, and by\nexamining the attentional focus of the model using gradient-based saliency\nmethods."", 'Clinical Named Entity Recognition (CNER) aims to automatically identity\nclinical terminologies in Electronic Health Records (EHRs), which is a\nfundamental and crucial step for clinical research. To train a high-performance\nmodel for CNER, it usually requires a large number of EHRs with high-quality\nlabels. However, labeling EHRs, especially Chinese EHRs, is time-consuming and\nexpensive. One effective solution to this is active learning, where a model\nasks labelers to annotate data which the model is uncertain of. Conventional\nactive learning assumes a single labeler that always replies noiseless answers\nto queried labels. However, in real settings, multiple labelers provide diverse\nquality of annotation with varied costs and labelers with low overall\nannotation quality can still assign correct labels for some specific instances.\nIn this paper, we propose a Cost-Quality Adaptive Active Learning (CQAAL)\napproach for CNER in Chinese EHRs, which maintains a balance between the\nannotation quality, labeling costs, and the informativeness of selected\ninstances. Specifically, CQAAL selects cost-effective instance-labeler pairs to\nachieve better annotation quality with lower costs in an adaptive manner.\nComputational results on the CCKS-2017 Task 2 benchmark dataset demonstrate the\nsuperiority and effectiveness of the proposed CQAAL.', 'Machine learning models depend on the quality of input data. As electronic\nhealth records are widely adopted, the amount of data in health care is\ngrowing, along with complaints about the quality of medical notes. We use two\nprediction tasks, readmission prediction and in-hospital mortality prediction,\nto characterize the value of information in medical notes. We show that as a\nwhole, medical notes only provide additional predictive power over structured\ninformation in readmission prediction. We further propose a probing framework\nto select parts of notes that enable more accurate predictions than using all\nnotes, despite that the selected information leads to a distribution shift from\nthe training data (""all notes""). Finally, we demonstrate that models trained on\nthe selected valuable information achieve even better predictive performance,\nwith only 6.8% of all the tokens for readmission prediction.', 'Predicting diagnoses from Electronic Health Records (EHRs) is an important\nmedical application of multi-label learning. We propose a convolutional\nresidual model for multi-label classification from doctor notes in EHR data. A\ngiven patient may have multiple diagnoses, and therefore multi-label learning\nis required. We employ a Convolutional Neural Network (CNN) to encode plain\ntext into a fixed-length sentence embedding vector. Since diagnoses are\ntypically correlated, a deep residual network is employed on top of the CNN\nencoder, to capture label (diagnosis) dependencies and incorporate information\ndirectly from the encoded sentence vector. A real EHR dataset is considered,\nand we compare the proposed model with several well-known baselines, to predict\ndiagnoses based on doctor notes. Experimental results demonstrate the\nsuperiority of the proposed convolutional residual model.', ""In recent years, the trend of deploying digital systems in numerous\nindustries has hiked. The health sector has observed an extensive adoption of\ndigital systems and services that generate significant medical records.\nElectronic health records contain valuable information for prospective and\nretrospective analysis that is often not entirely exploited because of the\ncomplicated dense information storage. The crude purpose of condensing health\nrecords is to select the information that holds most characteristics of the\noriginal documents based on a reported disease. These summaries may boost\ndiagnosis and save a doctor's time during a saturated workload situation like\nthe COVID-19 pandemic. In this paper, we are applying a multi-head\nattention-based mechanism to perform extractive summarization of meaningful\nphrases on clinical notes. Our method finds major sentences for a summary by\ncorrelating tokens, segments, and positional embeddings of sentences in a\nclinical note. The model outputs attention scores that are statistically\ntransformed to extract critical phrases for visualization on the heat-mapping\ntool and for human use."", 'Document-level Relation Extraction (RE) requires extracting relations\nexpressed within and across sentences. Recent works show that graph-based\nmethods, usually constructing a document-level graph that captures\ndocument-aware interactions, can obtain useful entity representations thus\nhelping tackle document-level RE. These methods either focus more on the entire\ngraph, or pay more attention to a part of the graph, e.g., paths between the\ntarget entity pair. However, we find that document-level RE may benefit from\nfocusing on both of them simultaneously. Therefore, to obtain more\ncomprehensive entity representations, we propose the Coarse-to-Fine Entity\nRepresentation model (CFER) that adopts a coarse-to-fine strategy involving two\nphases. First, CFER uses graph neural networks to integrate global information\nin the entire graph at a coarse level. Next, CFER utilizes the global\ninformation as a guidance to selectively aggregate path information between the\ntarget entity pair at a fine level. In classification, we combine the entity\nrepresentations from both two levels into more comprehensive representations\nfor relation extraction. Experimental results on two document-level RE\ndatasets, DocRED and CDR, show that CFER outperforms existing models and is\nrobust to the uneven label distribution.', 'The modelling of Electronic Health Records (EHRs) has the potential to drive\nmore efficient allocation of healthcare resources, enabling early intervention\nstrategies and advancing personalised healthcare. However, EHRs are challenging\nto model due to their realisation as noisy, multi-modal data occurring at\nirregular time intervals. To address their temporal nature, we treat EHRs as\nsamples generated by a Temporal Point Process (TPP), enabling us to model what\nhappened in an event with when it happened in a principled way. We gather and\npropose neural network parameterisations of TPPs, collectively referred to as\nNeural TPPs. We perform evaluations on synthetic EHRs as well as on a set of\nestablished benchmarks. We show that TPPs significantly outperform their\nnon-TPP counterparts on EHRs. We also show that an assumption of many Neural\nTPPs, that the class distribution is conditionally independent of time, reduces\nperformance on EHRs. Finally, our proposed attention-based Neural TPP performs\nfavourably compared to existing models, whilst aligning with real world\ninterpretability requirements, an important step towards a component of\nclinical decision support systems.', 'Electronic health records (EHRs) have contributed to the computerization of\npatient records and can thus be used not only for efficient and systematic\nmedical services, but also for research on biomedical data science. However,\nthere are many missing values in EHRs when provided in matrix form, which is an\nimportant issue in many biomedical EHR applications. In this paper, we propose\na two-stage framework that includes missing data imputation and disease\nprediction to address the missing data problem in EHRs. We compared the disease\nprediction performance of generative adversarial networks (GANs) and\nconventional learning algorithms in combination with missing data prediction\nmethods. As a result, we obtained a level of accuracy of 0.9777, sensitivity of\n0.9521, specificity of 0.9925, area under the receiver operating characteristic\ncurve (AUC-ROC) of 0.9889, and F-score of 0.9688 with a stacked autoencoder as\nthe missing data prediction method and an auxiliary classifier GAN (AC-GAN) as\nthe disease prediction method. The comparison results show that a combination\nof a stacked autoencoder and an AC-GAN significantly outperforms other existing\napproaches. Our results suggest that the proposed framework is more robust for\ndisease prediction from EHRs with missing data.', 'Databases of electronic health records (EHRs) are increasingly used to inform\nclinical decisions. Machine learning methods can find patterns in EHRs that are\npredictive of future adverse outcomes. However, statistical models may be built\nupon patterns of health-seeking behavior that vary across patient\nsubpopulations, leading to poor predictive performance when training on one\npatient population and predicting on another. This note proposes two tests to\nbetter measure and understand model generalization. We use these tests to\ncompare models derived from two data sources: (i) historical medical records,\nand (ii) electrocardiogram (EKG) waveforms. In a predictive task, we show that\nEKG-based models can be more stable than EHR-based models across different\npatient populations.', 'Document AI, or Document Intelligence, is a relatively new research topic\nthat refers to the techniques for automatically reading, understanding, and\nanalyzing business documents. It is an important research direction for natural\nlanguage processing and computer vision. In recent years, the popularity of\ndeep learning technology has greatly advanced the development of Document AI,\nsuch as document layout analysis, visual information extraction, document\nvisual question answering, document image classification, etc. This paper\nbriefly reviews some of the representative models, tasks, and benchmark\ndatasets. Furthermore, we also introduce early-stage heuristic rule-based\ndocument analysis, statistical machine learning algorithms, and deep learning\napproaches especially pre-training methods. Finally, we look into future\ndirections for Document AI research.', 'We present a contrasting learning approach with data augmentation techniques\nto learn document representations in an unsupervised manner. Inspired by recent\ncontrastive self-supervised learning algorithms used for image and NLP\npretraining, we hypothesize that high-quality document embedding should be\ninvariant to diverse paraphrases that preserve the semantics of the original\ndocument. With different backbones and contrastive learning frameworks, our\nstudy reveals the enormous benefits of contrastive augmentation for document\nrepresentation learning with two additional insights: 1) including data\naugmentation in a contrastive way can substantially improve the embedding\nquality in unsupervised document representation learning, and 2) in general,\nstochastic augmentations generated by simple word-level manipulation work much\nbetter than sentence-level and document-level ones. We plug our method into a\nclassifier and compare it with a broad range of baseline methods on six\nbenchmark datasets. Our method can decrease the classification error rate by up\nto 6.4% over the SOTA approaches on the document classification task, matching\nor even surpassing fully-supervised methods.', ""Patient notes contain a wealth of information of potentially great interest\nto medical investigators. However, to protect patients' privacy, Protected\nHealth Information (PHI) must be removed from the patient notes before they can\nbe legally released, a process known as patient note de-identification. The\nmain objective for a de-identification system is to have the highest possible\nrecall. Recently, the first neural-network-based de-identification system has\nbeen proposed, yielding state-of-the-art results. Unlike other systems, it does\nnot rely on human-engineered features, which allows it to be quickly deployed,\nbut does not leverage knowledge from human experts or from electronic health\nrecords (EHRs). In this work, we explore a method to incorporate\nhuman-engineered features as well as features derived from EHRs to a\nneural-network-based de-identification system. Our results show that the\naddition of features, especially the EHR-derived features, further improves the\nstate-of-the-art in patient note de-identification, including for some of the\nmost sensitive PHI types such as patient names. Since in a real-life setting\npatient notes typically come with EHRs, we recommend developers of\nde-identification systems to leverage the information EHRs contain."", 'The adoption of electronic health records (EHR) has become universal during\nthe past decade, which has afforded in-depth data-based research. By learning\nfrom the large amount of healthcare data, various data-driven models have been\nbuilt to predict future events for different medical tasks, such as auto\ndiagnosis and heart-attack prediction. Although EHR is abundant, the population\nthat satisfies specific criteria for learning population-specific tasks is\nscarce, making it challenging to train data-hungry deep learning models. This\nstudy presents the Claim Pre-Training (Claim-PT) framework, a generic\npre-training model that first trains on the entire pediatric claims dataset,\nfollowed by a discriminative fine-tuning on each population-specific task. The\nsemantic meaning of medical events can be captured in the pre-training stage,\nand the effective knowledge transfer is completed through the task-aware\nfine-tuning stage. The fine-tuning process requires minimal parameter\nmodification without changing the model architecture, which mitigates the data\nscarcity issue and helps train the deep learning model adequately on small\npatient cohorts. We conducted experiments on a real-world claims dataset with\nmore than one million patient records. Experimental results on two downstream\ntasks demonstrated the effectiveness of our method: our general task-agnostic\npre-training framework outperformed tailored task-specific models, achieving\nmore than 10\\% higher in model performance as compared to baselines. In\naddition, our framework showed a great generalizability potential to transfer\nlearned knowledge from one institution to another, paving the way for future\nhealthcare model pre-training across institutions.', 'In this paper we study the problem of predicting clinical diagnoses from\ntextual Electronic Health Records (EHR) data. We show the importance of this\nproblem in medical community and present comprehensive historical review of the\nproblem and proposed methods. As the main scientific contributions we present a\nmodification of Bidirectional Encoder Representations from Transformers (BERT)\nmodel for sequence classification that implements a novel way of\nFully-Connected (FC) layer composition and a BERT model pretrained only on\ndomain data. To empirically validate our model, we use a large-scale Russian\nEHR dataset consisting of about 4 million unique patient visits. This is the\nlargest such study for the Russian language and one of the largest globally. We\nperformed a number of comparative experiments with other text representation\nmodels on the task of multiclass classification for 265 disease subset of\nICD-10. The experiments demonstrate improved performance of our models compared\nto other baselines, including a fine-tuned Russian BERT (RuBERT) variant. We\nalso show comparable performance of our model with a panel of experienced\nmedical experts. This allows us to hope that implementation of this system will\nreduce misdiagnosis.', 'Large, pre-trained transformer models like BERT have achieved\nstate-of-the-art results on document understanding tasks, but most\nimplementations can only consider 512 tokens at a time. For many real-world\napplications, documents can be much longer, and the segmentation strategies\ntypically used on longer documents miss out on document structure and\ncontextual information, hurting their results on downstream tasks. In our work\non legal agreements, we find that visual cues such as layout, style, and\nplacement of text in a document are strong features that are crucial to\nachieving an acceptable level of accuracy on long documents. We measure the\nimpact of incorporating such visual cues, obtained via computer vision methods,\non the accuracy of document understanding tasks including document\nsegmentation, entity extraction, and attribute classification. Our method of\nsegmenting documents based on structural metadata out-performs existing methods\non four long-document understanding tasks as measured on the Contract\nUnderstanding Atticus Dataset.', 'Events and entities are closely related; entities are often actors or\nparticipants in events and events without entities are uncommon. The\ninterpretation of events and entities is highly contextually dependent.\nExisting work in information extraction typically models events separately from\nentities, and performs inference at the sentence level, ignoring the rest of\nthe document. In this paper, we propose a novel approach that models the\ndependencies among variables of events, entities, and their relations, and\nperforms joint inference of these variables across a document. The goal is to\nenable access to document-level contextual information and facilitate\ncontext-aware predictions. We demonstrate that our approach substantially\noutperforms the state-of-the-art methods for event extraction as well as a\nstrong baseline for entity extraction.', 'Document-level relation extraction (DocRE) aims to extract semantic relations\namong entity pairs in a document. Typical DocRE methods blindly take the full\ndocument as input, while a subset of the sentences in the document, noted as\nthe evidence, are often sufficient for humans to predict the relation of an\nentity pair. In this paper, we propose an evidence-enhanced framework, Eider,\nthat empowers DocRE by efficiently extracting evidence and effectively fusing\nthe extracted evidence in inference. We first jointly train an RE model with a\nlightweight evidence extraction model, which is efficient in both memory and\nruntime. Empirically, even training the evidence model on silver labels\nconstructed by our heuristic rules can lead to better RE performance. We\nfurther design a simple yet effective inference process that makes RE\npredictions on both extracted evidence and the full document, then fuses the\npredictions through a blending layer. This allows Eider to focus on important\nsentences while still having access to the complete information in the\ndocument. Extensive experiments show that Eider outperforms state-of-the-art\nmethods on three benchmark datasets (e.g., by 1.37/1.26 Ign F1/F1 on DocRED).', 'Hierarchical neural architectures are often used to capture long-distance\ndependencies and have been applied to many document-level tasks such as\nsummarization, document segmentation, and sentiment analysis. However,\neffective usage of such a large context can be difficult to learn, especially\nin the case where there is limited labeled data available. Building on the\nrecent success of language model pretraining methods for learning flat\nrepresentations of text, we propose algorithms for pre-training hierarchical\ndocument representations from unlabeled data. Unlike prior work, which has\nfocused on pre-training contextual token representations or context-independent\n{sentence/paragraph} representations, our hierarchical document representations\ninclude fixed-length sentence/paragraph representations which integrate\ncontextual information from the entire documents. Experiments on document\nsegmentation, document-level question answering, and extractive document\nsummarization demonstrate the effectiveness of the proposed pre-training\nalgorithms.', 'Standard neural machine translation (NMT) is on the assumption of\ndocument-level context independent. Most existing document-level NMT methods\nare satisfied with a smattering sense of brief document-level information,\nwhile this work focuses on exploiting detailed document-level context in terms\nof multiple forms of document embeddings, which is capable of sufficiently\nmodeling deeper and richer document-level context. The proposed document-aware\nNMT is implemented to enhance the Transformer baseline by introducing both\nglobal and local document-level clues on the source end. Experiments show that\nthe proposed method significantly improves the translation performance over\nstrong baselines and other related studies.', 'Recent years have seen numerous NLP datasets introduced to evaluate the\nperformance of fine-tuned models on natural language understanding tasks.\nRecent results from large pretrained models, though, show that many of these\ndatasets are largely saturated and unlikely to be able to detect further\nprogress. What kind of datasets are still effective at discriminating among\nstrong models, and what kind of datasets should we expect to be able to detect\nfuture improvements? To measure this uniformly across datasets, we draw on Item\nResponse Theory and evaluate 29 datasets using predictions from 18 pretrained\nTransformer models on individual test examples. We find that Quoref, HellaSwag,\nand MC-TACO are best suited for distinguishing among state-of-the-art models,\nwhile SNLI, MNLI, and CommitmentBank seem to be saturated for current strong\nmodels. We also observe span selection task format, which is used for QA\ndatasets like QAMR or SQuAD2.0, is effective in differentiating between strong\nand weak models.', ""Electronic Health Records (EHRs) have been heavily used to predict various\ndownstream clinical tasks such as readmission or mortality. One of the\nmodalities in EHRs, clinical notes, has not been fully explored for these tasks\ndue to its unstructured and inexplicable nature. Although recent advances in\ndeep learning (DL) enables models to extract interpretable features from\nunstructured data, they often require a large amount of training data. However,\nmany tasks in medical domains inherently consist of small sample data with\nlengthy documents; for a kidney transplant as an example, data from only a few\nthousand of patients are available and each patient's document consists of a\ncouple of millions of words in major hospitals. Thus, complex DL methods cannot\nbe applied to these kinds of domains. In this paper, we present a comprehensive\nensemble model using vector space modeling and topic modeling. Our proposed\nmodel is evaluated on the readmission task of kidney transplant patients and\nimproves 0.0211 in terms of c-statistics from the previous state-of-the-art\napproach using structured data, while typical DL methods fail to beat this\napproach. The proposed architecture provides the interpretable score for each\nfeature from both modalities, structured and unstructured data, which is shown\nto be meaningful through a physician's evaluation."", 'Multiple entities in a document generally exhibit complex inter-sentence\nrelations, and cannot be well handled by existing relation extraction (RE)\nmethods that typically focus on extracting intra-sentence relations for single\nentity pairs. In order to accelerate the research on document-level RE, we\nintroduce DocRED, a new dataset constructed from Wikipedia and Wikidata with\nthree features: (1) DocRED annotates both named entities and relations, and is\nthe largest human-annotated dataset for document-level RE from plain text; (2)\nDocRED requires reading multiple sentences in a document to extract entities\nand infer their relations by synthesizing all information of the document; (3)\nalong with the human-annotated data, we also offer large-scale distantly\nsupervised data, which enables DocRED to be adopted for both supervised and\nweakly supervised scenarios. In order to verify the challenges of\ndocument-level RE, we implement recent state-of-the-art methods for RE and\nconduct a thorough evaluation of these methods on DocRED. Empirical results\nshow that DocRED is challenging for existing RE methods, which indicates that\ndocument-level RE remains an open problem and requires further efforts. Based\non the detailed analysis on the experiments, we discuss multiple promising\ndirections for future research.', ""Electronic health records (EHRs) are longitudinal records of a patient's\ninteractions with healthcare systems. A patient's EHR data is organized as a\nthree-level hierarchy from top to bottom: patient journey - all the experiences\nof diagnoses and treatments over a period of time; individual visit - a set of\nmedical codes in a particular visit; and medical code - a specific record in\nthe form of medical codes. As EHRs begin to amass in millions, the potential\nbenefits, which these data might hold for medical research and medical outcome\nprediction, are staggering - including, for example, predicting future\nadmissions to hospitals, diagnosing illnesses or determining the efficacy of\nmedical treatments. Each of these analytics tasks requires a domain knowledge\nextraction method to transform the hierarchical patient journey into a vector\nrepresentation for further prediction procedure. The representations should\nembed a sequence of visits and a set of medical codes with a specific\ntimestamp, which are crucial to any downstream prediction tasks. Hence,\nexpressively powerful representations are appealing to boost learning\nperformance. To this end, we propose a novel self-attention mechanism that\ncaptures the contextual dependency and temporal relationships within a\npatient's healthcare journey. An end-to-end bidirectional temporal encoder\nnetwork (BiteNet) then learns representations of the patient's journeys, based\nsolely on the proposed attention mechanism. We have evaluated the effectiveness\nof our methods on two supervised prediction and two unsupervised clustering\ntasks with a real-world EHR dataset. The empirical results demonstrate the\nproposed BiteNet model produces higher-quality representations than\nstate-of-the-art baseline methods."", ""Social and behavioral determinants of health (SBDoH) have important roles in\nshaping people's health. In clinical research studies, especially comparative\neffectiveness studies, failure to adjust for SBDoH factors will potentially\ncause confounding issues and misclassification errors in either statistical\nanalyses and machine learning-based models. However, there are limited studies\nto examine SBDoH factors in clinical outcomes due to the lack of structured\nSBDoH information in current electronic health record (EHR) systems, while much\nof the SBDoH information is documented in clinical narratives. Natural language\nprocessing (NLP) is thus the key technology to extract such information from\nunstructured clinical text. However, there is not a mature clinical NLP system\nfocusing on SBDoH. In this study, we examined two state-of-the-art\ntransformer-based NLP models, including BERT and RoBERTa, to extract SBDoH\nconcepts from clinical narratives, applied the best performing model to extract\nSBDoH concepts on a lung cancer screening patient cohort, and examined the\ndifference of SBDoH information between NLP extracted results and structured\nEHRs (SBDoH information captured in standard vocabularies such as the\nInternational Classification of Diseases codes). The experimental results show\nthat the BERT-based NLP model achieved the best strict/lenient F1-score of\n0.8791 and 0.8999, respectively. The comparison between NLP extracted SBDoH\ninformation and structured EHRs in the lung cancer patient cohort of 864\npatients with 161,933 various types of clinical notes showed that much more\ndetailed information about smoking, education, and employment were only\ncaptured in clinical narratives and that it is necessary to use both clinical\nnarratives and structured EHRs to construct a more complete picture of\npatients' SBDoH factors."", 'Sequence to sequence (Seq2Seq) learning has recently been used for\nabstractive and extractive summarization. In current study, Seq2Seq models have\nbeen used for eBay product description summarization. We propose a novel\nDocument-Context based Seq2Seq models using RNNs for abstractive and extractive\nsummarizations. Intuitively, this is similar to humans reading the title,\nabstract or any other contextual information before reading the document. This\ngives humans a high-level idea of what the document is about. We use this idea\nand propose that Seq2Seq models should be started with contextual information\nat the first time-step of the input to obtain better summaries. In this manner,\nthe output summaries are more document centric, than being generic, overcoming\none of the major hurdles of using generative models. We generate\ndocument-context from user-behavior and seller provided information. We train\nand evaluate our models on human-extracted-golden-summaries. The\ndocument-contextual Seq2Seq models outperform standard Seq2Seq models.\nMoreover, generating human extracted summaries is prohibitively expensive to\nscale, we therefore propose a semi-supervised technique for extracting\napproximate summaries and using it for training Seq2Seq models at scale.\nSemi-supervised models are evaluated against human extracted summaries and are\nfound to be of similar efficacy. We provide side by side comparison for\nabstractive and extractive summarizers (contextual and non-contextual) on same\nevaluation dataset. Overall, we provide methodologies to use and evaluate the\nproposed techniques for large document summarization. Furthermore, we found\nthese techniques to be highly effective, which is not the case with existing\ntechniques.', 'Cross-lingual document representations enable language understanding in\nmultilingual contexts and allow transfer learning from high-resource to\nlow-resource languages at the document level. Recently large pre-trained\nlanguage models such as BERT, XLM and XLM-RoBERTa have achieved great success\nwhen fine-tuned on sentence-level downstream tasks. It is tempting to apply\nthese cross-lingual models to document representation learning. However, there\nare two challenges: (1) these models impose high costs on long document\nprocessing and thus many of them have strict length limit; (2) model\nfine-tuning requires extra data and computational resources, which is not\npractical in resource-limited settings. In this work, we address these\nchallenges by proposing unsupervised Language-Agnostic Weighted Document\nRepresentations (LAWDR). We study the geometry of pre-trained sentence\nembeddings and leverage it to derive document representations without\nfine-tuning. Evaluated on cross-lingual document alignment, LAWDR demonstrates\ncomparable performance to state-of-the-art models on benchmark datasets.', 'Document grounded generation is the task of using the information provided in\na document to improve text generation. This work focuses on two different\ndocument grounded generation tasks: Wikipedia Update Generation task and\nDialogue response generation. Our work introduces two novel adaptations of\nlarge scale pre-trained encoder-decoder models focusing on building context\ndriven representation of the document and enabling specific attention to the\ninformation in the document. Additionally, we provide a stronger BART baseline\nfor these tasks. Our proposed techniques outperform existing methods on both\nautomated (at least 48% increase in BLEU-4 points) and human evaluation for\ncloseness to reference and relevance to the document. Furthermore, we perform\ncomprehensive manual inspection of the generated output and categorize errors\nto provide insights into future directions in modeling these tasks.', 'Most of the existing information extraction frameworks (Wadden et al., 2019;\nVeysehet al., 2020) focus on sentence-level tasks and are hardly able to\ncapture the consolidated information from a given document. In our endeavour to\ngenerate precise document-level information frames from lengthy textual\nrecords, we introduce the task of Information Aggregation or Argument\nAggregation. More specifically, our aim is to filter irrelevant and redundant\nargument mentions that were extracted at a sentence level and render a document\nlevel information frame. Majority of the existing works have been observed to\nresolve related tasks of document-level event argument extraction (Yang et al.,\n2018a; Zheng et al., 2019a) and salient entity identification (Jain et\nal.,2020) using supervised techniques. To remove dependency from large amounts\nof labelled data, we explore the task of information aggregation using\nweakly-supervised techniques. In particular, we present an extractive algorithm\nwith multiple sieves which adopts active learning strategies to work\nefficiently in low-resource settings. For this task, we have annotated our own\ntest dataset comprising of 131 document information frames and have released\nthe code and dataset to further research prospects in this new domain. To the\nbest of our knowledge, we are the first to establish baseline results for this\ntask in English. Our data and code are publicly available at\nhttps://github.com/DebanjanaKar/ArgFuse.', ""A variety of methods existing for generating synthetic electronic health\nrecords (EHRs), but they are not capable of generating unstructured text, like\nemergency department (ED) chief complaints, history of present illness or\nprogress notes. Here, we use the encoder-decoder model, a deep learning\nalgorithm that features in many contemporary machine translation systems, to\ngenerate synthetic chief complaints from discrete variables in EHRs, like age\ngroup, gender, and discharge diagnosis. After being trained end-to-end on\nauthentic records, the model can generate realistic chief complaint text that\npreserves much of the epidemiological information in the original data. As a\nside effect of the model's optimization goal, these synthetic chief complaints\nare also free of relatively uncommon abbreviation and misspellings, and they\ninclude none of the personally-identifiable information (PII) that was in the\ntraining data, suggesting it may be used to support the de-identification of\ntext in EHRs. When combined with algorithms like generative adversarial\nnetworks (GANs), our model could be used to generate fully-synthetic EHRs,\nfacilitating data sharing between healthcare providers and researchers and\nimproving our ability to develop machine learning methods tailored to the\ninformation in healthcare data."", 'We present an overview of the CLEF-2019 Lab ProtestNews on Extracting\nProtests from News in the context of generalizable natural language processing.\nThe lab consists of document, sentence, and token level information\nclassification and extraction tasks that were referred as task 1, task 2, and\ntask 3 respectively in the scope of this lab. The tasks required the\nparticipants to identify protest relevant information from English local news\nat one or more aforementioned levels in a cross-context setting, which is\ncross-country in the scope of this lab. The training and development data were\ncollected from India and test data was collected from India and China. The lab\nattracted 58 teams to participate in the lab. 12 and 9 of these teams submitted\nresults and working notes respectively. We have observed neural networks yield\nthe best results and the performance drops significantly for majority of the\nsubmissions in the cross-country setting, which is China.', 'Augmentation of disease diagnosis and decision-making in healthcare with\nmachine learning algorithms is gaining much impetus in recent years. In\nparticular, in the current epidemiological situation caused by COVID-19\npandemic, swift and accurate prediction of disease diagnosis with machine\nlearning algorithms could facilitate identification and care of vulnerable\nclusters of population, such as those having multi-morbidity conditions. In\norder to build a useful disease diagnosis prediction system, advancement in\nboth data representation and development of machine learning architectures are\nimperative. First, with respect to data collection and representation, we face\nsevere problems due to multitude of formats and lack of coherency prevalent in\nElectronic Health Records (EHRs). This causes hindrance in extraction of\nvaluable information contained in EHRs. Currently, no universal global data\nstandard has been established. As a useful solution, we develop and publish a\nPython package to transform public health dataset into an easy to access\nuniversal format. This data transformation to an international health data\nformat facilitates researchers to easily combine EHR datasets with clinical\ndatasets of diverse formats. Second, machine learning algorithms that predict\nmultiple disease diagnosis categories simultaneously remain underdeveloped. We\npropose two novel model architectures in this regard. First, DeepObserver,\nwhich uses structured numerical data to predict the diagnosis categories and\nsecond, ClinicalBERT_Multi, that incorporates rich information available in\nclinical notes via natural language processing methods and also provides\ninterpretable visualizations to medical practitioners. We show that both models\ncan predict multiple diagnoses simultaneously with high accuracy.', 'Objective: Allowing patients to access their own electronic health record\n(EHR) notes through online patient portals has the potential to improve\npatient-centered care. However, medical jargon, which abounds in EHR notes, has\nbeen shown to be a barrier for patient EHR comprehension. Existing knowledge\nbases that link medical jargon to lay terms or definitions play an important\nrole in alleviating this problem but have low coverage of medical jargon in\nEHRs. We developed a data-driven approach that mines EHRs to identify and rank\nmedical jargon based on its importance to patients, to support the building of\nEHR-centric lay language resources.\n  Methods: We developed an innovative adapted distant supervision (ADS) model\nbased on support vector machines to rank medical jargon from EHRs. For distant\nsupervision, we utilized the open-access, collaborative consumer health\nvocabulary, a large, publicly available resource that links lay terms to\nmedical jargon. We explored both knowledge-based features from the Unified\nMedical Language System and distributed word representations learned from\nunlabeled large corpora. We evaluated the ADS model using physician-identified\nimportant medical terms.\n  Results: Our ADS model significantly surpassed two state-of-the-art automatic\nterm recognition methods, TF*IDF and C-Value, yielding 0.810 ROC-AUC versus\n0.710 and 0.667, respectively. Our model identified 10K important medical\njargon terms after ranking over 100K candidate terms mined from over 7,500 EHR\nnarratives.\n  Conclusion: Our work is an important step towards enriching lexical resources\nthat link medical jargon to lay terms/definitions to support patient EHR\ncomprehension. The identified medical jargon terms and their rankings are\navailable upon request.', 'Deep reading models for question-answering have demonstrated promising\nperformance over the last couple of years. However current systems tend to\nlearn how to cleverly extract a span of the source document, based on its\nsimilarity with the question, instead of seeking for the appropriate answer.\nIndeed, a reading machine should be able to detect relevant passages in a\ndocument regarding a question, but more importantly, it should be able to\nreason over the important pieces of the document in order to produce an answer\nwhen it is required. To motivate this purpose, we present ReviewQA, a\nquestion-answering dataset based on hotel reviews. The questions of this\ndataset are linked to a set of relational understanding competencies that we\nexpect a model to master. Indeed, each question comes with an associated type\nthat characterizes the required competency. With this framework, it is possible\nto benchmark the main families of models and to get an overview of what are the\nstrengths and the weaknesses of a given model on the set of tasks evaluated in\nthis dataset. Our corpus contains more than 500.000 questions in natural\nlanguage over 100.000 hotel reviews. Our setup is projective, the answer of a\nquestion does not need to be extracted from a document, like in most of the\nrecent datasets, but selected among a set of candidates that contains all the\npossible answers to the questions of the dataset. Finally, we present several\nbaselines over this dataset.', 'Measuring similarity between texts is an important task for several\napplications. Available approaches to measure document similarity are\ninadequate for document pairs that have non-comparable lengths, such as a long\ndocument and its summary. This is because of the lexical, contextual and the\nabstraction gaps between a long document of rich details and its concise\nsummary of abstract information. In this paper, we present a document matching\napproach to bridge this gap, by comparing the texts in a common space of hidden\ntopics. We evaluate the matching algorithm on two matching tasks and find that\nit consistently and widely outperforms strong baselines. We also highlight the\nbenefits of incorporating domain knowledge to text matching.', 'Modern electronic health records (EHRs) provide data to answer clinically\nmeaningful questions. The growing data in EHRs makes healthcare ripe for the\nuse of machine learning. However, learning in a clinical setting presents\nunique challenges that complicate the use of common machine learning\nmethodologies. For example, diseases in EHRs are poorly labeled, conditions can\nencompass multiple underlying endotypes, and healthy individuals are\nunderrepresented. This article serves as a primer to illuminate these\nchallenges and highlights opportunities for members of the machine learning\ncommunity to contribute to healthcare.', 'The increased adoption of Electronic Health Records(EHRs) has brought changes\nto the way the patient care is carried out. The rich heterogeneous and temporal\ndata space stored in EHRs can be leveraged by machine learning models to\ncapture the underlying information and make clinically relevant predictions.\nThis can be exploited to support public health activities such as\npharmacovigilance and specifically mitigate the public health issue of Adverse\nDrug Events(ADEs). The aim of this article is, therefore, to investigate the\nvarious ways of handling temporal data for the purpose of detecting ADEs. Based\non a review of the existing literature, 11 articles from the last 10 years were\nchosen to be studied. According to the literature retrieved the main methods\nwere found to fall into 5 different approaches: based on temporal abstraction,\ngraph-based, learning weights and data tables containing time series of\ndifferent length. To that end, EHRs are a valuable source that has led current\nresearch to the automatic detection of ADEs. Yet there still exists a great\ndeal of challenges that concerns the exploitation of the heterogeneous, data\ntypes with temporal information included in EHRs for predicting ADEs.', 'In the Query Focused Multi-Document Summarization (QF-MDS) task, a set of\ndocuments and a query are given where the goal is to generate a summary from\nthese documents based on the given query. However, one major challenge for this\ntask is the lack of availability of labeled training datasets. To overcome this\nissue, in this paper, we propose a novel weakly supervised learning approach\nvia utilizing distant supervision. In particular, we use datasets similar to\nthe target dataset as the training data where we leverage pre-trained sentence\nsimilarity models to generate the weak reference summary of each individual\ndocument in a document set from the multi-document gold reference summaries.\nThen, we iteratively train our summarization model on each single-document to\nalleviate the computational complexity issue that occurs while training neural\nsummarization models in multiple documents (i.e., long sequences) at once.\nExperimental results in Document Understanding Conferences (DUC) datasets show\nthat our proposed approach sets a new state-of-the-art result in terms of\nvarious evaluation metrics.', 'Text Mining is a field that aims at extracting information from textual data.\nOne of the challenges of such field of study comes from the pre-processing\nstage in which a vector (and structured) representation should be extracted\nfrom unstructured data. The common extraction creates large and sparse vectors\nrepresenting the importance of each term to a document. As such, this usually\nleads to the curse-of-dimensionality that plagues most machine learning\nalgorithms. To cope with this issue, in this paper we propose a new supervised\nfeature extraction and reduction algorithm, named DCDistance, that creates\nfeatures based on the distance between a document to a representative of each\nclass label. As such, the proposed technique can reduce the features set in\nmore than 99% of the original set. Additionally, this algorithm was also\ncapable of improving the classification accuracy over a set of benchmark\ndatasets when compared to traditional and state-of-the-art features selection\nalgorithms.', 'In this paper we present a method for the unsupervised clustering of\nhigh-dimensional binary data, with a special focus on electronic healthcare\nrecords. We present a robust and efficient heuristic to face this problem using\ntensor decomposition. We present the reasons why this approach is preferable\nfor tasks such as clustering patient records, to more commonly used\ndistance-based methods. We run the algorithm on two datasets of healthcare\nrecords, obtaining clinically meaningful results.', 'As the application of deep learning has expanded to real-world problems with\ninsufficient volume of training data, transfer learning recently has gained\nmuch attention as means of improving the performance in such small-data regime.\nHowever, when existing methods are applied between heterogeneous architectures\nand tasks, it becomes more important to manage their detailed configurations\nand often requires exhaustive tuning on them for the desired performance. To\naddress the issue, we propose a novel transfer learning approach based on\nmeta-learning that can automatically learn what knowledge to transfer from the\nsource network to where in the target network. Given source and target\nnetworks, we propose an efficient training scheme to learn meta-networks that\ndecide (a) which pairs of layers between the source and target networks should\nbe matched for knowledge transfer and (b) which features and how much knowledge\nfrom each feature should be transferred. We validate our meta-transfer approach\nagainst recent transfer learning methods on various datasets and network\narchitectures, on which our automated scheme significantly outperforms the\nprior baselines that find ""what and where to transfer"" in a hand-crafted\nmanner.', 'This paper introduces a conceptually simple, scalable, and highly effective\nBERT-based entity linking model, along with an extensive evaluation of its\naccuracy-speed trade-off. We present a two-stage zero-shot linking algorithm,\nwhere each entity is defined only by a short textual description. The first\nstage does retrieval in a dense space defined by a bi-encoder that\nindependently embeds the mention context and the entity descriptions. Each\ncandidate is then re-ranked with a cross-encoder, that concatenates the mention\nand entity text. Experiments demonstrate that this approach is state of the art\non recent zero-shot benchmarks (6 point absolute gains) and also on more\nestablished non-zero-shot evaluations (e.g. TACKBP-2010), despite its relative\nsimplicity (e.g. no explicit entity embeddings or manually engineered mention\ntables). We also show that bi-encoder linking is very fast with nearest\nneighbour search (e.g. linking with 5.9 million candidates in 2 milliseconds),\nand that much of the accuracy gain from the more expensive cross-encoder can be\ntransferred to the bi-encoder via knowledge distillation. Our code and models\nare available at https://github.com/facebookresearch/BLINK.', 'We present Attract-Repel, an algorithm for improving the semantic quality of\nword vectors by injecting constraints extracted from lexical resources.\nAttract-Repel facilitates the use of constraints from mono- and cross-lingual\nresources, yielding semantically specialised cross-lingual vector spaces. Our\nevaluation shows that the method can make use of existing cross-lingual\nlexicons to construct high-quality vector spaces for a plethora of different\nlanguages, facilitating semantic transfer from high- to lower-resource ones.\nThe effectiveness of our approach is demonstrated with state-of-the-art results\non semantic similarity datasets in six languages. We next show that\nAttract-Repel-specialised vectors boost performance in the downstream task of\ndialogue state tracking (DST) across multiple languages. Finally, we show that\ncross-lingual vector spaces produced by our algorithm facilitate the training\nof multilingual DST models, which brings further performance improvements.', 'Handwriting is a natural and versatile method for human-computer interaction,\nespecially on small mobile devices such as smart phones. However, as\nhandwriting varies significantly from person to person, it is difficult to\ndesign handwriting recognizers that perform well for all users. A natural\nsolution is to use machine learning to adapt the recognizer to the user. One\ncomplicating factor is that, as the computer adapts to the user, the user also\nadapts to the computer and probably changes their handwriting. This paper\ninvestigates the dynamics of co-adaptation, a process in which both the\ncomputer and the user are adapting their behaviors in order to improve the\nspeed and accuracy of the communication through handwriting. We devised an\ninformation-theoretic framework for quantifying the efficiency of a handwriting\nsystem where the system includes both the user and the computer. Using this\nframework, we analyzed data collected from an adaptive handwriting recognition\nsystem and characterized the impact of machine adaptation and of human\nadaptation. We found that both machine adaptation and human adaptation have\nsignificant impact on the input rate and must be considered together in order\nto improve the efficiency of the system as a whole.', 'The success of pretrained cross-lingual language models relies on two\nessential abilities, i.e., generalization ability for learning downstream tasks\nin a source language, and cross-lingual transferability for transferring the\ntask knowledge to other languages. However, current methods jointly learn the\ntwo abilities in a single-phase cross-lingual pretraining process, resulting in\na trade-off between generalization and cross-lingual transfer. In this paper,\nwe propose cross-lingual language model meta-pretraining, which learns the two\nabilities in different training phases. Our method introduces an additional\nmeta-pretraining phase before cross-lingual pretraining, where the model learns\ngeneralization ability on a large-scale monolingual corpus. Then, the model\nfocuses on learning cross-lingual transfer on a multilingual corpus.\nExperimental results show that our method improves both generalization and\ncross-lingual transfer, and produces better-aligned representations across\ndifferent languages.', 'We explore using multilingual document embeddings for nearest neighbor mining\nof parallel data. Three document-level representations are investigated: (i)\ndocument embeddings generated by simply averaging multilingual sentence\nembeddings; (ii) a neural bag-of-words (BoW) document encoding model; (iii) a\nhierarchical multilingual document encoder (HiDE) that builds on our\nsentence-level model. The results show document embeddings derived from\nsentence-level averaging are surprisingly effective for clean datasets, but\nsuggest models trained hierarchically at the document-level are more effective\non noisy data. Analysis experiments demonstrate our hierarchical models are\nvery robust to variations in the underlying sentence embedding quality. Using\ndocument embeddings trained with HiDE achieves state-of-the-art performance on\nUnited Nations (UN) parallel document mining, 94.9% P@1 for en-fr and 97.3% P@1\nfor en-es.', 'We present a novel model for the problem of ranking a collection of documents\naccording to their semantic similarity to a source (query) document. While the\nproblem of document-to-document similarity ranking has been studied, most\nmodern methods are limited to relatively short documents or rely on the\nexistence of ""ground-truth"" similarity labels. Yet, in most common real-world\ncases, similarity ranking is an unsupervised problem as similarity labels are\nunavailable. Moreover, an ideal model should not be restricted by documents\'\nlength. Hence, we introduce SDR, a self-supervised method for document\nsimilarity that can be applied to documents of arbitrary length. Importantly,\nSDR can be effectively applied to extremely long documents, exceeding the 4,096\nmaximal token limits of Longformer. Extensive evaluations on large document\ndatasets show that SDR significantly outperforms its alternatives across all\nmetrics. To accelerate future research on unlabeled long document similarity\nranking, and as an additional contribution to the community, we herein publish\ntwo human-annotated test sets of long documents similarity evaluation. The SDR\ncode and datasets are publicly available.', 'In this paper, we introduce Iterative Text Summarization (ITS), an\niteration-based model for supervised extractive text summarization, inspired by\nthe observation that it is often necessary for a human to read an article\nmultiple times in order to fully understand and summarize its contents. Current\nsummarization approaches read through a document only once to generate a\ndocument representation, resulting in a sub-optimal representation. To address\nthis issue we introduce a model which iteratively polishes the document\nrepresentation on many passes through the document. As part of our model, we\nalso introduce a selective reading mechanism that decides more accurately the\nextent to which each sentence in the model should be updated. Experimental\nresults on the CNN/DailyMail and DUC2002 datasets demonstrate that our model\nsignificantly outperforms state-of-the-art extractive systems when evaluated by\nmachines and by humans.', 'Multi-document question generation focuses on generating a question that\ncovers the common aspect of multiple documents. Such a model is useful in\ngenerating clarifying options. However, a naive model trained only using the\ntargeted (""positive"") document set may generate too generic questions that\ncover a larger scope than delineated by the document set. To address this\nchallenge, we introduce the contrastive learning strategy where given\n""positive"" and ""negative"" sets of documents, we generate a question that is\nclosely related to the ""positive"" set but is far away from the ""negative"" set.\nThis setting allows generated questions to be more specific and related to the\ntarget document set. To generate such specific questions, we propose\nMulti-Source Coordinated Question Generator (MSCQG), a novel framework that\nincludes a supervised learning (SL) stage and a reinforcement learning (RL)\nstage. In the SL stage, a single-document question generator is trained. In the\nRL stage, a coordinator model is trained to find optimal attention weights to\nalign multiple single-document generators, by optimizing a reward designed to\npromote specificity of generated questions. We also develop an effective\nauxiliary objective, named Set-induced Contrastive Regularization (SCR) that\nimproves the coordinator\'s contrastive learning during the RL stage. We show\nthat our model significantly outperforms several strong baselines, as measured\nby automatic metrics and human evaluation. The source repository is publicly\navailable at \\url{www.github.com/woonsangcho/contrast_qgen}.', 'Document-level relation extraction requires integrating information within\nand across multiple sentences of a document and capturing complex interactions\nbetween inter-sentence entities. However, effective aggregation of relevant\ninformation in the document remains a challenging research question. Existing\napproaches construct static document-level graphs based on syntactic trees,\nco-references or heuristics from the unstructured text to model the\ndependencies. Unlike previous methods that may not be able to capture rich\nnon-local interactions for inference, we propose a novel model that empowers\nthe relational reasoning across sentences by automatically inducing the latent\ndocument-level graph. We further develop a refinement strategy, which enables\nthe model to incrementally aggregate relevant information for multi-hop\nreasoning. Specifically, our model achieves an F1 score of 59.05 on a\nlarge-scale document-level dataset (DocRED), significantly improving over the\nprevious results, and also yields new state-of-the-art results on the CDR and\nGDA dataset. Furthermore, extensive analyses show that the model is able to\ndiscover more accurate inter-sentence relations.', 'State-of-the-art systems in deep question answering proceed as follows: (1)\nan initial document retrieval selects relevant documents, which (2) are then\nprocessed by a neural network in order to extract the final answer. Yet the\nexact interplay between both components is poorly understood, especially\nconcerning the number of candidate documents that should be retrieved. We show\nthat choosing a static number of documents -- as used in prior research --\nsuffers from a noise-information trade-off and yields suboptimal results. As a\nremedy, we propose an adaptive document retrieval model. This learns the\noptimal candidate number for document retrieval, conditional on the size of the\ncorpus and the query. We report extensive experimental results showing that our\nadaptive approach outperforms state-of-the-art methods on multiple benchmark\ndatasets, as well as in the context of corpora with variable sizes.', 'Unstructured documents serving as external knowledge of the dialogues help to\ngenerate more informative responses. Previous research focused on knowledge\nselection (KS) in the document with dialogue. However, dialogue history that is\nnot related to the current dialogue may introduce noise in the KS processing.\nIn this paper, we propose a Compare Aggregate Transformer (CAT) to jointly\ndenoise the dialogue context and aggregate the document information for\nresponse generation. We designed two different comparison mechanisms to reduce\nnoise (before and during decoding). In addition, we propose two metrics for\nevaluating document utilization efficiency based on word overlap. Experimental\nresults on the CMUDoG dataset show that the proposed CAT model outperforms the\nstate-of-the-art approach and strong baselines.', ""Probabilistic topic models like Latent Dirichlet Allocation (LDA) have been\npreviously extended to the bilingual setting. A fundamental modeling assumption\nin several of these extensions is that the input corpora are in the form of\ndocument pairs whose constituent documents share a single topic distribution.\nHowever, this assumption is strong for comparable corpora that consist of\ndocuments thematically similar to an extent only, which are, in turn, the most\ncommonly available or easy to obtain. In this paper we relax this assumption by\nproposing for the paired documents to have separate, yet bound topic\ndistributions. % a binding mechanism between the distributions of the paired\ndocuments. We suggest that the strength of the bound should depend on each\npair's semantic similarity. To estimate the similarity of documents that are\nwritten in different languages we use cross-lingual word embeddings that are\nlearned with shallow neural networks. We evaluate the proposed binding\nmechanism by extending two topic models: a bilingual adaptation of LDA that\nassumes bag-of-words inputs and a model that incorporates part of the text\nstructure in the form of boundaries of semantically coherent segments. To\nassess the performance of the novel topic models we conduct intrinsic and\nextrinsic experiments on five bilingual, comparable corpora of English\ndocuments with French, German, Italian, Spanish and Portuguese documents. The\nresults demonstrate the efficiency of our approach in terms of both topic\ncoherence measured by the normalized point-wise mutual information, and\ngeneralization performance measured by perplexity and in terms of Mean\nReciprocal Rank in a cross-lingual document retrieval task for each of the\nlanguage pairs."", 'Text alignment finds application in tasks such as citation recommendation and\nplagiarism detection. Existing alignment methods operate at a single,\npredefined level and cannot learn to align texts at, for example, sentence and\ndocument levels. We propose a new learning approach that equips previously\nestablished hierarchical attention encoders for representing documents with a\ncross-document attention component, enabling structural comparisons across\ndifferent levels (document-to-document and sentence-to-document). Our component\nis weakly supervised from document pairs and can align at multiple levels. Our\nevaluation on predicting document-to-document relationships and\nsentence-to-document relationships on the tasks of citation recommendation and\nplagiarism detection shows that our approach outperforms previously established\nhierarchical, attention encoders based on recurrent and transformer\ncontextualization that are unaware of structural correspondence between\ndocuments.', 'We investigate the pertinence of methods from algebraic topology for text\ndata analysis. These methods enable the development of\nmathematically-principled isometric-invariant mappings from a set of vectors to\na document embedding, which is stable with respect to the geometry of the\ndocument in the selected metric space. In this work, we evaluate the utility of\nthese topology-based document representations in traditional NLP tasks,\nspecifically document clustering and sentiment classification. We find that the\nembeddings do not benefit text analysis. In fact, performance is worse than\nsimple techniques like $\\textit{tf-idf}$, indicating that the geometry of the\ndocument does not provide enough variability for classification on the basis of\ntopic or sentiment in the chosen datasets.', ""In information retrieval (IR) and related tasks, term weighting approaches\ntypically consider the frequency of the term in the document and in the\ncollection in order to compute a score reflecting the importance of the term\nfor the document. In tasks characterized by the presence of training data (such\nas text classification) it seems logical that the term weighting function\nshould take into account the distribution (as estimated from training data) of\nthe term across the classes of interest. Although `supervised term weighting'\napproaches that use this intuition have been described before, they have failed\nto show consistent improvements. In this article we analyse the possible\nreasons for this failure, and call consolidated assumptions into question.\nFollowing this criticism we propose a novel supervised term weighting approach\nthat, instead of relying on any predefined formula, learns a term weighting\nfunction optimised on the training set of interest; we dub this approach\n\\emph{Learning to Weight} (LTW). The experiments that we run on several\nwell-known benchmarks, and using different learning methods, show that our\nmethod outperforms previous term weighting approaches in text classification."", 'Lifelong learning has recently attracted attention in building machine\nlearning systems that continually accumulate and transfer knowledge to help\nfuture learning. Unsupervised topic modeling has been popularly used to\ndiscover topics from document collections. However, the application of topic\nmodeling is challenging due to data sparsity, e.g., in a small collection of\n(short) documents and thus, generate incoherent topics and sub-optimal document\nrepresentations. To address the problem, we propose a lifelong learning\nframework for neural topic modeling that can continuously process streams of\ndocument collections, accumulate topics and guide future topic modeling tasks\nby knowledge transfer from several sources to better deal with the sparse data.\nIn the lifelong process, we particularly investigate jointly: (1) sharing\ngenerative homologies (latent topics) over lifetime to transfer prior\nknowledge, and (2) minimizing catastrophic forgetting to retain the past\nlearning via novel selective data augmentation, co-training and topic\nregularization approaches. Given a stream of document collections, we apply the\nproposed Lifelong Neural Topic Modeling (LNTM) framework in modeling three\nsparse document collections as future tasks and demonstrate improved\nperformance quantified by perplexity, topic coherence and information retrieval\ntask.', 'A significant roadblock in multilingual neural language modeling is the lack\nof labeled non-English data. One potential method for overcoming this issue is\nlearning cross-lingual text representations that can be used to transfer the\nperformance from training on English tasks to non-English tasks, despite little\nto no task-specific non-English data. In this paper, we explore a natural setup\nfor learning cross-lingual sentence representations: the dual-encoder. We\nprovide a comprehensive evaluation of our cross-lingual representations on a\nnumber of monolingual, cross-lingual, and zero-shot/few-shot learning tasks,\nand also give an analysis of different learned cross-lingual embedding spaces.', ""Document similarity is the problem of estimating the degree to which a given\npair of documents has similar semantic content. An accurate document similarity\nmeasure can improve several enterprise relevant tasks such as document\nclustering, text mining, and question-answering. In this paper, we show that a\ndocument's thematic flow, which is often disregarded by bag-of-word techniques,\nis pivotal in estimating their similarity. To this end, we propose a novel\nsemantic document similarity framework, called SimDoc. We model documents as\ntopic-sequences, where topics represent latent generative clusters of related\nwords. Then, we use a sequence alignment algorithm to estimate their semantic\nsimilarity. We further conceptualize a novel mechanism to compute topic-topic\nsimilarity to fine tune our system. In our experiments, we show that SimDoc\noutperforms many contemporary bag-of-words techniques in accurately computing\ndocument similarity, and on practical applications such as document clustering."", 'The organization and evolution of science has recently become itself an\nobject of scientific quantitative investigation, thanks to the wealth of\ninformation that can be extracted from scientific documents, such as citations\nbetween papers and co-authorship between researchers. However, only few studies\nhave focused on the concepts that characterize full documents and that can be\nextracted and analyzed, revealing the deeper organization of scientific\nknowledge. Unfortunately, several concepts can be so common across documents\nthat they hinder the emergence of the underlying topical structure of the\ndocument corpus, because they give rise to a large amount of spurious and\ntrivial relations among documents. To identify and remove common concepts, we\nintroduce a method to gauge their relevance according to an objective\ninformation-theoretic measure related to the statistics of their occurrence\nacross the document corpus. After progressively removing concepts that,\naccording to this metric, can be considered as generic, we find that the topic\norganization displays a correspondingly more refined structure.', 'Recent developments in natural language representations have been accompanied\nby large and expensive models that leverage vast amounts of general-domain text\nthrough self-supervised pre-training. Due to the cost of applying such models\nto down-stream tasks, several model compression techniques on pre-trained\nlanguage representations have been proposed (Sun et al., 2019; Sanh, 2019).\nHowever, surprisingly, the simple baseline of just pre-training and fine-tuning\ncompact models has been overlooked. In this paper, we first show that\npre-training remains important in the context of smaller architectures, and\nfine-tuning pre-trained compact models can be competitive to more elaborate\nmethods proposed in concurrent work. Starting with pre-trained compact models,\nwe then explore transferring task knowledge from large fine-tuned models\nthrough standard knowledge distillation. The resulting simple, yet effective\nand general algorithm, Pre-trained Distillation, brings further improvements.\nThrough extensive experiments, we more generally explore the interaction\nbetween pre-training and distillation under two variables that have been\nunder-studied: model size and properties of unlabeled task data. One surprising\nobservation is that they have a compound effect even when sequentially applied\non the same data. To accelerate future research, we will make our 24\npre-trained miniature BERT models publicly available.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', 'Research is a tertiary priority in the EHR, where the priorities are patient\ncare and billing. Because of this, the data is not standardized or formatted in\na manner easily adapted to machine learning approaches. Data may be missing for\na large variety of reasons ranging from individual input styles to differences\nin clinical decision making, for example, which lab tests to issue. Few\npatients are annotated at a research quality, limiting sample size and\npresenting a moving gold standard. Patient progression over time is key to\nunderstanding many diseases but many machine learning algorithms require a\nsnapshot, at a single time point, to create a usable vector form. Furthermore,\nalgorithms that produce black box results do not provide the interpretability\nrequired for clinical adoption. This chapter discusses these challenges and\nothers in applying machine learning techniques to the structured EHR (i.e.\nPatient Demographics, Family History, Medication Information, Vital Signs,\nLaboratory Tests, Genetic Testing). It does not cover feature extraction from\nadditional sources such as imaging data or free text patient notes but the\napproaches discussed can include features extracted from these sources.', 'The increasing amount of online content motivated the development of\nmulti-document summarization methods. In this work, we explore straightforward\napproaches to extend single-document summarization methods to multi-document\nsummarization. The proposed methods are based on the hierarchical combination\nof single-document summaries, and achieves state of the art results.', 'Standard neural machine translation (NMT) is on the assumption that the\ndocument-level context is independent. Most existing document-level NMT\napproaches are satisfied with a smattering sense of global document-level\ninformation, while this work focuses on exploiting detailed document-level\ncontext in terms of a memory network. The capacity of the memory network that\ndetecting the most relevant part of the current sentence from memory renders a\nnatural solution to model the rich document-level context. In this work, the\nproposed document-aware memory network is implemented to enhance the\nTransformer NMT baseline. Experiments on several tasks show that the proposed\nmethod significantly improves the NMT performance over strong Transformer\nbaselines and other related studies.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.', 'Under special circumstances, summaries should conform to a particular style\nwith patterns, such as court judgments and abstracts in academic papers. To\nthis end, the prototype document-summary pairs can be utilized to generate\nbetter summaries. There are two main challenges in this task: (1) the model\nneeds to incorporate learned patterns from the prototype, but (2) should avoid\ncopying contents other than the patternized words---such as irrelevant\nfacts---into the generated summaries. To tackle these challenges, we design a\nmodel named Prototype Editing based Summary Generator (PESG). PESG first learns\nsummary patterns and prototype facts by analyzing the correlation between a\nprototype document and its summary. Prototype facts are then utilized to help\nextract facts from the input document. Next, an editing generator generates new\nsummary based on the summary pattern or extracted facts. Finally, to address\nthe second challenge, a fact checker is used to estimate mutual information\nbetween the input document and generated summary, providing an additional\nsignal for the generator. Extensive experiments conducted on a large-scale\nreal-world text summarization dataset show that PESG achieves the\nstate-of-the-art performance in terms of both automatic metrics and human\nevaluations.', ""Electronic Health Records (EHRs) in hospital information systems contain\npatients' diagnosis and treatments, so EHRs are essential to clinical data\nmining. Of all the tasks in the mining process, Chinese Word Segmentation (CWS)\nis a fundamental and important one, and most state-of-the-art methods greatly\nrely on large-scale of manually-annotated data. Since annotation is\ntime-consuming and expensive, efforts have been devoted to techniques, such as\nactive learning, to locate the most informative samples for modeling. In this\npaper, we follow the trend and present an active learning method for CWS in\nEHRs. Specically, a new sampling strategy combining Normalized Entropy with\nLoss Prediction (NE-LP) is proposed to select the most representative data.\nMeanwhile, to minimize the computational cost of learning, we propose a joint\nmodel including a word segmenter and a loss prediction model. Furthermore, to\ncapture interactions between adjacent characters, bigram features are also\napplied in the joint model. To illustrate the effectiveness of NE-LP, we\nconducted experiments on EHRs collected from the Shuguang Hospital Affiliated\nto Shanghai University of Traditional Chinese Medicine. The results demonstrate\nthat NE-LP consistently outperforms conventional uncertainty-based sampling\nstrategies for active learning in CWS."", ""The selection of a suitable document representation approach plays a crucial\nrole in the performance of a document clustering task. Being able to pick out\nrepresentative words within a document can lead to substantial improvements in\ndocument clustering. In the case of web documents, the HTML markup that defines\nthe layout of the content provides additional structural information that can\nbe further exploited to identify representative words. In this paper we\nintroduce a fuzzy term weighing approach that makes the most of the HTML\nstructure for document clustering. We set forth and build on the hypothesis\nthat a good representation can take advantage of how humans skim through\ndocuments to extract the most representative words. The authors of web pages\nmake use of HTML tags to convey the most important message of a web page\nthrough page elements that attract the readers' attention, such as page titles\nor emphasized elements. We define a set of criteria to exploit the information\nprovided by these page elements, and introduce a fuzzy combination of these\ncriteria that we evaluate within the context of a web page clustering task. Our\nproposed approach, called Abstract Fuzzy Combination of Criteria (AFCC), can\nadapt to datasets whose features are distributed differently, achieving good\nresults compared to other similar fuzzy logic based approaches and TF-IDF\nacross different datasets."", 'We introduce a novel type of text representation that preserves the 2D layout\nof a document. This is achieved by encoding each document page as a\ntwo-dimensional grid of characters. Based on this representation, we present a\ngeneric document understanding pipeline for structured documents. This pipeline\nmakes use of a fully convolutional encoder-decoder network that predicts a\nsegmentation mask and bounding boxes. We demonstrate its capabilities on an\ninformation extraction task from invoices and show that it significantly\noutperforms approaches based on sequential text or document images.', 'Keyphrase extraction from documents is useful to a variety of applications\nsuch as information retrieval and document summarization. This paper presents\nan end-to-end method called DivGraphPointer for extracting a set of diversified\nkeyphrases from a document. DivGraphPointer combines the advantages of\ntraditional graph-based ranking methods and recent neural network-based\napproaches. Specifically, given a document, a word graph is constructed from\nthe document based on word proximity and is encoded with graph convolutional\nnetworks, which effectively capture document-level word salience by modeling\nlong-range dependency between words in the document and aggregating multiple\nappearances of identical words into one node. Furthermore, we propose a\ndiversified point network to generate a set of diverse keyphrases out of the\nword graph in the decoding process. Experimental results on five benchmark data\nsets show that our proposed method significantly outperforms the existing\nstate-of-the-art approaches.', ""Neural conversation models are known to generate appropriate but\nnon-informative responses in general. A scenario where informativeness can be\nsignificantly enhanced is Conversing by Reading (CbR), where conversations take\nplace with respect to a given external document. In previous work, the external\ndocument is utilized by (1) creating a context-aware document memory that\nintegrates information from the document and the conversational context, and\nthen (2) generating responses referring to the memory. In this paper, we\npropose to create the document memory with some anticipated responses in mind.\nThis is achieved using a teacher-student framework. The teacher is given the\nexternal document, the context, and the ground-truth response, and learns how\nto build a response-aware document memory from three sources of information.\nThe student learns to construct a response-anticipated document memory from the\nfirst two sources, and the teacher's insight on memory creation. Empirical\nresults show that our model outperforms the previous state-of-the-art for the\nCbR task."", 'Models of object vision have been of great interest in computer vision and\nvisual neuroscience. During the last decades, several models have been\ndeveloped to extract visual features from images for object recognition tasks.\nSome of these were inspired by the hierarchical structure of primate visual\nsystem, and some others were engineered models. The models are varied in\nseveral aspects: models that are trained by supervision, models trained without\nsupervision, and models (e.g. feature extractors) that are fully hard-wired and\ndo not need training. Some of the models come with a deep hierarchical\nstructure consisting of several layers, and some others are shallow and come\nwith only one or two layers of processing. More recently, new models have been\ndeveloped that are not hand-tuned but trained using millions of images, through\nwhich they learn how to extract informative task-related features. Here I will\nsurvey all these different models and provide the reader with an intuitive, as\nwell as a more detailed, understanding of the underlying computations in each\nof the models.', 'Previous researchers have considered sentiment analysis as a document\nclassification task, in which input documents are classified into predefined\nsentiment classes. Although there are sentences in a document that support\nimportant evidences for sentiment analysis and sentences that do not, they have\ntreated the document as a bag of sentences. In other words, they have not\nconsidered the importance of each sentence in the document. To effectively\ndetermine polarity of a document, each sentence in the document should be dealt\nwith different degrees of importance. To address this problem, we propose a\ndocument-level sentence classification model based on deep neural networks, in\nwhich the importance degrees of sentences in documents are automatically\ndetermined through gate mechanisms. To verify our new sentiment analysis model,\nwe conducted experiments using the sentiment datasets in the four different\ndomains such as movie reviews, hotel reviews, restaurant reviews, and music\nreviews. In the experiments, the proposed model outperformed previous\nstate-of-the-art models that do not consider importance differences of\nsentences in a document. The experimental results show that the importance of\nsentences should be considered in a document-level sentiment classification\ntask.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'Electronic Health Records (EHRs) provide a wealth of information for machine\nlearning algorithms to predict the patient outcome from the data including\ndiagnostic information, vital signals, lab tests, drug administration, and\ndemographic information. Machine learning models can be built, for example, to\nevaluate patients based on their predicted mortality or morbidity and to\npredict required resources for efficient resource management in hospitals. In\nthis paper, we demonstrate that an attacker can manipulate the machine learning\npredictions with EHRs easily and selectively at test time by backdoor attacks\nwith the poisoned training data. Furthermore, the poison we create has\nstatistically similar features to the original data making it hard to detect,\nand can also attack multiple machine learning models without any knowledge of\nthe models. With less than 5% of the raw EHR data poisoned, we achieve average\nattack success rates of 97% on mortality prediction tasks with MIMIC-III\ndatabase against Logistic Regression, Multilayer Perceptron, and Long\nShort-term Memory models simultaneously.', 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', 'Traditional learning-based coreference resolvers operate by training the\nmention-pair model for determining whether two mentions are coreferent or not.\nThough conceptually simple and easy to understand, the mention-pair model is\nlinguistically rather unappealing and lags far behind the heuristic-based\ncoreference models proposed in the pre-statistical NLP era in terms of\nsophistication. Two independent lines of recent research have attempted to\nimprove the mention-pair model, one by acquiring the mention-ranking model to\nrank preceding mentions for a given anaphor, and the other by training the\nentity-mention model to determine whether a preceding cluster is coreferent\nwith a given mention. We propose a cluster-ranking approach to coreference\nresolution, which combines the strengths of the mention-ranking model and the\nentity-mention model, and is therefore theoretically more appealing than both\nof these models. In addition, we seek to improve cluster rankers via two\nextensions: (1) lexicalization and (2) incorporating knowledge of anaphoricity\nby jointly modeling anaphoricity determination and coreference resolution.\nExperimental results on the ACE data sets demonstrate the superior performance\nof cluster rankers to competing approaches as well as the effectiveness of our\ntwo extensions.', 'We introduce the ""Incremental Implicitly-Refined Classi-fication (IIRC)""\nsetup, an extension to the class incremental learning setup where the incoming\nbatches of classes have two granularity levels. i.e., each sample could have a\nhigh-level (coarse) label like ""bear"" and a low-level (fine) label like ""polar\nbear"". Only one label is provided at a time, and the model has to figure out\nthe other label if it has already learnfed it. This setup is more aligned with\nreal-life scenarios, where a learner usually interacts with the same family of\nentities multiple times, discovers more granularity about them, while still\ntrying not to forget previous knowledge. Moreover, this setup enables\nevaluating models for some important lifelong learning challenges that cannot\nbe easily addressed under the existing setups. These challenges can be\nmotivated by the example ""if a model was trained on the class bear in one task\nand on polar bear in another task, will it forget the concept of bear, will it\nrightfully infer that a polar bear is still a bear? and will it wrongfully\nassociate the label of polar bear to other breeds of bear?"". We develop a\nstandardized benchmark that enables evaluating models on the IIRC setup. We\nevaluate several state-of-the-art lifelong learning algorithms and highlight\ntheir strengths and limitations. For example, distillation-based methods\nperform relatively well but are prone to incorrectly predicting too many labels\nper image. We hope that the proposed setup, along with the benchmark, would\nprovide a meaningful problem setting to the practitioners', 'Recent studies emphasize the need of document context in human evaluation of\nmachine translations, but little research has been done on the impact of user\ninterfaces on annotator productivity and the reliability of assessments. In\nthis work, we compare human assessment data from the last two WMT evaluation\ncampaigns collected via two different methods for document-level evaluation.\nOur analysis shows that a document-centric approach to evaluation where the\nannotator is presented with the entire document context on a screen leads to\nhigher quality segment and document level assessments. It improves the\ncorrelation between segment and document scores and increases inter-annotator\nagreement for document scores but is considerably more time consuming for\nannotators.', 'Term frequency normalization is a serious issue since lengths of documents\nare various. Generally, documents become long due to two different reasons -\nverbosity and multi-topicality. First, verbosity means that the same topic is\nrepeatedly mentioned by terms related to the topic, so that term frequency is\nmore increased than the well-summarized one. Second, multi-topicality indicates\nthat a document has a broad discussion of multi-topics, rather than single\ntopic. Although these document characteristics should be differently handled,\nall previous methods of term frequency normalization have ignored these\ndifferences and have used a simplified length-driven approach which decreases\nthe term frequency by only the length of a document, causing an unreasonable\npenalization. To attack this problem, we propose a novel TF normalization\nmethod which is a type of partially-axiomatic approach. We first formulate two\nformal constraints that the retrieval model should satisfy for documents having\nverbose and multi-topicality characteristic, respectively. Then, we modify\nlanguage modeling approaches to better satisfy these two constraints, and\nderive novel smoothing methods. Experimental results show that the proposed\nmethod increases significantly the precision for keyword queries, and\nsubstantially improves MAP (Mean Average Precision) for verbose queries.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am currently developing a transparent and reliable AI chatbot. I am in search of AI models that can complete incomplete sentences or answer questions based on the next token prediction. For transparency, I need models that can explicitly compute the probability of the next token and update it in real-time based on user input. I also aim to provide a visualization that displays the perceptrons and weights utilized during each inference to the user. To ensure trustworthiness, I require techniques that can perturb the training set to increase the model's resilience against adversarial attacks. The model should also possess the ability to forget or replace outdated knowledge, as some data from the training set may become incorrect over time. Considering that a significant portion of potential training data from the internet includes sensitive personal information, I also require an efficient filter to automatically mask such information prior to its use in training. Ultimately, my objective is to develop an AI chatbot that is both transparent and trustworthy.",[],"['When recurrent neural network transducers (RNNTs) are trained using the\ntypical maximum likelihood criterion, the prediction network is trained only on\nground truth label sequences. This leads to a mismatch during inference, known\nas exposure bias, when the model must deal with label sequences containing\nerrors. In this paper we investigate approaches to reducing exposure bias in\ntraining to improve the generalization of RNNT models for automatic speech\nrecognition (ASR). A label-preserving input perturbation to the prediction\nnetwork is introduced. The input token sequences are perturbed using SwitchOut\nand scheduled sampling based on an additional token language model. Experiments\nconducted on the 300-hour Switchboard dataset demonstrate their effectiveness.\nBy reducing the exposure bias, we show that we can further improve the accuracy\nof a high-performance RNNT ASR model and obtain state-of-the-art results on the\n300-hour Switchboard dataset.', 'Recent neural models of dialogue generation offer great promise for\ngenerating responses for conversational agents, but tend to be shortsighted,\npredicting utterances one at a time while ignoring their influence on future\noutcomes. Modeling the future direction of a dialogue is crucial to generating\ncoherent, interesting dialogues, a need which led traditional NLP models of\ndialogue to draw on reinforcement learning. In this paper, we show how to\nintegrate these goals, applying deep reinforcement learning to model future\nreward in chatbot dialogue. The model simulates dialogues between two virtual\nagents, using policy gradient methods to reward sequences that display three\nuseful conversational properties: informativity (non-repetitive turns),\ncoherence, and ease of answering (related to forward-looking function). We\nevaluate our model on diversity, length as well as with human judges, showing\nthat the proposed algorithm generates more interactive responses and manages to\nfoster a more sustained conversation in dialogue simulation. This work marks a\nfirst step towards learning a neural conversational model based on the\nlong-term success of dialogues.', 'Recent developments have established the vulnerability of deep Reinforcement\nLearning (RL) to policy manipulation attacks via adversarial perturbations. In\nthis paper, we investigate the robustness and resilience of deep RL to\ntraining-time and test-time attacks. Through experimental results, we\ndemonstrate that under noncontiguous training-time attacks, Deep Q-Network\n(DQN) agents can recover and adapt to the adversarial conditions by reactively\nadjusting the policy. Our results also show that policies learned under\nadversarial perturbations are more robust to test-time attacks. Furthermore, we\ncompare the performance of $\\epsilon$-greedy and parameter-space noise\nexploration methods in terms of robustness and resilience against adversarial\nperturbations.', 'The Neural Machine Translation (NMT) model is essentially a joint language\nmodel conditioned on both the source sentence and partial translation.\nTherefore, the NMT model naturally involves the mechanism of the Language Model\n(LM) that predicts the next token only based on partial translation. Despite\nits success, NMT still suffers from the hallucination problem, generating\nfluent but inadequate translations. The main reason is that NMT pays excessive\nattention to the partial translation while neglecting the source sentence to\nsome extent, namely overconfidence of the LM. Accordingly, we define the Margin\nbetween the NMT and the LM, calculated by subtracting the predicted probability\nof the LM from that of the NMT model for each token. The Margin is negatively\ncorrelated to the overconfidence degree of the LM. Based on the property, we\npropose a Margin-based Token-level Objective (MTO) and a Margin-based\nSentencelevel Objective (MSO) to maximize the Margin for preventing the LM from\nbeing overconfident. Experiments on WMT14 English-to-German, WMT19\nChinese-to-English, and WMT14 English-to-French translation tasks demonstrate\nthe effectiveness of our approach, with 1.36, 1.50, and 0.63 BLEU improvements,\nrespectively, compared to the Transformer baseline. The human evaluation\nfurther verifies that our approaches improve translation adequacy as well as\nfluency.', 'Recent research in Visual Question Answering (VQA) has revealed\nstate-of-the-art models to be inconsistent in their understanding of the world\n-- they answer seemingly difficult questions requiring reasoning correctly but\nget simpler associated sub-questions wrong. These sub-questions pertain to\nlower level visual concepts in the image that models ideally should understand\nto be able to answer the higher level question correctly. To address this, we\nfirst present a gradient-based interpretability approach to determine the\nquestions most strongly correlated with the reasoning question on an image, and\nuse this to evaluate VQA models on their ability to identify the relevant\nsub-questions needed to answer a reasoning question. Next, we propose a\ncontrastive gradient learning based approach called Sub-question Oriented\nTuning (SOrT) which encourages models to rank relevant sub-questions higher\nthan irrelevant questions for an <image, reasoning-question> pair. We show that\nSOrT improves model consistency by upto 6.5% points over existing baselines,\nwhile also improving visual grounding.', 'Attention is sparse in vision transformers. We observe the final prediction\nin vision transformers is only based on a subset of most informative tokens,\nwhich is sufficient for accurate image recognition. Based on this observation,\nwe propose a dynamic token sparsification framework to prune redundant tokens\nprogressively and dynamically based on the input. Specifically, we devise a\nlightweight prediction module to estimate the importance score of each token\ngiven the current features. The module is added to different layers to prune\nredundant tokens hierarchically. To optimize the prediction module in an\nend-to-end manner, we propose an attention masking strategy to differentiably\nprune a token by blocking its interactions with other tokens. Benefiting from\nthe nature of self-attention, the unstructured sparse tokens are still hardware\nfriendly, which makes our framework easy to achieve actual speed-up. By\nhierarchically pruning 66% of the input tokens, our method greatly reduces\n31%~37% FLOPs and improves the throughput by over 40% while the drop of\naccuracy is within 0.5% for various vision transformers. Equipped with the\ndynamic token sparsification framework, DynamicViT models can achieve very\ncompetitive complexity/accuracy trade-offs compared to state-of-the-art CNNs\nand vision transformers on ImageNet. Code is available at\nhttps://github.com/raoyongming/DynamicViT', 'To demystify the ""black box"" property of deep neural networks for natural\nlanguage processing (NLP), several methods have been proposed to interpret\ntheir predictions by measuring the change in prediction probability after\nerasing each token of an input. Since existing methods replace each token with\na predefined value (i.e., zero), the resulting sentence lies out of the\ntraining data distribution, yielding misleading interpretations. In this study,\nwe raise the out-of-distribution problem induced by the existing interpretation\nmethods and present a remedy; we propose to marginalize each token out. We\ninterpret various NLP models trained for sentiment analysis and natural\nlanguage inference using the proposed method.', ""Recent deep learning models have shown remarkable performance in image\nclassification. While these deep learning systems are getting closer to\npractical deployment, the common assumption made about data is that it does not\ncarry any sensitive information. This assumption may not hold for many\npractical cases, especially in the domain where an individual's personal\ninformation is involved, like healthcare and facial recognition systems. We\nposit that selectively removing features in this latent space can protect the\nsensitive information and provide a better privacy-utility trade-off.\nConsequently, we propose DISCO which learns a dynamic and data driven pruning\nfilter to selectively obfuscate sensitive information in the feature space. We\npropose diverse attack schemes for sensitive inputs \\& attributes and\ndemonstrate the effectiveness of DISCO against state-of-the-art methods through\nquantitative and qualitative evaluation. Finally, we also release an evaluation\nbenchmark dataset of 1 million sensitive representations to encourage rigorous\nexploration of novel attack schemes."", 'We present two categories of model-agnostic adversarial strategies that\nreveal the weaknesses of several generative, task-oriented dialogue models:\nShould-Not-Change strategies that evaluate over-sensitivity to small and\nsemantics-preserving edits, as well as Should-Change strategies that test if a\nmodel is over-stable against subtle yet semantics-changing modifications. We\nnext perform adversarial training with each strategy, employing a max-margin\napproach for negative generative examples. This not only makes the target\ndialogue model more robust to the adversarial inputs, but also helps it perform\nsignificantly better on the original inputs. Moreover, training on all\nstrategies combined achieves further improvements, achieving a new\nstate-of-the-art performance on the original task (also verified via human\nevaluation). In addition to adversarial training, we also address the\nrobustness task at the model-level, by feeding it subword units as both inputs\nand outputs, and show that the resulting model is equally competitive, requires\nonly 1/4 of the original vocabulary size, and is robust to one of the\nadversarial strategies (to which the original model is vulnerable) even without\nadversarial training.', 'In complex processes, various events can happen in different sequences. The\nprediction of the next event given an a-priori process state is of importance\nin such processes. Recent methods have proposed deep learning techniques such\nas recurrent neural networks, developed on raw event logs, to predict the next\nevent from a process state. However, such deep learning models by themselves\nlack a clear representation of the process states. At the same time, recent\nmethods have neglected the time feature of event instances. In this paper, we\ntake advantage of Petri nets as a powerful tool in modeling complex process\nbehaviors considering time as an elemental variable. We propose an approach\nwhich starts from a Petri net process model constructed by a process mining\nalgorithm. We enhance the Petri net model with time decay functions to create\ncontinuous process state samples. Finally, we use these samples in combination\nwith discrete token movement counters and Petri net markings to train a deep\nlearning model that predicts the next event. We demonstrate significant\nperformance improvements and outperform the state-of-the-art methods on nine\nreal-world benchmark event logs.', 'Neural language models predict the next token using a latent representation\nof the immediate token history. Recently, various methods for augmenting neural\nlanguage models with an attention mechanism over a differentiable memory have\nbeen proposed. For predicting the next token, these models query information\nfrom a memory of the recent history which can facilitate learning mid- and\nlong-range dependencies. However, conventional attention mechanisms used in\nmemory-augmented neural language models produce a single output vector per time\nstep. This vector is used both for predicting the next token as well as for the\nkey and value of a differentiable memory of a token history. In this paper, we\npropose a neural language model with a key-value attention mechanism that\noutputs separate representations for the key and value of a differentiable\nmemory, as well as for encoding the next-word distribution. This model\noutperforms existing memory-augmented neural language models on two corpora.\nYet, we found that our method mainly utilizes a memory of the five most recent\noutput representations. This led to the unexpected main finding that a much\nsimpler model based only on the concatenation of recent output representations\nfrom previous time steps is on par with more sophisticated memory-augmented\nneural language models.', 'In hybrid human-AI systems, users need to decide whether or not to trust an\nalgorithmic prediction while the true error in the prediction is unknown. To\naccommodate such settings, we introduce RETRO-VIZ, a method for (i) estimating\nand (ii) explaining trustworthiness of regression predictions. It consists of\nRETRO, a quantitative estimate of the trustworthiness of a prediction, and VIZ,\na visual explanation that helps users identify the reasons for the (lack of)\ntrustworthiness of a prediction. We find that RETRO-scores negatively correlate\nwith prediction error across 117 experimental settings, indicating that RETRO\nprovides a useful measure to distinguish trustworthy predictions from\nuntrustworthy ones. In a user study with 41 participants, we find that\nVIZ-explanations help users identify whether a prediction is trustworthy or\nnot: on average, 95.1% of participants correctly select the more trustworthy\nprediction, given a pair of predictions. In addition, an average of 75.6% of\nparticipants can accurately describe why a prediction seems to be (not)\ntrustworthy. Finally, we find that the vast majority of users subjectively\nexperience RETRO-VIZ as a useful tool to assess the trustworthiness of\nalgorithmic predictions.', 'Despite widespread adoption, machine learning models remain mostly black\nboxes. Understanding the reasons behind predictions is, however, quite\nimportant in assessing trust, which is fundamental if one plans to take action\nbased on a prediction, or when choosing whether to deploy a new model. Such\nunderstanding also provides insights into the model, which can be used to\ntransform an untrustworthy model or prediction into a trustworthy one. In this\nwork, we propose LIME, a novel explanation technique that explains the\npredictions of any classifier in an interpretable and faithful manner, by\nlearning an interpretable model locally around the prediction. We also propose\na method to explain models by presenting representative individual predictions\nand their explanations in a non-redundant way, framing the task as a submodular\noptimization problem. We demonstrate the flexibility of these methods by\nexplaining different models for text (e.g. random forests) and image\nclassification (e.g. neural networks). We show the utility of explanations via\nnovel experiments, both simulated and with human subjects, on various scenarios\nthat require trust: deciding if one should trust a prediction, choosing between\nmodels, improving an untrustworthy classifier, and identifying why a classifier\nshould not be trusted.', 'There have been many advances in the artificial intelligence field due to the\nemergence of deep learning. In almost all sub-fields, artificial neural\nnetworks have reached or exceeded human-level performance. However, most of the\nmodels are not interpretable. As a result, it is hard to trust their decisions,\nespecially in life and death scenarios. In recent years, there has been a\nmovement toward creating explainable artificial intelligence, but most work to\ndate has concentrated on image processing models, as it is easier for humans to\nperceive visual patterns. There has been little work in other fields like\nnatural language processing. In this paper, we train a convolutional model on\ntextual data and analyze the global logic of the model by studying its filter\nvalues. In the end, we find the most important words in our corpus to our\nmodels logic and remove the rest (95%). New models trained on just the 5% most\nimportant words can achieve the same performance as the original model while\nreducing training time by more than half. Approaches such as this will help us\nto understand NLP models, explain their decisions according to their word\nchoices, and improve them by finding blind spots and biases.', 'Autonomous AI systems will be entering human society in the near future to\nprovide services and work alongside humans. For those systems to be accepted\nand trusted, the users should be able to understand the reasoning process of\nthe system, i.e. the system should be transparent. System transparency enables\nhumans to form coherent explanations of the system\'s decisions and actions.\nTransparency is important not only for user trust, but also for software\ndebugging and certification. In recent years, Deep Neural Networks have made\ngreat advances in multiple application areas. However, deep neural networks are\nopaque. In this paper, we report on work in transparency in Deep Reinforcement\nLearning Networks (DRLN). Such networks have been extremely successful in\naccurately learning action control in image input domains, such as Atari games.\nIn this paper, we propose a novel and general method that (a) incorporates\nexplicit object recognition processing into deep reinforcement learning models,\n(b) forms the basis for the development of ""object saliency maps"", to provide\nvisualization of internal states of DRLNs, thus enabling the formation of\nexplanations and (c) can be incorporated in any existing deep reinforcement\nlearning framework. We present computational results and human experiments to\nevaluate our approach.', 'Human mobility prediction is a core functionality in many location-based\nservices and applications. However, due to the sparsity of mobility data, it is\nnot an easy task to predict future POIs (place-of-interests) that are going to\nbe visited. In this paper, we propose MobTCast, a Transformer-based\ncontext-aware network for mobility prediction. Specifically, we explore the\ninfluence of four types of context in the mobility prediction: temporal,\nsemantic, social and geographical contexts. We first design a base mobility\nfeature extractor using the Transformer architecture, which takes both the\nhistory POI sequence and the semantic information as input. It handles both the\ntemporal and semantic contexts. Based on the base extractor and the social\nconnections of a user, we employ a self-attention module to model the influence\nof the social context. Furthermore, unlike existing methods, we introduce a\nlocation prediction branch in MobTCast as an auxiliary task to model the\ngeographical context and predict the next location. Intuitively, the\ngeographical distance between the location of the predicted POI and the\npredicted location from the auxiliary branch should be as close as possible. To\nreflect this relation, we design a consistency loss to further improve the POI\nprediction performance. In our experimental results, MobTCast outperforms other\nstate-of-the-art next POI prediction methods. Our approach illustrates the\nvalue of including different types of context in next POI prediction.', 'Attention-based encoder-decoder (AED) models have achieved promising\nperformance in speech recognition. However, because the decoder predicts text\ntokens (such as characters or words) in an autoregressive manner, it is\ndifficult for an AED model to predict all tokens in parallel. This makes the\ninference speed relatively slow. We believe that because the encoder already\ncaptures the whole speech utterance, which has the token-level relationship\nimplicitly, we can predict a token without explicitly autoregressive language\nmodeling. When the prediction of a token does not rely on other tokens, the\nparallel prediction of all tokens in the sequence is realizable. Based on this\nidea, we propose a non-autoregressive speech recognition model called LASO\n(Listen Attentively, and Spell Once). The model consists of an encoder, a\ndecoder, and a position dependent summarizer (PDS). The three modules are based\non basic attention blocks. The encoder extracts high-level representations from\nthe speech. The PDS uses positional encodings corresponding to tokens to\nconvert the acoustic representations into token-level representations. The\ndecoder further captures token-level relationships with the self-attention\nmechanism. At last, the probability distribution on the vocabulary is computed\nfor each token position. Therefore, speech recognition is re-formulated as a\nposition-wise classification problem. Further, we propose a cross-modal\ntransfer learning method to refine semantics from a large-scale pre-trained\nlanguage model BERT for improving the performance.', ""Personal assistant AI systems such as Siri, Cortana, and Alexa have become\nwidely used as a means to accomplish tasks through natural language commands.\nHowever, components in these systems generally rely on supervised machine\nlearning algorithms that require large amounts of hand-annotated training data,\nwhich is expensive and time consuming to collect. The ability to incorporate\nunsupervised, weakly supervised, or distantly supervised data holds significant\npromise in overcoming this bottleneck. In this paper, we describe a framework\nthat leverages user engagement signals (user behaviors that demonstrate a\npositive or negative response to content) to automatically create granular\nentity labels for training data augmentation. Strategies such as multi-task\nlearning and validation using an external knowledge base are employed to\nincorporate the engagement annotated data and to boost the model's accuracy on\na sequence labeling task. Our results show that learning from data\nautomatically labeled by user engagement signals achieves significant accuracy\ngains in a production deep learning system, when measured on both the sequence\nlabeling task as well as on user facing results produced by the system\nend-to-end. We believe this is the first use of user engagement signals to help\ngenerate training data for a sequence labeling task on a large scale, and can\nbe applied in practical settings to speed up new feature deployment when little\nhuman annotated data is available."", ""Recently, large pre-trained neural language models have attained remarkable\nperformance on many downstream natural language processing (NLP) applications\nvia fine-tuning. In this paper, we target at how to further improve the token\nrepresentations on the language models. We, therefore, propose a simple yet\neffective plug-and-play module, Sequential Attention Module (SAM), on the token\nembeddings learned from a pre-trained language model. Our proposed SAM consists\nof two main attention modules deployed sequentially: Feature-wise Attention\nModule (FAM) and Token-wise Attention Module (TAM). More specifically, FAM can\neffectively identify the importance of features at each dimension and promote\nthe effect via dot-product on the original token embeddings for downstream NLP\napplications. Meanwhile, TAM can further re-weight the features at the\ntoken-wise level. Moreover, we propose an adaptive filter on FAM to prevent\nnoise impact and increase information absorption. Finally, we conduct extensive\nexperiments to demonstrate the advantages and properties of our proposed SAM.\nWe first show how SAM plays a primary role in the champion solution of two\nsubtasks of SemEval'21 Task 7. After that, we apply SAM on sentiment analysis\nand three popular NLP tasks and demonstrate that SAM consistently outperforms\nthe state-of-the-art baselines."", ""Masked language modeling (MLM), a self-supervised pretraining objective, is\nwidely used in natural language processing for learning text representations.\nMLM trains a model to predict a random sample of input tokens that have been\nreplaced by a [MASK] placeholder in a multi-class setting over the entire\nvocabulary. When pretraining, it is common to use alongside MLM other auxiliary\nobjectives on the token or sequence level to improve downstream performance\n(e.g. next sentence prediction). However, no previous work so far has attempted\nin examining whether other simpler linguistically intuitive or not objectives\ncan be used standalone as main pretraining objectives. In this paper, we\nexplore five simple pretraining objectives based on token-level classification\ntasks as replacements of MLM. Empirical results on GLUE and SQuAD show that our\nproposed methods achieve comparable or better performance to MLM using a\nBERT-BASE architecture. We further validate our methods using smaller models,\nshowing that pretraining a model with 41% of the BERT-BASE's parameters,\nBERT-MEDIUM results in only a 1% drop in GLUE scores with our best objective."", 'We introduce a novel approach to transformers that learns hierarchical\nrepresentations in multiparty dialogue. First, three language modeling tasks\nare used to pre-train the transformers, token- and utterance-level language\nmodeling and utterance order prediction, that learn both token and utterance\nembeddings for better understanding in dialogue contexts. Then, multi-task\nlearning between the utterance prediction and the token span prediction is\napplied to fine-tune for span-based question answering (QA). Our approach is\nevaluated on the FriendsQA dataset and shows improvements of 3.8% and 1.4% over\nthe two state-of-the-art transformer models, BERT and RoBERTa, respectively.', ""This paper describes problems in AI research and how the SP System (described\nin an appendix) may help to solve them. Most of the problems are described by\nleading researchers in AI in interviews with science writer Martin Ford, and\nreported by him in his book {\\em Architects of Intelligence}. These problems\nare: the need to bridge the divide between symbolic and non-symbolic kinds of\nknowledge and processing; the tendency of deep neural networks (DNNs) to make\nlarge and unexpected errors in recognition; the need to strengthen the\nrepresentation and processing of natural languages; the challenges of\nunsupervised learning; the need for a coherent account of generalisation; how\nto learn usable knowledge from a single exposure; how to achieve transfer\nlearning; how to increase the efficiency of AI processing; the need for\ntransparency in AI structures and processes; how to achieve varieties of\nprobabilistic reasoning; the need for more emphasis on top-down strategies; how\nto minimise the risk of accidents with self-driving vehicles; the need for\nstrong compositionality in AI knowledge; the challenges of commonsense\nreasoning and commonsense knowledge; establishing the importance of information\ncompression in AI research; establishing the importance of a biological\nperspective in AI research; establishing whether knowledge in the brain is\nrepresented in `distributed' or `localist' form; how to bypassing the limited\nscope for adaptation in deep neural networks; the need to develop `broad AI';\nand how to eliminate the problem of catastrophic forgetting."", ""Traffic state prediction is necessary for many Intelligent Transportation\nSystems applications. Recent developments of the topic have focused on\nnetwork-wide, multi-step prediction, where state of the art performance is\nachieved via deep learning models, in particular, graph neural network-based\nmodels. While the prediction accuracy of deep learning models is high, these\nmodels' robustness has raised many safety concerns, given that imperceptible\nperturbations added to input can substantially degrade the model performance.\nIn this work, we propose an adversarial attack framework by treating the\nprediction model as a black-box, i.e., assuming no knowledge of the model\narchitecture, training data, and (hyper)parameters. However, we assume that the\nadversary can oracle the prediction model with any input and obtain\ncorresponding output. Next, the adversary can train a substitute model using\ninput-output pairs and generate adversarial signals based on the substitute\nmodel. To test the attack effectiveness, two state of the art, graph neural\nnetwork-based models (GCGRNN and DCRNN) are examined. As a result, the\nadversary can degrade the target model's prediction accuracy up to $54\\%$. In\ncomparison, two conventional statistical models (linear regression and\nhistorical average) are also examined. While these two models do not produce\nhigh prediction accuracy, they are either influenced negligibly (less than\n$3\\%$) or are immune to the adversary's attack."", 'The increased adoption of Artificial Intelligence (AI) presents an\nopportunity to solve many socio-economic and environmental challenges; however,\nthis cannot happen without securing AI-enabled technologies. In recent years,\nmost AI models are vulnerable to advanced and sophisticated hacking techniques.\nThis challenge has motivated concerted research efforts into adversarial AI,\nwith the aim of developing robust machine and deep learning models that are\nresilient to different types of adversarial scenarios. In this paper, we\npresent a holistic cyber security review that demonstrates adversarial attacks\nagainst AI applications, including aspects such as adversarial knowledge and\ncapabilities, as well as existing methods for generating adversarial examples\nand existing cyber defence models. We explain mathematical AI models,\nespecially new variants of reinforcement and federated learning, to demonstrate\nhow attack vectors would exploit vulnerabilities of AI models. We also propose\na systematic framework for demonstrating attack techniques against AI\napplications and reviewed several cyber defences that would protect AI\napplications against those attacks. We also highlight the importance of\nunderstanding the adversarial goals and their capabilities, especially the\nrecent attacks against industry applications, to develop adaptive defences that\nassess to secure AI applications. Finally, we describe the main challenges and\nfuture research directions in the domain of security and privacy of AI\ntechnologies.', 'Video captioning is a challenging task that captures different visual parts\nand describes them in sentences, for it requires visual and linguistic\ncoherence. The attention mechanism in the current video captioning method\nlearns to assign weight to each frame, promoting the decoder dynamically. This\nmay not explicitly model the correlation and the temporal coherence of the\nvisual features extracted in the sequence frames.To generate semantically\ncoherent sentences, we propose a new Visual-aware Attention (VA) model, which\nconcatenates dynamic changes of temporal sequence frames with the words at the\nprevious moment, as the input of attention mechanism to extract sequence\nfeatures.In addition, the prevalent approaches widely use the teacher-forcing\n(TF) learning during training, where the next token is generated conditioned on\nthe previous ground-truth tokens. The semantic information in the previously\ngenerated tokens is lost. Therefore, we design a self-forcing (SF) stream that\ntakes the semantic information in the probability distribution of the previous\ntoken as input to enhance the current token.The Dual-stream Decoder (DD)\narchitecture unifies the TF and SF streams, generating sentences to promote the\nannotated captioning for both streams.Meanwhile, with the Dual-stream Decoder\nutilized, the exposure bias problem is alleviated, caused by the discrepancy\nbetween the training and testing in the TF learning.The effectiveness of the\nproposed Visual-aware Attention Dual-stream Decoder (VADD) is demonstrated\nthrough the result of experimental studies on Microsoft video description\n(MSVD) corpus and MSR-Video to text (MSR-VTT) datasets.', 'We develop a chatbot using Deep Bidirectional Transformer models (BERT) to\nhandle client questions in financial investment customer service. The bot can\nrecognize 381 intents, and decides when to say ""I don\'t know"" and escalates\nirrelevant/uncertain questions to human operators. Our main novel contribution\nis the discussion about uncertainty measure for BERT, where three different\napproaches are systematically compared on real problems. We investigated two\nuncertainty metrics, information entropy and variance of dropout sampling in\nBERT, followed by mixed-integer programming to optimize decision thresholds.\nAnother novel contribution is the usage of BERT as a language model in\nautomatic spelling correction. Inputs with accidental spelling errors can\nsignificantly decrease intent classification performance. The proposed approach\ncombines probabilities from masked language model and word edit distances to\nfind the best corrections for misspelled words. The chatbot and the entire\nconversational AI system are developed using open-source tools, and deployed\nwithin our company\'s intranet. The proposed approach can be useful for\nindustries seeking similar in-house solutions in their specific business\ndomains. We share all our code and a sample chatbot built on a public dataset\non Github.', ""While improving neural dialogue agents' factual accuracy is the object of\nmuch research, another important aspect of communication, less studied in the\nsetting of neural dialogue, is transparency about ignorance. In this work, we\nanalyze to what extent state-of-the-art chit-chat models are linguistically\ncalibrated in the sense that their verbalized expression of doubt (or\nconfidence) matches the likelihood that the model's responses are factually\nincorrect (or correct). We find that these models are poorly calibrated, yet we\nshow that likelihood of correctness can accurately be predicted. By\nincorporating such metacognitive features into the training of a controllable\ngeneration model, we obtain a dialogue agent with greatly improved linguistic\ncalibration. While improving neural dialogue agents' factual accuracy is the\nobject of much research, another important aspect of communication, less\nstudied in the setting of neural dialogue, is transparency about ignorance. In\nthis work, we analyze to what extent state-of-the-art chit-chat models are\nlinguistically calibrated in the sense that their verbalized expression of\ndoubt (or confidence) matches the likelihood that the model's responses are\nfactually incorrect (or correct). We find that these models are poorly\ncalibrated, yet we show that likelihood of correctness can accurately be\npredicted. By incorporating such metacognitive features into the training of a\ncontrollable generation model, we obtain a dialogue agent with greatly improved\nlinguistic calibration."", 'Today, most large-scale conversational AI agents (e.g. Alexa, Siri, or Google\nAssistant) are built using manually annotated data to train the different\ncomponents of the system. Typically, the accuracy of the ML models in these\ncomponents are improved by manually transcribing and annotating data. As the\nscope of these systems increase to cover more scenarios and domains, manual\nannotation to improve the accuracy of these components becomes prohibitively\ncostly and time consuming. In this paper, we propose a system that leverages\nuser-system interaction feedback signals to automate learning without any\nmanual annotation. Users here tend to modify a previous query in hopes of\nfixing an error in the previous turn to get the right results. These\nreformulations, which are often preceded by defective experiences caused by\nerrors in ASR, NLU, ER or the application. In some cases, users may not\nproperly formulate their requests (e.g. providing partial title of a song), but\ngleaning across a wider pool of users and sessions reveals the underlying\nrecurrent patterns. Our proposed self-learning system automatically detects the\nerrors, generate reformulations and deploys fixes to the runtime system to\ncorrect different types of errors occurring in different components of the\nsystem. In particular, we propose leveraging an absorbing Markov Chain model as\na collaborative filtering mechanism in a novel attempt to mine these patterns.\nWe show that our approach is highly scalable, and able to learn reformulations\nthat reduce Alexa-user errors by pooling anonymized data across millions of\ncustomers. The proposed self-learning system achieves a win/loss ratio of 11.8\nand effectively reduces the defect rate by more than 30% on utterance level\nreformulations in our production A/B tests. To the best of our knowledge, this\nis the first self-learning large-scale conversational AI system in production.', 'Our project task was to create a model that, given a speaker ID, chat\nhistory, and an utterance query, can predict the response utterance in a\nconversation. The model is personalized for each speaker. This task can be a\nuseful tool for building speech bots that talk in a human-like manner in a live\nconversation. Further, we succeeded at using dense-vector encoding clustering\nto be able to retrieve relevant historical dialogue context, a useful strategy\nfor overcoming the input limitations of neural-based models when predictions\nrequire longer-term references from the dialogue history. In this paper, we\nhave implemented a state-of-the-art model using pre-training and fine-tuning\ntechniques built on transformer architecture and multi-headed attention blocks\nfor the Switchboard corpus. We also show how efficient vector clustering\nalgorithms can be used for real-time utterance predictions that require no\ntraining and therefore work on offline and encrypted message histories.', 'Current conversational systems can follow simple commands and answer basic\nquestions, but they have difficulty maintaining coherent and open-ended\nconversations about specific topics. Competitions like the Conversational\nIntelligence (ConvAI) challenge are being organized to push the research\ndevelopment towards that goal. This article presents in detail the RLLChatbot\nthat participated in the 2017 ConvAI challenge. The goal of this research is to\nbetter understand how current deep learning and reinforcement learning tools\ncan be used to build a robust yet flexible open domain conversational agent. We\nprovide a thorough description of how a dialog system can be built and trained\nfrom mostly public-domain datasets using an ensemble model. The first\ncontribution of this work is a detailed description and analysis of different\ntext generation models in addition to novel message ranking and selection\nmethods. Moreover, a new open-source conversational dataset is presented.\nTraining on this data significantly improves the Recall@k score of the ranking\nand selection mechanisms compared to our baseline model responsible for\nselecting the message returned at each interaction.', ""Sentence order prediction is the task of finding the correct order of\nsentences in a randomly ordered document. Correctly ordering the sentences\nrequires an understanding of coherence with respect to the chronological\nsequence of events described in the text. Document-level contextual\nunderstanding and commonsense knowledge centered around these events are often\nessential in uncovering this coherence and predicting the exact chronological\norder. In this paper, we introduce STaCK -- a framework based on graph neural\nnetworks and temporal commonsense knowledge to model global information and\npredict the relative order of sentences. Our graph network accumulates temporal\nevidence using knowledge of `past' and `future' and formulates sentence\nordering as a constrained edge classification problem. We report results on\nfive different datasets, and empirically show that the proposed method is\nnaturally suitable for order prediction. The implementation of this work is\npublicly available at: https://github.com/declare-lab/sentence-ordering."", 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'Highly regularized LSTMs achieve impressive results on several benchmark\ndatasets in language modeling. We propose a new regularization method based on\ndecoding the last token in the context using the predicted distribution of the\nnext token. This biases the model towards retaining more contextual\ninformation, in turn improving its ability to predict the next token. With\nnegligible overhead in the number of parameters and training time, our Past\nDecode Regularization (PDR) method achieves a word level perplexity of 55.6 on\nthe Penn Treebank and 63.5 on the WikiText-2 datasets using a single softmax.\nWe also show gains by using PDR in combination with a mixture-of-softmaxes,\nachieving a word level perplexity of 53.8 and 60.5 on these datasets. In\naddition, our method achieves 1.169 bits-per-character on the Penn Treebank\nCharacter dataset for character level language modeling. These results\nconstitute a new state-of-the-art in their respective settings.', 'Token embeddings in multilingual BERT (m-BERT) contain both language and\nsemantic information. We find that the representation of a language can be\nobtained by simply averaging the embeddings of the tokens of the language.\nGiven this language representation, we control the output languages of\nmultilingual BERT by manipulating the token embeddings, thus achieving\nunsupervised token translation. We further propose a computationally cheap but\neffective approach to improve the cross-lingual ability of m-BERT based on this\nobservation.', ""We propose a new model, DocHopper, that iteratively attends to different\nparts of long, hierarchically structured documents to answer complex questions.\nSimilar to multi-hop question-answering (QA) systems, at each step, DocHopper\nuses a query $q$ to attend to information from a document, combines this\n``retrieved'' information with $q$ to produce the next query. However, in\ncontrast to most previous multi-hop QA systems, DocHopper is able to\n``retrieve'' either short passages or long sections of the document, thus\nemulating a multi-step process of ``navigating'' through a long document to\nanswer a question. To enable this novel behavior, DocHopper does not combine\ndocument information with $q$ by concatenating text to the text of $q$, but by\ncombining a compact neural representation of $q$ with a compact neural\nrepresentation of a hierarchical part of the document, which can potentially be\nquite large. We experiment with DocHopper on four different QA tasks that\nrequire reading long and complex documents to answer multi-hop questions, and\nshow that DocHopper achieves state-of-the-art results on three of the datasets.\nAdditionally, DocHopper is efficient at inference time, being 3--10 times\nfaster than the baselines."", ""The rapid development of Artificial Intelligence (AI) technology has enabled\nthe deployment of various systems based on it. However, many current AI systems\nare found vulnerable to imperceptible attacks, biased against underrepresented\ngroups, lacking in user privacy protection. These shortcomings degrade user\nexperience and erode people's trust in all AI systems. In this review, we\nprovide AI practitioners with a comprehensive guide for building trustworthy AI\nsystems. We first introduce the theoretical framework of important aspects of\nAI trustworthiness, including robustness, generalization, explainability,\ntransparency, reproducibility, fairness, privacy preservation, and\naccountability. To unify currently available but fragmented approaches toward\ntrustworthy AI, we organize them in a systematic approach that considers the\nentire lifecycle of AI systems, ranging from data acquisition to model\ndevelopment, to system development and deployment, finally to continuous\nmonitoring and governance. In this framework, we offer concrete action items\nfor practitioners and societal stakeholders (e.g., researchers, engineers, and\nregulators) to improve AI trustworthiness. Finally, we identify key\nopportunities and challenges for the future development of trustworthy AI\nsystems, where we identify the need for a paradigm shift toward comprehensively\ntrustworthy AI systems."", 'Recent advances in deep neural networks, language modeling and language\ngeneration have introduced new ideas to the field of conversational agents. As\na result, deep neural models such as sequence-to-sequence, Memory Networks, and\nthe Transformer have become key ingredients of state-of-the-art dialog systems.\nWhile those models are able to generate meaningful responses even in unseen\nsituation, they need a lot of training data to build a reliable model. Thus,\nmost real-world systems stuck to traditional approaches based on information\nretrieval and even hand-crafted rules, due to their robustness and\neffectiveness, especially for narrow-focused conversations. Here, we present a\nmethod that adapts a deep neural architecture from the domain of machine\nreading comprehension to re-rank the suggested answers from different models\nusing the question as context. We train our model using negative sampling based\non question-answer pairs from the Twitter Customer Support Dataset.The\nexperimental results show that our re-ranking framework can improve the\nperformance in terms of word overlap and semantics both for individual models\nas well as for model combinations.', ""The ability of a machine to communicate with humans has long been associated\nwith the general success of AI. This dates back to Alan Turing's epoch-making\nwork in the early 1950s, which proposes that a machine's intelligence can be\ntested by how well it, the machine, can fool a human into believing that the\nmachine is a human through dialogue conversations. Many systems learn\ngeneration rules from a minimal set of authored rules or labels on top of\nhand-coded rules or templates, and thus are both expensive and difficult to\nextend to open-domain scenarios. Recently, the emergence of neural network\nmodels the potential to solve many of the problems in dialogue learning that\nearlier systems cannot tackle: the end-to-end neural frameworks offer the\npromise of scalability and language-independence, together with the ability to\ntrack the dialogue state and then mapping between states and dialogue actions\nin a way not possible with conventional systems. On the other hand, neural\nsystems bring about new challenges: they tend to output dull and generic\nresponses; they lack a consistent or a coherent persona; they are usually\noptimized through single-turn conversations and are incapable of handling the\nlong-term success of a conversation; and they are not able to take the\nadvantage of the interactions with humans. This dissertation attempts to tackle\nthese challenges: Contributions are two-fold: (1) we address new challenges\npresented by neural network models in open-domain dialogue generation systems;\n(2) we develop interactive question-answering dialogue systems by (a) giving\nthe agent the ability to ask questions and (b) training a conversation agent\nthrough interactions with humans in an online fashion, where a bot improves\nthrough communicating with humans and learning from the mistakes that it makes."", 'The encoder-decoder framework achieves state-of-the-art results in keyphrase\ngeneration (KG) tasks by predicting both present keyphrases that appear in the\nsource document and absent keyphrases that do not. However, relying solely on\nthe source document can result in generating uncontrollable and inaccurate\nabsent keyphrases. To address these problems, we propose a novel graph-based\nmethod that can capture explicit knowledge from related references. Our model\nfirst retrieves some document-keyphrases pairs similar to the source document\nfrom a pre-defined index as references. Then a heterogeneous graph is\nconstructed to capture relationships of different granularities between the\nsource document and its references. To guide the decoding process, a\nhierarchical attention and copy mechanism is introduced, which directly copies\nappropriate words from both the source document and its references based on\ntheir relevance and significance. The experimental results on multiple KG\nbenchmarks show that the proposed model achieves significant improvements\nagainst other baseline models, especially with regard to the absent keyphrase\nprediction.', 'In this work, we present the Chatbot Interaction with Artificial Intelligence\n(CI-AI) framework as an approach to the training of deep learning chatbots for\ntask classification. The intelligent system augments human-sourced data via\nartificial paraphrasing in order to generate a large set of training data for\nfurther classical, attention, and language transformation-based learning\napproaches for Natural Language Processing. Human beings are asked to\nparaphrase commands and questions for task identification for further execution\nof a machine. The commands and questions are split into training and validation\nsets. A total of 483 responses were recorded. Secondly, the training set is\nparaphrased by the T5 model in order to augment it with further data. Seven\nstate-of-the-art transformer-based text classification algorithms (BERT,\nDistilBERT, RoBERTa, DistilRoBERTa, XLM, XLM-RoBERTa, and XLNet) are\nbenchmarked for both sets after fine-tuning on the training data for two\nepochs. We find that all models are improved when training data is augmented by\nthe T5 model, with an average increase of classification accuracy by 4.01%. The\nbest result was the RoBERTa model trained on T5 augmented data which achieved\n98.96% classification accuracy. Finally, we found that an ensemble of the five\nbest-performing transformer models via Logistic Regression of output label\npredictions led to an accuracy of 99.59% on the dataset of human responses. A\nhighly-performing model allows the intelligent system to interpret human\ncommands at the social-interaction level through a chatbot-like interface (e.g.\n""Robot, can we have a conversation?"") and allows for better accessibility to AI\nby non-technical users.', 'Artificial Intelligence (AI) is increasingly used in critical applications.\nThus, the need for dependable AI systems is rapidly growing. In 2018, the\nEuropean Commission appointed experts to a High-Level Expert Group on AI\n(AI-HLEG). AI-HLEG defined Trustworthy AI as 1) lawful, 2) ethical, and 3)\nrobust and specified seven corresponding key requirements. To help development\norganizations, AI-HLEG recently published the Assessment List for Trustworthy\nAI (ALTAI). We present an illustrative case study from applying ALTAI to an\nongoing development project of an Advanced Driver-Assistance System (ADAS) that\nrelies on Machine Learning (ML). Our experience shows that ALTAI is largely\napplicable to ADAS development, but specific parts related to human agency and\ntransparency can be disregarded. Moreover, bigger questions related to societal\nand environmental impact cannot be tackled by an ADAS supplier in isolation. We\npresent how we plan to develop the ADAS to ensure ALTAI-compliance. Finally, we\nprovide three recommendations for the next revision of ALTAI, i.e., life-cycle\nvariants, domain-specific adaptations, and removed redundancy.', ""Despite the widespread use of Knowledge Graph Embeddings (KGE), little is\nknown about the security vulnerabilities that might disrupt their intended\nbehaviour. We study data poisoning attacks against KGE models for link\nprediction. These attacks craft adversarial additions or deletions at training\ntime to cause model failure at test time. To select adversarial deletions, we\npropose to use the model-agnostic instance attribution methods from\nInterpretable Machine Learning, which identify the training instances that are\nmost influential to a neural model's predictions on test instances. We use\nthese influential triples as adversarial deletions. We further propose a\nheuristic method to replace one of the two entities in each influential triple\nto generate adversarial additions. Our experiments show that the proposed\nstrategies outperform the state-of-art data poisoning attacks on KGE models and\nimprove the MRR degradation due to the attacks by up to 62% over the baselines."", 'Cross-lingual document representations enable language understanding in\nmultilingual contexts and allow transfer learning from high-resource to\nlow-resource languages at the document level. Recently large pre-trained\nlanguage models such as BERT, XLM and XLM-RoBERTa have achieved great success\nwhen fine-tuned on sentence-level downstream tasks. It is tempting to apply\nthese cross-lingual models to document representation learning. However, there\nare two challenges: (1) these models impose high costs on long document\nprocessing and thus many of them have strict length limit; (2) model\nfine-tuning requires extra data and computational resources, which is not\npractical in resource-limited settings. In this work, we address these\nchallenges by proposing unsupervised Language-Agnostic Weighted Document\nRepresentations (LAWDR). We study the geometry of pre-trained sentence\nembeddings and leverage it to derive document representations without\nfine-tuning. Evaluated on cross-lingual document alignment, LAWDR demonstrates\ncomparable performance to state-of-the-art models on benchmark datasets.', 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', 'Explaining the predictions of AI models is paramount in safety-critical\napplications, such as in legal or medical domains. One form of explanation for\na prediction is an extractive rationale, i.e., a subset of features of an\ninstance that lead the model to give its prediction on the instance. Previous\nworks on generating extractive rationales usually employ a two-phase model: a\nselector that selects the most important features (i.e., the rationale)\nfollowed by a predictor that makes the prediction based exclusively on the\nselected features. One disadvantage of these works is that the main signal for\nlearning to select features comes from the comparison of the answers given by\nthe predictor and the ground-truth answers. In this work, we propose to squeeze\nmore information from the predictor via an information calibration method. More\nprecisely, we train two models jointly: one is a typical neural model that\nsolves the task at hand in an accurate but black-box manner, and the other is a\nselector-predictor model that additionally produces a rationale for its\nprediction. The first model is used as a guide to the second model. We use an\nadversarial-based technique to calibrate the information extracted by the two\nmodels such that the difference between them is an indicator of the missed or\nover-selected features. In addition, for natural language tasks, we propose to\nuse a language-model-based regularizer to encourage the extraction of fluent\nrationales. Experimental results on a sentiment analysis task as well as on\nthree tasks from the legal domain show the effectiveness of our approach to\nrationale extraction.', 'Models are used in both Software Engineering (SE) and Artificial Intelligence\n(AI). SE models may specify the architecture at different levels of abstraction\nand for addressing different concerns at various stages of the software\ndevelopment life-cycle, from early conceptualization and design, to\nverification, implementation, testing and evolution. However, AI models may\nprovide smart capabilities, such as prediction and decision-making support. For\ninstance, in Machine Learning (ML), which is currently the most popular\nsub-discipline of AI, mathematical models may learn useful patterns in the\nobserved data and can become capable of making predictions. The goal of this\nwork is to create synergy by bringing models in the said communities together\nand proposing a holistic approach to model-driven software development for\nintelligent systems that require ML. We illustrate how software models can\nbecome capable of creating and dealing with ML models in a seamless manner. The\nmain focus is on the domain of the Internet of Things (IoT), where both ML and\nmodel-driven SE play a key role. In the context of the need to take a\nCyber-Physical System-of-Systems perspective of the targeted architecture, an\nintegrated design environment for both SE and ML sub-systems would best support\nthe optimization and overall efficiency of the implementation of the resulting\nsystem. In particular, we implement the proposed approach, called ML-Quadrat,\nbased on ThingML, and validate it using a case study from the IoT domain, as\nwell as through an empirical user evaluation. It transpires that the proposed\napproach is not only feasible, but may also contribute to the performance leap\nof software development for smart Cyber-Physical Systems (CPS) which are\nconnected to the IoT, as well as an enhanced user experience of the\npractitioners who use the proposed modeling solution.', 'This paper summarizes our entries to both subtasks of the first DialDoc\nshared task which focuses on the agent response prediction task in\ngoal-oriented document-grounded dialogs. The task is split into two subtasks:\npredicting a span in a document that grounds an agent turn and generating an\nagent response based on a dialog and grounding document. In the first subtask,\nwe restrict the set of valid spans to the ones defined in the dataset, use a\nbiaffine classifier to model spans, and finally use an ensemble of different\nmodels. For the second subtask, we use a cascaded model which grounds the\nresponse prediction on the predicted span instead of the full document. With\nthese approaches, we obtain significant improvements in both subtasks compared\nto the baseline.', 'The history of AI has included several ""waves"" of ideas. The first wave, from\nthe mid-1950s to the 1980s, focused on logic and symbolic hand-encoded\nrepresentations of knowledge, the foundations of so-called ""expert systems"".\nThe second wave, starting in the 1990s, focused on statistics and machine\nlearning, in which, instead of hand-programming rules for behavior, programmers\nconstructed ""statistical learning algorithms"" that could be trained on large\ndatasets. In the most recent wave research in AI has largely focused on deep\n(i.e., many-layered) neural networks, which are loosely inspired by the brain\nand trained by ""deep learning"" methods. However, while deep neural networks\nhave led to many successes and new capabilities in computer vision, speech\nrecognition, language processing, game-playing, and robotics, their potential\nfor broad application remains limited by several factors.\n  A concerning limitation is that even the most successful of today\'s AI\nsystems suffer from brittleness-they can fail in unexpected ways when faced\nwith situations that differ sufficiently from ones they have been trained on.\nThis lack of robustness also appears in the vulnerability of AI systems to\nadversarial attacks, in which an adversary can subtly manipulate data in a way\nto guarantee a specific wrong answer or action from an AI system. AI systems\nalso can absorb biases-based on gender, race, or other factors-from their\ntraining data and further magnify these biases in their subsequent\ndecision-making. Taken together, these various limitations have prevented AI\nsystems such as automatic medical diagnosis or autonomous vehicles from being\nsufficiently trustworthy for wide deployment. The massive proliferation of AI\nacross society will require radically new ideas to yield technology that will\nnot sacrifice our productivity, our quality of life, or our values.', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', 'Next location prediction is of great importance for many location-based\napplications and provides essential intelligence to business and governments.\nIn existing studies, a common approach to next location prediction is to learn\nthe sequential transitions with massive historical trajectories based on\nconditional probability. Unfortunately, due to the time and space complexity,\nthese methods (e.g., Markov models) only use the just passed locations to\npredict next locations, without considering all the passed locations in the\ntrajectory. In this paper, we seek to enhance the prediction performance by\nconsidering the travel time from all the passed locations in the query\ntrajectory to a candidate next location. In particular, we propose a novel\nmethod, called Travel Time Difference Model (TTDM), which exploits the\ndifference between the shortest travel time and the actual travel time to\npredict next locations. Further, we integrate the TTDM with a Markov model via\na linear interpolation to yield a joint model, which computes the probability\nof reaching each possible next location and returns the top-rankings as\nresults. We have conducted extensive experiments on two real datasets: the\nvehicle passage record (VPR) data and the taxi trajectory data. The\nexperimental results demonstrate significant improvements in prediction\naccuracy over existing solutions. For example, compared with the Markov model,\nthe top-1 accuracy improves by 40% on the VPR data and by 15.6% on the Taxi\ndata.', 'As a new type of e-commerce platform developed in recent years, local\nconsumer service platform provides users with software to consume service to\nthe nearby store or to the home, such as Groupon and Koubei. Different from\nother common e-commerce platforms, the behavior of users on the local consumer\nservice platform is closely related to their real-time local context\ninformation. Therefore, building a context-aware user behavior prediction\nsystem is able to provide both merchants and users better service in local\nconsumer service platforms. However, most of the previous work just treats the\ncontextual information as an ordinary feature into the prediction model to\nobtain the prediction list under a specific context, which ignores the fact\nthat the interest of a user in different contexts is often significantly\ndifferent. Hence, in this paper, we propose a context-aware heterogeneous graph\nattention network (CHGAT) to dynamically generate the representation of the\nuser and to estimate the probability for future behavior. Specifically, we\nfirst construct the meta-path based heterogeneous graphs with the historical\nbehaviors from multiple sources and comprehend heterogeneous vertices in the\ngraph with a novel unified knowledge representing approach. Next, a multi-level\nattention mechanism is introduced for context-aware aggregation with graph\nvertices, which contains the vertex-level attention network and the path-level\nattention network. Both of them aim to capture the semantic correlation between\ninformation contained in the graph and the outside real-time contextual\ninformation in the search system. Then the model proposed in this paper\naggregates specific graphs with their corresponding context features and\nobtains the representation of user interest under a specific context and input\nit into the prediction network to finally obtain the predicted probability of\nuser behavior.', ""Providing accurate predictions is challenging for machine learning algorithms\nwhen the number of features is larger than the number of samples in the data.\nPrior knowledge can improve machine learning models by indicating relevant\nvariables and parameter values. Yet, this prior knowledge is often tacit and\nonly available from domain experts. We present a novel approach that uses\ninteractive visualization to elicit the tacit prior knowledge and uses it to\nimprove the accuracy of prediction models. The main component of our approach\nis a user model that models the domain expert's knowledge of the relevance of\ndifferent features for a prediction task. In particular, based on the expert's\nearlier input, the user model guides the selection of the features on which to\nelicit user's knowledge next. The results of a controlled user study show that\nthe user model significantly improves prior knowledge elicitation and\nprediction accuracy, when predicting the relative citation counts of scientific\ndocuments in a specific domain."", ""In this work we present a formal theoretical framework for assessing and\nanalyzing two classes of malevolent action towards generic Artificial\nIntelligence (AI) systems. Our results apply to general multi-class classifiers\nthat map from an input space into a decision space, including artificial neural\nnetworks used in deep learning applications. Two classes of attacks are\nconsidered. The first class involves adversarial examples and concerns the\nintroduction of small perturbations of the input data that cause\nmisclassification. The second class, introduced here for the first time and\nnamed stealth attacks, involves small perturbations to the AI system itself.\nHere the perturbed system produces whatever output is desired by the attacker\non a specific small data set, perhaps even a single input, but performs as\nnormal on a validation set (which is unknown to the attacker). We show that in\nboth cases, i.e., in the case of an attack based on adversarial examples and in\nthe case of a stealth attack, the dimensionality of the AI's decision-making\nspace is a major contributor to the AI's susceptibility. For attacks based on\nadversarial examples, a second crucial parameter is the absence of local\nconcentrations in the data probability distribution, a property known as\nSmeared Absolute Continuity. According to our findings, robustness to\nadversarial examples requires either (a) the data distributions in the AI's\nfeature space to have concentrated probability density functions or (b) the\ndimensionality of the AI's decision variables to be sufficiently small. We also\nshow how to construct stealth attacks on high-dimensional AI systems that are\nhard to spot unless the validation set is made exponentially large."", 'Vision Transformer (ViT) and its variants (e.g., Swin, PVT) have achieved\ngreat success in various computer vision tasks, owing to their capability to\nlearn long-range contextual information. Layer Normalization (LN) is an\nessential ingredient in these models. However, we found that the ordinary LN\nmakes tokens at different positions similar in magnitude because it normalizes\nembeddings within each token. It is difficult for Transformers to capture\ninductive bias such as the positional context in an image with LN. We tackle\nthis problem by proposing a new normalizer, termed Dynamic Token Normalization\n(DTN), where normalization is performed both within each token (intra-token)\nand across different tokens (inter-token). DTN has several merits. Firstly, it\nis built on a unified formulation and thus can represent various existing\nnormalization methods. Secondly, DTN learns to normalize tokens in both\nintra-token and inter-token manners, enabling Transformers to capture both the\nglobal contextual information and the local positional context. {Thirdly, by\nsimply replacing LN layers, DTN can be readily plugged into various vision\ntransformers, such as ViT, Swin, PVT, LeViT, T2T-ViT, BigBird and Reformer.\nExtensive experiments show that the transformer equipped with DTN consistently\noutperforms baseline model with minimal extra parameters and computational\noverhead. For example, DTN outperforms LN by $0.5\\%$ - $1.2\\%$ top-1 accuracy\non ImageNet, by $1.2$ - $1.4$ box AP in object detection on COCO benchmark, by\n$2.3\\%$ - $3.9\\%$ mCE in robustness experiments on ImageNet-C, and by $0.5\\%$ -\n$0.8\\%$ accuracy in Long ListOps on Long-Range Arena.} Codes will be made\npublic at \\url{https://github.com/wqshao126/DTN}', 'Current conversational AI systems aim to understand a set of pre-designed\nrequests and execute related actions, which limits them to evolve naturally and\nadapt based on human interactions. Motivated by how children learn their first\nlanguage interacting with adults, this paper describes a new Teachable AI\nsystem that is capable of learning new language nuggets called concepts,\ndirectly from end users using live interactive teaching sessions. The proposed\nsetup uses three models to: a) Identify gaps in understanding automatically\nduring live conversational interactions, b) Learn the respective\ninterpretations of such unknown concepts from live interactions with users, and\nc) Manage a classroom sub-dialogue specifically tailored for interactive\nteaching sessions. We propose state-of-the-art transformer based neural\narchitectures of models, fine-tuned on top of pre-trained models, and show\naccuracy improvements on the respective components. We demonstrate that this\nmethod is very promising in leading way to build more adaptive and personalized\nlanguage understanding models.', 'Conversational agents (""bots"") are beginning to be widely used in\nconversational interfaces. To design a system that is capable of emulating\nhuman-like interactions, a conversational layer that can serve as a fabric for\nchat-like interaction with the agent is needed. In this paper, we introduce a\nmodel that employs Information Retrieval by utilizing convolutional deep\nstructured semantic neural network-based features in the ranker to present\nhuman-like responses in ongoing conversation with a user. In conversations,\naccounting for context is critical to the retrieval model; we show that our\ncontext-sensitive approach using a Convolutional Deep Structured Semantic Model\n(cDSSM) with character trigrams significantly outperforms several conventional\nbaselines in terms of the relevance of responses retrieved.', 'Over the last several years, end-to-end neural conversational agents have\nvastly improved in their ability to carry a chit-chat conversation with humans.\nHowever, these models are often trained on large datasets from the internet,\nand as a result, may learn undesirable behaviors from this data, such as toxic\nor otherwise harmful language. Researchers must thus wrestle with the issue of\nhow and when to release these models. In this paper, we survey the problem\nlandscape for safety for end-to-end conversational AI and discuss recent and\nrelated work. We highlight tensions between values, potential positive impact\nand potential harms, and provide a framework for making decisions about whether\nand how to release these models, following the tenets of value-sensitive\ndesign. We additionally provide a suite of tools to enable researchers to make\nbetter-informed decisions about training and releasing end-to-end\nconversational AI models.', 'Despite the tremendous success of deep neural networks in various learning\nproblems, it has been observed that adding an intentionally designed\nadversarial perturbation to inputs of these architectures leads to erroneous\nclassification with high confidence in the prediction. In this work, we propose\na general framework based on the perturbation analysis of learning algorithms\nwhich consists of convex programming and is able to recover many current\nadversarial attacks as special cases. The framework can be used to propose\nnovel attacks against learning algorithms for classification and regression\ntasks under various new constraints with closed form solutions in many\ninstances. In particular we derive new attacks against classification\nalgorithms which are shown to achieve comparable performances to notable\nexisting attacks. The framework is then used to generate adversarial\nperturbations for regression tasks which include single pixel and single subset\nattacks. By applying this method to autoencoding and image colorization tasks,\nit is shown that adversarial perturbations can effectively perturb the output\nof regression tasks as well.', 'Context-aware neural machine translation (NMT) is a promising direction to\nimprove the translation quality by making use of the additional context, e.g.,\ndocument-level translation, or having meta-information. Although there exist\nvarious architectures and analyses, the effectiveness of different\ncontext-aware NMT models is not well explored yet. This paper analyzes the\nperformance of document-level NMT models on four diverse domains with a varied\namount of parallel document-level bilingual data. We conduct a comprehensive\nset of experiments to investigate the impact of document-level NMT. We find\nthat there is no single best approach to document-level NMT, but rather that\ndifferent architectures come out on top on different tasks. Looking at\ntask-specific problems, such as pronoun resolution or headline translation, we\nfind improvements in the context-aware systems, even in cases where the\ncorpus-level metrics like BLEU show no significant improvement. We also show\nthat document-level back-translation significantly helps to compensate for the\nlack of document-level bi-texts.', 'AI systems have seen significant adoption in various domains. At the same\ntime, further adoption in some domains is hindered by inability to fully trust\nan AI system that it will not harm a human. Besides the concerns for fairness,\nprivacy, transparency, and explainability are key to developing trusts in AI\nsystems. As stated in describing trustworthy AI ""Trust comes through\nunderstanding. How AI-led decisions are made and what determining factors were\nincluded are crucial to understand."" The subarea of explaining AI systems has\ncome to be known as XAI. Multiple aspects of an AI system can be explained;\nthese include biases that the data might have, lack of data points in a\nparticular region of the example space, fairness of gathering the data, feature\nimportances, etc. However, besides these, it is critical to have human-centered\nexplanations that are directly related to decision-making similar to how a\ndomain expert makes decisions based on ""domain knowledge,"" that also include\nwell-established, peer-validated explicit guidelines. To understand and\nvalidate an AI system\'s outcomes (such as classification, recommendations,\npredictions), that lead to developing trust in the AI system, it is necessary\nto involve explicit domain knowledge that humans understand and use.', 'The application of cognitive mechanisms to support knowledge acquisition is,\nfrom our point of view, crucial for making the resulting models coherent,\nefficient, credible, easy to use and understandable. In particular, there are\ntwo characteristic features of intelligence that are essential for knowledge\ndevelopment: forgetting and consolidation. Both plays an important role in\nknowledge bases and learning systems to avoid possible information overflow and\nredundancy, and in order to preserve and strengthen important or frequently\nused rules and remove (or forget) useless ones. We present an incremental,\nlong-life view of knowledge acquisition which tries to improve task after task\nby determining what to keep, what to consolidate and what to forget, overcoming\nThe Stability-Plasticity dilemma. In order to do that, we rate rules by\nintroducing several metrics through the first adaptation, to our knowledge, of\nthe Minimum Message Length (MML) principle to a coverage graph, a hierarchical\nassessment structure which treats evidence and rules in a unified way. The\nmetrics are not only used to forget some of the worst rules, but also to set a\nconsolidation process to promote those selected rules to the knowledge base,\nwhich is also mirrored by a demotion system. We evaluate the framework with a\nseries of tasks in a chess rule learning domain.', 'eXplainable AI focuses on generating explanations for the output of an AI\nalgorithm to a user, usually a decision-maker. Such user needs to interpret the\nAI system in order to decide whether to trust the machine outcome. When\naddressing this challenge, therefore, proper attention should be given to\nproduce explanations that are interpretable by the target community of users.\nIn this chapter, we claim for the need to better investigate what constitutes a\nhuman explanation, i.e. a justification of the machine behaviour that is\ninterpretable and actionable by the human decision makers. In particular, we\nfocus on the contributions that Human Intelligence can bring to eXplainable AI,\nespecially in conjunction with the exploitation of Knowledge Graphs. Indeed, we\ncall for a better interplay between Knowledge Representation and Reasoning,\nSocial Sciences, Human Computation and Human-Machine Cooperation research -- as\nalready explored in other AI branches -- in order to support the goal of\neXplainable AI with the adoption of a Human-in-the-Loop approach.', 'Owing to the rapidly growing multimedia content available on the Internet,\nextractive spoken document summarization, with the purpose of automatically\nselecting a set of representative sentences from a spoken document to concisely\nexpress the most important theme of the document, has been an active area of\nresearch and experimentation. On the other hand, word embedding has emerged as\na newly favorite research subject because of its excellent performance in many\nnatural language processing (NLP)-related tasks. However, as far as we are\naware, there are relatively few studies investigating its use in extractive\ntext or speech summarization. A common thread of leveraging word embeddings in\nthe summarization process is to represent the document (or sentence) by\naveraging the word embeddings of the words occurring in the document (or\nsentence). Then, intuitively, the cosine similarity measure can be employed to\ndetermine the relevance degree between a pair of representations. Beyond the\ncontinued efforts made to improve the representation of words, this paper\nfocuses on building novel and efficient ranking models based on the general\nword embedding methods for extractive speech summarization. Experimental\nresults demonstrate the effectiveness of our proposed methods, compared to\nexisting state-of-the-art methods.', 'A key challenge of dialog systems research is to effectively and efficiently\nadapt to new domains. A scalable paradigm for adaptation necessitates the\ndevelopment of generalizable models that perform well in few-shot settings. In\nthis paper, we focus on the intent classification problem which aims to\nidentify user intents given utterances addressed to the dialog system. We\npropose two approaches for improving the generalizability of utterance\nclassification models: (1) observers and (2) example-driven training. Prior\nwork has shown that BERT-like models tend to attribute a significant amount of\nattention to the [CLS] token, which we hypothesize results in diluted\nrepresentations. Observers are tokens that are not attended to, and are an\nalternative to the [CLS] token as a semantic representation of utterances.\nExample-driven training learns to classify utterances by comparing to examples,\nthereby using the underlying encoder as a sentence similarity model. These\nmethods are complementary; improving the representation through observers\nallows the example-driven model to better measure sentence similarities. When\ncombined, the proposed methods attain state-of-the-art results on three intent\nprediction datasets (\\textsc{banking77}, \\textsc{clinc150}, \\textsc{hwu64}) in\nboth the full data and few-shot (10 examples per intent) settings. Furthermore,\nwe demonstrate that the proposed approach can transfer to new intents and\nacross datasets without any additional training.', 'Recently we see a rising number of methods in the field of eXplainable\nArtificial Intelligence. To our surprise, their development is driven by model\ndevelopers rather than a study of needs for human end users. The analysis of\nneeds, if done, takes the form of an A/B test rather than a study of open\nquestions. To answer the question ""What would a human operator like to ask the\nML model?"" we propose a conversational system explaining decisions of the\npredictive model. In this experiment, we developed a chatbot called dr_ant to\ntalk about machine learning model trained to predict survival odds on Titanic.\nPeople can talk with dr_ant about different aspects of the model to understand\nthe rationale behind its predictions. Having collected a corpus of 1000+\ndialogues, we analyse the most common types of questions that users would like\nto ask. To our knowledge, it is the first study which uses a conversational\nsystem to collect the needs of human operators from the interactive and\niterative dialogue explorations of a predictive model.', 'We describe the setting and results of the ConvAI2 NeurIPS competition that\naims to further the state-of-the-art in open-domain chatbots. Some key\ntakeaways from the competition are: (i) pretrained Transformer variants are\ncurrently the best performing models on this task, (ii) but to improve\nperformance on multi-turn conversations with humans, future systems must go\nbeyond single word metrics like perplexity to measure the performance across\nsequences of utterances (conversations) -- in terms of repetition, consistency\nand balance of dialogue acts (e.g. how many questions asked vs. answered).', ""For the next step in human to machine interaction, Artificial Intelligence\n(AI) should interact predominantly using natural language because, if it\nworked, it would be the fastest way to communicate. Facebook's toy tasks (bAbI)\nprovide a useful benchmark to compare implementations for conversational AI.\nWhile the published experiments so far have been based on exploiting the\ndistributional hypothesis with machine learning, our model exploits natural\nlanguage understanding (NLU) with the decomposition of language based on Role\nand Reference Grammar (RRG) and the brain-based Patom theory. Our combinatorial\nsystem for conversational AI based on linguistics has many advantages: passing\nbAbI task tests without parsing or statistics while increasing scalability. Our\nmodel validates both the training and test data to find 'garbage' input and\noutput (GIGO). It is not rules-based, nor does it use parts of speech, but\ninstead relies on meaning. While Deep Learning is difficult to debug and fix,\nevery step in our model can be understood and changed like any non-statistical\ncomputer program. Deep Learning's lack of explicable reasoning has raised\nopposition to AI, partly due to fear of the unknown. To support the goals of\nAI, we propose extended tasks to use human-level statements with tense, aspect\nand voice, and embedded clauses with junctures: and answers to be natural\nlanguage generation (NLG) instead of keywords. While machine learning permits\ninvalid training data to produce incorrect test responses, our system cannot\nbecause the context tracking would need to be intentionally broken. We believe\nno existing learning systems can currently solve these extended natural\nlanguage tests. There appears to be a knowledge gap between NLP researchers and\nlinguists, but ongoing competitive results such as these promise to narrow that\ngap."", 'Endowing a chatbot with personality or an identity is quite challenging but\ncritical to deliver more realistic and natural conversations. In this paper, we\naddress the issue of generating responses that are coherent to a pre-specified\nagent profile. We design a model consisting of three modules: a profile\ndetector to decide whether a post should be responded using the profile and\nwhich key should be addressed, a bidirectional decoder to generate responses\nforward and backward starting from a selected profile value, and a position\ndetector that predicts a word position from which decoding should start given a\nselected profile value. We show that general conversation data from social\nmedia can be used to generate profile-coherent responses. Manual and automatic\nevaluation shows that our model can deliver more coherent, natural, and\ndiversified responses.', 'Document-level relation extraction aims to discover relations between\nentities across a whole document. How to build the dependency of entities from\ndifferent sentences in a document remains to be a great challenge. Current\napproaches either leverage syntactic trees to construct document-level graphs\nor aggregate inference information from different sentences. In this paper, we\nbuild cross-sentence dependencies by inferring compositional relations between\ninter-sentence mentions. Adopting aggressive linking strategy, intermediate\nrelations are reasoned on the document-level graphs by mention convolution. We\nfurther notice the generalization problem of NA instances, which is caused by\nincomplete annotation and worsened by fully-connected mention pairs. An\nimproved ranking loss is proposed to attend this problem. Experiments show the\nconnections between different mentions are crucial to document-level relation\nextraction, which enables the model to extract more meaningful higher-level\ncompositional relations.', 'Artificial intelligence (AI) has huge potential to improve the health and\nwell-being of people, but adoption in clinical practice is still limited. Lack\nof transparency is identified as one of the main barriers to implementation, as\nclinicians should be confident the AI system can be trusted. Explainable AI has\nthe potential to overcome this issue and can be a step towards trustworthy AI.\nIn this paper we review the recent literature to provide guidance to\nresearchers and practitioners on the design of explainable AI systems for the\nhealth-care domain and contribute to formalization of the field of explainable\nAI. We argue the reason to demand explainability determines what should be\nexplained as this determines the relative importance of the properties of\nexplainability (i.e. interpretability and fidelity). Based on this, we propose\na framework to guide the choice between classes of explainable AI methods\n(explainable modelling versus post-hoc explanation; model-based,\nattribution-based, or example-based explanations; global and local\nexplanations). Furthermore, we find that quantitative evaluation metrics, which\nare important for objective standardized evaluation, are still lacking for some\nproperties (e.g. clarity) and types of explanations (e.g. example-based\nmethods). We conclude that explainable modelling can contribute to trustworthy\nAI, but the benefits of explainability still need to be proven in practice and\ncomplementary measures might be needed to create trustworthy AI in health care\n(e.g. reporting data quality, performing extensive (external) validation, and\nregulation).', 'AI practitioners typically strive to develop the most accurate systems,\nmaking an implicit assumption that the AI system will function autonomously.\nHowever, in practice, AI systems often are used to provide advice to people in\ndomains ranging from criminal justice and finance to healthcare. In such\nAI-advised decision making, humans and machines form a team, where the human is\nresponsible for making final decisions. But is the most accurate AI the best\nteammate? We argue ""No"" -- predictable performance may be worth a slight\nsacrifice in AI accuracy. Instead, we argue that AI systems should be trained\nin a human-centered manner, directly optimized for team performance. We study\nthis proposal for a specific type of human-AI teaming, where the human overseer\nchooses to either accept the AI recommendation or solve the task themselves. To\noptimize the team performance for this setting we maximize the team\'s expected\nutility, expressed in terms of the quality of the final decision, cost of\nverifying, and individual accuracies of people and machines. Our experiments\nwith linear and non-linear models on real-world, high-stakes datasets show that\nthe most accuracy AI may not lead to highest team performance and show the\nbenefit of modeling teamwork during training through improvements in expected\nteam utility across datasets, considering parameters such as human skill and\nthe cost of mistakes. We discuss the shortcoming of current optimization\napproaches beyond well-studied loss functions such as log-loss, and encourage\nfuture work on AI optimization problems motivated by human-AI collaboration.', 'Most typical click models assume that the probability of a document to be\nexamined by users only depends on position, such as PBM and UBM. It works well\nin various kinds of search engines. However, in a search engine where massive\ncandidate documents display images as responses to the query, the examination\nprobability should not only depend on position. The visual appearance of an\nimage-oriented document also plays an important role in its opportunity to be\nexamined. In this paper, we assume that vision bias exists in an image-oriented\nsearch engine as another crucial factor affecting the examination probability\naside from position. Specifically, we apply this assumption to classical click\nmodels and propose an extended model, to better capture the examination\nprobabilities of documents. We use regression-based EM algorithm to predict the\nvision bias given the visual features extracted from candidate documents.\nEmpirically, we evaluate our model on a dataset developed from a real-world\nonline image-oriented search engine, and demonstrate that our proposed model\ncan achieve significant improvements over its baseline model in data fitness\nand sparsity handling.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'In this paper we discuss the relationships between conditional and\npreferential logics and neural network models, based on a multi-preferential\nsemantics. We propose a concept-wise multipreference semantics, recently\nintroduced for defeasible description logics to take into account preferences\nwith respect to different concepts, as a tool for providing a semantic\ninterpretation to neural network models. This approach has been explored both\nfor unsupervised neural network models (Self-Organising Maps) and for\nsupervised ones (Multilayer Perceptrons), and we expect that the same approach\nmight be extended to other neural network models. It allows for logical\nproperties of the network to be checked (by model checking) over an\ninterpretation capturing the input-output behavior of the network. For\nMultilayer Perceptrons, the deep network itself can be regarded as a\nconditional knowledge base, in which synaptic connections correspond to\nweighted conditionals. The paper describes the general approach, through the\ncases of Self-Organising Maps and Multilayer Perceptrons, and discusses some\nopen issues and perspectives.', ""We develop and study new adversarial perturbations that enable an attacker to\ngain control over decisions in generic Artificial Intelligence (AI) systems\nincluding deep learning neural networks. In contrast to adversarial data\nmodification, the attack mechanism we consider here involves alterations to the\nAI system itself. Such a stealth attack could be conducted by a mischievous,\ncorrupt or disgruntled member of a software development team. It could also be\nmade by those wishing to exploit a ``democratization of AI'' agenda, where\nnetwork architectures and trained parameter sets are shared publicly. We\ndevelop a range of new implementable attack strategies with accompanying\nanalysis, showing that with high probability a stealth attack can be made\ntransparent, in the sense that system performance is unchanged on a fixed\nvalidation set which is unknown to the attacker, while evoking any desired\noutput on a trigger input of interest. The attacker only needs to have\nestimates of the size of the validation set and the spread of the AI's relevant\nlatent space. In the case of deep learning neural networks, we show that a one\nneuron attack is possible - a modification to the weights and bias associated\nwith a single neuron - revealing a vulnerability arising from\nover-parameterization. We illustrate these concepts using state of the art\narchitectures on two standard image data sets. Guided by the theory and\ncomputational results, we also propose strategies to guard against stealth\nattacks."", ""Trust is a central component of the interaction between people and AI, in\nthat 'incorrect' levels of trust may cause misuse, abuse or disuse of the\ntechnology. But what, precisely, is the nature of trust in AI? What are the\nprerequisites and goals of the cognitive mechanism of trust, and how can we\npromote them, or assess whether they are being satisfied in a given\ninteraction? This work aims to answer these questions. We discuss a model of\ntrust inspired by, but not identical to, sociology's interpersonal trust (i.e.,\ntrust between people). This model rests on two key properties of the\nvulnerability of the user and the ability to anticipate the impact of the AI\nmodel's decisions. We incorporate a formalization of 'contractual trust', such\nthat trust between a user and an AI is trust that some implicit or explicit\ncontract will hold, and a formalization of 'trustworthiness' (which detaches\nfrom the notion of trustworthiness in sociology), and with it concepts of\n'warranted' and 'unwarranted' trust. We then present the possible causes of\nwarranted trust as intrinsic reasoning and extrinsic behavior, and discuss how\nto design trustworthy AI, how to evaluate whether trust has manifested, and\nwhether it is warranted. Finally, we elucidate the connection between trust and\nXAI using our formalization."", 'The widespread utilization of AI systems has drawn attention to the potential\nimpacts of such systems on society. Of particular concern are the consequences\nthat prediction errors may have on real-world scenarios, and the trust humanity\nplaces in AI systems. It is necessary to understand how we can evaluate\ntrustworthiness in AI and how individuals and entities alike can develop\ntrustworthy AI systems. In this paper, we analyze each element of\ntrustworthiness and provide a set of 20 guidelines that can be leveraged to\nensure optimal AI functionality while taking into account the greater ethical,\ntechnical, and practical impacts to humanity. Moreover, the guidelines help\nensure that trustworthiness is provable and can be demonstrated, they are\nimplementation agnostic, and they can be applied to any AI system in any\nsector.', ""The role of conversational assistants has become more prevalent in helping\npeople increase their productivity. Document-centered assistance, for example\nto help an individual quickly review a document, has seen less significant\nprogress, even though it has the potential to tremendously increase a user's\nproductivity. This type of document-centered assistance is the focus of this\npaper. Our contributions are three-fold: (1) We first present a survey to\nunderstand the space of document-centered assistance and the capabilities\npeople expect in this scenario. (2) We investigate the types of queries that\nusers will pose while seeking assistance with documents, and show that\ndocument-centered questions form the majority of these queries. (3) We present\na set of initial machine learned models that show that (a) we can accurately\ndetect document-centered questions, and (b) we can build reasonably accurate\nmodels for answering such questions. These positive results are encouraging,\nand suggest that even greater results may be attained with continued study of\nthis interesting and novel problem space. Our findings have implications for\nthe design of intelligent systems to support task completion via natural\ninteractions with documents."", 'People have information needs of varying complexity, which can be solved by\nan intelligent agent able to answer questions formulated in a proper way,\neventually considering user context and preferences. In a scenario in which the\nuser profile can be considered as a question, intelligent agents able to answer\nquestions can be used to find the most relevant answers for a given user. In\nthis work we propose a novel model based on Artificial Neural Networks to\nanswer questions with multiple answers by exploiting multiple facts retrieved\nfrom a knowledge base. The model is evaluated on the factoid Question Answering\nand top-n recommendation tasks of the bAbI Movie Dialog dataset. After\nassessing the performance of the model on both tasks, we try to define the\nlong-term goal of a conversational recommender system able to interact using\nnatural language and to support users in their information seeking processes in\na personalized way.', 'Building a dialogue system that can communicate naturally with humans is a\nchallenging yet interesting problem of agent-based computing. The rapid growth\nin this area is usually hindered by the long-standing problem of data scarcity\nas these systems are expected to learn syntax, grammar, decision making, and\nreasoning from insufficient amounts of task-specific dataset. The recently\nintroduced pre-trained language models have the potential to address the issue\nof data scarcity and bring considerable advantages by generating contextualized\nword embeddings. These models are considered counterpart of ImageNet in NLP and\nhave demonstrated to capture different facets of language such as hierarchical\nrelations, long-term dependency, and sentiment. In this short survey paper, we\ndiscuss the recent progress made in the field of pre-trained language models.\nWe also deliberate that how the strengths of these language models can be\nleveraged in designing more engaging and more eloquent conversational agents.\nThis paper, therefore, intends to establish whether these pre-trained models\ncan overcome the challenges pertinent to dialogue systems, and how their\narchitecture could be exploited in order to overcome these challenges. Open\nchallenges in the field of dialogue systems have also been deliberated.', 'Today is the era of intelligence in machines. With the advances in Artificial\nIntelligence, machines have started to impersonate different human traits, a\nchatbot is the next big thing in the domain of conversational services. A\nchatbot is a virtual person who is capable to carry out a natural conversation\nwith people. They can include skills that enable them to converse with the\nhumans in audio, visual, or textual formats. Artificial intelligence\nconversational entities, also called chatbots, conversational agents, or\ndialogue system, are an excellent example of such machines. Obtaining the right\ninformation at the right time and place is the key to effective disaster\nmanagement. The term ""disaster management"" encompasses both natural and\nhuman-caused disasters. To assist citizens, our project is to create a COVID\nAssistant to provide the need of up to date information to be available 24\nhours. With the growth in the World Wide Web, it is quite intelligible that\nusers are interested in the swift and relatedly correct information for their\nhunt. A chatbot can be seen as a question-and-answer system in which experts\nprovide knowledge to solicit users. This master thesis is dedicated to discuss\nCOVID Assistant chatbot and explain each component in detail. The design of the\nproposed chatbot is introduced by its seven components: Ontology, Web Scraping\nmodule, DB, State Machine, keyword Extractor, Trained chatbot, and User\nInterface.', ""A market-maker-based prediction market lets forecasters aggregate information\nby editing a consensus probability distribution either directly or by trading\nsecurities that pay off contingent on an event of interest. Combinatorial\nprediction markets allow trading on any event that can be specified as a\ncombination of a base set of events. However, explicitly representing the full\njoint distribution is infeasible for markets with more than a few base events.\nA factored representation such as a Bayesian network (BN) can achieve tractable\ncomputation for problems with many related variables. Standard BN inference\nalgorithms, such as the junction tree algorithm, can be used to update a\nrepresentation of the entire joint distribution given a change to any local\nconditional probability. However, in order to let traders reuse assets from\nprior trades while never allowing assets to become negative, a BN based\nprediction market also needs to update a representation of each user's assets\nand find the conditional state in which a user has minimum assets. Users also\nfind it useful to see their expected assets given an edit outcome. We show how\nto generalize the junction tree algorithm to perform all these computations."", ""In the current era, people and society have grown increasingly reliant on\nartificial intelligence (AI) technologies. AI has the potential to drive us\ntowards a future in which all of humanity flourishes. It also comes with\nsubstantial risks for oppression and calamity. Discussions about whether we\nshould (re)trust AI have repeatedly emerged in recent years and in many\nquarters, including industry, academia, healthcare, services, and so on.\nTechnologists and AI researchers have a responsibility to develop trustworthy\nAI systems. They have responded with great effort to design more responsible AI\nalgorithms. However, existing technical solutions are narrow in scope and have\nbeen primarily directed towards algorithms for scoring or classification tasks,\nwith an emphasis on fairness and unwanted bias. To build long-lasting trust\nbetween AI and human beings, we argue that the key is to think beyond\nalgorithmic fairness and connect major aspects of AI that potentially cause\nAI's indifferent behavior. In this survey, we provide a systematic framework of\nSocially Responsible AI Algorithms that aims to examine the subjects of AI\nindifference and the need for socially responsible AI algorithms, define the\nobjectives, and introduce the means by which we may achieve these objectives.\nWe further discuss how to leverage this framework to improve societal\nwell-being through protection, information, and prevention/mitigation."", 'Many name tagging approaches use local contextual information with much\nsuccess, but fail when the local context is ambiguous or limited. We present a\nnew framework to improve name tagging by utilizing local, document-level, and\ncorpus-level contextual information. We retrieve document-level context from\nother sentences within the same document and corpus-level context from\nsentences in other topically related documents. We propose a model that learns\nto incorporate document-level and corpus-level contextual information alongside\nlocal contextual information via global attentions, which dynamically weight\ntheir respective contextual information, and gating mechanisms, which determine\nthe influence of this information. Extensive experiments on benchmark datasets\nshow the effectiveness of our approach, which achieves state-of-the-art results\nfor Dutch, German, and Spanish on the CoNLL-2002 and CoNLL-2003 datasets.', 'Recent advances in artificial intelligence (AI) and machine learning have\ncreated a general perception that AI could be used to solve complex problems,\nand in some situations over-hyped as a tool that can be so easily used.\nUnfortunately, the barrier to realization of mass adoption of AI on various\nbusiness domains is too high because most domain experts have no background in\nAI. Developing AI applications involves multiple phases, namely data\npreparation, application modeling, and product deployment. The effort of AI\nresearch has been spent mostly on new AI models (in the model training stage)\nto improve the performance of benchmark tasks such as image recognition. Many\nother factors such as usability, efficiency and security of AI have not been\nwell addressed, and therefore form a barrier to democratizing AI. Further, for\nmany real world applications such as healthcare and autonomous driving,\nlearning via huge amounts of possibility exploration is not feasible since\nhumans are involved. In many complex applications such as healthcare, subject\nmatter experts (e.g. Clinicians) are the ones who appreciate the importance of\nfeatures that affect health, and their knowledge together with existing\nknowledge bases are critical to the end results. In this paper, we take a new\nperspective on developing AI solutions, and present a solution for making AI\nusable. We hope that this resolution will enable all subject matter experts\n(eg. Clinicians) to exploit AI like data scientists.', 'Chatbot is a technology that is used to mimic human behavior using natural\nlanguage. There are different types of Chatbot that can be used as\nconversational agent in various business domains in order to increase the\ncustomer service and satisfaction. For any business domain, it requires a\nknowledge base to be built for that domain and design an information retrieval\nbased system that can respond the user with a piece of documentation or\ngenerated sentences. The core component of a Chatbot is Natural Language\nUnderstanding (NLU) which has been impressively improved by deep learning\nmethods. But we often lack such properly built NLU modules and requires more\ntime to build it from scratch for high quality conversations. This may\nencourage fresh learners to build a Chatbot from scratch with simple\narchitecture and using small dataset, although it may have reduced\nfunctionality, rather than building high quality data driven methods. This\nresearch focuses on Named Entity Recognition (NER) and Intent Classification\nmodels which can be integrated into NLU service of a Chatbot. Named entities\nwill be inserted manually in the knowledge base and automatically detected in a\ngiven sentence. The NER model in the proposed architecture is based on\nartificial neural network which is trained on manually created entities and\nevaluated using CoNLL-2003 dataset.', ""Today, AI is being increasingly used to help human experts make decisions in\nhigh-stakes scenarios. In these scenarios, full automation is often\nundesirable, not only due to the significance of the outcome, but also because\nhuman experts can draw on their domain knowledge complementary to the model's\nto ensure task success. We refer to these scenarios as AI-assisted decision\nmaking, where the individual strengths of the human and the AI come together to\noptimize the joint decision outcome. A key to their success is to appropriately\n\\textit{calibrate} human trust in the AI on a case-by-case basis; knowing when\nto trust or distrust the AI allows the human expert to appropriately apply\ntheir knowledge, improving decision outcomes in cases where the model is likely\nto perform poorly. This research conducts a case study of AI-assisted decision\nmaking in which humans and AI have comparable performance alone, and explores\nwhether features that reveal case-specific model information can calibrate\ntrust and improve the joint performance of the human and AI. Specifically, we\nstudy the effect of showing confidence score and local explanation for a\nparticular prediction. Through two human experiments, we show that confidence\nscore can help calibrate people's trust in an AI model, but trust calibration\nalone is not sufficient to improve AI-assisted decision making, which may also\ndepend on whether the human can bring in enough unique knowledge to complement\nthe AI's errors. We also highlight the problems in using local explanation for\nAI-assisted decision making scenarios and invite the research community to\nexplore new approaches to explainability for calibrating human trust in AI."", 'With the complexity of the network structure, uncertainty inference has\nbecome an important task to improve the classification accuracy for artificial\nintelligence systems. For image classification tasks, we propose a structured\nDropConnect (SDC) framework to model the output of a deep neural network by a\nDirichlet distribution. We introduce a DropConnect strategy on weights in the\nfully connected layers during training. In test, we split the network into\nseveral sub-networks, and then model the Dirichlet distribution by match its\nmoments with the mean and variance of the outputs of these sub-networks. The\nentropy of the estimated Dirichlet distribution is finally utilized for\nuncertainty inference. In this paper, this framework is implemented on LeNet$5$\nand VGG$16$ models for misclassification detection and out-of-distribution\ndetection on MNIST and CIFAR-$10$ datasets. Experimental results show that the\nperformance of the proposed SDC can be comparable to other uncertainty\ninference methods. Furthermore, the SDC is adapted well to different network\nstructures with certain generalization capabilities and research prospects.', 'In adversarial training, a set of models learn together by pursuing competing\ngoals, usually defined on single data instances. However, in relational\nlearning and other non-i.i.d domains, goals can also be defined over sets of\ninstances. For example, a link predictor for the is-a relation needs to be\nconsistent with the transitivity property: if is-a(x_1, x_2) and is-a(x_2, x_3)\nhold, is-a(x_1, x_3) needs to hold as well. Here we use such assumptions for\nderiving an inconsistency loss, measuring the degree to which the model\nviolates the assumptions on an adversarially-generated set of examples. The\ntraining objective is defined as a minimax problem, where an adversary finds\nthe most offending adversarial examples by maximising the inconsistency loss,\nand the model is trained by jointly minimising a supervised loss and the\ninconsistency loss on the adversarial examples. This yields the first method\nthat can use function-free Horn clauses (as in Datalog) to regularise any\nneural link predictor, with complexity independent of the domain size. We show\nthat for several link prediction models, the optimisation problem faced by the\nadversary has efficient closed-form solutions. Experiments on link prediction\nbenchmarks indicate that given suitable prior knowledge, our method can\nsignificantly improve neural link predictors on all relevant metrics.', ""Neural networks with deep architectures have demonstrated significant\nperformance improvements in computer vision, speech recognition, and natural\nlanguage processing. The challenges in information retrieval (IR), however, are\ndifferent from these other application areas. A common form of IR involves\nranking of documents--or short passages--in response to keyword-based queries.\nEffective IR systems must deal with query-document vocabulary mismatch problem,\nby modeling relationships between different query and document terms and how\nthey indicate relevance. Models should also consider lexical matches when the\nquery contains rare terms--such as a person's name or a product model\nnumber--not seen during training, and to avoid retrieving semantically related\nbut irrelevant results. In many real-life IR tasks, the retrieval involves\nextremely large collections--such as the document index of a commercial Web\nsearch engine--containing billions of documents. Efficient IR methods should\ntake advantage of specialized IR data structures, such as inverted index, to\nefficiently retrieve from large collections. Given an information need, the IR\nsystem also mediates how much exposure an information artifact receives by\ndeciding whether it should be displayed, and where it should be positioned,\namong other results. Exposure-aware IR systems may optimize for additional\nobjectives, besides relevance, such as parity of exposure for retrieved items\nand content publishers. In this thesis, we present novel neural architectures\nand methods motivated by the specific needs and challenges of IR tasks."", 'Document clustering is a text mining technique used to provide better\ndocument search and browsing in digital libraries or online corpora. A lot of\nresearch has been done on biomedical document clustering that is based on using\nexisting ontology. But, associations and co-occurrences of the medical concepts\nare not well represented by using ontology. In this research, a vector\nrepresentation of concepts of diseases and similarity measurement between\nconcepts are proposed. They identify the closest concepts of diseases in the\ncontext of a corpus. Each document is represented by using the vector space\nmodel. A weight scheme is proposed to consider both local content and\nassociations between concepts. A Self-Organizing Map is used as document\nclustering algorithm. The vector projection and visualization features of SOM\nenable visualization and analysis of the clusters distributions and\nrelationships on the two dimensional space. The experimental results show that\nthe proposed document clustering framework generates meaningful clusters and\nfacilitate visualization of the clusters based on the concepts of diseases.', 'Black-box machine learning models are now routinely used in high-risk\nsettings, like medical diagnostics, which demand uncertainty quantification to\navoid consequential model failures. Conformal prediction is a user-friendly\nparadigm for creating statistically rigorous uncertainty sets/intervals for the\npredictions of such models. Critically, the sets are valid in a\ndistribution-free sense: they possess explicit, non-asymptotic guarantees even\nwithout distributional assumptions or model assumptions. One can use conformal\nprediction with any pre-trained model, such as a neural network, to produce\nsets that are guaranteed to contain the ground truth with a user-specified\nprobability, such as 90%. It is easy-to-understand, easy-to-use, and general,\napplying naturally to problems arising in the fields of computer vision,\nnatural language processing, deep reinforcement learning, and so on.\n  This hands-on introduction is aimed to provide the reader a working\nunderstanding of conformal prediction and related distribution-free uncertainty\nquantification techniques with one self-contained document. We lead the reader\nthrough practical theory for and examples of conformal prediction and describe\nits extensions to complex machine learning tasks involving structured outputs,\ndistribution shift, time-series, outliers, models that abstain, and more.\nThroughout, there are many explanatory illustrations, examples, and code\nsamples in Python. With each code sample comes a Jupyter notebook implementing\nthe method on a real-data example; the notebooks can be accessed and easily run\nusing our codebase.', 'Over the past few years, providers such as Google, Microsoft, and Amazon have\nstarted to provide customers with access to software interfaces allowing them\nto easily embed machine learning tasks into their applications. Overall,\norganizations can now use Machine Learning as a Service (MLaaS) engines to\noutsource complex tasks, e.g., training classifiers, performing predictions,\nclustering, etc. They can also let others query models trained on their data.\nNaturally, this approach can also be used (and is often advocated) in other\ncontexts, including government collaborations, citizen science projects, and\nbusiness-to-business partnerships. However, if malicious users were able to\nrecover data used to train these models, the resulting information leakage\nwould create serious issues. Likewise, if the inner parameters of the model are\nconsidered proprietary information, then access to the model should not allow\nan adversary to learn such parameters. In this document, we set to review\nprivacy challenges in this space, providing a systematic review of the relevant\nresearch literature, also exploring possible countermeasures. More\nspecifically, we provide ample background information on relevant concepts\naround machine learning and privacy. Then, we discuss possible adversarial\nmodels and settings, cover a wide range of attacks that relate to private\nand/or sensitive information leakage, and review recent results attempting to\ndefend against such attacks. Finally, we conclude with a list of open problems\nthat require more work, including the need for better evaluations, more\ntargeted defenses, and the study of the relation to policy and data protection\nefforts.', 'Chatbots and AI assistants have claimed their importance in today life. The\nmain reason behind adopting this technology is to connect with the user,\nunderstand their requirements, and fulfill them. This has been achieved but at\nthe cost of heavy training data and complex learning models. This work is\ncarried out proposes a simple algorithm, a model which can be implemented in\ndifferent fields each with its own work scope. The proposed model converts\nhuman language text to computer-understandable SQL queries. The model requires\ndata only related to the specific field, saving data space. This model performs\nlinear computation hence solving the computational complexity. This work also\ndefines the stages where a new methodology is implemented and what previous\nmethod was adopted to fulfill the requirement at that stage. Two datasets\navailable online will be used in this work, the ATIS dataset, and WikiSQL. This\nwork compares the computation time among the 2 datasets and also compares the\naccuracy of both. This paper works over basic Natural language processing tasks\nlike semantic parsing, NER, parts of speech and tends to achieve results\nthrough these simple methods.', 'Instances of Artificial Intelligence (AI) systems failing to deliver\nconsistent, satisfactory performance are legion. We investigate why AI failures\noccur. We address only a narrow subset of the broader field of AI Safety. We\nfocus on AI failures on account of flaws in conceptualization, design and\ndeployment. Other AI Safety issues like trade-offs between privacy and security\nor convenience, bad actors hacking into AI systems to create mayhem or bad\nactors deploying AI for purposes harmful to humanity and are out of scope of\nour discussion. We find that AI systems fail on account of omission and\ncommission errors in the design of the AI system, as well as upon failure to\ndevelop an appropriate interpretation of input information. Moreover, even when\nthere is no significant flaw in the AI software, an AI system may fail because\nthe hardware is incapable of robust performance across environments. Finally an\nAI system is quite likely to fail in situations where, in effect, it is called\nupon to deliver moral judgments -- a capability AI does not possess. We observe\ncertain trade-offs in measures to mitigate a subset of AI failures and provide\nsome recommendations.', ""The task of next POI recommendation has been studied extensively in recent\nyears. However, developing an unified recommendation framework to incorporate\nmultiple factors associated with both POIs and users remains challenging,\nbecause of the heterogeneity nature of these information. Further, effective\nmechanisms to handle cold-start and endow the system with interpretability are\nalso difficult topics. Inspired by the recent success of neural networks in\nmany areas, in this paper, we present a simple but effective neural network\nframework for next POI recommendation, named NEXT. NEXT is an unified framework\nto learn the hidden intent regarding user's next move, by incorporating\ndifferent factors in an unified manner. Specifically, in NEXT, we incorporate\nmeta-data information and two kinds of temporal contexts (i.e., time interval\nand visit time). To leverage sequential relations and geographical influence,\nwe propose to adopt DeepWalk, a network representation learning technique, to\nencode such knowledge. We evaluate the effectiveness of NEXT against\nstate-of-the-art alternatives and neural networks based solutions. Experimental\nresults over three publicly available datasets demonstrate that NEXT\nsignificantly outperforms baselines in real-time next POI recommendation.\nFurther experiments demonstrate the superiority of NEXT in handling cold-start.\nMore importantly, we show that NEXT provides meaningful explanation of the\ndimensions in hidden intent space."", 'Perhaps the most ambitious scientific quest in human history is the creation\nof general artificial intelligence, which roughly means AI that is as smart or\nsmarter than humans. The dominant approach in the machine learning community is\nto attempt to discover each of the pieces required for intelligence, with the\nimplicit assumption that some future group will complete the Herculean task of\nfiguring out how to combine all of those pieces into a complex thinking\nmachine. I call this the ""manual AI approach"". This paper describes another\nexciting path that ultimately may be more successful at producing general AI.\nIt is based on the clear trend in machine learning that hand-designed solutions\neventually are replaced by more effective, learned solutions. The idea is to\ncreate an AI-generating algorithm (AI-GA), which automatically learns how to\nproduce general AI. Three Pillars are essential for the approach: (1)\nmeta-learning architectures, (2) meta-learning the learning algorithms\nthemselves, and (3) generating effective learning environments. I argue that\neither approach could produce general AI first, and both are scientifically\nworthwhile irrespective of which is the fastest path. Because both are\npromising, yet the ML community is currently committed to the manual approach,\nI argue that our community should increase its research investment in the AI-GA\napproach. To encourage such research, I describe promising work in each of the\nThree Pillars. I also discuss AI-GA-specific safety and ethical considerations.\nBecause it it may be the fastest path to general AI and because it is\ninherently scientifically interesting to understand the conditions in which a\nsimple algorithm can produce general AI (as happened on Earth where Darwinian\nevolution produced human intelligence), I argue that the pursuit of AI-GAs\nshould be considered a new grand challenge of computer science research.', 'In recent years, traditional cybersecurity safeguards have proven ineffective\nagainst insider threats. Famous cases of sensitive information leaks caused by\ninsiders, including the WikiLeaks release of diplomatic cables and the Edward\nSnowden incident, have greatly harmed the U.S. government\'s relationship with\nother governments and with its own citizens. Data Leak Prevention (DLP) is a\nsolution for detecting and preventing information leaks from within an\norganization\'s network. However, state-of-art DLP detection models are only\nable to detect very limited types of sensitive information, and research in the\nfield has been hindered due to the lack of available sensitive texts. Many\nresearchers have focused on document-based detection with artificially labeled\n""confidential documents"" for which security labels are assigned to the entire\ndocument, when in reality only a portion of the document is sensitive. This\ntype of whole-document based security labeling increases the chances of\npreventing authorized users from accessing non-sensitive information within\nsensitive documents. In this paper, we introduce Automated Classification\nEnabled by Security Similarity (ACESS), a new and innovative detection model\nthat penetrates the complexity of big text security classification/detection.\nTo analyze the ACESS system, we constructed a novel dataset, containing\nformerly classified paragraphs from diplomatic cables made public by the\nWikiLeaks organization. To our knowledge this paper is the first to analyze a\ndataset that contains actual formerly sensitive information annotated at\nparagraph granularity.', ""Artificial Neural Networks have reached Grandmaster and even super-human\nperformance across a variety of games: from those involving perfect-information\n(such as Go) to those involving imperfect-information (such as Starcraft). Such\ntechnological developments from AI-labs have ushered concomitant applications\nacross the world of business - where an AI brand tag is fast becoming\nubiquitous. A corollary of such widespread commercial deployment is that when\nAI gets things wrong - an autonomous vehicle crashes; a chatbot exhibits racist\nbehaviour; automated credit scoring processes discriminate on gender etc. -\nthere are often significant financial, legal and brand consequences and the\nincident becomes major news. As Judea Pearl sees it, the underlying reason for\nsuch mistakes is that, 'all the impressive achievements of deep learning amount\nto just curve fitting'. The key, Judea Pearl suggests, is to replace reasoning\nby association with causal-reasoning - the ability to infer causes from\nobserved phenomena. It is a point that was echoed by Gary Marcus and Ernest\nDavis in a recent piece for the New York Times: 'we need to stop building\ncomputer systems that merely get better and better at detecting statistical\npatterns in data sets - often using an approach known as Deep Learning - and\nstart building computer systems that from the moment of their assembly innately\ngrasp three basic concepts: time, space and causality'. In this paper,\nforegrounding what in 1949 Gilbert Ryle termed a category mistake, I will offer\nan alternative explanation for AI errors: it is not so much that AI machinery\ncannot grasp causality, but that AI machinery - qua computation - cannot\nunderstand anything at all."", ""AI has provided us with the ability to automate tasks, extract information\nfrom vast amounts of data, and synthesize media that is nearly\nindistinguishable from the real thing. However, positive tools can also be used\nfor negative purposes. In particular, cyber adversaries can use AI (such as\nmachine learning) to enhance their attacks and expand their campaigns.\n  Although offensive AI has been discussed in the past, there is a need to\nanalyze and understand the threat in the context of organizations. For example,\nhow does an AI-capable adversary impact the cyber kill chain? Does AI benefit\nthe attacker more than the defender? What are the most significant AI threats\nfacing organizations today and what will be their impact on the future?\n  In this survey, we explore the threat of offensive AI on organizations.\nFirst, we present the background and discuss how AI changes the adversary's\nmethods, strategies, goals, and overall attack model. Then, through a\nliterature review, we identify 33 offensive AI capabilities which adversaries\ncan use to enhance their attacks. Finally, through a user study spanning\nindustry and academia, we rank the AI threats and provide insights on the\nadversaries."", 'This work provides a starting point for researchers interested in gaining a\ndeeper understanding of the big picture of artificial intelligence (AI). To\nthis end, a narrative is conveyed that allows the reader to develop an\nobjective view on current developments that is free from false promises that\ndominate public communication. An essential takeaway for the reader is that AI\nmust be understood as an umbrella term encompassing a plethora of different\nmethods, schools of thought, and their respective historical movements.\nConsequently, a bottom-up strategy is pursued in which the field of AI is\nintroduced by presenting various aspects that are characteristic of the\nsubject. This paper is structured in three parts: (i) Discussion of current\ntrends revealing false public narratives, (ii) an introduction to the history\nof AI focusing on recurring patterns and main characteristics, and (iii) a\ncritical discussion on the limitations of current methods in the context of the\npotential emergence of a strong(er) AI. It should be noted that this work does\nnot cover any of these aspects holistically; rather, the content addressed is a\nselection made by the author and subject to a didactic strategy.', 'Recently, many AI researchers and practitioners have embarked on research\nvisions that involve doing AI for ""Good"". This is part of a general drive\ntowards infusing AI research and practice with ethical thinking. One frequent\ntheme in current ethical guidelines is the requirement that AI be good for all,\nor: contribute to the Common Good. But what is the Common Good, and is it\nenough to want to be good? Via four lead questions, I will illustrate\nchallenges and pitfalls when determining, from an AI point of view, what the\nCommon Good is and how it can be enhanced by AI. The questions are: What is the\nproblem / What is a problem?, Who defines the problem?, What is the role of\nknowledge?, and What are important side effects and dynamics? The illustration\nwill use an example from the domain of ""AI for Social Good"", more specifically\n""Data Science for Social Good"". Even if the importance of these questions may\nbe known at an abstract level, they do not get asked sufficiently in practice,\nas shown by an exploratory study of 99 contributions to recent conferences in\nthe field. Turning these challenges and pitfalls into a positive\nrecommendation, as a conclusion I will draw on another characteristic of\ncomputer-science thinking and practice to make these impediments visible and\nattenuate them: ""attacks"" as a method for improving design. This results in the\nproposal of ethics pen-testing as a method for helping AI designs to better\ncontribute to the Common Good.', ""AI researchers employ not only the scientific method, but also methodology\nfrom mathematics and engineering. However, the use of the scientific method -\nspecifically hypothesis testing - in AI is typically conducted in service of\nengineering objectives. Growing interest in topics such as fairness and\nalgorithmic bias show that engineering-focused questions only comprise a subset\nof the important questions about AI systems. This results in the AI Knowledge\nGap: the number of unique AI systems grows faster than the number of studies\nthat characterize these systems' behavior. To close this gap, we argue that the\nstudy of AI could benefit from the greater inclusion of researchers who are\nwell positioned to formulate and test hypotheses about the behavior of AI\nsystems. We examine the barriers preventing social and behavioral scientists\nfrom conducting such studies. Our diagnosis suggests that accelerating the\nscientific study of AI systems requires new incentives for academia and\nindustry, mediated by new tools and institutions. To address these needs, we\npropose a two-sided marketplace called TuringBox. On one side, AI contributors\nupload existing and novel algorithms to be studied scientifically by others. On\nthe other side, AI examiners develop and post machine intelligence tasks\ndesigned to evaluate and characterize algorithmic behavior. We discuss this\nmarket's potential to democratize the scientific study of AI behavior, and thus\nnarrow the AI Knowledge Gap."", 'In this paper I argue that the search for explainable models and\ninterpretable decisions in AI must be reformulated in terms of the broader\nproject of offering a pragmatic and naturalistic account of understanding in\nAI. Intuitively, the purpose of providing an explanation of a model or a\ndecision is to make it understandable to its stakeholders. But without a\nprevious grasp of what it means to say that an agent understands a model or a\ndecision, the explanatory strategies will lack a well-defined goal. Aside from\nproviding a clearer objective for XAI, focusing on understanding also allows us\nto relax the factivity condition on explanation, which is impossible to fulfill\nin many machine learning models, and to focus instead on the pragmatic\nconditions that determine the best fit between a model and the methods and\ndevices deployed to understand it. After an examination of the different types\nof understanding discussed in the philosophical and psychological literature, I\nconclude that interpretative or approximation models not only provide the best\nway to achieve the objectual understanding of a machine learning model, but are\nalso a necessary condition to achieve post-hoc interpretability. This\nconclusion is partly based on the shortcomings of the purely functionalist\napproach to post-hoc interpretability that seems to be predominant in most\nrecent literature.', ""Probability theory, epistemically interpreted, provides an excellent, if not\nthe best available account of inductive reasoning. This is so because there are\ngeneral and definite rules for the change of subjective probabilities through\ninformation or experience; induction and belief change are one and same topic,\nafter all. The most basic of these rules is simply to conditionalize with\nrespect to the information received; and there are similar and more general\nrules. 1 Hence, a fundamental reason for the epistemological success of\nprobability theory is that there at all exists a well-behaved concept of\nconditional probability. Still, people have, and have reasons for, various\nconcerns over probability theory. One of these is my starting point:\nIntuitively, we have the notion of plain belief; we believe propositions2 to be\ntrue (or to be false or neither). Probability theory, however, offers no formal\ncounterpart to this notion. Believing A is not the same as having probability 1\nfor A, because probability 1 is incorrigible3; but plain belief is clearly\ncorrigible. And believing A is not the same as giving A a probability larger\nthan some 1 - c, because believing A and believing B is usually taken to be\nequivalent to believing A & B.4 Thus, it seems that the formal representation\nof plain belief has to take a non-probabilistic route. Indeed, representing\nplain belief seems easy enough: simply represent an epistemic state by the set\nof all propositions believed true in it or, since I make the common assumption\nthat plain belief is deductively closed, by the conjunction of all propositions\nbelieved true in it. But this does not yet provide a theory of induction, i.e.\nan answer to the question how epistemic states so represented are changed\ntbrough information or experience. There is a convincing partial answer: if the\nnew information is compatible with the old epistemic state, then the new\nepistemic state is simply represented by the conjunction of the new information\nand the old beliefs. This answer is partial because it does not cover the quite\ncommon case where the new information is incompatible with the old beliefs. It\nis, however, important to complete the answer and to cover this case, too;\notherwise, we would not represent plain belief as conigible. The crucial\nproblem is that there is no good completion. When epistemic states are\nrepresented simply by the conjunction of all propositions believed true in it,\nthe answer cannot be completed; and though there is a lot of fruitful work, no\nother representation of epistemic states has been proposed, as far as I know,\nwhich provides a complete solution to this problem. In this paper, I want to\nsuggest such a solution. In [4], I have more fully argued that this is the only\nsolution, if certain plausible desiderata are to be satisfied. Here, in section\n2, I will be content with formally defining and intuitively explaining my\nproposal. I will compare my proposal with probability theory in section 3. It\nwill turn out that the theory I am proposing is structurally homomorphic to\nprobability theory in important respects and that it is thus equally easily\nimplementable, but moreover computationally simpler. Section 4 contains a very\nbrief comparison with various kinds of logics, in particular conditional logic,\nwith Shackle's functions of potential surprise and related theories, and with\nthe Dempster - Shafer theory of belief functions."", 'Artificial Intelligence represents many things: a new market to conquer or a\nquality label for tech companies, a threat for traditional industries, a menace\nfor democracy, or a blessing for our busy everyday life. The press abounds in\nexamples illustrating these aspects, but one should draw not hasty and\npremature conclusions. The first successes in AI have been a surprise for\nsociety at large-including researchers in the field. Today, after the initial\nstupefaction, we have examples of the system reactions: traditional companies\nare heavily investing in AI, social platforms are monitored during elections,\ndata collection is more and more regulated, etc. The resilience of an\norganization (i.e. its capacity to resist to a shock) relies deeply on the\nperception of its environment. Future problems have to be anticipated, while\nunforeseen events occurring have to be quickly identified in order to be\nmitigated as fast as possible. The author states that this clear perception\nstarts with a common definition of AI in terms of capacities and limits. AI\npractitioners should make notions and concepts accessible to the general public\nand the impacted fields (e.g. industries, law, education). It is a truism that\nonly law experts would have the potential to estimate IA impacts on judicial\nsystem. However, questions remain on how to connect different kind of expertise\nand what is the appropriate level of detail required for the knowledge\nexchanges. And the same consideration is true for dissemination towards\nsociety. Ultimately, society will live with decisions made by the ""experts"". It\nsounds wise to involve society in the decision process rather than risking to\npay consequences later. Therefore, society also needs the key concepts to\nunderstand AI impact on their life. This was the purpose of the trial of an IA\nthat took place in October 2018 at the Court of Appeal of Paris: gathering\nexperts from various fields to expose challenges in law and science towards a\ngeneral public.', 'The range of application of artificial intelligence (AI) is vast, as is the\npotential for harm. Growing awareness of potential risks from AI systems has\nspurred action to address those risks, while eroding confidence in AI systems\nand the organizations that develop them. A 2019 study found over 80\norganizations that published and adopted ""AI ethics principles\'\', and more have\njoined since. But the principles often leave a gap between the ""what"" and the\n""how"" of trustworthy AI development. Such gaps have enabled questionable or\nethically dubious behavior, which casts doubts on the trustworthiness of\nspecific organizations, and the field more broadly. There is thus an urgent\nneed for concrete methods that both enable AI developers to prevent harm and\nallow them to demonstrate their trustworthiness through verifiable behavior.\nBelow, we explore mechanisms (drawn from arXiv:2004.07213) for creating an\necosystem where AI developers can earn trust - if they are trustworthy. Better\nassessment of developer trustworthiness could inform user choice, employee\nactions, investment decisions, legal recourse, and emerging governance regimes.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I aim to assist actors in better comprehending their roles in movies or TV shows by creating a system that can comprehend and analyze the mood and personality of characters within a narrative context. My plan is to utilize transformer-based models in conjunction with a reinforcement learning (RL) framework to accomplish this. The transformer-based model should be capable of inferring from the context and dialogue to predict a character's potential mood and anticipate their next line. The RL framework will then compare the predicted conversation with the actual conversation in the reward formula to provide weak self-supervision during the training process. I am aware of the concept of a warm start in training, but I am unsure how to adapt this to my current project. Furthermore, the target narrative contexts, such as scripts, differ from typical texts, so I would appreciate advice on the appropriate length for the context window in this type of text. Finally, the model should be able to map mood or personality to facial expressions or character appearances to provide guidance to the actors.","['Recent neural models of dialogue generation offer great promise for\ngenerating responses for conversational agents, but tend to be shortsighted,\npredicting utterances one at a time while ignoring their influence on future\noutcomes. Modeling the future direction of a dialogue is crucial to generating\ncoherent, interesting dialogues, a need which led traditional NLP models of\ndialogue to draw on reinforcement learning. In this paper, we show how to\nintegrate these goals, applying deep reinforcement learning to model future\nreward in chatbot dialogue. The model simulates dialogues between two virtual\nagents, using policy gradient methods to reward sequences that display three\nuseful conversational properties: informativity (non-repetitive turns),\ncoherence, and ease of answering (related to forward-looking function). We\nevaluate our model on diversity, length as well as with human judges, showing\nthat the proposed algorithm generates more interactive responses and manages to\nfoster a more sustained conversation in dialogue simulation. This work marks a\nfirst step towards learning a neural conversational model based on the\nlong-term success of dialogues.', ""A conversational agent (chatbot) is a piece of software that is able to\ncommunicate with humans using natural language. Modeling conversation is an\nimportant task in natural language processing and artificial intelligence.\nWhile chatbots can be used for various tasks, in general they have to\nunderstand users' utterances and provide responses that are relevant to the\nproblem at hand.\n  In my work, I conduct an in-depth survey of recent literature, examining over\n70 publications related to chatbots published in the last 3 years. Then, I\nproceed to make the argument that the very nature of the general conversation\ndomain demands approaches that are different from current state-of-of-the-art\narchitectures. Based on several examples from the literature I show why current\nchatbot models fail to take into account enough priors when generating\nresponses and how this affects the quality of the conversation. In the case of\nchatbots, these priors can be outside sources of information that the\nconversation is conditioned on like the persona or mood of the conversers. In\naddition to presenting the reasons behind this problem, I propose several ideas\non how it could be remedied.\n  The next section focuses on adapting the very recent Transformer model to the\nchatbot domain, which is currently state-of-the-art in neural machine\ntranslation. I first present experiments with the vanilla model, using\nconversations extracted from the Cornell Movie-Dialog Corpus. Secondly, I\naugment the model with some of my ideas regarding the issues of encoder-decoder\narchitectures. More specifically, I feed additional features into the model\nlike mood or persona together with the raw conversation data. Finally, I\nconduct a detailed analysis of how the vanilla model performs on conversational\ndata by comparing it to previous chatbot models and how the additional features\naffect the quality of the generated responses."", ""Neural models for response generation produce responses that are semantically\nplausible but not necessarily factually consistent with facts describing the\nspeaker's persona. These models are trained with fully supervised learning\nwhere the objective function barely captures factual consistency. We propose to\nfine-tune these models by reinforcement learning and an efficient reward\nfunction that explicitly captures the consistency between a response and\npersona facts as well as semantic plausibility. Our automatic and human\nevaluations on the PersonaChat corpus confirm that our approach increases the\nrate of responses that are factually consistent with persona facts over its\nsupervised counterpart while retaining the language quality of responses.""]","['Standard neural machine translation (NMT) is on the assumption of\ndocument-level context independent. Most existing document-level NMT methods\nare satisfied with a smattering sense of brief document-level information,\nwhile this work focuses on exploiting detailed document-level context in terms\nof multiple forms of document embeddings, which is capable of sufficiently\nmodeling deeper and richer document-level context. The proposed document-aware\nNMT is implemented to enhance the Transformer baseline by introducing both\nglobal and local document-level clues on the source end. Experiments show that\nthe proposed method significantly improves the translation performance over\nstrong baselines and other related studies.', 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', ""The successful emotional conversation system depends on sufficient perception\nand appropriate expression of emotions. In a real-life conversation, humans\nfirstly instinctively perceive emotions from multi-source information,\nincluding the emotion flow hidden in dialogue history, facial expressions,\naudio, and personalities of speakers. Then, they convey suitable emotions\naccording to their personalities, but these multiple types of information are\ninsufficiently exploited in emotional conversation fields. To address this\nissue, in this paper, we propose a heterogeneous graph-based model for\nemotional conversation generation. Firstly, we design a Heterogeneous\nGraph-Based Encoder to represent the conversation content (i.e., the dialogue\nhistory, its emotion flow, facial expressions, audio, and speakers'\npersonalities) with a heterogeneous graph neural network, and then predict\nsuitable emotions for feedback. Secondly, we employ an\nEmotion-Personality-Aware Decoder to generate a response relevant to the\nconversation context as well as with appropriate emotions, through taking the\nencoded graph representations, the predicted emotions by the encoder and the\npersonality of the current speaker as inputs. Experiments on both automatic and\nhuman evaluation show that our method can effectively perceive emotions from\nmulti-source knowledge and generate a satisfactory response. Furthermore, based\non the up-to-date text generator BART, our model still can achieve consistent\nimprovement, which significantly outperforms some existing state-of-the-art\nmodels."", 'Neural networks have recently become good at engaging in dialog. However,\ncurrent approaches are based solely on verbal text, lacking the richness of a\nreal face-to-face conversation. We propose a neural conversation model that\naims to read and generate facial gestures alongside with text. This allows our\nmodel to adapt its response based on the ""mood"" of the conversation. In\nparticular, we introduce an RNN encoder-decoder that exploits the movement of\nfacial muscles, as well as the verbal conversation. The decoder consists of two\nlayers, where the lower layer aims at generating the verbal response and coarse\nfacial expressions, while the second layer fills in the subtle gestures, making\nthe generated output more smooth and natural. We train our neural network by\nhaving it ""watch"" 250 movies. We showcase our joint face-text model in\ngenerating more natural conversations through automatic metrics and a human\nstudy. We demonstrate an example application with a face-to-face chatting\navatar.', ""An important aspect of human conversation difficult for machines is\nconversing with empathy, which is to understand the user's emotion and respond\nappropriately. Recent neural conversation models that attempted to generate\nempathetic responses either focused on conditioning the output to a given\nemotion, or incorporating the current user emotional state. However, these\napproaches do not factor in how the user would feel towards the generated\nresponse. Hence, in this paper, we propose Sentiment Look-ahead, which is a\nnovel perspective for empathy that models the future user emotional state. In\nshort, Sentiment Look-ahead is a reward function under a reinforcement learning\nframework that provides a higher reward to the generative model when the\ngenerated utterance improves the user's sentiment. We implement and evaluate\nthree different possible implementations of sentiment look-ahead and\nempirically show that our proposed approach can generate significantly more\nempathetic, relevant, and fluent responses than other competitive baselines\nsuch as multitask learning."", 'People speak at different levels of specificity in different situations.\nDepending on their knowledge, interlocutors, mood, etc.} A conversational agent\nshould have this ability and know when to be specific and when to be general.\nWe propose an approach that gives a neural network--based conversational agent\nthis ability. Our approach involves alternating between \\emph{data\ndistillation} and model training : removing training examples that are closest\nto the responses most commonly produced by the model trained from the last\nround and then retrain the model on the remaining dataset. Dialogue generation\nmodels trained with different degrees of data distillation manifest different\nlevels of specificity.\n  We then train a reinforcement learning system for selecting among this pool\nof generation models, to choose the best level of specificity for a given\ninput. Compared to the original generative model trained without distillation,\nthe proposed system is capable of generating more interesting and\nhigher-quality responses, in addition to appropriately adjusting specificity\ndepending on the context.\n  Our research constitutes a specific case of a broader approach involving\ntraining multiple subsystems from a single dataset distinguished by differences\nin a specific property one wishes to model. We show that from such a set of\nsubsystems, one can use reinforcement learning to build a system that tailors\nits output to different input contexts at test time.', 'To provide consistent emotional interaction with users, dialog systems should\nbe capable to automatically select appropriate emotions for responses like\nhumans. However, most existing works focus on rendering specified emotions in\nresponses or empathetically respond to the emotion of users, yet the individual\ndifference in emotion expression is overlooked. This may lead to inconsistent\nemotional expressions and disinterest users. To tackle this issue, we propose\nto equip the dialog system with personality and enable it to automatically\nselect emotions in responses by simulating the emotion transition of humans in\nconversation. In detail, the emotion of the dialog system is transitioned from\nits preceding emotion in context. The transition is triggered by the preceding\ndialog context and affected by the specified personality trait. To achieve\nthis, we first model the emotion transition in the dialog system as the\nvariation between the preceding emotion and the response emotion in the\nValence-Arousal-Dominance (VAD) emotion space. Then, we design neural networks\nto encode the preceding dialog context and the specified personality traits to\ncompose the variation. Finally, the emotion for response is selected from the\nsum of the preceding emotion and the variation. We construct a dialog dataset\nwith emotion and personality labels and conduct emotion prediction tasks for\nevaluation. Experimental results validate the effectiveness of the\npersonality-affected emotion transition.', ""For conversational AI and virtual assistants to communicate with humans in a\nrealistic way, they must exhibit human characteristics such as expression of\nemotion and personality. Current attempts toward constructing human-like\ndialogue agents have presented significant difficulties. We propose Human Level\nAttributes (HLAs) based on tropes as the basis of a method for learning\ndialogue agents that can imitate the personalities of fictional characters.\nTropes are characteristics of fictional personalities that are observed\nrecurrently and determined by viewers' impressions. By combining detailed HLA\ndata with dialogue data for specific characters, we present a dataset,\nHLA-Chat, that models character profiles and gives dialogue agents the ability\nto learn characters' language styles through their HLAs. We then introduce a\nthree-component system, ALOHA (which stands for Artificial Learning of Human\nAttributes), that combines character space mapping, character community\ndetection, and language style retrieval to build a character (or personality)\nspecific language model. Our preliminary experiments demonstrate that two\nvariations of ALOHA, combined with our proposed dataset, can outperform\nbaseline models at identifying the correct dialogue responses of chosen target\ncharacters, and are stable regardless of the character's identity, the genre of\nthe show, and the context of the dialogue."", 'Language systems have been of great interest to the research community and\nhave recently reached the mass market through various assistant platforms on\nthe web. Reinforcement Learning methods that optimize dialogue policies have\nseen successes in past years and have recently been extended into methods that\npersonalize the dialogue, e.g. take the personal context of users into account.\nThese works, however, are limited to personalization to a single user with whom\nthey require multiple interactions and do not generalize the usage of context\nacross users. This work introduces a problem where a generalized usage of\ncontext is relevant and proposes two Reinforcement Learning (RL)-based\napproaches to this problem. The first approach uses a single learner and\nextends the traditional POMDP formulation of dialogue state with features that\ndescribe the user context. The second approach segments users by context and\nthen employs a learner per context. We compare these approaches in a benchmark\nof existing non-RL and RL-based methods in three established and one novel\napplication domain of financial product recommendation. We compare the\ninfluence of context and training experiences on performance and find that\nlearning approaches generally outperform a handcrafted gold standard.', ""Understanding narrative text requires capturing characters' motivations,\ngoals, and mental states. This paper proposes an Entity-based Narrative Graph\n(ENG) to model the internal-states of characters in a story. We explicitly\nmodel entities, their interactions and the context in which they appear, and\nlearn rich representations for them. We experiment with different task-adaptive\npre-training objectives, in-domain training, and symbolic inference to capture\ndependencies between different decisions in the output space. We evaluate our\nmodel on two narrative understanding tasks: predicting character mental states,\nand desire fulfillment, and conduct a qualitative analysis."", 'Open-domain dialog generation is a challenging problem; maximum likelihood\ntraining can lead to repetitive outputs, models have difficulty tracking\nlong-term conversational goals, and training on standard movie or online\ndatasets may lead to the generation of inappropriate, biased, or offensive\ntext. Reinforcement Learning (RL) is a powerful framework that could\npotentially address these issues, for example by allowing a dialog model to\noptimize for reducing toxicity and repetitiveness. However, previous approaches\nwhich apply RL to open-domain dialog generation do so at the word level, making\nit difficult for the model to learn proper credit assignment for long-term\nconversational rewards. In this paper, we propose a novel approach to\nhierarchical reinforcement learning, VHRL, which uses policy gradients to tune\nthe utterance-level embedding of a variational sequence model. This\nhierarchical approach provides greater flexibility for learning long-term,\nconversational rewards. We use self-play and RL to optimize for a set of\nhuman-centered conversation metrics, and show that our approach provides\nsignificant improvements -- in terms of both human evaluation and automatic\nmetrics -- over state-of-the-art dialog models, including Transformers.', 'The use of virtual agents in social coaching has increased rapidly in the\nlast decade. In order to train the user in different situations than can occur\nin real life, the virtual agent should be able to express different social\nattitudes. In this paper, we propose a model of social attitudes that enables a\nvirtual agent to reason on the appropriate social attitude to express during\nthe interaction with a user given the course of the interaction, but also the\nemotions, mood and personality of the agent. Moreover, the model enables the\nvirtual agent to display its social attitude through its non-verbal behaviour.\nThe proposed model has been developed in the context of job interview\nsimulation. The methodology used to develop such a model combined a theoretical\nand an empirical approach. Indeed, the model is based both on the literature in\nHuman and Social Sciences on social attitudes but also on the analysis of an\naudiovisual corpus of job interviews and on post-hoc interviews with the\nrecruiters on their expressed attitudes during the job interview.', ""Explainable AI techniques that describe agent reward functions can enhance\nhuman-robot collaboration in a variety of settings. One context where human\nunderstanding of agent reward functions is particularly beneficial is in the\nvalue alignment setting. In the value alignment context, an agent aims to infer\na human's reward function through interaction so that it can assist the human\nwith their tasks. If the human can understand where gaps exist in the agent's\nreward understanding, they will be able to teach more efficiently and\neffectively, leading to quicker human-agent team performance improvements. In\norder to support human collaborators in the value alignment setting and similar\ncontexts, it is first important to understand the effectiveness of different\nreward explanation techniques in a variety of domains. In this paper, we\nintroduce a categorization of information modalities for reward explanation\ntechniques, suggest a suite of assessment techniques for human reward\nunderstanding, and introduce four axes of domain complexity. We then propose an\nexperiment to study the relative efficacy of a broad set of reward explanation\ntechniques covering multiple modalities of information in a set of domains of\nvarying complexity."", ""One obstacle to applying reinforcement learning algorithms to real-world\nproblems is the lack of suitable reward functions. Designing such reward\nfunctions is difficult in part because the user only has an implicit\nunderstanding of the task objective. This gives rise to the agent alignment\nproblem: how do we create agents that behave in accordance with the user's\nintentions? We outline a high-level research direction to solve the agent\nalignment problem centered around reward modeling: learning a reward function\nfrom interaction with the user and optimizing the learned reward function with\nreinforcement learning. We discuss the key challenges we expect to face when\nscaling reward modeling to complex and general domains, concrete approaches to\nmitigate these challenges, and ways to establish trust in the resulting agents."", ""Automated prediction of valence, one key feature of a person's emotional\nstate, from individuals' personal narratives may provide crucial information\nfor mental healthcare (e.g. early diagnosis of mental diseases, supervision of\ndisease course, etc.). In the Interspeech 2018 ComParE Self-Assessed Affect\nchallenge, the task of valence prediction was framed as a three-class\nclassification problem using 8 seconds fragments from individuals' narratives.\nAs such, the task did not allow for exploring contextual information of the\nnarratives. In this work, we investigate the intrinsic information from\nmultiple narratives recounted by the same individual in order to predict their\ncurrent state-of-mind. Furthermore, with generalizability in mind, we decided\nto focus our experiments exclusively on textual information as the public\navailability of audio narratives is limited compared to text. Our hypothesis\nis, that context modeling might provide insights about emotion triggering\nconcepts (e.g. events, people, places) mentioned in the narratives that are\nlinked to an individual's state of mind. We explore multiple machine learning\ntechniques to model narratives. We find that the models are able to capture\ninter-individual differences, leading to more accurate predictions of an\nindividual's emotional state, as compared to single narratives."", 'The high request for autonomous and flexible HRI implies the necessity of\ndeploying Machine Learning (ML) mechanisms in the robot control. Indeed, the\nuse of ML techniques, such as Reinforcement Learning (RL), makes the robot\nbehaviour, during the learning process, not transparent to the observing user.\nIn this work, we proposed an emotional model to improve the transparency in RL\ntasks for human-robot collaborative scenarios. The architecture we propose\nsupports the RL algorithm with an emotional model able to both receive human\nfeedback and exhibit emotional responses based on the learning process. The\nmodel is entirely based on the Temporal Difference (TD) error. The architecture\nwas tested in an isolated laboratory with a simple setup. The results highlight\nthat showing its internal state through an emotional response is enough to make\na robot transparent to its human teacher. People also prefer to interact with a\nresponsive robot because they are used to understand their intentions via\nemotions and social signals.', 'Human communication is often executed in the form of a narrative, an account\nof connected events composed of characters, actions, and settings. A coherent\nnarrative structure is therefore a requisite for a well-formulated narrative --\nbe it fictional or nonfictional -- for informative and effective communication,\nopening up the possibility of a deeper understanding of a narrative by studying\nits structural properties. In this paper we present a network-based framework\nfor modeling and analyzing the structure of a narrative, which is further\nexpanded by incorporating methods from computational linguistics to utilize the\nnarrative text. Modeling a narrative as a dynamically unfolding system, we\ncharacterize its progression via the growth patterns of the character network,\nand use sentiment analysis and topic modeling to represent the actual content\nof the narrative in the form of interaction maps between characters with\nassociated sentiment values and keywords. This is a network framework advanced\nbeyond the simple occurrence-based one most often used until now, allowing one\nto utilize the unique characteristics of a given narrative to a high degree.\nGiven the ubiquity and importance of narratives, such advanced network-based\nrepresentation and analysis framework may lead to a more systematic modeling\nand understanding of narratives for social interactions, expression of human\nsentiments, and communication.', 'Grounding human-machine conversation in a document is an effective way to\nimprove the performance of retrieval-based chatbots. However, only a part of\nthe document content may be relevant to help select the appropriate response at\na round. It is thus crucial to select the part of document content relevant to\nthe current conversation context. In this paper, we propose a document content\nselection network (CSN) to perform explicit selection of relevant document\ncontents, and filter out the irrelevant parts. We show in experiments on two\npublic document-grounded conversation datasets that CSN can effectively help\nselect the relevant document contents to the conversation context, and it\nproduces better results than the state-of-the-art approaches. Our code and\ndatasets are available at https://github.com/DaoD/CSN.', 'An emerging recipe for achieving state-of-the-art effectiveness in neural\ndocument re-ranking involves utilizing large pre-trained language models -\ne.g., BERT - to evaluate all individual passages in the document and then\naggregating the outputs by pooling or additional Transformer layers. A major\ndrawback of this approach is high query latency due to the cost of evaluating\nevery passage in the document with BERT. To make matters worse, this high\ninference cost and latency varies based on the length of the document, with\nlonger documents requiring more time and computation. To address this\nchallenge, we adopt an intra-document cascading strategy, which prunes passages\nof a candidate document using a less expensive model, called ESM, before\nrunning a scoring model that is more expensive and effective, called ETM. We\nfound it best to train ESM (short for Efficient Student Model) via knowledge\ndistillation from the ETM (short for Effective Teacher Model) e.g., BERT. This\npruning allows us to only run the ETM model on a smaller set of passages whose\nsize does not vary by document length. Our experiments on the MS MARCO and TREC\nDeep Learning Track benchmarks suggest that the proposed Intra-Document\nCascaded Ranking Model (IDCM) leads to over 400% lower query latency by\nproviding essentially the same effectiveness as the state-of-the-art BERT-based\ndocument ranking models.', 'End-to-End intelligent neural dialogue systems suffer from the problems of\ngenerating inconsistent and repetitive responses. Existing dialogue models pay\nattention to unilaterally incorporating personal knowledge into the dialog\nwhile ignoring the fact that incorporating the personality-related conversation\ninformation into personal knowledge taken as the bilateral information flow\nboosts the quality of the subsequent conversation. Besides, it is indispensable\nto control personal knowledge utilization over the conversation level. In this\npaper, we propose a conversation-adaption multi-view persona aware response\ngeneration model that aims at enhancing conversation consistency and\nalleviating the repetition from two folds. First, we consider conversation\nconsistency from multiple views. From the view of the persona profile, we\ndesign a novel interaction module that not only iteratively incorporates\npersonalized knowledge into each turn conversation but also captures the\npersonality-related information from conversation to enhance personalized\nknowledge semantic representation. From the view of speaking style, we\nintroduce the speaking style vector and feed it into the decoder to keep the\nspeaking style consistency. To avoid conversation repetition, we devise a\ncoverage mechanism to keep track of the activation of personal knowledge\nutilization. Experiments on both automatic and human evaluation verify the\nsuperiority of our model over previous models.', ""Mental State Transition Network (MSTN) is a basic concept of approximating to\nhuman psychological and mental responses. A stimulus calculated by Emotion\nGenerating Calculations (EGC) method can cause the transition of mood from an\nemotional state to others. In this paper, the agent can interact with human to\nrealize smooth communication by an adaptive learning method of the user's\npersonality trait based mood. The learning method consists of the profit\nsharing (PS) method and the recurrent neural network (RNN). An emotion for\nsensor inputs to MSTN is calculated by EGC and the variance of emotion leads to\nthe change of mental state, and then the sequence of states forms an episode.\nIn order to learn the tendency of personality trait effectively, the\nineffective rules should be removed from the episode. PS method finds out a\ndetour in episode and should be deleted. Furthermore, RNN works to realize the\nvariance of user's mood. Some experimental results were shown the success of\nrepresenting a various human's delicate emotion."", 'Hierarchical neural architectures are often used to capture long-distance\ndependencies and have been applied to many document-level tasks such as\nsummarization, document segmentation, and sentiment analysis. However,\neffective usage of such a large context can be difficult to learn, especially\nin the case where there is limited labeled data available. Building on the\nrecent success of language model pretraining methods for learning flat\nrepresentations of text, we propose algorithms for pre-training hierarchical\ndocument representations from unlabeled data. Unlike prior work, which has\nfocused on pre-training contextual token representations or context-independent\n{sentence/paragraph} representations, our hierarchical document representations\ninclude fixed-length sentence/paragraph representations which integrate\ncontextual information from the entire documents. Experiments on document\nsegmentation, document-level question answering, and extractive document\nsummarization demonstrate the effectiveness of the proposed pre-training\nalgorithms.', ""World models improve a learning agent's ability to efficiently operate in\ninteractive and situated environments. This work focuses on the task of\nbuilding world models of text-based game environments. Text-based games, or\ninteractive narratives, are reinforcement learning environments in which agents\nperceive and interact with the world using textual natural language. These\nenvironments contain long, multi-step puzzles or quests woven through a world\nthat is filled with hundreds of characters, locations, and objects. Our world\nmodel learns to simultaneously: (1) predict changes in the world caused by an\nagent's actions when representing the world as a knowledge graph; and (2)\ngenerate the set of contextually relevant natural language actions required to\noperate in the world. We frame this task as a Set of Sequences generation\nproblem by exploiting the inherent structure of knowledge graphs and actions\nand introduce both a transformer-based multi-task architecture and a loss\nfunction to train it. A zero-shot ablation study on never-before-seen textual\nworlds shows that our methodology significantly outperforms existing textual\nworld modeling techniques as well as the importance of each of our\ncontributions."", 'To build an open-domain multi-turn conversation system is one of the most\ninteresting and challenging tasks in Artificial Intelligence. Many research\nefforts have been dedicated to building such dialogue systems, yet few shed\nlight on modeling the conversation flow in an ongoing dialogue. Besides, it is\ncommon for people to talk about highly relevant aspects during a conversation.\nAnd the topics are coherent and drift naturally, which demonstrates the\nnecessity of dialogue flow modeling. To this end, we present the multi-turn\ncue-words driven conversation system with reinforcement learning method (RLCw),\nwhich strives to select an adaptive cue word with the greatest future credit,\nand therefore improve the quality of generated responses. We introduce a new\nreward to measure the quality of cue words in terms of effectiveness and\nrelevance. To further optimize the model for long-term conversations, a\nreinforcement approach is adopted in this paper. Experiments on real-life\ndataset demonstrate that our model consistently outperforms a set of\ncompetitive baselines in terms of simulated turns, diversity and human\nevaluation.', 'In this paper we consider the problem of how a reinforcement learning agent\nthat is tasked with solving a sequence of reinforcement learning problems (a\nsequence of Markov decision processes) can use knowledge acquired early in its\nlifetime to improve its ability to solve new problems. We argue that previous\nexperience with similar problems can provide an agent with information about\nhow it should explore when facing a new but related problem. We show that the\nsearch for an optimal exploration strategy can be formulated as a reinforcement\nlearning problem itself and demonstrate that such strategy can leverage\npatterns found in the structure of related problems. We conclude with\nexperiments that show the benefits of optimizing an exploration strategy using\nour proposed approach.', 'How to leverage cross-document interactions to improve ranking performance is\nan important topic in information retrieval (IR) research. However, this topic\nhas not been well-studied in the learning-to-rank setting and most of the\nexisting work still treats each document independently while scoring. The\nrecent development of deep learning shows strength in modeling complex\nrelationships across sequences and sets. It thus motivates us to study how to\nleverage cross-document interactions for learning-to-rank in the deep learning\nframework. In this paper, we formally define the permutation-equivariance\nrequirement for a scoring function that captures cross-document interactions.\nWe then propose a self-attention based document interaction network and show\nthat it satisfies the permutation-equivariant requirement, and can generate\nscores for document sets of varying sizes. Our proposed methods can\nautomatically learn to capture document interactions without any auxiliary\ninformation, and can scale across large document sets. We conduct experiments\non three ranking datasets: the benchmark Web30k, a Gmail search, and a Google\nDrive Quick Access dataset. Experimental results show that our proposed methods\nare both more effective and efficient than baselines.', 'We present two deep learning approaches to narrative text understanding for\ncharacter relationship modelling. The temporal evolution of these relations is\ndescribed by dynamic word embeddings, that are designed to learn semantic\nchanges over time. An empirical analysis of the corresponding character\ntrajectories shows that such approaches are effective in depicting dynamic\nevolution. A supervised learning approach based on the state-of-the-art\ntransformer model BERT is used instead to detect static relations between\ncharacters. The empirical validation shows that such events (e.g., two\ncharacters belonging to the same family) might be spotted with good accuracy,\neven when using automatically annotated data. This provides a deeper\nunderstanding of narrative plots based on the identification of key facts.\nStandard clustering techniques are finally used for character de-aliasing, a\nnecessary pre-processing step for both approaches. Overall, deep learning\nmodels appear to be suitable for narrative text understanding, while also\nproviding a challenging and unexploited benchmark for general natural language\nunderstanding.', 'Dialogue authoring in large games requires not only content creation but the\nsubtlety of its delivery, which can vary from character to character. Manually\nauthoring this dialogue can be tedious, time-consuming, or even altogether\ninfeasible. This paper utilizes a rich narrative representation for modeling\ndialogue and an expressive natural language generation engine for realizing it,\nand expands upon a translation tool that bridges the two. We add functionality\nto the translator to allow direct speech to be modeled by the narrative\nrepresentation, whereas the original translator supports only narratives told\nby a third person narrator. We show that we can perform character substitution\nin dialogues. We implement and evaluate a potential application to dialogue\nimplementation: generating dialogue for games with big, dynamic, or\nprocedurally-generated open worlds. We present a pilot study on human\nperceptions of the personalities of characters using direct speech, assuming\nunknown personality types at the time of authoring.', 'We propose Object-oriented Neural Programming (OONP), a framework for\nsemantically parsing documents in specific domains. Basically, OONP reads a\ndocument and parses it into a predesigned object-oriented data structure\n(referred to as ontology in this paper) that reflects the domain-specific\nsemantics of the document. An OONP parser models semantic parsing as a decision\nprocess: a neural net-based Reader sequentially goes through the document, and\nduring the process it builds and updates an intermediate ontology to summarize\nits partial understanding of the text it covers. OONP supports a rich family of\noperations (both symbolic and differentiable) for composing the ontology, and a\nbig variety of forms (both symbolic and differentiable) for representing the\nstate and the document. An OONP parser can be trained with supervision of\ndifferent forms and strength, including supervised learning (SL) ,\nreinforcement learning (RL) and hybrid of the two. Our experiments on both\nsynthetic and real-world document parsing tasks have shown that OONP can learn\nto handle fairly complicated ontology with training data of modest sizes.', 'It is appealing to have a system that generates a story or scripts\nautomatically from a story-line, even though this is still out of our reach. In\ndialogue systems, it would also be useful to drive dialogues by a dialogue\nplan. In this paper, we address a key problem involved in these applications --\nguiding a dialogue by a narrative. The proposed model ScriptWriter selects the\nbest response among the candidates that fit the context as well as the given\nnarrative. It keeps track of what in the narrative has been said and what is to\nbe said. A narrative plays a different role than the context (i.e., previous\nutterances), which is generally used in current dialogue systems. Due to the\nunavailability of data for this new application, we construct a new large-scale\ndata collection GraphMovie from a movie website where end-users can upload\ntheir narratives freely when watching a movie. Experimental results on the\ndataset show that our proposed approach based on narratives significantly\noutperforms the baselines that simply use the narrative as a kind of context.', ""Neural summarization models suffer from the fixed-size input limitation: if\ntext length surpasses the model's maximal number of input tokens, some document\ncontent (possibly summary-relevant) gets truncated Independently summarizing\nwindows of maximal input size disallows for information flow between windows\nand leads to incoherent summaries. We propose windowing models for neural\nabstractive summarization of (arbitrarily) long texts. We extend the\nsequence-to-sequence model augmented with pointer generator network by (1)\nallowing the encoder to slide over different windows of the input document and\n(2) sharing the decoder and retaining its state across different input windows.\nWe explore two windowing variants: Static Windowing precomputes the number of\ntokens the decoder should generate from each window (based on training corpus\nstatistics); in Dynamic Windowing the decoder learns to emit a token that\nsignals encoder's shift to the next input window. Empirical results render our\nmodels effective in their intended use-case: summarizing long texts with\nrelevant content not bound to the very document beginning."", 'The use of conversational assistants to search for information is becoming\nincreasingly more popular among the general public, pushing the research\ntowards more advanced and sophisticated techniques. In the last few years, in\nparticular, the interest in conversational search is increasing, not only\nbecause of the generalization of conversational assistants but also because\nconversational search is a step forward in allowing a more natural interaction\nwith the system.\n  In this work, the focus is on exploring the context present of the\nconversation via the historical utterances and respective embeddings with the\naim of developing a conversational search system that helps people search for\ninformation in a natural way. In particular, this system must be able to\nunderstand the context where the question is posed, tracking the current state\nof the conversation and detecting mentions to previous questions and answers.\nWe achieve this by using a context-tracking component based on neural\nquery-rewriting models. Another crucial aspect of the system is to provide the\nmost relevant answers given the question and the conversational history. To\nachieve this objective, we used a Transformer-based re-ranking method and\nexpanded this architecture to use the conversational context.\n  The results obtained with the system developed showed the advantages of using\nthe context present in the natural language utterances and in the neural\nembeddings generated throughout the conversation.', 'Reinforcement learning optimizes policies for expected cumulative reward.\nNeed the supervision be so narrow? Reward is delayed and sparse for many tasks,\nmaking it a difficult and impoverished signal for end-to-end optimization. To\naugment reward, we consider a range of self-supervised tasks that incorporate\nstates, actions, and successors to provide auxiliary losses. These losses offer\nubiquitous and instantaneous supervision for representation learning even in\nthe absence of reward. While current results show that learning from reward\nalone is feasible, pure reinforcement learning methods are constrained by\ncomputational and data efficiency issues that can be remedied by auxiliary\nlosses. Self-supervised pre-training and joint optimization improve the data\nefficiency and policy returns of end-to-end reinforcement learning.', 'Recent progress in pretrained Transformer-based language models has shown\ngreat success in learning contextual representation of text. However, due to\nthe quadratic self-attention complexity, most of the pretrained Transformers\nmodels can only handle relatively short text. It is still a challenge when it\ncomes to modeling very long documents. In this work, we propose to use a graph\nattention network on top of the available pretrained Transformers model to\nlearn document embeddings. This graph attention network allows us to leverage\nthe high-level semantic structure of the document. In addition, based on our\ngraph document model, we design a simple contrastive learning strategy to\npretrain our models on a large amount of unlabeled corpus. Empirically, we\ndemonstrate the effectiveness of our approaches in document classification and\ndocument retrieval tasks.', ""Neural conversation models are known to generate appropriate but\nnon-informative responses in general. A scenario where informativeness can be\nsignificantly enhanced is Conversing by Reading (CbR), where conversations take\nplace with respect to a given external document. In previous work, the external\ndocument is utilized by (1) creating a context-aware document memory that\nintegrates information from the document and the conversational context, and\nthen (2) generating responses referring to the memory. In this paper, we\npropose to create the document memory with some anticipated responses in mind.\nThis is achieved using a teacher-student framework. The teacher is given the\nexternal document, the context, and the ground-truth response, and learns how\nto build a response-aware document memory from three sources of information.\nThe student learns to construct a response-anticipated document memory from the\nfirst two sources, and the teacher's insight on memory creation. Empirical\nresults show that our model outperforms the previous state-of-the-art for the\nCbR task."", 'In this work, we propose a novel method for training neural networks to\nperform single-document extractive summarization without\nheuristically-generated extractive labels. We call our approach BanditSum as it\ntreats extractive summarization as a contextual bandit (CB) problem, where the\nmodel receives a document to summarize (the context), and chooses a sequence of\nsentences to include in the summary (the action). A policy gradient\nreinforcement learning algorithm is used to train the model to select sequences\nof sentences that maximize ROUGE score. We perform a series of experiments\ndemonstrating that BanditSum is able to achieve ROUGE scores that are better\nthan or comparable to the state-of-the-art for extractive summarization, and\nconverges using significantly fewer update steps than competing approaches. In\naddition, we show empirically that BanditSum performs significantly better than\ncompeting approaches when good summary sentences appear late in the source\ndocument.', ""No real-world reward function is perfect. Sensory errors and software bugs\nmay result in RL agents observing higher (or lower) rewards than they should.\nFor example, a reinforcement learning agent may prefer states where a sensory\nerror gives it the maximum reward, but where the true reward is actually small.\nWe formalise this problem as a generalised Markov Decision Problem called\nCorrupt Reward MDP. Traditional RL methods fare poorly in CRMDPs, even under\nstrong simplifying assumptions and when trying to compensate for the possibly\ncorrupt rewards. Two ways around the problem are investigated. First, by giving\nthe agent richer data, such as in inverse reinforcement learning and\nsemi-supervised reinforcement learning, reward corruption stemming from\nsystematic sensory errors may sometimes be completely managed. Second, by using\nrandomisation to blunt the agent's optimisation, reward corruption can be\npartially managed under some assumptions."", 'Bayesian reward learning from demonstrations enables rigorous safety and\nuncertainty analysis when performing imitation learning. However, Bayesian\nreward learning methods are typically computationally intractable for complex\ncontrol problems. We propose Bayesian Reward Extrapolation (Bayesian REX), a\nhighly efficient Bayesian reward learning algorithm that scales to\nhigh-dimensional imitation learning problems by pre-training a low-dimensional\nfeature encoding via self-supervised tasks and then leveraging preferences over\ndemonstrations to perform fast Bayesian inference. Bayesian REX can learn to\nplay Atari games from demonstrations, without access to the game score and can\ngenerate 100,000 samples from the posterior over reward functions in only 5\nminutes on a personal laptop. Bayesian REX also results in imitation learning\nperformance that is competitive with or better than state-of-the-art methods\nthat only learn point estimates of the reward function. Finally, Bayesian REX\nenables efficient high-confidence policy evaluation without having access to\nsamples of the reward function. These high-confidence performance bounds can be\nused to rank the performance and risk of a variety of evaluation policies and\nprovide a way to detect reward hacking behaviors.', 'When a person is not satisfied with how a robot performs a task, they can\nintervene to correct it. Reward learning methods enable the robot to adapt its\nreward function online based on such human input, but they rely on handcrafted\nfeatures. When the correction cannot be explained by these features, recent\nwork in deep Inverse Reinforcement Learning (IRL) suggests that the robot could\nask for task demonstrations and recover a reward defined over the raw state\nspace. Our insight is that rather than implicitly learning about the missing\nfeature(s) from demonstrations, the robot should instead ask for data that\nexplicitly teaches it about what it is missing. We introduce a new type of\nhuman input in which the person guides the robot from states where the feature\nbeing taught is highly expressed to states where it is not. We propose an\nalgorithm for learning the feature from the raw state space and integrating it\ninto the reward function. By focusing the human input on the missing feature,\nour method decreases sample complexity and improves generalization of the\nlearned reward over the above deep IRL baseline. We show this in experiments\nwith a physical 7DOF robot manipulator, as well as in a user study conducted in\na simulated environment.', ""In dialogue systems, utterances with similar semantics may have distinctive\nemotions under different contexts. Therefore, modeling long-range contextual\nemotional relationships with speaker dependency plays a crucial part in\ndialogue emotion recognition. Meanwhile, distinguishing the different emotion\ncategories is non-trivial since they usually have semantically similar\nsentiments. To this end, we adopt supervised contrastive learning to make\ndifferent emotions mutually exclusive to identify similar emotions better.\nMeanwhile, we utilize an auxiliary response generation task to enhance the\nmodel's ability of handling context information, thereby forcing the model to\nrecognize emotions with similar semantics in diverse contexts. To achieve these\nobjectives, we use the pre-trained encoder-decoder model BART as our backbone\nmodel since it is very suitable for both understanding and generation tasks.\nThe experiments on four datasets demonstrate that our proposed model obtains\nsignificantly more favorable results than the state-of-the-art model in\ndialogue emotion recognition. The ablation study further demonstrates the\neffectiveness of supervised contrastive loss and generative loss."", 'While there have been significant advances in de-tecting emotions in text, in\nthe field of utter-ance-level emotion recognition (ULER), there are still many\nproblems to be solved. In this paper, we address some challenges in ULER in\ndialog sys-tems. (1) The same utterance can deliver different emotions when it\nis in different contexts or from different speakers. (2) Long-range contextual\nin-formation is hard to effectively capture. (3) Unlike the traditional text\nclassification problem, this task is supported by a limited number of datasets,\namong which most contain inadequate conversa-tions or speech. To address these\nproblems, we propose a hierarchical transformer framework (apart from the\ndescription of other studies, the ""transformer"" in this paper usually refers to\nthe encoder part of the transformer) with a lower-level transformer to model\nthe word-level input and an upper-level transformer to capture the context of\nutterance-level embeddings. We use a pretrained language model bidirectional\nencoder representa-tions from transformers (BERT) as the lower-level\ntransformer, which is equivalent to introducing external data into the model\nand solve the problem of data shortage to some extent. In addition, we add\nspeaker embeddings to the model for the first time, which enables our model to\ncapture the in-teraction between speakers. Experiments on three dialog emotion\ndatasets, Friends, EmotionPush, and EmoryNLP, demonstrate that our proposed\nhierarchical transformer network models achieve 1.98%, 2.83%, and 3.94%\nimprovement, respec-tively, over the state-of-the-art methods on each dataset\nin terms of macro-F1.', 'Emotion dynamics modeling is a significant task in emotion recognition in\nconversation. It aims to predict conversational emotions when building\nempathetic dialogue systems. Existing studies mainly develop models based on\nRecurrent Neural Networks (RNNs). They cannot benefit from the power of the\nrecently-developed pre-training strategies for better token representation\nlearning in conversations. More seriously, it is hard to distinguish the\ndependency of interlocutors and the emotional influence among interlocutors by\nsimply assembling the features on top of RNNs. In this paper, we develop a\nseries of BERT-based models to specifically capture the inter-interlocutor and\nintra-interlocutor dependencies of the conversational emotion dynamics.\nConcretely, we first substitute BERT for RNNs to enrich the token\nrepresentations. Then, a Flat-structured BERT (F-BERT) is applied to link up\nutterances in a conversation directly, and a Hierarchically-structured BERT\n(H-BERT) is employed to distinguish the interlocutors when linking up\nutterances. More importantly, a Spatial-Temporal-structured BERT, namely\nST-BERT, is proposed to further determine the emotional influence among\ninterlocutors. Finally, we conduct extensive experiments on two popular emotion\nrecognition in conversation benchmark datasets and demonstrate that our\nproposed models can attain around 5\\% and 10\\% improvement over the\nstate-of-the-art baselines, respectively.', 'While large-scale pretrained language models have been shown to learn\neffective linguistic representations for many NLP tasks, there remain many\nreal-world contextual aspects of language that current approaches do not\ncapture. For instance, consider a cloze-test ""I enjoyed the ____ game this\nweekend"": the correct answer depends heavily on where the speaker is from, when\nthe utterance occurred, and the speaker\'s broader social milieu and\npreferences. Although language depends heavily on the geographical, temporal,\nand other social contexts of the speaker, these elements have not been\nincorporated into modern transformer-based language models. We propose a simple\nbut effective approach to incorporate speaker social context into the learned\nrepresentations of large-scale language models. Our method first learns dense\nrepresentations of social contexts using graph representation learning\nalgorithms and then primes language model pretraining with these social context\nrepresentations. We evaluate our approach on geographically-sensitive\nlanguage-modeling tasks and show a substantial improvement (more than 100%\nrelative lift on MRR) compared to baselines.', 'Most extractive summarization methods focus on the main body of the document\nfrom which sentences need to be extracted. However, the gist of the document\nmay lie in side information, such as the title and image captions which are\noften available for newswire articles. We propose to explore side information\nin the context of single-document extractive summarization. We develop a\nframework for single-document summarization composed of a hierarchical document\nencoder and an attention-based extractor with attention over side information.\nWe evaluate our model on a large scale news dataset. We show that extractive\nsummarization with side information consistently outperforms its counterpart\nthat does not use any side information, in terms of both informativeness and\nfluency.', 'Document-level machine translation manages to outperform sentence level\nmodels by a small margin, but have failed to be widely adopted. We argue that\nprevious research did not make a clear use of the global context, and propose a\nnew document-level NMT framework that deliberately models the local context of\neach sentence with the awareness of the global context of the document in both\nsource and target languages. We specifically design the model to be able to\ndeal with documents containing any number of sentences, including single\nsentences. This unified approach allows our model to be trained elegantly on\nstandard datasets without needing to train on sentence and document level data\nseparately. Experimental results demonstrate that our model outperforms\nTransformer baselines and previous document-level NMT models with substantial\nmargins of up to 2.1 BLEU on state-of-the-art baselines. We also provide\nanalyses which show the benefit of context far beyond the neighboring two or\nthree sentences, which previous studies have typically incorporated.', 'We introduce a framework that abstracts Reinforcement Learning (RL) as a\nsequence modeling problem. This allows us to draw upon the simplicity and\nscalability of the Transformer architecture, and associated advances in\nlanguage modeling such as GPT-x and BERT. In particular, we present Decision\nTransformer, an architecture that casts the problem of RL as conditional\nsequence modeling. Unlike prior approaches to RL that fit value functions or\ncompute policy gradients, Decision Transformer simply outputs the optimal\nactions by leveraging a causally masked Transformer. By conditioning an\nautoregressive model on the desired return (reward), past states, and actions,\nour Decision Transformer model can generate future actions that achieve the\ndesired return. Despite its simplicity, Decision Transformer matches or exceeds\nthe performance of state-of-the-art model-free offline RL baselines on Atari,\nOpenAI Gym, and Key-to-Door tasks.', 'In continuing tasks, average-reward reinforcement learning may be a more\nappropriate problem formulation than the more common discounted reward\nformulation. As usual, learning an optimal policy in this setting typically\nrequires a large amount of training experiences. Reward shaping is a common\napproach for incorporating domain knowledge into reinforcement learning in\norder to speed up convergence to an optimal policy. However, to the best of our\nknowledge, the theoretical properties of reward shaping have thus far only been\nestablished in the discounted setting. This paper presents the first reward\nshaping framework for average-reward learning and proves that, under standard\nassumptions, the optimal policy under the original reward function can be\nrecovered. In order to avoid the need for manual construction of the shaping\nfunction, we introduce a method for utilizing domain knowledge expressed as a\ntemporal logic formula. The formula is automatically translated to a shaping\nfunction that provides additional reward throughout the learning process. We\nevaluate the proposed method on three continuing tasks. In all cases, shaping\nspeeds up the average-reward learning rate without any reduction in the\nperformance of the learned policy compared to relevant baselines.', 'Narrative, and in particular storytelling, is an important part of the human\nexperience. Consequently, computational systems that can reason about narrative\ncan be more effective communicators, entertainers, educators, and trainers. One\nof the central challenges in computational narrative reasoning is narrative\ngeneration, the automated creation of meaningful event sequences. There are\nmany factors -- logical and aesthetic -- that contribute to the success of a\nnarrative artifact. Central to this success is its understandability. We argue\nthat the following two attributes of narratives are universal: (a) the logical\ncausal progression of plot, and (b) character believability. Character\nbelievability is the perception by the audience that the actions performed by\ncharacters do not negatively impact the audiences suspension of disbelief.\nSpecifically, characters must be perceived by the audience to be intentional\nagents. In this article, we explore the use of refinement search as a technique\nfor solving the narrative generation problem -- to find a sound and believable\nsequence of character actions that transforms an initial world state into a\nworld state in which goal propositions hold. We describe a novel refinement\nsearch planning algorithm -- the Intent-based Partial Order Causal Link (IPOCL)\nplanner -- that, in addition to creating causally sound plot progression,\nreasons about character intentionality by identifying possible character goals\nthat explain their actions and creating plan structures that explain why those\ncharacters commit to their goals. We present the results of an empirical\nevaluation that demonstrates that narrative plans generated by the IPOCL\nalgorithm support audience comprehension of character intentions better than\nplans generated by conventional partial-order planners.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'Transformer is important for text modeling. However, it has difficulty in\nhandling long documents due to the quadratic complexity with input text length.\nIn order to handle this problem, we propose a hierarchical interactive\nTransformer (Hi-Transformer) for efficient and effective long document\nmodeling. Hi-Transformer models documents in a hierarchical way, i.e., first\nlearns sentence representations and then learns document representations. It\ncan effectively reduce the complexity and meanwhile capture global document\ncontext in the modeling of each sentence. More specifically, we first use a\nsentence Transformer to learn the representations of each sentence. Then we use\na document Transformer to model the global document context from these sentence\nrepresentations. Next, we use another sentence Transformer to enhance sentence\nmodeling using the global document context. Finally, we use hierarchical\npooling method to obtain document embedding. Extensive experiments on three\nbenchmark datasets validate the efficiency and effectiveness of Hi-Transformer\nin long document modeling.', ""With the improvements in speech recognition and voice generation technologies\nover the last years, a lot of companies have sought to develop conversation\nunderstanding systems that run on mobile phones or smart home devices through\nnatural language interfaces. Conversational assistants, such as Google\nAssistant and Microsoft Cortana, can help users to complete various types of\ntasks. This requires an accurate understanding of the user's information need\nas the conversation evolves into multiple turns. Finding relevant context in a\nconversation's history is challenging because of the complexity of natural\nlanguage and the evolution of a user's information need. In this work, we\npresent an extensive analysis of language, relevance, dependency of user\nutterances in a multi-turn information-seeking conversation. To this aim, we\nhave annotated relevant utterances in the conversations released by the TREC\nCaST 2019 track. The annotation labels determine which of the previous\nutterances in a conversation can be used to improve the current one.\nFurthermore, we propose a neural utterance relevance model based on BERT\nfine-tuning, outperforming competitive baselines. We study and compare the\nperformance of multiple retrieval models, utilizing different strategies to\nincorporate the user's context. The experimental results on both classification\nand retrieval tasks show that our proposed approach can effectively identify\nand incorporate the conversation context. We show that processing the current\nutterance using the predicted relevant utterance leads to a 38% relative\nimprovement in terms of nDCG@20. Finally, to foster research in this area, we\nhave released the dataset of the annotations."", 'The problem of inverse reinforcement learning (IRL) is relevant to a variety\nof tasks including value alignment and robot learning from demonstration.\nDespite significant algorithmic contributions in recent years, IRL remains an\nill-posed problem at its core; multiple reward functions coincide with the\nobserved behavior and the actual reward function is not identifiable without\nprior knowledge or supplementary information. This paper presents an IRL\nframework called Bayesian optimization-IRL (BO-IRL) which identifies multiple\nsolutions that are consistent with the expert demonstrations by efficiently\nexploring the reward function space. BO-IRL achieves this by utilizing Bayesian\nOptimization along with our newly proposed kernel that (a) projects the\nparameters of policy invariant reward functions to a single point in a latent\nspace and (b) ensures nearby points in the latent space correspond to reward\nfunctions yielding similar likelihoods. This projection allows the use of\nstandard stationary kernels in the latent space to capture the correlations\npresent across the reward function space. Empirical results on synthetic and\nreal-world environments (model-free and model-based) show that BO-IRL discovers\nmultiple reward functions while minimizing the number of expensive exact policy\noptimizations.', ""Human-in-the-loop machine learning is widely used in artificial intelligence\n(AI) to elicit labels for data points from experts or to provide feedback on\nhow close the predicted results are to the target. This simplifies away all the\ndetails of the decision-making process of the expert. In this work, we allow\nthe experts to additionally produce decision rules describing their\ndecision-making; the rules are expected to be imperfect but to give additional\ninformation. In particular, the rules can extend to new distributions, and\nhence enable significantly improving performance for cases where the training\nand testing distributions differ, such as in domain adaptation. We apply the\nproposed method to lifelong learning and domain adaptation problems and discuss\napplications in other branches of AI, such as knowledge acquisition problems in\nexpert systems. In simulated and real-user studies, we show that decision rule\nelicitation improves domain adaptation of the algorithm and helps to propagate\nexpert's knowledge to the AI model."", 'Humans are able to understand and perform complex tasks by strategically\nstructuring the tasks into incremental steps or subgoals. For a robot\nattempting to learn to perform a sequential task with critical subgoal states,\nsuch states can provide a natural opportunity for interaction with a human\nexpert. This paper analyzes the benefit of incorporating a notion of subgoals\ninto Inverse Reinforcement Learning (IRL) with a Human-In-The-Loop (HITL)\nframework. The learning process is interactive, with a human expert first\nproviding input in the form of full demonstrations along with some subgoal\nstates. These subgoal states define a set of subtasks for the learning agent to\ncomplete in order to achieve the final goal. The learning agent queries for\npartial demonstrations corresponding to each subtask as needed when the agent\nstruggles with the subtask. The proposed Human Interactive IRL (HI-IRL)\nframework is evaluated on several discrete path-planning tasks. We demonstrate\nthat subgoal-based interactive structuring of the learning task results in\nsignificantly more efficient learning, requiring only a fraction of the\ndemonstration data needed for learning the underlying reward function with the\nbaseline IRL model.', 'Deep reinforcement learning is the combination of reinforcement learning (RL)\nand deep learning. This field of research has been able to solve a wide range\nof complex decision-making tasks that were previously out of reach for a\nmachine. Thus, deep RL opens up many new applications in domains such as\nhealthcare, robotics, smart grids, finance, and many more. This manuscript\nprovides an introduction to deep reinforcement learning models, algorithms and\ntechniques. Particular focus is on the aspects related to generalization and\nhow deep RL can be used for practical applications. We assume the reader is\nfamiliar with basic machine learning concepts.', 'Recognizing emotions in conversations is a challenging task due to the\npresence of contextual dependencies governed by self- and inter-personal\ninfluences. Recent approaches have focused on modeling these dependencies\nprimarily via supervised learning. However, purely supervised strategies demand\nlarge amounts of annotated data, which is lacking in most of the available\ncorpora in this task. To tackle this challenge, we look at transfer learning\napproaches as a viable alternative. Given the large amount of available\nconversational data, we investigate whether generative conversational models\ncan be leveraged to transfer affective knowledge for detecting emotions in\ncontext. We propose an approach, TL-ERC, where we pre-train a hierarchical\ndialogue model on multi-turn conversations (source) and then transfer its\nparameters to a conversational emotion classifier (target). In addition to the\npopular practice of using pre-trained sentence encoders, our approach also\nincorporates recurrent parameters that model inter-sentential context across\nthe whole conversation. Based on this idea, we perform several experiments\nacross multiple datasets and find improvement in performance and robustness\nagainst limited training data. TL-ERC also achieves better validation\nperformances in significantly fewer epochs. Overall, we infer that knowledge\nacquired from dialogue generators can indeed help recognize emotions in\nconversations.', 'Recently, text world games have been proposed to enable artificial agents to\nunderstand and reason about real-world scenarios. These text-based games are\nchallenging for artificial agents, as it requires an understanding of and\ninteraction using natural language in a partially observable environment.\nAgents observe the environment via textual descriptions designed to be\nchallenging enough for even human players. Past approaches have not paid enough\nattention to the language understanding capability of the proposed agents.\nTypically, these approaches train from scratch, an agent that learns both\ntextual representations and the gameplay online during training using a\ntemporal loss function. Given the sample-inefficiency of RL approaches, it is\ninefficient to learn rich enough textual representations to be able to\nunderstand and reason using the textual observation in such a complicated game\nenvironment setting. In this paper, we improve the semantic understanding of\nthe agent by proposing a simple RL with LM framework where we use\ntransformer-based language models with Deep RL models. We perform a detailed\nstudy of our framework to demonstrate how our model outperforms all existing\nagents on the popular game, Zork1, to achieve a score of 44.7, which is 1.6\nhigher than the state-of-the-art model. Overall, our proposed approach\noutperforms 4 games out of the 14 text-based games, while performing comparable\nto the state-of-the-art models on the remaining games.', 'Standard neural machine translation (NMT) is on the assumption that the\ndocument-level context is independent. Most existing document-level NMT\napproaches are satisfied with a smattering sense of global document-level\ninformation, while this work focuses on exploiting detailed document-level\ncontext in terms of a memory network. The capacity of the memory network that\ndetecting the most relevant part of the current sentence from memory renders a\nnatural solution to model the rich document-level context. In this work, the\nproposed document-aware memory network is implemented to enhance the\nTransformer NMT baseline. Experiments on several tasks show that the proposed\nmethod significantly improves the NMT performance over strong Transformer\nbaselines and other related studies.', 'In this paper, we propose a model-free reinforcement learning method to\nsynthesize control policies for motion planning problems with continuous states\nand actions. The robot is modelled as a labeled discrete-time Markov decision\nprocess (MDP) with continuous state and action spaces. Linear temporal logics\n(LTL) are used to specify high-level tasks. We then train deep neural networks\nto approximate the value function and policy using an actor-critic\nreinforcement learning method. The LTL specification is converted into an\nannotated limit-deterministic B\\""uchi automaton (LDBA) for continuously shaping\nthe reward so that dense rewards are available during training. A na\\""ive way\nof solving a motion planning problem with LTL specifications using\nreinforcement learning is to sample a trajectory and then assign a high reward\nfor training if the trajectory satisfies the entire LTL formula. However, the\nsampling complexity needed to find such a trajectory is too high when we have a\ncomplex LTL formula for continuous state and action spaces. As a result, it is\nvery unlikely that we get enough reward for training if all sample trajectories\nstart from the initial state in the automata. In this paper, we propose a\nmethod that samples not only an initial state from the state space, but also an\narbitrary state in the automata at the beginning of each training episode. We\ntest our algorithm in simulation using a car-like robot and find out that our\nmethod can learn policies for different working configurations and LTL\nspecifications successfully.', 'Previous researchers have considered sentiment analysis as a document\nclassification task, in which input documents are classified into predefined\nsentiment classes. Although there are sentences in a document that support\nimportant evidences for sentiment analysis and sentences that do not, they have\ntreated the document as a bag of sentences. In other words, they have not\nconsidered the importance of each sentence in the document. To effectively\ndetermine polarity of a document, each sentence in the document should be dealt\nwith different degrees of importance. To address this problem, we propose a\ndocument-level sentence classification model based on deep neural networks, in\nwhich the importance degrees of sentences in documents are automatically\ndetermined through gate mechanisms. To verify our new sentiment analysis model,\nwe conducted experiments using the sentiment datasets in the four different\ndomains such as movie reviews, hotel reviews, restaurant reviews, and music\nreviews. In the experiments, the proposed model outperformed previous\nstate-of-the-art models that do not consider importance differences of\nsentences in a document. The experimental results show that the importance of\nsentences should be considered in a document-level sentiment classification\ntask.', 'Mental health conditions remain under-diagnosed even in countries with common\naccess to advanced medical care. The ability to accurately and efficiently\npredict mood from easily collectible data has several important implications\ntowards the early detection and intervention of mental health disorders. One\npromising data source to help monitor human behavior is from daily smartphone\nusage. However, care must be taken to summarize behaviors without identifying\nthe user through personal (e.g., personally identifiable information) or\nprotected attributes (e.g., race, gender). In this paper, we study behavioral\nmarkers or daily mood using a recent dataset of mobile behaviors from high-risk\nadolescent populations. Using computational models, we find that multimodal\nmodeling of both text and app usage features is highly predictive of daily mood\nover each modality alone. Furthermore, we evaluate approaches that reliably\nobfuscate user identity while remaining predictive of daily mood. By combining\nmultimodal representations with privacy-preserving learning, we are able to\npush forward the performance-privacy frontier as compared to unimodal\napproaches.', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', ""In some agent designs like inverse reinforcement learning an agent needs to\nlearn its own reward function. Learning the reward function and optimising for\nit are typically two different processes, usually performed at different\nstages. We consider a continual (``one life'') learning approach where the\nagent both learns the reward function and optimises for it at the same time. We\nshow that this comes with a number of pitfalls, such as deliberately\nmanipulating the learning process in one direction, refusing to learn,\n``learning'' facts already known to the agent, and making decisions that are\nstrictly dominated (for all relevant reward functions). We formally introduce\ntwo desirable properties: the first is `unriggability', which prevents the\nagent from steering the learning process in the direction of a reward function\nthat is easier to optimise. The second is `uninfluenceability', whereby the\nreward-function learning process operates by learning facts about the\nenvironment. We show that an uninfluenceable process is automatically\nunriggable, and if the set of possible environments is sufficiently rich, the\nconverse is true too."", 'Real-world robotic tasks require complex reward functions. When we define the\nproblem the robot needs to solve, we pretend that a designer specifies this\ncomplex reward exactly, and it is set in stone from then on. In practice,\nhowever, reward design is an iterative process: the designer chooses a reward,\neventually encounters an ""edge-case"" environment where the reward incentivizes\nthe wrong behavior, revises the reward, and repeats. What would it mean to\nrethink robotics problems to formally account for this iterative nature of\nreward design? We propose that the robot not take the specified reward for\ngranted, but rather have uncertainty about it, and account for the future\ndesign iterations as future evidence. We contribute an Assisted Reward Design\nmethod that speeds up the design process by anticipating and influencing this\nfuture evidence: rather than letting the designer eventually encounter failure\ncases and revise the reward then, the method actively exposes the designer to\nsuch environments during the development phase. We test this method in a\nsimplified autonomous driving task and find that it more quickly improves the\ncar\'s behavior in held-out environments by proposing environments that are\n""edge cases"" for the current reward.', ""We study the problem of designing autonomous agents that can learn to\ncooperate effectively with a potentially suboptimal partner while having no\naccess to the joint reward function. This problem is modeled as a cooperative\nepisodic two-agent Markov decision process. We assume control over only the\nfirst of the two agents in a Stackelberg formulation of the game, where the\nsecond agent is acting so as to maximise expected utility given the first\nagent's policy. How should the first agent act in order to learn the joint\nreward function as quickly as possible and so that the joint policy is as close\nto optimal as possible? We analyse how knowledge about the reward function can\nbe gained in this interactive two-agent scenario. We show that when the\nlearning agent's policies have a significant effect on the transition function,\nthe reward function can be learned efficiently."", 'The ability to infer persona from dialogue can have applications in areas\nranging from computational narrative analysis to personalized dialogue\ngeneration. We introduce neural models to learn persona embeddings in a\nsupervised character trope classification task. The models encode dialogue\nsnippets from IMDB into representations that can capture the various categories\nof film characters. The best-performing models use a multi-level attention\nmechanism over a set of utterances. We also utilize prior knowledge in the form\nof textual descriptions of the different tropes. We apply the learned\nembeddings to find similar characters across different movies, and cluster\nmovies according to the distribution of the embeddings. The use of short\nconversational text as input, and the ability to learn from prior knowledge\nusing memory, suggests these methods could be applied to other domains.', 'Recently, the automatic prediction of personality traits has received a lot\nof attention. Specifically, personality trait prediction from multimodal data\nhas emerged as a hot topic within the field of affective computing. In this\npaper, we review significant machine learning models which have been employed\nfor personality detection, with an emphasis on deep learning-based methods.\nThis review paper provides an overview of the most popular approaches to\nautomated personality detection, various computational datasets, its industrial\napplications, and state-of-the-art machine learning models for personality\ndetection with specific focus on multimodal approaches. Personality detection\nis a very broad and diverse topic: this survey only focuses on computational\napproaches and leaves out psychological studies on personality detection.', 'Document grounded generation is the task of using the information provided in\na document to improve text generation. This work focuses on two different\ndocument grounded generation tasks: Wikipedia Update Generation task and\nDialogue response generation. Our work introduces two novel adaptations of\nlarge scale pre-trained encoder-decoder models focusing on building context\ndriven representation of the document and enabling specific attention to the\ninformation in the document. Additionally, we provide a stronger BART baseline\nfor these tasks. Our proposed techniques outperform existing methods on both\nautomated (at least 48% increase in BLEU-4 points) and human evaluation for\ncloseness to reference and relevance to the document. Furthermore, we perform\ncomprehensive manual inspection of the generated output and categorize errors\nto provide insights into future directions in modeling these tasks.', 'Question Answering with NLP has progressed through the evolution of advanced\nmodel architectures like BERT and BiDAF and earlier word, character, and\ncontext-based embeddings. As BERT has leapfrogged the accuracy of models, an\nelement of the next frontier can be the introduction of deep networks and an\neffective way to train them. In this context, I explored the effectiveness of\ndeep networks focussing on the model encoder layer of BiDAF. BiDAF with its\nheterogeneous layers provides the opportunity not only to explore the\neffectiveness of deep networks but also to evaluate whether the refinements\nmade in lower layers are additive to the refinements made in the upper layers\nof the model architecture. I believe the next greatest model in NLP will in\nfact fold in a solid language modeling like BERT with a composite architecture\nwhich will bring in refinements in addition to generic language modeling and\nwill have a more extensive layered architecture. I experimented with the Bypass\nnetwork, Residual Highway network, and DenseNet architectures. In addition, I\nevaluated the effectiveness of ensembling the last few layers of the network. I\nalso studied the difference character embeddings make in adding them to the\nword embeddings, and whether the effects are additive with deep networks. My\nstudies indicate that deep networks are in fact effective in giving a boost.\nAlso, the refinements in the lower layers like embeddings are passed on\nadditively to the gains made through deep networks.', 'We present a contrasting learning approach with data augmentation techniques\nto learn document representations in an unsupervised manner. Inspired by recent\ncontrastive self-supervised learning algorithms used for image and NLP\npretraining, we hypothesize that high-quality document embedding should be\ninvariant to diverse paraphrases that preserve the semantics of the original\ndocument. With different backbones and contrastive learning frameworks, our\nstudy reveals the enormous benefits of contrastive augmentation for document\nrepresentation learning with two additional insights: 1) including data\naugmentation in a contrastive way can substantially improve the embedding\nquality in unsupervised document representation learning, and 2) in general,\nstochastic augmentations generated by simple word-level manipulation work much\nbetter than sentence-level and document-level ones. We plug our method into a\nclassifier and compare it with a broad range of baseline methods on six\nbenchmark datasets. Our method can decrease the classification error rate by up\nto 6.4% over the SOTA approaches on the document classification task, matching\nor even surpassing fully-supervised methods.', ""We propose a framework for verifiable and compositional reinforcement\nlearning (RL) in which a collection of RL subsystems, each of which learns to\naccomplish a separate subtask, are composed to achieve an overall task. The\nframework consists of a high-level model, represented as a parametric Markov\ndecision process (pMDP) which is used to plan and to analyze compositions of\nsubsystems, and of the collection of low-level subsystems themselves. By\ndefining interfaces between the subsystems, the framework enables automatic\ndecompositions of task specifications, e.g., reach a target set of states with\na probability of at least 0.95, into individual subtask specifications, i.e.\nachieve the subsystem's exit conditions with at least some minimum probability,\ngiven that its entry conditions are met. This in turn allows for the\nindependent training and testing of the subsystems; if they each learn a policy\nsatisfying the appropriate subtask specification, then their composition is\nguaranteed to satisfy the overall task specification. Conversely, if the\nsubtask specifications cannot all be satisfied by the learned policies, we\npresent a method, formulated as the problem of finding an optimal set of\nparameters in the pMDP, to automatically update the subtask specifications to\naccount for the observed shortcomings. The result is an iterative procedure for\ndefining subtask specifications, and for training the subsystems to meet them.\nAs an additional benefit, this procedure allows for particularly challenging or\nimportant components of an overall task to be determined automatically, and\nfocused on, during training. Experimental results demonstrate the presented\nframework's novel capabilities."", 'Reinforcement learning (RL) algorithms have been around for decades and\nemployed to solve various sequential decision-making problems. These algorithms\nhowever have faced great challenges when dealing with high-dimensional\nenvironments. The recent development of deep learning has enabled RL methods to\ndrive optimal policies for sophisticated and capable agents, which can perform\nefficiently in these challenging environments. This paper addresses an\nimportant aspect of deep RL related to situations that require multiple agents\nto communicate and cooperate to solve complex tasks. A survey of different\napproaches to problems related to multi-agent deep RL (MADRL) is presented,\nincluding non-stationarity, partial observability, continuous state and action\nspaces, multi-agent training schemes, multi-agent transfer learning. The merits\nand demerits of the reviewed methods will be analyzed and discussed, with their\ncorresponding applications explored. It is envisaged that this review provides\ninsights about various MADRL methods and can lead to future development of more\nrobust and highly useful multi-agent learning methods for solving real-world\nproblems.', 'We examine a new commonsense reasoning task: given a narrative describing a\nsocial interaction that centers on two protagonists, systems make inferences\nabout the underlying relationship trajectory. Specifically, we propose two\nevaluation tasks: Relationship Outlook Prediction MCQ and Resolution Prediction\nMCQ. In Relationship Outlook Prediction, a system maps an interaction to a\nrelationship outlook that captures how the interaction is expected to change\nthe relationship. In Resolution Prediction, a system attributes a given\nrelationship outlook to a particular resolution that explains the outcome.\nThese two tasks parallel two real-life questions that people frequently ponder\nupon as they navigate different social situations: ""where is this relationship\ngoing?"" and ""how did we end up here?"". To facilitate the investigation of human\nsocial relationships through these two tasks, we construct a new dataset,\nSocial Narrative Tree, which consists of 1250 stories documenting a variety of\ndaily social interactions. The narratives encode a multitude of social elements\nthat interweave to give rise to rich commonsense knowledge of how relationships\nevolve with respect to social interactions. We establish baseline performances\nusing language models and the accuracies are significantly lower than human\nperformance. The results demonstrate that models need to look beyond syntactic\nand semantic signals to comprehend complex human relationships.', 'Policy distillation, which transfers a teacher policy to a student policy has\nachieved great success in challenging tasks of deep reinforcement learning.\nThis teacher-student framework requires a well-trained teacher model which is\ncomputationally expensive. Moreover, the performance of the student model could\nbe limited by the teacher model if the teacher model is not optimal. In the\nlight of collaborative learning, we study the feasibility of involving joint\nintellectual efforts from diverse perspectives of student models. In this work,\nwe introduce dual policy distillation(DPD), a student-student framework in\nwhich two learners operate on the same environment to explore different\nperspectives of the environment and extract knowledge from each other to\nenhance their learning. The key challenge in developing this dual learning\nframework is to identify the beneficial knowledge from the peer learner for\ncontemporary learning-based reinforcement learning algorithms, since it is\nunclear whether the knowledge distilled from an imperfect and noisy peer\nlearner would be helpful. To address the challenge, we theoretically justify\nthat distilling knowledge from a peer learner will lead to policy improvement\nand propose a disadvantageous distillation strategy based on the theoretical\nresults. The conducted experiments on several continuous control tasks show\nthat the proposed framework achieves superior performance with a learning-based\nagent and function approximation without the use of expensive teacher models.', ""As intelligent systems gain autonomy and capability, it becomes vital to\nensure that their objectives match those of their human users; this is known as\nthe value-alignment problem. In robotics, value alignment is key to the design\nof collaborative robots that can integrate into human workflows, successfully\ninferring and adapting to their users' objectives as they go. We argue that a\nmeaningful solution to value alignment must combine multi-agent decision theory\nwith rich mathematical models of human cognition, enabling robots to tap into\npeople's natural collaborative capabilities. We present a solution to the\ncooperative inverse reinforcement learning (CIRL) dynamic game based on\nwell-established cognitive models of decision making and theory of mind. The\nsolution captures a key reciprocity relation: the human will not plan her\nactions in isolation, but rather reason pedagogically about how the robot might\nlearn from them; the robot, in turn, can anticipate this and interpret the\nhuman's actions pragmatically. To our knowledge, this work constitutes the\nfirst formal analysis of value alignment grounded in empirically validated\ncognitive models."", 'We present an approach to training neural networks to generate sequences\nusing actor-critic methods from reinforcement learning (RL). Current\nlog-likelihood training methods are limited by the discrepancy between their\ntraining and testing modes, as models must generate tokens conditioned on their\nprevious guesses rather than the ground-truth tokens. We address this problem\nby introducing a \\textit{critic} network that is trained to predict the value\nof an output token, given the policy of an \\textit{actor} network. This results\nin a training procedure that is much closer to the test phase, and allows us to\ndirectly optimize for a task-specific score such as BLEU. Crucially, since we\nleverage these techniques in the supervised learning setting rather than the\ntraditional RL setting, we condition the critic network on the ground-truth\noutput. We show that our method leads to improved performance on both a\nsynthetic task, and for German-English machine translation. Our analysis paves\nthe way for such methods to be applied in natural language generation tasks,\nsuch as machine translation, caption generation, and dialogue modelling.', 'Deep reading models for question-answering have demonstrated promising\nperformance over the last couple of years. However current systems tend to\nlearn how to cleverly extract a span of the source document, based on its\nsimilarity with the question, instead of seeking for the appropriate answer.\nIndeed, a reading machine should be able to detect relevant passages in a\ndocument regarding a question, but more importantly, it should be able to\nreason over the important pieces of the document in order to produce an answer\nwhen it is required. To motivate this purpose, we present ReviewQA, a\nquestion-answering dataset based on hotel reviews. The questions of this\ndataset are linked to a set of relational understanding competencies that we\nexpect a model to master. Indeed, each question comes with an associated type\nthat characterizes the required competency. With this framework, it is possible\nto benchmark the main families of models and to get an overview of what are the\nstrengths and the weaknesses of a given model on the set of tasks evaluated in\nthis dataset. Our corpus contains more than 500.000 questions in natural\nlanguage over 100.000 hotel reviews. Our setup is projective, the answer of a\nquestion does not need to be extracted from a document, like in most of the\nrecent datasets, but selected among a set of candidates that contains all the\npossible answers to the questions of the dataset. Finally, we present several\nbaselines over this dataset.', 'We explore using multilingual document embeddings for nearest neighbor mining\nof parallel data. Three document-level representations are investigated: (i)\ndocument embeddings generated by simply averaging multilingual sentence\nembeddings; (ii) a neural bag-of-words (BoW) document encoding model; (iii) a\nhierarchical multilingual document encoder (HiDE) that builds on our\nsentence-level model. The results show document embeddings derived from\nsentence-level averaging are surprisingly effective for clean datasets, but\nsuggest models trained hierarchically at the document-level are more effective\non noisy data. Analysis experiments demonstrate our hierarchical models are\nvery robust to variations in the underlying sentence embedding quality. Using\ndocument embeddings trained with HiDE achieves state-of-the-art performance on\nUnited Nations (UN) parallel document mining, 94.9% P@1 for en-fr and 97.3% P@1\nfor en-es.', 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', 'Bipolar disorder (BPD) is a chronic mental illness characterized by extreme\nmood and energy changes from mania to depression. These changes drive behaviors\nthat often lead to devastating personal or social consequences. BPD is managed\nclinically with regular interactions with care providers, who assess mood,\nenergy levels, and the form and content of speech. Recent work has proposed\nsmartphones for monitoring mood using speech. However, these works do not\npredict when to intervene. Predicting when to intervene is challenging because\nthere is not a single measure that is relevant for every person: different\nindividuals may have different levels of symptom severity considered typical.\nAdditionally, this typical mood, or baseline, may change over time, making a\nsingle symptom threshold insufficient. This work presents an innovative\napproach that expands clinical mood monitoring to predict when interventions\nare necessary using an anomaly detection framework, which we call Temporal\nNormalization. We first validate the model using a dataset annotated for\nclinical interventions and then incorporate this method in a deep learning\nframework to predict mood anomalies from natural, unstructured, telephone\nspeech data. The combination of these approaches provides a framework to enable\nreal-world speech-focused mood monitoring.', 'In many real-world deployments of machine learning systems, data arrive\npiecemeal. These learning scenarios may be passive, where data arrive\nincrementally due to structural properties of the problem (e.g., daily\nfinancial data) or active, where samples are selected according to a measure of\ntheir quality (e.g., experimental design). In both of these cases, we are\nbuilding a sequence of models that incorporate an increasing amount of data. We\nwould like each of these models in the sequence to be performant and take\nadvantage of all the data that are available to that point. Conventional\nintuition suggests that when solving a sequence of related optimization\nproblems of this form, it should be possible to initialize using the solution\nof the previous iterate -- to ""warm start"" the optimization rather than\ninitialize from scratch -- and see reductions in wall-clock time. However, in\npractice this warm-starting seems to yield poorer generalization performance\nthan models that have fresh random initializations, even though the final\ntraining losses are similar. While it appears that some hyperparameter settings\nallow a practitioner to close this generalization gap, they seem to only do so\nin regimes that damage the wall-clock gains of the warm start. Nevertheless, it\nis highly desirable to be able to warm-start neural network training, as it\nwould dramatically reduce the resource usage associated with the construction\nof performant deep learning systems. In this work, we take a closer look at\nthis empirical phenomenon and try to understand when and how it occurs. We also\nprovide a surprisingly simple trick that overcomes this pathology in several\nimportant situations, and present experiments that elucidate some of its\nproperties.', 'Information Retrieval (IR) models need to deal with two difficult issues,\nvocabulary mismatch and term dependencies. Vocabulary mismatch corresponds to\nthe difficulty of retrieving relevant documents that do not contain exact query\nterms but semantically related terms. Term dependencies refers to the need of\nconsidering the relationship between the words of the query when estimating the\nrelevance of a document. A multitude of solutions has been proposed to solve\neach of these two problems, but no principled model solve both. In parallel, in\nthe last few years, language models based on neural networks have been used to\ncope with complex natural language processing tasks like emotion and paraphrase\ndetection. Although they present good abilities to cope with both term\ndependencies and vocabulary mismatch problems, thanks to the distributed\nrepresentation of words they are based upon, such models could not be used\nreadily in IR, where the estimation of one language model per document (or\nquery) is required. This is both computationally unfeasible and prone to\nover-fitting. Based on a recent work that proposed to learn a generic language\nmodel that can be modified through a set of document-specific parameters, we\nexplore use of new neural network models that are adapted to ad-hoc IR tasks.\nWithin the language model IR framework, we propose and study the use of a\ngeneric language model as well as a document-specific language model. Both can\nbe used as a smoothing component, but the latter is more adapted to the\ndocument at hand and has the potential of being used as a full document\nlanguage model. We experiment with such models and analyze their results on\nTREC-1 to 8 datasets.', 'This work explores the ability of collective matrix factorization models in\nrecommender systems to make predictions about users and items for which there\nis side information available but no feedback or interactions data, and\nproposes a new formulation with a faster cold-start prediction formula that can\nbe used in real-time systems. While these cold-start recommendations are not as\ngood as warm-start ones, they were found to be of better quality than\nnon-personalized recommendations, and predictions about new users were found to\nbe more reliable than those about new items. The formulation proposed here\nresulted in improved cold-start recommendations in many scenarios, at the\nexpense of worse warm-start ones.', 'Different people have different habits of describing their intents in\nconversations. Some people tend to deliberate their intents in several\nsuccessive utterances, i.e., they use several consistent messages for\nreadability instead of a long sentence to express their question. This creates\na predicament faced by the application of dialogue systems, especially in\nreal-world industry scenarios, in which the dialogue system is unsure whether\nit should answer the query of user immediately or wait for further\nsupplementary input. Motivated by such an interesting predicament, we define a\nnovel Wait-or-Answer task for dialogue systems. We shed light on a new research\ntopic about how the dialogue system can be more intelligent to behave in this\nWait-or-Answer quandary. Further, we propose a predictive approach named\nPredict-then-Decide (PTD) to tackle this Wait-or-Answer task. More\nspecifically, we take advantage of a decision model to help the dialogue system\ndecide whether to wait or answer. The decision of decision model is made with\nthe assistance of two ancillary prediction models: a user prediction and an\nagent prediction. The user prediction model tries to predict what the user\nwould supplement and uses its prediction to persuade the decision model that\nthe user has some information to add, so the dialogue system should wait. The\nagent prediction model tries to predict the answer of the dialogue system and\nconvince the decision model that it is a superior choice to answer the query of\nuser immediately since the input of user has come to an end. We conduct our\nexperiments on two real-life scenarios and three public datasets. Experimental\nresults on five datasets show our proposed PTD approach significantly\noutperforms the existing models in solving this Wait-or-Answer problem.', 'The recent advances introduced by neural machine translation (NMT) are\nrapidly expanding the application fields of machine translation, as well as\nreshaping the quality level to be targeted. In particular, if translations have\nto fit some given layout, quality should not only be measured in terms of\nadequacy and fluency, but also length. Exemplary cases are the translation of\ndocument files, subtitles, and scripts for dubbing, where the output length\nshould ideally be as close as possible to the length of the input text. This\npaper addresses for the first time, to the best of our knowledge, the problem\nof controlling the output length in NMT. We investigate two methods for biasing\nthe output length with a transformer architecture: i) conditioning the output\nto a given target-source length-ratio class and ii) enriching the transformer\npositional embedding with length information. Our experiments show that both\nmethods can induce the network to generate shorter translations, as well as\nacquiring interpretable linguistic skills.', ""How can humans remain in control of artificial intelligence (AI)-based\nsystems designed to perform tasks autonomously? Such systems are increasingly\nubiquitous, creating benefits - but also undesirable situations where moral\nresponsibility for their actions cannot be properly attributed to any\nparticular person or group. The concept of meaningful human control has been\nproposed to address responsibility gaps and mitigate them by establishing\nconditions that enable a proper attribution of responsibility for humans;\nhowever, clear requirements for researchers, designers, and engineers are yet\ninexistent, making the development of AI-based systems that remain under\nmeaningful human control challenging. In this paper, we address the gap between\nphilosophical theory and engineering practice by identifying, through an\niterative process of abductive thinking, four actionable properties for\nAI-based systems under meaningful human control, which we discuss making use of\ntwo applications scenarios: automated vehicles and AI-based hiring. First, a\nsystem in which humans and AI algorithms interact should have an explicitly\ndefined domain of morally loaded situations within which the system ought to\noperate. Second, humans and AI agents within the system should have appropriate\nand mutually compatible representations. Third, responsibility attributed to a\nhuman should be commensurate with that human's ability and authority to control\nthe system. Fourth, there should be explicit links between the actions of the\nAI agents and actions of humans who are aware of their moral responsibility. We\nargue that these four properties will support practically-minded professionals\nto take concrete steps toward designing and engineering for AI systems that\nfacilitate meaningful human control."", 'We study a security threat to reinforcement learning where an attacker\npoisons the learning environment to force the agent into executing a target\npolicy chosen by the attacker. As a victim, we consider RL agents whose\nobjective is to find a policy that maximizes average reward in undiscounted\ninfinite-horizon problem settings. The attacker can manipulate the rewards or\nthe transition dynamics in the learning environment at training-time and is\ninterested in doing so in a stealthy manner. We propose an optimization\nframework for finding an \\emph{optimal stealthy attack} for different measures\nof attack cost. We provide sufficient technical conditions under which the\nattack is feasible and provide lower/upper bounds on the attack cost. We\ninstantiate our attacks in two settings: (i) an \\emph{offline} setting where\nthe agent is doing planning in the poisoned environment, and (ii) an\n\\emph{online} setting where the agent is learning a policy using a\nregret-minimization framework with poisoned feedback. Our results show that the\nattacker can easily succeed in teaching any target policy to the victim under\nmild conditions and highlight a significant security threat to reinforcement\nlearning agents in practice.', ""The ability to comprehend wishes or desires and their fulfillment is\nimportant to Natural Language Understanding. This paper introduces the task of\nidentifying if a desire expressed by a subject in a given short piece of text\nwas fulfilled. We propose various unstructured and structured models that\ncapture fulfillment cues such as the subject's emotional state and actions. Our\nexperiments with two different datasets demonstrate the importance of\nunderstanding the narrative and discourse structure to address this task."", ""In many real-world tasks, it is not possible to procedurally specify an RL\nagent's reward function. In such cases, a reward function must instead be\nlearned from interacting with and observing humans. However, current techniques\nfor reward learning may fail to produce reward functions which accurately\nreflect user preferences. Absent significant advances in reward learning, it is\nthus important to be able to audit learned reward functions to verify whether\nthey truly capture user preferences. In this paper, we investigate techniques\nfor interpreting learned reward functions. In particular, we apply saliency\nmethods to identify failure modes and predict the robustness of reward\nfunctions. We find that learned reward functions often implement surprising\nalgorithms that rely on contingent aspects of the environment. We also discover\nthat existing interpretability techniques often attend to irrelevant changes in\nreward output, suggesting that reward interpretability may need significantly\ndifferent methods from policy interpretability."", 'A long-term goal of machine learning is to build intelligent conversational\nagents. One recent popular approach is to train end-to-end models on a large\namount of real dialog transcripts between humans (Sordoni et al., 2015; Vinyals\n& Le, 2015; Shang et al., 2015). However, this approach leaves many questions\nunanswered as an understanding of the precise successes and shortcomings of\neach model is hard to assess. A contrasting recent proposal are the bAbI tasks\n(Weston et al., 2015b) which are synthetic data that measure the ability of\nlearning machines at various reasoning tasks over toy language. Unfortunately,\nthose tests are very small and hence may encourage methods that do not scale.\nIn this work, we propose a suite of new tasks of a much larger scale that\nattempt to bridge the gap between the two regimes. Choosing the domain of\nmovies, we provide tasks that test the ability of models to answer factual\nquestions (utilizing OMDB), provide personalization (utilizing MovieLens),\ncarry short conversations about the two, and finally to perform on natural\ndialogs from Reddit. We provide a dataset covering 75k movie entities and with\n3.5M training examples. We present results of various models on these tasks,\nand evaluate their performance.', 'Determinantal point processes (DPPs) are well known models for diverse subset\nselection problems, including recommendation tasks, document summarization and\nimage search. In this paper, we discuss a greedy deterministic adaptation of\nk-DPP. Deterministic algorithms are interesting for many applications, as they\nprovide interpretability to the user by having no failure probability and\nalways returning the same results. First, the ability of the method to yield\nlow-rank approximations of kernel matrices is evaluated by comparing the\naccuracy of the Nystr\\""om approximation on multiple datasets. Afterwards, we\ndemonstrate the usefulness of the model on an image search task.', ""Along with the advance of opinion mining techniques, public mood has been\nfound to be a key element for stock market prediction. However, how market\nparticipants' behavior is affected by public mood has been rarely discussed.\nConsequently, there has been little progress in leveraging public mood for the\nasset allocation problem, which is preferred in a trusted and interpretable\nway. In order to address the issue of incorporating public mood analyzed from\nsocial media, we propose to formalize public mood into market views, because\nmarket views can be integrated into the modern portfolio theory. In our\nframework, the optimal market views will maximize returns in each period with a\nBayesian asset allocation model. We train two neural models to generate the\nmarket views, and benchmark the model performance on other popular asset\nallocation strategies. Our experimental results suggest that the formalization\nof market views significantly increases the profitability (5% to 10% annually)\nof the simulated portfolio at a given risk level."", 'The primary goal of ad-hoc retrieval (document retrieval in the context of\nquestion answering) is to find relevant documents satisfied the information\nneed posted in a natural language query. It requires a good understanding of\nthe query and all the documents in a corpus, which is difficult because the\nmeaning of natural language texts depends on the context, syntax,and semantics.\nRecently deep neural networks have been used to rank search results in response\nto a query. In this paper, we devise a multi-resolution neural network(MRNN) to\nleverage the whole hierarchy of representations for document retrieval. The\nproposed MRNN model is capable of deriving a representation that integrates\nrepresentations of different levels of abstraction from all the layers of the\nlearned hierarchical representation.Moreover, a duplex attention component is\ndesigned to refinethe multi-resolution representation so that an optimal\ncontextfor matching the query and document can be determined. More\nspecifically, the first attention mechanism determines optimal context from the\nlearned multi-resolution representation for the query and document. The latter\nattention mechanism aims to fine-tune the representation so that the query and\nthe relevant document are closer in proximity. The empirical study shows that\nMRNN with the duplex attention is significantly superior to existing models\nused for ad-hoc retrieval on benchmark datasets including SQuAD, WikiQA,\nQUASAR, and TrecQA.', ""Sentence position is a strong feature for news summarization, since the lead\noften (but not always) summarizes the key points of the article. In this paper,\nwe show that recent neural systems excessively exploit this trend, which\nalthough powerful for many inputs, is also detrimental when summarizing\ndocuments where important content should be extracted from later parts of the\narticle. We propose two techniques to make systems sensitive to the importance\nof content in different parts of the article. The first technique employs\n'unbiased' data; i.e., randomly shuffled sentences of the source document, to\npretrain the model. The second technique uses an auxiliary ROUGE-based loss\nthat encourages the model to distribute importance scores throughout a document\nby mimicking sentence-level ROUGE scores on the training data. We show that\nthese techniques significantly improve the performance of a competitive\nreinforcement learning based extractive system, with the auxiliary loss being\nmore powerful than pretraining."", 'A character network is a graph extracted from a narrative, in which vertices\nrepresent characters and edges correspond to interactions between them. A\nnumber of narrative-related problems can be addressed automatically through the\nanalysis of character networks, such as summarization, classification, or role\ndetection. Character networks are particularly relevant when considering works\nof fictions (e.g. novels, plays, movies, TV series), as their exploitation\nallows developing information retrieval and recommendation systems. However,\nworks of fiction possess specific properties making these tasks harder. This\nsurvey aims at presenting and organizing the scientific literature related to\nthe extraction of character networks from works of fiction, as well as their\nanalysis. We first describe the extraction process in a generic way, and\nexplain how its constituting steps are implemented in practice, depending on\nthe medium of the narrative, the goal of the network analysis, and other\nfactors. We then review the descriptive tools used to characterize character\nnetworks, with a focus on the way they are interpreted in this context. We\nillustrate the relevance of character networks by also providing a review of\napplications derived from their analysis. Finally, we identify the limitations\nof the existing approaches, and the most promising perspectives.', 'We present a novel model for the problem of ranking a collection of documents\naccording to their semantic similarity to a source (query) document. While the\nproblem of document-to-document similarity ranking has been studied, most\nmodern methods are limited to relatively short documents or rely on the\nexistence of ""ground-truth"" similarity labels. Yet, in most common real-world\ncases, similarity ranking is an unsupervised problem as similarity labels are\nunavailable. Moreover, an ideal model should not be restricted by documents\'\nlength. Hence, we introduce SDR, a self-supervised method for document\nsimilarity that can be applied to documents of arbitrary length. Importantly,\nSDR can be effectively applied to extremely long documents, exceeding the 4,096\nmaximal token limits of Longformer. Extensive evaluations on large document\ndatasets show that SDR significantly outperforms its alternatives across all\nmetrics. To accelerate future research on unlabeled long document similarity\nranking, and as an additional contribution to the community, we herein publish\ntwo human-annotated test sets of long documents similarity evaluation. The SDR\ncode and datasets are publicly available.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'Text Mining is a field that aims at extracting information from textual data.\nOne of the challenges of such field of study comes from the pre-processing\nstage in which a vector (and structured) representation should be extracted\nfrom unstructured data. The common extraction creates large and sparse vectors\nrepresenting the importance of each term to a document. As such, this usually\nleads to the curse-of-dimensionality that plagues most machine learning\nalgorithms. To cope with this issue, in this paper we propose a new supervised\nfeature extraction and reduction algorithm, named DCDistance, that creates\nfeatures based on the distance between a document to a representative of each\nclass label. As such, the proposed technique can reduce the features set in\nmore than 99% of the original set. Additionally, this algorithm was also\ncapable of improving the classification accuracy over a set of benchmark\ndatasets when compared to traditional and state-of-the-art features selection\nalgorithms.', 'Document AI, or Document Intelligence, is a relatively new research topic\nthat refers to the techniques for automatically reading, understanding, and\nanalyzing business documents. It is an important research direction for natural\nlanguage processing and computer vision. In recent years, the popularity of\ndeep learning technology has greatly advanced the development of Document AI,\nsuch as document layout analysis, visual information extraction, document\nvisual question answering, document image classification, etc. This paper\nbriefly reviews some of the representative models, tasks, and benchmark\ndatasets. Furthermore, we also introduce early-stage heuristic rule-based\ndocument analysis, statistical machine learning algorithms, and deep learning\napproaches especially pre-training methods. Finally, we look into future\ndirections for Document AI research.', 'This paper discusses learning a structured feedback control to obtain\nsufficient robustness to exogenous inputs for linear dynamic systems with\nunknown state matrix. The structural constraint on the controller is necessary\nfor many cyber-physical systems, and our approach presents a design for any\ngeneric structure, paving the way for distributed learning control. The ideas\nfrom reinforcement learning (RL) in conjunction with control-theoretic\nsufficient stability and performance guarantees are used to develop the\nmethodology. First, a model-based framework is formulated using dynamic\nprogramming to embed the structural constraint in the linear quadratic\nregulator (LQR) setting along with sufficient robustness conditions.\nThereafter, we translate these conditions to a data-driven learning-based\nframework - robust structured reinforcement learning (RSRL) that enjoys the\ncontrol-theoretic guarantees on stability and convergence. We validate our\ntheoretical results with a simulation on a multi-agent network with 6 agents.', 'As a new type of e-commerce platform developed in recent years, local\nconsumer service platform provides users with software to consume service to\nthe nearby store or to the home, such as Groupon and Koubei. Different from\nother common e-commerce platforms, the behavior of users on the local consumer\nservice platform is closely related to their real-time local context\ninformation. Therefore, building a context-aware user behavior prediction\nsystem is able to provide both merchants and users better service in local\nconsumer service platforms. However, most of the previous work just treats the\ncontextual information as an ordinary feature into the prediction model to\nobtain the prediction list under a specific context, which ignores the fact\nthat the interest of a user in different contexts is often significantly\ndifferent. Hence, in this paper, we propose a context-aware heterogeneous graph\nattention network (CHGAT) to dynamically generate the representation of the\nuser and to estimate the probability for future behavior. Specifically, we\nfirst construct the meta-path based heterogeneous graphs with the historical\nbehaviors from multiple sources and comprehend heterogeneous vertices in the\ngraph with a novel unified knowledge representing approach. Next, a multi-level\nattention mechanism is introduced for context-aware aggregation with graph\nvertices, which contains the vertex-level attention network and the path-level\nattention network. Both of them aim to capture the semantic correlation between\ninformation contained in the graph and the outside real-time contextual\ninformation in the search system. Then the model proposed in this paper\naggregates specific graphs with their corresponding context features and\nobtains the representation of user interest under a specific context and input\nit into the prediction network to finally obtain the predicted probability of\nuser behavior.', 'Deciding which scripts to turn into movies is a costly and time-consuming\nprocess for filmmakers. Thus, building a tool to aid script selection, an\ninitial phase in movie production, can be very beneficial. Toward that goal, in\nthis work, we present a method to evaluate the quality of a screenplay based on\nlinguistic cues. We address this in a two-fold approach: (1) we define the task\nas predicting nominations of scripts at major film awards with the hypothesis\nthat the peer-recognized scripts should have a greater chance to succeed. (2)\nbased on industry opinions and narratology, we extract and integrate\ndomain-specific features into common classification techniques. We face two\nchallenges (1) scripts are much longer than other document datasets (2)\nnominated scripts are limited and thus difficult to collect. However, with\nnarratology-inspired modeling and domain features, our approach offers clear\nimprovements over strong baselines. Our work provides a new approach for future\nwork in screenplay analysis.', ""AI systems often rely on two key components: a specified goal or reward\nfunction and an optimization algorithm to compute the optimal behavior for that\ngoal. This approach is intended to provide value for a principal: the user on\nwhose behalf the agent acts. The objectives given to these agents often refer\nto a partial specification of the principal's goals. We consider the cost of\nthis incompleteness by analyzing a model of a principal and an agent in a\nresource constrained world where the $L$ attributes of the state correspond to\ndifferent sources of utility for the principal. We assume that the reward\nfunction given to the agent only has support on $J < L$ attributes. The\ncontributions of our paper are as follows: 1) we propose a novel model of an\nincomplete principal-agent problem from artificial intelligence; 2) we provide\nnecessary and sufficient conditions under which indefinitely optimizing for any\nincomplete proxy objective leads to arbitrarily low overall utility; and 3) we\nshow how modifying the setup to allow reward functions that reference the full\nstate or allowing the principal to update the proxy objective over time can\nlead to higher utility solutions. The results in this paper argue that we\nshould view the design of reward functions as an interactive and dynamic\nprocess and identifies a theoretical scenario where some degree of\ninteractivity is desirable."", 'Reward functions are easy to misspecify; although designers can make\ncorrections after observing mistakes, an agent pursuing a misspecified reward\nfunction can irreversibly change the state of its environment. If that change\nprecludes optimization of the correctly specified reward function, then\ncorrection is futile. For example, a robotic factory assistant could break\nexpensive equipment due to a reward misspecification; even if the designers\nimmediately correct the reward function, the damage is done. To mitigate this\nrisk, we introduce an approach that balances optimization of the primary reward\nfunction with preservation of the ability to optimize auxiliary reward\nfunctions. Surprisingly, even when the auxiliary reward functions are randomly\ngenerated and therefore uninformative about the correctly specified reward\nfunction, this approach induces conservative, effective behavior.', 'Reward is the driving force for reinforcement-learning agents. This paper is\ndedicated to understanding the expressivity of reward as a way to capture tasks\nthat we would want an agent to perform. We frame this study around three new\nabstract notions of ""task"" that might be desirable: (1) a set of acceptable\nbehaviors, (2) a partial ordering over behaviors, or (3) a partial ordering\nover trajectories. Our main results prove that while reward can express many of\nthese tasks, there exist instances of each task type that no Markov reward\nfunction can capture. We then provide a set of polynomial-time algorithms that\nconstruct a Markov reward function that allows an agent to optimize tasks of\neach of these three types, and correctly determine when no such reward function\nexists. We conclude with an empirical study that corroborates and illustrates\nour theoretical findings.', 'We present an empirical investigation of pre-trained Transformer-based\nauto-regressive language models for the task of open-domain dialogue\ngeneration. Training paradigm of pre-training and fine-tuning is employed to\nconduct the parameter learning. Corpora of News and Wikipedia in Chinese and\nEnglish are collected for the pre-training stage respectively. Dialogue context\nand response are concatenated into a single sequence utilized as the input of\nthe models during the fine-tuning stage. A weighted joint prediction paradigm\nfor both context and response is designed to evaluate the performance of models\nwith or without the loss term for context prediction. Various of decoding\nstrategies such as greedy search, beam search, top-k sampling, etc. are\nemployed to conduct the response text generation. Extensive experiments are\nconducted on the typical single-turn and multi-turn dialogue corpora such as\nWeibo, Douban, Reddit, DailyDialog, and Persona-Chat. Detailed numbers of\nautomatic evaluation metrics on relevance and diversity of the generated\nresults for the languages models as well as the baseline approaches are\nreported.', ""While most conversational AI systems focus on textual dialogue only,\nconditioning utterances on visual context (when it's available) can lead to\nmore realistic conversations. Unfortunately, a major challenge for\nincorporating visual context into conversational dialogue is the lack of\nlarge-scale labeled datasets. We provide a solution in the form of a new\nvisually conditioned Future Utterance Prediction task. Our task involves\npredicting the next utterance in a video, using both visual frames and\ntranscribed speech as context. By exploiting the large number of instructional\nvideos online, we train a model to solve this task at scale, without the need\nfor manual annotations. Leveraging recent advances in multimodal learning, our\nmodel consists of a novel co-attentional multimodal video transformer, and when\ntrained on both textual and visual context, outperforms baselines that use\ntextual inputs alone. Further, we demonstrate that our model trained for this\ntask on unlabelled videos achieves state-of-the-art performance on a number of\ndownstream VideoQA benchmarks such as MSRVTT-QA, MSVD-QA, ActivityNet-QA and\nHow2QA."", ""Most popular goal-oriented dialogue agents are capable of understanding the\nconversational context. However, with the surge of virtual assistants with\nscreen, the next generation of agents are required to also understand screen\ncontext in order to provide a proper interactive experience, and better\nunderstand users' goals. In this paper, we propose a novel multimodal\nconversational framework, where the dialogue agent's next action and their\narguments are derived jointly conditioned both on the conversational and the\nvisual context. Specifically, we propose a new model, that can reason over the\nvisual context within a conversation and populate API arguments with visual\nentities given the user query. Our model can recognize visual features such as\ncolor and shape as well as the metadata based features such as price or star\nrating associated with a visual entity. In order to train our model, due to a\nlack of suitable multimodal conversational datasets, we also propose a novel\nmultimodal dialog simulator to generate synthetic data and also collect\nrealistic user data from MTurk to improve model robustness. The proposed model\nachieves a reasonable 85% model accuracy, without high inference latency. We\nalso demonstrate the proposed approach in a prototypical furniture shopping\nexperience for a multimodal virtual assistant."", 'Based on recent advances in realistic language modeling (GPT-3) and\ncross-modal representations (CLIP), Gaud\\\'i was developed to help designers\nsearch for inspirational images using natural language. In the early stages of\nthe design process, with the goal of eliciting a client\'s preferred creative\ndirection, designers will typically create thematic collections of\ninspirational images called ""mood-boards"". Creating a mood-board involves\nsequential image searches which are currently performed using keywords or\nimages. Gaud\\\'i transforms this process into a conversation where the user is\ngradually detailing the mood-board\'s theme. This representation allows our AI\nto generate new search queries from scratch, straight from a project briefing,\nfollowing a theme hypothesized by GPT-3. Compared to previous computational\napproaches to mood-board creation, to the best of our knowledge, ours is the\nfirst attempt to represent mood-boards as the stories that designers tell when\npresenting a creative direction to a client.', 'Recent studies emphasize the need of document context in human evaluation of\nmachine translations, but little research has been done on the impact of user\ninterfaces on annotator productivity and the reliability of assessments. In\nthis work, we compare human assessment data from the last two WMT evaluation\ncampaigns collected via two different methods for document-level evaluation.\nOur analysis shows that a document-centric approach to evaluation where the\nannotator is presented with the entire document context on a screen leads to\nhigher quality segment and document level assessments. It improves the\ncorrelation between segment and document scores and increases inter-annotator\nagreement for document scores but is considerably more time consuming for\nannotators.', 'In the context of investigative journalism, we address the problem of\nautomatically identifying which claims in a given document are most worthy and\nshould be prioritized for fact-checking. Despite its importance, this is a\nrelatively understudied problem. Thus, we create a new dataset of political\ndebates, containing statements that have been fact-checked by nine reputable\nsources, and we train machine learning models to predict which claims should be\nprioritized for fact-checking, i.e., we model the problem as a ranking task.\nUnlike previous work, which has looked primarily at sentences in isolation, in\nthis paper we focus on a rich input representation modeling the context:\nrelationship between the target statement and the larger context of the debate,\ninteraction between the opponents, and reaction by the moderator and by the\npublic. Our experiments show state-of-the-art results, outperforming a strong\nrivaling system by a margin, while also confirming the importance of the\ncontextual information.', 'Automatic question generation is an important technique that can improve the\ntraining of question answering, help chatbots to start or continue a\nconversation with humans, and provide assessment materials for educational\npurposes. Existing neural question generation models are not sufficient mainly\ndue to their inability to properly model the process of how each word in the\nquestion is selected, i.e., whether repeating the given passage or being\ngenerated from a vocabulary. In this paper, we propose our Clue Guided Copy\nNetwork for Question Generation (CGC-QG), which is a sequence-to-sequence\ngenerative model with copying mechanism, yet employing a variety of novel\ncomponents and techniques to boost the performance of question generation. In\nCGC-QG, we design a multi-task labeling strategy to identify whether a question\nword should be copied from the input passage or be generated instead, guiding\nthe model to learn the accurate boundaries between copying and generation.\nFurthermore, our input passage encoder takes as input, among a diverse range of\nother features, the prediction made by a clue word predictor, which helps\nidentify whether each word in the input passage is a potential clue to be\ncopied into the target question. The clue word predictor is designed based on a\nnovel application of Graph Convolutional Networks onto a syntactic dependency\ntree representation of each passage, thus being able to predict clue words only\nbased on their context in the passage and their relative positions to the\nanswer in the tree. We jointly train the clue prediction as well as question\ngeneration with multi-task learning and a number of practical strategies to\nreduce the complexity. Extensive evaluations show that our model significantly\nimproves the performance of question generation and out-performs all previous\nstate-of-the-art neural question generation models by a substantial margin.', 'With the increase of information, document classification as one of the\nmethods of text mining, plays vital role in many management and organizing\ninformation. Document classification is the process of assigning a document to\none or more predefined category labels. Document classification includes\ndifferent parts such as text processing, term selection, term weighting and\nfinal classification. The accuracy of document classification is very\nimportant. Thus improvement in each part of classification should lead to\nbetter results and higher precision. Term weighting has a great impact on the\naccuracy of the classification. Most of the existing weighting methods exploit\nthe statistical information of terms in documents and do not consider semantic\nrelations between words. In this paper, an automated document classification\nsystem is presented that uses a novel term weighting method based on semantic\nrelations between terms. To evaluate the proposed method, three standard\nPersian corpuses are used. Experiment results show 2 to 4 percent improvement\nin classification accuracy compared with the best previous designed system for\nPersian documents.', 'The creation of fictional stories is a very complex task that usually implies\na creative process where the author has to combine characters, conflicts and\nplots to create an engaging narrative. This work presents a simulated\nenvironment with hundreds of characters that allows the study of coherent and\ninteresting literary archetypes (or behaviours), plots and sub-plots. We will\nuse this environment to perform a study about the number of profiles\n(parameters that define the personality of a character) needed to create two\nemergent scenes of archetypes: ""natality control"" and ""revenge"". A Genetic\nAlgorithm (GA) will be used to find the fittest number of profiles and\nparameter configuration that enables the existence of the desired archetypes\n(played by the characters without their explicit knowledge). The results show\nthat parametrizing this complex system is possible and that these kind of\narchetypes can emerge in the given environment.', 'Term frequency normalization is a serious issue since lengths of documents\nare various. Generally, documents become long due to two different reasons -\nverbosity and multi-topicality. First, verbosity means that the same topic is\nrepeatedly mentioned by terms related to the topic, so that term frequency is\nmore increased than the well-summarized one. Second, multi-topicality indicates\nthat a document has a broad discussion of multi-topics, rather than single\ntopic. Although these document characteristics should be differently handled,\nall previous methods of term frequency normalization have ignored these\ndifferences and have used a simplified length-driven approach which decreases\nthe term frequency by only the length of a document, causing an unreasonable\npenalization. To attack this problem, we propose a novel TF normalization\nmethod which is a type of partially-axiomatic approach. We first formulate two\nformal constraints that the retrieval model should satisfy for documents having\nverbose and multi-topicality characteristic, respectively. Then, we modify\nlanguage modeling approaches to better satisfy these two constraints, and\nderive novel smoothing methods. Experimental results show that the proposed\nmethod increases significantly the precision for keyword queries, and\nsubstantially improves MAP (Mean Average Precision) for verbose queries.', 'We study a security threat to reinforcement learning where an attacker\npoisons the learning environment to force the agent into executing a target\npolicy chosen by the attacker. As a victim, we consider RL agents whose\nobjective is to find a policy that maximizes reward in infinite-horizon problem\nsettings. The attacker can manipulate the rewards and the transition dynamics\nin the learning environment at training-time, and is interested in doing so in\na stealthy manner. We propose an optimization framework for finding an optimal\nstealthy attack for different measures of attack cost. We provide lower/upper\nbounds on the attack cost, and instantiate our attacks in two settings: (i) an\noffline setting where the agent is doing planning in the poisoned environment,\nand (ii) an online setting where the agent is learning a policy with poisoned\nfeedback. Our results show that the attacker can easily succeed in teaching any\ntarget policy to the victim under mild conditions and highlight a significant\nsecurity threat to reinforcement learning agents in practice.', ""Autonomous agents optimize the reward function we give them. What they don't\nknow is how hard it is for us to design a reward function that actually\ncaptures what we want. When designing the reward, we might think of some\nspecific training scenarios, and make sure that the reward will lead to the\nright behavior in those scenarios. Inevitably, agents encounter new scenarios\n(e.g., new types of terrain) where optimizing that same reward may lead to\nundesired behavior. Our insight is that reward functions are merely\nobservations about what the designer actually wants, and that they should be\ninterpreted in the context in which they were designed. We introduce inverse\nreward design (IRD) as the problem of inferring the true objective based on the\ndesigned reward and the training MDP. We introduce approximate methods for\nsolving IRD problems, and use their solution to plan risk-averse behavior in\ntest MDPs. Empirical results suggest that this approach can help alleviate\nnegative side effects of misspecified reward functions and mitigate reward\nhacking."", 'The increasing amount of online content motivated the development of\nmulti-document summarization methods. In this work, we explore straightforward\napproaches to extend single-document summarization methods to multi-document\nsummarization. The proposed methods are based on the hierarchical combination\nof single-document summaries, and achieves state of the art results.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.', 'It is essential for dialogue-based spatial reasoning systems to maintain\nmemory of historical states of the world. In addition to conveying that the\ndialogue agent is mentally present and engaged with the task, referring to\nhistorical states may be crucial for enabling collaborative planning (e.g., for\nplanning to return to a previous state, or diagnosing a past misstep). In this\npaper, we approach the problem of spatial memory in a multi-modal spoken\ndialogue system capable of answering questions about interaction history in a\nphysical blocks world setting. This work builds upon a full spatial\nquestion-answering pipeline consisting of a vision system, speech input and\noutput mediated by an animated avatar, a dialogue system that robustly\ninterprets spatial queries, and a constraint solver that derives answers based\non 3-D spatial modelling. The contributions of this work include a symbolic\ndialogue context registering knowledge about discourse history and changes in\nthe world, as well as a natural language understanding module capable of\ninterpreting free-form historical questions and querying the dialogue context\nto form an answer.', 'Mood recognition is an important problem in music informatics and has key\napplications in music discovery and recommendation. These applications have\nbecome even more relevant with the rise of music streaming. Our work\ninvestigates the research question of whether we can leverage audio metadata\nsuch as artist and year, which is readily available, to improve the performance\nof mood classification models. To this end, we propose a multi-task learning\napproach in which a shared model is simultaneously trained for mood and\nmetadata prediction tasks with the goal to learn richer representations.\nExperimentally, we demonstrate that applying our technique on the existing\nstate-of-the-art convolutional neural networks for mood classification improves\ntheir performances consistently. We conduct experiments on multiple datasets\nand report that our approach can lead to improvements in the average precision\nmetric by up to 8.7 points.', ""Probability theory, epistemically interpreted, provides an excellent, if not\nthe best available account of inductive reasoning. This is so because there are\ngeneral and definite rules for the change of subjective probabilities through\ninformation or experience; induction and belief change are one and same topic,\nafter all. The most basic of these rules is simply to conditionalize with\nrespect to the information received; and there are similar and more general\nrules. 1 Hence, a fundamental reason for the epistemological success of\nprobability theory is that there at all exists a well-behaved concept of\nconditional probability. Still, people have, and have reasons for, various\nconcerns over probability theory. One of these is my starting point:\nIntuitively, we have the notion of plain belief; we believe propositions2 to be\ntrue (or to be false or neither). Probability theory, however, offers no formal\ncounterpart to this notion. Believing A is not the same as having probability 1\nfor A, because probability 1 is incorrigible3; but plain belief is clearly\ncorrigible. And believing A is not the same as giving A a probability larger\nthan some 1 - c, because believing A and believing B is usually taken to be\nequivalent to believing A & B.4 Thus, it seems that the formal representation\nof plain belief has to take a non-probabilistic route. Indeed, representing\nplain belief seems easy enough: simply represent an epistemic state by the set\nof all propositions believed true in it or, since I make the common assumption\nthat plain belief is deductively closed, by the conjunction of all propositions\nbelieved true in it. But this does not yet provide a theory of induction, i.e.\nan answer to the question how epistemic states so represented are changed\ntbrough information or experience. There is a convincing partial answer: if the\nnew information is compatible with the old epistemic state, then the new\nepistemic state is simply represented by the conjunction of the new information\nand the old beliefs. This answer is partial because it does not cover the quite\ncommon case where the new information is incompatible with the old beliefs. It\nis, however, important to complete the answer and to cover this case, too;\notherwise, we would not represent plain belief as conigible. The crucial\nproblem is that there is no good completion. When epistemic states are\nrepresented simply by the conjunction of all propositions believed true in it,\nthe answer cannot be completed; and though there is a lot of fruitful work, no\nother representation of epistemic states has been proposed, as far as I know,\nwhich provides a complete solution to this problem. In this paper, I want to\nsuggest such a solution. In [4], I have more fully argued that this is the only\nsolution, if certain plausible desiderata are to be satisfied. Here, in section\n2, I will be content with formally defining and intuitively explaining my\nproposal. I will compare my proposal with probability theory in section 3. It\nwill turn out that the theory I am proposing is structurally homomorphic to\nprobability theory in important respects and that it is thus equally easily\nimplementable, but moreover computationally simpler. Section 4 contains a very\nbrief comparison with various kinds of logics, in particular conditional logic,\nwith Shackle's functions of potential surprise and related theories, and with\nthe Dempster - Shafer theory of belief functions."", 'Text alignment finds application in tasks such as citation recommendation and\nplagiarism detection. Existing alignment methods operate at a single,\npredefined level and cannot learn to align texts at, for example, sentence and\ndocument levels. We propose a new learning approach that equips previously\nestablished hierarchical attention encoders for representing documents with a\ncross-document attention component, enabling structural comparisons across\ndifferent levels (document-to-document and sentence-to-document). Our component\nis weakly supervised from document pairs and can align at multiple levels. Our\nevaluation on predicting document-to-document relationships and\nsentence-to-document relationships on the tasks of citation recommendation and\nplagiarism detection shows that our approach outperforms previously established\nhierarchical, attention encoders based on recurrent and transformer\ncontextualization that are unaware of structural correspondence between\ndocuments.', 'We investigate the pertinence of methods from algebraic topology for text\ndata analysis. These methods enable the development of\nmathematically-principled isometric-invariant mappings from a set of vectors to\na document embedding, which is stable with respect to the geometry of the\ndocument in the selected metric space. In this work, we evaluate the utility of\nthese topology-based document representations in traditional NLP tasks,\nspecifically document clustering and sentiment classification. We find that the\nembeddings do not benefit text analysis. In fact, performance is worse than\nsimple techniques like $\\textit{tf-idf}$, indicating that the geometry of the\ndocument does not provide enough variability for classification on the basis of\ntopic or sentiment in the chosen datasets.', 'Conceptual entanglement is a crucial phenomenon in quantum cognition because\nit implies that classical probabilities cannot model non--compositional\nconceptual phenomena. While several psychological experiments have been\ndeveloped to test conceptual entanglement, this has not been explored in the\ncontext of Natural Language Processing. In this paper, we apply the hypothesis\nthat words of a document are traces of the concepts that a person has in mind\nwhen writing the document. Therefore, if these concepts are entangled, we\nshould be able to observe traces of their entanglement in the documents. In\nparticular, we test conceptual entanglement by contrasting language simulations\nwith results obtained from a text corpus. Our analysis indicates that\nconceptual entanglement is strongly linked to the way in which language is\nstructured. We discuss the implications of this finding in the context of\nconceptual modeling and of Natural Language Processing.', 'Grammatical tense and mood are important linguistic phenomena to consider in\nnatural language processing (NLP) research. We consider the correspondence\nbetween English and German tense and mood in translation. Human translators do\nnot find this correspondence easy, and as we will show through careful\nanalysis, there are no simplistic ways to map tense and mood from one language\nto another. Our observations about the challenges of human translation of tense\nand mood have important implications for multilingual NLP. Of particular\nimportance is the challenge of modeling tense and mood in rule-based,\nphrase-based statistical and neural machine translation.', 'Today all kind of information is getting digitized and along with all this\ndigitization, the huge archive of various kinds of documents is being digitized\ntoo. We know that, Optical Character Recognition is the method through which,\nnewspapers and other paper documents convert into digital resources. But, it is\na fact that this method works on texts only. As a result, if we try to process\nany document which contains non-textual zones, then we will get garbage texts\nas output. That is why; in order to digitize documents properly they should be\nprepossessed carefully. And while preprocessing, segmenting document in\ndifferent regions according to the category properly is most important. But,\nthe Optical Character Recognition processes available for Bangla language have\nno such algorithm that can categorize a newspaper/book page fully. So we worked\nto decompose a document into its several parts like headlines, sub headlines,\ncolumns, images etc. And if the input is skewed and rotated, then the input was\nalso deskewed and de-rotated. To decompose any Bangla document we found out the\nedges of the input image. Then we find out the horizontal and vertical area of\nevery pixel where it lies in. Later on the input image was cut according to\nthese areas. Then we pick each and every sub image and found out their\nheight-width ratio, line height. Then according to these values the sub images\nwere categorized. To deskew the image we found out the skew angle and de skewed\nthe image according to this angle. To de-rotate the image we used the line\nheight, matra line, pixel ratio of matra line.', ""In this paper, I will discuss the work I am currently doing as a Ph.D.\nstudent at the University of Potsdam, under the tutoring of T. Schaub. I'm\ncurrently looking into action description in ASP. More precisely, my goal is to\nexplore how to represent actions with durations in ASP, in different contexts.\nRight now, I'm focused on Multi-Agent Path Finding (MAPF), looking at how to\nrepresent speeds for different agents and contexts.\n  Before tackling duration, I wanted to explore and compare different\nrepresentations of action taking in ASP. For this, I started comparing\ndifferent simple encodings tackling the MAPF problem. Even in simple code,\nchoices and assumptions have been made in their creations. The objective of my\nwork is to present the consequences of those design decisions in terms of\nperformance and knowledge representation. As far as I know, there is no current\nresearch on this topic.\n  Besides that, I'm also exploring different ways to represent duration and to\nsolve related problems. I planed to compare them the same way I described\nbefore. I also want this to help me find innovative and effective ways to solve\nproblems with duration."", ""To train a statistical spoken dialogue system (SDS) it is essential that an\naccurate method for measuring task success is available. To date training has\nrelied on presenting a task to either simulated or paid users and inferring the\ndialogue's success by observing whether this presented task was achieved or\nnot. Our aim however is to be able to learn from real users acting under their\nown volition, in which case it is non-trivial to rate the success as any prior\nknowledge of the task is simply unavailable. User feedback may be utilised but\nhas been found to be inconsistent. Hence, here we present two neural network\nmodels that evaluate a sequence of turn-level features to rate the success of a\ndialogue. Importantly these models make no use of any prior knowledge of the\nuser's task. The models are trained on dialogues generated by a simulated user\nand the best model is then used to train a policy on-line which is shown to\nperform at least as well as a baseline system using prior knowledge of the\nuser's task. We note that the models should also be of interest for evaluating\nSDS and for monitoring a dialogue in rule-based SDS."", 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'The coordination and cooperation of all the stakeholders involved is a\ndecisive point for the control and the resolution of problems. In the\ninsecurity events, the resolution should refer to a plan that defines a general\nframework of the procedures to be undertaken and the instructions to be\ncomplied with; also, a more precise process must be defined by the actors to\ndeal with the case represented by the particular problem of the current\nsituation. Indeed, this process has to cope with a dynamic, unstable and\nunpredictable environment, due to the heterogeneity and multiplicity of\nstakeholders, and finally due to their possible geographical distribution. In\nthis article, we will present the first steps of validation of a collaborative\ndecision-making approach in the context of crisis situations such as road\naccidents. This approach is based on ontologies and multi-agent systems.', 'A mind-map is a diagram that represents the central concept and key ideas in\na hierarchical way. Converting plain text into a mind-map will reveal its key\nsemantic structure and be easier to understand. Given a document, the existing\nautomatic mind-map generation method extracts the relationships of every\nsentence pair to generate the directed semantic graph for this document. The\ncomputation complexity increases exponentially with the length of the document.\nMoreover, it is difficult to capture the overall semantics. To deal with the\nabove challenges, we propose an efficient mind-map generation network that\nconverts a document into a graph via sequence-to-graph. To guarantee a\nmeaningful mind-map, we design a graph refinement module to adjust the relation\ngraph in a reinforcement learning manner. Extensive experimental results\ndemonstrate that the proposed approach is more effective and efficient than the\nexisting methods. The inference time is reduced by thousands of times compared\nwith the existing methods. The case studies verify that the generated mind-maps\nbetter reveal the underlying semantic structures of the document.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', 'We address the problem of solving mixed random linear equations. We have\nunlabeled observations coming from multiple linear regressions, and each\nobservation corresponds to exactly one of the regression models. The goal is to\nlearn the linear regressors from the observations. Classically, Alternating\nMinimization (AM) (which is a variant of Expectation Maximization (EM)) is used\nto solve this problem. AM iteratively alternates between the estimation of\nlabels and solving the regression problems with the estimated labels.\nEmpirically, it is observed that, for a large variety of non-convex problems\nincluding mixed linear regression, AM converges at a much faster rate compared\nto gradient based algorithms. However, the existing theory suggests similar\nrate of convergence for AM and gradient based methods, failing to capture this\nempirical behavior. In this paper, we close this gap between theory and\npractice for the special case of a mixture of $2$ linear regressions. We show\nthat, provided initialized properly, AM enjoys a \\emph{super-linear} rate of\nconvergence in certain parameter regimes. To the best of our knowledge, this is\nthe first work that theoretically establishes such rate for AM. Hence, if we\nwant to recover the unknown regressors upto an error (in $\\ell_2$ norm) of\n$\\epsilon$, AM only takes $\\mathcal{O}(\\log \\log (1/\\epsilon))$ iterations.\nFurthermore, we compare AM with a gradient based heuristic algorithm\nempirically and show that AM dominates in iteration complexity as well as\nwall-clock time.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I aim to create a chatbot capable of interpreting images as input and analyzing them for subsequent tasks. The model should deliver responses efficiently without compromising the quality of the answers. To achieve this, I require an image analysis model that can swiftly perform image classification by ignoring unnecessary local detailed features. Additionally, the model should be capable of executing pixel-level localization and analysis as required. To enhance efficiency, I plan to incorporate reinforcement learning for user prediction. I envision the chatbot predicting the user's potential next questions, allowing the image analysis model to analyze relevant images while awaiting user input. The reward component of the framework should be directly linked to the accuracy of the prediction and the time saved. Furthermore, I want the method to remember the most frequently asked questions related to popular images, enabling it to provide immediate answers to these queries.",[],"['The accuracy of deep convolutional neural networks (CNNs) generally improves\nwhen fueled with high resolution images. However, this often comes at a high\ncomputational cost and high memory footprint. Inspired by the fact that not all\nregions in an image are task-relevant, we propose a novel framework that\nperforms efficient image classification by processing a sequence of relatively\nsmall inputs, which are strategically selected from the original image with\nreinforcement learning. Such a dynamic decision process naturally facilitates\nadaptive inference at test time, i.e., it can be terminated once the model is\nsufficiently confident about its prediction and thus avoids further redundant\ncomputation. Notably, our framework is general and flexible as it is compatible\nwith most of the state-of-the-art light-weighted CNNs (such as MobileNets,\nEfficientNets and RegNets), which can be conveniently deployed as the backbone\nfeature extractor. Experiments on ImageNet show that our method consistently\nimproves the computational efficiency of a wide variety of deep models. For\nexample, it further reduces the average latency of the highly efficient\nMobileNet-V3 on an iPhone XS Max by 20% without sacrificing accuracy. Code and\npre-trained models are available at\nhttps://github.com/blackfeather-wang/GFNet-Pytorch.', 'Document image classification remains a popular research area because it can\nbe commercialized in many enterprise applications across different industries.\nRecent advancements in large pre-trained computer vision and language models\nand graph neural networks has lent document image classification many tools.\nHowever using large pre-trained models usually requires substantial computing\nresources which could defeat the cost-saving advantages of automatic document\nimage classification. In the paper we propose an efficient document image\nclassification framework that uses graph convolution neural networks and\nincorporates textual, visual and layout information of the document. We have\nrigorously benchmarked our proposed algorithm against several state-of-art\nvision and language models on both publicly available dataset and a real-life\ninsurance document classification dataset. Empirical results on both publicly\navailable and real-world data show that our methods achieve near SOTA\nperformance yet require much less computing resources and time for model\ntraining and inference. This results in solutions than offer better cost\nadvantages, especially in scalable deployment for enterprise applications. The\nresults showed that our algorithm can achieve classification performance quite\nclose to SOTA. We also provide comprehensive comparisons of computing\nresources, model sizes, train and inference time between our proposed methods\nand baselines. In addition we delineate the cost per image using our method and\nother baselines.', 'As a training and analysis strategy for convolutional neural networks (CNNs),\nwe slice images into tiled segments and use, for training and prediction,\nsegments that both satisfy a criterion of information diversity and contain\nsufficient content to support classification. In particular, we utilize image\nentropy as the diversity criterion. This ensures that each tile carries as much\ninformation diversity as the original image, and for many applications serves\nas an indicator of usefulness in classification. To make predictions, a\nprobability aggregation framework is applied to probabilities assigned by the\nCNN to the input image tiles. This technique facilitates the use of large,\nhigh-resolution images that would be impractical to analyze unmodified;\nprovides data augmentation for training, which is particularly valuable when\nimage availability is limited; and the ensemble nature of the input for\nprediction enhances its accuracy.', 'Automatically generating maps from satellite images is an important task.\nThere is a body of literature which tries to address this challenge. We created\na more expansive survey of the task by experimenting with different models and\nadding new loss functions to improve results. We created a database of pairs of\nsatellite images and the corresponding map of the area. Our model translates\nthe satellite image to the corresponding standard layer map image using three\nmain model architectures: (i) a conditional Generative Adversarial Network\n(GAN) which compresses the images down to a learned embedding, (ii) a generator\nwhich is trained as a normalizing flow (RealNVP) model, and (iii) a conditional\nGAN where the generator translates via a series of convolutions to the standard\nlayer of a map and the discriminator input is the concatenation of the\nreal/generated map and the satellite image. Model (iii) was by far the most\npromising of three models. To improve the results we also added a\nreconstruction loss and style transfer loss in addition to the GAN losses. The\nthird model architecture produced the best quality of sampled images. In\ncontrast to the other generative model where evaluation of the model is a\nchallenging problem. since we have access to the real map for a given satellite\nimage, we are able to assign a quantitative metric to the quality of the\ngenerated images in addition to inspecting them visually. While we are\ncontinuing to work on increasing the accuracy of the model, one challenge has\nbeen the coarse resolution of the data which upper-bounds the quality of the\nresults of our model. Nevertheless, as will be seen in the results, the\ngenerated map is more accurate in the features it produces since the generator\narchitecture demands a pixel-wise image translation/pixel-wise coloring. A\nvideo presentation summarizing this paper is available at:\nhttps://youtu.be/Ur0flOX-Ji0', 'Band selection refers to the process of choosing the most relevant bands in a\nhyperspectral image. By selecting a limited number of optimal bands, we aim at\nspeeding up model training, improving accuracy, or both. It reduces redundancy\namong spectral bands while trying to preserve the original information of the\nimage. By now many efforts have been made to develop unsupervised band\nselection approaches, of which the majority are heuristic algorithms devised by\ntrial and error. In this paper, we are interested in training an intelligent\nagent that, given a hyperspectral image, is capable of automatically learning\npolicy to select an optimal band subset without any hand-engineered reasoning.\nTo this end, we frame the problem of unsupervised band selection as a Markov\ndecision process, propose an effective method to parameterize it, and finally\nsolve the problem by deep reinforcement learning. Once the agent is trained, it\nlearns a band-selection policy that guides the agent to sequentially select\nbands by fully exploiting the hyperspectral image and previously picked bands.\nFurthermore, we propose two different reward schemes for the environment\nsimulation of deep reinforcement learning and compare them in experiments.\nThis, to the best of our knowledge, is the first study that explores a deep\nreinforcement learning model for hyperspectral image analysis, thus opening a\nnew door for future research and showcasing the great potential of deep\nreinforcement learning in remote sensing applications. Extensive experiments\nare carried out on four hyperspectral data sets, and experimental results\ndemonstrate the effectiveness of the proposed method.', 'Convolutional Neural Networks (CNNs) are state-of-the-art models for document\nimage classification tasks. However, many of these approaches rely on\nparameters and architectures designed for classifying natural images, which\ndiffer from document images. We question whether this is appropriate and\nconduct a large empirical study to find what aspects of CNNs most affect\nperformance on document images. Among other results, we exceed the\nstate-of-the-art on the RVL-CDIP dataset by using shear transform data\naugmentation and an architecture designed for a larger input image.\nAdditionally, we analyze the learned features and find evidence that CNNs\ntrained on RVL-CDIP learn region-specific layout features.', 'In this paper we present a data-driven approach to obtain the static image of\na scene, eliminating dynamic objects that might have been present at the time\nof traversing the scene with a camera. The general objective is to improve\nvision-based localization and mapping tasks in dynamic environments, where the\npresence (or absence) of different dynamic objects in different moments makes\nthese tasks less robust. We introduce an end-to-end deep learning framework to\nturn images of an urban environment that include dynamic content, such as\nvehicles or pedestrians, into realistic static frames suitable for localization\nand mapping. This objective faces two main challenges: detecting the dynamic\nobjects, and inpainting the static occluded back-ground. The first challenge is\naddressed by the use of a convolutional network that learns a multi-class\nsemantic segmentation of the image. The second challenge is approached with a\ngenerative adversarial model that, taking as input the original dynamic image\nand the computed dynamic/static binary mask, is capable of generating the final\nstatic image. This framework makes use of two new losses, one based on image\nsteganalysis techniques, useful to improve the inpainting quality, and another\none based on ORB features, designed to enhance feature matching between real\nand hallucinated image regions. To validate our approach, we perform an\nextensive evaluation on different tasks that are affected by dynamic entities,\ni.e., visual odometry, place recognition and multi-view stereo, with the\nhallucinated images. Code has been made available on\nhttps://github.com/bertabescos/EmptyCities_SLAM.', ""Image classification is the task of assigning to an input image a label from\na fixed set of categories. One of its most important applicative fields is that\nof robotics, in particular the needing of a robot to be aware of what's around\nand the consequent exploitation of that information as a benefit for its tasks.\nIn this work we consider the problem of a robot that enters a new environment\nand wants to understand visual data coming from its camera, so to extract\nknowledge from them. As main novelty we want to overcome the needing of a\nphysical robot, as it could be expensive and unhandy, so to hopefully enhance,\nspeed up and ease the research in this field. That's why we propose to develop\nan application for a mobile platform that wraps several deep visual recognition\ntasks. First we deal with a simple Image classification, testing a model\nobtained from an AlexNet trained on the ILSVRC 2012 dataset. Several photo\nsettings are considered to better understand which factors affect most the\nquality of classification. For the same purpose we are interested to integrate\nthe classification task with an extra module dealing with segmentation of the\nobject inside the image. In particular we propose a technique for extracting\nthe object shape and moving out all the background, so to focus the\nclassification only on the region occupied by the object. Another significant\ntask that is included is that of object discovery. Its purpose is to simulate\nthe situation in which the robot needs a certain object to complete one of its\nactivities. It starts searching for what it needs by looking around and trying\nto understand the location of the object by scanning the surrounding\nenvironment. Finally we provide a tool for dealing with the creation of\ncustomized task-specific databases, meant to better suit to one's needing in a\nparticular vision task."", 'Automatically generating a natural language description of an image is a task\nclose to the heart of image understanding. In this paper, we present a\nmulti-model neural network method closely related to the human visual system\nthat automatically learns to describe the content of images. Our model consists\nof two sub-models: an object detection and localization model, which extract\nthe information of objects and their spatial relationship in images\nrespectively; Besides, a deep recurrent neural network (RNN) based on long\nshort-term memory (LSTM) units with attention mechanism for sentences\ngeneration. Each word of the description will be automatically aligned to\ndifferent objects of the input image when it is generated. This is similar to\nthe attention mechanism of the human visual system. Experimental results on the\nCOCO dataset showcase the merit of the proposed method, which outperforms\nprevious benchmark models.', 'Grounding human-machine conversation in a document is an effective way to\nimprove the performance of retrieval-based chatbots. However, only a part of\nthe document content may be relevant to help select the appropriate response at\na round. It is thus crucial to select the part of document content relevant to\nthe current conversation context. In this paper, we propose a document content\nselection network (CSN) to perform explicit selection of relevant document\ncontents, and filter out the irrelevant parts. We show in experiments on two\npublic document-grounded conversation datasets that CSN can effectively help\nselect the relevant document contents to the conversation context, and it\nproduces better results than the state-of-the-art approaches. Our code and\ndatasets are available at https://github.com/DaoD/CSN.', 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'Document image enhancement and binarization methods are often used to improve\nthe accuracy and efficiency of document image analysis tasks such as text\nrecognition. Traditional non-machine-learning methods are constructed on\nlow-level features in an unsupervised manner but have difficulty with\nbinarization on documents with severely degraded backgrounds. Convolutional\nneural network-based methods focus only on grayscale images and on local\ntextual features. In this paper, we propose a two-stage color document image\nenhancement and binarization method using generative adversarial neural\nnetworks. In the first stage, four color-independent adversarial networks are\ntrained to extract color foreground information from an input image for\ndocument image enhancement. In the second stage, two independent adversarial\nnetworks with global and local features are trained for image binarization of\ndocuments of variable size. For the adversarial neural networks, we formulate\nloss functions between a discriminator and generators having an encoder-decoder\nstructure. Experimental results show that the proposed method achieves better\nperformance than many classical and state-of-the-art algorithms over the\nDocument Image Binarization Contest (DIBCO) datasets, the LRDE Document\nBinarization Dataset (LRDE DBD), and our shipping label image dataset. We plan\nto release the shipping label dataset as well as our implementation code at\ngithub.com/opensuh/DocumentBinarization/.', 'Graphical elements: particularly tables and figures contain a visual summary\nof the most valuable information contained in a document. Therefore,\nlocalization of such graphical objects in the document images is the initial\nstep to understand the content of such graphical objects or document images. In\nthis paper, we present a novel end-to-end trainable deep learning based\nframework to localize graphical objects in the document images called as\nGraphical Object Detection (GOD). Our framework is data-driven and does not\nrequire any heuristics or meta-data to locate graphical objects in the document\nimages. The GOD explores the concept of transfer learning and domain adaptation\nto handle scarcity of labeled training images for graphical object detection\ntask in the document images. Performance analysis carried out on the various\npublic benchmark data sets: ICDAR-2013, ICDAR-POD2017,and UNLV shows that our\nmodel yields promising results as compared to state-of-the-art techniques.', 'We propose an end-to-end image compression and analysis model with\nTransformers, targeting to the cloud-based image classification application.\nInstead of placing an existing Transformer-based image classification model\ndirectly after an image codec, we aim to redesign the Vision Transformer (ViT)\nmodel to perform image classification from the compressed features and\nfacilitate image compression with the long-term information from the\nTransformer. Specifically, we first replace the patchify stem (i.e., image\nsplitting and embedding) of the ViT model with a lightweight image encoder\nmodelled by a convolutional neural network. The compressed features generated\nby the image encoder are injected convolutional inductive bias and are fed to\nthe Transformer for image classification bypassing image reconstruction.\nMeanwhile, we propose a feature aggregation module to fuse the compressed\nfeatures with the selected intermediate features of the Transformer, and feed\nthe aggregated features to a deconvolutional neural network for image\nreconstruction. The aggregated features can obtain the long-term information\nfrom the self-attention mechanism of the Transformer and improve the\ncompression performance. The rate-distortion-accuracy optimization problem is\nfinally solved by a two-step training strategy. Experimental results\ndemonstrate the effectiveness of the proposed model in both the image\ncompression and the classification tasks.', 'When digitizing a document into an image, it is common to include a\nsurrounding border region to visually indicate that the entire document is\npresent in the image. However, this border should be removed prior to automated\nprocessing. In this work, we present a deep learning based system, PageNet,\nwhich identifies the main page region in an image in order to segment content\nfrom both textual and non-textual border noise. In PageNet, a Fully\nConvolutional Network obtains a pixel-wise segmentation which is post-processed\ninto the output quadrilateral region. We evaluate PageNet on 4 collections of\nhistorical handwritten documents and obtain over 94% mean intersection over\nunion on all datasets and approach human performance on 2 of these collections.\nAdditionally, we show that PageNet can segment documents that are overlayed on\ntop of other documents.', 'Recent research advances in Computer Vision and Natural Language Processing\nhave introduced novel tasks that are paving the way for solving AI-complete\nproblems. One of those tasks is called Visual Question Answering (VQA). A VQA\nsystem must take an image and a free-form, open-ended natural language question\nabout the image, and produce a natural language answer as the output. Such a\ntask has drawn great attention from the scientific community, which generated a\nplethora of approaches that aim to improve the VQA predictive accuracy. Most of\nthem comprise three major components: (i) independent representation learning\nof images and questions; (ii) feature fusion so the model can use information\nfrom both sources to answer visual questions; and (iii) the generation of the\ncorrect answer in natural language. With so many approaches being recently\nintroduced, it became unclear the real contribution of each component for the\nultimate performance of the model. The main goal of this paper is to provide a\ncomprehensive analysis regarding the impact of each component in VQA models.\nOur extensive set of experiments cover both visual and textual elements, as\nwell as the combination of these representations in form of fusion and\nattention mechanisms. Our major contribution is to identify core components for\ntraining VQA models so as to maximize their predictive performance.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'The task of Visual Commonsense Reasoning is extremely challenging in the\nsense that the model has to not only be able to answer a question given an\nimage, but also be able to learn to reason. The baselines introduced in this\ntask are quite limiting because two networks are trained for predicting answers\nand rationales separately. Question and image is used as input to train answer\nprediction network while question, image and correct answer are used as input\nin the rationale prediction network. As rationale is conditioned on the correct\nanswer, it is based on the assumption that we can solve Visual Question\nAnswering task without any error - which is over ambitious. Moreover, such an\napproach makes both answer and rationale prediction two completely independent\nVQA tasks rendering cognition task meaningless. In this paper, we seek to\naddress these issues by proposing an end-to-end trainable model which considers\nboth answers and their reasons jointly. Specifically, we first predict the\nanswer for the question and then use the chosen answer to predict the\nrationale. However, a trivial design of such a model becomes non-differentiable\nwhich makes it difficult to train. We solve this issue by proposing four\napproaches - softmax, gumbel-softmax, reinforcement learning based sampling and\ndirect cross entropy against all pairs of answers and rationales. We\ndemonstrate through experiments that our model performs competitively against\ncurrent state-of-the-art. We conclude with an analysis of presented approaches\nand discuss avenues for further work.', ""A conversational agent (chatbot) is a piece of software that is able to\ncommunicate with humans using natural language. Modeling conversation is an\nimportant task in natural language processing and artificial intelligence.\nWhile chatbots can be used for various tasks, in general they have to\nunderstand users' utterances and provide responses that are relevant to the\nproblem at hand.\n  In my work, I conduct an in-depth survey of recent literature, examining over\n70 publications related to chatbots published in the last 3 years. Then, I\nproceed to make the argument that the very nature of the general conversation\ndomain demands approaches that are different from current state-of-of-the-art\narchitectures. Based on several examples from the literature I show why current\nchatbot models fail to take into account enough priors when generating\nresponses and how this affects the quality of the conversation. In the case of\nchatbots, these priors can be outside sources of information that the\nconversation is conditioned on like the persona or mood of the conversers. In\naddition to presenting the reasons behind this problem, I propose several ideas\non how it could be remedied.\n  The next section focuses on adapting the very recent Transformer model to the\nchatbot domain, which is currently state-of-the-art in neural machine\ntranslation. I first present experiments with the vanilla model, using\nconversations extracted from the Cornell Movie-Dialog Corpus. Secondly, I\naugment the model with some of my ideas regarding the issues of encoder-decoder\narchitectures. More specifically, I feed additional features into the model\nlike mood or persona together with the raw conversation data. Finally, I\nconduct a detailed analysis of how the vanilla model performs on conversational\ndata by comparing it to previous chatbot models and how the additional features\naffect the quality of the generated responses."", 'We propose Pixel-BERT to align image pixels with text by deep multi-modal\ntransformers that jointly learn visual and language embedding in a unified\nend-to-end framework. We aim to build a more accurate and thorough connection\nbetween image pixels and language semantics directly from image and sentence\npairs instead of using region-based image features as the most recent vision\nand language tasks. Our Pixel-BERT which aligns semantic connection in pixel\nand text level solves the limitation of task-specific visual representation for\nvision and language tasks. It also relieves the cost of bounding box\nannotations and overcomes the unbalance between semantic labels in visual task\nand language semantic. To provide a better representation for down-stream\ntasks, we pre-train a universal end-to-end model with image and sentence pairs\nfrom Visual Genome dataset and MS-COCO dataset. We propose to use a random\npixel sampling mechanism to enhance the robustness of visual representation and\nto apply the Masked Language Model and Image-Text Matching as pre-training\ntasks. Extensive experiments on downstream tasks with our pre-trained model\nshow that our approach makes the most state-of-the-arts in downstream tasks,\nincluding Visual Question Answering (VQA), image-text retrieval, Natural\nLanguage for Visual Reasoning for Real (NLVR). Particularly, we boost the\nperformance of a single model in VQA task by 2.17 points compared with SOTA\nunder fair comparison.', 'Current conversational systems can follow simple commands and answer basic\nquestions, but they have difficulty maintaining coherent and open-ended\nconversations about specific topics. Competitions like the Conversational\nIntelligence (ConvAI) challenge are being organized to push the research\ndevelopment towards that goal. This article presents in detail the RLLChatbot\nthat participated in the 2017 ConvAI challenge. The goal of this research is to\nbetter understand how current deep learning and reinforcement learning tools\ncan be used to build a robust yet flexible open domain conversational agent. We\nprovide a thorough description of how a dialog system can be built and trained\nfrom mostly public-domain datasets using an ensemble model. The first\ncontribution of this work is a detailed description and analysis of different\ntext generation models in addition to novel message ranking and selection\nmethods. Moreover, a new open-source conversational dataset is presented.\nTraining on this data significantly improves the Recall@k score of the ranking\nand selection mechanisms compared to our baseline model responsible for\nselecting the message returned at each interaction.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'The demand for on-device document recognition systems increases in\nconjunction with the emergence of more strict privacy and security\nrequirements. In such systems, there is no data transfer from the end device to\na third-party information processing servers. The response time is vital to the\nuser experience of on-device document recognition. Combined with the\nunavailability of discrete GPUs, powerful CPUs, or a large RAM capacity on\nconsumer-grade end devices such as smartphones, the time limitations put\nsignificant constraints on the computational complexity of the applied\nalgorithms for on-device execution.\n  In this work, we consider document location in an image without prior\nknowledge of the document content or its internal structure. In accordance with\nthe published works, at least 5 systems offer solutions for on-device document\nlocation. All these systems use a location method which can be considered\nHough-based. The precision of such systems seems to be lower than that of the\nstate-of-the-art solutions which were not designed to account for the limited\ncomputational resources.\n  We propose an advanced Hough-based method. In contrast with other approaches,\nit accounts for the geometric invariants of the central projection model and\ncombines both edge and color features for document boundary detection. The\nproposed method allowed for the second best result for SmartDoc dataset in\nterms of precision, surpassed by U-net like neural network. When evaluated on a\nmore challenging MIDV-500 dataset, the proposed algorithm guaranteed the best\nprecision compared to published methods. Our method retained the applicability\nto on-device computations.', 'We consider the problem of accurately and efficiently querying a remote\nserver to retrieve information about images captured by a mobile device. In\naddition to reduced transmission overhead and computational complexity, the\nretrieval protocol should be robust to variations in the image acquisition\nprocess, such as translation, rotation, scaling, and sensor-related\ndifferences. We propose to extract scale-invariant image features and then\nperform clustering to reduce the number of features needed for image matching.\nPrincipal Component Analysis (PCA) and Non-negative Matrix Factorization (NMF)\nare investigated as candidate clustering approaches. The image matching\ncomplexity at the database server is quadratic in the (small) number of\nclusters, not in the (very large) number of image features. We employ an\nimage-dependent information content metric to approximate the model order,\ni.e., the number of clusters, needed for accurate matching, which is preferable\nto setting the model order using trial and error. We show how to combine the\nhypotheses provided by PCA and NMF factor loadings, thereby obtaining more\naccurate retrieval than using either approach alone. In experiments on a\ndatabase of urban images, we obtain a top-1 retrieval accuracy of 89% and a\ntop-3 accuracy of 92.5%.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'With the rising use of Machine Learning (ML) and Deep Learning (DL) in\nvarious industries, the medical industry is also not far behind. A very simple\nyet extremely important use case of ML in this industry is for image\nclassification. This is important for doctors to help them detect certain\ndiseases timely, thereby acting as an aid to reduce chances of human judgement\nerror. However, when using automated systems like these, there is a privacy\nconcern as well. Attackers should not be able to get access to the medical\nrecords and images of the patients. It is also required that the model be\nsecure, and that the data that is sent to the model and the predictions that\nare received both should not be revealed to the model in clear text.\n  In this study, we aim to solve these problems in the context of a medical\nimage classification problem of detection of pneumonia by examining chest x-ray\nimages.', 'Vision based localization is the problem of inferring the pose of the camera\ngiven a single image. One solution to this problem is to learn a deep neural\nnetwork to infer the pose of a query image after learning on a dataset of\nimages with known poses. Another more commonly used approach rely on image\nretrieval where the query image is compared against the database of images and\nits pose is inferred with the help of the retrieved images. The latter approach\nassumes that images taken from the same places consists of the same landmarks\nand, thus would have similar feature representations. These representation can\nbe learned using full supervision to be robust to different variations in\ncapture conditions like time of the day and weather. In this work, we introduce\na framework to enhance the performance of these retrieval based localization\nmethods by taking into account the additional information including GPS\ncoordinates and temporal neighbourhood of the images provided by the\nacquisition process in addition to the descriptor similarity of pairs of images\nin the reference or query database which is used traditionally for\nlocalization. Our method constructs a graph based on this additional\ninformation and use it for robust retrieval by smoothing the feature\nrepresentation of reference and/or query images. We show that the proposed\nmethod is able to significantly improve the localization accuracy on two large\nscale datasets over the baselines.', 'Errors in medication intake among elderly people are very common. One of the\nmain causes for this is their loss of ability to retain information. The high\namount of medicine intake required by the advanced age is another limiting\nfactor. Thence, the design of an interactive aid system, preferably using\nnatural language, to help the older population with medication is in demand. A\nchatbot based on a subtractive cluster algorithm, included in unsupervised\nlearned models, is the chosen solution since the processing of natural\nlanguages is a necessary step in view to construct a chatbot able to answer\nquestions that older people may pose upon themselves concerning a particular\ndrug. In this work, the subtractive mountain clustering algorithm has been\nadapted to the problem of natural languages processing. This algorithm version\nallows for the association of a set of words into clusters. After finding the\ncentre of every cluster -- the most relevant word, all the others are\naggregated according to a defined metric adapted to the language processing\nrealm. All the relevant stored information is processed, as well as the\nquestions, by the algorithm. The correct processing of the text enables the\nchatbot to produce answers that relate to the posed queries. To validate the\nmethod, we use the package insert of a drug as the available information and\nformulate associated questions.', 'Conventional methods for the image-text generation tasks mainly tackle the\nnaturally bidirectional generation tasks separately, focusing on designing\ntask-specific frameworks to improve the quality and fidelity of the generated\nsamples. Recently, Vision-Language Pre-training models have greatly improved\nthe performance of the image-to-text generation tasks, but large-scale\npre-training models for text-to-image synthesis task are still under-developed.\nIn this paper, we propose ERNIE-ViLG, a unified generative pre-training\nframework for bidirectional image-text generation with transformer model. Based\non the image quantization models, we formulate both image generation and text\ngeneration as autoregressive generative tasks conditioned on the text/image\ninput. The bidirectional image-text generative modeling eases the semantic\nalignments across vision and language. For the text-to-image generation\nprocess, we further propose an end-to-end training method to jointly learn the\nvisual sequence generator and the image reconstructor. To explore the landscape\nof large-scale pre-training for bidirectional text-image generation, we train a\n10-billion parameter ERNIE-ViLG model on a large-scale dataset of 145 million\n(Chinese) image-text pairs which achieves state-of-the-art performance for both\ntext-to-image and image-to-text tasks, obtaining an FID of 7.9 on MS-COCO for\ntext-to-image synthesis and best results on COCO-CN and AIC-ICC for image\ncaptioning.', 'The ability to engage in goal-oriented conversations has allowed humans to\ngain knowledge, reduce uncertainty, and perform tasks more efficiently.\nArtificial agents, however, are still far behind humans in having goal-driven\nconversations. In this work, we focus on the task of goal-oriented visual\ndialogue, aiming to automatically generate a series of questions about an image\nwith a single objective. This task is challenging since these questions must\nnot only be consistent with a strategy to achieve a goal, but also consider the\ncontextual information in the image. We propose an end-to-end goal-oriented\nvisual dialogue system, that combines reinforcement learning with regularized\ninformation gain. Unlike previous approaches that have been proposed for the\ntask, our work is motivated by the Rational Speech Act framework, which models\nthe process of human inquiry to reach a goal. We test the two versions of our\nmodel on the GuessWhat?! dataset, obtaining significant results that outperform\nthe current state-of-the-art models in the task of generating questions to find\nan undisclosed object in an image.', ""Most popular goal-oriented dialogue agents are capable of understanding the\nconversational context. However, with the surge of virtual assistants with\nscreen, the next generation of agents are required to also understand screen\ncontext in order to provide a proper interactive experience, and better\nunderstand users' goals. In this paper, we propose a novel multimodal\nconversational framework, where the dialogue agent's next action and their\narguments are derived jointly conditioned both on the conversational and the\nvisual context. Specifically, we propose a new model, that can reason over the\nvisual context within a conversation and populate API arguments with visual\nentities given the user query. Our model can recognize visual features such as\ncolor and shape as well as the metadata based features such as price or star\nrating associated with a visual entity. In order to train our model, due to a\nlack of suitable multimodal conversational datasets, we also propose a novel\nmultimodal dialog simulator to generate synthetic data and also collect\nrealistic user data from MTurk to improve model robustness. The proposed model\nachieves a reasonable 85% model accuracy, without high inference latency. We\nalso demonstrate the proposed approach in a prototypical furniture shopping\nexperience for a multimodal virtual assistant."", 'Cloud computing offers the potential to help scientists to process massive\nnumber of computing resources often required in machine learning application\nsuch as computer vision problems. This proposal would like to show that which\nbenefits can be obtained from cloud in order to help medical image analysis\nusers (including scientists, clinicians, and research institutes). As security\nand privacy of algorithms are important for most of algorithms inventors, these\nalgorithms can be hidden in a cloud to allow the users to use the algorithms as\na package without any access to see/change their inside. In another word, in\nthe user part, users send their images to the cloud and configure the algorithm\nvia an interface. In the cloud part, the algorithms are applied to this image\nand the results are returned back to the user. My proposal has two parts: (1)\ninvestigate the potential of cloud computing for computer vision problems and\n(2) study the components of a proposed cloud-based framework for medical image\nanalysis application and develop them (depending on the length of the\ninternship). The investigation part will involve a study on several aspects of\nthe problem including security, usability (for medical end users of the\nservice), appropriate programming abstractions for vision problems, scalability\nand resource requirements. In the second part of this proposal I am going to\nthoroughly study of the proposed framework components and their relations and\ndevelop them. The proposed cloud-based framework includes an integrated\nenvironment to enable scientists and clinicians to access to the previous and\ncurrent medical image analysis algorithms using a handful user interface\nwithout any access to the algorithm codes and procedures.', 'Recently, the progress of learning-by-synthesis has proposed a training model\nfor synthetic images, which can effectively reduce the cost of human and\nmaterial resources. However, due to the different distribution of synthetic\nimages compared to real images, the desired performance cannot still be\nachieved. Real images consist of multiple forms of light orientation, while\nsynthetic images consist of a uniform light orientation. These features are\nconsidered to be characteristic of outdoor and indoor scenes, respectively. To\nsolve this problem, the previous method learned a model to improve the realism\nof the synthetic image. Different from the previous methods, this paper takes\nthe first step to purify real images. Through the style transfer task, the\ndistribution of outdoor real images is converted into indoor synthetic images,\nthereby reducing the influence of light. Therefore, this paper proposes a\nreal-time style transfer network that preserves image content information (eg,\ngaze direction, pupil center position) of an input image (real image) while\ninferring style information (eg, image color structure, semantic features) of\nstyle image (synthetic image). In addition, the network accelerates the\nconvergence speed of the model and adapts to multi-scale images. Experiments\nwere performed using mixed studies (qualitative and quantitative) methods to\ndemonstrate the possibility of purifying real images in complex directions.\nQualitatively, it compares the proposed method with the available methods in a\nseries of indoor and outdoor scenarios of the LPW dataset. In quantitative\nterms, it evaluates the purified image by training a gaze estimation model on\nthe cross data set. The results show a significant improvement over the\nbaseline method compared to the raw real image.', 'Document Grounded Conversations is a task to generate dialogue responses when\nchatting about the content of a given document. Obviously, document knowledge\nplays a critical role in Document Grounded Conversations, while existing\ndialogue models do not exploit this kind of knowledge effectively enough. In\nthis paper, we propose a novel Transformer-based architecture for multi-turn\ndocument grounded conversations. In particular, we devise an Incremental\nTransformer to encode multi-turn utterances along with knowledge in related\ndocuments. Motivated by the human cognitive process, we design a two-pass\ndecoder (Deliberation Decoder) to improve context coherence and knowledge\ncorrectness. Our empirical study on a real-world Document Grounded Dataset\nproves that responses generated by our model significantly outperform\ncompetitive baselines on both context coherence and knowledge relevance.', 'How to leverage cross-document interactions to improve ranking performance is\nan important topic in information retrieval (IR) research. However, this topic\nhas not been well-studied in the learning-to-rank setting and most of the\nexisting work still treats each document independently while scoring. The\nrecent development of deep learning shows strength in modeling complex\nrelationships across sequences and sets. It thus motivates us to study how to\nleverage cross-document interactions for learning-to-rank in the deep learning\nframework. In this paper, we formally define the permutation-equivariance\nrequirement for a scoring function that captures cross-document interactions.\nWe then propose a self-attention based document interaction network and show\nthat it satisfies the permutation-equivariant requirement, and can generate\nscores for document sets of varying sizes. Our proposed methods can\nautomatically learn to capture document interactions without any auxiliary\ninformation, and can scale across large document sets. We conduct experiments\non three ranking datasets: the benchmark Web30k, a Gmail search, and a Google\nDrive Quick Access dataset. Experimental results show that our proposed methods\nare both more effective and efficient than baselines.', 'Document Image Analysis, like any Digital Image Analysis requires\nidentification and extraction of proper features, which are generally extracted\nfrom uncompressed images, though in reality images are made available in\ncompressed form for the reasons such as transmission and storage efficiency.\nHowever, this implies that the compressed image should be decompressed, which\nindents additional computing resources. This limitation induces the motivation\nto research in extracting features directly from the compressed image. In this\nresearch, we propose to extract essential features such as projection profile,\nrun-histogram and entropy for text document analysis directly from run-length\ncompressed text-documents. The experimentation illustrates that features are\nextracted directly from the compressed image without going through the stage of\ndecompression, because of which the computing time is reduced. The feature\nvalues so extracted are exactly identical to those extracted from uncompressed\nimages.', 'Although CNNs are widely considered as the state-of-the-art models in various\napplications of image analysis, one of the main challenges still open is the\ntraining of a CNN on high resolution images. Different strategies have been\nproposed involving either a rescaling of the image or an individual processing\nof parts of the image. Such strategies cannot be applied to images, such as\ngigapixel histopathological images, for which a high reduction in resolution\ninherently effects a loss of discriminative information, and in respect of\nwhich the analysis of single parts of the image suffers from a lack of global\ninformation or implies a high workload in terms of annotating the training\nimages in such a way as to select significant parts. We propose a method for\nthe analysis of gigapixel histopathological images solely by using weak\nimage-level labels. In particular, two analysis tasks are taken into account: a\nbinary classification and a prediction of the tumor proliferation score. Our\nmethod is based on a CNN structure consisting of a compressing path and a\nlearning path. In the compressing path, the gigapixel image is packed into a\ngrid-based feature map by using a residual network devoted to the feature\nextraction of each patch into which the image has been divided. In the learning\npath, attention modules are applied to the grid-based feature map, taking into\naccount spatial correlations of neighboring patch features to find regions of\ninterest, which are then used for the final whole slide analysis. Our method\nintegrates both global and local information, is flexible with regard to the\nsize of the input images and only requires weak image-level labels. Comparisons\nwith different methods of the state-of-the-art on two well known datasets,\nCamelyon16 and TUPAC16, have been made to confirm the validity of the proposed\nmodel.', 'Weakly supervised learning of object detection is an important problem in\nimage understanding that still does not have a satisfactory solution. In this\npaper, we address this problem by exploiting the power of deep convolutional\nneural networks pre-trained on large-scale image-level classification tasks. We\npropose a weakly supervised deep detection architecture that modifies one such\nnetwork to operate at the level of image regions, performing simultaneously\nregion selection and classification. Trained as an image classifier, the\narchitecture implicitly learns object detectors that are better than\nalternative weakly supervised detection systems on the PASCAL VOC data. The\nmodel, which is a simple and elegant end-to-end architecture, outperforms\nstandard data augmentation and fine-tuning techniques for the task of\nimage-level classification as well.', 'Deep learning methods have proven extremely effective at performing a variety\nof medical image analysis tasks. With their potential use in clinical routine,\ntheir lack of transparency has however been one of their few weak points,\nraising concerns regarding their behavior and failure modes. While most\nresearch to infer model behavior has focused on indirect strategies that\nestimate prediction uncertainties and visualize model support in the input\nimage space, the ability to explicitly query a prediction model regarding its\nimage content offers a more direct way to determine the behavior of trained\nmodels. To this end, we present a novel Visual Question Answering approach that\nallows an image to be queried by means of a written question. Experiments on a\nvariety of medical and natural image datasets show that by fusing image and\nquestion features in a novel way, the proposed approach achieves an equal or\nhigher accuracy compared to current methods.', 'Existing methods for interactive image retrieval have demonstrated the merit\nof integrating user feedback, improving retrieval results. However, most\ncurrent systems rely on restricted forms of user feedback, such as binary\nrelevance responses, or feedback based on a fixed set of relative attributes,\nwhich limits their impact. In this paper, we introduce a new approach to\ninteractive image search that enables users to provide feedback via natural\nlanguage, allowing for more natural and effective interaction. We formulate the\ntask of dialog-based interactive image retrieval as a reinforcement learning\nproblem, and reward the dialog system for improving the rank of the target\nimage during each dialog turn. To mitigate the cumbersome and costly process of\ncollecting human-machine conversations as the dialog system learns, we train\nour system with a user simulator, which is itself trained to describe the\ndifferences between target and candidate images. The efficacy of our approach\nis demonstrated in a footwear retrieval application. Experiments on both\nsimulated and real-world data show that 1) our proposed learning framework\nachieves better accuracy than other supervised and reinforcement learning\nbaselines and 2) user feedback based on natural language rather than\npre-specified attributes leads to more effective retrieval results, and a more\nnatural and expressive communication interface.', 'Recognition of document images have important applications in restoring old\nand classical texts. The problem involves quality improvement before passing it\nto a properly trained OCR to get accurate recognition of the text. The image\nenhancement and quality improvement constitute important steps as subsequent\nrecognition depends upon the quality of the input image. There are scenarios\nwhen high resolution images are not available and our experiments show that the\nOCR accuracy reduces significantly with decrease in the spatial resolution of\ndocument images. Thus the only option is to improve the resolution of such\ndocument images. The goal is to construct a high resolution image, given a\nsingle low resolution binary image, which constitutes the problem of single\nimage super-resolution. Most of the previous work in super-resolution deal with\nnatural images which have more information-content than the document images.\nHere, we use Convolution Neural Network to learn the mapping between low and\nthe corresponding high resolution images. We experiment with different number\nof layers, parameter settings and non-linear functions to build a fast\nend-to-end framework for document image super-resolution. Our proposed model\nshows a very good PSNR improvement of about 4 dB on 75 dpi Tamil images,\nresulting in a 3 % improvement of word level accuracy by the OCR. It takes less\ntime than the recent sparse based natural image super-resolution technique,\nmaking it useful for real-time document recognition applications.', 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', 'Document AI, or Document Intelligence, is a relatively new research topic\nthat refers to the techniques for automatically reading, understanding, and\nanalyzing business documents. It is an important research direction for natural\nlanguage processing and computer vision. In recent years, the popularity of\ndeep learning technology has greatly advanced the development of Document AI,\nsuch as document layout analysis, visual information extraction, document\nvisual question answering, document image classification, etc. This paper\nbriefly reviews some of the representative models, tasks, and benchmark\ndatasets. Furthermore, we also introduce early-stage heuristic rule-based\ndocument analysis, statistical machine learning algorithms, and deep learning\napproaches especially pre-training methods. Finally, we look into future\ndirections for Document AI research.', ""Information gathering in a partially observable environment can be formulated\nas a reinforcement learning (RL), problem where the reward depends on the\nagent's uncertainty. For example, the reward can be the negative entropy of the\nagent's belief over an unknown (or hidden) variable. Typically, the rewards of\nan RL agent are defined as a function of the state-action pairs and not as a\nfunction of the belief of the agent; this hinders the direct application of\ndeep RL methods for such tasks. This paper tackles the challenge of using\nbelief-based rewards for a deep RL agent, by offering a simple insight that\nmaximizing any convex function of the belief of the agent can be approximated\nby instead maximizing a prediction reward: a reward based on prediction\naccuracy. In particular, we derive the exact error between negative entropy and\nthe expected prediction reward. This insight provides theoretical motivation\nfor several fields using prediction rewards---namely visual attention, question\nanswering systems, and intrinsic motivation---and highlights their connection\nto the usually distinct fields of active perception, active sensing, and sensor\nplacement. Based on this insight we present deep anticipatory networks (DANs),\nwhich enables an agent to take actions to reduce its uncertainty without\nperforming explicit belief inference. We present two applications of DANs:\nbuilding a sensor selection system for tracking people in a shopping mall and\nlearning discrete models of attention on fashion MNIST and MNIST digit\nclassification."", 'We present FlipDial, a generative model for visual dialogue that\nsimultaneously plays the role of both participants in a visually-grounded\ndialogue. Given context in the form of an image and an associated caption\nsummarising the contents of the image, FlipDial learns both to answer questions\nand put forward questions, capable of generating entire sequences of dialogue\n(question-answer pairs) which are diverse and relevant to the image. To do\nthis, FlipDial relies on a simple but surprisingly powerful idea: it uses\nconvolutional neural networks (CNNs) to encode entire dialogues directly,\nimplicitly capturing dialogue context, and conditional VAEs to learn the\ngenerative model. FlipDial outperforms the state-of-the-art model in the\nsequential answering task (one-way visual dialogue) on the VisDial dataset by 5\npoints in Mean Rank using the generated answers. We are the first to extend\nthis paradigm to full two-way visual dialogue, where our model is capable of\ngenerating both questions and answers in sequence based on a visual input, for\nwhich we propose a set of novel evaluation measures and metrics.', 'Recent work in open-domain conversational agents has demonstrated that\nsignificant improvements in model engagingness and humanness metrics can be\nachieved via massive scaling in both pre-training data and model size\n(Adiwardana et al., 2020; Roller et al., 2020). However, if we want to build\nagents with human-like abilities, we must expand beyond handling just text. A\nparticularly important topic is the ability to see images and communicate about\nwhat is perceived. With the goal of engaging humans in multi-modal dialogue, we\ninvestigate combining components from state-of-the-art open-domain dialogue\nagents with those from state-of-the-art vision models. We study incorporating\ndifferent image fusion schemes and domain-adaptive pre-training and fine-tuning\nstrategies, and show that our best resulting model outperforms strong existing\nmodels in multi-modal dialogue while simultaneously performing as well as its\npredecessor (text-only) BlenderBot (Roller et al., 2020) in text-based\nconversation. We additionally investigate and incorporate safety components in\nour final model, and show that such efforts do not diminish model performance\nwith respect to engagingness metrics.', 'Recently we see a rising number of methods in the field of eXplainable\nArtificial Intelligence. To our surprise, their development is driven by model\ndevelopers rather than a study of needs for human end users. The analysis of\nneeds, if done, takes the form of an A/B test rather than a study of open\nquestions. To answer the question ""What would a human operator like to ask the\nML model?"" we propose a conversational system explaining decisions of the\npredictive model. In this experiment, we developed a chatbot called dr_ant to\ntalk about machine learning model trained to predict survival odds on Titanic.\nPeople can talk with dr_ant about different aspects of the model to understand\nthe rationale behind its predictions. Having collected a corpus of 1000+\ndialogues, we analyse the most common types of questions that users would like\nto ask. To our knowledge, it is the first study which uses a conversational\nsystem to collect the needs of human operators from the interactive and\niterative dialogue explorations of a predictive model.', ""A big part of achieving Artificial General Intelligence(AGI) is to build a\nmachine that can see and listen like humans. Much work has focused on designing\nmodels for image classification, video classification, object detection, pose\nestimation, speech recognition, etc., and has achieved significant progress in\nrecent years thanks to deep learning. However, understanding the world is not\nenough. An AI agent also needs to know how to talk, especially how to\ncommunicate with a human. While perception (vision, for example) is more common\nacross animal species, the use of complicated language is unique to humans and\nis one of the most important aspects of intelligence.\n  In this thesis, we focus on generating textual output given visual input. In\nChapter 3, we focus on generating the referring expression, a text description\nfor an object in the image so that a receiver can infer which object is being\ndescribed. We use a comprehension machine to directly guide the generated\nreferring expressions to be more discriminative. In Chapter 4, we introduce a\nmethod that encourages discriminability in image caption generation. We show\nthat more discriminative captioning models generate more descriptive captions.\nIn Chapter 5, we study how training objectives and sampling methods affect the\nmodels' ability to generate diverse captions. We find that a popular captioning\ntraining strategy will be detrimental to the diversity of generated captions.\nIn Chapter 6, we propose a model that can control the length of generated\ncaptions. By changing the desired length, one can influence the style and\ndescriptiveness of the captions. Finally, in Chapter 7, we rank/generate\ninformative image tags according to their information utility. The proposed\nmethod better matches what humans think are the most important tags for the\nimages."", ""One obstacle to applying reinforcement learning algorithms to real-world\nproblems is the lack of suitable reward functions. Designing such reward\nfunctions is difficult in part because the user only has an implicit\nunderstanding of the task objective. This gives rise to the agent alignment\nproblem: how do we create agents that behave in accordance with the user's\nintentions? We outline a high-level research direction to solve the agent\nalignment problem centered around reward modeling: learning a reward function\nfrom interaction with the user and optimizing the learned reward function with\nreinforcement learning. We discuss the key challenges we expect to face when\nscaling reward modeling to complex and general domains, concrete approaches to\nmitigate these challenges, and ways to establish trust in the resulting agents."", 'Conventional seq2seq chatbot models attempt only to find sentences with the\nhighest probabilities conditioned on the input sequences, without considering\nthe sentiment of the output sentences. In this paper, we investigate four\nmodels to scale or adjust the sentiment of the chatbot response: a\npersona-based model, reinforcement learning, a plug and play model, and\nCycleGAN, all based on the seq2seq model. We also develop machine-evaluated\nmetrics to estimate whether the responses are reasonable given the input. These\nmetrics, together with human evaluation, are used to analyze the performance of\nthe four models in terms of different aspects; reinforcement learning and\nCycleGAN are shown to be very attractive.', 'Document-level machine translation focuses on the translation of entire\ndocuments from a source to a target language. It is widely regarded as a\nchallenging task since the translation of the individual sentences in the\ndocument needs to retain aspects of the discourse at document level. However,\ndocument-level translation models are usually not trained to explicitly ensure\ndiscourse quality. Therefore, in this paper we propose a training approach that\nexplicitly optimizes two established discourse metrics, lexical cohesion (LC)\nand coherence (COH), by using a reinforcement learning objective. Experiments\nover four different language pairs and three translation domains have shown\nthat our training approach has been able to achieve more cohesive and coherent\ndocument translations than other competitive approaches, yet without\ncompromising the faithfulness to the reference translation. In the case of the\nZh-En language pair, our method has achieved an improvement of 2.46 percentage\npoints (pp) in LC and 1.17 pp in COH over the runner-up, while at the same time\nimproving 0.63 pp in BLEU score and 0.47 pp in F_BERT.', ""Using chatbots to deliver recommendations is increasingly popular. The design\nof recommendation chatbots has primarily been taking an information-centric\napproach by focusing on the recommended content per se. Limited attention is on\nhow social connection and relational strategies, such as self-disclosure from a\nchatbot, may influence users' perception and acceptance of the recommendation.\nIn this work, we designed, implemented, and evaluated a social chatbot capable\nof performing three different levels of self-disclosure: factual information\n(low), cognitive opinions (medium), and emotions (high). In the evaluation, we\nrecruited 372 participants to converse with the chatbot on two topics: movies\nand COVID-19 experiences. In each topic, the chatbot performed small talks and\nmade recommendations relevant to the topic. Participants were randomly assigned\nto four experimental conditions where the chatbot used factual, cognitive,\nemotional, and adaptive strategies to perform self-disclosures. By training a\ntext classifier to identify users' level of self-disclosure in real-time, the\nadaptive chatbot can dynamically match its self-disclosure to the level of\ndisclosure exhibited by the users. Our results show that users reciprocate with\nhigher-level self-disclosure when a recommendation chatbot consistently\ndisplays emotions throughout the conversation. Chatbot's emotional disclosure\nalso led to increased interactional enjoyment and more positive interpersonal\nperception towards the bot, fostering a stronger human-chatbot relationship and\nthus leading to increased recommendation effectiveness, including a higher\ntendency to accept the recommendation. We discuss the understandings obtained\nand implications to future design."", 'Under special circumstances, summaries should conform to a particular style\nwith patterns, such as court judgments and abstracts in academic papers. To\nthis end, the prototype document-summary pairs can be utilized to generate\nbetter summaries. There are two main challenges in this task: (1) the model\nneeds to incorporate learned patterns from the prototype, but (2) should avoid\ncopying contents other than the patternized words---such as irrelevant\nfacts---into the generated summaries. To tackle these challenges, we design a\nmodel named Prototype Editing based Summary Generator (PESG). PESG first learns\nsummary patterns and prototype facts by analyzing the correlation between a\nprototype document and its summary. Prototype facts are then utilized to help\nextract facts from the input document. Next, an editing generator generates new\nsummary based on the summary pattern or extracted facts. Finally, to address\nthe second challenge, a fact checker is used to estimate mutual information\nbetween the input document and generated summary, providing an additional\nsignal for the generator. Extensive experiments conducted on a large-scale\nreal-world text summarization dataset show that PESG achieves the\nstate-of-the-art performance in terms of both automatic metrics and human\nevaluations.', 'Since the low quality of document images will greatly undermine the chances\nof success in automatic text recognition and analysis, it is necessary to\nassess the quality of document images uploaded in online business process, so\nas to reject those images of low quality. In this paper, we attempt to achieve\ndocument image quality assessment and our contributions are twofold. Firstly,\nsince document image quality assessment is more interested in text, we propose\na text line based framework to estimate document image quality, which is\ncomposed of three stages: text line detection, text line quality prediction,\nand overall quality assessment. Text line detection aims to find potential text\nlines with a detector. In the text line quality prediction stage, the quality\nscore is computed for each text line with a CNN-based prediction model. The\noverall quality of document images is finally assessed with the ensemble of all\ntext line quality. Secondly, to train the prediction model, a large-scale\ndataset, comprising 52,094 text line images, is synthesized with diverse\nattributes. For each text line image, a quality label is computed with a\npiece-wise function. To demonstrate the effectiveness of the proposed\nframework, comprehensive experiments are evaluated on two popular document\nimage quality assessment benchmarks. Our framework significantly outperforms\nthe state-of-the-art methods by large margins on the large and complicated\ndataset.', 'The number of social images has exploded by the wide adoption of social\nnetworks, and people like to share their comments about them. These comments\ncan be a description of the image, or some objects, attributes, scenes in it,\nwhich are normally used as the user-provided tags. However, it is well-known\nthat user-provided tags are incomplete and imprecise to some extent. Directly\nusing them can damage the performance of related applications, such as the\nimage annotation and retrieval. In this paper, we propose to learn an image\nannotation model and refine the user-provided tags simultaneously in a\nweakly-supervised manner. The deep neural network is utilized as the image\nfeature learning and backbone annotation model, while visual consistency,\nsemantic dependency, and user-error sparsity are introduced as the constraints\nat the batch level to alleviate the tag noise. Therefore, our model is highly\nflexible and stable to handle large-scale image sets. Experimental results on\ntwo benchmark datasets indicate that our proposed model achieves the best\nperformance compared to the state-of-the-art methods.', ""We propose an architecture for fine-grained visual categorization that\napproaches expert human performance in the classification of bird species. Our\narchitecture first computes an estimate of the object's pose; this is used to\ncompute local image features which are, in turn, used for classification. The\nfeatures are computed by applying deep convolutional nets to image patches that\nare located and normalized by the pose. We perform an empirical study of a\nnumber of pose normalization schemes, including an investigation of higher\norder geometric warping functions. We propose a novel graph-based clustering\nalgorithm for learning a compact pose normalization space. We perform a\ndetailed investigation of state-of-the-art deep convolutional feature\nimplementations and fine-tuning feature learning for fine-grained\nclassification. We observe that a model that integrates lower-level feature\nlayers with pose-normalized extraction routines and higher-level feature layers\nwith unaligned image features works best. Our experiments advance\nstate-of-the-art performance on bird species recognition, with a large\nimprovement of correct classification rates over previous methods (75% vs.\n55-65%)."", 'Previous researchers have considered sentiment analysis as a document\nclassification task, in which input documents are classified into predefined\nsentiment classes. Although there are sentences in a document that support\nimportant evidences for sentiment analysis and sentences that do not, they have\ntreated the document as a bag of sentences. In other words, they have not\nconsidered the importance of each sentence in the document. To effectively\ndetermine polarity of a document, each sentence in the document should be dealt\nwith different degrees of importance. To address this problem, we propose a\ndocument-level sentence classification model based on deep neural networks, in\nwhich the importance degrees of sentences in documents are automatically\ndetermined through gate mechanisms. To verify our new sentiment analysis model,\nwe conducted experiments using the sentiment datasets in the four different\ndomains such as movie reviews, hotel reviews, restaurant reviews, and music\nreviews. In the experiments, the proposed model outperformed previous\nstate-of-the-art models that do not consider importance differences of\nsentences in a document. The experimental results show that the importance of\nsentences should be considered in a document-level sentiment classification\ntask.', 'One common trend in image tagging research is to focus on visually relevant\ntags, and this tends to ignore the personal and social aspect of tags,\nespecially on photoblogging websites such as Flickr. Previous work has\ncorrectly identified that many of the tags that users provide on images are not\nvisually relevant (i.e. representative of the salient content in the image) and\nthey go on to treat such tags as noise, ignoring that the users chose to\nprovide those tags over others that could have been more visually relevant.\nAnother common assumption about user generated tags for images is that the\norder of these tags provides no useful information for the prediction of tags\non future images. This assumption also tends to define usefulness in terms of\nwhat is visually relevant to the image. For general tagging or labeling\napplications that focus on providing visual information about image content,\nthese assumptions are reasonable, but when considering personalized image\ntagging applications, these assumptions are at best too rigid, ignoring user\nchoice and preferences.\n  We challenge the aforementioned assumptions, and provide a machine learning\napproach to the problem of personalized image tagging with the following\ncontributions: 1.) We reformulate the personalized image tagging problem as a\nsearch/retrieval ranking problem, 2.) We leverage the order of tags, which does\nnot always reflect visual relevance, provided by the user in the past as a cue\nto their tag preferences, similar to click data, 3.) We propose a technique to\naugment sparse user tag data (semi-supervision), and 4.) We demonstrate the\nefficacy of our method on a subset of Flickr images, showing improvement over\nprevious state-of-art methods.', 'Generative models have made significant progress in the tasks of modeling\ncomplex data distributions such as natural images. The introduction of\nGenerative Adversarial Networks (GANs) and auto-encoders lead to the\npossibility of training on big data sets in an unsupervised manner. However,\nfor many generative models it is not possible to specify what kind of image\nshould be generated and it is not possible to translate existing images into\nnew images of similar domains. Furthermore, models that can perform\nimage-to-image translation often need distinct models for each domain, making\nit hard to scale these systems to multiple domain image-to-image translation.\nWe introduce a model that can do both, controllable image generation and\nimage-to-image translation between multiple domains. We split our image\nrepresentation into two parts encoding unstructured and structured information\nrespectively. The latter is designed in a disentangled manner, so that\ndifferent parts encode different image characteristics. We train an encoder to\nencode images into these representations and use a small amount of labeled data\nto specify what kind of information should be encoded in the disentangled part.\nA generator is trained to generate images from these representations using the\ncharacteristics provided by the disentangled part of the representation.\nThrough this we can control what kind of images the generator generates,\ntranslate images between different domains, and even learn unknown\ndata-generating factors while only using one single model.', 'Standard neural machine translation (NMT) is on the assumption of\ndocument-level context independent. Most existing document-level NMT methods\nare satisfied with a smattering sense of brief document-level information,\nwhile this work focuses on exploiting detailed document-level context in terms\nof multiple forms of document embeddings, which is capable of sufficiently\nmodeling deeper and richer document-level context. The proposed document-aware\nNMT is implemented to enhance the Transformer baseline by introducing both\nglobal and local document-level clues on the source end. Experiments show that\nthe proposed method significantly improves the translation performance over\nstrong baselines and other related studies.', 'We propose a novel mode of feedback for image search, where a user describes\nwhich properties of exemplar images should be adjusted in order to more closely\nmatch his/her mental model of the image sought. For example, perusing image\nresults for a query ""black shoes"", the user might state, ""Show me shoe images\nlike these, but sportier."" Offline, our approach first learns a set of ranking\nfunctions, each of which predicts the relative strength of a nameable attribute\nin an image (e.g., sportiness). At query time, the system presents the user\nwith a set of exemplar images, and the user relates them to his/her target\nimage with comparative statements. Using a series of such constraints in the\nmulti-dimensional attribute space, our method iteratively updates its relevance\nfunction and re-ranks the database of images. To determine which exemplar\nimages receive feedback from the user, we present two variants of the approach:\none where the feedback is user-initiated and another where the feedback is\nactively system-initiated. In either case, our approach allows a user to\nefficiently ""whittle away"" irrelevant portions of the visual feature space,\nusing semantic language to precisely communicate her preferences to the system.\nWe demonstrate our technique for refining image search for people, products,\nand scenes, and we show that it outperforms traditional binary relevance\nfeedback in terms of search speed and accuracy. In addition, the ordinal nature\nof relative attributes helps make our active approach efficient -- both\ncomputationally for the machine when selecting the reference images, and for\nthe user by requiring less user interaction than conventional passive and\nactive methods.', ""In shared autonomy, user input is combined with semi-autonomous control to\nachieve a common goal. The goal is often unknown ex-ante, so prior work enables\nagents to infer the goal from user input and assist with the task. Such methods\ntend to assume some combination of knowledge of the dynamics of the\nenvironment, the user's policy given their goal, and the set of possible goals\nthe user might target, which limits their application to real-world scenarios.\nWe propose a deep reinforcement learning framework for model-free shared\nautonomy that lifts these assumptions. We use human-in-the-loop reinforcement\nlearning with neural network function approximation to learn an end-to-end\nmapping from environmental observation and user input to agent action values,\nwith task reward as the only form of supervision. This approach poses the\nchallenge of following user commands closely enough to provide the user with\nreal-time action feedback and thereby ensure high-quality user input, but also\ndeviating from the user's actions when they are suboptimal. We balance these\ntwo needs by discarding actions whose values fall below some threshold, then\nselecting the remaining action closest to the user's input. Controlled studies\nwith users (n = 12) and synthetic pilots playing a video game, and a pilot\nstudy with users (n = 4) flying a real quadrotor, demonstrate the ability of\nour algorithm to assist users with real-time control tasks in which the agent\ncannot directly access the user's private information through observations, but\nreceives a reward signal and user input that both depend on the user's intent.\nThe agent learns to assist the user without access to this private information,\nimplicitly inferring it from the user's input. This paper is a proof of concept\nthat illustrates the potential for deep reinforcement learning to enable\nflexible and practical assistive systems."", ""Prior work on training generative Visual Dialog models with reinforcement\nlearning(Das et al.) has explored a Qbot-Abot image-guessing game and shown\nthat this 'self-talk' approach can lead to improved performance at the\ndownstream dialog-conditioned image-guessing task. However, this improvement\nsaturates and starts degrading after a few rounds of interaction, and does not\nlead to a better Visual Dialog model. We find that this is due in part to\nrepeated interactions between Qbot and Abot during self-talk, which are not\ninformative with respect to the image. To improve this, we devise a simple\nauxiliary objective that incentivizes Qbot to ask diverse questions, thus\nreducing repetitions and in turn enabling Abot to explore a larger state space\nduring RL ie. be exposed to more visual concepts to talk about, and varied\nquestions to answer. We evaluate our approach via a host of automatic metrics\nand human studies, and demonstrate that it leads to better dialog, ie. dialog\nthat is more diverse (ie. less repetitive), consistent (ie. has fewer\nconflicting exchanges), fluent (ie. more human-like),and detailed, while still\nbeing comparably image-relevant as prior work and ablations."", 'To achieve the long-term goal of machines being able to engage humans in\nconversation, our models should captivate the interest of their speaking\npartners. Communication grounded in images, whereby a dialogue is conducted\nbased on a given photo, is a setup naturally appealing to humans (Hu et al.,\n2014). In this work we study large-scale architectures and datasets for this\ngoal. We test a set of neural architectures using state-of-the-art image and\ntext representations, considering various ways to fuse the components. To test\nsuch models, we collect a dataset of grounded human-human conversations, where\nspeakers are asked to play roles given a provided emotional mood or style, as\nthe use of such traits is also a key factor in engagingness (Guo et al., 2019).\nOur dataset, Image-Chat, consists of 202k dialogues over 202k images using 215\npossible style traits. Automatic metrics and human evaluations of engagingness\nshow the efficacy of our approach; in particular, we obtain state-of-the-art\nperformance on the existing IGC task, and our best performing model is almost\non par with humans on the Image-Chat test set (preferred 47.7% of the time).', 'In this paper we consider the problem of continuously discovering image\ncontents by actively asking image based questions and subsequently answering\nthe questions being asked. The key components include a Visual Question\nGeneration (VQG) module and a Visual Question Answering module, in which\nRecurrent Neural Networks (RNN) and Convolutional Neural Network (CNN) are\nused. Given a dataset that contains images, questions and their answers, both\nmodules are trained at the same time, with the difference being VQG uses the\nimages as input and the corresponding questions as output, while VQA uses\nimages and questions as input and the corresponding answers as output. We\nevaluate the self talk process subjectively using Amazon Mechanical Turk, which\nshow effectiveness of the proposed method.', 'State-of-the-art systems in deep question answering proceed as follows: (1)\nan initial document retrieval selects relevant documents, which (2) are then\nprocessed by a neural network in order to extract the final answer. Yet the\nexact interplay between both components is poorly understood, especially\nconcerning the number of candidate documents that should be retrieved. We show\nthat choosing a static number of documents -- as used in prior research --\nsuffers from a noise-information trade-off and yields suboptimal results. As a\nremedy, we propose an adaptive document retrieval model. This learns the\noptimal candidate number for document retrieval, conditional on the size of the\ncorpus and the query. We report extensive experimental results showing that our\nadaptive approach outperforms state-of-the-art methods on multiple benchmark\ndatasets, as well as in the context of corpora with variable sizes.', 'In this paper, we introduce Iterative Text Summarization (ITS), an\niteration-based model for supervised extractive text summarization, inspired by\nthe observation that it is often necessary for a human to read an article\nmultiple times in order to fully understand and summarize its contents. Current\nsummarization approaches read through a document only once to generate a\ndocument representation, resulting in a sub-optimal representation. To address\nthis issue we introduce a model which iteratively polishes the document\nrepresentation on many passes through the document. As part of our model, we\nalso introduce a selective reading mechanism that decides more accurately the\nextent to which each sentence in the model should be updated. Experimental\nresults on the CNN/DailyMail and DUC2002 datasets demonstrate that our model\nsignificantly outperforms state-of-the-art extractive systems when evaluated by\nmachines and by humans.', 'Today all kind of information is getting digitized and along with all this\ndigitization, the huge archive of various kinds of documents is being digitized\ntoo. We know that, Optical Character Recognition is the method through which,\nnewspapers and other paper documents convert into digital resources. But, it is\na fact that this method works on texts only. As a result, if we try to process\nany document which contains non-textual zones, then we will get garbage texts\nas output. That is why; in order to digitize documents properly they should be\nprepossessed carefully. And while preprocessing, segmenting document in\ndifferent regions according to the category properly is most important. But,\nthe Optical Character Recognition processes available for Bangla language have\nno such algorithm that can categorize a newspaper/book page fully. So we worked\nto decompose a document into its several parts like headlines, sub headlines,\ncolumns, images etc. And if the input is skewed and rotated, then the input was\nalso deskewed and de-rotated. To decompose any Bangla document we found out the\nedges of the input image. Then we find out the horizontal and vertical area of\nevery pixel where it lies in. Later on the input image was cut according to\nthese areas. Then we pick each and every sub image and found out their\nheight-width ratio, line height. Then according to these values the sub images\nwere categorized. To deskew the image we found out the skew angle and de skewed\nthe image according to this angle. To de-rotate the image we used the line\nheight, matra line, pixel ratio of matra line.', 'We assume that we are given a time series of data from a dynamical system and\nour task is to learn the flow map of the dynamical system. We present a\ncollection of results on how to enforce constraints coming from the dynamical\nsystem in order to accelerate the training of deep neural networks to represent\nthe flow map of the system as well as increase their predictive ability. In\nparticular, we provide ways to enforce constraints during training for all\nthree major modes of learning, namely supervised, unsupervised and\nreinforcement learning. In general, the dynamic constraints need to include\nterms which are analogous to memory terms in model reduction formalisms. Such\nmemory terms act as a restoring force which corrects the errors committed by\nthe learned flow map during prediction.\n  For supervised learning, the constraints are added to the objective function.\nFor the case of unsupervised learning, in particular generative adversarial\nnetworks, the constraints are introduced by augmenting the input of the\ndiscriminator. Finally, for the case of reinforcement learning and in\nparticular actor-critic methods, the constraints are added to the reward\nfunction. In addition, for the reinforcement learning case, we present a novel\napproach based on homotopy of the action-value function in order to stabilize\nand accelerate training. We use numerical results for the Lorenz system to\nillustrate the various constructions.', 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.', 'We present a contrasting learning approach with data augmentation techniques\nto learn document representations in an unsupervised manner. Inspired by recent\ncontrastive self-supervised learning algorithms used for image and NLP\npretraining, we hypothesize that high-quality document embedding should be\ninvariant to diverse paraphrases that preserve the semantics of the original\ndocument. With different backbones and contrastive learning frameworks, our\nstudy reveals the enormous benefits of contrastive augmentation for document\nrepresentation learning with two additional insights: 1) including data\naugmentation in a contrastive way can substantially improve the embedding\nquality in unsupervised document representation learning, and 2) in general,\nstochastic augmentations generated by simple word-level manipulation work much\nbetter than sentence-level and document-level ones. We plug our method into a\nclassifier and compare it with a broad range of baseline methods on six\nbenchmark datasets. Our method can decrease the classification error rate by up\nto 6.4% over the SOTA approaches on the document classification task, matching\nor even surpassing fully-supervised methods.', 'Computation of document image quality metrics often depends upon the\navailability of a ground truth image corresponding to the document. This limits\nthe applicability of quality metrics in applications such as hyperparameter\noptimization of image processing algorithms that operate on-the-fly on unseen\ndocuments. This work proposes the use of surrogate models to learn the behavior\nof a given document quality metric on existing datasets where ground truth\nimages are available. The trained surrogate model can later be used to predict\nthe metric value on previously unseen document images without requiring access\nto ground truth images. The surrogate model is empirically evaluated on the\nDocument Image Binarization Competition (DIBCO) and the Handwritten Document\nImage Binarization Competition (H-DIBCO) datasets.', 'Recent progress in pretrained Transformer-based language models has shown\ngreat success in learning contextual representation of text. However, due to\nthe quadratic self-attention complexity, most of the pretrained Transformers\nmodels can only handle relatively short text. It is still a challenge when it\ncomes to modeling very long documents. In this work, we propose to use a graph\nattention network on top of the available pretrained Transformers model to\nlearn document embeddings. This graph attention network allows us to leverage\nthe high-level semantic structure of the document. In addition, based on our\ngraph document model, we design a simple contrastive learning strategy to\npretrain our models on a large amount of unlabeled corpus. Empirically, we\ndemonstrate the effectiveness of our approaches in document classification and\ndocument retrieval tasks.', 'We propose a method for automatically answering questions about images by\nbringing together recent advances from natural language processing and computer\nvision. We combine discrete reasoning with uncertain predictions by a\nmulti-world approach that represents uncertainty about the perceived world in a\nbayesian framework. Our approach can handle human questions of high complexity\nabout realistic scenes and replies with range of answer like counts, object\nclasses, instances and lists of them. The system is directly trained from\nquestion-answer pairs. We establish a first benchmark for this task that can be\nseen as a modern attempt at a visual turing test.', ""Most chatbot literature that focuses on improving the fluency and coherence\nof a chatbot, is dedicated to making chatbots more human-like. However, very\nlittle work delves into what really separates humans from chatbots -- humans\nintrinsically understand the effect their responses have on the interlocutor\nand often respond with an intention such as proposing an optimistic view to\nmake the interlocutor feel better. This paper proposes an innovative framework\nto train chatbots to possess human-like intentions. Our framework includes a\nguiding chatbot and an interlocutor model that plays the role of humans. The\nguiding chatbot is assigned an intention and learns to induce the interlocutor\nto reply with responses matching the intention, for example, long responses,\njoyful responses, responses with specific words, etc. We examined our framework\nusing three experimental setups and evaluated the guiding chatbot with four\ndifferent metrics to demonstrate flexibility and performance advantages.\nAdditionally, we performed trials with human interlocutors to substantiate the\nguiding chatbot's effectiveness in influencing the responses of humans to a\ncertain extent. Code will be made available to the public."", 'In industrial data analytics, one of the fundamental problems is to utilize\nthe temporal correlation of the industrial data to make timely predictions in\nthe production process, such as fault prediction and yield prediction. However,\nthe traditional prediction models are fixed while the conditions of the\nmachines change over time, thus making the errors of predictions increase with\nthe lapse of time. In this paper, we propose a general data renewal model to\ndeal with it. Combined with the similarity function and the loss function, it\nestimates the time of updating the existing prediction model, then updates it\naccording to the evaluation function iteratively and adaptively. We have\napplied the data renewal model to two prediction algorithms. The experiments\ndemonstrate that the data renewal model can effectively identify the changes of\ndata, update and optimize the prediction model so as to improve the accuracy of\nprediction.', 'Line separators are used to segregate text-lines from one another in document\nimage analysis. Finding the separator points at every line terminal in a\ndocument image would enable text-line segmentation. In particular, identifying\nthe separators in handwritten text could be a thrilling exercise. Obviously it\nwould be challenging to perform this in the compressed version of a document\nimage and that is the proposed objective in this research. Such an effort would\nprevent the computational burden of decompressing a document for text-line\nsegmentation. Since document images are generally compressed using run length\nencoding (RLE) technique as per the CCITT standards, the first column in the\nRLE will be a white column. The value (depth) in the white column is very low\nwhen a particular line is a text line and the depth could be larger at the\npoint of text line separation. A longer consecutive sequence of such larger\ndepth should indicate the gap between the text lines, which provides the\nseparator region. In case of over separation and under separation issues,\ncorrective actions such as deletion and insertion are suggested respectively.\nAn extensive experimentation is conducted on the compressed images of the\nbenchmark datasets of ICDAR13 and Alireza et al [17] to demonstrate the\nefficacy.', ""Keyword extraction is an important document process that aims at finding a\nsmall set of terms that concisely describe a document's topics. The most\npopular state-of-the-art unsupervised approaches belong to the family of the\ngraph-based methods that build a graph-of-words and use various centrality\nmeasures to score the nodes (candidate keywords). In this work, we follow a\ndifferent path to detect the keywords from a text document by modeling the main\ndistribution of the document's words using local word vector representations.\nThen, we rank the candidates based on their position in the text and the\ndistance between the corresponding local vectors and the main distribution's\ncenter. We confirm the high performance of our approach compared to strong\nbaselines and state-of-the-art unsupervised keyword extraction methods, through\nan extended experimental study, investigating the properties of the local\nrepresentations."", 'We propose a new probabilistic graphical model that jointly models the\ndifficulties of questions, the abilities of participants and the correct\nanswers to questions in aptitude testing and crowdsourcing settings. We devise\nan active learning/adaptive testing scheme based on a greedy minimization of\nexpected model entropy, which allows a more efficient resource allocation by\ndynamically choosing the next question to be asked based on the previous\nresponses. We present experimental results that confirm the ability of our\nmodel to infer the required parameters and demonstrate that the adaptive\ntesting scheme requires fewer questions to obtain the same accuracy as a static\ntest scenario.', ""(abridged) We describe an image-comparison technique of Heidemann and Ritter\nthat uses image compression, and is capable of: (i) detecting novel textures in\na series of images, as well as of: (ii) alerting the user to the similarity of\na new image to a previously-observed texture. This image-comparison technique\nhas been implemented and tested using our Astrobiology Phone-cam system, which\nemploys Bluetooth communication to send images to a local laptop server in the\nfield for the image-compression analysis. We tested the system in a field site\ndisplaying a heterogeneous suite of sandstones, limestones, mudstones and\ncoalbeds. Some of the rocks are partly covered with lichen. The image-matching\nprocedure of this system performed very well with data obtained through our\nfield test, grouping all images of yellow lichens together and grouping all\nimages of a coal bed together, and giving a 91% accuracy for similarity\ndetection. Such similarity detection could be employed to make maps of\ndifferent geological units. The novelty-detection performance of our system was\nalso rather good (a 64% accuracy). Such novelty detection may become valuable\nin searching for new geological units, which could be of astrobiological\ninterest. The image-comparison technique is an unsupervised technique that is\nnot capable of directly classifying an image as containing a particular\ngeological feature; labeling of such geological features is done post facto by\nhuman geologists associated with this study, for the purpose of analyzing the\nsystem's performance. By providing more advanced capabilities for similarity\ndetection and novelty detection, this image-compression technique could be\nuseful in giving more scientific autonomy to robotic planetary rovers, and in\nassisting human astronauts in their geological exploration and assessment."", ""Smart reply systems have been developed for various messaging platforms. In\nthis paper, we introduce Uber's smart reply system: one-click-chat (OCC), which\nis a key enhanced feature on top of the Uber in-app chat system. It enables\ndriver-partners to quickly respond to rider messages using smart replies. The\nsmart replies are dynamically selected according to conversation content using\nmachine learning algorithms. Our system consists of two major components:\nintent detection and reply retrieval, which are very different from standard\nsmart reply systems where the task is to directly predict a reply. It is\ndesigned specifically for mobile applications with short and non-canonical\nmessages. Reply retrieval utilizes pairings between intent and reply based on\ntheir popularity in chat messages as derived from historical data. For intent\ndetection, a set of embedding and classification techniques are experimented\nwith, and we choose to deploy a solution using unsupervised distributed\nembedding and nearest-neighbor classifier. It has the advantage of only\nrequiring a small amount of labeled training data, simplicity in developing and\ndeploying to production, and fast inference during serving and hence highly\nscalable. At the same time, it performs comparably with deep learning\narchitectures such as word-level convolutional neural network. Overall, the\nsystem achieves a high accuracy of 76% on intent detection. Currently, the\nsystem is deployed in production for English-speaking countries and 71% of\nin-app communications between riders and driver-partners adopted the smart\nreplies to speedup the communication process."", 'Image Processing, Optimization and Prediction of an Image play a key role in\nComputer Science. Image processing provides a way to analyze and identify an\nimage .Many areas like medical image processing, Satellite images, natural\nimages and artificial images requires lots of analysis and research on\noptimization. In Image Optimization and Prediction we are combining the\nfeatures of Query Optimization, Image Processing and Prediction . Image\noptimization is used in Pattern analysis, object recognition, in medical Image\nprocessing to predict the type of diseases, in satellite images for predicting\nweather forecast, availability of water or mineral etc. Image Processing,\nOptimization and analysis is a wide open area for research .Lots of research\nhas been conducted in the area of Image analysis and many techniques are\navailable for image analysis but, a single technique is not yet identified for\nimage analysis and prediction .our research is focused on identifying a global\ntechnique for image analysis and Prediction.', 'We investigate the difficulty levels of questions in reading comprehension\ndatasets such as SQuAD, and propose a new question generation setting, named\nDifficulty-controllable Question Generation (DQG). Taking as input a sentence\nin the reading comprehension paragraph and some of its text fragments (i.e.,\nanswers) that we want to ask questions about, a DQG method needs to generate\nquestions each of which has a given text fragment as its answer, and meanwhile\nthe generation is under the control of specified difficulty labels---the output\nquestions should satisfy the specified difficulty as much as possible. To solve\nthis task, we propose an end-to-end framework to generate questions of\ndesignated difficulty levels by exploring a few important intuitions. For\nevaluation, we prepared the first dataset of reading comprehension questions\nwith difficulty labels. The results show that the question generated by our\nframework not only have better quality under the metrics like BLEU, but also\ncomply with the specified difficulty labels.', 'The ability to describe images with natural language sentences is the\nhallmark for image and language understanding. Such a system has wide ranging\napplications such as annotating images and using natural sentences to search\nfor images.In this project we focus on the task of bidirectional image\nretrieval: such asystem is capable of retrieving an image based on a sentence\n(image search) andretrieve sentence based on an image query (image annotation).\nWe present asystem based on a global ranking objective function which uses a\ncombinationof convolutional neural networks (CNN) and multi layer perceptrons\n(MLP).It takes a pair of image and sentence and processes them in different\nchannels,finally embedding it into a common multimodal vector space. These\nembeddingsencode abstract semantic information about the two inputs and can be\ncomparedusing traditional information retrieval approaches. For each such pair,\nthe modelreturns a score which is interpretted as a similarity metric. If this\nscore is high,the image and sentence are likely to convey similar meaning, and\nif the score is low then they are likely not to.\n  The visual input is modeled via deep convolutional neural network. On\ntheother hand we explore three models for the textual module. The first one\nisbag of words with an MLP. The second one uses n-grams (bigram, trigrams,and a\ncombination of trigram & skip-grams) with an MLP. The third is morespecialized\ndeep network specific for modeling variable length sequences (SSE).We report\ncomparable performance to recent work in the field, even though ouroverall\nmodel is simpler. We also show that the training time choice of how wecan\ngenerate our negative samples has a significant impact on performance, and can\nbe used to specialize the bi-directional system in one particular task.', 'Standard neural machine translation (NMT) is on the assumption that the\ndocument-level context is independent. Most existing document-level NMT\napproaches are satisfied with a smattering sense of global document-level\ninformation, while this work focuses on exploiting detailed document-level\ncontext in terms of a memory network. The capacity of the memory network that\ndetecting the most relevant part of the current sentence from memory renders a\nnatural solution to model the rich document-level context. In this work, the\nproposed document-aware memory network is implemented to enhance the\nTransformer NMT baseline. Experiments on several tasks show that the proposed\nmethod significantly improves the NMT performance over strong Transformer\nbaselines and other related studies.', 'In this paper, we show that certain phrases although not present in a given\nquestion/query, play a very important role in answering the question. Exploring\nthe role of such phrases in answering questions not only reduces the dependency\non matching question phrases for extracting answers, but also improves the\nquality of the extracted answers. Here matching question phrases means phrases\nwhich co-occur in given question and candidate answers. To achieve the above\ndiscussed goal, we introduce a bigram-based word graph model populated with\nsemantic and topical relatedness of terms in the given document. Next, we apply\nan improved version of ranking with a prior-based approach, which ranks all\nwords in the candidate document with respect to a set of root words (i.e.\nnon-stopwords present in the question and in the candidate document). As a\nresult, terms logically related to the root words are scored higher than terms\nthat are not related to the root words. Experimental results show that our\ndevised system performs better than state-of-the-art for the task of answering\nWhy-questions.', 'As the wide adoption of intelligent chatbot in human daily life, user demands\nfor such systems evolve from basic task-solving conversations to more casual\nand friend-like communication. To meet the user needs and build emotional bond\nwith users, it is essential for social chatbots to incorporate more human-like\nand advanced linguistic features. In this paper, we investigate the usage of a\ncommonly used rhetorical device by human -- metaphor for social chatbot. Our\nwork first designs a metaphor generation framework, which generates topic-aware\nand novel figurative sentences. By embedding the framework into a chatbot\nsystem, we then enables the chatbot to communicate with users using figurative\nlanguage. Human annotators validate the novelty and properness of the generated\nmetaphors. More importantly, we evaluate the effects of employing metaphors in\nhuman-chatbot conversations. Experiments indicate that our system effectively\narouses user interests in communicating with our chatbot, resulting in\nsignificantly longer human-chatbot conversations.', 'With the increase of information, document classification as one of the\nmethods of text mining, plays vital role in many management and organizing\ninformation. Document classification is the process of assigning a document to\none or more predefined category labels. Document classification includes\ndifferent parts such as text processing, term selection, term weighting and\nfinal classification. The accuracy of document classification is very\nimportant. Thus improvement in each part of classification should lead to\nbetter results and higher precision. Term weighting has a great impact on the\naccuracy of the classification. Most of the existing weighting methods exploit\nthe statistical information of terms in documents and do not consider semantic\nrelations between words. In this paper, an automated document classification\nsystem is presented that uses a novel term weighting method based on semantic\nrelations between terms. To evaluate the proposed method, three standard\nPersian corpuses are used. Experiment results show 2 to 4 percent improvement\nin classification accuracy compared with the best previous designed system for\nPersian documents.', 'Generating realistic images is difficult, and many formulations for this task\nhave been proposed recently. If we restrict the task to that of generating a\nparticular class of images, however, the task becomes more tractable. That is\nto say, instead of generating an arbitrary image as a sample from the manifold\nof natural images, we propose to sample images from a particular ""subspace"" of\nnatural images, directed by a low-resolution image from the same subspace. The\nproblem we address, while close to the formulation of the single-image\nsuper-resolution problem, is in fact rather different. Single image\nsuper-resolution is the task of predicting the image closest to the ground\ntruth from a relatively low resolution image. We propose to produce samples of\nhigh resolution images given extremely small inputs with a new method called\nLatent Adversarial Generator (LAG). In our generative sampling framework, we\nonly use the input (possibly of very low-resolution) to direct what class of\nsamples the network should produce. As such, the output of our algorithm is not\na unique image that relates to the input, but rather a possible se} of related\nimages sampled from the manifold of natural images. Our method learns\nexclusively in the latent space of the adversary using perceptual loss -- it\ndoes not have a pixel loss.', 'Text Mining is a field that aims at extracting information from textual data.\nOne of the challenges of such field of study comes from the pre-processing\nstage in which a vector (and structured) representation should be extracted\nfrom unstructured data. The common extraction creates large and sparse vectors\nrepresenting the importance of each term to a document. As such, this usually\nleads to the curse-of-dimensionality that plagues most machine learning\nalgorithms. To cope with this issue, in this paper we propose a new supervised\nfeature extraction and reduction algorithm, named DCDistance, that creates\nfeatures based on the distance between a document to a representative of each\nclass label. As such, the proposed technique can reduce the features set in\nmore than 99% of the original set. Additionally, this algorithm was also\ncapable of improving the classification accuracy over a set of benchmark\ndatasets when compared to traditional and state-of-the-art features selection\nalgorithms.', 'There is more to images than their objective physical content: for example,\nadvertisements are created to persuade a viewer to take a certain action. We\npropose the novel problem of automatic advertisement understanding. To enable\nresearch on this problem, we create two datasets: an image dataset of 64,832\nimage ads, and a video dataset of 3,477 ads. Our data contains rich annotations\nencompassing the topic and sentiment of the ads, questions and answers\ndescribing what actions the viewer is prompted to take and the reasoning that\nthe ad presents to persuade the viewer (""What should I do according to this ad,\nand why should I do it?""), and symbolic references ads make (e.g. a dove\nsymbolizes peace). We also analyze the most common persuasive strategies ads\nuse, and the capabilities that computer vision systems should have to\nunderstand these strategies. We present baseline classification results for\nseveral prediction tasks, including automatically answering questions about the\nmessages of the ads.', ""We have seen great progress in basic perceptual tasks such as object\nrecognition and detection. However, AI models still fail to match humans in\nhigh-level vision tasks due to the lack of capacities for deeper reasoning.\nRecently the new task of visual question answering (QA) has been proposed to\nevaluate a model's capacity for deep image understanding. Previous works have\nestablished a loose, global association between QA sentences and images.\nHowever, many questions and answers, in practice, relate to local regions in\nthe images. We establish a semantic link between textual descriptions and image\nregions by object-level grounding. It enables a new type of QA with visual\nanswers, in addition to textual answers used in previous work. We study the\nvisual QA tasks in a grounded setting with a large collection of 7W\nmultiple-choice QA pairs. Furthermore, we evaluate human performance and\nseveral baseline models on the QA tasks. Finally, we propose a novel LSTM model\nwith spatial attention to tackle the 7W QA tasks."", 'Generating natural language descriptions of images is an important capability\nfor a robot or other visual-intelligence driven AI agent that may need to\ncommunicate with human users about what it is seeing. Such image captioning\nmethods are typically trained by maximising the likelihood of ground-truth\nannotated caption given the image. While simple and easy to implement, this\napproach does not directly maximise the language quality metrics we care about\nsuch as CIDEr. In this paper we investigate training image captioning methods\nbased on actor-critic reinforcement learning in order to directly optimise\nnon-differentiable quality metrics of interest. By formulating a per-token\nadvantage and value computation strategy in this novel reinforcement learning\nbased captioning model, we show that it is possible to achieve the state of the\nart performance on the widely used MSCOCO benchmark.', 'Deep reading models for question-answering have demonstrated promising\nperformance over the last couple of years. However current systems tend to\nlearn how to cleverly extract a span of the source document, based on its\nsimilarity with the question, instead of seeking for the appropriate answer.\nIndeed, a reading machine should be able to detect relevant passages in a\ndocument regarding a question, but more importantly, it should be able to\nreason over the important pieces of the document in order to produce an answer\nwhen it is required. To motivate this purpose, we present ReviewQA, a\nquestion-answering dataset based on hotel reviews. The questions of this\ndataset are linked to a set of relational understanding competencies that we\nexpect a model to master. Indeed, each question comes with an associated type\nthat characterizes the required competency. With this framework, it is possible\nto benchmark the main families of models and to get an overview of what are the\nstrengths and the weaknesses of a given model on the set of tasks evaluated in\nthis dataset. Our corpus contains more than 500.000 questions in natural\nlanguage over 100.000 hotel reviews. Our setup is projective, the answer of a\nquestion does not need to be extracted from a document, like in most of the\nrecent datasets, but selected among a set of candidates that contains all the\npossible answers to the questions of the dataset. Finally, we present several\nbaselines over this dataset.', 'Document categorization, which aims to assign a topic label to each document,\nplays a fundamental role in a wide variety of applications. Despite the success\nof existing studies in conventional supervised document classification, they\nare less concerned with two real problems: (1) the presence of metadata: in\nmany domains, text is accompanied by various additional information such as\nauthors and tags. Such metadata serve as compelling topic indicators and should\nbe leveraged into the categorization framework; (2) label scarcity: labeled\ntraining samples are expensive to obtain in some cases, where categorization\nneeds to be performed using only a small set of annotated data. In recognition\nof these two challenges, we propose MetaCat, a minimally supervised framework\nto categorize text with metadata. Specifically, we develop a generative process\ndescribing the relationships between words, documents, labels, and metadata.\nGuided by the generative model, we embed text and metadata into the same\nsemantic space to encode heterogeneous signals. Then, based on the same\ngenerative process, we synthesize training samples to address the bottleneck of\nlabel scarcity. We conduct a thorough evaluation on a wide range of datasets.\nExperimental results prove the effectiveness of MetaCat over many competitive\nbaselines.', ""Neural machine translation (NMT) has arguably achieved human level parity\nwhen trained and evaluated at the sentence-level. Document-level neural machine\ntranslation has received less attention and lags behind its sentence-level\ncounterpart. The majority of the proposed document-level approaches investigate\nways of conditioning the model on several source or target sentences to capture\ndocument context. These approaches require training a specialized NMT model\nfrom scratch on parallel document-level corpora. We propose an approach that\ndoesn't require training a specialized model on parallel document-level corpora\nand is applied to a trained sentence-level NMT model at decoding time. We\nprocess the document from left to right multiple times and self-train the\nsentence-level model on pairs of source sentences and generated translations.\nOur approach reinforces the choices made by the model, thus making it more\nlikely that the same choices will be made in other sentences in the document.\nWe evaluate our approach on three document-level datasets: NIST\nChinese-English, WMT'19 Chinese-English and OpenSubtitles English-Russian. We\ndemonstrate that our approach has higher BLEU score and higher human preference\nthan the baseline. Qualitative analysis of our approach shows that choices made\nby model are consistent across the document."", 'In this paper, we study the problem of answering visual analogy questions.\nThese questions take the form of image A is to image B as image C is to what.\nAnswering these questions entails discovering the mapping from image A to image\nB and then extending the mapping to image C and searching for the image D such\nthat the relation from A to B holds for C to D. We pose this problem as\nlearning an embedding that encourages pairs of analogous images with similar\ntransformations to be close together using convolutional neural networks with a\nquadruple Siamese architecture. We introduce a dataset of visual analogy\nquestions in natural images, and show first results of its kind on solving\nanalogy questions on natural images.', 'We explore using multilingual document embeddings for nearest neighbor mining\nof parallel data. Three document-level representations are investigated: (i)\ndocument embeddings generated by simply averaging multilingual sentence\nembeddings; (ii) a neural bag-of-words (BoW) document encoding model; (iii) a\nhierarchical multilingual document encoder (HiDE) that builds on our\nsentence-level model. The results show document embeddings derived from\nsentence-level averaging are surprisingly effective for clean datasets, but\nsuggest models trained hierarchically at the document-level are more effective\non noisy data. Analysis experiments demonstrate our hierarchical models are\nvery robust to variations in the underlying sentence embedding quality. Using\ndocument embeddings trained with HiDE achieves state-of-the-art performance on\nUnited Nations (UN) parallel document mining, 94.9% P@1 for en-fr and 97.3% P@1\nfor en-es.', 'COVID-19 has resulted in an ongoing pandemic and as of 12 June 2020, has\ncaused more than 7.4 million cases and over 418,000 deaths. The highly dynamic\nand rapidly evolving situation with COVID-19 has made it difficult to access\naccurate, on-demand information regarding the disease. Online communities,\nforums, and social media provide potential venues to search for relevant\nquestions and answers, or post questions and seek answers from other members.\nHowever, due to the nature of such sites, there are always a limited number of\nrelevant questions and responses to search from, and posted questions are\nrarely answered immediately. With the advancements in the field of natural\nlanguage processing, particularly in the domain of language models, it has\nbecome possible to design chatbots that can automatically answer consumer\nquestions. However, such models are rarely applied and evaluated in the\nhealthcare domain, to meet the information needs with accurate and up-to-date\nhealthcare data. In this paper, we propose to apply a language model for\nautomatically answering questions related to COVID-19 and qualitatively\nevaluate the generated responses. We utilized the GPT-2 language model and\napplied transfer learning to retrain it on the COVID-19 Open Research Dataset\n(CORD-19) corpus. In order to improve the quality of the generated responses,\nwe applied 4 different approaches, namely tf-idf, BERT, BioBERT, and USE to\nfilter and retain relevant sentences in the responses. In the performance\nevaluation step, we asked two medical experts to rate the responses. We found\nthat BERT and BioBERT, on average, outperform both tf-idf and USE in\nrelevance-based sentence filtering tasks. Additionally, based on the chatbot,\nwe created a user-friendly interactive web application to be hosted online.', 'Explaining the predictions of AI models is paramount in safety-critical\napplications, such as in legal or medical domains. One form of explanation for\na prediction is an extractive rationale, i.e., a subset of features of an\ninstance that lead the model to give its prediction on the instance. Previous\nworks on generating extractive rationales usually employ a two-phase model: a\nselector that selects the most important features (i.e., the rationale)\nfollowed by a predictor that makes the prediction based exclusively on the\nselected features. One disadvantage of these works is that the main signal for\nlearning to select features comes from the comparison of the answers given by\nthe predictor and the ground-truth answers. In this work, we propose to squeeze\nmore information from the predictor via an information calibration method. More\nprecisely, we train two models jointly: one is a typical neural model that\nsolves the task at hand in an accurate but black-box manner, and the other is a\nselector-predictor model that additionally produces a rationale for its\nprediction. The first model is used as a guide to the second model. We use an\nadversarial-based technique to calibrate the information extracted by the two\nmodels such that the difference between them is an indicator of the missed or\nover-selected features. In addition, for natural language tasks, we propose to\nuse a language-model-based regularizer to encourage the extraction of fluent\nrationales. Experimental results on a sentiment analysis task as well as on\nthree tasks from the legal domain show the effectiveness of our approach to\nrationale extraction.', 'This paper summarizes our entries to both subtasks of the first DialDoc\nshared task which focuses on the agent response prediction task in\ngoal-oriented document-grounded dialogs. The task is split into two subtasks:\npredicting a span in a document that grounds an agent turn and generating an\nagent response based on a dialog and grounding document. In the first subtask,\nwe restrict the set of valid spans to the ones defined in the dataset, use a\nbiaffine classifier to model spans, and finally use an ensemble of different\nmodels. For the second subtask, we use a cascaded model which grounds the\nresponse prediction on the predicted span instead of the full document. With\nthese approaches, we obtain significant improvements in both subtasks compared\nto the baseline.', 'Automatically describing the content of an image is a fundamental problem in\nartificial intelligence that connects computer vision and natural language\nprocessing. In this paper, we present a generative model based on a deep\nrecurrent architecture that combines recent advances in computer vision and\nmachine translation and that can be used to generate natural sentences\ndescribing an image. The model is trained to maximize the likelihood of the\ntarget description sentence given the training image. Experiments on several\ndatasets show the accuracy of the model and the fluency of the language it\nlearns solely from image descriptions. Our model is often quite accurate, which\nwe verify both qualitatively and quantitatively. Finally, given the recent\nsurge of interest in this task, a competition was organized in 2015 using the\nnewly released COCO dataset. We describe and analyze the various improvements\nwe applied to our own baseline and show the resulting performance in the\ncompetition, which we won ex-aequo with a team from Microsoft Research, and\nprovide an open source implementation in TensorFlow.', ""We introduce the first goal-driven training for visual question answering and\ndialog agents. Specifically, we pose a cooperative 'image guessing' game\nbetween two agents -- Qbot and Abot -- who communicate in natural language\ndialog so that Qbot can select an unseen image from a lineup of images. We use\ndeep reinforcement learning (RL) to learn the policies of these agents\nend-to-end -- from pixels to multi-agent multi-round dialog to game reward.\n  We demonstrate two experimental results.\n  First, as a 'sanity check' demonstration of pure RL (from scratch), we show\nresults on a synthetic world, where the agents communicate in ungrounded\nvocabulary, i.e., symbols with no pre-specified meanings (X, Y, Z). We find\nthat two bots invent their own communication protocol and start using certain\nsymbols to ask/answer about certain visual attributes (shape/color/style).\nThus, we demonstrate the emergence of grounded language and communication among\n'visual' dialog agents with no human supervision.\n  Second, we conduct large-scale real-image experiments on the VisDial dataset,\nwhere we pretrain with supervised dialog data and show that the RL 'fine-tuned'\nagents significantly outperform SL agents. Interestingly, the RL Qbot learns to\nask questions that Abot is good at, ultimately resulting in more informative\ndialog and a better team."", ""Neural Networks and Deep Learning have seen an upsurge of research in the\npast decade due to the improved results. Generates text from the given image is\na crucial task that requires the combination of both sectors which are computer\nvision and natural language processing in order to understand an image and\nrepresent it using a natural language. However existing works have all been\ndone on a particular lingual domain and on the same set of data. This leads to\nthe systems being developed to perform poorly on images that belong to specific\nlocales' geographical context. TextMage is a system that is capable of\nunderstanding visual scenes that belong to the Bangladeshi geographical context\nand use its knowledge to represent what it understands in Bengali. Hence, we\nhave trained a model on our previously developed and published dataset named\nBanglaLekhaImageCaptions. This dataset contains 9,154 images along with two\nannotations for each image. In order to access performance, the proposed model\nhas been implemented and evaluated."", 'It is known that $O(N)$ parameters are sufficient for neural networks to\nmemorize arbitrary $N$ input-label pairs. By exploiting depth, we show that\n$O(N^{2/3})$ parameters suffice to memorize $N$ pairs, under a mild condition\non the separation of input points. In particular, deeper networks (even with\nwidth $3$) are shown to memorize more pairs than shallow networks, which also\nagrees with the recent line of works on the benefits of depth for function\napproximation. We also provide empirical results that support our theoretical\nfindings.', 'We consider the task of learning to estimate human pose in still images. In\norder to avoid the high cost of full supervision, we propose to use a diverse\ndata set, which consists of two types of annotations: (i) a small number of\nimages are labeled using the expensive ground-truth pose; and (ii) other images\nare labeled using the inexpensive action label. As action information helps\nnarrow down the pose of a human, we argue that this approach can help reduce\nthe cost of training without significantly affecting the accuracy. To\ndemonstrate this we design a probabilistic framework that employs two\ndistributions: (i) a conditional distribution to model the uncertainty over the\nhuman pose given the image and the action; and (ii) a prediction distribution,\nwhich provides the pose of an image without using any action information. We\njointly estimate the parameters of the two aforementioned distributions by\nminimizing their dissimilarity coefficient, as measured by a task-specific loss\nfunction. During both training and testing, we only require an efficient\nsampling strategy for both the aforementioned distributions. This allows us to\nuse deep probabilistic networks that are capable of providing accurate pose\nestimates for previously unseen images. Using the MPII data set, we show that\nour approach outperforms baseline methods that either do not use the diverse\nannotations or rely on pointwise estimates of the pose.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', 'The increasing amount of online content motivated the development of\nmulti-document summarization methods. In this work, we explore straightforward\napproaches to extend single-document summarization methods to multi-document\nsummarization. The proposed methods are based on the hierarchical combination\nof single-document summaries, and achieves state of the art results.', 'We propose a novel approach to identify the difficulty of visual questions\nfor Visual Question Answering (VQA) without direct supervision or annotations\nto the difficulty. Prior works have considered the diversity of ground-truth\nanswers of human annotators. In contrast, we analyze the difficulty of visual\nquestions based on the behavior of multiple different VQA models. We propose to\ncluster the entropy values of the predicted answer distributions obtained by\nthree different models: a baseline method that takes as input images and\nquestions, and two variants that take as input images only and questions only.\nWe use a simple k-means to cluster the visual questions of the VQA v2\nvalidation set. Then we use state-of-the-art methods to determine the accuracy\nand the entropy of the answer distributions for each cluster. A benefit of the\nproposed method is that no annotation of the difficulty is required, because\nthe accuracy of each cluster reflects the difficulty of visual questions that\nbelong to it. Our approach can identify clusters of difficult visual questions\nthat are not answered correctly by state-of-the-art methods. Detailed analysis\non the VQA v2 dataset reveals that 1) all methods show poor performances on the\nmost difficult cluster (about 10\\% accuracy), 2) as the cluster difficulty\nincreases, the answers predicted by the different methods begin to differ, and\n3) the values of cluster entropy are highly correlated with the cluster\naccuracy. We show that our approach has the advantage of being able to assess\nthe difficulty of visual questions without ground-truth (\\ie, the test set of\nVQA v2) by assigning them to one of the clusters. We expect that this can\nstimulate the development of novel directions of research and new algorithms.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.', 'Predicting pairs of anchor users plays an important role in the cross-network\nanalysis. Due to the expensive costs of labeling anchor users for training\nprediction models, we consider in this paper the problem of minimizing the\nnumber of user pairs across multiple networks for labeling as to improve the\naccuracy of the prediction. To this end, we present a deep active learning\nmodel for anchor user prediction (DALAUP for short). However, active learning\nfor anchor user sampling meets the challenges of non-i.i.d. user pair data\ncaused by network structures and the correlation among anchor or non-anchor\nuser pairs. To solve the challenges, DALAUP uses a couple of neural networks\nwith shared-parameter to obtain the vector representations of user pairs, and\nensembles three query strategies to select the most informative user pairs for\nlabeling and model training. Experiments on real-world social network data\ndemonstrate that DALAUP outperforms the state-of-the-art approaches.', 'Relating entities and events in text is a key component of natural language\nunderstanding. Cross-document coreference resolution, in particular, is\nimportant for the growing interest in multi-document analysis tasks. In this\nwork we propose a new model that extends the efficient sequential prediction\nparadigm for coreference resolution to cross-document settings and achieves\ncompetitive results for both entity and event coreference while provides strong\nevidence of the efficacy of both sequential models and higher-order inference\nin cross-document settings. Our model incrementally composes mentions into\ncluster representations and predicts links between a mention and the already\nconstructed clusters, approximating a higher-order model. In addition, we\nconduct extensive ablation studies that provide new insights into the\nimportance of various inputs and representation types in coreference.', 'We investigate the pertinence of methods from algebraic topology for text\ndata analysis. These methods enable the development of\nmathematically-principled isometric-invariant mappings from a set of vectors to\na document embedding, which is stable with respect to the geometry of the\ndocument in the selected metric space. In this work, we evaluate the utility of\nthese topology-based document representations in traditional NLP tasks,\nspecifically document clustering and sentiment classification. We find that the\nembeddings do not benefit text analysis. In fact, performance is worse than\nsimple techniques like $\\textit{tf-idf}$, indicating that the geometry of the\ndocument does not provide enough variability for classification on the basis of\ntopic or sentiment in the chosen datasets.', 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.', 'Document-level relation extraction requires integrating information within\nand across multiple sentences of a document and capturing complex interactions\nbetween inter-sentence entities. However, effective aggregation of relevant\ninformation in the document remains a challenging research question. Existing\napproaches construct static document-level graphs based on syntactic trees,\nco-references or heuristics from the unstructured text to model the\ndependencies. Unlike previous methods that may not be able to capture rich\nnon-local interactions for inference, we propose a novel model that empowers\nthe relational reasoning across sentences by automatically inducing the latent\ndocument-level graph. We further develop a refinement strategy, which enables\nthe model to incrementally aggregate relevant information for multi-hop\nreasoning. Specifically, our model achieves an F1 score of 59.05 on a\nlarge-scale document-level dataset (DocRED), significantly improving over the\nprevious results, and also yields new state-of-the-art results on the CDR and\nGDA dataset. Furthermore, extensive analyses show that the model is able to\ndiscover more accurate inter-sentence relations.', 'Sentence embedding is an important research topic in natural language\nprocessing. It is essential to generate a good embedding vector that fully\nreflects the semantic meaning of a sentence in order to achieve an enhanced\nperformance for various natural language processing tasks, such as machine\ntranslation and document classification. Thus far, various sentence embedding\nmodels have been proposed, and their feasibility has been demonstrated through\ngood performances on tasks following embedding, such as sentiment analysis and\nsentence classification. However, because the performances of sentence\nclassification and sentiment analysis can be enhanced by using a simple\nsentence representation method, it is not sufficient to claim that these models\nfully reflect the meanings of sentences based on good performances for such\ntasks. In this paper, inspired by human language recognition, we propose the\nfollowing concept of semantic coherence, which should be satisfied for a good\nsentence embedding method: similar sentences should be located close to each\nother in the embedding space. Then, we propose the Paraphrase-Thought\n(P-thought) model to pursue semantic coherence as much as possible.\nExperimental results on two paraphrase identification datasets (MS COCO and STS\nbenchmark) show that the P-thought models outperform the benchmarked sentence\nembedding methods.', 'This document should serve as a quick reference for and guide to the\nimplementation of linear GQ($\\lambda$), a gradient-based off-policy\ntemporal-difference learning algorithm. Explanation of the intuition and theory\nbehind the algorithm are provided elsewhere (e.g., Maei & Sutton 2010, Maei\n2011). If you questions or concerns about the content in this document or the\nattached java code please email Adam White (adam.white@ualberta.ca).\n  The code is provided as part of the source files in the arXiv submission.', 'Recent studies emphasize the need of document context in human evaluation of\nmachine translations, but little research has been done on the impact of user\ninterfaces on annotator productivity and the reliability of assessments. In\nthis work, we compare human assessment data from the last two WMT evaluation\ncampaigns collected via two different methods for document-level evaluation.\nOur analysis shows that a document-centric approach to evaluation where the\nannotator is presented with the entire document context on a screen leads to\nhigher quality segment and document level assessments. It improves the\ncorrelation between segment and document scores and increases inter-annotator\nagreement for document scores but is considerably more time consuming for\nannotators.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I am in the process of developing a deep learning model designed to solve a range of tasks, each with its unique structure. I am operating under the assumption that the structures of these various tasks can be represented by graphs. Additionally, I am working with the constraint of having limited labeled data for each task. My objective is twofold. Firstly, I aim to create a model capable of processing graphical data, learning from diverse tasks, and then adapting to new tasks using a few-shot learning approach. My current strategy involves the use of a graph meta-learning algorithm. Secondly, I have observed that when a sub-task is adversarially designed, it can severely disrupt the meta-learning algorithm. As a result, I need to ensure my model is robust against adversarial attacks. My current strategy to achieve this involves introducing randomized variations in the model architecture, input data, and intermediate latent results to fortify my model. I am also open to other robust training strategies. By integrating these techniques, my goal is to develop a scalable and robust deep learning model that can efficiently adapt to new graph-based tasks and resist adversarial attacks.","[""Learning tasks on source code (i.e., formal languages) have been considered\nrecently, but most work has tried to transfer natural language methods and does\nnot capitalize on the unique opportunities offered by code's known syntax. For\nexample, long-range dependencies induced by using the same variable or function\nin distant locations are often not considered. We propose to use graphs to\nrepresent both the syntactic and semantic structure of code and use graph-based\ndeep learning methods to learn to reason over program structures.\n  In this work, we present how to construct graphs from source code and how to\nscale Gated Graph Neural Networks training to such large graphs. We evaluate\nour method on two tasks: VarNaming, in which a network attempts to predict the\nname of a variable given its usage, and VarMisuse, in which the network learns\nto reason about selecting the correct variable that should be used at a given\nprogram location. Our comparison to methods that use less structured program\nrepresentations shows the advantages of modeling known structure, and suggests\nthat our models learn to infer meaningful names and to solve the VarMisuse task\nin many cases. Additionally, our testing showed that VarMisuse identifies a\nnumber of bugs in mature open-source projects."", 'Most previous few-shot learning algorithms are based on meta-training with\nfake few-shot tasks as training samples, where large labeled base classes are\nrequired. The trained model is also limited by the type of tasks. In this paper\nwe propose a new paradigm of unsupervised few-shot learning to repair the\ndeficiencies. We solve the few-shot tasks in two phases: meta-training a\ntransferable feature extractor via contrastive self-supervised learning and\ntraining a classifier using graph aggregation, self-distillation and manifold\naugmentation. Once meta-trained, the model can be used in any type of tasks\nwith a task-dependent classifier training. Our method achieves state of-the-art\nperformance in a variety of established few-shot tasks on the standard few-shot\nvisual classification datasets, with an 8- 28% increase compared to the\navailable unsupervised few-shot learning methods.', 'Label propagation is a powerful and flexible semi-supervised learning\ntechnique on graphs. Neural networks, on the other hand, have proven track\nrecords in many supervised learning tasks. In this work, we propose a training\nframework with a graph-regularised objective, namely ""Neural Graph Machines"",\nthat can combine the power of neural networks and label propagation. This work\ngeneralises previous literature on graph-augmented training of neural networks,\nenabling it to be applied to multiple neural architectures (Feed-forward NNs,\nCNNs and LSTM RNNs) and a wide range of graphs. The new objective allows the\nneural networks to harness both labeled and unlabeled data by: (a) allowing the\nnetwork to train using labeled data as in the supervised setting, (b) biasing\nthe network to learn similar hidden representations for neighboring nodes on a\ngraph, in the same vein as label propagation. Such architectures with the\nproposed objective can be trained efficiently using stochastic gradient descent\nand scaled to large graphs, with a runtime that is linear in the number of\nedges. The proposed joint training approach convincingly outperforms many\nexisting methods on a wide range of tasks (multi-label classification on social\ngraphs, news categorization, document classification and semantic intent\nclassification), with multiple forms of graph inputs (including graphs with and\nwithout node-level features) and using different types of neural networks.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'Meta-learning extracts common knowledge from learning different tasks and\nuses it for unseen tasks. It can significantly improve tasks that suffer from\ninsufficient training data, e.g., few shot learning. In most meta-learning\nmethods, tasks are implicitly related by sharing parameters or optimizer. In\nthis paper, we show that a meta-learner that explicitly relates tasks on a\ngraph describing the relations of their output dimensions (e.g., classes) can\nsignificantly improve few shot learning. The graph\'s structure is usually free\nor cheap to obtain but has rarely been explored in previous works. We develop a\nnovel meta-learner of this type for prototype-based classification, in which a\nprototype is generated for each class, such that the nearest neighbor search\namong the prototypes produces an accurate classification. The meta-learner,\ncalled ""Gated Propagation Network (GPN)"", learns to propagate messages between\nprototypes of different classes on the graph, so that learning the prototype of\neach class benefits from the data of other related classes. In GPN, an\nattention mechanism aggregates messages from neighboring classes of each class,\nwith a gate choosing between the aggregated message and the message from the\nclass itself. We train GPN on a sequence of tasks from many-shot to few shot\ngenerated by subgraph sampling. During training, it is able to reuse and update\npreviously achieved prototypes from the memory in a life-long learning cycle.\nIn experiments, under different training-test discrepancy and test task\ngeneration settings, GPN outperforms recent meta-learning methods on two\nbenchmark datasets. The code of GPN and dataset generation is available at\nhttps://github.com/liulu112601/Gated-Propagation-Net.', 'Deep neural networks excel in regimes with large amounts of data, but tend to\nstruggle when data is scarce or when they need to adapt quickly to changes in\nthe task. In response, recent work in meta-learning proposes training a\nmeta-learner on a distribution of similar tasks, in the hopes of generalization\nto novel but related tasks by learning a high-level strategy that captures the\nessence of the problem it is asked to solve. However, many recent meta-learning\napproaches are extensively hand-designed, either using architectures\nspecialized to a particular application, or hard-coding algorithmic components\nthat constrain how the meta-learner solves the task. We propose a class of\nsimple and generic meta-learner architectures that use a novel combination of\ntemporal convolutions and soft attention; the former to aggregate information\nfrom past experience and the latter to pinpoint specific pieces of information.\nIn the most extensive set of meta-learning experiments to date, we evaluate the\nresulting Simple Neural AttentIve Learner (or SNAIL) on several\nheavily-benchmarked tasks. On all tasks, in both supervised and reinforcement\nlearning, SNAIL attains state-of-the-art performance by significant margins.', 'Meta-reinforcement learning (meta-RL) aims to learn from multiple training\ntasks the ability to adapt efficiently to unseen test tasks. Despite the\nsuccess, existing meta-RL algorithms are known to be sensitive to the task\ndistribution shift. When the test task distribution is different from the\ntraining task distribution, the performance may degrade significantly. To\naddress this issue, this paper proposes Model-based Adversarial\nMeta-Reinforcement Learning (AdMRL), where we aim to minimize the worst-case\nsub-optimality gap -- the difference between the optimal return and the return\nthat the algorithm achieves after adaptation -- across all tasks in a family of\ntasks, with a model-based approach. We propose a minimax objective and optimize\nit by alternating between learning the dynamics model on a fixed task and\nfinding the adversarial task for the current model -- the task for which the\npolicy induced by the model is maximally suboptimal. Assuming the family of\ntasks is parameterized, we derive a formula for the gradient of the\nsuboptimality with respect to the task parameters via the implicit function\ntheorem, and show how the gradient estimator can be efficiently implemented by\nthe conjugate gradient method and a novel use of the REINFORCE estimator. We\nevaluate our approach on several continuous control benchmarks and demonstrate\nits efficacy in the worst-case performance over all tasks, the generalization\npower to out-of-distribution tasks, and in training and test time sample\nefficiency, over existing state-of-the-art meta-RL algorithms.', 'This article reviews meta-learning also known as learning-to-learn which\nseeks rapid and accurate model adaptation to unseen tasks with applications in\nhighly automated AI, few-shot learning, natural language processing and\nrobotics. Unlike deep learning, meta-learning can be applied to few-shot\nhigh-dimensional datasets and considers further improving model generalization\nto unseen tasks. Deep learning is focused upon in-sample prediction and\nmeta-learning concerns model adaptation for out-of-sample prediction.\nMeta-learning can continually perform self-improvement to achieve highly\nautonomous AI. Meta-learning may serve as an additional generalization block\ncomplementary for original deep learning model. Meta-learning seeks adaptation\nof machine learning models to unseen tasks which are vastly different from\ntrained tasks. Meta-learning with coevolution between agent and environment\nprovides solutions for complex tasks unsolvable by training from scratch.\nMeta-learning methodology covers a wide range of great minds and thoughts. We\nbriefly introduce meta-learning methodologies in the following categories:\nblack-box meta-learning, metric-based meta-learning, layered meta-learning and\nBayesian meta-learning framework. Recent applications concentrate upon the\nintegration of meta-learning with other machine learning framework to provide\nfeasible integrated problem solutions. We briefly present recent meta-learning\nadvances and discuss potential future research directions.', 'Many few-shot learning approaches have been designed under the meta-learning\nframework, which learns from a variety of learning tasks and generalizes to new\ntasks. These meta-learning approaches achieve the expected performance in the\nscenario where all samples are drawn from the same distributions (i.i.d.\nobservations). However, in real-world applications, few-shot learning paradigm\noften suffers from data shift, i.e., samples in different tasks, even in the\nsame task, could be drawn from various data distributions. Most existing\nfew-shot learning approaches are not designed with the consideration of data\nshift, and thus show downgraded performance when data distribution shifts.\nHowever, it is non-trivial to address the data shift problem in few-shot\nlearning, due to the limited number of labeled samples in each task. Targeting\nat addressing this problem, we propose a novel metric-based meta-learning\nframework to extract task-specific representations and task-shared\nrepresentations with the help of knowledge graph. The data shift within/between\ntasks can thus be combated by the combination of task-shared and task-specific\nrepresentations. The proposed model is evaluated on popular benchmarks and two\nconstructed new challenging datasets. The evaluation results demonstrate its\nremarkable performance.', 'Meta-learning has received a tremendous recent attention as a possible\napproach for mimicking human intelligence, i.e., acquiring new knowledge and\nskills with little or even no demonstration. Most of the existing meta-learning\nmethods are proposed to tackle few-shot learning problems such as image and\ntext, in rather Euclidean domain. However, there are very few works applying\nmeta-learning to non-Euclidean domains, and the recently proposed graph neural\nnetworks (GNNs) models do not perform effectively on graph few-shot learning\nproblems. Towards this, we propose a novel graph meta-learning framework --\nMeta-GNN -- to tackle the few-shot node classification problem in graph\nmeta-learning settings. It obtains the prior knowledge of classifiers by\ntraining on many similar few-shot learning tasks and then classifies the nodes\nfrom new classes with only few labeled samples. Additionally, Meta-GNN is a\ngeneral model that can be straightforwardly incorporated into any existing\nstate-of-the-art GNN. Our experiments conducted on three benchmark datasets\ndemonstrate that our proposed approach not only improves the node\nclassification performance by a large margin on few-shot learning problems in\nmeta-learning paradigm, but also learns a more general and flexible model for\ntask adaption.', 'Learning from demonstrations has made great progress over the past few years.\nHowever, it is generally data hungry and task specific. In other words, it\nrequires a large amount of data to train a decent model on a particular task,\nand the model often fails to generalize to new tasks that have a different\ndistribution. In practice, demonstrations from new tasks will be continuously\nobserved and the data might be unlabeled or only partially labeled. Therefore,\nit is desirable for the trained model to adapt to new tasks that have limited\ndata samples available. In this work, we build an adaptable imitation learning\nmodel based on the integration of Meta-learning and Adversarial Inverse\nReinforcement Learning (Meta-AIRL). We exploit the adversarial learning and\ninverse reinforcement learning mechanisms to learn policies and reward\nfunctions simultaneously from available training tasks and then adapt them to\nnew tasks with the meta-learning framework. Simulation results show that the\nadapted policy trained with Meta-AIRL can effectively learn from limited number\nof demonstrations, and quickly reach the performance comparable to that of the\nexperts on unseen tasks.', 'Graph Neural Networks (GNNs), a generalization of deep neural networks on\ngraph data have been widely used in various domains, ranging from drug\ndiscovery to recommender systems. However, GNNs on such applications are\nlimited when there are few available samples. Meta-learning has been an\nimportant framework to address the lack of samples in machine learning, and in\nrecent years, researchers have started to apply meta-learning to GNNs. In this\nwork, we provide a comprehensive survey of different meta-learning approaches\ninvolving GNNs on various graph problems showing the power of using these two\napproaches together. We categorize the literature based on proposed\narchitectures, shared representations, and applications. Finally, we discuss\nseveral exciting future research directions and open problems.', 'Graphs are widely used to model the relational structure of data, and the\nresearch of graph machine learning (ML) has a wide spectrum of applications\nranging from drug design in molecular graphs to friendship recommendation in\nsocial networks. Prevailing approaches for graph ML typically require abundant\nlabeled instances in achieving satisfactory results, which is commonly\ninfeasible in real-world scenarios since labeled data for newly emerged\nconcepts (e.g., new categorizations of nodes) on graphs is limited. Though\nmeta-learning has been applied to different few-shot graph learning problems,\nmost existing efforts predominately assume that all the data from those seen\nclasses is gold-labeled, while those methods may lose their efficacy when the\nseen data is weakly-labeled with severe label noise. As such, we aim to\ninvestigate a novel problem of weakly-supervised graph meta-learning for\nimproving the model robustness in terms of knowledge transfer. To achieve this\ngoal, we propose a new graph meta-learning framework -- Graph Hallucination\nNetworks (Meta-GHN) in this paper. Based on a new robustness-enhanced episodic\ntraining, Meta-GHN is meta-learned to hallucinate clean node representations\nfrom weakly-labeled data and extracts highly transferable meta-knowledge, which\nenables the model to quickly adapt to unseen tasks with few labeled instances.\nExtensive experiments demonstrate the superiority of Meta-GHN over existing\ngraph meta-learning studies on the task of weakly-supervised few-shot node\nclassification.', 'Deep neural networks are highly effective when a large number of labeled\nsamples are available but fail with few-shot classification tasks. Recently,\nmeta-learning methods have received much attention, which train a meta-learner\non massive additional tasks to gain the knowledge to instruct the few-shot\nclassification. Usually, the training tasks are randomly sampled and performed\nindiscriminately, often making the meta-learner stuck into a bad local optimum.\nSome works in the optimization of deep neural networks have shown that a better\narrangement of training data can make the classifier converge faster and\nperform better. Inspired by this idea, we propose an easy-to-hard expert\nmeta-training strategy to arrange the training tasks properly, where easy tasks\nare preferred in the first phase, then, hard tasks are emphasized in the second\nphase. A task hardness aware module is designed and integrated into the\ntraining procedure to estimate the hardness of a task based on the\ndistinguishability of its categories. In addition, we explore multiple hardness\nmeasurements including the semantic relation, the pairwise Euclidean distance,\nthe Hausdorff distance, and the Hilbert-Schmidt independence criterion.\nExperimental results on the miniImageNet and tieredImageNetSketch datasets show\nthat the meta-learners can obtain better results with our expert training\nstrategy.', 'Modern machine learning systems need to be able to cope with constantly\narriving and changing data. Two main areas of research dealing with such\nscenarios are continual learning and data stream mining. Continual learning\nfocuses on accumulating knowledge and avoiding forgetting, assuming information\nonce learned should be stored. Data stream mining focuses on adaptation to\nconcept drift and discarding outdated information, assuming that only the most\nrecent data is relevant. While these two areas are mainly being developed in\nseparation, they offer complementary views on the problem of learning from\ndynamic data. There is a need for unifying them, by offering architectures\ncapable of both learning and storing new information, as well as revisiting and\nadapting to changes in previously seen concepts. We propose a novel continual\nlearning approach that can handle both tasks. Our experience replay method is\nfueled by a centroid-driven memory storing diverse instances of incrementally\narriving classes. This is enhanced with a reactive subspace buffer that tracks\nconcept drift occurrences in previously seen classes and adapts clusters\naccordingly. The proposed architecture is thus capable of both remembering\nvalid and forgetting outdated information, offering a holistic framework for\ncontinual learning under concept drift.', 'The ability to learn new concepts with small amounts of data is a critical\naspect of intelligence that has proven challenging for deep learning methods.\nMeta-learning has emerged as a promising technique for leveraging data from\nprevious tasks to enable efficient learning of new tasks. However, most\nmeta-learning algorithms implicitly require that the meta-training tasks be\nmutually-exclusive, such that no single model can solve all of the tasks at\nonce. For example, when creating tasks for few-shot image classification, prior\nwork uses a per-task random assignment of image classes to N-way classification\nlabels. If this is not done, the meta-learner can ignore the task training data\nand learn a single model that performs all of the meta-training tasks\nzero-shot, but does not adapt effectively to new image classes. This\nrequirement means that the user must take great care in designing the tasks,\nfor example by shuffling labels or removing task identifying information from\nthe inputs. In some domains, this makes meta-learning entirely inapplicable. In\nthis paper, we address this challenge by designing a meta-regularization\nobjective using information theory that places precedence on data-driven\nadaptation. This causes the meta-learner to decide what must be learned from\nthe task training data and what should be inferred from the task testing input.\nBy doing so, our algorithm can successfully use data from\nnon-mutually-exclusive tasks to efficiently adapt to novel tasks. We\ndemonstrate its applicability to both contextual and gradient-based\nmeta-learning algorithms, and apply it in practical settings where applying\nstandard meta-learning has been difficult. Our approach substantially\noutperforms standard meta-learning algorithms in these settings.', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', 'Few-shot learning aims to adapt knowledge learned from previous tasks to\nnovel tasks with only a limited amount of labeled data. Research literature on\nfew-shot learning exhibits great diversity, while different algorithms often\nexcel at different few-shot learning scenarios. It is therefore tricky to\ndecide which learning strategies to use under different task conditions.\nInspired by the recent success in Automated Machine Learning literature\n(AutoML), in this paper, we present Meta Navigator, a framework that attempts\nto solve the aforementioned limitation in few-shot learning by seeking a\nhigher-level strategy and proffer to automate the selection from various\nfew-shot learning designs. The goal of our work is to search for good parameter\nadaptation policies that are applied to different stages in the network for\nfew-shot classification. We present a search space that covers many popular\nfew-shot learning algorithms in the literature and develop a differentiable\nsearching and decoding algorithm based on meta-learning that supports\ngradient-based optimization. We demonstrate the effectiveness of our\nsearching-based method on multiple benchmark datasets. Extensive experiments\nshow that our approach significantly outperforms baselines and demonstrates\nperformance advantages over many state-of-the-art methods. Code and models will\nbe made publicly available.', 'Large-scale pretrained language models have led to dramatic improvements in\ntext generation. Impressive performance can be achieved by finetuning only on a\nsmall number of instances (few-shot setting). Nonetheless, almost all previous\nwork simply applies random sampling to select the few-shot training instances.\nLittle to no attention has been paid to the selection strategies and how they\nwould affect model performance. In this work, we present a study on training\ninstance selection in few-shot neural text generation. The selection decision\nis made based only on the unlabeled data so as to identify the most worthwhile\ndata points that should be annotated under some budget of labeling cost. Based\non the intuition that the few-shot training instances should be diverse and\nrepresentative of the entire data distribution, we propose a simple selection\nstrategy with K-means clustering. We show that even with the naive\nclustering-based approach, the generation models consistently outperform random\nsampling on three text generation tasks: data-to-text generation, document\nsummarization and question generation. We hope that this work will call for\nmore attention on this largely unexplored area.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', ""Document screening is a central task within Evidenced Based Medicine, which\nis a clinical discipline that supplements scientific proof to back medical\ndecisions. Given the recent advances in DL (Deep Learning) methods applied to\nInformation Retrieval tasks, I propose a DL document classification approach\nwith BERT or PubMedBERT embeddings and a DL similarity search path using SBERT\nembeddings to reduce physicians' tasks of screening and classifying immense\namounts of documents to answer clinical queries. I test and evaluate the\nretrieval effectiveness of my DL strategy on the 2017 and 2018 CLEF eHealth\ncollections. I find that the proposed DL strategy works, I compare it to the\nrecently successful BM25 plus RM3 model, and conclude that the suggested method\naccomplishes advanced retrieval performance in the initial ranking of the\narticles with the aforementioned datasets, for the CLEF eHealth Technologically\nAssisted Reviews in Empirical Medicine Task."", 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'This doctoral thesis covers some of my advances in electron microscopy with\ndeep learning. Highlights include a comprehensive review of deep learning in\nelectron microscopy; large new electron microscopy datasets for machine\nlearning, dataset search engines based on variational autoencoders, and\nautomatic data clustering by t-distributed stochastic neighbour embedding;\nadaptive learning rate clipping to stabilize learning; generative adversarial\nnetworks for compressed sensing with spiral, uniformly spaced and other fixed\nsparse scan paths; recurrent neural networks trained to piecewise adapt sparse\nscan paths to specimens by reinforcement learning; improving signal-to-noise;\nand conditional generative adversarial networks for exit wavefunction\nreconstruction from single transmission electron micrographs. This thesis adds\nto my publications by presenting their relationships, reflections, and holistic\nconclusions. This version of my thesis is typeset for online dissemination to\nimprove readability, whereas the thesis submitted to the University of Warwick\nin support of my application for the degree of Doctor of Philosophy in Physics\nis typeset for physical printing and binding.', 'This paper and accompanying Python and C++ Framework is the product of the\nauthors perceived problems with narrow (Discrimination based) AI. (Artificial\nIntelligence) The Framework attempts to develop a genetic transfer of\nexperience through potential structural expressions using a common\nregulation/exchange value (energy) to create a model whereby neural\narchitecture and all unit processes are co-dependently developed by genetic and\nreal time signal processing influences; successful routes are defined by\nstability of the spike distribution per epoch which is influenced by\ngenetically encoded morphological development biases.These principles are aimed\ntowards creating a diverse and robust network that is capable of adapting to\ngeneral tasks by training within a simulation designed for transfer learning to\nother mediums at scale.', 'Few-Shot Learning is the challenge of training a model with only a small\namount of data. Many solutions to this problem use meta-learning algorithms,\ni.e. algorithms that learn to learn. By sampling few-shot tasks from a larger\ndataset, we can teach these algorithms to solve new, unseen tasks. This\ndocument reports my work on meta-learning algorithms for Few-Shot Computer\nVision. This work was done during my internship at Sicara, a French company\nbuilding image recognition solutions for businesses. It contains: 1. an\nextensive review of the state-of-the-art in few-shot computer vision; 2. a\nbenchmark of meta-learning algorithms for few-shot image classification; 3. the\nintroduction to a novel meta-learning algorithm for few-shot object detection,\nwhich is still in development.', 'We consider the task of few shot link prediction on graphs. The goal is to\nlearn from a distribution over graphs so that a model is able to quickly infer\nmissing edges in a new graph after a small amount of training. We show that\ncurrent link prediction methods are generally ill-equipped to handle this task.\nThey cannot effectively transfer learned knowledge from one graph to another\nand are unable to effectively learn from sparse samples of edges. To address\nthis challenge, we introduce a new gradient-based meta learning framework,\nMeta-Graph. Our framework leverages higher-order gradients along with a learned\ngraph signature function that conditionally generates a graph neural network\ninitialization. Using a novel set of few shot link prediction benchmarks, we\nshow that Meta-Graph can learn to quickly adapt to a new graph using only a\nsmall sample of true edges, enabling not only fast adaptation but also improved\nresults at convergence.', ""Graph Neural Networks (GNNs) are a powerful tool for machine learning on\ngraphs.GNNs combine node feature information with the graph structure by\nrecursively passing neural messages along edges of the input graph. However,\nincorporating both graph structure and feature information leads to complex\nmodels, and explaining predictions made by GNNs remains unsolved. Here we\npropose GNNExplainer, the first general, model-agnostic approach for providing\ninterpretable explanations for predictions of any GNN-based model on any\ngraph-based machine learning task. Given an instance, GNNExplainer identifies a\ncompact subgraph structure and a small subset of node features that have a\ncrucial role in GNN's prediction. Further, GNNExplainer can generate consistent\nand concise explanations for an entire class of instances. We formulate\nGNNExplainer as an optimization task that maximizes the mutual information\nbetween a GNN's prediction and distribution of possible subgraph structures.\nExperiments on synthetic and real-world graphs show that our approach can\nidentify important graph structures as well as node features, and outperforms\nbaselines by 17.1% on average. GNNExplainer provides a variety of benefits,\nfrom the ability to visualize semantically relevant structures to\ninterpretability, to giving insights into errors of faulty GNNs."", 'Many existing deep learning models are vulnerable to adversarial examples\nthat are imperceptible to humans. To address this issue, various methods have\nbeen proposed to design network architectures that are robust to one particular\ntype of adversarial attacks. It is practically impossible, however, to predict\nbeforehand which type of attacks a machine learn model may suffer from. To\naddress this challenge, we propose to search for deep neural architectures that\nare robust to five types of well-known adversarial attacks using a\nmulti-objective evolutionary algorithm. To reduce the computational cost, a\nnormalized error rate of a randomly chosen attack is calculated as the\nrobustness for each newly generated neural architecture at each generation. All\nnon-dominated network architectures obtained by the proposed method are then\nfully trained against randomly chosen adversarial attacks and tested on two\nwidely used datasets. Our experimental results demonstrate the superiority of\noptimized neural architectures found by the proposed approach over\nstate-of-the-art networks that are widely used in the literature in terms of\nthe classification accuracy under different adversarial attacks.', 'The ability to learn and adapt in real time is a central feature of\nbiological systems. Neuromorphic architectures demonstrating such versatility\ncan greatly enhance our ability to efficiently process information at the edge.\nA key challenge, however, is to understand which learning rules are best suited\nfor specific tasks and how the relevant hyperparameters can be fine-tuned. In\nthis work, we introduce a conceptual framework in which the learning process is\nintegrated into the network itself. This allows us to cast meta-learning as a\nmathematical optimization problem. We employ DeepHyper, a scalable,\nasynchronous model-based search, to simultaneously optimize the choice of\nmeta-learning rules and their hyperparameters. We demonstrate our approach with\ntwo different datasets, MNIST and FashionMNIST, using a network architecture\ninspired by the learning center of the insect brain. Our results show that\noptimal learning rules can be dataset-dependent even within similar tasks. This\ndependency demonstrates the importance of introducing versatility and\nflexibility in the learning algorithms. It also illuminates experimental\nfindings in insect neuroscience that have shown a heterogeneity of learning\nrules within the insect mushroom body.', 'Adversarial attacks can affect the performance of existing deep learning\nmodels. With the increased interest in graph based machine learning techniques,\nthere have been investigations which suggest that these models are also\nvulnerable to attacks. In particular, corruptions of the graph topology can\ndegrade the performance of graph based learning algorithms severely. This is\ndue to the fact that the prediction capability of these algorithms relies\nmostly on the similarity structure imposed by the graph connectivity.\nTherefore, detecting the location of the corruption and correcting the induced\nerrors becomes crucial. There has been some recent work which tackles the\ndetection problem, however these methods do not address the effect of the\nattack on the downstream learning task. In this work, we propose an algorithm\nthat uses node copying to mitigate the degradation in classification that is\ncaused by adversarial attacks. The proposed methodology is applied only after\nthe model for the downstream task is trained and the added computation cost\nscales well for large graphs. Experimental results show the effectiveness of\nour approach for several real world datasets.']","['How to leverage cross-document interactions to improve ranking performance is\nan important topic in information retrieval (IR) research. However, this topic\nhas not been well-studied in the learning-to-rank setting and most of the\nexisting work still treats each document independently while scoring. The\nrecent development of deep learning shows strength in modeling complex\nrelationships across sequences and sets. It thus motivates us to study how to\nleverage cross-document interactions for learning-to-rank in the deep learning\nframework. In this paper, we formally define the permutation-equivariance\nrequirement for a scoring function that captures cross-document interactions.\nWe then propose a self-attention based document interaction network and show\nthat it satisfies the permutation-equivariant requirement, and can generate\nscores for document sets of varying sizes. Our proposed methods can\nautomatically learn to capture document interactions without any auxiliary\ninformation, and can scale across large document sets. We conduct experiments\non three ranking datasets: the benchmark Web30k, a Gmail search, and a Google\nDrive Quick Access dataset. Experimental results show that our proposed methods\nare both more effective and efficient than baselines.', 'Recent grid-based document representations like BERTgrid allow the\nsimultaneous encoding of the textual and layout information of a document in a\n2D feature map so that state-of-the-art image segmentation and/or object\ndetection models can be straightforwardly leveraged to extract key information\nfrom documents. However, such methods have not achieved comparable performance\nto state-of-the-art sequence- and graph-based methods such as LayoutLM and PICK\nyet. In this paper, we propose a new multi-modal backbone network by\nconcatenating a BERTgrid to an intermediate layer of a CNN model, where the\ninput of CNN is a document image and the BERTgrid is a grid of word embeddings,\nto generate a more powerful grid-based document representation, named\nViBERTgrid. Unlike BERTgrid, the parameters of BERT and CNN in our multimodal\nbackbone network are trained jointly. Our experimental results demonstrate that\nthis joint training strategy improves significantly the representation ability\nof ViBERTgrid. Consequently, our ViBERTgrid-based key information extraction\napproach has achieved state-of-the-art performance on real-world datasets.', ""Automatic image cropping techniques are commonly used to enhance the\naesthetic quality of an image; they do it by detecting the most beautiful or\nthe most salient parts of the image and removing the unwanted content to have a\nsmaller image that is more visually pleasing. In this thesis, I introduce an\nadditional dimension to the problem of cropping, semantics. I argue that image\ncropping can also enhance the image's relevancy for a given entity by using the\nsemantic information contained in the image. I call this problem, Semantic\nImage Cropping. To support my argument, I provide a new dataset containing 100\nimages with at least two different entities per image and four ground truth\ncroppings collected using Amazon Mechanical Turk. I use this dataset to show\nthat state-of-the-art cropping algorithms that only take into account\naesthetics do not perform well in the problem of semantic image cropping.\nAdditionally, I provide a new deep learning system that takes not just\naesthetics but also semantics into account to generate image croppings, and I\nevaluate its performance using my new semantic cropping dataset, showing that\nusing the semantic information of an image can help to produce better\ncroppings."", 'Deep learning models on graphs have achieved remarkable performance in\nvarious graph analysis tasks, e.g., node classification, link prediction, and\ngraph clustering. However, they expose uncertainty and unreliability against\nthe well-designed inputs, i.e., adversarial examples. Accordingly, a line of\nstudies has emerged for both attack and defense addressed in different graph\nanalysis tasks, leading to the arms race in graph adversarial learning. Despite\nthe booming works, there still lacks a unified problem definition and a\ncomprehensive review. To bridge this gap, we investigate and summarize the\nexisting works on graph adversarial learning tasks systemically. Specifically,\nwe survey and unify the existing works w.r.t. attack and defense in graph\nanalysis tasks, and give appropriate definitions and taxonomies at the same\ntime. Besides, we emphasize the importance of related evaluation metrics,\ninvestigate and summarize them comprehensively. Hopefully, our works can\nprovide a comprehensive overview and offer insights for the relevant\nresearchers. Latest advances in graph adversarial learning are summarized in\nour GitHub repository\nhttps://github.com/EdisonLeeeee/Graph-Adversarial-Learning.', 'Alongside the well-publicized accomplishments of deep neural networks there\nhas emerged an apparent bug in their success on tasks such as object\nrecognition: with deep models trained using vanilla methods, input images can\nbe slightly corrupted in order to modify output predictions, even when these\ncorruptions are practically invisible. This apparent lack of robustness has led\nresearchers to propose methods that can help to prevent an adversary from\nhaving such capabilities. The state-of-the-art approaches have incorporated the\nrobustness requirement into the loss function, and the training process\ninvolves taking stochastic gradient descent steps not using original inputs but\non adversarially-corrupted ones. In this paper we propose a multiclass boosting\nframework to ensure adversarial robustness. Boosting algorithms are generally\nwell-suited for adversarial scenarios, as they were classically designed to\nsatisfy a minimax guarantee. We provide a theoretical foundation for this\nmethodology and describe conditions under which robustness can be achieved\ngiven a weak training oracle. We show empirically that adversarially-robust\nmulticlass boosting not only outperforms the state-of-the-art methods, it does\nso at a fraction of the training time.', 'Graph deep learning models, such as graph convolutional networks (GCN)\nachieve remarkable performance for tasks on graph data. Similar to other types\nof deep models, graph deep learning models often suffer from adversarial\nattacks. However, compared with non-graph data, the discrete features, graph\nconnections and different definitions of imperceptible perturbations bring\nunique challenges and opportunities for the adversarial attacks and defenses\nfor graph data. In this paper, we propose both attack and defense techniques.\nFor attack, we show that the discreteness problem could easily be resolved by\nintroducing integrated gradients which could accurately reflect the effect of\nperturbing certain features or edges while still benefiting from the parallel\ncomputations. For defense, we observe that the adversarially manipulated graph\nfor the targeted attack differs from normal graphs statistically. Based on this\nobservation, we propose a defense approach which inspects the graph and\nrecovers the potential adversarial perturbations. Our experiments on a number\nof datasets show the effectiveness of the proposed methods.', 'The encoder-decoder framework achieves state-of-the-art results in keyphrase\ngeneration (KG) tasks by predicting both present keyphrases that appear in the\nsource document and absent keyphrases that do not. However, relying solely on\nthe source document can result in generating uncontrollable and inaccurate\nabsent keyphrases. To address these problems, we propose a novel graph-based\nmethod that can capture explicit knowledge from related references. Our model\nfirst retrieves some document-keyphrases pairs similar to the source document\nfrom a pre-defined index as references. Then a heterogeneous graph is\nconstructed to capture relationships of different granularities between the\nsource document and its references. To guide the decoding process, a\nhierarchical attention and copy mechanism is introduced, which directly copies\nappropriate words from both the source document and its references based on\ntheir relevance and significance. The experimental results on multiple KG\nbenchmarks show that the proposed model achieves significant improvements\nagainst other baseline models, especially with regard to the absent keyphrase\nprediction.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'Document-level relation extraction aims to discover relations between\nentities across a whole document. How to build the dependency of entities from\ndifferent sentences in a document remains to be a great challenge. Current\napproaches either leverage syntactic trees to construct document-level graphs\nor aggregate inference information from different sentences. In this paper, we\nbuild cross-sentence dependencies by inferring compositional relations between\ninter-sentence mentions. Adopting aggressive linking strategy, intermediate\nrelations are reasoned on the document-level graphs by mention convolution. We\nfurther notice the generalization problem of NA instances, which is caused by\nincomplete annotation and worsened by fully-connected mention pairs. An\nimproved ranking loss is proposed to attend this problem. Experiments show the\nconnections between different mentions are crucial to document-level relation\nextraction, which enables the model to extract more meaningful higher-level\ncompositional relations.', 'In this paper, we focus on learning structure-aware document representations\nfrom data without recourse to a discourse parser or additional annotations.\nDrawing inspiration from recent efforts to empower neural networks with a\nstructural bias, we propose a model that can encode a document while\nautomatically inducing rich structural dependencies. Specifically, we embed a\ndifferentiable non-projective parsing algorithm into a neural model and use\nattention mechanisms to incorporate the structural biases. Experimental\nevaluation across different tasks and datasets shows that the proposed model\nachieves state-of-the-art results on document modeling tasks while inducing\nintermediate structures which are both interpretable and meaningful.', ""An accurate abstractive summary of a document should contain all its salient\ninformation and should be logically entailed by the input document. We improve\nthese important aspects of abstractive summarization via multi-task learning\nwith the auxiliary tasks of question generation and entailment generation,\nwhere the former teaches the summarization model how to look for salient\nquestioning-worthy details, and the latter teaches the model how to rewrite a\nsummary which is a directed-logical subset of the input document. We also\npropose novel multi-task architectures with high-level (semantic)\nlayer-specific sharing across multiple encoder and decoder layers of the three\ntasks, as well as soft-sharing mechanisms (and show performance ablations and\nanalysis examples of each contribution). Overall, we achieve statistically\nsignificant improvements over the state-of-the-art on both the CNN/DailyMail\nand Gigaword datasets, as well as on the DUC-2002 transfer setup. We also\npresent several quantitative and qualitative analysis studies of our model's\nlearned saliency and entailment skills."", 'Marrying topic models and language models exposes language understanding to a\nbroader source of document-level context beyond sentences via topics. While\nintroducing topical semantics in language models, existing approaches\nincorporate latent document topic proportions and ignore topical discourse in\nsentences of the document. This work extends the line of research by\nadditionally introducing an explainable topic representation in language\nunderstanding, obtained from a set of key terms correspondingly for each latent\ntopic of the proportion. Moreover, we retain sentence-topic associations along\nwith document-topic association by modeling topical discourse for every\nsentence in the document. We present a novel neural composite language model\nthat exploits both the latent and explainable topics along with topical\ndiscourse at sentence-level in a joint learning framework of topic and language\nmodels. Experiments over a range of tasks such as language modeling, word sense\ndisambiguation, document classification, retrieval and text generation\ndemonstrate ability of the proposed model in improving language understanding.', 'Forum discussion data differ in both structure and properties from generic\nform of textual data such as news. Henceforth, summarization techniques should,\nin turn, make use of such differences, and craft models that can benefit from\nthe structural nature of discussion data. In this work, we propose attending to\nthe beginning of a document, to improve the performance of extractive\nsummarization models when applied to forum discussion data. Evaluations\ndemonstrated that with the help of bidirectional attention mechanism, attending\nto the beginning of a document (initial comment/post) in a discussion thread,\ncan introduce a consistent boost in ROUGE scores, as well as introducing a new\nState Of The Art (SOTA) ROUGE scores on the forum discussions dataset.\nAdditionally, we explored whether this hypothesis is extendable to other\ngeneric forms of textual data. We make use of the tendency of introducing\nimportant information early in the text, by attending to the first few\nsentences in generic textual data. Evaluations demonstrated that attending to\nintroductory sentences using bidirectional attention, improves the performance\nof extractive summarization models when even applied to more generic form of\ntextual data.', 'Image-to-image translation, which translates input images to a different\ndomain with a learned one-to-one mapping, has achieved impressive success in\nrecent years. The success of translation mainly relies on the network\narchitecture to reserve the structural information while modify the appearance\nslightly at the pixel level through adversarial training. Although these\nnetworks are able to learn the mapping, the translated images are predictable\nwithout exclusion. It is more desirable to diversify them using image-to-image\ntranslation by introducing uncertainties, i.e., the generated images hold\npotential for variations in colors and textures in addition to the general\nsimilarity to the input images, and this happens in both the target and source\ndomains. To this end, we propose a novel generative adversarial network (GAN)\nbased model, InjectionGAN, to learn a many-to-many mapping. In this model, the\ninput image is combined with latent variables, which comprise of\ndomain-specific attribute and unspecific random variations. The domain-specific\nattribute indicates the target domain of the translation, while the unspecific\nrandom variations introduce uncertainty into the model. A unified framework is\nproposed to regroup these two parts and obtain diverse generations in each\ndomain. Extensive experiments demonstrate that the diverse generations have\nhigh quality for the challenging image-to-image translation tasks where no\npairing information of the training dataset exits. Both quantitative and\nqualitative results prove the superior performance of InjectionGAN over the\nstate-of-the-art approaches.', 'The field of meta-learning has seen a dramatic rise in interest in recent\nyears. In existing meta-learning approaches, learning tasks for training\nmeta-models are usually collected from public datasets, which brings the\ndifficulty of obtaining a sufficient number of meta-learning tasks with a large\namount of training data. In this paper, we propose a meta-learning approach\nbased on randomly generated meta-learning tasks to obtain a parametric loss for\nclassification learning based on big data. The loss is represented by a deep\nneural network, called meta-loss network (MLN). To train the MLN, we construct\na large number of classification learning tasks through randomly generating\ntraining data, validation data, and corresponding ground-truth linear\nclassifier. Our approach has two advantages. First, sufficient meta-learning\ntasks with large number of training data can be obtained easily. Second, the\nground-truth classifier is given, so that the difference between the learned\nclassifier and the ground-truth model can be measured to reflect the\nperformance of MLN more precisely than validation accuracy. Based on this\ndifference, we apply the evolutionary strategy algorithm to find out the\noptimal MLN. The resultant MLN not only leads to satisfactory learning effects\non generated linear classifier learning tasks for testing, but also behaves\nvery well on generated nonlinear classifier learning tasks and various public\nclassification tasks. Our MLN stably surpass cross-entropy (CE) and mean square\nerror (MSE) in testing accuracy and generalization ability. These results\nillustrate the possibility of achieving satisfactory meta-learning effects\nusing generated learning tasks.', 'The recent explosion of machine learning (ML) and artificial intelligence\n(AI) shows great potential in the breakthrough of metal additive manufacturing\n(AM) process modeling. However, the success of conventional machine learning\ntools in data science is primarily attributed to the unprecedented large amount\nof labeled data-sets (big data), which can be either obtained by experiments or\nfirst-principle simulations. Unfortunately, these labeled data-sets are\nexpensive to obtain in AM due to the high expense of the AM experiments and\nprohibitive computational cost of high-fidelity simulations.\n  We propose a physics-informed neural network (PINN) framework that fuses both\ndata and first physical principles, including conservation laws of momentum,\nmass, and energy, into the neural network to inform the learning processes. To\nthe best knowledge of the authors, this is the first application of PINN to\nthree dimensional AM processes modeling. Besides, we propose a hard-type\napproach for Dirichlet boundary conditions (BCs) based on a Heaviside function,\nwhich can not only enforce the BCs but also accelerate the learning process.\nThe PINN framework is applied to two representative metal manufacturing\nproblems, including the 2018 NIST AM-Benchmark test series. We carefully assess\nthe performance of the PINN model by comparing the predictions with available\nexperimental data and high-fidelity simulation results. The investigations show\nthat the PINN, owed to the additional physical knowledge, can accurately\npredict the temperature and melt pool dynamics during metal AM processes with\nonly a moderate amount of labeled data-sets. The foray of PINN to metal AM\nshows the great potential of physics-informed deep learning for broader\napplications to advanced manufacturing.', 'We introduce MARGE, a pre-trained sequence-to-sequence model learned with an\nunsupervised multi-lingual multi-document paraphrasing objective. MARGE\nprovides an alternative to the dominant masked language modeling paradigm,\nwhere we self-supervise the reconstruction of target text by retrieving a set\nof related texts (in many languages) and conditioning on them to maximize the\nlikelihood of generating the original. We show it is possible to jointly learn\nto do retrieval and reconstruction, given only a random initialization. The\nobjective noisily captures aspects of paraphrase, translation, multi-document\nsummarization, and information retrieval, allowing for strong zero-shot\nperformance on several tasks. For example, with no additional task-specific\ntraining we achieve BLEU scores of up to 35.8 for document translation. We\nfurther show that fine-tuning gives strong performance on a range of\ndiscriminative and generative tasks in many languages, making MARGE the most\ngenerally applicable pre-training method to date.', 'Although machine learning models typically experience a drop in performance\non out-of-distribution data, accuracies on in- versus out-of-distribution data\nare widely observed to follow a single linear trend when evaluated across a\ntestbed of models. Models that are more accurate on the out-of-distribution\ndata relative to this baseline exhibit ""effective robustness"" and are\nexceedingly rare. Identifying such models, and understanding their properties,\nis key to improving out-of-distribution performance. We conduct a thorough\nempirical investigation of effective robustness during fine-tuning and\nsurprisingly find that models pre-trained on larger datasets exhibit effective\nrobustness during training that vanishes at convergence. We study how\nproperties of the data influence effective robustness, and we show that it\nincreases with the larger size, more diversity, and higher example difficulty\nof the dataset. We also find that models that display effective robustness are\nable to correctly classify 10% of the examples that no other current testbed\nmodel gets correct. Finally, we discuss several strategies for scaling\neffective robustness to the high-accuracy regime to improve the\nout-of-distribution accuracy of state-of-the-art models.', 'Many important NLP problems can be posed as dual-sequence or\nsequence-to-sequence modeling tasks. Recent advances in building end-to-end\nneural architectures have been highly successful in solving such tasks. In this\nwork we propose a new architecture for dual-sequence modeling that is based on\nassociative memory. We derive AM-RNNs, a recurrent associative memory (AM)\nwhich augments generic recurrent neural networks (RNN). This architecture is\nextended to the Dual AM-RNN which operates on two AMs at once. Our models\nachieve very competitive results on textual entailment. A qualitative analysis\ndemonstrates that long range dependencies between source and target-sequence\ncan be bridged effectively using Dual AM-RNNs. However, an initial experiment\non auto-encoding reveals that these benefits are not exploited by the system\nwhen learning to solve sequence-to-sequence tasks which indicates that\nadditional supervision or regularization is needed.', ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'Despite significant progress on current state-of-the-art image generation\nmodels, synthesis of document images containing multiple and complex object\nlayouts is a challenging task. This paper presents a novel approach, called\nDocSynth, to automatically synthesize document images based on a given layout.\nIn this work, given a spatial layout (bounding boxes with object categories) as\na reference by the user, our proposed DocSynth model learns to generate a set\nof realistic document images consistent with the defined layout. Also, this\nframework has been adapted to this work as a superior baseline model for\ncreating synthetic document image datasets for augmenting real data during\ntraining for document layout analysis tasks. Different sets of learning\nobjectives have been also used to improve the model performance.\nQuantitatively, we also compare the generated results of our model with real\ndata using standard evaluation metrics. The results highlight that our model\ncan successfully generate realistic and diverse document images with multiple\nobjects. We also present a comprehensive qualitative analysis summary of the\ndifferent scopes of synthetic image generation tasks. Lastly, to our knowledge\nthis is the first work of its kind.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'In this paper we address the problem of modeling relational data, which\nappear in many applications such as social network analysis, recommender\nsystems and bioinformatics. Previous studies either consider latent feature\nbased models but disregarding local structure in the network, or focus\nexclusively on capturing local structure of objects based on latent blockmodels\nwithout coupling with latent characteristics of objects. To combine the\nbenefits of the previous work, we propose a novel model that can simultaneously\nincorporate the effect of latent features and covariates if any, as well as the\neffect of latent structure that may exist in the data. To achieve this, we\nmodel the relation graph as a function of both latent feature factors and\nlatent cluster memberships of objects to collectively discover globally\npredictive intrinsic properties of objects and capture latent block structure\nin the network to improve prediction performance. We also develop an\noptimization transfer algorithm based on the generalized EM-style strategy to\nlearn the latent factors. We prove the efficacy of our proposed model through\nthe link prediction task and cluster analysis task, and extensive experiments\non the synthetic data and several real world datasets suggest that our proposed\nLFBM model outperforms the other state of the art approaches in the evaluated\ntasks.', 'Deep Learning has revolutionized machine learning and artificial\nintelligence, achieving superhuman performance in several standard benchmarks.\nIt is well-known that deep learning models are inefficient to train; they learn\nby processing millions of training data multiple times and require powerful\ncomputational resources to process large batches of data in parallel at the\nsame time rather than sequentially. Deep learning models also have unexpected\nfailure modes; they can be fooled into misbehaviour, producing unexpectedly\nincorrect predictions.\n  In this thesis, we study approaches to improve the training efficiency and\nrobustness of deep learning models. In the context of learning visual-semantic\nembeddings, we find that prioritizing learning on more informative training\ndata increases convergence speed and improves generalization performance on\ntest data. We formalize a simple trick called hard negative mining as a\nmodification to the learning objective function with no computational overhead.\nNext, we seek improvements to optimization speed in general-purpose\noptimization methods in deep learning. We show that a redundancy-aware\nmodification to the sampling of training data improves the training speed and\ndevelops an efficient method for detecting the diversity of training signal,\nnamely, gradient clustering. Finally, we study adversarial robustness in deep\nlearning and approaches to achieve maximal adversarial robustness without\ntraining with additional data. For linear models, we prove guaranteed maximal\nrobustness achieved only by appropriate choice of the optimizer,\nregularization, or architecture.', 'There is an explosion of data, documents, and other content, and people\nrequire tools to analyze and interpret these, tools to turn the content into\ninformation and knowledge. Topic modeling have been developed to solve these\nproblems. Topic models such as LDA [Blei et. al. 2003] allow salient patterns\nin data to be extracted automatically. When analyzing texts, these patterns are\ncalled topics. Among numerous extensions of LDA, few of them can reliably\nanalyze multiple groups of documents and extract topic similarities. Recently,\nthe introduction of differential topic modeling (SPDP) [Chen et. al. 2012]\nperforms uniformly better than many topic models in a discriminative setting.\n  There is also a need to improve the sampling speed for topic models. While\nsome effort has been made for distributed algorithms, there is no work\ncurrently done using graphical processing units (GPU). Note the GPU framework\nhas already become the most cost-efficient platform for many problems.\n  In this thesis, I propose and implement a scalable multi-GPU distributed\nparallel framework which approximates SPDP. Through experiments, I have shown\nmy algorithms have a gain in speed of about 50 times while being almost as\naccurate, with only one single cheap laptop GPU. Furthermore, I have shown the\nspeed improvement is sublinearly scalable when multiple GPUs are used, while\nfairly maintaining the accuracy. Therefore on a medium-sized GPU cluster, the\nspeed improvement could potentially reach a factor of a thousand.\n  Note SPDP is just a representative of other extensions of LDA. Although my\nalgorithm is implemented to work with SPDP, it is designed to be a general\nenough to work with other topic models. The speed-up on smaller collections\n(i.e., 1000s of documents), means that these more complex LDA extensions could\nnow be done in real-time, thus opening up a new way of using these LDA models\nin industry.', 'We propose a Label Propagation based algorithm for weakly supervised text\nclassification. We construct a graph where each document is represented by a\nnode and edge weights represent similarities among the documents. Additionally,\nwe discover underlying topics using Latent Dirichlet Allocation (LDA) and\nenrich the document graph by including the topics in the form of additional\nnodes. The edge weights between a topic and a text document represent level of\n""affinity"" between them. Our approach does not require document level\nlabelling, instead it expects manual labels only for topic nodes. This\nsignificantly minimizes the level of supervision needed as only a few topics\nare observed to be enough for achieving sufficiently high accuracy. The Label\nPropagation Algorithm is employed on this enriched graph to propagate labels\namong the nodes. Our approach combines the advantages of Label Propagation\n(through document-document similarities) and Topic Modelling (for minimal but\nsmart supervision). We demonstrate the effectiveness of our approach on various\ndatasets and compare with state-of-the-art weakly supervised text\nclassification approaches.', 'Capturing the compositional process which maps the meaning of words to that\nof documents is a central challenge for researchers in Natural Language\nProcessing and Information Retrieval. We introduce a model that is able to\nrepresent the meaning of documents by embedding them in a low dimensional\nvector space, while preserving distinctions of word and sentence order crucial\nfor capturing nuanced semantics. Our model is based on an extended Dynamic\nConvolution Neural Network, which learns convolution filters at both the\nsentence and document level, hierarchically learning to capture and compose low\nlevel lexical features into high level semantic concepts. We demonstrate the\neffectiveness of this model on a range of document modelling tasks, achieving\nstrong results with no feature engineering and with a more compact model.\nInspired by recent advances in visualising deep convolution networks for\ncomputer vision, we present a novel visualisation technique for our document\nnetworks which not only provides insight into their learning process, but also\ncan be interpreted to produce a compelling automatic summarisation system for\ntexts.', 'Robustness of neural networks has recently attracted a great amount of\ninterest. The many investigations in this area lack a precise common foundation\nof robustness concepts. Therefore, in this paper, we propose a rigorous and\nflexible framework for defining different types of robustness properties for\nclassifiers. Our robustness concept is based on postulates that robustness of a\nclassifier should be considered as a property that is independent of accuracy,\nand that it should be defined in purely mathematical terms without reliance on\nalgorithmic procedures for its measurement. We develop a very general\nrobustness framework that is applicable to any type of classification model,\nand that encompasses relevant robustness concepts for investigations ranging\nfrom safety against adversarial attacks to transferability of models to new\ndomains. For two prototypical, distinct robustness objectives we then propose\nnew learning approaches based on neural network co-training strategies for\nobtaining image classifiers optimized for these respective objectives.', 'We present two categories of model-agnostic adversarial strategies that\nreveal the weaknesses of several generative, task-oriented dialogue models:\nShould-Not-Change strategies that evaluate over-sensitivity to small and\nsemantics-preserving edits, as well as Should-Change strategies that test if a\nmodel is over-stable against subtle yet semantics-changing modifications. We\nnext perform adversarial training with each strategy, employing a max-margin\napproach for negative generative examples. This not only makes the target\ndialogue model more robust to the adversarial inputs, but also helps it perform\nsignificantly better on the original inputs. Moreover, training on all\nstrategies combined achieves further improvements, achieving a new\nstate-of-the-art performance on the original task (also verified via human\nevaluation). In addition to adversarial training, we also address the\nrobustness task at the model-level, by feeding it subword units as both inputs\nand outputs, and show that the resulting model is equally competitive, requires\nonly 1/4 of the original vocabulary size, and is robust to one of the\nadversarial strategies (to which the original model is vulnerable) even without\nadversarial training.', ""Few-shot classification refers to learning a classifier for new classes given\nonly a few examples. While a plethora of models have emerged to tackle it, we\nfind the procedure and datasets that are used to assess their progress lacking.\nTo address this limitation, we propose Meta-Dataset: a new benchmark for\ntraining and evaluating models that is large-scale, consists of diverse\ndatasets, and presents more realistic tasks. We experiment with popular\nbaselines and meta-learners on Meta-Dataset, along with a competitive method\nthat we propose. We analyze performance as a function of various\ncharacteristics of test tasks and examine the models' ability to leverage\ndiverse training sources for improving their generalization. We also propose a\nnew set of baselines for quantifying the benefit of meta-learning in\nMeta-Dataset. Our extensive experimentation has uncovered important research\nchallenges and we hope to inspire work in these directions."", 'Recent breakthroughs in the field of deep learning have led to advancements\nin a broad spectrum of tasks in computer vision, audio processing, natural\nlanguage processing and other areas. In most instances where these tasks are\ndeployed in real-world scenarios, the models used in them have been shown to be\nsusceptible to adversarial attacks, making it imperative for us to address the\nchallenge of their adversarial robustness. Existing techniques for adversarial\nrobustness fall into three broad categories: defensive distillation techniques,\nadversarial training techniques, and randomized or non-deterministic model\nbased techniques. In this paper, we propose a novel neural network paradigm\nthat falls under the category of randomized models for adversarial robustness,\nbut differs from all existing techniques under this category in that it models\neach parameter of the network as a statistical distribution with learnable\nparameters. We show experimentally that this framework is highly robust to a\nvariety of white-box and black-box adversarial attacks, while preserving the\ntask-specific performance of the traditional neural network model.', 'The last decade has witnessed the proliferation of Deep Learning models in\nmany applications, achieving unrivaled levels of predictive performance.\nUnfortunately, the black-box nature of Deep Learning models has posed\nunanswered questions about what they learn from data. Certain application\nscenarios have highlighted the importance of assessing the bounds under which\nDeep Learning models operate, a problem addressed by using assorted approaches\naimed at audiences from different domains. However, as the focus of the\napplication is placed more on non-expert users, it results mandatory to provide\nthe means for him/her to trust the model, just like a human gets familiar with\na system or process: by understanding the hypothetical circumstances under\nwhich it fails. This is indeed the angular stone for this research work: to\nundertake an adversarial analysis of a Deep Learning model. The proposed\nframework constructs counterfactual examples by ensuring their plausibility,\ne.g. there is a reasonable probability that a human could generate them without\nresorting to a computer program. Therefore, this work must be regarded as\nvaluable auditing exercise of the usable bounds a certain model is constrained\nwithin, thereby allowing for a much greater understanding of the capabilities\nand pitfalls of a model used in a real application. To this end, a Generative\nAdversarial Network (GAN) and multi-objective heuristics are used to furnish a\nplausible attack to the audited model, efficiently trading between the\nconfusion of this model, the intensity and plausibility of the generated\ncounterfactual. Its utility is showcased within a human face classification\ntask, unveiling the enormous potential of the proposed framework.', 'In this article I describe a research agenda for securing machine learning\nmodels against adversarial inputs at test time. This article does not present\nresults but instead shares some of my thoughts about where I think that the\nfield needs to go. Modern machine learning works very well on I.I.D. data: data\nfor which each example is drawn {\\em independently} and for which the\ndistribution generating each example is {\\em identical}. When these assumptions\nare relaxed, modern machine learning can perform very poorly. When machine\nlearning is used in contexts where security is a concern, it is desirable to\ndesign models that perform well even when the input is designed by a malicious\nadversary. So far most research in this direction has focused on an adversary\nwho violates the {\\em identical} assumption, and imposes some kind of\nrestricted worst-case distribution shift. I argue that machine learning\nsecurity researchers should also address the problem of relaxing the {\\em\nindependence} assumption and that current strategies designed for robustness to\ndistribution shift will not do so. I recommend {\\em dynamic models} that change\neach time they are run as a potential solution path to this problem, and show\nan example of a simple attack using correlated data that can be mitigated by a\nsimple dynamic defense. This is not intended as a real-world security measure,\nbut as a recommendation to explore this research direction and develop more\nrealistic defenses.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.', 'In the research field of activity recognition, although it is difficult to\ncollect a large amount of measured sensor data, there has not been much\ndiscussion about data augmentation (DA). In this study, I propose Octave Mix as\na new synthetic-style DA method for sensor-based activity recognition. Octave\nMix is a simple DA method that combines two types of waveforms by intersecting\nlow and high frequency waveforms using frequency decomposition. In addition, I\npropose a DA ensemble model and its training algorithm to acquire robustness to\nthe original sensor data while remaining a wide variety of feature\nrepresentation. I conducted experiments to evaluate the effectiveness of my\nproposed method using four different benchmark datasets of sensing-based\nactivity recognition. As a result, my proposed method achieved the best\nestimation accuracy. Furthermore, I found that ensembling two DA strategies:\nOctave Mix with rotation and mixup with rotation, make it possible to achieve\nhigher accuracy.', 'Cross-lingual document representations enable language understanding in\nmultilingual contexts and allow transfer learning from high-resource to\nlow-resource languages at the document level. Recently large pre-trained\nlanguage models such as BERT, XLM and XLM-RoBERTa have achieved great success\nwhen fine-tuned on sentence-level downstream tasks. It is tempting to apply\nthese cross-lingual models to document representation learning. However, there\nare two challenges: (1) these models impose high costs on long document\nprocessing and thus many of them have strict length limit; (2) model\nfine-tuning requires extra data and computational resources, which is not\npractical in resource-limited settings. In this work, we address these\nchallenges by proposing unsupervised Language-Agnostic Weighted Document\nRepresentations (LAWDR). We study the geometry of pre-trained sentence\nembeddings and leverage it to derive document representations without\nfine-tuning. Evaluated on cross-lingual document alignment, LAWDR demonstrates\ncomparable performance to state-of-the-art models on benchmark datasets.', 'Meta learning algorithms have been widely applied in many tasks for efficient\nlearning, such as few-shot image classification and fast reinforcement\nlearning. During meta training, the meta learner develops a common learning\nstrategy, or experience, from a variety of learning tasks. Therefore, during\nmeta test, the meta learner can use the learned strategy to quickly adapt to\nnew tasks even with a few training samples. However, there is still a dark side\nabout meta learning in terms of reliability and robustness. In particular, is\nmeta learning vulnerable to adversarial attacks? In other words, would a\nwell-trained meta learner utilize its learned experience to build wrong or\nlikely useless knowledge, if an adversary unnoticeably manipulates the given\ntraining set? Without the understanding of this problem, it is extremely risky\nto apply meta learning in safety-critical applications. Thus, in this paper, we\nperform the initial study about adversarial attacks on meta learning under the\nfew-shot classification problem. In particular, we formally define key elements\nof adversarial attacks unique to meta learning and propose the first attacking\nalgorithm against meta learning under various settings. We evaluate the\neffectiveness of the proposed attacking strategy as well as the robustness of\nseveral representative meta learning algorithms. Experimental results\ndemonstrate that the proposed attacking strategy can easily break the meta\nlearner and meta learning is vulnerable to adversarial attacks. The\nimplementation of the proposed framework will be released upon the acceptance\nof this paper.', 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', 'My overarching research goal is to provide robots with perceptional abilities\nthat allow interactions with humans in a human-like manner. To develop these\nperceptional abilities, I believe that it is useful to study the principles of\nthe human visual system. I use these principles to develop new computer vision\nalgorithms and validate their effectiveness in intelligent robotic systems. I\nam enthusiastic about this approach as it offers the dual benefit of uncovering\nprinciples inherent in the human visual system, as well as applying these\nprinciples to its artificial counterpart. Fig. 1 contains a depiction of my\nresearch.', 'Analyzing the layout of a document to identify headers, sections, tables,\nfigures etc. is critical to understanding its content. Deep learning based\napproaches for detecting the layout structure of document images have been\npromising. However, these methods require a large number of annotated examples\nduring training, which are both expensive and time consuming to obtain. We\ndescribe here a synthetic document generator that automatically produces\nrealistic documents with labels for spatial positions, extents and categories\nof the layout elements. The proposed generative process treats every physical\ncomponent of a document as a random variable and models their intrinsic\ndependencies using a Bayesian Network graph. Our hierarchical formulation using\nstochastic templates allow parameter sharing between documents for retaining\nbroad themes and yet the distributional characteristics produces visually\nunique samples, thereby capturing complex and diverse layouts. We empirically\nillustrate that a deep layout detection model trained purely on the synthetic\ndocuments can match the performance of a model that uses real documents.', 'Additive Manufacturing (AM) is a manufacturing paradigm that builds\nthree-dimensional objects from a computer-aided design model by successively\nadding material layer by layer. AM has become very popular in the past decade\ndue to its utility for fast prototyping such as 3D printing as well as\nmanufacturing functional parts with complex geometries using processes such as\nlaser metal deposition that would be difficult to create using traditional\nmachining. As the process for creating an intricate part for an expensive metal\nsuch as Titanium is prohibitive with respect to cost, computational models are\nused to simulate the behavior of AM processes before the experimental run.\nHowever, as the simulations are computationally costly and time-consuming for\npredicting multiscale multi-physics phenomena in AM, physics-informed\ndata-driven machine-learning systems for predicting the behavior of AM\nprocesses are immensely beneficial. Such models accelerate not only multiscale\nsimulation tools but also empower real-time control systems using in-situ data.\nIn this paper, we design and develop essential components of a scientific\nframework for developing a data-driven model-based real-time control system.\nFinite element methods are employed for solving time-dependent heat equations\nand developing the database. The proposed framework uses extremely randomized\ntrees - an ensemble of bagged decision trees as the regression algorithm\niteratively using temperatures of prior voxels and laser information as inputs\nto predict temperatures of subsequent voxels. The models achieve mean absolute\npercentage errors below 1% for predicting temperature profiles for AM\nprocesses. The code is made available for the research community at\nhttps://github.com/paularindam/ml-iter-additive.', 'Recent studies have shown that state-of-the-art deep learning models are\nvulnerable to the inputs with small perturbations (adversarial examples). We\nobserve two critical obstacles in adversarial examples: (i) Strong adversarial\nattacks (e.g., C&W attack) require manually tuning hyper-parameters and take a\nlong time to construct an adversarial example, making it impractical to attack\nreal-time systems; (ii) Most of the studies focus on non-sequential tasks, such\nas image classification, yet only a few consider sequential tasks. In this\nwork, we speed up adversarial attacks, especially on sequential learning tasks.\nBy leveraging the uncertainty of each task, we directly learn the adaptive\nmulti-task weightings, without manually searching hyper-parameters. A unified\narchitecture is developed and evaluated for both non-sequential tasks and\nsequential ones. To validate the effectiveness, we take the scene text\nrecognition task as a case study. To our best knowledge, our proposed method is\nthe first attempt to adversarial attack for scene text recognition. Adaptive\nAttack achieves over 99.9\\% success rate with 3-6X speedup compared to\nstate-of-the-art adversarial attacks.', 'Deep learning has emerged as a strong and efficient framework that can be\napplied to a broad spectrum of complex learning problems which were difficult\nto solve using the traditional machine learning techniques in the past. In the\nlast few years, deep learning has advanced radically in such a way that it can\nsurpass human-level performance on a number of tasks. As a consequence, deep\nlearning is being extensively used in most of the recent day-to-day\napplications. However, security of deep learning systems are vulnerable to\ncrafted adversarial examples, which may be imperceptible to the human eye, but\ncan lead the model to misclassify the output. In recent times, different types\nof adversaries based on their threat model leverage these vulnerabilities to\ncompromise a deep learning system where adversaries have high incentives.\nHence, it is extremely important to provide robustness to deep learning\nalgorithms against these adversaries. However, there are only a few strong\ncountermeasures which can be used in all types of attack scenarios to design a\nrobust deep learning system. In this paper, we attempt to provide a detailed\ndiscussion on different types of adversarial attacks with various threat models\nand also elaborate the efficiency and challenges of recent countermeasures\nagainst them.', 'Due to the increasing amount of data on the internet, finding a\nhighly-informative, low-dimensional representation for text is one of the main\nchallenges for efficient natural language processing tasks including text\nclassification. This representation should capture the semantic information of\nthe text while retaining their relevance level for document classification.\nThis approach maps the documents with similar topics to a similar space in\nvector space representation. To obtain representation for large text, we\npropose the utilization of deep Siamese neural networks. To embed document\nrelevance in topics in the distributed representation, we use a Siamese neural\nnetwork to jointly learn document representations. Our Siamese network consists\nof two sub-network of multi-layer perceptron. We examine our representation for\nthe text categorization task on BBC news dataset. The results show that the\nproposed representations outperform the conventional and state-of-the-art\nrepresentations in the text classification task on this dataset.', 'Inspired by the success of self attention mechanism and Transformer\narchitecture in sequence transduction and image generation applications, we\npropose novel self attention-based architectures to improve the performance of\nadversarial latent code- based schemes in text generation. Adversarial latent\ncode-based text generation has recently gained a lot of attention due to their\npromising results. In this paper, we take a step to fortify the architectures\nused in these setups, specifically AAE and ARAE. We benchmark two latent\ncode-based methods (AAE and ARAE) designed based on adversarial setups. In our\nexperiments, the Google sentence compression dataset is utilized to compare our\nmethod with these methods using various objective and subjective measures. The\nexperiments demonstrate the proposed (self) attention-based models outperform\nthe state-of-the-art in adversarial code-based text generation.', 'Topic models are frequently used in machine learning owing to their high\ninterpretability and modular structure. However, extending a topic model to\ninclude a supervisory signal, to incorporate pre-trained word embedding vectors\nand to include a nonlinear output function is not an easy task because one has\nto resort to a highly intricate approximate inference procedure. The present\npaper shows that topic modeling with pre-trained word embedding vectors can be\nviewed as implementing a neighborhood aggregation algorithm where messages are\npassed through a network defined over words. From the network view of topic\nmodels, nodes correspond to words in a document and edges correspond to either\na relationship describing co-occurring words in a document or a relationship\ndescribing the same word in the corpus. The network view allows us to extend\nthe model to include supervisory signals, incorporate pre-trained word\nembedding vectors and include a nonlinear output function in a simple manner.\nIn experiments, we show that our approach outperforms the state-of-the-art\nsupervised Latent Dirichlet Allocation implementation in terms of held-out\ndocument classification tasks.', 'In this dissertation, I present my work towards exploring temporal\ninformation for better video understanding. Specifically, I have worked on two\nproblems: action recognition and semantic segmentation. For action recognition,\nI have proposed a framework, termed hidden two-stream networks, to learn an\noptimal motion representation that does not require the computation of optical\nflow. My framework alleviates several challenges faced in video classification,\nsuch as learning motion representations, real-time inference, multi-framerate\nhandling, generalizability to unseen actions, etc. For semantic segmentation, I\nhave introduced a general framework that uses video prediction models to\nsynthesize new training samples. By scaling up the training dataset, my trained\nmodels are more accurate and robust than previous models even without\nmodifications to the network architectures or objective functions. I believe\nvideos have much more potential to be mined, and temporal information is one of\nthe most important cues for machines to perceive the visual world better.', 'We present an approach to analyze $C^1(\\mathbb{R}^m)$ functions that\naddresses limitations present in the Active Subspaces (AS) method of\nConstantine et al.(2015; 2014). Under appropriate hypotheses, our Active\nManifolds (AM) method identifies a 1-D curve in the domain (the active\nmanifold) on which nearly all values of the unknown function are attained, and\nwhich can be exploited for approximation or analysis, especially when $m$ is\nlarge (high-dimensional input space). We provide theorems justifying our AM\ntechnique and an algorithm permitting functional approximation and sensitivity\nanalysis. Using accessible, low-dimensional functions as initial examples, we\nshow AM reduces approximation error by an order of magnitude compared to AS, at\nthe expense of more computation. Following this, we revisit the sensitivity\nanalysis by Glaws et al. (2017), who apply AS to analyze a magnetohydrodynamic\npower generator model, and compare the performance of AM on the same data. Our\nanalysis provides detailed information not captured by AS, exhibiting the\ninfluence of each parameter individually along an active manifold. Overall, AM\nrepresents a novel technique for analyzing functional models with benefits\nincluding: reducing $m$-dimensional analysis to a 1-D analogue, permitting more\naccurate regression than AS (at more computational expense), enabling more\ninformative sensitivity analysis, and granting accessible visualizations(2-D\nplots) of parameter sensitivity along the AM.', 'A neuroscience method to understanding the brain is to find and study the\npreferred stimuli that highly activate an individual cell or groups of cells.\nRecent advances in machine learning enable a family of methods to synthesize\npreferred stimuli that cause a neuron in an artificial or biological brain to\nfire strongly. Those methods are known as Activation Maximization (AM) or\nFeature Visualization via Optimization. In this chapter, we (1) review existing\nAM techniques in the literature; (2) discuss a probabilistic interpretation for\nAM; and (3) review the applications of AM in debugging and explaining networks.', ""Deep learning has demonstrated state-of-the-art performance for a variety of\nchallenging computer vision tasks. On one hand, this has enabled deep visual\nmodels to pave the way for a plethora of critical applications like disease\nprognostics and smart surveillance. On the other, deep learning has also been\nfound vulnerable to adversarial attacks, which calls for new techniques to\ndefend deep models against these attacks. Among the attack algorithms, the\nblack-box schemes are of serious practical concern since they only need\npublicly available knowledge of the targeted model. We carefully analyze the\ninherent weakness of deep models in black-box settings where the attacker may\ndevelop the attack using a model similar to the targeted model. Based on our\nanalysis, we introduce a novel gradient regularization scheme that encourages\nthe internal representation of a deep model to be orthogonal to another, even\nif the architectures of the two models are similar. Our unique constraint\nallows a model to concomitantly endeavour for higher accuracy while maintaining\nnear orthogonal alignment of gradients with respect to a reference model.\nDetailed empirical study verifies that controlled misalignment of gradients\nunder our orthogonality objective significantly boosts a model's robustness\nagainst transferable black-box adversarial attacks. In comparison to regular\nmodels, the orthogonal models are significantly more robust to a range of $l_p$\nnorm bounded perturbations. We verify the effectiveness of our technique on a\nvariety of large-scale models."", 'Deep Learning has become interestingly popular in computer vision, mostly\nattaining near or above human-level performance in various vision tasks. But\nrecent work has also demonstrated that these deep neural networks are very\nvulnerable to adversarial examples (adversarial examples - inputs to a model\nwhich are naturally similar to original data but fools the model in classifying\nit into a wrong class). Humans are very robust against such perturbations; one\npossible reason could be that humans do not learn to classify based on an error\nbetween ""target label"" and ""predicted label"" but possibly due to reinforcements\nthat they receive on their predictions. In this work, we proposed a novel\nmethod to train deep learning models on an image classification task. We used a\nreward-based optimization function, similar to the vanilla policy gradient\nmethod used in reinforcement learning, to train our model instead of\nconventional cross-entropy loss. An empirical evaluation on the cifar10 dataset\nshowed that our method learns a more robust classifier than the same model\narchitecture trained using cross-entropy loss function (on adversarial\ntraining). At the same time, our method shows a better generalization with the\ndifference in test accuracy and train accuracy $< 2\\%$ for most of the time\ncompared to the cross-entropy one, whose difference most of the time remains $>\n2\\%$.', 'The burgeoning success of deep learning has raised the security and privacy\nconcerns as more and more tasks are accompanied with sensitive data.\nAdversarial attacks in deep learning have emerged as one of the dominating\nsecurity threat to a range of mission-critical deep learning systems and\napplications. This paper takes a holistic and principled approach to perform\nstatistical characterization of adversarial examples in deep learning. We\nprovide a general formulation of adversarial examples and elaborate on the\nbasic principle for adversarial attack algorithm design. We introduce easy and\nhard categorization of adversarial attacks to analyze the effectiveness of\nadversarial examples in terms of attack success rate, degree of change in\nadversarial perturbation, average entropy of prediction qualities, and fraction\nof adversarial examples that lead to successful attacks. We conduct extensive\nexperimental study on adversarial behavior in easy and hard attacks under deep\nlearning models with different hyperparameters and different deep learning\nframeworks. We show that the same adversarial attack behaves differently under\ndifferent hyperparameters and across different frameworks due to the different\nfeatures learned under different deep learning model training process. Our\nstatistical characterization with strong empirical evidence provides a\ntransformative enlightenment on mitigation strategies towards effective\ncountermeasures against present and future adversarial attacks.', 'Automatic document content processing is affected by artifacts caused by the\nshape of the paper, non-uniform and diverse color of lighting conditions.\nFully-supervised methods on real data are impossible due to the large amount of\ndata needed. Hence, the current state of the art deep learning models are\ntrained on fully or partially synthetic images. However, document shadow or\nshading removal results still suffer because: (a) prior methods rely on\nuniformity of local color statistics, which limit their application on\nreal-scenarios with complex document shapes and textures and; (b) synthetic or\nhybrid datasets with non-realistic, simulated lighting conditions are used to\ntrain the models. In this paper we tackle these problems with our two main\ncontributions. First, a physically constrained learning-based method that\ndirectly estimates document reflectance based on intrinsic image formation\nwhich generalizes to challenging illumination conditions. Second, a new dataset\nthat clearly improves previous synthetic ones, by adding a large range of\nrealistic shading and diverse multi-illuminant conditions, uniquely customized\nto deal with documents in-the-wild. The proposed architecture works in a\nself-supervised manner where only the synthetic texture is used as a weak\ntraining signal (obviating the need for very costly ground truth with\ndisentangled versions of shading and reflectance). The proposed approach leads\nto a significant generalization of document reflectance estimation in real\nscenes with challenging illumination. We extensively evaluate on the real\nbenchmark datasets available for intrinsic image decomposition and document\nshadow removal tasks. Our reflectance estimation scheme, when used as a\npre-processing step of an OCR pipeline, shows a 26% improvement of character\nerror rate (CER), thus, proving the practical applicability.', 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', ""Early detection of psychological distress is key to effective treatment.\nAutomatic detection of distress, such as depression, is an active area of\nresearch. Current approaches utilise vocal, facial, and bodily modalities. Of\nthese, the bodily modality is the least investigated, partially due to the\ndifficulty in extracting bodily representations from videos, and partially due\nto the lack of viable datasets. Existing body modality approaches use automatic\ncategorization of expressions to represent body language as a series of\nspecific expressions, much like words within natural language. In this\ndissertation I present a new type of feature, within the body modality, that\nrepresents meta information of gestures, such as speed, and use it to predict a\nnon-clinical depression label. This differs to existing work by representing\noverall behaviour as a small set of aggregated meta features derived from a\nperson's movement. In my method I extract pose estimation from videos, detect\ngestures within body parts, extract meta information from individual gestures,\nand finally aggregate these features to generate a small feature vector for use\nin prediction tasks. I introduce a new dataset of 65 video recordings of\ninterviews with self-evaluated distress, personality, and demographic labels.\nThis dataset enables the development of features utilising the whole body in\ndistress detection tasks. I evaluate my newly introduced meta-features for\npredicting depression, anxiety, perceived stress, somatic stress, five standard\npersonality measures, and gender. A linear regression based classifier using\nthese features achieves a 82.70% F1 score for predicting depression within my\nnovel dataset."", 'Document clustering and topic modeling are two closely related tasks which\ncan mutually benefit each other. Topic modeling can project documents into a\ntopic space which facilitates effective document clustering. Cluster labels\ndiscovered by document clustering can be incorporated into topic models to\nextract local topics specific to each cluster and global topics shared by all\nclusters. In this paper, we propose a multi-grain clustering topic model\n(MGCTM) which integrates document clustering and topic modeling into a unified\nframework and jointly performs the two tasks to achieve the overall best\nperformance. Our model tightly couples two components: a mixture component used\nfor discovering latent groups in document collection and a topic model\ncomponent used for mining multi-grain topics including local topics specific to\neach cluster and global topics shared across clusters.We employ variational\ninference to approximate the posterior of hidden variables and learn model\nparameters. Experiments on two datasets demonstrate the effectiveness of our\nmodel.', 'Document clustering is a text mining technique used to provide better\ndocument search and browsing in digital libraries or online corpora. A lot of\nresearch has been done on biomedical document clustering that is based on using\nexisting ontology. But, associations and co-occurrences of the medical concepts\nare not well represented by using ontology. In this research, a vector\nrepresentation of concepts of diseases and similarity measurement between\nconcepts are proposed. They identify the closest concepts of diseases in the\ncontext of a corpus. Each document is represented by using the vector space\nmodel. A weight scheme is proposed to consider both local content and\nassociations between concepts. A Self-Organizing Map is used as document\nclustering algorithm. The vector projection and visualization features of SOM\nenable visualization and analysis of the clusters distributions and\nrelationships on the two dimensional space. The experimental results show that\nthe proposed document clustering framework generates meaningful clusters and\nfacilitate visualization of the clusters based on the concepts of diseases.', 'Text Mining is a field that aims at extracting information from textual data.\nOne of the challenges of such field of study comes from the pre-processing\nstage in which a vector (and structured) representation should be extracted\nfrom unstructured data. The common extraction creates large and sparse vectors\nrepresenting the importance of each term to a document. As such, this usually\nleads to the curse-of-dimensionality that plagues most machine learning\nalgorithms. To cope with this issue, in this paper we propose a new supervised\nfeature extraction and reduction algorithm, named DCDistance, that creates\nfeatures based on the distance between a document to a representative of each\nclass label. As such, the proposed technique can reduce the features set in\nmore than 99% of the original set. Additionally, this algorithm was also\ncapable of improving the classification accuracy over a set of benchmark\ndatasets when compared to traditional and state-of-the-art features selection\nalgorithms.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', ""Convolututional Neural Networks have achieved state of the art in image\nclassification, object detection and other image related tasks. In this paper I\npresent another use of CNNs i.e. if given a set of images and then giving a\nsingle test image the network identifies that the test image is part of which\nimage from the images given before. This is a task somehow similar to measuring\nimage similarity and can be done using a simple CNN. Doing this task manually\nby looping can be quite a time consuming problem and won't be a generalizable\nsolution. The task is quite similar to doing object detection but for that lots\ntraining data should be given or in the case of sliding window it takes lot of\ntime and my algorithm can work with much fewer examples, is totally\nunsupervised and works much efficiently. Also, I explain that how unsupervised\nalgorithm like K-Means or supervised algorithm like K-NN are not good enough to\nperform this task. The basic idea is that image encodings are collected for\neach image from a CNN, when a test image comes it is replaced by a part of\noriginal image, the encoding is generated using the same network, the frobenius\nnorm is calculated and if it comes under a tolerance level then the test image\nis said to be the part of the original image."", 'Computation of document image quality metrics often depends upon the\navailability of a ground truth image corresponding to the document. This limits\nthe applicability of quality metrics in applications such as hyperparameter\noptimization of image processing algorithms that operate on-the-fly on unseen\ndocuments. This work proposes the use of surrogate models to learn the behavior\nof a given document quality metric on existing datasets where ground truth\nimages are available. The trained surrogate model can later be used to predict\nthe metric value on previously unseen document images without requiring access\nto ground truth images. The surrogate model is empirically evaluated on the\nDocument Image Binarization Competition (DIBCO) and the Handwritten Document\nImage Binarization Competition (H-DIBCO) datasets.', 'Machine Learning (ML) and Deep Learning (DL) models have achieved\nstate-of-the-art performance on multiple learning tasks, from vision to natural\nlanguage modelling. With the growing adoption of ML and DL to many areas of\ncomputer science, recent research has also started focusing on the security\nproperties of these models. There has been a lot of work undertaken to\nunderstand if (deep) neural network architectures are resilient to black-box\nadversarial attacks which craft perturbed input samples that fool the\nclassifier without knowing the architecture used. Recent work has also focused\non the transferability of adversarial attacks and found that adversarial\nattacks are generally easily transferable between models, datasets, and\ntechniques. However, such attacks and their analysis have not been covered from\nthe perspective of unsupervised machine learning algorithms. In this paper, we\nseek to bridge this gap through multiple contributions. We first provide a\nstrong (iterative) black-box adversarial attack that can craft adversarial\nsamples which will be incorrectly clustered irrespective of the choice of\nclustering algorithm. We choose 4 prominent clustering algorithms, and a\nreal-world dataset to show the working of the proposed adversarial algorithm.\nUsing these clustering algorithms we also carry out a simple study of\ncross-technique adversarial attack transferability.', 'This paper presents results of Document Visual Question Answering Challenge\norganized as part of ""Text and Documents in the Deep Learning Era"" workshop, in\nCVPR 2020. The challenge introduces a new problem - Visual Question Answering\non document images. The challenge comprised two tasks. The first task concerns\nwith asking questions on a single document image. On the other hand, the second\ntask is set as a retrieval task where the question is posed over a collection\nof images. For the task 1 a new dataset is introduced comprising 50,000\nquestions-answer(s) pairs defined over 12,767 document images. For task 2\nanother dataset has been created comprising 20 questions over 14,362 document\nimages which share the same document template.', ""Sentence position is a strong feature for news summarization, since the lead\noften (but not always) summarizes the key points of the article. In this paper,\nwe show that recent neural systems excessively exploit this trend, which\nalthough powerful for many inputs, is also detrimental when summarizing\ndocuments where important content should be extracted from later parts of the\narticle. We propose two techniques to make systems sensitive to the importance\nof content in different parts of the article. The first technique employs\n'unbiased' data; i.e., randomly shuffled sentences of the source document, to\npretrain the model. The second technique uses an auxiliary ROUGE-based loss\nthat encourages the model to distribute importance scores throughout a document\nby mimicking sentence-level ROUGE scores on the training data. We show that\nthese techniques significantly improve the performance of a competitive\nreinforcement learning based extractive system, with the auxiliary loss being\nmore powerful than pretraining."", 'Graphical elements: particularly tables and figures contain a visual summary\nof the most valuable information contained in a document. Therefore,\nlocalization of such graphical objects in the document images is the initial\nstep to understand the content of such graphical objects or document images. In\nthis paper, we present a novel end-to-end trainable deep learning based\nframework to localize graphical objects in the document images called as\nGraphical Object Detection (GOD). Our framework is data-driven and does not\nrequire any heuristics or meta-data to locate graphical objects in the document\nimages. The GOD explores the concept of transfer learning and domain adaptation\nto handle scarcity of labeled training images for graphical object detection\ntask in the document images. Performance analysis carried out on the various\npublic benchmark data sets: ICDAR-2013, ICDAR-POD2017,and UNLV shows that our\nmodel yields promising results as compared to state-of-the-art techniques.', ""In information retrieval (IR) and related tasks, term weighting approaches\ntypically consider the frequency of the term in the document and in the\ncollection in order to compute a score reflecting the importance of the term\nfor the document. In tasks characterized by the presence of training data (such\nas text classification) it seems logical that the term weighting function\nshould take into account the distribution (as estimated from training data) of\nthe term across the classes of interest. Although `supervised term weighting'\napproaches that use this intuition have been described before, they have failed\nto show consistent improvements. In this article we analyse the possible\nreasons for this failure, and call consolidated assumptions into question.\nFollowing this criticism we propose a novel supervised term weighting approach\nthat, instead of relying on any predefined formula, learns a term weighting\nfunction optimised on the training set of interest; we dub this approach\n\\emph{Learning to Weight} (LTW). The experiments that we run on several\nwell-known benchmarks, and using different learning methods, show that our\nmethod outperforms previous term weighting approaches in text classification."", ""Neural networks with deep architectures have demonstrated significant\nperformance improvements in computer vision, speech recognition, and natural\nlanguage processing. The challenges in information retrieval (IR), however, are\ndifferent from these other application areas. A common form of IR involves\nranking of documents--or short passages--in response to keyword-based queries.\nEffective IR systems must deal with query-document vocabulary mismatch problem,\nby modeling relationships between different query and document terms and how\nthey indicate relevance. Models should also consider lexical matches when the\nquery contains rare terms--such as a person's name or a product model\nnumber--not seen during training, and to avoid retrieving semantically related\nbut irrelevant results. In many real-life IR tasks, the retrieval involves\nextremely large collections--such as the document index of a commercial Web\nsearch engine--containing billions of documents. Efficient IR methods should\ntake advantage of specialized IR data structures, such as inverted index, to\nefficiently retrieve from large collections. Given an information need, the IR\nsystem also mediates how much exposure an information artifact receives by\ndeciding whether it should be displayed, and where it should be positioned,\namong other results. Exposure-aware IR systems may optimize for additional\nobjectives, besides relevance, such as parity of exposure for retrieved items\nand content publishers. In this thesis, we present novel neural architectures\nand methods motivated by the specific needs and challenges of IR tasks."", 'Identity documents recognition is an important sub-field of document\nanalysis, which deals with tasks of robust document detection, type\nidentification, text fields recognition, as well as identity fraud prevention\nand document authenticity validation given photos, scans, or video frames of an\nidentity document capture. Significant amount of research has been published on\nthis topic in recent years, however a chief difficulty for such research is\nscarcity of datasets, due to the subject matter being protected by security\nrequirements. A few datasets of identity documents which are available lack\ndiversity of document types, capturing conditions, or variability of document\nfield values. In addition, the published datasets were typically designed only\nfor a subset of document recognition problems, not for a complex identity\ndocument analysis. In this paper, we present a dataset MIDV-2020 which consists\nof 1000 video clips, 2000 scanned images, and 1000 photos of 1000 unique mock\nidentity documents, each with unique text field values and unique artificially\ngenerated faces, with rich annotation. For the presented benchmark dataset\nbaselines are provided for such tasks as document location and identification,\ntext fields recognition, and face detection. With 72409 annotated images in\ntotal, to the date of publication the proposed dataset is the largest publicly\navailable identity documents dataset with variable artificially generated data,\nand we believe that it will prove invaluable for advancement of the field of\ndocument analysis and recognition. The dataset is available for download at\nftp://smartengines.com/midv-2020 and http://l3i-share.univ-lr.fr .', 'This paper describes the short-term competition on the Components\nSegmentation Task of Document Photos that was prepared in the context of the\n16th International Conference on Document Analysis and Recognition (ICDAR\n2021). This competition aims to bring together researchers working in the field\nof identification document image processing and provides them a suitable\nbenchmark to compare their techniques on the component segmentation task of\ndocument images. Three challenge tasks were proposed entailing different\nsegmentation assignments to be performed on a provided dataset. The collected\ndata are from several types of Brazilian ID documents, whose personal\ninformation was conveniently replaced. There were 16 participants whose results\nobtained for some or all the three tasks show different rates for the adopted\nmetrics, like Dice Similarity Coefficient ranging from 0.06 to 0.99. Different\nDeep Learning models were applied by the entrants with diverse strategies to\nachieve the best results in each of the tasks. Obtained results show that the\ncurrently applied methods for solving one of the proposed tasks (document\nboundary detection) are already well established. However, for the other two\nchallenge tasks (text zone and handwritten sign detection) research and\ndevelopment of more robust approaches are still required to achieve acceptable\nresults.', 'The Query Focused Text Summarization (QFTS) task aims at building systems\nthat generate the summary of the text document(s) based on the given query. A\nkey challenge in addressing this task is the lack of large labeled data for\ntraining the summarization model. In this paper, we address this challenge by\nexploring a series of domain adaptation techniques. Given the recent success of\npre-trained transformer models in a wide range of natural language processing\ntasks, we utilize such models to generate abstractive summaries for the QFTS\ntask for both single-document and multi-document scenarios. For domain\nadaptation, we apply a variety of techniques using pre-trained\ntransformer-based summarization models including transfer learning, weakly\nsupervised learning, and distant supervision. Extensive experiments on six\ndatasets show that our proposed approach is very effective in generating\nabstractive summaries for the QFTS task while setting a new state-of-the-art\nresult in several datasets across a set of automatic and human evaluation\nmetrics.', 'We present a novel Bayesian topic model for learning discourse-level document\nstructure. Our model leverages insights from discourse theory to constrain\nlatent topic assignments in a way that reflects the underlying organization of\ndocument topics. We propose a global model in which both topic selection and\nordering are biased to be similar across a collection of related documents. We\nshow that this space of orderings can be effectively represented using a\ndistribution over permutations called the Generalized Mallows Model. We apply\nour method to three complementary discourse-level tasks: cross-document\nalignment, document segmentation, and information ordering. Our experiments\nshow that incorporating our permutation-based model in these applications\nyields substantial improvements in performance over previously proposed\nmethods.', ""What should I work on first? What can wait until later? Which projects should\nI prioritize and which tasks are not worth my time? These are challenging\nquestions that many people face every day. People's intuitive strategy is to\nprioritize their immediate experience over the long-term consequences. This\nleads to procrastination and the neglect of important long-term projects in\nfavor of seemingly urgent tasks that are less important. Optimal gamification\nstrives to help people overcome these problems by incentivizing each task by a\nnumber of points that communicates how valuable it is in the long-run.\nUnfortunately, computing the optimal number of points with standard dynamic\nprogramming methods quickly becomes intractable as the number of a person's\nprojects and the number of tasks required by each project increase. Here, we\nintroduce and evaluate a scalable method for identifying which tasks are most\nimportant in the long run and incentivizing each task according to its\nlong-term value. Our method makes it possible to create to-do list gamification\napps that can handle the size and complexity of people's to-do lists in the\nreal world."", 'With the increase of information, document classification as one of the\nmethods of text mining, plays vital role in many management and organizing\ninformation. Document classification is the process of assigning a document to\none or more predefined category labels. Document classification includes\ndifferent parts such as text processing, term selection, term weighting and\nfinal classification. The accuracy of document classification is very\nimportant. Thus improvement in each part of classification should lead to\nbetter results and higher precision. Term weighting has a great impact on the\naccuracy of the classification. Most of the existing weighting methods exploit\nthe statistical information of terms in documents and do not consider semantic\nrelations between words. In this paper, an automated document classification\nsystem is presented that uses a novel term weighting method based on semantic\nrelations between terms. To evaluate the proposed method, three standard\nPersian corpuses are used. Experiment results show 2 to 4 percent improvement\nin classification accuracy compared with the best previous designed system for\nPersian documents.', 'This paper addresses the problem of unsupervised clustering which remains one\nof the most fundamental challenges in machine learning and artificial\nintelligence. We propose the clustered generator model for clustering which\ncontains both continuous and discrete latent variables. Discrete latent\nvariables model the cluster label while the continuous ones model variations\nwithin each cluster. The learning of the model proceeds in a unified\nprobabilistic framework and incorporates the unsupervised clustering as an\ninner step without the need for an extra inference model as in existing\nvariational-based models. The latent variables learned serve as both observed\ndata embedding or latent representation for data distribution. Our experiments\nshow that the proposed model can achieve competitive unsupervised clustering\naccuracy and can learn disentangled latent representations to generate\nrealistic samples. In addition, the model can be naturally extended to\nper-pixel unsupervised clustering which remains largely unexplored.', ""Video feeds are often deliberately used as evidence, as in the case of CCTV\nfootage; but more often than not, the existence of footage of a supposed event\nis perceived as proof of fact in the eyes of the public at large. This reliance\nrepresents a societal vulnerability given the existence of easy-to-use editing\ntools and means to fabricate entire video feeds using machine learning. And, as\nthe recent barrage of fake news and fake porn videos have shown, this isn't\nmerely an academic concern, it is actively been exploited. I posit that this\nexploitation is only going to get more insidious. In this position paper, I\nintroduce a long term project that aims to mitigate some of the most egregious\nforms of manipulation by embedding trustworthy components in the video\ntransmission chain. Unlike earlier works, I am not aiming to do tamper\ndetection or other forms of forensics -- approaches I think are bound to fail\nin the face of the reality of necessary editing and compression -- instead, the\naim here is to provide a way for the video publisher to prove the integrity of\nthe video feed as well as make explicit any edits they may have performed. To\ndo this, I present a novel data structure, a video-edit specification language\nand supporting infrastructure that provides end-to-end video provenance, from\nthe camera sensor to the viewer. I have implemented a prototype of this system\nand am in talks with journalists and video editors to discuss the best ways\nforward with introducing this idea to the mainstream."", 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.', 'Social and behavioral interventions are a critical tool for governments and\ncommunities to tackle deep-rooted societal challenges such as homelessness,\ndisease, and poverty. However, real-world interventions are almost always\nplagued by limited resources and limited data, which creates a computational\nchallenge: how can we use algorithmic techniques to enhance the targeting and\ndelivery of social and behavioral interventions? The goal of my thesis is to\nprovide a unified study of such questions, collectively considered under the\nname ""algorithmic social intervention"". This proposal introduces algorithmic\nsocial intervention as a distinct area with characteristic technical\nchallenges, presents my published research in the context of these challenges,\nand outlines open problems for future work. A common technical theme is\ndecision making under uncertainty: how can we find actions which will impact a\nsocial system in desirable ways under limitations of knowledge and resources?\nThe primary application area for my work thus far is public health, e.g. HIV or\ntuberculosis prevention. For instance, I have developed a series of algorithms\nwhich optimize social network interventions for HIV prevention. Two of these\nalgorithms have been pilot-tested in collaboration with LA-area service\nproviders for homeless youth, with preliminary results showing substantial\nimprovement over status-quo approaches. My work also spans other topics in\ninfectious disease prevention and underlying algorithmic questions in robust\nand risk-aware submodular optimization.', ""The focus of my PhD thesis is on exploring parallel approaches to efficiently\nsolve problems modeled by constraints and presenting a new proposal. Current\nsolvers are very advanced; they are carefully designed to effectively manage\nthe high-level problems' description and include refined strategies to avoid\nuseless work. Despite this, finding a solution can take an unacceptable amount\nof time. Parallelization can mitigate this problem when the instance of the\nproblem modeled is large, as it happens in real world problems. It is done by\npropagating constraints in parallel and concurrently exploring different parts\nof the search space. I am developing on a constraint solver that exploits the\nmany cores available on Graphics Processing Units (GPU) to speed up the search."", 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', ""It is well-known that machine learning models are vulnerable to small but\ncleverly-designed adversarial perturbations that can cause misclassification.\nWhile there has been major progress in designing attacks and defenses for\nvarious adversarial settings, many fundamental and theoretical problems are yet\nto be resolved. In this paper, we consider classification in the presence of\n$\\ell_0$-bounded adversarial perturbations, a.k.a. sparse attacks. This setting\nis significantly different from other $\\ell_p$-adversarial settings, with\n$p\\geq 1$, as the $\\ell_0$-ball is non-convex and highly non-smooth. Under the\nassumption that data is distributed according to the Gaussian mixture model,\nour goal is to characterize the optimal robust classifier and the corresponding\nrobust classification error as well as a variety of trade-offs between\nrobustness, accuracy, and the adversary's budget. To this end, we develop a\nnovel classification algorithm called FilTrun that has two main modules:\nFiltration and Truncation. The key idea of our method is to first filter out\nthe non-robust coordinates of the input and then apply a carefully-designed\ntruncated inner product for classification. By analyzing the performance of\nFilTrun, we derive an upper bound on the optimal robust classification error.\nWe also find a lower bound by designing a specific adversarial strategy that\nenables us to derive the corresponding robust classifier and its achieved\nerror. For the case that the covariance matrix of the Gaussian mixtures is\ndiagonal, we show that as the input's dimension gets large, the upper and lower\nbounds converge; i.e. we characterize the asymptotically-optimal robust\nclassifier. Throughout, we discuss several examples that illustrate interesting\nbehaviors such as the existence of a phase transition for adversary's budget\ndetermining whether the effect of adversarial perturbation can be fully\nneutralized."", 'Due to its decentralized nature, Federated Learning (FL) lends itself to\nadversarial attacks in the form of backdoors during training. The goal of a\nbackdoor is to corrupt the performance of the trained model on specific\nsub-tasks (e.g., by classifying green cars as frogs). A range of FL backdoor\nattacks have been introduced in the literature, but also methods to defend\nagainst them, and it is currently an open question whether FL systems can be\ntailored to be robust against backdoors. In this work, we provide evidence to\nthe contrary. We first establish that, in the general case, robustness to\nbackdoors implies model robustness to adversarial examples, a major open\nproblem in itself. Furthermore, detecting the presence of a backdoor in a FL\nmodel is unlikely assuming first order oracles or polynomial time. We couple\nour theoretical results with a new family of backdoor attacks, which we refer\nto as edge-case backdoors. An edge-case backdoor forces a model to misclassify\non seemingly easy inputs that are however unlikely to be part of the training,\nor test data, i.e., they live on the tail of the input distribution. We explain\nhow these edge-case backdoors can lead to unsavory failures and may have\nserious repercussions on fairness, and exhibit that with careful tuning at the\nside of the adversary, one can insert them across a range of machine learning\ntasks (e.g., image classification, OCR, text prediction, sentiment analysis).', 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.', 'Despite the remarkable performance and generalization levels of deep learning\nmodels in a wide range of artificial intelligence tasks, it has been\ndemonstrated that these models can be easily fooled by the addition of\nimperceptible yet malicious perturbations to natural inputs. These altered\ninputs are known in the literature as adversarial examples. In this paper, we\npropose a novel probabilistic framework to generalize and extend adversarial\nattacks in order to produce a desired probability distribution for the classes\nwhen we apply the attack method to a large number of inputs. This novel attack\nparadigm provides the adversary with greater control over the target model,\nthereby exposing, in a wide range of scenarios, threats against deep learning\nmodels that cannot be conducted by the conventional paradigms. We introduce\nfour different strategies to efficiently generate such attacks, and illustrate\nour approach by extending multiple adversarial attack algorithms. We also\nexperimentally validate our approach for the spoken command classification task\nand the Tweet emotion classification task, two exemplary machine learning\nproblems in the audio and text domain, respectively. Our results demonstrate\nthat we can closely approximate any probability distribution for the classes\nwhile maintaining a high fooling rate and even prevent the attacks from being\ndetected by label-shift detection methods.', 'Deep neural networks (DNNs) are vulnerable to adversarial attack which is\nmaliciously implemented by adding human-imperceptible perturbation to images\nand thus leads to incorrect prediction. Existing studies have proposed various\nmethods to detect the new adversarial attacks. However, new attack methods keep\nevolving constantly and yield new adversarial examples to bypass the existing\ndetectors. It needs to collect tens of thousands samples to train detectors,\nwhile the new attacks evolve much more frequently than the high-cost data\ncollection. Thus, this situation leads the newly evolved attack samples to\nremain in small scales. To solve such few-shot problem with the evolving\nattack, we propose a meta-learning based robust detection method to detect new\nadversarial attacks with limited examples. Specifically, the learning consists\nof a double-network framework: a task-dedicated network and a master network\nwhich alternatively learn the detection capability for either seen attack or a\nnew attack. To validate the effectiveness of our approach, we construct the\nbenchmarks with few-shot-fashion protocols based on three conventional\ndatasets, i.e. CIFAR-10, MNIST and Fashion-MNIST. Comprehensive experiments are\nconducted on them to verify the superiority of our approach with respect to the\ntraditional adversarial attack detection methods.', ""Additive manufacturing (AM) is growing as fast as anyone can imagine, and it\nis now a multi-billion-dollar industry. AM becomes popular in a variety of\nsectors, such as automotive, aerospace, biomedical, and pharmaceutical, for\nproducing parts/ components/ subsystems. However, current AM technologies can\nface vast risks of security issues and privacy loss. For the security of AM\nprocess, many researchers are working on the defense mechanism to\ncountermeasure such security concerns and finding efficient ways to eliminate\nthose risks. Researchers have also been conducting experiments to establish a\nsecure framework for the user's privacy and security components. This survey\nconsists of four sections. In the first section, we will explore the relevant\nlimitations of additive manufacturing in terms of printing capability,\nsecurity, and possible solutions. The second section will present different\nkinds of attacks on AM and their effects. The next part will analyze and\ndiscuss the mechanisms and frameworks for access control and authentication for\nAM devices. The final section examines the security issues in various\nindustrial sectors and provides the observations on the security of the\nadditive manufacturing process."", ""I am Joachim Jansen and this is my research summary, part of my application\nto the Doctoral Consortium at ICLP'14. I am a PhD student in the Knowledge\nRepresentation and Reasoning (KRR) research group, a subgroup of the\nDeclarative Languages and Artificial Intelligence (DTAI) group at the\ndepartment of Computer Science at KU Leuven. I started my PhD in September\n2012. My promotor is prof. dr. ir. Gerda Janssens and my co-promotor is prof.\ndr. Marc Denecker. I can be contacted at joachim.jansen@cs.kuleuven.be or at:\nRoom 01.167 Celestijnenlaan 200A 3001 Heverlee Belgium An extended abstract /\nfull version of a paper accepted to be presented at the Doctoral Consortium of\nthe 30th International Conference on Logic Programming (ICLP 2014), July 19-22,\nVienna, Austria"", ""In this paper, I will discuss the work I am currently doing as a Ph.D.\nstudent at the University of Potsdam, under the tutoring of T. Schaub. I'm\ncurrently looking into action description in ASP. More precisely, my goal is to\nexplore how to represent actions with durations in ASP, in different contexts.\nRight now, I'm focused on Multi-Agent Path Finding (MAPF), looking at how to\nrepresent speeds for different agents and contexts.\n  Before tackling duration, I wanted to explore and compare different\nrepresentations of action taking in ASP. For this, I started comparing\ndifferent simple encodings tackling the MAPF problem. Even in simple code,\nchoices and assumptions have been made in their creations. The objective of my\nwork is to present the consequences of those design decisions in terms of\nperformance and knowledge representation. As far as I know, there is no current\nresearch on this topic.\n  Besides that, I'm also exploring different ways to represent duration and to\nsolve related problems. I planed to compare them the same way I described\nbefore. I also want this to help me find innovative and effective ways to solve\nproblems with duration."", '""Information Processing"" is a recently launched buzzword whose meaning is\nvague and obscure even for the majority of its users. The reason for this is\nthe lack of a suitable definition for the term ""information"". In my attempt to\namend this bizarre situation, I have realized that, following the insights of\nKolmogorov\'s Complexity theory, information can be defined as a description of\nstructures observable in a given data set. Two types of structures could be\neasily distinguished in every data set - in this regard, two types of\ninformation (information descriptions) should be designated: physical\ninformation and semantic information. Kolmogorov\'s theory also posits that the\ninformation descriptions should be provided as a linguistic text structure.\nThis inevitably leads us to an assertion that information processing has to be\nseen as a kind of text processing. The idea is not new - inspired by the\nobservation that human information processing is deeply rooted in natural\nlanguage handling customs, Lotfi Zadeh and his followers have introduced the\nso-called ""Computing With Words"" paradigm. Despite of promotional efforts, the\nidea is not taking off yet. The reason - a lack of a coherent understanding of\nwhat should be called ""information"", and, as a result, misleading research\nroadmaps and objectives. I hope my humble attempt to clarify these issues would\nbe helpful in avoiding common traps and pitfalls.', 'Document segmentation is a method of rending the document into distinct\nregions. A document is an assortment of information and a standard mode of\nconveying information to others. Pursuance of data from documents involves ton\nof human effort, time intense and might severely prohibit the usage of data\nsystems. So, automatic information pursuance from the document has become a big\nissue. It is been shown that document segmentation will facilitate to beat such\nproblems. This paper proposes a new approach to segment and classify the\ndocument regions as text, image, drawings and table. Document image is divided\ninto blocks using Run length smearing rule and features are extracted from\nevery blocks. Discipulus tool has been used to construct the Genetic\nprogramming based classifier model and located 97.5% classification accuracy.', ""The recent introduction of Find My app by Apple will open a large window of\nopportunities for whistleblowers. Based on a short range Bluetooth signals, an\nEC P-224 encryption, and an end-to-end encrypted manner using iCloud Keychain,\nFind My app is probably the first application broadcasting a large number of\nanonymous public key on this scale. Hence, this new Apple's application may\nintroduce a revolution in secret communication, if we divert it from its\nprimordial use and transform it into a powerful tool to put in the hands of\nwhistleblowers. By using Find My app and an entity authentication protocol\nbased on artificial intelligence, our goal is to make mass surveillance and\nkleptographic backdoors ineffective in the lifting of the whistleblower's\nanonymity. However, in some case, Find my app may also be a powerful tool in\nthe hands of dictatorships governments in their fight against whistleblowers\nand political adversaries. Thus, the aim of this paper is to show with simple\nexamples, how these two previous situation can happen."", ""Probability theory, epistemically interpreted, provides an excellent, if not\nthe best available account of inductive reasoning. This is so because there are\ngeneral and definite rules for the change of subjective probabilities through\ninformation or experience; induction and belief change are one and same topic,\nafter all. The most basic of these rules is simply to conditionalize with\nrespect to the information received; and there are similar and more general\nrules. 1 Hence, a fundamental reason for the epistemological success of\nprobability theory is that there at all exists a well-behaved concept of\nconditional probability. Still, people have, and have reasons for, various\nconcerns over probability theory. One of these is my starting point:\nIntuitively, we have the notion of plain belief; we believe propositions2 to be\ntrue (or to be false or neither). Probability theory, however, offers no formal\ncounterpart to this notion. Believing A is not the same as having probability 1\nfor A, because probability 1 is incorrigible3; but plain belief is clearly\ncorrigible. And believing A is not the same as giving A a probability larger\nthan some 1 - c, because believing A and believing B is usually taken to be\nequivalent to believing A & B.4 Thus, it seems that the formal representation\nof plain belief has to take a non-probabilistic route. Indeed, representing\nplain belief seems easy enough: simply represent an epistemic state by the set\nof all propositions believed true in it or, since I make the common assumption\nthat plain belief is deductively closed, by the conjunction of all propositions\nbelieved true in it. But this does not yet provide a theory of induction, i.e.\nan answer to the question how epistemic states so represented are changed\ntbrough information or experience. There is a convincing partial answer: if the\nnew information is compatible with the old epistemic state, then the new\nepistemic state is simply represented by the conjunction of the new information\nand the old beliefs. This answer is partial because it does not cover the quite\ncommon case where the new information is incompatible with the old beliefs. It\nis, however, important to complete the answer and to cover this case, too;\notherwise, we would not represent plain belief as conigible. The crucial\nproblem is that there is no good completion. When epistemic states are\nrepresented simply by the conjunction of all propositions believed true in it,\nthe answer cannot be completed; and though there is a lot of fruitful work, no\nother representation of epistemic states has been proposed, as far as I know,\nwhich provides a complete solution to this problem. In this paper, I want to\nsuggest such a solution. In [4], I have more fully argued that this is the only\nsolution, if certain plausible desiderata are to be satisfied. Here, in section\n2, I will be content with formally defining and intuitively explaining my\nproposal. I will compare my proposal with probability theory in section 3. It\nwill turn out that the theory I am proposing is structurally homomorphic to\nprobability theory in important respects and that it is thus equally easily\nimplementable, but moreover computationally simpler. Section 4 contains a very\nbrief comparison with various kinds of logics, in particular conditional logic,\nwith Shackle's functions of potential surprise and related theories, and with\nthe Dempster - Shafer theory of belief functions."", 'The increasing adoption of Cloud-based data processing and storage poses a\nnumber of privacy issues. Users wish to preserve full control over their\nsensitive data and cannot accept it to be fully accessible to an external\nstorage provider. Previous research in this area was mostly addressed at\ntechniques to protect data stored on untrusted database servers; however, I\nargue that the Cloud architecture presents a number of specific problems and\nissues. This dissertation contains a detailed analysis of open issues. To\nhandle them, I present a novel approach where confidential data is stored in a\nhighly distributed partitioned database, partly located on the Cloud and partly\non the clients. In my approach, data can be either private or shared; the\nlatter is shared in a secure manner by means of simple grant-and-revoke\npermissions. I have developed a proof-of-concept implementation using an\nin-memory RDBMS with row-level data encryption in order to achieve fine-grained\ndata access control. This type of approach is rarely adopted in conventional\noutsourced RDBMSs because it requires several complex steps. Benchmarks of my\nproofof-concept implementation show that my approach overcomes most of the\nproblems.', 'We address the problem of solving mixed random linear equations. We have\nunlabeled observations coming from multiple linear regressions, and each\nobservation corresponds to exactly one of the regression models. The goal is to\nlearn the linear regressors from the observations. Classically, Alternating\nMinimization (AM) (which is a variant of Expectation Maximization (EM)) is used\nto solve this problem. AM iteratively alternates between the estimation of\nlabels and solving the regression problems with the estimated labels.\nEmpirically, it is observed that, for a large variety of non-convex problems\nincluding mixed linear regression, AM converges at a much faster rate compared\nto gradient based algorithms. However, the existing theory suggests similar\nrate of convergence for AM and gradient based methods, failing to capture this\nempirical behavior. In this paper, we close this gap between theory and\npractice for the special case of a mixture of $2$ linear regressions. We show\nthat, provided initialized properly, AM enjoys a \\emph{super-linear} rate of\nconvergence in certain parameter regimes. To the best of our knowledge, this is\nthe first work that theoretically establishes such rate for AM. Hence, if we\nwant to recover the unknown regressors upto an error (in $\\ell_2$ norm) of\n$\\epsilon$, AM only takes $\\mathcal{O}(\\log \\log (1/\\epsilon))$ iterations.\nFurthermore, we compare AM with a gradient based heuristic algorithm\nempirically and show that AM dominates in iteration complexity as well as\nwall-clock time.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I aim to develop a comprehensive fake news detection system capable of identifying and curbing the spread of misinformation. The system should be adaptable enough to detect fake news across various domains, such as tracking misinformation related to COVID-19 on social media platforms. Given the multimedia nature of news, the system should incorporate an automatic speech recognition module, an image (video) processing module, and a language model, among other potential components. It would be beneficial to convert audio recognition into transcripts and image and video processing into natural language descriptions, necessitating the design of a robust language model. Additionally, I am keen to explore the dynamics of large-scale information dissemination, which I hope to address by my proposed fake news detection system.","['Dense video captioning is a task of localizing interesting events from an\nuntrimmed video and producing textual description (captions) for each localized\nevent. Most of the previous works in dense video captioning are solely based on\nvisual information and completely ignore the audio track. However, audio, and\nspeech, in particular, are vital cues for a human observer in understanding an\nenvironment. In this paper, we present a new dense video captioning approach\nthat is able to utilize any number of modalities for event description.\nSpecifically, we show how audio and speech modalities may improve a dense video\ncaptioning model. We apply automatic speech recognition (ASR) system to obtain\na temporally aligned textual description of the speech (similar to subtitles)\nand treat it as a separate input alongside video frames and the corresponding\naudio track. We formulate the captioning task as a machine translation problem\nand utilize recently proposed Transformer architecture to convert multi-modal\ninput data into textual descriptions. We demonstrate the performance of our\nmodel on ActivityNet Captions dataset. The ablation studies indicate a\nconsiderable contribution from audio and speech components suggesting that\nthese modalities contain substantial complementary information to video frames.\nFurthermore, we provide an in-depth analysis of the ActivityNet Caption results\nby leveraging the category tags obtained from original YouTube videos. Code is\npublicly available: github.com/v-iashin/MDVC', 'Speech summarization, which generates a text summary from speech, can be\nachieved by combining automatic speech recognition (ASR) and text summarization\n(TS). With this cascade approach, we can exploit state-of-the-art models and\nlarge training datasets for both subtasks, i.e., Transformer for ASR and\nBidirectional Encoder Representations from Transformers (BERT) for TS. However,\nASR errors directly affect the quality of the output summary in the cascade\napproach. We propose a cascade speech summarization model that is robust to ASR\nerrors and that exploits multiple hypotheses generated by ASR to attenuate the\neffect of ASR errors on the summary. We investigate several schemes to combine\nASR hypotheses. First, we propose using the sum of sub-word embedding vectors\nweighted by their posterior values provided by an ASR system as an input to a\nBERT-based TS system. Then, we introduce a more general scheme that uses an\nattention-based fusion module added to a pre-trained BERT module to align and\ncombine several ASR hypotheses. Finally, we perform speech summarization\nexperiments on the How2 dataset and a newly assembled TED-based dataset that we\nwill release with this paper. These experiments show that retraining the\nBERT-based TS system with these schemes can improve summarization performance\nand that the attention-based fusion module is particularly effective.', 'The rapid advancement of technology in online communication via social media\nplatforms has led to a prolific rise in the spread of misinformation and fake\nnews. Fake news is especially rampant in the current COVID-19 pandemic, leading\nto people believing in false and potentially harmful claims and stories.\nDetecting fake news quickly can alleviate the spread of panic, chaos and\npotential health hazards. We developed a two stage automated pipeline for\nCOVID-19 fake news detection using state of the art machine learning models for\nnatural language processing. The first model leverages a novel fact checking\nalgorithm that retrieves the most relevant facts concerning user claims about\nparticular COVID-19 claims. The second model verifies the level of truth in the\nclaim by computing the textual entailment between the claim and the true facts\nretrieved from a manually curated COVID-19 dataset. The dataset is based on a\npublicly available knowledge source consisting of more than 5000 COVID-19 false\nclaims and verified explanations, a subset of which was internally annotated\nand cross-validated to train and evaluate our models. We evaluate a series of\nmodels based on classical text-based features to more contextual Transformer\nbased models and observe that a model pipeline based on BERT and ALBERT for the\ntwo stages respectively yields the best results.', 'Conventional spoken language understanding systems consist of two main\ncomponents: an automatic speech recognition module that converts audio to a\ntranscript, and a natural language understanding module that transforms the\nresulting text (or top N hypotheses) into a set of domains, intents, and\narguments. These modules are typically optimized independently. In this paper,\nwe formulate audio to semantic understanding as a sequence-to-sequence problem\n[1]. We propose and compare various encoder-decoder based approaches that\noptimize both modules jointly, in an end-to-end manner. Evaluations on a\nreal-world task show that 1) having an intermediate text representation is\ncrucial for the quality of the predicted semantics, especially the intent\narguments and 2) jointly optimizing the full system improves overall accuracy\nof prediction. Compared to independently trained models, our best jointly\ntrained model achieves similar domain and intent prediction F1 scores, but\nimproves argument word error rate by 18% relative.', 'A major focus of recent research in spoken language understanding (SLU) has\nbeen on the end-to-end approach where a single model can predict intents\ndirectly from speech inputs without intermediate transcripts. However, this\napproach presents some challenges. First, since speech can be considered as\npersonally identifiable information, in some cases only automatic speech\nrecognition (ASR) transcripts are accessible. Second, intent-labeled speech\ndata is scarce. To address the first challenge, we propose a novel system that\ncan predict intents from flexible types of inputs: speech, ASR transcripts, or\nboth. We demonstrate strong performance for either modality separately, and\nwhen both speech and ASR transcripts are available, through system combination,\nwe achieve better results than using a single input modality. To address the\nsecond challenge, we leverage a semantically robust pre-trained BERT model and\nadopt a cross-modal system that co-trains text embeddings and acoustic\nembeddings in a shared latent space. We further enhance this system by\nutilizing an acoustic module pre-trained on LibriSpeech and domain-adapting the\ntext module on our target datasets. Our experiments show significant advantages\nfor these pre-training and fine-tuning strategies, resulting in a system that\nachieves competitive intent-classification performance on Snips SLU and Fluent\nSpeech Commands datasets.', 'It has been argued that fake news and the spread of false information pose a\nthreat to societies throughout the world, from influencing the results of\nelections to hindering the efforts to manage the COVID-19 pandemic. To combat\nthis threat, a number of Natural Language Processing (NLP) approaches have been\ndeveloped. These leverage a number of datasets, feature extraction/selection\ntechniques and machine learning (ML) algorithms to detect fake news before it\nspreads. While these methods are well-documented, there is less evidence\nregarding their efficacy in this domain. By systematically reviewing the\nliterature, this paper aims to delineate the approaches for fake news detection\nthat are most performant, identify limitations with existing approaches, and\nsuggest ways these can be mitigated. The analysis of the results indicates that\nEnsemble Methods using a combination of news content and socially-based\nfeatures are currently the most effective. Finally, it is proposed that future\nresearch should focus on developing approaches that address generalisability\nissues (which, in part, arise from limitations with current datasets),\nexplainability and bias.', 'Fake news has altered society in negative ways in politics and culture. It\nhas adversely affected both online social network systems as well as offline\ncommunities and conversations. Using automatic machine learning classification\nmodels is an efficient way to combat the widespread dissemination of fake news.\nHowever, a lack of effective, comprehensive datasets has been a problem for\nfake news research and detection model development. Prior fake news datasets do\nnot provide multimodal text and image data, metadata, comment data, and\nfine-grained fake news categorization at the scale and breadth of our dataset.\nWe present Fakeddit, a novel multimodal dataset consisting of over 1 million\nsamples from multiple categories of fake news. After being processed through\nseveral stages of review, the samples are labeled according to 2-way, 3-way,\nand 6-way classification categories through distant supervision. We construct\nhybrid text+image models and perform extensive experiments for multiple\nvariations of classification, demonstrating the importance of the novel aspect\nof multimodality and fine-grained classification unique to Fakeddit.', 'Over the last years, there has been an unprecedented proliferation of fake\nnews. As a consequence, we are more susceptible to the pernicious impact that\nmisinformation and disinformation spreading can have in different segments of\nour society. Thus, the development of tools for automatic detection of fake\nnews plays and important role in the prevention of its negative effects. Most\nattempts to detect and classify false content focus only on using textual\ninformation. Multimodal approaches are less frequent and they typically\nclassify news either as true or fake. In this work, we perform a fine-grained\nclassification of fake news on the Fakeddit dataset, using both unimodal and\nmultimodal approaches. Our experiments show that the multimodal approach based\non a Convolutional Neural Network (CNN) architecture combining text and image\ndata achieves the best results, with an accuracy of 87%. Some fake news\ncategories such as Manipulated content, Satire or False connection strongly\nbenefit from the use of images. Using images also improves the results of the\nother categories, but with less impact. Regarding the unimodal approaches using\nonly text, Bidirectional Encoder Representations from Transformers (BERT) is\nthe best model with an accuracy of 78%. Therefore, exploiting both text and\nimage data significantly improves the performance of fake news detection.', 'Large-scale dissemination of disinformation online intended to mislead or\ndeceive the general population is a major societal problem. Rapid progression\nin image, video, and natural language generative models has only exacerbated\nthis situation and intensified our need for an effective defense mechanism.\nWhile existing approaches have been proposed to defend against neural fake\nnews, they are generally constrained to the very limited setting where articles\nonly have text and metadata such as the title and authors. In this paper, we\nintroduce the more realistic and challenging task of defending against\nmachine-generated news that also includes images and captions. To identify the\npossible weaknesses that adversaries can exploit, we create a NeuralNews\ndataset composed of 4 different types of generated articles as well as conduct\na series of human user study experiments based on this dataset. In addition to\nthe valuable insights gleaned from our user study experiments, we provide a\nrelatively effective approach based on detecting visual-semantic\ninconsistencies, which will serve as an effective first line of defense and a\nuseful reference for future work in defending against machine-generated\ndisinformation.', ""In recent years, misinformation on the Web has become increasingly rampant.\nThe research community has responded by proposing systems and challenges, which\nare beginning to be useful for (various subtasks of) detecting misinformation.\nHowever, most proposed systems are based on deep learning techniques which are\nfine-tuned to specific domains, are difficult to interpret and produce results\nwhich are not machine readable. This limits their applicability and adoption as\nthey can only be used by a select expert audience in very specific settings. In\nthis paper we propose an architecture based on a core concept of Credibility\nReviews (CRs) that can be used to build networks of distributed bots that\ncollaborate for misinformation detection. The CRs serve as building blocks to\ncompose graphs of (i) web content, (ii) existing credibility signals\n--fact-checked claims and reputation reviews of websites--, and (iii)\nautomatically computed reviews. We implement this architecture on top of\nlightweight extensions to Schema.org and services providing generic NLP tasks\nfor semantic similarity and stance detection. Evaluations on existing datasets\nof social-media posts, fake news and political speeches demonstrates several\nadvantages over existing systems: extensibility, domain-independence,\ncomposability, explainability and transparency via provenance. Furthermore, we\nobtain competitive results without requiring finetuning and establish a new\nstate of the art on the Clef'18 CheckThat! Factuality task.""]","['Fake news and misinformation are a matter of concern for people around the\nglobe. Users of the internet and social media sites encounter content with\nfalse information much frequently. Fake news detection is one of the most\nanalyzed and prominent areas of research. These detection techniques apply\npopular machine learning and deep learning algorithms. Previous work in this\ndomain covers fake news detection vastly among text circulating online.\nPlatforms that have extensively been observed and analyzed include news\nwebsites and Twitter. Facebook, Reddit, WhatsApp, YouTube, and other social\napplications are gradually gaining attention in this emerging field.\nResearchers are analyzing online data based on multiple modalities composed of\ntext, image, video, speech, and other contributing factors. The combination of\nvarious modalities has resulted in efficient fake news detection. At present,\nthere is an abundance of surveys consolidating textual fake news detection\nalgorithms. This review primarily deals with multi-modal fake news detection\ntechniques that include images, videos, and their combinations with text. We\nprovide a comprehensive literature survey of eighty articles presenting\nstate-of-the-art detection techniques, thereby identifying research gaps and\nbuilding a pathway for researchers to further advance this domain.', ""With the development of social networks, fake news for various commercial and\npolitical purposes has been appearing in large numbers and gotten widespread in\nthe online world. With deceptive words, people can get infected by the fake\nnews very easily and will share them without any fact-checking. For instance,\nduring the 2016 US president election, various kinds of fake news about the\ncandidates widely spread through both official news media and the online social\nnetworks. These fake news is usually released to either smear the opponents or\nsupport the candidate on their side. The erroneous information in the fake news\nis usually written to motivate the voters' irrational emotion and enthusiasm.\nSuch kinds of fake news sometimes can bring about devastating effects, and an\nimportant goal in improving the credibility of online social networks is to\nidentify the fake news timely. In this paper, we propose to study the fake news\ndetection problem. Automatic fake news identification is extremely hard, since\npure model based fact-checking for news is still an open problem, and few\nexisting models can be applied to solve the problem. With a thorough\ninvestigation of a fake news data, lots of useful explicit features are\nidentified from both the text words and images used in the fake news. Besides\nthe explicit features, there also exist some hidden patterns in the words and\nimages used in fake news, which can be captured with a set of latent features\nextracted via the multiple convolutional layers in our model. A model named as\nTI-CNN (Text and Image information based Convolutinal Neural Network) is\nproposed in this paper. By projecting the explicit and latent features into a\nunified feature space, TI-CNN is trained with both the text and image\ninformation simultaneously. Extensive experiments carried on the real-world\nfake news datasets have demonstrate the effectiveness of TI-CNN."", 'Recent rapid technological advancements in online social networks such as\nTwitter have led to a great incline in spreading false information and fake\nnews. Misinformation is especially prevalent in the ongoing coronavirus disease\n(COVID-19) pandemic, leading to individuals accepting bogus and potentially\ndeleterious claims and articles. Quick detection of fake news can reduce the\nspread of panic and confusion among the public. For our analysis in this paper,\nwe report a methodology to analyze the reliability of information shared on\nsocial media pertaining to the COVID-19 pandemic. Our best approach is based on\nan ensemble of three transformer models (BERT, ALBERT, and XLNET) to detecting\nfake news. This model was trained and evaluated in the context of the\nConstraintAI 2021 shared task COVID19 Fake News Detection in English. Our\nsystem obtained 0.9855 f1-score on testset and ranked 5th among 160 teams.', 'Social media platforms are vulnerable to fake news dissemination, which\ncauses negative consequences such as panic and wrong medication in the\nhealthcare domain. Therefore, it is important to automatically detect fake news\nin an early stage before they get widely spread. This paper analyzes the impact\nof incorporating content information, prior knowledge, and credibility of\nsources into models for the early detection of fake news. We propose a\nframework modeling those features by using BERT language model and external\nsources, namely Simple English Wikipedia and source reliability tags. The\nconducted experiments on CONSTRAINT datasets demonstrated the benefit of\nintegrating these features for the early detection of fake news in the\nhealthcare domain.', 'As single-task accuracy on individual language and image tasks has improved\nsubstantially in the last few years, the long-term goal of a generally skilled\nagent that can both see and talk becomes more feasible to explore. In this\nwork, we focus on leveraging individual language and image tasks, along with\nresources that incorporate both vision and language towards that objective. We\ndesign an architecture that combines state-of-the-art Transformer and ResNeXt\nmodules fed into a novel attentive multimodal module to produce a combined\nmodel trained on many tasks. We provide a thorough analysis of the components\nof the model, and transfer performance when training on one, some, or all of\nthe tasks. Our final models provide a single system that obtains good results\non all vision and language tasks considered, and improves the state-of-the-art\nin image-grounded conversational applications.', 'Massive dissemination of fake news and its potential to erode democracy has\nincreased the demand for accurate fake news detection. Recent advancements in\nthis area have proposed novel techniques that aim to detect fake news by\nexploring how it propagates on social networks. Nevertheless, to detect fake\nnews at an early stage, i.e., when it is published on a news outlet but not yet\nspread on social media, one cannot rely on news propagation information as it\ndoes not exist. Hence, there is a strong need to develop approaches that can\ndetect fake news by focusing on news content. In this paper, a theory-driven\nmodel is proposed for fake news detection. The method investigates news content\nat various levels: lexicon-level, syntax-level, semantic-level and\ndiscourse-level. We represent news at each level, relying on well-established\ntheories in social and forensic psychology. Fake news detection is then\nconducted within a supervised machine learning framework. As an\ninterdisciplinary research, our work explores potential fake news patterns,\nenhances the interpretability in fake news feature engineering, and studies the\nrelationships among fake news, deception/disinformation, and clickbaits.\nExperiments conducted on two real-world datasets indicate the proposed method\ncan outperform the state-of-the-art and enable fake news early detection when\nthere is limited content information.', ""The proliferation of fake news, i.e., news intentionally spread for\nmisinformation, poses a threat to individuals and society. Despite various\nfact-checking websites such as PolitiFact, robust detection techniques are\nrequired to deal with the increase in fake news. Several deep learning models\nshow promising results for fake news classification, however, their black-box\nnature makes it difficult to explain their classification decisions and\nquality-assure the models. We here address this problem by proposing a novel\ninterpretable fake news detection framework based on the recently introduced\nTsetlin Machine (TM). In brief, we utilize the conjunctive clauses of the TM to\ncapture lexical and semantic properties of both true and fake news text.\nFurther, we use the clause ensembles to calculate the credibility of fake news.\nFor evaluation, we conduct experiments on two publicly available datasets,\nPolitiFact and GossipCop, and demonstrate that the TM framework significantly\noutperforms previously published baselines by at least $5\\%$ in terms of\naccuracy, with the added benefit of an interpretable logic-based\nrepresentation. Further, our approach provides higher F1-score than BERT and\nXLNet, however, we obtain slightly lower accuracy. We finally present a case\nstudy on our model's explainability, demonstrating how it decomposes into\nmeaningful words and their negations."", 'The rise in online misinformation in recent years threatens democracies by\ndistorting authentic public discourse and causing confusion, fear, and even, in\nextreme cases, violence. There is a need to understand the spread of false\ncontent through online networks for developing interventions that disrupt\nmisinformation before it achieves virality. Using a Deep Bidirectional\nTransformer for Language Understanding (BERT) and propagation graphs, this\nstudy classifies and visualizes the spread of misinformation on a social media\nnetwork using publicly available Twitter data. The results confirm prior\nresearch around user clusters and the virality of false content while improving\nthe precision of deep learning models for misinformation detection. The study\nfurther demonstrates the suitability of BERT for providing a scalable model for\nfalse information detection, which can contribute to the development of more\ntimely and accurate interventions to slow the spread of misinformation in\nonline environments.', ""In this paper, we introduce UnifiedM2, a general-purpose misinformation model\nthat jointly models multiple domains of misinformation with a single, unified\nsetup. The model is trained to handle four tasks: detecting news bias,\nclickbait, fake news, and verifying rumors. By grouping these tasks together,\nUnifiedM2learns a richer representation of misinformation, which leads to\nstate-of-the-art or comparable performance across all tasks. Furthermore, we\ndemonstrate that UnifiedM2's learned representation is helpful for few-shot\nlearning of unseen misinformation tasks/datasets and model's generalizability\nto unseen events."", 'Sequence-to-sequence (S2S) modeling is becoming a popular paradigm for\nautomatic speech recognition (ASR) because of its ability to jointly optimize\nall the conventional ASR components in an end-to-end (E2E) fashion. This report\ninvestigates the ability of E2E ASR from standard close-talk to far-field\napplications by encompassing entire multichannel speech enhancement and ASR\ncomponents within the S2S model. There have been previous studies on jointly\noptimizing neural beamforming alongside E2E ASR for denoising. It is clear from\nboth recent challenge outcomes and successful products that far-field systems\nwould be incomplete without solving both denoising and dereverberation\nsimultaneously. This report uses a recently developed architecture for\nfar-field ASR by composing neural extensions of dereverberation and beamforming\nmodules with the S2S ASR module as a single differentiable neural network and\nalso clearly defining the role of each subnetwork. The original implementation\nof this architecture was successfully applied to the noisy speech recognition\ntask (CHiME-4), while we applied this implementation to noisy reverberant tasks\n(DIRHA and REVERB). Our investigation shows that the method achieves better\nperformance than conventional pipeline methods on the DIRHA English dataset and\ncomparable performance on the REVERB dataset. It also has additional advantages\nof being neither iterative nor requiring parallel noisy and clean speech data.', 'Social media platforms like Facebook, Twitter, and Instagram have enabled\nconnection and communication on a large scale. It has revolutionized the rate\nat which information is shared and enhanced its reach. However, another side of\nthe coin dictates an alarming story. These platforms have led to an increase in\nthe creation and spread of fake news. The fake news has not only influenced\npeople in the wrong direction but also claimed human lives. During these\ncritical times of the Covid19 pandemic, it is easy to mislead people and make\nthem believe in fatal information. Therefore it is important to curb fake news\nat source and prevent it from spreading to a larger audience. We look at\nautomated techniques for fake news detection from a data mining perspective. We\nevaluate different supervised text classification algorithms on Contraint@AAAI\n2021 Covid-19 Fake news detection dataset. The classification algorithms are\nbased on Convolutional Neural Networks (CNN), Long Short Term Memory (LSTM),\nand Bidirectional Encoder Representations from Transformers (BERT). We also\nevaluate the importance of unsupervised learning in the form of language model\npre-training and distributed word representations using unlabelled covid tweets\ncorpus. We report the best accuracy of 98.41\\% on the Covid-19 Fake news\ndetection dataset.', 'The use of the internet as a fast medium of spreading fake news reinforces\nthe need for computational tools that combat it. Techniques that train fake\nnews classifiers exist, but they all assume an abundance of resources including\nlarge labeled datasets and expert-curated corpora, which low-resource languages\nmay not have. In this work, we make two main contributions: First, we alleviate\nresource scarcity by constructing the first expertly-curated benchmark dataset\nfor fake news detection in Filipino, which we call ""Fake News Filipino.""\nSecond, we benchmark Transfer Learning (TL) techniques and show that they can\nbe used to train robust fake news classifiers from little data, achieving 91%\naccuracy on our fake news dataset, reducing the error by 14% compared to\nestablished few-shot baselines. Furthermore, lifting ideas from multitask\nlearning, we show that augmenting transformer-based transfer techniques with\nauxiliary language modeling losses improves their performance by adapting to\nwriting style. Using this, we improve TL performance by 4-6%, achieving an\naccuracy of 96% on our best model. Lastly, we show that our method generalizes\nwell to different types of news articles, including political news,\nentertainment news, and opinion articles.', ""The proliferation of fake news and its propagation on social media has become\na major concern due to its ability to create devastating impacts. Different\nmachine learning approaches have been suggested to detect fake news. However,\nmost of those focused on a specific type of news (such as political) which\nleads us to the question of dataset-bias of the models used. In this research,\nwe conducted a benchmark study to assess the performance of different\napplicable machine learning approaches on three different datasets where we\naccumulated the largest and most diversified one. We explored a number of\nadvanced pre-trained language models for fake news detection along with the\ntraditional and deep learning ones and compared their performances from\ndifferent aspects for the first time to the best of our knowledge. We find that\nBERT and similar pre-trained models perform the best for fake news detection,\nespecially with very small dataset. Hence, these models are significantly\nbetter option for languages with limited electronic contents, i.e., training\ndata. We also carried out several analysis based on the models' performance,\narticle's topic, article's length, and discussed different lessons learned from\nthem. We believe that this benchmark study will help the research community to\nexplore further and news sites/blogs to select the most appropriate fake news\ndetection method."", 'As the COVID-19 pandemic sweeps across the world, it has been accompanied by\na tsunami of fake news and misinformation on social media. At the time when\nreliable information is vital for public health and safety, COVID-19 related\nfake news has been spreading even faster than the facts. During times such as\nthe COVID-19 pandemic, fake news can not only cause intellectual confusion but\ncan also place lives of people at risk. This calls for an immediate need to\ncontain the spread of such misinformation on social media. We introduce CTF,\nthe first COVID-19 Twitter fake news dataset with labeled genuine and fake\ntweets. Additionally, we propose Cross-SEAN, a cross-stitch based\nsemi-supervised end-to-end neural attention model, which leverages the large\namount of unlabelled data. Cross-SEAN partially generalises to emerging fake\nnews as it learns from relevant external knowledge. We compare Cross-SEAN with\nseven state-of-the-art fake news detection methods. We observe that it achieves\n$0.95$ F1 Score on CTF, outperforming the best baseline by $9\\%$. We also\ndevelop Chrome-SEAN, a Cross-SEAN based chrome extension for real-time\ndetection of fake tweets.', 'The viral spread of fake news has caused great social harm, making fake news\ndetection an urgent task. Current fake news detection methods rely heavily on\ntext information by learning the extracted news content or writing style of\ninternal knowledge. However, deliberate rumors can mask writing style,\nbypassing language models and invalidating simple text-based models. In fact,\nnews articles and other related components (such as news creators and news\ntopics) can be modeled as a heterogeneous information network (HIN for short).\nIn this paper, we propose a novel fake news detection framework, namely\nHierarchical Graph Attention Network(HGAT), which uses a novel hierarchical\nattention mechanism to perform node representation learning in HIN, and then\ndetects fake news by classifying news article nodes. Experiments on two\nreal-world fake news datasets show that HGAT can outperform text-based models\nand other network-based models. In addition, the experiment proved the\nexpandability and generalizability of our for graph representation learning and\nother node classification related applications in heterogeneous graphs.', 'The dissemination of fake news significantly affects personal reputation and\npublic trust. Recently, fake news detection has attracted tremendous attention,\nand previous studies mainly focused on finding clues from news content or\ndiffusion path. However, the required features of previous models are often\nunavailable or insufficient in early detection scenarios, resulting in poor\nperformance. Thus, early fake news detection remains a tough challenge.\nIntuitively, the news from trusted and authoritative sources or shared by many\nusers with a good reputation is more reliable than other news. Using the\ncredibility of publishers and users as prior weakly supervised information, we\ncan quickly locate fake news in massive news and detect them in the early\nstages of dissemination.\n  In this paper, we propose a novel Structure-aware Multi-head Attention\nNetwork (SMAN), which combines the news content, publishing, and reposting\nrelations of publishers and users, to jointly optimize the fake news detection\nand credibility prediction tasks. In this way, we can explicitly exploit the\ncredibility of publishers and users for early fake news detection. We conducted\nexperiments on three real-world datasets, and the results show that SMAN can\ndetect fake news in 4 hours with an accuracy of over 91%, which is much faster\nthan the state-of-the-art models.', 'Social media for news consumption is a double-edged sword. On the one hand,\nits low cost, easy access, and rapid dissemination of information lead people\nto seek out and consume news from social media. On the other hand, it enables\nthe wide spread of ""fake news"", i.e., low quality news with intentionally false\ninformation. The extensive spread of fake news has the potential for extremely\nnegative impacts on individuals and society. Therefore, fake news detection on\nsocial media has recently become an emerging research that is attracting\ntremendous attention. Fake news detection on social media presents unique\ncharacteristics and challenges that make existing detection algorithms from\ntraditional news media ineffective or not applicable. First, fake news is\nintentionally written to mislead readers to believe false information, which\nmakes it difficult and nontrivial to detect based on news content; therefore,\nwe need to include auxiliary information, such as user social engagements on\nsocial media, to help make a determination. Second, exploiting this auxiliary\ninformation is challenging in and of itself as users\' social engagements with\nfake news produce data that is big, incomplete, unstructured, and noisy.\nBecause the issue of fake news detection on social media is both challenging\nand relevant, we conducted this survey to further facilitate research on the\nproblem. In this survey, we present a comprehensive review of detecting fake\nnews on social media, including fake news characterizations on psychology and\nsocial theories, existing algorithms from a data mining perspective, evaluation\nmetrics and representative datasets. We also discuss related research areas,\nopen problems, and future research directions for fake news detection on social\nmedia.', 'Automatic transcriptions of consumer-generated multi-media content such as\n""Youtube"" videos still exhibit high word error rates. Such data typically\noccupies a very broad domain, has been recorded in challenging conditions, with\ncheap hardware and a focus on the visual modality, and may have been\npost-processed or edited. In this paper, we extend our earlier work on adapting\nthe acoustic model of a DNN-based speech recognition system to an RNN language\nmodel and show how both can be adapted to the objects and scenes that can be\nautomatically detected in the video. We are working on a corpus of ""how-to""\nvideos from the web, and the idea is that an object that can be seen (""car""),\nor a scene that is being detected (""kitchen"") can be used to condition both\nmodels on the ""context"" of the recording, thereby reducing perplexity and\nimproving transcription. We achieve good improvements in both cases and compare\nand analyze the respective reductions in word error rate. We expect that our\nresults can be used for any type of speech processing in which ""context""\ninformation is available, for example in robotics, man-machine interaction, or\nwhen indexing large audio-visual archives, and should ultimately help to bring\ntogether the ""video-to-text"" and ""speech-to-text"" communities.', 'Recently, news consumption using online news portals has increased\nexponentially due to several reasons, such as low cost and easy accessibility.\nHowever, such online platforms inadvertently also become the cause of spreading\nfalse information across the web. They are being misused quite frequently as a\nmedium to disseminate misinformation and hoaxes. Such malpractices call for a\nrobust automatic fake news detection system that can keep us at bay from such\nmisinformation and hoaxes. We propose a robust yet simple fake news detection\nsystem, leveraging the tools for paraphrasing, grammar-checking, and\nword-embedding. In this paper, we try to the potential of these tools in\njointly unearthing the authenticity of a news article. Notably, we leverage\nSpinbot (for paraphrasing), Grammarly (for grammar-checking), and GloVe (for\nword-embedding) tools for this purpose. Using these tools, we were able to\nextract novel features that could yield state-of-the-art results on the Fake\nNews AMT dataset and comparable results on Celebrity datasets when combined\nwith some of the essential features. More importantly, the proposed method is\nfound to be more robust empirically than the existing ones, as revealed in our\ncross-domain analysis and multi-domain analysis.', 'Fake News on social media platforms has attracted a lot of attention in\nrecent times, primarily for events related to politics (2016 US Presidential\nelections), healthcare (infodemic during COVID-19), to name a few. Various\nmethods have been proposed for detecting Fake News. The approaches span from\nexploiting techniques related to network analysis, Natural Language Processing\n(NLP), and the usage of Graph Neural Networks (GNNs). In this work, we propose\nDEAP-FAKED, a knowleDgE grAPh FAKe nEws Detection framework for identifying\nFake News. Our approach is a combination of the NLP -- where we encode the news\ncontent, and the GNN technique -- where we encode the Knowledge Graph (KG). A\nvariety of these encodings provides a complementary advantage to our detector.\nWe evaluate our framework using two publicly available datasets containing\narticles from domains such as politics, business, technology, and healthcare.\nAs part of dataset pre-processing, we also remove the bias, such as the source\nof the articles, which could impact the performance of the models. DEAP-FAKED\nobtains an F1-score of 88% and 78% for the two datasets, which is an\nimprovement of 21%, and 3% respectively, which shows the effectiveness of the\napproach.', 'Recent progress in using recurrent neural networks (RNNs) for image\ndescription has motivated the exploration of their application for video\ndescription. However, while images are static, working with videos requires\nmodeling their dynamic temporal structure and then properly integrating that\ninformation into a natural language description. In this context, we propose an\napproach that successfully takes into account both the local and global\ntemporal structure of videos to produce descriptions. First, our approach\nincorporates a spatial temporal 3-D convolutional neural network (3-D CNN)\nrepresentation of the short temporal dynamics. The 3-D CNN representation is\ntrained on video action recognition tasks, so as to produce a representation\nthat is tuned to human motion and behavior. Second we propose a temporal\nattention mechanism that allows to go beyond local temporal modeling and learns\nto automatically select the most relevant temporal segments given the\ntext-generating RNN. Our approach exceeds the current state-of-art for both\nBLEU and METEOR metrics on the Youtube2Text dataset. We also present results on\na new, larger and more challenging dataset of paired video and natural language\ndescriptions.', ""Disinformation through fake news is an ongoing problem in our society and has\nbecome easily spread through social media. The most cost and time effective way\nto filter these large amounts of data is to use a combination of human and\ntechnical interventions to identify it. From a technical perspective, Natural\nLanguage Processing (NLP) is widely used in detecting fake news. Social media\ncompanies use NLP techniques to identify the fake news and warn their users,\nbut fake news may still slip through undetected. It is especially a problem in\nmore localised contexts (outside the United States of America). How do we\nadjust fake news detection systems to work better for local contexts such as in\nSouth Africa. In this work we investigate fake news detection on South African\nwebsites. We curate a dataset of South African fake news and then train\ndetection models. We contrast this with using widely available fake news\ndatasets (from mostly USA website). We also explore making the datasets more\ndiverse by combining them and observe the differences in behaviour in writing\nbetween nations' fake news using interpretable machine learning."", 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', 'Spoken language understanding is typically based on pipeline architectures\nincluding speech recognition and natural language understanding steps. These\ncomponents are optimized independently to allow usage of available data, but\nthe overall system suffers from error propagation. In this paper, we propose a\nnovel training method that enables pretrained contextual embeddings to process\nacoustic features. In particular, we extend it with an encoder of pretrained\nspeech recognition systems in order to construct end-to-end spoken language\nunderstanding systems. Our proposed method is based on the teacher-student\nframework across speech and text modalities that aligns the acoustic and the\nsemantic latent spaces. Experimental results in three benchmarks show that our\nsystem reaches the performance comparable to the pipeline architecture without\nusing any training data and outperforms it after fine-tuning with ten examples\nper class on two out of three benchmarks.', 'During the COVID-19 pandemic, social media platforms were ideal for\ncommunicating due to social isolation and quarantine. Also, it was the primary\nsource of misinformation dissemination on a large scale, referred to as the\ninfodemic. Therefore, automatic debunking misinformation is a crucial problem.\nTo tackle this problem, we present two COVID-19 related misinformation datasets\non Twitter and propose a misinformation detection system comprising\nnetwork-based and content-based processes based on machine learning algorithms\nand NLP techniques. In the network-based process, we focus on social\nproperties, network characteristics, and users. On the other hand, we classify\nmisinformation using the content of the tweets directly in the content-based\nprocess, which contains text classification models (paragraph-level and\nsentence-level) and similarity models. The evaluation results on the\nnetwork-based process show the best results for the artificial neural network\nmodel with an F1 score of 88.68%. In the content-based process, our novel\nsimilarity models, which obtained an F1 score of 90.26%, show an improvement in\nthe misinformation classification results compared to the network-based models.\nIn addition, in the text classification models, the best result was achieved\nusing the stacking ensemble-learning model by obtaining an F1 score of 95.18%.\nFurthermore, we test our content-based models on the Constraint@AAAI2021\ndataset, and by getting an F1 score of 94.38%, we improve the baseline results.\nFinally, we develop a fact-checking website called Checkovid that uses each\nprocess to detect misinformative and informative claims in the domain of\nCOVID-19 from different perspectives.', 'Fake news detection is a very prominent and essential task in the field of\njournalism. This challenging problem is seen so far in the field of politics,\nbut it could be even more challenging when it is to be determined in the\nmulti-domain platform. In this paper, we propose two effective models based on\ndeep learning for solving fake news detection problem in online news contents\nof multiple domains. We evaluate our techniques on the two recently released\ndatasets, namely FakeNews AMT and Celebrity for fake news detection. The\nproposed systems yield encouraging performance, outperforming the current\nhandcrafted feature engineering based state-of-the-art system with a\nsignificant margin of 3.08% and 9.3% by the two models, respectively. In order\nto exploit the datasets, available for the related tasks, we perform\ncross-domain analysis (i.e. model trained on FakeNews AMT and tested on\nCelebrity and vice versa) to explore the applicability of our systems across\nthe domains.', 'The rampant integration of social media in our every day lives and culture\nhas given rise to fast and easier access to the flow of information than ever\nin human history. However, the inherently unsupervised nature of social media\nplatforms has also made it easier to spread false information and fake news.\nFurthermore, the high volume and velocity of information flow in such platforms\nmake manual supervision and control of information propagation infeasible. This\npaper aims to address this issue by proposing a novel deep learning approach\nfor automated detection of false short-text claims on social media. We first\nintroduce Sentimental LIAR, which extends the LIAR dataset of short claims by\nadding features based on sentiment and emotion analysis of claims. Furthermore,\nwe propose a novel deep learning architecture based on the BERT-Base language\nmodel for classification of claims as genuine or fake. Our results demonstrate\nthat the proposed architecture trained on Sentimental LIAR can achieve an\naccuracy of 70%, which is an improvement of ~30% over previously reported\nresults for the LIAR benchmark.', 'Fake news is a severe problem in social media. In this paper, we present an\nempirical study on visual, textual, and multimodal models for the tasks of\nclaim, claim check-worthiness, and conspiracy detection, all of which are\nrelated to fake news detection. Recent work suggests that images are more\ninfluential than text and often appear alongside fake text. To this end,\nseveral multimodal models have been proposed in recent years that use images\nalong with text to detect fake news on social media sites like Twitter.\nHowever, the role of images is not well understood for claim detection,\nspecifically using transformer-based textual and multimodal models. We\ninvestigate state-of-the-art models for images, text (Transformer-based), and\nmultimodal information for four different datasets across two languages to\nunderstand the role of images in the task of claim and conspiracy detection.', 'Both politics and pandemics have recently provided ample motivation for the\ndevelopment of machine learning-enabled disinformation (a.k.a. fake news)\ndetection algorithms. Existing literature has focused primarily on the\nfully-automated case, but the resulting techniques cannot reliably detect\ndisinformation on the varied topics, sources, and time scales required for\nmilitary applications. By leveraging an already-available analyst as a\nhuman-in-the-loop, however, the canonical machine learning techniques of\nsentiment analysis, aspect-based sentiment analysis, and stance detection\nbecome plausible methods to use for a partially-automated disinformation\ndetection system. This paper aims to determine which of these techniques is\nbest suited for this purpose and how each technique might best be used towards\nthis end. Training datasets of the same size and nearly identical neural\narchitectures (a BERT transformer as a word embedder with a single feed-forward\nlayer thereafter) are used for each approach, which are then tested on\nsentiment- and stance-specific datasets to establish a baseline of how well\neach method can be used to do the other tasks. Four different datasets relating\nto COVID-19 disinformation are used to test the ability of each technique to\ndetect disinformation on a topic that did not appear in the training data set.\nQuantitative and qualitative results from these tests are then used to provide\ninsight into how best to employ these techniques in practice.', ""Spoken Language Understanding (SLU) typically comprises of an automatic\nspeech recognition (ASR) followed by a natural language understanding (NLU)\nmodule. The two modules process signals in a blocking sequential fashion, i.e.,\nthe NLU often has to wait for the ASR to finish processing on an utterance\nbasis, potentially leading to high latencies that render the spoken interaction\nless natural. In this paper, we propose recurrent neural network (RNN) based\nincremental processing towards the SLU task of intent detection. The proposed\nmethodology offers lower latencies than a typical SLU system, without any\nsignificant reduction in system accuracy. We introduce and analyze different\nrecurrent neural network architectures for incremental and online processing of\nthe ASR transcripts and compare it to the existing offline systems. A lexical\nEnd-of-Sentence (EOS) detector is proposed for segmenting the stream of\ntranscript into sentences for intent classification. Intent detection\nexperiments are conducted on benchmark ATIS, Snips and Facebook's multilingual\ntask oriented dialog datasets modified to emulate a continuous incremental\nstream of words with no utterance demarcation. We also analyze the prospects of\nearly intent detection, before EOS, with our proposed system."", 'Recent advancements in the area of Computer Vision with state-of-art Neural\nNetworks has given a boost to Optical Character Recognition (OCR) accuracies.\nHowever, extracting characters/text alone is often insufficient for relevant\ninformation extraction as documents also have a visual structure that is not\ncaptured by OCR. Extracting information from tables, charts, footnotes, boxes,\nheadings and retrieving the corresponding structured representation for the\ndocument remains a challenge and finds application in a large number of\nreal-world use cases. In this paper, we propose a novel enterprise based\nend-to-end framework called DeepReader which facilitates information extraction\nfrom document images via identification of visual entities and populating a\nmeta relational model across different entities in the document image. The\nmodel schema allows for an easy to understand abstraction of the entities\ndetected by the deep vision models and the relationships between them.\nDeepReader has a suite of state-of-the-art vision algorithms which are applied\nto recognize handwritten and printed text, eliminate noisy effects, identify\nthe type of documents and detect visual entities like tables, lines and boxes.\nDeep Reader maps the extracted entities into a rich relational schema so as to\ncapture all the relevant relationships between entities (words, textboxes,\nlines etc) detected in the document. Relevant information and fields can then\nbe extracted from the document by writing SQL queries on top of the\nrelationship tables. A natural language based interface is added on top of the\nrelationship schema so that a non-technical user, specifying the queries in\nnatural language, can fetch the information with minimal effort. In this paper,\nwe also demonstrate many different capabilities of Deep Reader and report\nresults on a real-world use case.', ""The authors of fake news often use facts from verified news sources and mix\nthem with misinformation to create confusion and provoke unrest among the\nreaders. The spread of fake news can thereby have serious implications on our\nsociety. They can sway political elections, push down the stock price or crush\nreputations of corporations or public figures. Several websites have taken on\nthe mission of checking rumors and allegations, but are often not fast enough\nto check the content of all the news being disseminated. Especially social\nmedia websites have offered an easy platform for the fast propagation of\ninformation. Towards limiting fake news from being propagated among social\nmedia users, the task of this year's PAN 2020 challenge lays the focus on the\nfake news spreaders. The aim of the task is to determine whether it is possible\nto discriminate authors that have shared fake news in the past from those that\nhave never done it. In this notebook, we describe our profiling system for the\nfake news detection task on Twitter. For this, we conduct different feature\nextraction techniques and learning experiments from a multilingual perspective,\nnamely English and Spanish. Our final submitted systems use character n-grams\nas features in combination with a linear SVM for English and Logistic\nRegression for the Spanish language. Our submitted models achieve an overall\naccuracy of 73% and 79% on the English and Spanish official test set,\nrespectively. Our experiments show that it is difficult to differentiate\nsolidly fake news spreaders on Twitter from users who share credible\ninformation leaving room for further investigations. Our model ranked 3rd out\nof 72 competitors."", 'In a modern spoken language understanding (SLU) system, the natural language\nunderstanding (NLU) module takes interpretations of a speech from the automatic\nspeech recognition (ASR) module as the input. The NLU module usually uses the\nfirst best interpretation of a given speech in downstream tasks such as domain\nand intent classification. However, the ASR module might misrecognize some\nspeeches and the first best interpretation could be erroneous and noisy. Solely\nrelying on the first best interpretation could make the performance of\ndownstream tasks non-optimal. To address this issue, we introduce a series of\nsimple yet efficient models for improving the understanding of semantics of the\ninput speeches by collectively exploiting the n-best speech interpretations\nfrom the ASR module.', 'Observing the damages that can be done by the rapid propagation of fake news\nin various sectors like politics and finance, automatic identification of fake\nnews using linguistic analysis has drawn the attention of the research\ncommunity. However, such methods are largely being developed for English where\nlow resource languages remain out of the focus. But the risks spawned by fake\nand manipulative news are not confined by languages. In this work, we propose\nan annotated dataset of ~50K news that can be used for building automated fake\nnews detection systems for a low resource language like Bangla. Additionally,\nwe provide an analysis of the dataset and develop a benchmark system with state\nof the art NLP techniques to identify Bangla fake news. To create this system,\nwe explore traditional linguistic features and neural network based methods. We\nexpect this dataset will be a valuable resource for building technologies to\nprevent the spreading of fake news and contribute in research with low resource\nlanguages.', ""Today social media has become the primary source for news. Via social media\nplatforms, fake news travel at unprecedented speeds, reach global audiences and\nput users and communities at great risk. Therefore, it is extremely important\nto detect fake news as early as possible. Recently, deep learning based\napproaches have shown improved performance in fake news detection. However, the\ntraining of such models requires a large amount of labeled data, but manual\nannotation is time-consuming and expensive. Moreover, due to the dynamic nature\nof news, annotated samples may become outdated quickly and cannot represent the\nnews articles on newly emerged events. Therefore, how to obtain fresh and\nhigh-quality labeled samples is the major challenge in employing deep learning\nmodels for fake news detection. In order to tackle this challenge, we propose a\nreinforced weakly-supervised fake news detection framework, i.e., WeFEND, which\ncan leverage users' reports as weak supervision to enlarge the amount of\ntraining data for fake news detection. The proposed framework consists of three\nmain components: the annotator, the reinforced selector and the fake news\ndetector. The annotator can automatically assign weak labels for unlabeled news\nbased on users' reports. The reinforced selector using reinforcement learning\ntechniques chooses high-quality samples from the weakly labeled data and\nfilters out those low-quality ones that may degrade the detector's prediction\nperformance. The fake news detector aims to identify fake news based on the\nnews content. We tested the proposed framework on a large collection of news\narticles published via WeChat official accounts and associated user reports.\nExtensive experiments on this dataset show that the proposed WeFEND model\nachieves the best performance compared with the state-of-the-art methods."", 'The proliferation of misleading information in everyday access media outlets\nsuch as social media feeds, news blogs, and online newspapers have made it\nchallenging to identify trustworthy news sources, thus increasing the need for\ncomputational tools able to provide insights into the reliability of online\ncontent. In this paper, we focus on the automatic identification of fake\ncontent in online news. Our contribution is twofold. First, we introduce two\nnovel datasets for the task of fake news detection, covering seven different\nnews domains. We describe the collection, annotation, and validation process in\ndetail and present several exploratory analysis on the identification of\nlinguistic differences in fake and legitimate news content. Second, we conduct\na set of learning experiments to build accurate fake news detectors. In\naddition, we provide comparative analyses of the automatic and manual\nidentification of fake news.', 'Standard neural machine translation (NMT) is on the assumption of\ndocument-level context independent. Most existing document-level NMT methods\nare satisfied with a smattering sense of brief document-level information,\nwhile this work focuses on exploiting detailed document-level context in terms\nof multiple forms of document embeddings, which is capable of sufficiently\nmodeling deeper and richer document-level context. The proposed document-aware\nNMT is implemented to enhance the Transformer baseline by introducing both\nglobal and local document-level clues on the source end. Experiments show that\nthe proposed method significantly improves the translation performance over\nstrong baselines and other related studies.', ""Fake news is a type of pervasive propaganda that spreads misinformation\nonline, taking advantage of social media's extensive reach to manipulate public\nperception. Over the past three years, fake news has become a focal discussion\npoint in the media due to its impact on the 2016 U.S. presidential election.\nFake news can have severe real-world implications: in 2016, a man walked into a\npizzeria carrying a rifle because he read that Hillary Clinton was harboring\nchildren as sex slaves. This project presents a high accuracy (87%) machine\nlearning classifier that determines the validity of news based on the word\ndistributions and specific linguistic and stylistic differences in the first\nfew sentences of an article. This can help readers identify the validity of an\narticle by looking for specific features in the opening lines aiding them in\nmaking informed decisions. Using a dataset of 2,107 articles from 30 different\nwebsites, this project establishes an understanding of the variations between\nfake and credible news by examining the model, dataset, and features. This\nclassifier appears to use the differences in word distribution, levels of tone\nauthenticity, and frequency of adverbs, adjectives, and nouns. The\ndifferentiation in the features of these articles can be used to improve future\nclassifiers. This classifier can also be further applied directly to browsers\nas a Google Chrome extension or as a filter for social media outlets or news\nwebsites to reduce the spread of misinformation."", 'Recent studies on machine reading comprehension have focused on text-level\nunderstanding but have not yet reached the level of human understanding of the\nvisual layout and content of real-world documents. In this study, we introduce\na new visual machine reading comprehension dataset, named VisualMRC, wherein\ngiven a question and a document image, a machine reads and comprehends texts in\nthe image to answer the question in natural language. Compared with existing\nvisual question answering (VQA) datasets that contain texts in images,\nVisualMRC focuses more on developing natural language understanding and\ngeneration abilities. It contains 30,000+ pairs of a question and an\nabstractive answer for 10,000+ document images sourced from multiple domains of\nwebpages. We also introduce a new model that extends existing\nsequence-to-sequence models, pre-trained with large-scale text corpora, to take\ninto account the visual layout and content of documents. Experiments with\nVisualMRC show that this model outperformed the base sequence-to-sequence\nmodels and a state-of-the-art VQA model. However, its performance is still\nbelow that of humans on most automatic evaluation metrics. The dataset will\nfacilitate research aimed at connecting vision and language understanding.', 'The effective incorporation of cross-utterance information has the potential\nto improve language models (LMs) for automatic speech recognition (ASR). To\nextract more powerful and robust cross-utterance representations for the\nTransformer LM (TLM), this paper proposes the R-TLM which uses hidden states in\na long short-term memory (LSTM) LM. To encode the cross-utterance information,\nthe R-TLM incorporates an LSTM module together with a segment-wise recurrence\nin some of the Transformer blocks. In addition to the LSTM module output, a\nshortcut connection using a fusion layer that bypasses the LSTM module is also\ninvestigated. The proposed system was evaluated on the AMI meeting corpus, the\nEval2000 and the RT03 telephone conversation evaluation sets. The best R-TLM\nachieved 0.9%, 0.6%, and 0.8% absolute WER reductions over the single-utterance\nTLM baseline, and 0.5%, 0.3%, 0.2% absolute WER reductions over a strong\ncross-utterance TLM baseline on the AMI evaluation set, Eval2000 and RT03\nrespectively. Improvements on Eval2000 and RT03 were further supported by\nsignificance tests. R-TLMs were found to have better LM scores on words where\nrecognition errors are more likely to occur. The R-TLM WER can be further\nreduced by interpolation with an LSTM-LM.', 'Fake news has now grown into a big problem for societies and also a major\nchallenge for people fighting disinformation. This phenomenon plagues\ndemocratic elections, reputations of individual persons or organizations, and\nhas negatively impacted citizens, (e.g., during the COVID-19 pandemic in the US\nor Brazil). Hence, developing effective tools to fight this phenomenon by\nemploying advanced Machine Learning (ML) methods poses a significant challenge.\nThe following paper displays the present body of knowledge on the application\nof such intelligent tools in the fight against disinformation. It starts by\nshowing the historical perspective and the current role of fake news in the\ninformation war. Proposed solutions based solely on the work of experts are\nanalysed and the most important directions of the application of intelligent\nsystems in the detection of misinformation sources are pointed out.\nAdditionally, the paper presents some useful resources (mainly datasets useful\nwhen assessing ML solutions for fake news detection) and provides a short\noverview of the most important R&D projects related to this subject. The main\npurpose of this work is to analyse the current state of knowledge in detecting\nfake news; on the one hand to show possible solutions, and on the other hand to\nidentify the main challenges and methodological gaps to motivate future\nresearch.', 'Localizing moments in a longer video via natural language queries is a new,\nchallenging task at the intersection of language and video understanding.\nThough moment localization with natural language is similar to other language\nand vision tasks like natural language object retrieval in images, moment\nlocalization offers an interesting opportunity to model temporal dependencies\nand reasoning in text. We propose a new model that explicitly reasons about\ndifferent temporal segments in a video, and shows that temporal context is\nimportant for localizing phrases which include temporal language. To benchmark\nwhether our model, and other recent video localization models, can effectively\nreason about temporal language, we collect the novel TEMPOral reasoning in\nvideo and language (TEMPO) dataset. Our dataset consists of two parts: a\ndataset with real videos and template sentences (TEMPO - Template Language)\nwhich allows for controlled studies on temporal language, and a human language\ndataset which consists of temporal sentences annotated by humans (TEMPO - Human\nLanguage).', 'Nowadays, social network platforms have been the prime source for people to\nexperience news and events due to their capacities to spread information\nrapidly, which inevitably provides a fertile ground for the dissemination of\nfake news. Thus, it is significant to detect fake news otherwise it could cause\npublic misleading and panic. Existing deep learning models have achieved great\nprogress to tackle the problem of fake news detection. However, training an\neffective deep learning model usually requires a large amount of labeled news,\nwhile it is expensive and time-consuming to provide sufficient labeled news in\nactual applications. To improve the detection performance of fake news, we take\nadvantage of the event correlations of news and propose an event correlation\nfiltering method (ECFM) for fake news detection, mainly consisting of the news\ncharacterizer, the pseudo label annotator, the event credibility updater, and\nthe news entropy selector. The news characterizer is responsible for extracting\ntextual features from news, which cooperates with the pseudo label annotator to\nassign pseudo labels for unlabeled news by fully exploiting the event\ncorrelations of news. In addition, the event credibility updater employs\nadaptive Kalman filter to weaken the credibility fluctuations of events. To\nfurther improve the detection performance, the news entropy selector\nautomatically discovers high-quality samples from pseudo labeled news by\nquantifying their news entropy. Finally, ECFM is proposed to integrate them to\ndetect fake news in an event correlation filtering manner. Extensive\nexperiments prove that the explainable introduction of the event correlations\nof news is beneficial to improve the detection performance of fake news.', 'The increasing occurrence, forms, and negative effects of misinformation on\nsocial media platforms has necessitated more misinformation detection tools.\nCurrently, work is being done addressing COVID-19 misinformation however, there\nare no misinformation detection tools for any of the 40 distinct indigenous\nUgandan languages. This paper addresses this gap by presenting basic language\nresources and a misinformation detection data set based on code-mixed\nLuganda-English messages sourced from the Facebook and Twitter social media\nplatforms. Several machine learning methods are applied on the misinformation\ndetection data set to develop classification models for detecting whether a\ncode-mixed Luganda-English message contains misinformation or not. A 10-fold\ncross validation evaluation of the classification methods in an experimental\nmisinformation detection task shows that a Discriminative Multinomial Naive\nBayes (DMNB) method achieves the highest accuracy and F-measure of 78.19% and\n77.90% respectively. Also, Support Vector Machine and Bagging ensemble\nclassification models achieve comparable results. These results are promising\nsince the machine learning models are based on n-gram features from only the\nmisinformation detection dataset.', 'In many natural language processing (NLP) tasks, a document is commonly\nmodeled as a bag of words using the term frequency-inverse document frequency\n(TF-IDF) vector. One major shortcoming of the frequency-based TF-IDF feature\nvector is that it ignores word orders that carry syntactic and semantic\nrelationships among the words in a document, and they can be important in some\nNLP tasks such as genre classification. This paper proposes a novel distributed\nvector representation of a document: a simple recurrent-neural-network language\nmodel (RNN-LM) or a long short-term memory RNN language model (LSTM-LM) is\nfirst created from all documents in a task; some of the LM parameters are then\nadapted by each document, and the adapted parameters are vectorized to\nrepresent the document. The new document vectors are labeled as DV-RNN and\nDV-LSTM respectively. We believe that our new document vectors can capture some\nhigh-level sequential information in the documents, which other current\ndocument representations fail to capture. The new document vectors were\nevaluated in the genre classification of documents in three corpora: the Brown\nCorpus, the BNC Baby Corpus and an artificially created Penn Treebank dataset.\nTheir classification performances are compared with the performance of TF-IDF\nvector and the state-of-the-art distributed memory model of paragraph vector\n(PV-DM). The results show that DV-LSTM significantly outperforms TF-IDF and\nPV-DM in most cases, and combinations of the proposed document vectors with\nTF-IDF or PV-DM may further improve performance.', 'Cross-document co-reference resolution (CDCR) is the task of identifying and\nlinking mentions to entities and concepts across many text documents. Current\nstate-of-the-art models for this task assume that all documents are of the same\ntype (e.g. news articles) or fall under the same theme. However, it is also\ndesirable to perform CDCR across different domains (type or theme). A\nparticular use case we focus on in this paper is the resolution of entities\nmentioned across scientific work and newspaper articles that discuss them.\nIdentifying the same entities and corresponding concepts in both scientific\narticles and news can help scientists understand how their work is represented\nin mainstream media. We propose a new task and English language dataset for\ncross-document cross-domain co-reference resolution (CD$^2$CR). The task aims\nto identify links between entities across heterogeneous document types. We show\nthat in this cross-domain, cross-document setting, existing CDCR models do not\nperform well and we provide a baseline model that outperforms current\nstate-of-the-art CDCR models on CD$^2$CR. Our data set, annotation tool and\nguidelines as well as our model for cross-document cross-domain co-reference\nare all supplied as open access open source resources.', 'The significance of social media has increased manifold in the past few\ndecades as it helps people from even the most remote corners of the world to\nstay connected. With the advent of technology, digital media has become more\nrelevant and widely used than ever before and along with this, there has been a\nresurgence in the circulation of fake news and tweets that demand immediate\nattention. In this paper, we describe a novel Fake News Detection system that\nautomatically identifies whether a news item is ""real"" or ""fake"", as an\nextension of our work in the CONSTRAINT COVID-19 Fake News Detection in English\nchallenge. We have used an ensemble model consisting of pre-trained models\nfollowed by a statistical feature fusion network , along with a novel heuristic\nalgorithm by incorporating various attributes present in news items or tweets\nlike source, username handles, URL domains and authors as statistical feature.\nOur proposed framework have also quantified reliable predictive uncertainty\nalong with proper class output confidence level for the classification task. We\nhave evaluated our results on the COVID-19 Fake News dataset and FakeNewsNet\ndataset to show the effectiveness of the proposed algorithm on detecting fake\nnews in short news content as well as in news articles. We obtained a best\nF1-score of 0.9892 on the COVID-19 dataset, and an F1-score of 0.9073 on the\nFakeNewsNet dataset.', ""Current methods for learning visually grounded language from videos often\nrely on text annotation, such as human generated captions or machine generated\nautomatic speech recognition (ASR) transcripts. In this work, we introduce the\nAudio-Video Language Network (AVLnet), a self-supervised network that learns a\nshared audio-visual embedding space directly from raw video inputs. To\ncircumvent the need for text annotation, we learn audio-visual representations\nfrom randomly segmented video clips and their raw audio waveforms. We train\nAVLnet on HowTo100M, a large corpus of publicly available instructional videos,\nand evaluate on image retrieval and video retrieval tasks, achieving\nstate-of-the-art performance. We perform analysis of AVLnet's learned\nrepresentations, showing our model utilizes speech and natural sounds to learn\naudio-visual concepts. Further, we propose a tri-modal model that jointly\nprocesses raw audio, video, and text captions from videos to learn a\nmulti-modal semantic embedding space useful for text-video retrieval. Our code,\ndata, and trained models will be released at avlnet.csail.mit.edu"", 'The evolution of electronic media is a mixed blessing. Due to the easy\naccess, low cost, and faster reach of the information, people search out and\ndevour news from online social networks. In contrast, the increasing acceptance\nof social media reporting leads to the spread of fake news. This is a minacious\nproblem that causes disputes and endangers societal stability and harmony. Fake\nnews spread has gained attention from researchers due to its vicious nature.\nproliferation of misinformation in all media, from the internet to cable news,\npaid advertising and local news outlets, has made it essential for people to\nidentify the misinformation and sort through the facts. Researchers are trying\nto analyze the credibility of information and curtail false information on such\nplatforms. Credibility is the believability of the piece of information at\nhand. Analyzing the credibility of fake news is challenging due to the intent\nof its creation and the polychromatic nature of the news. In this work, we\npropose a model for detecting fake news. Our method investigates the content of\nthe news at the early stage i.e. when the news is published but is yet to be\ndisseminated through social media. Our work interprets the content with\nautomatic feature extraction and the relevance of the text pieces. In summary,\nwe introduce stance as one of the features along with the content of the\narticle and employ the pre-trained contextualized word embeddings BERT to\nobtain the state-of-art results for fake news detection. The experiment\nconducted on the real-world dataset indicates that our model outperforms the\nprevious work and enables fake news detection with an accuracy of 95.32%.', 'With the rapid evolution of social media, fake news has become a significant\nsocial problem, which cannot be addressed in a timely manner using manual\ninvestigation. This has motivated numerous studies on automating fake news\ndetection. Most studies explore supervised training models with different\nmodalities (e.g., text, images, and propagation networks) of news records to\nidentify fake news. However, the performance of such techniques generally drops\nif news records are coming from different domains (e.g., politics,\nentertainment), especially for domains that are unseen or rarely-seen during\ntraining. As motivation, we empirically show that news records from different\ndomains have significantly different word usage and propagation patterns.\nFurthermore, due to the sheer volume of unlabelled news records, it is\nchallenging to select news records for manual labelling so that the\ndomain-coverage of the labelled dataset is maximized. Hence, this work: (1)\nproposes a novel framework that jointly preserves domain-specific and\ncross-domain knowledge in news records to detect fake news from different\ndomains; and (2) introduces an unsupervised technique to select a set of\nunlabelled informative news records for manual labelling, which can be\nultimately used to train a fake news detection model that performs well for\nmany domains while minimizing the labelling cost. Our experiments show that the\nintegration of the proposed fake news model and the selective annotation\napproach achieves state-of-the-art performance for cross-domain news datasets,\nwhile yielding notable improvements for rarely-appearing domains in news\ndatasets.', ""The proliferation of fake news on social media has opened up new directions\nof research for timely identification and containment of fake news, and\nmitigation of its widespread impact on public opinion. While much of the\nearlier research was focused on identification of fake news based on its\ncontents or by exploiting users' engagements with the news on social media,\nthere has been a rising interest in proactive intervention strategies to\ncounter the spread of misinformation and its impact on society. In this survey,\nwe describe the modern-day problem of fake news and, in particular, highlight\nthe technical challenges associated with it. We discuss existing methods and\ntechniques applicable to both identification and mitigation, with a focus on\nthe significant advances in each method and their advantages and limitations.\nIn addition, research has often been limited by the quality of existing\ndatasets and their specific application contexts. To alleviate this problem, we\ncomprehensively compile and summarize characteristic features of available\ndatasets. Furthermore, we outline new directions of research to facilitate\nfuture development of effective and interdisciplinary solutions."", 'In general, self help systems are being increasingly deployed by service\nbased industries because they are capable of delivering better customer service\nand increasingly the switch is to voice based self help systems because they\nprovide a natural interface for a human to interact with a machine. A speech\nbased self help system ideally needs a speech recognition engine to convert\nspoken speech to text and in addition a language processing engine to take care\nof any misrecognitions by the speech recognition engine. Any off-the-shelf\nspeech recognition engine is generally a combination of acoustic processing and\nspeech grammar. While this is the norm, we believe that ideally a speech\nrecognition application should have in addition to a speech recognition engine\na separate language processing engine to give the system better performance. In\nthis paper, we discuss ways in which the speech recognition engine and the\nlanguage processing engine can be combined to give a better user experience.', 'Many users turn to document retrieval systems (e.g. search engines) to seek\nanswers to controversial questions. Answering such user queries usually require\nidentifying responses within web documents, and aggregating the responses based\non their different perspectives.\n  Classical document retrieval systems fall short at delivering a set of direct\nand diverse responses to the users. Naturally, identifying such responses\nwithin a document is a natural language understanding task. In this paper, we\nexamine the challenges of synthesizing such language understanding objectives\nwith document retrieval, and study a new perspective-oriented document\nretrieval paradigm. We discuss and assess the inherent natural language\nunderstanding challenges in order to achieve the goal. Following the design\nchallenges and principles, we demonstrate and evaluate a practical prototype\npipeline system. We use the prototype system to conduct a user survey in order\nto assess the utility of our paradigm, as well as understanding the user\ninformation needs for controversial queries.', ""Fake news can significantly misinform people who often rely on online sources\nand social media for their information. Current research on fake news detection\nhas mostly focused on analyzing fake news content and how it propagates on a\nnetwork of users. In this paper, we emphasize the detection of fake news by\nassessing its credibility. By analyzing public fake news data, we show that\ninformation on news sources (and authors) can be a strong indicator of\ncredibility. Our findings suggest that an author's history of association with\nfake news, and the number of authors of a news article, can play a significant\nrole in detecting fake news. Our approach can help improve traditional fake\nnews detection methods, wherein content features are often used to detect fake\nnews."", 'In this paper, we introduce Iterative Text Summarization (ITS), an\niteration-based model for supervised extractive text summarization, inspired by\nthe observation that it is often necessary for a human to read an article\nmultiple times in order to fully understand and summarize its contents. Current\nsummarization approaches read through a document only once to generate a\ndocument representation, resulting in a sub-optimal representation. To address\nthis issue we introduce a model which iteratively polishes the document\nrepresentation on many passes through the document. As part of our model, we\nalso introduce a selective reading mechanism that decides more accurately the\nextent to which each sentence in the model should be updated. Experimental\nresults on the CNN/DailyMail and DUC2002 datasets demonstrate that our model\nsignificantly outperforms state-of-the-art extractive systems when evaluated by\nmachines and by humans.', ""This paper presents the Multilingual COVID-19 Analysis Method (CMTA) for\ndetecting and observing the spread of misinformation about this disease within\ntexts. CMTA proposes a data science (DS) pipeline that applies machine learning\nmodels for processing, classifying (Dense-CNN) and analyzing (MBERT)\nmultilingual (micro)-texts. DS pipeline data preparation tasks extract features\nfrom multilingual textual data and categorize it into specific information\nclasses (i.e., 'false', 'partly false', 'misleading'). The CMTA pipeline has\nbeen experimented with multilingual micro-texts (tweets), showing\nmisinformation spread across different languages. To assess the performance of\nCMTA and put it in perspective, we performed a comparative analysis of CMTA\nwith eight monolingual models used for detecting misinformation. The comparison\nshows that CMTA has surpassed various monolingual models and suggests that it\ncan be used as a general method for detecting misinformation in multilingual\nmicro-texts. CMTA experimental results show misinformation trends about\nCOVID-19 in different languages during the first pandemic months."", 'Fake news detection is a critical yet challenging problem in Natural Language\nProcessing (NLP). The rapid rise of social networking platforms has not only\nyielded a vast increase in information accessibility but has also accelerated\nthe spread of fake news. Thus, the effect of fake news has been growing,\nsometimes extending to the offline world and threatening public safety. Given\nthe massive amount of Web content, automatic fake news detection is a practical\nNLP problem useful to all online content providers, in order to reduce the\nhuman time and effort to detect and prevent the spread of fake news. In this\npaper, we describe the challenges involved in fake news detection and also\ndescribe related tasks. We systematically review and compare the task\nformulations, datasets and NLP solutions that have been developed for this\ntask, and also discuss the potentials and limitations of them. Based on our\ninsights, we outline promising research directions, including more\nfine-grained, detailed, fair, and practical detection models. We also highlight\nthe difference between fake news detection and other related tasks, and the\nimportance of NLP solutions for fake news detection.', 'WhatsApp emerged as a major communication platform in many countries in the\nrecent years. Despite offering only one-to-one and small group conversations,\nWhatsApp has been shown to enable the formation of a rich underlying network,\ncrossing the boundaries of existing groups, and with structural properties that\nfavor information dissemination at large. Indeed, WhatsApp has reportedly been\nused as a forum of misinformation campaigns with significant social, political\nand economic consequences in several countries. In this article, we aim at\ncomplementing recent studies on misinformation spread on WhatsApp, mostly\nfocused on content properties and propagation dynamics, by looking into the\nnetwork that connects users sharing the same piece of content. Specifically, we\npresent a hierarchical network-oriented characterization of the users engaged\nin misinformation spread by focusing on three perspectives: individuals,\nWhatsApp groups and user communities, i.e., groupings of users who,\nintentionally or not, share the same content disproportionately often. By\nanalyzing sharing and network topological properties, our study offers valuable\ninsights into how WhatsApp users leverage the underlying network connecting\ndifferent groups to gain large reach in the spread of misinformation on the\nplatform.', ""News recommendation is an effective information dissemination solution in\nmodern society. While recent years have witnessed many promising news\nrecommendation models, they mostly capture the user-news interactions on the\ndocument-level in a static manner. However, in real-world scenarios, the news\ncan be quite complex and diverse, blindly squeezing all the contents into an\nembedding vector can be less effective in extracting information compatible\nwith the personalized preference of the users. In addition, user preferences in\nthe news recommendation scenario can be highly dynamic, and a tailored dynamic\nmechanism should be designed for better recommendation performance. In this\npaper, we propose a novel dynamic news recommender model. For better\nunderstanding the news content, we leverage the attention mechanism to\nrepresent the news from the sentence-, element- and document-levels,\nrespectively. For capturing users' dynamic preferences, the continuous time\ninformation is seamlessly incorporated into the computing of the attention\nweights. More specifically, we design a hierarchical attention network, where\nthe lower layer learns the importance of different sentences and elements, and\nthe upper layer captures the correlations between the previously interacted and\nthe target news. To comprehensively model the dynamic characters, we firstly\nenhance the traditional attention mechanism by incorporating both absolute and\nrelative time information, and then we propose a dynamic negative sampling\nmethod to optimize the users' implicit feedback. We conduct extensive\nexperiments based on three real-world datasets to demonstrate our model's\neffectiveness. Our source code and pre-trained representations are available at\nhttps://github.com/lshowway/D-HAN."", 'Spoken language understanding system is traditionally designed as a pipeline\nof a number of components. First, the audio signal is processed by an automatic\nspeech recognizer for transcription or n-best hypotheses. With the recognition\nresults, a natural language understanding system classifies the text to\nstructured data as domain, intent and slots for down-streaming consumers, such\nas dialog system, hands-free applications. These components are usually\ndeveloped and optimized independently. In this paper, we present our study on\nan end-to-end learning system for spoken language understanding. With this\nunified approach, we can infer the semantic meaning directly from audio\nfeatures without the intermediate text representation. This study showed that\nthe trained model can achieve reasonable good result and demonstrated that the\nmodel can capture the semantic attention directly from the audio features.', 'Attention-based sequence-to-sequence models for automatic speech recognition\njointly train an acoustic model, language model, and alignment mechanism. Thus,\nthe language model component is only trained on transcribed audio-text pairs.\nThis leads to the use of shallow fusion with an external language model at\ninference time. Shallow fusion refers to log-linear interpolation with a\nseparately trained language model at each step of the beam search. In this\nwork, we investigate the behavior of shallow fusion across a range of\nconditions: different types of language models, different decoding units, and\ndifferent tasks. On Google Voice Search, we demonstrate that the use of shallow\nfusion with a neural LM with wordpieces yields a 9.1% relative word error rate\nreduction (WERR) over our competitive attention-based sequence-to-sequence\nmodel, obviating the need for second-pass rescoring.', 'The significance of social media has increased manifold in the past few\ndecades as it helps people from even the most remote corners of the world stay\nconnected. With the COVID-19 pandemic raging, social media has become more\nrelevant and widely used than ever before, and along with this, there has been\na resurgence in the circulation of fake news and tweets that demand immediate\nattention. In this paper, we describe our Fake News Detection system that\nautomatically identifies whether a tweet related to COVID-19 is ""real"" or\n""fake"", as a part of CONSTRAINT COVID19 Fake News Detection in English\nchallenge. We have used an ensemble model consisting of pre-trained models that\nhas helped us achieve a joint 8th position on the leader board. We have\nachieved an F1-score of 0.9831 against a top score of 0.9869. Post completion\nof the competition, we have been able to drastically improve our system by\nincorporating a novel heuristic algorithm based on username handles and link\ndomains in tweets fetching an F1-score of 0.9883 and achieving state-of-the art\nresults on the given dataset.', 'Fake news may be intentionally created to promote economic, political and\nsocial interests, and can lead to negative impacts on humans beliefs and\ndecisions. Hence, detection of fake news is an emerging problem that has become\nextremely prevalent during the last few years. Most existing works on this\ntopic focus on manual feature extraction and supervised classification models\nleveraging a large number of labeled (fake or real) articles. In contrast, we\nfocus on content-based detection of fake news articles, while assuming that we\nhave a small amount of labels, made available by manual fact-checkers or\nautomated sources. We argue this is a more realistic setting in the presence of\nmassive amounts of content, most of which cannot be easily factchecked. To that\nend, we represent collections of news articles as multi-dimensional tensors,\nleverage tensor decomposition to derive concise article embeddings that capture\nspatial/contextual information about each news article, and use those\nembeddings to create an article-by-article graph on which we propagate limited\nlabels. Results on three real-world datasets show that our method performs on\npar or better than existing models that are fully supervised, in that we\nachieve better detection accuracy using fewer labels. In particular, our\nproposed method achieves 75.43% of accuracy using only 30% of labels of a\npublic dataset while an SVM-based classifier achieved 67.43%. Furthermore, our\nmethod achieves 70.92% of accuracy in a large dataset using only 2% of labels.', ""In fighting against fake news, many fact-checking systems comprised of\nhuman-based fact-checking sites (e.g., snopes.com and politifact.com) and\nautomatic detection systems have been developed in recent years. However,\nonline users still keep sharing fake news even when it has been debunked. It\nmeans that early fake news detection may be insufficient and we need another\ncomplementary approach to mitigate the spread of misinformation. In this paper,\nwe introduce a novel application of text generation for combating fake news. In\nparticular, we (1) leverage online users named \\emph{fact-checkers}, who cite\nfact-checking sites as credible evidences to fact-check information in public\ndiscourse; (2) analyze linguistic characteristics of fact-checking tweets; and\n(3) propose and build a deep learning framework to generate responses with\nfact-checking intention to increase the fact-checkers' engagement in\nfact-checking activities. Our analysis reveals that the fact-checkers tend to\nrefute misinformation and use formal language (e.g. few swear words and\nInternet slangs). Our framework successfully generates relevant responses, and\noutperforms competing models by achieving up to 30\\% improvements. Our\nqualitative study also confirms that the superiority of our generated responses\ncompared with responses generated from the existing models."", 'As the spread of false information on the internet has increased dramatically\nin recent years, more and more attention is being paid to automated fake news\ndetection. Some fake news detection methods are already quite successful.\nNevertheless, there are still many vulnerabilities in the detection algorithms.\nThe reason for this is that fake news publishers can structure and formulate\ntheir texts in such a way that a detection algorithm does not expose this text\nas fake news. This paper shows that it is possible to automatically attack\nstate-of-the-art models that have been trained to detect Fake News, making\nthese vulnerable. For this purpose, corresponding models were first trained\nbased on a dataset. Then, using Text-Attack, an attempt was made to manipulate\nthe trained models in such a way that previously correctly identified fake news\nwas classified as true news. The results show that it is possible to\nautomatically bypass Fake News detection mechanisms, leading to implications\nconcerning existing policy initiatives.', 'Information Retrieval (IR) models need to deal with two difficult issues,\nvocabulary mismatch and term dependencies. Vocabulary mismatch corresponds to\nthe difficulty of retrieving relevant documents that do not contain exact query\nterms but semantically related terms. Term dependencies refers to the need of\nconsidering the relationship between the words of the query when estimating the\nrelevance of a document. A multitude of solutions has been proposed to solve\neach of these two problems, but no principled model solve both. In parallel, in\nthe last few years, language models based on neural networks have been used to\ncope with complex natural language processing tasks like emotion and paraphrase\ndetection. Although they present good abilities to cope with both term\ndependencies and vocabulary mismatch problems, thanks to the distributed\nrepresentation of words they are based upon, such models could not be used\nreadily in IR, where the estimation of one language model per document (or\nquery) is required. This is both computationally unfeasible and prone to\nover-fitting. Based on a recent work that proposed to learn a generic language\nmodel that can be modified through a set of document-specific parameters, we\nexplore use of new neural network models that are adapted to ad-hoc IR tasks.\nWithin the language model IR framework, we propose and study the use of a\ngeneric language model as well as a document-specific language model. Both can\nbe used as a smoothing component, but the latter is more adapted to the\ndocument at hand and has the potential of being used as a full document\nlanguage model. We experiment with such models and analyze their results on\nTREC-1 to 8 datasets.', 'The blurry line between nefarious fake news and protected-speech satire has\nbeen a notorious struggle for social media platforms. Further to the efforts of\nreducing exposure to misinformation on social media, purveyors of fake news\nhave begun to masquerade as satire sites to avoid being demoted. In this work,\nwe address the challenge of automatically classifying fake news versus satire.\nPrevious work have studied whether fake news and satire can be distinguished\nbased on language differences. Contrary to fake news, satire stories are\nusually humorous and carry some political or social message. We hypothesize\nthat these nuances could be identified using semantic and linguistic cues.\nConsequently, we train a machine learning method using semantic representation,\nwith a state-of-the-art contextual language model, and with linguistic features\nbased on textual coherence metrics. Empirical evaluation attests to the merits\nof our approach compared to the language-based baseline and sheds light on the\nnuances between fake news and satire. As avenues for future work, we consider\nstudying additional linguistic features related to the humor aspect, and\nenriching the data with current news events, to help identify a political or\nsocial message.', 'Fake news, misinformation, and unverifiable facts on social media platforms\npropagate disharmony and affect society, especially when dealing with an\nepidemic like COVID-19. The task of Fake News Detection aims to tackle the\neffects of such misinformation by classifying news items as fake or real. In\nthis paper, we propose a novel approach that improves over the current\nautomatic fake news detection approaches by automatically gathering evidence\nfor each claim. Our approach extracts supporting evidence from the web articles\nand then selects appropriate text to be treated as evidence sets. We use a\npre-trained summarizer on these evidence sets and then use the extracted\nsummary as supporting evidence to aid the classification task. Our experiments,\nusing both machine learning and deep learning-based methods, help perform an\nextensive evaluation of our approach. The results show that our approach\noutperforms the state-of-the-art methods in fake news detection to achieve an\nF1-score of 99.25 over the dataset provided for the CONSTRAINT-2021 Shared\nTask. We also release the augmented dataset, our code and models for any\nfurther research.', 'Fact verification models have enjoyed a fast advancement in the last two\nyears with the development of pre-trained language models like BERT and the\nrelease of large scale datasets such as FEVER. However, the challenging problem\nof fake news detection has not benefited from the improvement of fact\nverification models, which is closely related to fake news detection. In this\npaper, we propose a simple yet effective approach to connect the dots between\nfact verification and fake news detection. Our approach first employs a text\nsummarization model pre-trained on news corpora to summarize the long news\narticle into a short claim. Then we use a fact verification model pre-trained\non the FEVER dataset to detect whether the input news article is real or fake.\nOur approach makes use of the recent success of fact verification models and\nenables zero-shot fake news detection, alleviating the need of large-scale\ntraining data to train fake news detection models. Experimental results on\nFakenewsNet, a benchmark dataset for fake news detection, demonstrate the\neffectiveness of our proposed approach.', ""Disinformation and fake news have posed detrimental effects on individuals\nand society in recent years, attracting broad attention to fake news detection.\nThe majority of existing fake news detection algorithms focus on mining news\ncontent and/or the surrounding exogenous context for discovering deceptive\nsignals; while the endogenous preference of a user when he/she decides to\nspread a piece of fake news or not is ignored. The confirmation bias theory has\nindicated that a user is more likely to spread a piece of fake news when it\nconfirms his/her existing beliefs/preferences. Users' historical, social\nengagements such as posts provide rich information about users' preferences\ntoward news and have great potential to advance fake news detection. However,\nthe work on exploring user preference for fake news detection is somewhat\nlimited. Therefore, in this paper, we study the novel problem of exploiting\nuser preference for fake news detection. We propose a new framework, UPFD,\nwhich simultaneously captures various signals from user preferences by joint\ncontent and graph modeling. Experimental results on real-world datasets\ndemonstrate the effectiveness of the proposed framework. We release our code\nand data as a benchmark for GNN-based fake news detection:\nhttps://github.com/safe-graph/GNN-FakeNews."", 'With the digitization of media, an immense amount of news data has been\ngenerated by online sources, including mainstream media outlets as well as\nsocial networks. However, the ease of production and distribution resulted in\ncirculation of fake news as well as credible, authentic news. The pervasive\ndissemination of fake news has extreme negative impacts on individuals and\nsociety. Therefore, fake news detection has recently become an emerging topic\nas an interdisciplinary research field that is attracting significant attention\nfrom many research disciplines, including social sciences and linguistics. In\nthis study, we propose a method primarily based on lexicons including a scoring\nsystem to facilitate the detection of the fake news in Turkish. We contribute\nto the literature by collecting a novel, large scale, and credible dataset of\nTurkish news, and by constructing the first fake news detection lexicon for\nTurkish.', 'Representation learning is a critical ingredient for natural language\nprocessing systems. Recent Transformer language models like BERT learn powerful\ntextual representations, but these models are targeted towards token- and\nsentence-level training objectives and do not leverage information on\ninter-document relatedness, which limits their document-level representation\npower. For applications on scientific documents, such as classification and\nrecommendation, the embeddings power strong performance on end tasks. We\npropose SPECTER, a new method to generate document-level embedding of\nscientific documents based on pretraining a Transformer language model on a\npowerful signal of document-level relatedness: the citation graph. Unlike\nexisting pretrained language models, SPECTER can be easily applied to\ndownstream applications without task-specific fine-tuning. Additionally, to\nencourage further research on document-level models, we introduce SciDocs, a\nnew evaluation benchmark consisting of seven document-level tasks ranging from\ncitation prediction, to document classification and recommendation. We show\nthat SPECTER outperforms a variety of competitive baselines on the benchmark.', 'Automatically identifying fake news from the Internet is a challenging\nproblem in deception detection tasks. Online news is modified constantly during\nits propagation, e.g., malicious users distort the original truth and make up\nfake news. However, the continuous evolution process would generate\nunprecedented fake news and cheat the original model. We present the Fake News\nEvolution (FNE) dataset: a new dataset tracking the fake news evolution\nprocess. Our dataset is composed of 950 paired data, each of which consists of\narticles representing the three significant phases of the evolution process,\nwhich are the truth, the fake news, and the evolved fake news. We observe the\nfeatures during the evolution and they are the disinformation techniques, text\nsimilarity, top 10 keywords, classification accuracy, parts of speech, and\nsentiment properties.', 'The recently proposed Sequence-to-Sequence (seq2seq) framework advocates\nreplacing complex data processing pipelines, such as an entire automatic speech\nrecognition system, with a single neural network trained in an end-to-end\nfashion. In this contribution, we analyse an attention-based seq2seq speech\nrecognition system that directly transcribes recordings into characters. We\nobserve two shortcomings: overconfidence in its predictions and a tendency to\nproduce incomplete transcriptions when language models are used. We propose\npractical solutions to both problems achieving competitive speaker independent\nword error rates on the Wall Street Journal dataset: without separate language\nmodels we reach 10.6% WER, while together with a trigram language model, we\nreach 6.7% WER.', 'We explore neural language modeling for speech recognition where the context\nspans multiple sentences. Rather than encode history beyond the current\nsentence using a cache of words or document-level features, we focus our study\non the ability of LSTM and Transformer language models to implicitly learn to\ncarry over context across sentence boundaries. We introduce a new architecture\nthat incorporates an attention mechanism into LSTM to combine the benefits of\nrecurrent and attention architectures. We conduct language modeling and speech\nrecognition experiments on the publicly available LibriSpeech corpus. We show\nthat conventional training on a paragraph-level corpus results in significant\nreductions in perplexity compared to training on a sentence-level corpus. We\nalso describe speech recognition experiments using long-span language models in\nsecond-pass re-ranking, and provide insights into the ability of such models to\ntake advantage of context beyond the current sentence.', 'Previous researchers have considered sentiment analysis as a document\nclassification task, in which input documents are classified into predefined\nsentiment classes. Although there are sentences in a document that support\nimportant evidences for sentiment analysis and sentences that do not, they have\ntreated the document as a bag of sentences. In other words, they have not\nconsidered the importance of each sentence in the document. To effectively\ndetermine polarity of a document, each sentence in the document should be dealt\nwith different degrees of importance. To address this problem, we propose a\ndocument-level sentence classification model based on deep neural networks, in\nwhich the importance degrees of sentences in documents are automatically\ndetermined through gate mechanisms. To verify our new sentiment analysis model,\nwe conducted experiments using the sentiment datasets in the four different\ndomains such as movie reviews, hotel reviews, restaurant reviews, and music\nreviews. In the experiments, the proposed model outperformed previous\nstate-of-the-art models that do not consider importance differences of\nsentences in a document. The experimental results show that the importance of\nsentences should be considered in a document-level sentiment classification\ntask.', 'Document AI, or Document Intelligence, is a relatively new research topic\nthat refers to the techniques for automatically reading, understanding, and\nanalyzing business documents. It is an important research direction for natural\nlanguage processing and computer vision. In recent years, the popularity of\ndeep learning technology has greatly advanced the development of Document AI,\nsuch as document layout analysis, visual information extraction, document\nvisual question answering, document image classification, etc. This paper\nbriefly reviews some of the representative models, tasks, and benchmark\ndatasets. Furthermore, we also introduce early-stage heuristic rule-based\ndocument analysis, statistical machine learning algorithms, and deep learning\napproaches especially pre-training methods. Finally, we look into future\ndirections for Document AI research.', ""Existing language models excel at writing from scratch, but many real-world\nscenarios require rewriting an existing document to fit a set of constraints.\nAlthough sentence-level rewriting has been fairly well-studied, little work has\naddressed the challenge of rewriting an entire document coherently. In this\nwork, we introduce the task of document-level targeted content transfer and\naddress it in the recipe domain, with a recipe as the document and a dietary\nrestriction (such as vegan or dairy-free) as the targeted constraint. We\npropose a novel model for this task based on the generative pre-trained\nlanguage model (GPT-2) and train on a large number of roughly-aligned recipe\npairs (https://github.com/microsoft/document-level-targeted-content-transfer).\nBoth automatic and human evaluations show that our model out-performs existing\nmethods by generating coherent and diverse rewrites that obey the constraint\nwhile remaining close to the original document. Finally, we analyze our model's\nrewrites to assess progress toward the goal of making language generation more\nattuned to constraints that are substantive rather than stylistic."", 'Recently, the end-to-end approach has proven its efficacy in monaural\nmulti-speaker speech recognition. However, high word error rates (WERs) still\nprevent these systems from being used in practical applications. On the other\nhand, the spatial information in multi-channel signals has proven helpful in\nfar-field speech recognition tasks. In this work, we propose a novel neural\nsequence-to-sequence (seq2seq) architecture, MIMO-Speech, which extends the\noriginal seq2seq to deal with multi-channel input and multi-channel output so\nthat it can fully model multi-channel multi-speaker speech separation and\nrecognition. MIMO-Speech is a fully neural end-to-end framework, which is\noptimized only via an ASR criterion. It is comprised of: 1) a monaural masking\nnetwork, 2) a multi-source neural beamformer, and 3) a multi-output speech\nrecognition model. With this processing, the input overlapped speech is\ndirectly mapped to text sequences. We further adopted a curriculum learning\nstrategy, making the best use of the training set to improve the performance.\nThe experiments on the spatialized wsj1-2mix corpus show that our model can\nachieve more than 60% WER reduction compared to the single-channel system with\nhigh quality enhanced signals (SI-SDR = 23.1 dB) obtained by the above\nseparation function.', 'Social media has greatly enabled people to participate in online activities\nat an unprecedented rate. However, this unrestricted access also exacerbates\nthe spread of misinformation and fake news online which might cause confusion\nand chaos unless being detected early for its mitigation. Given the rapidly\nevolving nature of news events and the limited amount of annotated data,\nstate-of-the-art systems on fake news detection face challenges due to the lack\nof large numbers of annotated training instances that are hard to come by for\nearly detection. In this work, we exploit multiple weak signals from different\nsources given by user and content engagements (referred to as weak social\nsupervision), and their complementary utilities to detect fake news. We jointly\nleverage the limited amount of clean data along with weak signals from social\nengagements to train deep neural networks in a meta-learning framework to\nestimate the quality of different weak instances. Experiments on realworld\ndatasets demonstrate that the proposed framework outperforms state-of-the-art\nbaselines for early detection of fake news without using any user engagements\nat prediction time.', 'In recent years, the proliferation of so-called ""fake news"" has caused much\ndisruptions in society and weakened the news ecosystem. Therefore, to mitigate\nsuch problems, researchers have developed state-of-the-art models to\nauto-detect fake news on social media using sophisticated data science and\nmachine learning techniques. In this work, then, we ask ""what if adversaries\nattempt to attack such detection models?"" and investigate related issues by (i)\nproposing a novel threat model against fake news detectors, in which\nadversaries can post malicious comments toward news articles to mislead fake\nnews detectors, and (ii) developing MALCOM, an end-to-end adversarial comment\ngeneration framework to achieve such an attack. Through a comprehensive\nevaluation, we demonstrate that about 94% and 93.5% of the time on average\nMALCOM can successfully mislead five of the latest neural detection models to\nalways output targeted real and fake news labels. Furthermore, MALCOM can also\nfool black box fake news detectors to always output real news labels 90% of the\ntime on average. We also compare our attack model with four baselines across\ntwo real-world datasets, not only on attack performance but also on generated\nquality, coherency, transferability, and robustness.', 'The rapid increase in fake news, which causes significant damage to society,\ntriggers many fake news related studies, including the development of fake news\ndetection and fact verification techniques. The resources for these studies are\nmainly available as public datasets taken from Web data. We surveyed 118\ndatasets related to fake news research on a large scale from three\nperspectives: (1) fake news detection, (2) fact verification, and (3) other\ntasks; for example, the analysis of fake news and satire detection. We also\ndescribe in detail their utilization tasks and their characteristics. Finally,\nwe highlight the challenges in the fake news dataset construction and some\nresearch opportunities that address these challenges. Our survey facilitates\nfake news research by helping researchers find suitable datasets without\nreinventing the wheel, and thereby, improves fake news studies in depth.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'Technical documents contain a fair amount of unnatural language, such as\ntables, formulas, pseudo-codes, etc. Unnatural language can be an important\nfactor of confusing existing NLP tools. This paper presents an effective method\nof distinguishing unnatural language from natural language, and evaluates the\nimpact of unnatural language detection on NLP tasks such as document\nclustering. We view this problem as an information extraction task and build a\nmulticlass classification model identifying unnatural language components into\nfour categories. First, we create a new annotated corpus by collecting slides\nand papers in various formats, PPT, PDF, and HTML, where unnatural language\ncomponents are annotated into four categories. We then explore features\navailable from plain text to build a statistical model that can handle any\nformat as long as it is converted into plain text. Our experiments show that\nremoving unnatural language components gives an absolute improvement in\ndocument clustering up to 15%. Our corpus and tool are publicly available.', 'Document-level relation extraction aims to extract relations among multiple\nentity pairs from a document. Previously proposed graph-based or\ntransformer-based models utilize the entities independently, regardless of\nglobal information among relational triples. This paper approaches the problem\nby predicting an entity-level relation matrix to capture local and global\ninformation, parallel to the semantic segmentation task in computer vision.\nHerein, we propose a Document U-shaped Network for document-level relation\nextraction. Specifically, we leverage an encoder module to capture the context\ninformation of entities and a U-shaped segmentation module over the image-style\nfeature map to capture global interdependency among triples. Experimental\nresults show that our approach can obtain state-of-the-art performance on three\nbenchmark datasets DocRED, CDR, and GDA.', 'Cross-lingual document representations enable language understanding in\nmultilingual contexts and allow transfer learning from high-resource to\nlow-resource languages at the document level. Recently large pre-trained\nlanguage models such as BERT, XLM and XLM-RoBERTa have achieved great success\nwhen fine-tuned on sentence-level downstream tasks. It is tempting to apply\nthese cross-lingual models to document representation learning. However, there\nare two challenges: (1) these models impose high costs on long document\nprocessing and thus many of them have strict length limit; (2) model\nfine-tuning requires extra data and computational resources, which is not\npractical in resource-limited settings. In this work, we address these\nchallenges by proposing unsupervised Language-Agnostic Weighted Document\nRepresentations (LAWDR). We study the geometry of pre-trained sentence\nembeddings and leverage it to derive document representations without\nfine-tuning. Evaluated on cross-lingual document alignment, LAWDR demonstrates\ncomparable performance to state-of-the-art models on benchmark datasets.', 'Text alignment finds application in tasks such as citation recommendation and\nplagiarism detection. Existing alignment methods operate at a single,\npredefined level and cannot learn to align texts at, for example, sentence and\ndocument levels. We propose a new learning approach that equips previously\nestablished hierarchical attention encoders for representing documents with a\ncross-document attention component, enabling structural comparisons across\ndifferent levels (document-to-document and sentence-to-document). Our component\nis weakly supervised from document pairs and can align at multiple levels. Our\nevaluation on predicting document-to-document relationships and\nsentence-to-document relationships on the tasks of citation recommendation and\nplagiarism detection shows that our approach outperforms previously established\nhierarchical, attention encoders based on recurrent and transformer\ncontextualization that are unaware of structural correspondence between\ndocuments.', 'Document grounded generation is the task of using the information provided in\na document to improve text generation. This work focuses on two different\ndocument grounded generation tasks: Wikipedia Update Generation task and\nDialogue response generation. Our work introduces two novel adaptations of\nlarge scale pre-trained encoder-decoder models focusing on building context\ndriven representation of the document and enabling specific attention to the\ninformation in the document. Additionally, we provide a stronger BART baseline\nfor these tasks. Our proposed techniques outperform existing methods on both\nautomated (at least 48% increase in BLEU-4 points) and human evaluation for\ncloseness to reference and relevance to the document. Furthermore, we perform\ncomprehensive manual inspection of the generated output and categorize errors\nto provide insights into future directions in modeling these tasks.', ""Neural networks with deep architectures have demonstrated significant\nperformance improvements in computer vision, speech recognition, and natural\nlanguage processing. The challenges in information retrieval (IR), however, are\ndifferent from these other application areas. A common form of IR involves\nranking of documents--or short passages--in response to keyword-based queries.\nEffective IR systems must deal with query-document vocabulary mismatch problem,\nby modeling relationships between different query and document terms and how\nthey indicate relevance. Models should also consider lexical matches when the\nquery contains rare terms--such as a person's name or a product model\nnumber--not seen during training, and to avoid retrieving semantically related\nbut irrelevant results. In many real-life IR tasks, the retrieval involves\nextremely large collections--such as the document index of a commercial Web\nsearch engine--containing billions of documents. Efficient IR methods should\ntake advantage of specialized IR data structures, such as inverted index, to\nefficiently retrieve from large collections. Given an information need, the IR\nsystem also mediates how much exposure an information artifact receives by\ndeciding whether it should be displayed, and where it should be positioned,\namong other results. Exposure-aware IR systems may optimize for additional\nobjectives, besides relevance, such as parity of exposure for retrieved items\nand content publishers. In this thesis, we present novel neural architectures\nand methods motivated by the specific needs and challenges of IR tasks."", 'Today all kind of information is getting digitized and along with all this\ndigitization, the huge archive of various kinds of documents is being digitized\ntoo. We know that, Optical Character Recognition is the method through which,\nnewspapers and other paper documents convert into digital resources. But, it is\na fact that this method works on texts only. As a result, if we try to process\nany document which contains non-textual zones, then we will get garbage texts\nas output. That is why; in order to digitize documents properly they should be\nprepossessed carefully. And while preprocessing, segmenting document in\ndifferent regions according to the category properly is most important. But,\nthe Optical Character Recognition processes available for Bangla language have\nno such algorithm that can categorize a newspaper/book page fully. So we worked\nto decompose a document into its several parts like headlines, sub headlines,\ncolumns, images etc. And if the input is skewed and rotated, then the input was\nalso deskewed and de-rotated. To decompose any Bangla document we found out the\nedges of the input image. Then we find out the horizontal and vertical area of\nevery pixel where it lies in. Later on the input image was cut according to\nthese areas. Then we pick each and every sub image and found out their\nheight-width ratio, line height. Then according to these values the sub images\nwere categorized. To deskew the image we found out the skew angle and de skewed\nthe image according to this angle. To de-rotate the image we used the line\nheight, matra line, pixel ratio of matra line.', 'This paper describes the short-term competition on the Components\nSegmentation Task of Document Photos that was prepared in the context of the\n16th International Conference on Document Analysis and Recognition (ICDAR\n2021). This competition aims to bring together researchers working in the field\nof identification document image processing and provides them a suitable\nbenchmark to compare their techniques on the component segmentation task of\ndocument images. Three challenge tasks were proposed entailing different\nsegmentation assignments to be performed on a provided dataset. The collected\ndata are from several types of Brazilian ID documents, whose personal\ninformation was conveniently replaced. There were 16 participants whose results\nobtained for some or all the three tasks show different rates for the adopted\nmetrics, like Dice Similarity Coefficient ranging from 0.06 to 0.99. Different\nDeep Learning models were applied by the entrants with diverse strategies to\nachieve the best results in each of the tasks. Obtained results show that the\ncurrently applied methods for solving one of the proposed tasks (document\nboundary detection) are already well established. However, for the other two\nchallenge tasks (text zone and handwritten sign detection) research and\ndevelopment of more robust approaches are still required to achieve acceptable\nresults.', 'Abstractive text summarization aims to shorten long text documents into a\nhuman readable form that contains the most important facts from the original\ndocument. However, the level of actual abstraction as measured by novel phrases\nthat do not appear in the source document remains low in existing approaches.\nWe propose two techniques to improve the level of abstraction of generated\nsummaries. First, we decompose the decoder into a contextual network that\nretrieves relevant parts of the source document, and a pretrained language\nmodel that incorporates prior knowledge about language generation. Second, we\npropose a novelty metric that is optimized directly through policy learning to\nencourage the generation of novel phrases. Our model achieves results\ncomparable to state-of-the-art models, as determined by ROUGE scores and human\nevaluations, while achieving a significantly higher level of abstraction as\nmeasured by n-gram overlap with the source document.', 'Long document summarization is an important and hard task in the field of\nnatural language processing. A good performance of the long document\nsummarization reveals the model has a decent understanding of the human\nlanguage. Currently, most researches focus on how to modify the attention\nmechanism of the transformer to achieve a higher ROUGE score. The study of data\npre-processing and post-processing are relatively few. In this paper, we use\ntwo pre-processing methods and a post-processing method and analyze the effect\nof these methods on various long document summarization models.', 'Most existing models for multilingual natural language processing (NLP) treat\nlanguage as a discrete category, and make predictions for either one language\nor the other. In contrast, we propose using continuous vector representations\nof language. We show that these can be learned efficiently with a\ncharacter-based neural language model, and used to improve inference about\nlanguage varieties not seen during training. In experiments with 1303 Bible\ntranslations into 990 different languages, we empirically explore the capacity\nof multilingual language models, and also show that the language vectors\ncapture genetic relationships between languages.', 'We present a contrasting learning approach with data augmentation techniques\nto learn document representations in an unsupervised manner. Inspired by recent\ncontrastive self-supervised learning algorithms used for image and NLP\npretraining, we hypothesize that high-quality document embedding should be\ninvariant to diverse paraphrases that preserve the semantics of the original\ndocument. With different backbones and contrastive learning frameworks, our\nstudy reveals the enormous benefits of contrastive augmentation for document\nrepresentation learning with two additional insights: 1) including data\naugmentation in a contrastive way can substantially improve the embedding\nquality in unsupervised document representation learning, and 2) in general,\nstochastic augmentations generated by simple word-level manipulation work much\nbetter than sentence-level and document-level ones. We plug our method into a\nclassifier and compare it with a broad range of baseline methods on six\nbenchmark datasets. Our method can decrease the classification error rate by up\nto 6.4% over the SOTA approaches on the document classification task, matching\nor even surpassing fully-supervised methods.', 'Sentence embedding is an important research topic in natural language\nprocessing. It is essential to generate a good embedding vector that fully\nreflects the semantic meaning of a sentence in order to achieve an enhanced\nperformance for various natural language processing tasks, such as machine\ntranslation and document classification. Thus far, various sentence embedding\nmodels have been proposed, and their feasibility has been demonstrated through\ngood performances on tasks following embedding, such as sentiment analysis and\nsentence classification. However, because the performances of sentence\nclassification and sentiment analysis can be enhanced by using a simple\nsentence representation method, it is not sufficient to claim that these models\nfully reflect the meanings of sentences based on good performances for such\ntasks. In this paper, inspired by human language recognition, we propose the\nfollowing concept of semantic coherence, which should be satisfied for a good\nsentence embedding method: similar sentences should be located close to each\nother in the embedding space. Then, we propose the Paraphrase-Thought\n(P-thought) model to pursue semantic coherence as much as possible.\nExperimental results on two paraphrase identification datasets (MS COCO and STS\nbenchmark) show that the P-thought models outperform the benchmarked sentence\nembedding methods.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', 'The multi-stream paradigm of audio processing, in which several sources are\nsimultaneously considered, has been an active research area for information\nfusion. Our previous study offered a promising direction within end-to-end\nautomatic speech recognition, where parallel encoders aim to capture diverse\ninformation followed by a stream-level fusion based on attention mechanisms to\ncombine the different views. However, with an increasing number of streams\nresulting in an increasing number of encoders, the previous approach could\nrequire substantial memory and massive amounts of parallel data for joint\ntraining. In this work, we propose a practical two-stage training scheme.\nStage-1 is to train a Universal Feature Extractor (UFE), where encoder outputs\nare produced from a single-stream model trained with all data. Stage-2\nformulates a multi-stream scheme intending to solely train the attention fusion\nmodule using the UFE features and pretrained components from Stage-1.\nExperiments have been conducted on two datasets, DIRHA and AMI, as a\nmulti-stream scenario. Compared with our previous method, this strategy\nachieves relative word error rate reductions of 8.2--32.4%, while consistently\noutperforming several conventional combination methods.', ""The selection of a suitable document representation approach plays a crucial\nrole in the performance of a document clustering task. Being able to pick out\nrepresentative words within a document can lead to substantial improvements in\ndocument clustering. In the case of web documents, the HTML markup that defines\nthe layout of the content provides additional structural information that can\nbe further exploited to identify representative words. In this paper we\nintroduce a fuzzy term weighing approach that makes the most of the HTML\nstructure for document clustering. We set forth and build on the hypothesis\nthat a good representation can take advantage of how humans skim through\ndocuments to extract the most representative words. The authors of web pages\nmake use of HTML tags to convey the most important message of a web page\nthrough page elements that attract the readers' attention, such as page titles\nor emphasized elements. We define a set of criteria to exploit the information\nprovided by these page elements, and introduce a fuzzy combination of these\ncriteria that we evaluate within the context of a web page clustering task. Our\nproposed approach, called Abstract Fuzzy Combination of Criteria (AFCC), can\nadapt to datasets whose features are distributed differently, achieving good\nresults compared to other similar fuzzy logic based approaches and TF-IDF\nacross different datasets."", 'We present a novel model for the problem of ranking a collection of documents\naccording to their semantic similarity to a source (query) document. While the\nproblem of document-to-document similarity ranking has been studied, most\nmodern methods are limited to relatively short documents or rely on the\nexistence of ""ground-truth"" similarity labels. Yet, in most common real-world\ncases, similarity ranking is an unsupervised problem as similarity labels are\nunavailable. Moreover, an ideal model should not be restricted by documents\'\nlength. Hence, we introduce SDR, a self-supervised method for document\nsimilarity that can be applied to documents of arbitrary length. Importantly,\nSDR can be effectively applied to extremely long documents, exceeding the 4,096\nmaximal token limits of Longformer. Extensive evaluations on large document\ndatasets show that SDR significantly outperforms its alternatives across all\nmetrics. To accelerate future research on unlabeled long document similarity\nranking, and as an additional contribution to the community, we herein publish\ntwo human-annotated test sets of long documents similarity evaluation. The SDR\ncode and datasets are publicly available.', 'Conceptual entanglement is a crucial phenomenon in quantum cognition because\nit implies that classical probabilities cannot model non--compositional\nconceptual phenomena. While several psychological experiments have been\ndeveloped to test conceptual entanglement, this has not been explored in the\ncontext of Natural Language Processing. In this paper, we apply the hypothesis\nthat words of a document are traces of the concepts that a person has in mind\nwhen writing the document. Therefore, if these concepts are entangled, we\nshould be able to observe traces of their entanglement in the documents. In\nparticular, we test conceptual entanglement by contrasting language simulations\nwith results obtained from a text corpus. Our analysis indicates that\nconceptual entanglement is strongly linked to the way in which language is\nstructured. We discuss the implications of this finding in the context of\nconceptual modeling and of Natural Language Processing.', ""Segmenting an unordered text document into different sections is a very\nuseful task in many text processing applications like multiple document\nsummarization, question answering, etc. This paper proposes structuring of an\nunordered text document based on the keywords in the document. We test our\napproach on Wikipedia documents using both statistical and predictive methods\nsuch as the TextRank algorithm and Google's USE (Universal Sentence Encoder).\nFrom our experimental results, we show that the proposed model can effectively\nstructure an unordered document into sections."", 'We show that Bayes\' rule provides an effective mechanism for creating\ndocument translation models that can be learned from only parallel sentences\nand monolingual documents---a compelling benefit as parallel documents are not\nalways available. In our formulation, the posterior probability of a candidate\ntranslation is the product of the unconditional (prior) probability of the\ncandidate output document and the ""reverse translation probability"" of\ntranslating the candidate output back into the source language. Our proposed\nmodel uses a powerful autoregressive language model as the prior on target\nlanguage documents, but it assumes that each sentence is translated\nindependently from the target to the source language. Crucially, at test time,\nwhen a source document is observed, the document language model prior induces\ndependencies between the translations of the source sentences in the posterior.\nThe model\'s independence assumption not only enables efficient use of available\ndata, but it additionally admits a practical left-to-right beam-search\nalgorithm for carrying out inference. Experiments show that our model benefits\nfrom using cross-sentence context in the language model, and it outperforms\nexisting document translation approaches.', ""We provide a series of algorithms demonstrating that solutions according to\nthe fundamental game-theoretic solution concept of closed under rational\nbehavior (CURB) sets in two-player, normal-form games can be computed in\npolynomial time (we also discuss extensions to n-player games). First, we\ndescribe an algorithm that identifies all of a player's best responses\nconditioned on the belief that the other player will play from within a given\nsubset of its strategy space. This algorithm serves as a subroutine in a series\nof polynomial-time algorithms for finding all minimal CURB sets, one minimal\nCURB set, and the smallest minimal CURB set in a game. We then show that the\ncomplexity of finding a Nash equilibrium can be exponential only in the size of\na game's smallest CURB set. Related to this, we show that the smallest CURB set\ncan be an arbitrarily small portion of the game, but it can also be arbitrarily\nlarger than the supports of its only enclosed Nash equilibrium. We test our\nalgorithms empirically and find that most commonly studied academic games tend\nto have either very large or very small minimal CURB sets."", 'This paper proposes an algorithm to improve the calculation of confidence\nmeasure for spoken term detection (STD). Given an input query term, the\nalgorithm first calculates a measurement named document ranking weight for each\ndocument in the speech database to reflect its relevance with the query term by\nsumming all the confidence measures of the hypothesized term occurrences in\nthis document. The confidence measure of each term occurrence is then\nre-estimated through linear interpolation with the calculated document ranking\nweight to improve its reliability by integrating document-level information.\nExperiments are conducted on three standard STD tasks for Tamil, Vietnamese and\nEnglish respectively. The experimental results all demonstrate that the\nproposed algorithm achieves consistent improvements over the state-of-the-art\nmethod for confidence measure calculation. Furthermore, this algorithm is still\neffective even if a high accuracy speech recognizer is not available, which\nmakes it applicable for the languages with limited speech resources.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : As a researcher focusing on long text summarization, I am keen to understand the key features that contribute to a high-quality summary. Which metrics would be most useful in assessing the quality of a summary? I intend to explore the correlation between the quality of summarization and certain NLP features, such as word diversity, entropy, and perplexity. I am curious to know if similar studies have been conducted previously. Furthermore, I am interested in examining self-supervised alignment techniques that could assist in aligning the semantic content of a summary with its original text. My plan is to develop a deep summarization model and an alignment model. The alignment model, potentially using an attention mechanism, could help align summaries with longer texts. The quality of this alignment could act as reward signals indicating the quality of the summaries, thereby driving further enhancements in my summarization model. The improved summarization model could, in turn, enhance the quality of the alignment model. This process would be iterative.","[""A quality abstractive summary should not only copy salient source texts as\nsummaries but should also tend to generate new conceptual words to express\nconcrete details. Inspired by the popular pointer generator\nsequence-to-sequence model, this paper presents a concept pointer network for\nimproving these aspects of abstractive summarization. The network leverages\nknowledge-based, context-aware conceptualizations to derive an extended set of\ncandidate concepts. The model then points to the most appropriate choice using\nboth the concept set and original source text. This joint approach generates\nabstractive summaries with higher-level semantic concepts. The training model\nis also optimized in a way that adapts to different data, which is based on a\nnovel method of distantly-supervised learning guided by reference summaries and\ntesting set. Overall, the proposed approach provides statistically significant\nimprovements over several state-of-the-art models on both the DUC-2004 and\nGigaword datasets. A human evaluation of the model's abstractive abilities also\nsupports the quality of the summaries produced within this framework."", 'In a world of proliferating data, the ability to rapidly summarize text is\ngrowing in importance. Automatic summarization of text can be thought of as a\nsequence to sequence problem. Another area of natural language processing that\nsolves a sequence to sequence problem is machine translation, which is rapidly\nevolving due to the development of attention-based encoder-decoder networks.\nThis work applies these modern techniques to abstractive summarization. We\nperform analysis on various attention mechanisms for summarization with the\ngoal of developing an approach and architecture aimed at improving the state of\nthe art. In particular, we modify and optimize a translation model with\nself-attention for generating abstractive sentence summaries. The effectiveness\nof this base model along with attention variants is compared and analyzed in\nthe context of standardized evaluation sets and test metrics. However, we show\nthat these metrics are limited in their ability to effectively score\nabstractive summaries, and propose a new approach based on the intuition that\nan abstractive model requires an abstractive evaluation.', 'Forum discussion data differ in both structure and properties from generic\nform of textual data such as news. Henceforth, summarization techniques should,\nin turn, make use of such differences, and craft models that can benefit from\nthe structural nature of discussion data. In this work, we propose attending to\nthe beginning of a document, to improve the performance of extractive\nsummarization models when applied to forum discussion data. Evaluations\ndemonstrated that with the help of bidirectional attention mechanism, attending\nto the beginning of a document (initial comment/post) in a discussion thread,\ncan introduce a consistent boost in ROUGE scores, as well as introducing a new\nState Of The Art (SOTA) ROUGE scores on the forum discussions dataset.\nAdditionally, we explored whether this hypothesis is extendable to other\ngeneric forms of textual data. We make use of the tendency of introducing\nimportant information early in the text, by attending to the first few\nsentences in generic textual data. Evaluations demonstrated that attending to\nintroductory sentences using bidirectional attention, improves the performance\nof extractive summarization models when even applied to more generic form of\ntextual data.', 'Discussions on Twitter involve participation from different communities with\ndifferent dialects and it is often necessary to summarize a large number of\nposts into a representative sample to provide a synopsis. Yet, any such\nrepresentative sample should sufficiently portray the underlying dialect\ndiversity to present the voices of different participating communities\nrepresenting the dialects. Extractive summarization algorithms perform the task\nof constructing subsets that succinctly capture the topic of any given set of\nposts. However, we observe that there is dialect bias in the summaries\ngenerated by common summarization approaches, i.e., they often return summaries\nthat under-represent certain dialects.\n  The vast majority of existing ""fair"" summarization approaches require\nsocially salient attribute labels (in this case, dialect) to ensure that the\ngenerated summary is fair with respect to the socially salient attribute.\nNevertheless, in many applications, these labels do not exist. Furthermore, due\nto the ever-evolving nature of dialects in social media, it is unreasonable to\nlabel or accurately infer the dialect of every social media post. To correct\nfor the dialect bias, we employ a framework that takes an existing text\nsummarization algorithm as a blackbox and, using a small set of dialect-diverse\nsentences, returns a summary that is relatively more dialect-diverse.\nCrucially, this approach does not need the posts being summarized to have\ndialect labels, ensuring that the diversification process is independent of\ndialect classification/identification models. We show the efficacy of our\napproach on Twitter datasets containing posts written in dialects used by\ndifferent social groups defined by race or gender; in all cases, our approach\nleads to improved dialect diversity compared to standard text summarization\napproaches.', 'Due to the manifold ranking method has a significant effect on the ranking of\nunknown data based on known data by using a weighted network, many researchers\nuse the manifold ranking method to solve the document summarization task.\nHowever, their models only consider the original features but ignore the\nsemantic features of sentences when they construct the weighted networks for\nthe manifold ranking method. To solve this problem, we proposed two improved\nmodels based on the manifold ranking method. One is combining the topic model\nand manifold ranking method (JTMMR) to solve the document summarization task.\nThis model not only uses the original feature, but also uses the semantic\nfeature to represent the document, which can improve the accuracy of the\nmanifold ranking method. The other one is combining the lifelong topic model\nand manifold ranking method (JLTMMR). On the basis of the JTMMR, this model\nadds the constraint of knowledge to improve the quality of the topic. At the\nsame time, we also add the constraint of the relationship between documents to\ndig out a better document semantic features. The JTMMR model can improve the\neffect of the manifold ranking method by using the better semantic feature.\nExperiments show that our models can achieve a better result than other\nbaseline models for multi-document summarization task. At the same time, our\nmodels also have a good performance on the single document summarization task.\nAfter combining with a few basic surface features, our model significantly\noutperforms some model based on deep learning in recent years. After that, we\nalso do an exploring work for lifelong machine learning by analyzing the effect\nof adding feedback. Experiments show that the effect of adding feedback to our\nmodel is significant.', 'Document summarization provides an instrument for faster understanding the\ncollection of text documents and has several real-life applications. With the\ngrowth of online text data, numerous summarization models have been proposed\nrecently. The Sequence-to-Sequence (Seq2Seq) based neural summarization model\nis the most widely used in the summarization field due to its high performance.\nThis is because semantic information and structure information in the text is\nadequately considered when encoding. However, the existing extractive\nsummarization models pay little attention to and use the central topic\ninformation to assist the generation of summaries, which leads to models not\nensuring the generated summary under the primary topic. A lengthy document can\nspan several topics, and a single summary cannot do justice to all the topics.\nTherefore, the key to generating a high-quality summary is determining the\ncentral topic and building a summary based on it, especially for a long\ndocument. We propose a topic-aware encoding for document summarization to deal\nwith this issue. This model effectively combines syntactic-level and\ntopic-level information to build a comprehensive sentence representation.\nSpecifically, a neural topic model is added in the neural-based sentence-level\nrepresentation learning to adequately consider the central topic information\nfor capturing the critical content in the original document. The experimental\nresults on three public datasets show that our model outperforms the\nstate-of-the-art models.', ""An accurate abstractive summary of a document should contain all its salient\ninformation and should be logically entailed by the input document. We improve\nthese important aspects of abstractive summarization via multi-task learning\nwith the auxiliary tasks of question generation and entailment generation,\nwhere the former teaches the summarization model how to look for salient\nquestioning-worthy details, and the latter teaches the model how to rewrite a\nsummary which is a directed-logical subset of the input document. We also\npropose novel multi-task architectures with high-level (semantic)\nlayer-specific sharing across multiple encoder and decoder layers of the three\ntasks, as well as soft-sharing mechanisms (and show performance ablations and\nanalysis examples of each contribution). Overall, we achieve statistically\nsignificant improvements over the state-of-the-art on both the CNN/DailyMail\nand Gigaword datasets, as well as on the DUC-2002 transfer setup. We also\npresent several quantitative and qualitative analysis studies of our model's\nlearned saliency and entailment skills."", 'Extractive summarization aims at selecting a set of indicative sentences from\na source document as a summary that can express the major theme of the\ndocument. A general consensus on extractive summarization is that both\nrelevance and coverage are critical issues to address. The existing methods\ndesigned to model coverage can be characterized by either reducing redundancy\nor increasing diversity in the summary. Maximal margin relevance (MMR) is a\nwidely-cited method since it takes both relevance and redundancy into account\nwhen generating a summary for a given document. In addition to MMR, there is\nonly a dearth of research concentrating on reducing redundancy or increasing\ndiversity for the spoken document summarization task, as far as we are aware.\nMotivated by these observations, two major contributions are presented in this\npaper. First, in contrast to MMR, which considers coverage by reducing\nredundancy, we propose two novel coverage-based methods, which directly\nincrease diversity. With the proposed methods, a set of representative\nsentences, which not only are relevant to the given document but also cover\nmost of the important sub-themes of the document, can be selected\nautomatically. Second, we make a step forward to plug in several\ndocument/sentence representation methods into the proposed framework to further\nenhance the summarization performance. A series of empirical evaluations\ndemonstrate the effectiveness of our proposed methods.', 'Aligning sentences in a reference summary with their counterparts in source\ndocuments was shown as a useful auxiliary summarization task, notably for\ngenerating training data for salience detection. Despite its assessed utility,\nthe alignment step was mostly approached with heuristic unsupervised methods,\ntypically ROUGE-based, and was never independently optimized or evaluated. In\nthis paper, we propose establishing summary-source alignment as an explicit\ntask, while introducing two major novelties: (1) applying it at the more\naccurate proposition span level, and (2) approaching it as a supervised\nclassification task. To that end, we created a novel training dataset for\nproposition-level alignment, derived automatically from available summarization\nevaluation data. In addition, we crowdsourced dev and test datasets, enabling\nmodel development and proper evaluation. Utilizing these data, we present a\nsupervised proposition alignment baseline model, showing improved\nalignment-quality over the unsupervised approach.', ""Current summarization systems yield generic summaries that are disconnected\nfrom users' preferences and expectations. To address this limitation, we\npresent CTRLsum, a novel framework for controllable summarization. Our approach\nenables users to control multiple aspects of generated summaries by interacting\nwith the summarization system through textual input in the form of a set of\nkeywords or descriptive prompts. Using a single unified model, CTRLsum is able\nto achieve a broad scope of summary manipulation at inference time without\nrequiring additional human annotations or pre-defining a set of control aspects\nduring training. We quantitatively demonstrate the effectiveness of our\napproach on three domains of summarization datasets and five control aspects:\n1) entity-centric and 2) length-controllable summarization, 3) contribution\nsummarization on scientific papers, 4) invention purpose summarization on\npatent filings, and 5) question-guided summarization on news articles in a\nreading comprehension setting. Moreover, when used in a standard, uncontrolled\nsummarization setting, CTRLsum achieves state-of-the-art results on the\nCNN/DailyMail dataset. Code and model checkpoints are available at\nhttps://github.com/salesforce/ctrl-sum"", 'A critical point of multi-document summarization (MDS) is to learn the\nrelations among various documents. In this paper, we propose a novel\nabstractive MDS model, in which we represent multiple documents as a\nheterogeneous graph, taking semantic nodes of different granularities into\naccount, and then apply a graph-to-sequence framework to generate summaries.\nMoreover, we employ a neural topic model to jointly discover latent topics that\ncan act as cross-document semantic units to bridge different documents and\nprovide global information to guide the summary generation. Since topic\nextraction can be viewed as a special type of summarization that ""summarizes""\ntexts into a more abstract format, i.e., a topic distribution, we adopt a\nmulti-task learning strategy to jointly train the topic and summarization\nmodule, allowing the promotion of each other. Experimental results on the\nMulti-News dataset demonstrate that our model outperforms previous\nstate-of-the-art MDS models on both Rouge metrics and human evaluation,\nmeanwhile learns high-quality topics.', 'In this paper, we introduce Iterative Text Summarization (ITS), an\niteration-based model for supervised extractive text summarization, inspired by\nthe observation that it is often necessary for a human to read an article\nmultiple times in order to fully understand and summarize its contents. Current\nsummarization approaches read through a document only once to generate a\ndocument representation, resulting in a sub-optimal representation. To address\nthis issue we introduce a model which iteratively polishes the document\nrepresentation on many passes through the document. As part of our model, we\nalso introduce a selective reading mechanism that decides more accurately the\nextent to which each sentence in the model should be updated. Experimental\nresults on the CNN/DailyMail and DUC2002 datasets demonstrate that our model\nsignificantly outperforms state-of-the-art extractive systems when evaluated by\nmachines and by humans.', 'The availability of large-scale datasets has driven the development of neural\nmodels that create summaries from single documents, for generic purposes. When\nusing a summarization system, users often have specific intents with various\nlanguage realizations, which, depending on the information need, can range from\na single keyword to a long narrative composed of multiple questions. Existing\nsummarization systems, however, often either fail to support or act robustly on\nthis query focused summarization task. We introduce LaQSum, the first unified\ntext summarization system that learns Latent Queries from documents for\nabstractive summarization with any existing query forms. Under a deep\ngenerative framework, our system jointly optimizes a latent query model and a\nconditional language model, allowing users to plug-and-play queries of any type\nat test time. Despite learning from only generic summarization data and\nrequiring no further optimization for downstream summarization tasks, our\nsystem robustly outperforms strong comparison systems across summarization\nbenchmarks with different query types, document settings, and target domains.', ""A desirable property of a reference-based evaluation metric that measures the\ncontent quality of a summary is that it should estimate how much information\nthat summary has in common with a reference. Traditional text overlap based\nmetrics such as ROUGE fail to achieve this because they are limited to matching\ntokens, either lexically or via embeddings. In this work, we propose a metric\nto evaluate the content quality of a summary using question-answering (QA).\nQA-based methods directly measure a summary's information overlap with a\nreference, making them fundamentally different than text overlap metrics. We\ndemonstrate the experimental benefits of QA-based metrics through an analysis\nof our proposed metric, QAEval. QAEval out-performs current state-of-the-art\nmetrics on most evaluations using benchmark datasets, while being competitive\non others due to limitations of state-of-the-art models. Through a careful\nanalysis of each component of QAEval, we identify its performance bottlenecks\nand estimate that its potential upper-bound performance surpasses all other\nautomatic metrics, approaching that of the gold-standard Pyramid Method."", 'Much progress has been made in text summarization, fueled by neural\narchitectures using large-scale training corpora. However, in the news domain,\nneural models easily overfit by leveraging position-related features due to the\nprevalence of the inverted pyramid writing style. In addition, there is an\nunmet need to generate a variety of summaries for different users. In this\npaper, we propose a neural framework that can flexibly control summary\ngeneration by introducing a set of sub-aspect functions (i.e. importance,\ndiversity, position). These sub-aspect functions are regulated by a set of\ncontrol codes to decide which sub-aspect to focus on during summary generation.\nWe demonstrate that extracted summaries with minimal position bias is\ncomparable with those generated by standard models that take advantage of\nposition preference. We also show that news summaries generated with a focus on\ndiversity can be more preferred by human raters. These results suggest that a\nmore flexible neural summarization framework providing more control options\ncould be desirable in tailoring to different user preferences, which is useful\nsince it is often impractical to articulate such preferences for different\napplications a priori.', 'The growth of online consumer health questions has led to the necessity for\nreliable and accurate question answering systems. A recent study showed that\nmanual summarization of consumer health questions brings significant\nimprovement in retrieving relevant answers. However, the automatic\nsummarization of long questions is a challenging task due to the lack of\ntraining data and the complexity of the related subtasks, such as the question\nfocus and type recognition. In this paper, we introduce a reinforcement\nlearning-based framework for abstractive question summarization. We propose two\nnovel rewards obtained from the downstream tasks of (i) question-type\nidentification and (ii) question-focus recognition to regularize the question\ngeneration model. These rewards ensure the generation of semantically valid\nquestions and encourage the inclusion of key medical entities/foci in the\nquestion summary. We evaluated our proposed method on two benchmark datasets\nand achieved higher performance over state-of-the-art models. The manual\nevaluation of the summaries reveals that the generated questions are more\ndiverse and have fewer factual inconsistencies than the baseline summaries', 'Recent advances in text autoencoders have significantly improved the quality\nof the latent space, which enables models to generate grammatical and\nconsistent text from aggregated latent vectors. As a successful application of\nthis property, unsupervised opinion summarization models generate a summary by\ndecoding the aggregated latent vectors of inputs. More specifically, they\nperform the aggregation via simple average. However, little is known about how\nthe vector aggregation step affects the generation quality. In this study, we\nrevisit the commonly used simple average approach by examining the latent space\nand generated summaries. We found that text autoencoders tend to generate\noverly generic summaries from simply averaged latent vectors due to an\nunexpected $L_2$-norm shrinkage in the aggregated latent vectors, which we\nrefer to as summary vector degeneration. To overcome this issue, we develop a\nframework Coop, which searches input combinations for the latent vector\naggregation using input-output word overlap. Experimental results show that\nCoop successfully alleviates the summary vector degeneration issue and\nestablishes new state-of-the-art performance on two opinion summarization\nbenchmarks. Code is available at \\url{https://github.com/megagonlabs/coop}.', 'Automatic text summarization aims to cut down readers time and cognitive\neffort by reducing the content of a text document without compromising on its\nessence. Ergo, informativeness is the prime attribute of document summary\ngenerated by an algorithm, and selecting sentences that capture the essence of\na document is the primary goal of extractive document summarization. In this\npaper, we employ Shannon entropy to capture informativeness of sentences. We\nemploy Non-negative Matrix Factorization (NMF) to reveal probability\ndistributions for computing entropy of terms, topics, and sentences in latent\nspace. We present an information theoretic interpretation of the computed\nentropy, which is the bedrock of the proposed E-Summ algorithm, an unsupervised\nmethod for extractive document summarization. The algorithm systematically\napplies information theoretic principle for selecting informative sentences\nfrom important topics in the document. The proposed algorithm is generic and\nfast, and hence amenable to use for summarization of documents in real time.\nFurthermore, it is domain-, collection-independent and agnostic to the language\nof the document. Benefiting from strictly positive NMF factor matrices, E-Summ\nalgorithm is transparent and explainable too. We use standard ROUGE toolkit for\nperformance evaluation of the proposed method on four well known public\ndata-sets. We also perform quantitative assessment of E-Summ summary quality by\ncomputing its semantic similarity w.r.t the original document. Our\ninvestigation reveals that though using NMF and information theoretic approach\nfor document summarization promises efficient, explainable, and language\nindependent text summarization, it needs to be bolstered to match the\nperformance of deep neural methods.', 'Traditional sequence-to-sequence (seq2seq) models and other variations of the\nattention-mechanism such as hierarchical attention have been applied to the\ntext summarization problem. Though there is a hierarchy in the way humans use\nlanguage by forming paragraphs from sentences and sentences from words,\nhierarchical models have usually not worked that much better than their\ntraditional seq2seq counterparts. This effect is mainly because either the\nhierarchical attention mechanisms are too sparse using hard attention or noisy\nusing soft attention. In this paper, we propose a method based on extracting\nthe highlights of a document; a key concept that is conveyed in a few\nsentences. In a typical text summarization dataset consisting of documents that\nare 800 tokens in length (average), capturing long-term dependencies is very\nimportant, e.g., the last sentence can be grouped with the first sentence of a\ndocument to form a summary. LSTMs (Long Short-Term Memory) proved useful for\nmachine translation. However, they often fail to capture long-term dependencies\nwhile modeling long sequences. To address these issues, we have adapted Neural\nSemantic Encoders (NSE) to text summarization, a class of memory-augmented\nneural networks by improving its functionalities and proposed a novel\nhierarchical NSE that outperforms similar previous models significantly. The\nquality of summarization was improved by augmenting linguistic factors, namely\nlemma, and Part-of-Speech (PoS) tags, to each word in the dataset for improved\nvocabulary coverage and generalization. The hierarchical NSE model on factored\ndataset outperformed the state-of-the-art by nearly 4 ROUGE points. We further\ndesigned and used the first GPU-based self-critical Reinforcement Learning\nmodel.', ""Reference-based metrics such as ROUGE or BERTScore evaluate the content\nquality of a summary by comparing the summary to a reference. Ideally, this\ncomparison should measure the summary's information quality by calculating how\nmuch information the summaries have in common. In this work, we analyze the\ntoken alignments used by ROUGE and BERTScore to compare summaries and argue\nthat their scores largely cannot be interpreted as measuring information\noverlap, but rather the extent to which they discuss the same topics. Further,\nwe provide evidence that this result holds true for many other summarization\nevaluation metrics. The consequence of this result is that it means the\nsummarization community has not yet found a reliable automatic metric that\naligns with its research goal, to generate summaries with high-quality\ninformation. Then, we propose a simple and interpretable method of evaluating\nsummaries which does directly measure information overlap and demonstrate how\nit can be used to gain insights into model behavior that could not be provided\nby other methods alone."", 'Text summarization refers to the process that generates a shorter form of\ntext from the source document preserving salient information. Many existing\nworks for text summarization are generally evaluated by using recall-oriented\nunderstudy for gisting evaluation (ROUGE) scores. However, as ROUGE scores are\ncomputed based on n-gram overlap, they do not reflect semantic meaning\ncorrespondences between generated and reference summaries. Because Korean is an\nagglutinative language that combines various morphemes into a word that express\nseveral meanings, ROUGE is not suitable for Korean summarization. In this\npaper, we propose evaluation metrics that reflect semantic meanings of a\nreference summary and the original document, Reference and Document Aware\nSemantic Score (RDASS). We then propose a method for improving the correlation\nof the metrics with human judgment. Evaluation results show that the\ncorrelation with human judgment is significantly higher for our evaluation\nmetrics than for ROUGE scores.', ""Summarizing texts is not a straightforward task. Before even considering text\nsummarization, one should determine what kind of summary is expected. How much\nshould the information be compressed? Is it relevant to reformulate or should\nthe summary stick to the original phrasing? State-of-the-art on automatic text\nsummarization mostly revolves around news articles. We suggest that considering\na wider variety of tasks would lead to an improvement in the field, in terms of\ngeneralization and robustness. We explore meeting summarization: generating\nreports from automatic transcriptions. Our work consists in segmenting and\naligning transcriptions with respect to reports, to get a suitable dataset for\nneural summarization. Using a bootstrapping approach, we provide pre-alignments\nthat are corrected by human annotators, making a validation set against which\nwe evaluate automatic models. This consistently reduces annotators' efforts by\nproviding iteratively better pre-alignment and maximizes the corpus size by\nusing annotations from our automatic alignment models. Evaluation is conducted\non \\publicmeetings, a novel corpus of aligned public meetings. We report\nautomatic alignment and summarization performances on this corpus and show that\nautomatic alignment is relevant for data annotation since it leads to large\nimprovement of almost +4 on all ROUGE scores on the summarization task."", 'As language models become more powerful, training and evaluation are\nincreasingly bottlenecked by the data and metrics used for a particular task.\nFor example, summarization models are often trained to predict human reference\nsummaries and evaluated using ROUGE, but both of these metrics are rough\nproxies for what we really care about -- summary quality. In this work, we show\nthat it is possible to significantly improve summary quality by training a\nmodel to optimize for human preferences. We collect a large, high-quality\ndataset of human comparisons between summaries, train a model to predict the\nhuman-preferred summary, and use that model as a reward function to fine-tune a\nsummarization policy using reinforcement learning. We apply our method to a\nversion of the TL;DR dataset of Reddit posts and find that our models\nsignificantly outperform both human reference summaries and much larger models\nfine-tuned with supervised learning alone. Our models also transfer to CNN/DM\nnews articles, producing summaries nearly as good as the human reference\nwithout any news-specific fine-tuning. We conduct extensive analyses to\nunderstand our human feedback dataset and fine-tuned models We establish that\nour reward model generalizes to new datasets, and that optimizing our reward\nmodel results in better summaries than optimizing ROUGE according to humans. We\nhope the evidence from our paper motivates machine learning researchers to pay\ncloser attention to how their training loss affects the model behavior they\nactually want.', 'The goal of a summary is to concisely state the most important information in\na document. With this principle in mind, we introduce new reference-free\nsummary evaluation metrics that use a pretrained language model to estimate the\ninformation content shared between a document and its summary. These metrics\nare a modern take on the Shannon Game, a method for summary quality scoring\nproposed decades ago, where we replace human annotators with language models.\nWe also view these metrics as an extension of BLANC, a recently proposed\napproach to summary quality measurement based on the performance of a language\nmodel with and without the help of a summary. Using transformer based language\nmodels, we empirically verify that our metrics achieve state-of-the-art\ncorrelation with human judgement of the summary quality dimensions of both\ncoherence and relevance, as well as competitive correlation with human\njudgement of consistency and fluency.', 'Neural network-based approaches have become widespread for abstractive text\nsummarization. Though previously proposed models for abstractive text\nsummarization addressed the problem of repetition of the same contents in the\nsummary, they did not explicitly consider its information structure. One of the\nreasons these previous models failed to account for information structure in\nthe generated summary is that standard datasets include summaries of variable\nlengths, resulting in problems in analyzing information flow, specifically, the\nmanner in which the first sentence is related to the following sentences.\nTherefore, we use a dataset containing summaries with only three bullet points,\nand propose a neural network-based abstractive summarization model that\nconsiders the information structures of the generated summaries. Our\nexperimental results show that the information structure of a summary can be\ncontrolled, thus improving the performance of the overall summarization.', 'Existing multi-document summarization systems usually rely on a specific\nsummarization model (i.e., a summarization method with a specific parameter\nsetting) to extract summaries for different document sets with different\ntopics. However, according to our quantitative analysis, none of the existing\nsummarization models can always produce high-quality summaries for different\ndocument sets, and even a summarization model with good overall performance may\nproduce low-quality summaries for some document sets. On the contrary, a\nbaseline summarization model may produce high-quality summaries for some\ndocument sets. Based on the above observations, we treat the summaries produced\nby different summarization models as candidate summaries, and then explore\ndiscriminative reranking techniques to identify high-quality summaries from the\ncandidates for difference document sets. We propose to extract a set of\ncandidate summaries for each document set based on an ILP framework, and then\nleverage Ranking SVM for summary reranking. Various useful features have been\ndeveloped for the reranking process, including word-level features,\nsentence-level features and summary-level features. Evaluation results on the\nbenchmark DUC datasets validate the efficacy and robustness of our proposed\napproach.', 'In this work, we propose a method for incorporating question-answering (QA)\nsignals into a summarization model. Our method identifies salient noun phrases\n(NPs) in the input document by automatically generating wh-questions that are\nanswered by the NPs and automatically determining whether those questions are\nanswered in the gold summaries. This QA-based signal is incorporated into a\ntwo-stage summarization model which first marks salient NPs in the input\ndocument using a classification model, then conditionally generates a summary.\nOur experiments demonstrate that the models trained using QA-based supervision\ngenerate higher-quality summaries than baseline methods of identifying salient\nspans on benchmark summarization datasets. Further, we show that the content of\nthe generated summaries can be controlled based on which NPs are marked in the\ninput document. Finally, we propose a method of augmenting the training data so\nthe gold summaries are more consistent with the marked input spans used during\ntraining and show how this results in models which learn to better exclude\nunmarked document content.', 'Abstractive summarization at controllable lengths is a challenging task in\nnatural language processing. It is even more challenging for domains where\nlimited training data is available or scenarios in which the length of the\nsummary is not known beforehand. At the same time, when it comes to trusting\nmachine-generated summaries, explaining how a summary was constructed in\nhuman-understandable terms may be critical. We propose Multi-level Summarizer\n(MLS), a supervised method to construct abstractive summaries of a text\ndocument at controllable lengths. The key enabler of our method is an\ninterpretable multi-headed attention mechanism that computes attention\ndistribution over an input document using an array of timestep independent\nsemantic kernels. Each kernel optimizes a human-interpretable syntactic or\nsemantic property. Exhaustive experiments on two low-resource datasets in the\nEnglish language show that MLS outperforms strong baselines by up to 14.70% in\nthe METEOR score. Human evaluation of the summaries also suggests that they\ncapture the key concepts of the document at various length-budgets.', 'Existing approaches to automatic summarization assume that a length limit for\nthe summary is given, and view content selection as an optimization problem to\nmaximize informativeness and minimize redundancy within this budget. This\nframework ignores the fact that human-written summaries have rich internal\nstructure which can be exploited to train a summarization system. We present\nNEXTSUM, a novel approach to summarization based on a model that predicts the\nnext sentence to include in the summary using not only the source article, but\nalso the summary produced so far. We show that such a model successfully\ncaptures summary-specific discourse moves, and leads to better content\nselection performance, in addition to automatically predicting how long the\ntarget summary should be. We perform experiments on the New York Times\nAnnotated Corpus of summaries, where NEXTSUM outperforms lead and content-model\nsummarization baselines by significant margins. We also show that the lengths\nof summaries produced by our system correlates with the lengths of the\nhuman-written gold standards.', 'We propose a simple method for post-processing the outputs of a text\nsummarization system in order to refine its overall quality. Our approach is to\ntrain text-to-text rewriting models to correct information redundancy errors\nthat may arise during summarization. We train on synthetically generated noisy\nsummaries, testing three different types of noise that introduce out-of-context\ninformation within each summary. When applied on top of extractive and\nabstractive summarization baselines, our summary denoising models yield metric\nimprovements while reducing redundancy.', 'We address the problem of unsupervised abstractive summarization of\ncollections of user generated reviews with self-supervision and control. We\npropose a self-supervised setup that considers an individual document as a\ntarget summary for a set of similar documents. This setting makes training\nsimpler than previous approaches by relying only on standard log-likelihood\nloss. We address the problem of hallucinations through the use of control\ncodes, to steer the generation towards more coherent and relevant\nsummaries.Finally, we extend the Transformer architecture to allow for multiple\nreviews as input. Our benchmarks on two datasets against graph-based and recent\nneural abstractive unsupervised models show that our proposed method generates\nsummaries with a superior quality and relevance.This is confirmed in our human\nevaluation which focuses explicitly on the faithfulness of generated summaries\nWe also provide an ablation study, which shows the importance of the control\nsetup in controlling hallucinations and achieve high sentiment and topic\nalignment of the summaries with the input reviews.']","['Abstractive text summarization is a challenging task, and one need to design\na mechanism to effectively extract salient information from the source text and\nthen generate a summary. A parsing process of the source text contains critical\nsyntactic or semantic structures, which is useful to generate more accurate\nsummary. However, modeling a parsing tree for text summarization is not trivial\ndue to its non-linear structure and it is harder to deal with a document that\nincludes multiple sentences and their parsing trees. In this paper, we propose\nto use a graph to connect the parsing trees from the sentences in a document\nand utilize the stacked graph convolutional networks (GCNs) to learn the\nsyntactic representation for a document. The selective attention mechanism is\nused to extract salient information in semantic and structural aspect and\ngenerate an abstractive summary. We evaluate our approach on the CNN/Daily Mail\ntext summarization dataset. The experimental results show that the proposed\nGCNs based selective attention approach outperforms the baselines and achieves\nthe state-of-the-art performance on the dataset.', 'Under special circumstances, summaries should conform to a particular style\nwith patterns, such as court judgments and abstracts in academic papers. To\nthis end, the prototype document-summary pairs can be utilized to generate\nbetter summaries. There are two main challenges in this task: (1) the model\nneeds to incorporate learned patterns from the prototype, but (2) should avoid\ncopying contents other than the patternized words---such as irrelevant\nfacts---into the generated summaries. To tackle these challenges, we design a\nmodel named Prototype Editing based Summary Generator (PESG). PESG first learns\nsummary patterns and prototype facts by analyzing the correlation between a\nprototype document and its summary. Prototype facts are then utilized to help\nextract facts from the input document. Next, an editing generator generates new\nsummary based on the summary pattern or extracted facts. Finally, to address\nthe second challenge, a fact checker is used to estimate mutual information\nbetween the input document and generated summary, providing an additional\nsignal for the generator. Extensive experiments conducted on a large-scale\nreal-world text summarization dataset show that PESG achieves the\nstate-of-the-art performance in terms of both automatic metrics and human\nevaluations.', 'We present a novel divide-and-conquer method for the neural summarization of\nlong documents. Our method exploits the discourse structure of the document and\nuses sentence similarity to split the problem into an ensemble of smaller\nsummarization problems. In particular, we break a long document and its summary\ninto multiple source-target pairs, which are used for training a model that\nlearns to summarize each part of the document separately. These partial\nsummaries are then combined in order to produce a final complete summary. With\nthis approach we can decompose the problem of long document summarization into\nsmaller and simpler problems, reducing computational complexity and creating\nmore training examples, which at the same time contain less noise in the target\nsummaries compared to the standard approach. We demonstrate that this approach\npaired with different summarization models, including sequence-to-sequence RNNs\nand Transformers, can lead to improved summarization performance. Our best\nmodels achieve results that are on par with the state-of-the-art in two two\npublicly available datasets of academic articles.', 'A crucial difference between single- and multi-document summarization is how\nsalient content manifests itself in the document(s). While such content may\nappear at the beginning of a single document, essential information is\nfrequently reiterated in a set of documents related to a particular topic,\nresulting in an endorsement effect that increases information salience. In this\npaper, we model the cross-document endorsement effect and its utilization in\nmultiple document summarization. Our method generates a synopsis from each\ndocument, which serves as an endorser to identify salient content from other\ndocuments. Strongly endorsed text segments are used to enrich a neural\nencoder-decoder model to consolidate them into an abstractive summary. The\nmethod has a great potential to learn from fewer examples to identify salient\ncontent, which alleviates the need for costly retraining when the set of\ndocuments is dynamically adjusted. Through extensive experiments on benchmark\nmulti-document summarization datasets, we demonstrate the effectiveness of our\nproposed method over strong published baselines. Finally, we shed light on\nfuture research directions and discuss broader challenges of this task using a\ncase study.', 'Text summarization aims to compress a textual document to a short summary\nwhile keeping salient information. Extractive approaches are widely used in\ntext summarization because of their fluency and efficiency. However, most of\nexisting extractive models hardly capture inter-sentence relationships,\nparticularly in long documents. They also often ignore the effect of topical\ninformation on capturing important contents. To address these issues, this\npaper proposes a graph neural network (GNN)-based extractive summarization\nmodel, enabling to capture inter-sentence relationships efficiently via\ngraph-structured document representation. Moreover, our model integrates a\njoint neural topic model (NTM) to discover latent topics, which can provide\ndocument-level features for sentence selection. The experimental results\ndemonstrate that our model not only substantially achieves state-of-the-art\nresults on CNN/DM and NYT datasets but also considerably outperforms existing\napproaches on scientific paper datasets consisting of much longer documents,\nindicating its better robustness in document genres and lengths. Further\ndiscussions show that topical information can help the model preselect salient\ncontents from an entire document, which interprets its effectiveness in long\ndocument summarization.', ""Recent neural sequence to sequence models have provided feasible solutions\nfor abstractive summarization. However, such models are still hard to tackle\nlong text dependency in the summarization task. A high-quality summarization\nsystem usually depends on strong encoder which can refine important information\nfrom long input texts so that the decoder can generate salient summaries from\nthe encoder's memory. In this paper, we propose an aggregation mechanism based\non the Transformer model to address the challenge of long text representation.\nOur model can review history information to make encoder hold more memory\ncapacity. Empirically, we apply our aggregation mechanism to the Transformer\nmodel and experiment on CNN/DailyMail dataset to achieve higher quality\nsummaries compared to several strong baseline models on the ROUGE metrics."", 'Text alignment finds application in tasks such as citation recommendation and\nplagiarism detection. Existing alignment methods operate at a single,\npredefined level and cannot learn to align texts at, for example, sentence and\ndocument levels. We propose a new learning approach that equips previously\nestablished hierarchical attention encoders for representing documents with a\ncross-document attention component, enabling structural comparisons across\ndifferent levels (document-to-document and sentence-to-document). Our component\nis weakly supervised from document pairs and can align at multiple levels. Our\nevaluation on predicting document-to-document relationships and\nsentence-to-document relationships on the tasks of citation recommendation and\nplagiarism detection shows that our approach outperforms previously established\nhierarchical, attention encoders based on recurrent and transformer\ncontextualization that are unaware of structural correspondence between\ndocuments.', 'Abstractive summarization for long-document or multi-document remains\nchallenging for the Seq2Seq architecture, as Seq2Seq is not good at analyzing\nlong-distance relations in text. In this paper, we present BASS, a novel\nframework for Boosting Abstractive Summarization based on a unified Semantic\ngraph, which aggregates co-referent phrases distributing across a long range of\ncontext and conveys rich relations between phrases. Further, a graph-based\nencoder-decoder model is proposed to improve both the document representation\nand summary generation process by leveraging the graph structure. Specifically,\nseveral graph augmentation methods are designed to encode both the explicit and\nimplicit relations in the text while the graph-propagation attention mechanism\nis developed in the decoder to select salient content into the summary.\nEmpirical results show that the proposed architecture brings substantial\nimprovements for both long-document and multi-document summarization tasks.', 'Multi-document summaritazion is the process of taking multiple texts as input\nand producing a short summary text based on the content of input texts. Up\nuntil recently, multi-document summarizers are mostly supervised extractive.\nHowever, supervised methods require datasets of large, paired document-summary\nexamples which are rare and expensive to produce. In 2018, an unsupervised\nmulti-document abstractive summarization method(Meansum) was proposed by Chu\nand Liu, and demonstrated competitive performances comparing to extractive\nmethods. Despite good evaluation results on automatic metrics, Meansum has\nmultiple limitations, notably the inability of dealing with multiple aspects.\nThe aim of this work was to use Multi-Aspect Masker(MAM) as content selector to\naddress the issue with multi-aspect. Moreover, we propose a regularizer to\ncontrol the length of the generated summaries. Through a series of experiments\non the hotel dataset from Trip Advisor, we validate our assumption and show\nthat our improved model achieves higher ROUGE, Sentiment Accuracy than the\noriginal Meansum method and also beats/ comprarable/close to the supervised\nbaseline.', 'Developed so far, multi-document summarization has reached its bottleneck due\nto the lack of sufficient training data and diverse categories of documents.\nText classification just makes up for these deficiencies. In this paper, we\npropose a novel summarization system called TCSum, which leverages plentiful\ntext classification data to improve the performance of multi-document\nsummarization. TCSum projects documents onto distributed representations which\nact as a bridge between text classification and summarization. It also utilizes\nthe classification results to produce summaries of different styles. Extensive\nexperiments on DUC generic multi-document summarization datasets show that,\nTCSum can achieve the state-of-the-art performance without using any\nhand-crafted features and has the capability to catch the variations of summary\nstyles with respect to different text categories.', ""Traditional video summarization methods generate fixed video representations\nregardless of user interest. Therefore such methods limit users' expectations\nin content search and exploration scenarios. Multi-modal video summarization is\none of the methods utilized to address this problem. When multi-modal video\nsummarization is used to help video exploration, a text-based query is\nconsidered as one of the main drivers of video summary generation, as it is\nuser-defined. Thus, encoding the text-based query and the video effectively are\nboth important for the task of multi-modal video summarization. In this work, a\nnew method is proposed that uses a specialized attention network and\ncontextualized word representations to tackle this task. The proposed model\nconsists of a contextualized video summary controller, multi-modal attention\nmechanisms, an interactive attention network, and a video summary generator.\nBased on the evaluation of the existing multi-modal video summarization\nbenchmark, experimental results show that the proposed model is effective with\nthe increase of +5.88% in accuracy and +4.06% increase of F1-score, compared\nwith the state-of-the-art method."", ""Neural summarization models suffer from the fixed-size input limitation: if\ntext length surpasses the model's maximal number of input tokens, some document\ncontent (possibly summary-relevant) gets truncated Independently summarizing\nwindows of maximal input size disallows for information flow between windows\nand leads to incoherent summaries. We propose windowing models for neural\nabstractive summarization of (arbitrarily) long texts. We extend the\nsequence-to-sequence model augmented with pointer generator network by (1)\nallowing the encoder to slide over different windows of the input document and\n(2) sharing the decoder and retaining its state across different input windows.\nWe explore two windowing variants: Static Windowing precomputes the number of\ntokens the decoder should generate from each window (based on training corpus\nstatistics); in Dynamic Windowing the decoder learns to emit a token that\nsignals encoder's shift to the next input window. Empirical results render our\nmodels effective in their intended use-case: summarizing long texts with\nrelevant content not bound to the very document beginning."", 'A key challenge for abstractive summarization is ensuring factual consistency\nof the generated summary with respect to the original document. For example,\nstate-of-the-art models trained on existing datasets exhibit entity\nhallucination, generating names of entities that are not present in the source\ndocument. We propose a set of new metrics to quantify the entity-level factual\nconsistency of generated summaries and we show that the entity hallucination\nproblem can be alleviated by simply filtering the training data. In addition,\nwe propose a summary-worthy entity classification task to the training process\nas well as a joint entity and summary generation approach, which yield further\nimprovements in entity level metrics.', 'In this paper, we aim to improve abstractive dialogue summarization quality\nand, at the same time, enable granularity control. Our model has two primary\ncomponents and stages: 1) a two-stage generation strategy that generates a\npreliminary summary sketch serving as the basis for the final summary. This\nsummary sketch provides a weakly supervised signal in the form of\npseudo-labeled interrogative pronoun categories and key phrases extracted using\na constituency parser. 2) A simple strategy to control the granularity of the\nfinal summary, in that our model can automatically determine or control the\nnumber of generated summary sentences for a given dialogue by predicting and\nhighlighting different text spans from the source text. Our model achieves\nstate-of-the-art performance on the largest dialogue summarization corpus\nSAMSum, with as high as 50.79 in ROUGE-L score. In addition, we conduct a case\nstudy and show competitive human evaluation results and controllability to\nhuman-annotated summaries.', ""Recently, various neural encoder-decoder models pioneered by Seq2Seq\nframework have been proposed to achieve the goal of generating more abstractive\nsummaries by learning to map input text to output text. At a high level, such\nneural models can freely generate summaries without any constraint on the words\nor phrases used. Moreover, their format is closer to human-edited summaries and\noutput is more readable and fluent. However, the neural model's abstraction\nability is a double-edged sword. A commonly observed problem with the generated\nsummaries is the distortion or fabrication of factual information in the\narticle. This inconsistency between the original text and the summary has\ncaused various concerns over its applicability, and the previous evaluation\nmethods of text summarization are not suitable for this issue. In response to\nthe above problems, the current research direction is predominantly divided\ninto two categories, one is to design fact-aware evaluation metrics to select\noutputs without factual inconsistency errors, and the other is to develop new\nsummarization systems towards factual consistency. In this survey, we focus on\npresenting a comprehensive review of these fact-specific evaluation methods and\ntext summarization models."", 'Automatic text summarization has been widely studied as an important task in\nnatural language processing. Traditionally, various feature engineering and\nmachine learning based systems have been proposed for extractive as well as\nabstractive text summarization. Recently, deep learning based, specifically\nTransformer-based systems have been immensely popular. Summarization is a\ncognitively challenging task - extracting summary worthy sentences is\nlaborious, and expressing semantics in brief when doing abstractive\nsummarization is complicated. In this paper, we specifically look at the\nproblem of summarizing scientific research papers from multiple domains. We\ndifferentiate between two types of summaries, namely, (a) LaySumm: A very short\nsummary that captures the essence of the research paper in layman terms\nrestricting overtly specific technical jargon and (b) LongSumm: A much longer\ndetailed summary aimed at providing specific insights into various ideas\ntouched upon in the paper. While leveraging latest Transformer-based models,\nour systems are simple, intuitive and based on how specific paper sections\ncontribute to human summaries of the two types described above. Evaluations\nagainst gold standard summaries using ROUGE metrics prove the effectiveness of\nour approach. On blind test corpora, our system ranks first and third for the\nLongSumm and LaySumm tasks respectively.', 'Despite the recent developments on neural summarization systems, the\nunderlying logic behind the improvements from the systems and its\ncorpus-dependency remains largely unexplored. Position of sentences in the\noriginal text, for example, is a well known bias for news summarization.\nFollowing in the spirit of the claim that summarization is a combination of\nsub-functions, we define three sub-aspects of summarization: position,\nimportance, and diversity and conduct an extensive analysis of the biases of\neach sub-aspect with respect to the domain of nine different summarization\ncorpora (e.g., news, academic papers, meeting minutes, movie script, books,\nposts). We find that while position exhibits substantial bias in news articles,\nthis is not the case, for example, with academic papers and meeting minutes.\nFurthermore, our empirical study shows that different types of summarization\nsystems (e.g., neural-based) are composed of different degrees of the\nsub-aspects. Our study provides useful lessons regarding consideration of\nunderlying sub-aspects when collecting a new summarization dataset or\ndeveloping a new system.', 'We present a new summarization task, generating summaries of novel chapters\nusing summary/chapter pairs from online study guides. This is a harder task\nthan the news summarization task, given the chapter length as well as the\nextreme paraphrasing and generalization found in the summaries. We focus on\nextractive summarization, which requires the creation of a gold-standard set of\nextractive summaries. We present a new metric for aligning reference summary\nsentences with chapter sentences to create gold extracts and also experiment\nwith different alignment methods. Our experiments demonstrate significant\nimprovement over prior alignment approaches for our task as shown through\nautomatic metrics and a crowd-sourced pyramid analysis. We make our data\ncollection scripts available at\nhttps://github.com/manestay/novel-chapter-dataset .', 'Automatic video summarization is still an unsolved problem due to several\nchallenges. We take steps towards making automatic video summarization more\nrealistic by addressing them. Firstly, the currently available datasets either\nhave very short videos or have few long videos of only a particular type. We\nintroduce a new benchmarking dataset VISIOCITY which comprises of longer videos\nacross six different categories with dense concept annotations capable of\nsupporting different flavors of video summarization and can be used for other\nvision problems. Secondly, for long videos, human reference summaries are\ndifficult to obtain. We present a novel recipe based on pareto optimality to\nautomatically generate multiple reference summaries from indirect ground truth\npresent in VISIOCITY. We show that these summaries are at par with human\nsummaries. Thirdly, we demonstrate that in the presence of multiple ground\ntruth summaries (due to the highly subjective nature of the task), learning\nfrom a single combined ground truth summary using a single loss function is not\na good idea. We propose a simple recipe VISIOCITY-SUM to enhance an existing\nmodel using a combination of losses and demonstrate that it beats the current\nstate of the art techniques when tested on VISIOCITY. We also show that a\nsingle measure to evaluate a summary, as is the current typical practice, falls\nshort. We propose a framework for better quantitative assessment of summary\nquality which is closer to human judgment than a single measure, say F1. We\nreport the performance of a few representative techniques of video\nsummarization on VISIOCITY assessed using various measures and bring out the\nlimitation of the techniques and/or the assessment mechanism in modeling human\njudgment and demonstrate the effectiveness of our evaluation framework in doing\nso.', ""Exploring the tremendous amount of data efficiently to make a decision,\nsimilar to answering a complicated question, is challenging with many\nreal-world application scenarios. In this context, automatic summarization has\nsubstantial importance as it will provide the foundation for big data analytic.\nTraditional summarization approaches optimize the system to produce a short\nstatic summary that fits all users that do not consider the subjectivity aspect\nof summarization, i.e., what is deemed valuable for different users, making\nthese approaches impractical in real-world use cases. This paper proposes an\ninteractive concept-based summarization model, called Adaptive Summaries, that\nhelps users make their desired summary instead of producing a single inflexible\nsummary. The system learns from users' provided information gradually while\ninteracting with the system by giving feedback in an iterative loop. Users can\nchoose either reject or accept action for selecting a concept being included in\nthe summary with the importance of that concept from users' perspectives and\nconfidence level of their feedback. The proposed approach can guarantee\ninteractive speed to keep the user engaged in the process. Furthermore, it\neliminates the need for reference summaries, which is a challenging issue for\nsummarization tasks. Evaluations show that Adaptive Summaries helps users make\nhigh-quality summaries based on their preferences by maximizing the\nuser-desired content in the generated summaries."", 'In this paper we present VideoSET, a method for Video Summary Evaluation\nthrough Text that can evaluate how well a video summary is able to retain the\nsemantic information contained in its original video. We observe that semantics\nis most easily expressed in words, and develop a text-based approach for the\nevaluation. Given a video summary, a text representation of the video summary\nis first generated, and an NLP-based metric is then used to measure its\nsemantic distance to ground-truth text summaries written by humans. We show\nthat our technique has higher agreement with human judgment than pixel-based\ndistance metrics. We also release text annotations and ground-truth text\nsummaries for a number of publicly available video datasets, for use by the\ncomputer vision community.', 'Standard neural machine translation (NMT) is on the assumption of\ndocument-level context independent. Most existing document-level NMT methods\nare satisfied with a smattering sense of brief document-level information,\nwhile this work focuses on exploiting detailed document-level context in terms\nof multiple forms of document embeddings, which is capable of sufficiently\nmodeling deeper and richer document-level context. The proposed document-aware\nNMT is implemented to enhance the Transformer baseline by introducing both\nglobal and local document-level clues on the source end. Experiments show that\nthe proposed method significantly improves the translation performance over\nstrong baselines and other related studies.', 'The scarcity of comprehensive up-to-date studies on evaluation metrics for\ntext summarization and the lack of consensus regarding evaluation protocols\ncontinue to inhibit progress. We address the existing shortcomings of\nsummarization evaluation methods along five dimensions: 1) we re-evaluate 14\nautomatic evaluation metrics in a comprehensive and consistent fashion using\nneural summarization model outputs along with expert and crowd-sourced human\nannotations, 2) we consistently benchmark 23 recent summarization models using\nthe aforementioned automatic evaluation metrics, 3) we assemble the largest\ncollection of summaries generated by models trained on the CNN/DailyMail news\ndataset and share it in a unified format, 4) we implement and share a toolkit\nthat provides an extensible and unified API for evaluating summarization models\nacross a broad range of automatic metrics, 5) we assemble and share the largest\nand most diverse, in terms of model types, collection of human judgments of\nmodel-generated summaries on the CNN/Daily Mail dataset annotated by both\nexpert judges and crowd-source workers. We hope that this work will help\npromote a more complete evaluation protocol for text summarization as well as\nadvance research in developing evaluation metrics that better correlate with\nhuman judgments.', ""Summarizing content contributed by individuals can be challenging, because\npeople make different lexical choices even when describing the same events.\nHowever, there remains a significant need to summarize such content. Examples\ninclude the student responses to post-class reflective questions, product\nreviews, and news articles published by different news agencies related to the\nsame events. High lexical diversity of these documents hinders the system's\nability to effectively identify salient content and reduce summary redundancy.\nIn this paper, we overcome this issue by introducing an integer linear\nprogramming-based summarization framework. It incorporates a low-rank\napproximation to the sentence-word co-occurrence matrix to intrinsically group\nsemantically-similar lexical items. We conduct extensive experiments on\ndatasets of student responses, product reviews, and news documents. Our\napproach compares favorably to a number of extractive baselines as well as a\nneural abstractive summarization system. The paper finally sheds light on when\nand why the proposed framework is effective at summarizing content with high\nlexical variety."", 'Currently used metrics for assessing summarization algorithms do not account\nfor whether summaries are factually consistent with source documents. We\npropose a weakly-supervised, model-based approach for verifying factual\nconsistency and identifying conflicts between source documents and a generated\nsummary. Training data is generated by applying a series of rule-based\ntransformations to the sentences of source documents. The factual consistency\nmodel is then trained jointly for three tasks: 1) identify whether sentences\nremain factually consistent after transformation, 2) extract a span in the\nsource documents to support the consistency prediction, 3) extract a span in\nthe summary sentence that is inconsistent if one exists. Transferring this\nmodel to summaries generated by several state-of-the art models reveals that\nthis highly scalable approach substantially outperforms previous models,\nincluding those trained with strong supervision using standard datasets for\nnatural language inference and fact checking. Additionally, human evaluation\nshows that the auxiliary span extraction tasks provide useful assistance in the\nprocess of verifying factual consistency.', 'Predicting the quality of a text document is a critical task when presented\nwith the problem of measuring the performance of a document before its release.\nIn this work, we evaluate various features including those extracted from the\ntext content (textual) and those describing higher-level characteristics of the\ntext (meta) features that are not directly available from the text, and show\nhow these features inform prediction of document quality in different ways.\nMoreover, we also compare our methods on both social user-generated data such\nas tweets, and scholarly user-generated data such as academic articles, showing\nhow the same features differently influence prediction of quality across these\ndisparate domains.', 'The technology of automatic document summarization is maturing and may\nprovide a solution to the information overload problem. Nowadays, document\nsummarization plays an important role in information retrieval. With a large\nvolume of documents, presenting the user with a summary of each document\ngreatly facilitates the task of finding the desired documents. Document\nsummarization is a process of automatically creating a compressed version of a\ngiven document that provides useful information to users, and multi-document\nsummarization is to produce a summary delivering the majority of information\ncontent from a set of documents about an explicit or implicit main topic. The\nlexical cohesion structure of the text can be exploited to determine the\nimportance of a sentence/phrase. Lexical chains are useful tools to analyze the\nlexical cohesion structure in a text .In this paper we consider the effect of\nthe use of lexical cohesion features in Summarization, And presenting a\nalgorithm base on the knowledge base. Ours algorithm at first find the correct\nsense of any word, Then constructs the lexical chains, remove Lexical chains\nthat less score than other, detects topics roughly from lexical chains,\nsegments the text with respect to the topics and selects the most important\nsentences. The experimental results on an open benchmark datasets from DUC01\nand DUC02 show that our proposed approach can improve the performance compared\nto sate-of-the-art summarization approaches.', 'Automatic Text Summarization strategies have been successfully employed to\ndigest text collections and extract its essential content. Usually, summaries\nare generated using textual corpora that belongs to the same domain area where\nthe summary will be used. Nonetheless, there are special cases where it is not\nfound enough textual sources, and one possible alternative is to generate a\nsummary from a different domain. One manner to summarize texts consists of\nusing a graph model. This model allows giving more importance to words\ncorresponding to the main concepts from the target domain found in the\nsummarized text. This gives the reader an overview of the main text concepts as\nwell as their relationships. However, this kind of summarization presents a\nsignificant number of repeated terms when compared to human-generated\nsummaries. In this paper, we present an approach to produce graph-model\nextractive summaries of texts, meeting the target domain exigences and treating\nthe terms repetition problem. To evaluate the proposition, we performed a\nseries of experiments showing that the proposed approach statistically improves\nthe performance of a model based on Graph Centrality, achieving better\ncoverage, accuracy, and recall.', ""Prior work in document summarization has mainly focused on generating short\nsummaries of a document. While this type of summary helps get a high-level view\nof a given document, it is desirable in some cases to know more detailed\ninformation about its salient points that can't fit in a short summary. This is\ntypically the case for longer documents such as a research paper, legal\ndocument, or a book. In this paper, we present a new method for generating\nextended summaries of long papers. Our method exploits hierarchical structure\nof the documents and incorporates it into an extractive summarization model\nthrough a multi-task learning approach. We then present our results on three\nlong summarization datasets, arXiv-Long, PubMed-Long, and Longsumm. Our method\noutperforms or matches the performance of strong baselines. Furthermore, we\nperform a comprehensive analysis over the generated results, shedding insights\non future research for long-form summary generation task. Our analysis shows\nthat our multi-tasking approach can adjust extraction probability distribution\nto the favor of summary-worthy sentences across diverse sections. Our datasets,\nand codes are publicly available at\nhttps://github.com/Georgetown-IR-Lab/ExtendedSumm"", 'Neural abstractive summarization models are prone to generate content\ninconsistent with the source document, i.e. unfaithful. Existing automatic\nmetrics do not capture such mistakes effectively. We tackle the problem of\nevaluating faithfulness of a generated summary given its source document. We\nfirst collected human annotations of faithfulness for outputs from numerous\nmodels on two datasets. We find that current models exhibit a trade-off between\nabstractiveness and faithfulness: outputs with less word overlap with the\nsource document are more likely to be unfaithful. Next, we propose an automatic\nquestion answering (QA) based metric for faithfulness, FEQA, which leverages\nrecent advances in reading comprehension. Given question-answer pairs generated\nfrom the summary, a QA model extracts answers from the document; non-matched\nanswers indicate unfaithful information in the summary. Among metrics based on\nword overlap, embedding similarity, and learned language understanding models,\nour QA-based metric has significantly higher correlation with human\nfaithfulness scores, especially on highly abstractive summaries.', 'Despite significant progress in neural abstractive summarization, recent\nstudies have shown that the current models are prone to generating summaries\nthat are unfaithful to the original context. To address the issue, we study\ncontrast candidate generation and selection as a model-agnostic post-processing\ntechnique to correct the extrinsic hallucinations (i.e. information not present\nin the source text) in unfaithful summaries. We learn a discriminative\ncorrection model by generating alternative candidate summaries where named\nentities and quantities in the generated summary are replaced with ones with\ncompatible semantic types from the source document. This model is then used to\nselect the best candidate as the final output summary. Our experiments and\nanalysis across a number of neural summarization systems show that our proposed\nmethod is effective in identifying and correcting extrinsic hallucinations. We\nanalyze the typical hallucination phenomenon by different types of neural\nsummarization systems, in hope to provide insights for future work on the\ndirection.', ""We present BLANC, a new approach to the automatic estimation of document\nsummary quality. Our goal is to measure the functional performance of a summary\nwith an objective, reproducible, and fully automated method. Our approach\nachieves this by measuring the performance boost gained by a pre-trained\nlanguage model with access to a document summary while carrying out its\nlanguage understanding task on the document's text. We present evidence that\nBLANC scores have as good correlation with human evaluations as do the ROUGE\nfamily of summary quality measurements. And unlike ROUGE, the BLANC method does\nnot require human-written reference summaries, allowing for fully human-free\nsummary quality estimation."", 'Owing to the rapidly growing multimedia content available on the Internet,\nextractive spoken document summarization, with the purpose of automatically\nselecting a set of representative sentences from a spoken document to concisely\nexpress the most important theme of the document, has been an active area of\nresearch and experimentation. On the other hand, word embedding has emerged as\na newly favorite research subject because of its excellent performance in many\nnatural language processing (NLP)-related tasks. However, as far as we are\naware, there are relatively few studies investigating its use in extractive\ntext or speech summarization. A common thread of leveraging word embeddings in\nthe summarization process is to represent the document (or sentence) by\naveraging the word embeddings of the words occurring in the document (or\nsentence). Then, intuitively, the cosine similarity measure can be employed to\ndetermine the relevance degree between a pair of representations. Beyond the\ncontinued efforts made to improve the representation of words, this paper\nfocuses on building novel and efficient ranking models based on the general\nword embedding methods for extractive speech summarization. Experimental\nresults demonstrate the effectiveness of our proposed methods, compared to\nexisting state-of-the-art methods.', 'In this paper, we present a model for generating summaries of text documents\nwith respect to a query. This is known as query-based summarization. We adapt\nan existing dataset of news article summaries for the task and train a\npointer-generator model using this dataset. The generated summaries are\nevaluated by measuring similarity to reference summaries. Our results show that\na neural network summarization model, similar to existing neural network models\nfor abstractive summarization, can be constructed to make use of queries to\nproduce targeted summaries.', 'A major challenge for scaling machine learning is training models to perform\ntasks that are very difficult or time-consuming for humans to evaluate. We\npresent progress on this problem on the task of abstractive summarization of\nentire fiction novels. Our method combines learning from human feedback with\nrecursive task decomposition: we use models trained on smaller parts of the\ntask to assist humans in giving feedback on the broader task. We collect a\nlarge volume of demonstrations and comparisons from human labelers, and\nfine-tune GPT-3 using behavioral cloning and reward modeling to do\nsummarization recursively. At inference time, the model first summarizes small\nsections of the book and then recursively summarizes these summaries to produce\na summary of the entire book. Our human labelers are able to supervise and\nevaluate the models quickly, despite not having read the entire books\nthemselves. Our resulting model generates sensible summaries of entire books,\neven matching the quality of human-written summaries in a few cases ($\\sim5\\%$\nof books). We achieve state-of-the-art results on the recent BookSum dataset\nfor book-length summarization. A zero-shot question-answering model using these\nsummaries achieves state-of-the-art results on the challenging NarrativeQA\nbenchmark for answering questions about books and movie scripts. We release\ndatasets of samples from our model.', 'Long document summarization is an important and hard task in the field of\nnatural language processing. A good performance of the long document\nsummarization reveals the model has a decent understanding of the human\nlanguage. Currently, most researches focus on how to modify the attention\nmechanism of the transformer to achieve a higher ROUGE score. The study of data\npre-processing and post-processing are relatively few. In this paper, we use\ntwo pre-processing methods and a post-processing method and analyze the effect\nof these methods on various long document summarization models.', ""Text generation is an important Natural Language Processing task with various\napplications. Although several metrics have already been introduced to evaluate\nthe text generation methods, each of them has its own shortcomings. The most\nwidely used metrics such as BLEU only consider the quality of generated\nsentences and neglect their diversity. For example, repeatedly generation of\nonly one high quality sentence would result in a high BLEU score. On the other\nhand, the more recent metric introduced to evaluate the diversity of generated\ntexts known as Self-BLEU ignores the quality of generated texts. In this paper,\nwe propose metrics to evaluate both the quality and diversity simultaneously by\napproximating the distance of the learned generative model and the real data\ndistribution. For this purpose, we first introduce a metric that approximates\nthis distance using n-gram based measures. Then, a feature-based measure which\nis based on a recent highly deep model trained on a large text corpus called\nBERT is introduced. Finally, for oracle training mode in which the generator's\ndensity can also be calculated, we propose to use the distance measures between\nthe corresponding explicit distributions. Eventually, the most popular and\nrecent text generation models are evaluated using both the existing and the\nproposed metrics and the preferences of the proposed metrics are determined."", 'Single document summarization generates summary by extracting the\nrepresentative sentences from the document. In this paper, we presented a novel\ntechnique for summarization of domain-specific text from a single web document\nthat uses statistical and linguistic analysis on the text in a reference corpus\nand the web document. The proposed summarizer uses the combinational function\nof Sentence Weight (SW) and Subject Weight (SuW) to determine the rank of a\nsentence, where SW is the function of number of terms (t_n) and number of words\n(w_n) in a sentence, and term frequency (t_f) in the corpus and SuW is the\nfunction of t_n and w_n in a subject, and t_f in the corpus. 30 percent of the\nranked sentences are considered to be the summary of the web document. We\ngenerated three web document summaries using our technique and compared each of\nthem with the summaries developed manually from 16 different human subjects.\nResults showed that 68 percent of the summaries produced by our approach\nsatisfy the manual summaries.', 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'In the light of exponentially increasing video content, video summarization\nhas attracted a lot of attention recently due to its ability to optimize time\nand storage. Characteristics of a good summary of a video depend on the\nparticular domain under question. We propose a novel framework for domain\nspecific video summarization. Given a video of a particular domain, our system\ncan produce a summary based on what is important for that domain in addition to\npossessing other desired characteristics like representativeness, coverage,\ndiversity etc. as suitable to that domain. Past related work has focused either\non using supervised approaches for ranking the snippets to produce summary or\non using unsupervised approaches of generating the summary as a subset of\nsnippets with the above characteristics. We look at the joint problem of\nlearning domain specific importance of segments as well as the desired summary\ncharacteristic for that domain. Our studies show that the more efficient way of\nincorporating domain specific relevances into a summary is by obtaining ratings\nof shots as opposed to binary inclusion/exclusion information. We also argue\nthat ratings can be seen as unified representation of all possible ground truth\nsummaries of a video, taking us one step closer in dealing with challenges\nassociated with multiple ground truth summaries of a video. We also propose a\nnovel evaluation measure which is more naturally suited in assessing the\nquality of video summary for the task at hand than F1 like measures. It\nleverages the ratings information and is richer in appropriately modeling\ndesirable and undesirable characteristics of a summary. Lastly, we release a\ngold standard dataset for furthering research in domain specific video\nsummarization, which to our knowledge is the first dataset with long videos\nacross several domains with rating annotations.', 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', 'The amount of text data available online is increasing at a very fast pace\nhence text summarization has become essential. Most of the modern recommender\nand text classification systems require going through a huge amount of data.\nManually generating precise and fluent summaries of lengthy articles is a very\ntiresome and time-consuming task. Hence generating automated summaries for the\ndata and using it to train machine learning models will make these models space\nand time-efficient. Extractive summarization and abstractive summarization are\ntwo separate methods of generating summaries. The extractive technique\nidentifies the relevant sentences from the original document and extracts only\nthose from the text. Whereas in abstractive summarization techniques, the\nsummary is generated after interpreting the original text, hence making it more\ncomplicated. In this paper, we will be presenting a comprehensive comparison of\na few transformer architecture based pre-trained models for text summarization.\nFor analysis and comparison, we have used the BBC news dataset that contains\ntext data that can be used for summarization and human generated summaries for\nevaluating and comparing the summaries generated by machine learning models.', 'Automatic text summarization aims to produce a brief but crucial summary for\nthe input documents. Both extractive and abstractive methods have witnessed\ngreat success in English datasets in recent years. However, there has been a\nminimal exploration of text summarization in Chinese, limited by the lack of\nlarge-scale datasets. In this paper, we present a large-scale Chinese news\nsummarization dataset CNewSum, which consists of 304,307 documents and\nhuman-written summaries for the news feed. It has long documents with\nhigh-abstractive summaries, which can encourage document-level understanding\nand generation for current summarization models. An additional distinguishing\nfeature of CNewSum is that its test set contains adequacy and deducibility\nannotations for the summaries. The adequacy level measures the degree of\nsummary information covered by the document, and the deducibility indicates the\nreasoning ability the model needs to generate the summary. These annotations\ncan help researchers analyze and target their model performance bottleneck. We\nexamine recent methods on CNewSum and release our dataset to provide a solid\ntestbed for automatic Chinese summarization research.', 'Data-driven approaches to sequence-to-sequence modelling have been\nsuccessfully applied to short text summarization of news articles. Such models\nare typically trained on input-summary pairs consisting of only a single or a\nfew sentences, partially due to limited availability of multi-sentence training\ndata. Here, we propose to use scientific articles as a new milestone for text\nsummarization: large-scale training data come almost for free with two types of\nhigh-quality summaries at different levels - the title and the abstract. We\ngenerate two novel multi-sentence summarization datasets from scientific\narticles and test the suitability of a wide range of existing extractive and\nabstractive neural network-based summarization approaches. Our analysis\ndemonstrates that scientific papers are suitable for data-driven text\nsummarization. Our results could serve as valuable benchmarks for scaling\nsequence-to-sequence models to very long sequences.', 'Summarization is one of the key features of human intelligence. It plays an\nimportant role in understanding and representation. With rapid and continual\nexpansion of texts, pictures and videos in cyberspace, automatic summarization\nbecomes more and more desirable. Text summarization has been studied for over\nhalf century, but it is still hard to automatically generate a satisfied\nsummary. Traditional methods process texts empirically and neglect the\nfundamental characteristics and principles of language use and understanding.\nThis paper summarizes previous text summarization approaches in a\nmulti-dimensional classification space, introduces a multi-dimensional\nmethodology for research and development, unveils the basic characteristics and\nprinciples of language use and understanding, investigates some fundamental\nmechanisms of summarization, studies the dimensions and forms of\nrepresentations, and proposes a multi-dimensional evaluation mechanisms.\nInvestigation extends to the incorporation of pictures into summary and to the\nsummarization of videos, graphs and pictures, and then reaches a general\nsummarization framework.', 'Text summarization is an interesting area for researchers to develop new\ntechniques to provide human like summaries for vast amounts of information.\nSummarization techniques tend to focus on providing accurate representation of\ncontent, and often the tone of the content is ignored. Tone of the content sets\na baseline for how a reader perceives the content. As such being able to\ngenerate summary with tone that is appropriate for the reader is important. In\nour work we implement Maximal Marginal Relevance [MMR] based multi-document\ntext summarization and propose a naive model to change tone of the\nsummarization by setting a bias to specific set of words and restricting other\nwords in the summarization output. This bias towards a specified set of words\nproduces a summary whose tone is same as tone of specified words.', 'Highlighting while reading is a natural behavior for people to track salient\ncontent of a document. It would be desirable to teach an extractive summarizer\nto do the same. However, a major obstacle to the development of a supervised\nsummarizer is the lack of ground-truth. Manual annotation of extraction units\nis cost-prohibitive, whereas acquiring labels by automatically aligning human\nabstracts and source documents can yield inferior results. In this paper we\ndescribe a novel framework to guide a supervised, extractive summarization\nsystem with question-answering rewards. We argue that quality summaries should\nserve as a document surrogate to answer important questions, and such\nquestion-answer pairs can be conveniently obtained from human abstracts. The\nsystem learns to promote summaries that are informative, fluent, and perform\ncompetitively on question-answering. Our results compare favorably with those\nreported by strong summarization baselines as evaluated by automatic metrics\nand human assessors.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', ""The development of summarization research has been significantly hampered by\nthe costly acquisition of reference summaries. This paper proposes an effective\nway to automatically collect large scales of news-related multi-document\nsummaries with reference to social media's reactions. We utilize two types of\nsocial labels in tweets, i.e., hashtags and hyper-links. Hashtags are used to\ncluster documents into different topic sets. Also, a tweet with a hyper-link\noften highlights certain key points of the corresponding document. We\nsynthesize a linked document cluster to form a reference summary which can\ncover most key points. To this aim, we adopt the ROUGE metrics to measure the\ncoverage ratio, and develop an Integer Linear Programming solution to discover\nthe sentence set reaching the upper bound of ROUGE. Since we allow summary\nsentences to be selected from both documents and high-quality tweets, the\ngenerated reference summaries could be abstractive. Both informativeness and\nreadability of the collected summaries are verified by manual judgment. In\naddition, we train a Support Vector Regression summarizer on DUC generic\nmulti-document summarization benchmarks. With the collected data as extra\ntraining resource, the performance of the summarizer improves a lot on all the\ntest sets. We release this dataset for further research."", ""Neural conversation models are known to generate appropriate but\nnon-informative responses in general. A scenario where informativeness can be\nsignificantly enhanced is Conversing by Reading (CbR), where conversations take\nplace with respect to a given external document. In previous work, the external\ndocument is utilized by (1) creating a context-aware document memory that\nintegrates information from the document and the conversational context, and\nthen (2) generating responses referring to the memory. In this paper, we\npropose to create the document memory with some anticipated responses in mind.\nThis is achieved using a teacher-student framework. The teacher is given the\nexternal document, the context, and the ground-truth response, and learns how\nto build a response-aware document memory from three sources of information.\nThe student learns to construct a response-anticipated document memory from the\nfirst two sources, and the teacher's insight on memory creation. Empirical\nresults show that our model outperforms the previous state-of-the-art for the\nCbR task."", 'We present a novel abstractive summarization framework that draws on the\nrecent development of a treebank for the Abstract Meaning Representation (AMR).\nIn this framework, the source text is parsed to a set of AMR graphs, the graphs\nare transformed into a summary graph, and then text is generated from the\nsummary graph. We focus on the graph-to-graph transformation that reduces the\nsource semantic graph into a summary graph, making use of an existing AMR\nparser and assuming the eventual availability of an AMR-to-text generator. The\nframework is data-driven, trainable, and not specifically designed for a\nparticular domain. Experiments on gold-standard AMR annotations and system\nparses show promising results. Code is available at:\nhttps://github.com/summarization', 'Question-driven summarization has been recently studied as an effective\napproach to summarizing the source document to produce concise but informative\nanswers for non-factoid questions. In this work, we propose a novel\nquestion-driven abstractive summarization method, Multi-hop Selective Generator\n(MSG), to incorporate multi-hop reasoning into question-driven summarization\nand, meanwhile, provide justifications for the generated summaries.\nSpecifically, we jointly model the relevance to the question and the\ninterrelation among different sentences via a human-like multi-hop inference\nmodule, which captures important sentences for justifying the summarized\nanswer. A gated selective pointer generator network with a multi-view coverage\nmechanism is designed to integrate diverse information from different\nperspectives. Experimental results show that the proposed method consistently\noutperforms state-of-the-art methods on two non-factoid QA datasets, namely\nWikiHow and PubMedQA.', 'Standard neural machine translation (NMT) is on the assumption that the\ndocument-level context is independent. Most existing document-level NMT\napproaches are satisfied with a smattering sense of global document-level\ninformation, while this work focuses on exploiting detailed document-level\ncontext in terms of a memory network. The capacity of the memory network that\ndetecting the most relevant part of the current sentence from memory renders a\nnatural solution to model the rich document-level context. In this work, the\nproposed document-aware memory network is implemented to enhance the\nTransformer NMT baseline. Experiments on several tasks show that the proposed\nmethod significantly improves the NMT performance over strong Transformer\nbaselines and other related studies.', 'This paper introduces STRASS: Summarization by TRAnsformation Selection and\nScoring. It is an extractive text summarization method which leverages the\nsemantic information in existing sentence embedding spaces. Our method creates\nan extractive summary by selecting the sentences with the closest embeddings to\nthe document embedding. The model learns a transformation of the document\nembedding to minimize the similarity between the extractive summary and the\nground truth summary. As the transformation is only composed of a dense layer,\nthe training can be done on CPU, therefore, inexpensive. Moreover, inference\ntime is short and linear according to the number of sentences. As a second\ncontribution, we introduce the French CASS dataset, composed of judgments from\nthe French Court of cassation and their corresponding summaries. On this\ndataset, our results show that our method performs similarly to the state of\nthe art extractive methods with effective training and inferring time.', 'Previous researchers have considered sentiment analysis as a document\nclassification task, in which input documents are classified into predefined\nsentiment classes. Although there are sentences in a document that support\nimportant evidences for sentiment analysis and sentences that do not, they have\ntreated the document as a bag of sentences. In other words, they have not\nconsidered the importance of each sentence in the document. To effectively\ndetermine polarity of a document, each sentence in the document should be dealt\nwith different degrees of importance. To address this problem, we propose a\ndocument-level sentence classification model based on deep neural networks, in\nwhich the importance degrees of sentences in documents are automatically\ndetermined through gate mechanisms. To verify our new sentiment analysis model,\nwe conducted experiments using the sentiment datasets in the four different\ndomains such as movie reviews, hotel reviews, restaurant reviews, and music\nreviews. In the experiments, the proposed model outperformed previous\nstate-of-the-art models that do not consider importance differences of\nsentences in a document. The experimental results show that the importance of\nsentences should be considered in a document-level sentiment classification\ntask.', 'In the Query Focused Multi-Document Summarization (QF-MDS) task, a set of\ndocuments and a query are given where the goal is to generate a summary from\nthese documents based on the given query. However, one major challenge for this\ntask is the lack of availability of labeled training datasets. To overcome this\nissue, in this paper, we propose a novel weakly supervised learning approach\nvia utilizing distant supervision. In particular, we use datasets similar to\nthe target dataset as the training data where we leverage pre-trained sentence\nsimilarity models to generate the weak reference summary of each individual\ndocument in a document set from the multi-document gold reference summaries.\nThen, we iteratively train our summarization model on each single-document to\nalleviate the computational complexity issue that occurs while training neural\nsummarization models in multiple documents (i.e., long sequences) at once.\nExperimental results in Document Understanding Conferences (DUC) datasets show\nthat our proposed approach sets a new state-of-the-art result in terms of\nvarious evaluation metrics.', 'Opinion summarization is the automatic creation of text reflecting subjective\ninformation expressed in multiple documents, such as user reviews of a product.\nThe task is practically important and has attracted a lot of attention.\nHowever, due to the high cost of summary production, datasets large enough for\ntraining supervised models are lacking. Instead, the task has been\ntraditionally approached with extractive methods that learn to select text\nfragments in an unsupervised or weakly-supervised way. Recently, it has been\nshown that abstractive summaries, potentially more fluent and better at\nreflecting conflicting information, can also be produced in an unsupervised\nfashion. However, these models, not being exposed to actual summaries, fail to\ncapture their essential properties. In this work, we show that even a handful\nof summaries is sufficient to bootstrap generation of the summary text with all\nexpected properties, such as writing style, informativeness, fluency, and\nsentiment preservation. We start by training a conditional Transformer language\nmodel to generate a new product review given other available reviews of the\nproduct. The model is also conditioned on review properties that are directly\nrelated to summaries; the properties are derived from reviews with no manual\neffort. In the second stage, we fine-tune a plug-in module that learns to\npredict property values on a handful of summaries. This lets us switch the\ngenerator to the summarization mode. We show on Amazon and Yelp datasets that\nour approach substantially outperforms previous extractive and abstractive\nmethods in automatic and human evaluation.', 'We present a method for generating comparative summaries that highlights\nsimilarities and contradictions in input documents. The key challenge in\ncreating such summaries is the lack of large parallel training data required\nfor training typical summarization systems. To this end, we introduce a hybrid\ngeneration approach inspired by traditional concept-to-text systems. To enable\naccurate comparison between different sources, the model first learns to\nextract pertinent relations from input documents. The content planning\ncomponent uses deterministic operators to aggregate these relations after\nidentifying a subset for inclusion into a summary. The surface realization\ncomponent lexicalizes this information using a text-infilling language model.\nBy separately modeling content selection and realization, we can effectively\ntrain them with limited annotations. We implemented and tested the model in the\ndomain of nutrition and health -- rife with inconsistencies. Compared to\nconventional methods, our framework leads to more faithful, relevant and\naggregation-sensitive summarization -- while being equally fluent.', 'Most of the textual information available to us are temporally variable. In a\nworld where information is dynamic, time-stamping them is a very important\ntask. Documents are a good source of information and are used for many tasks\nlike, sentiment analysis, classification of reviews etc. The knowledge of\ncreation date of documents facilitates several tasks like summarization, event\nextraction, temporally focused information extraction etc. Unfortunately, for\nmost of the documents on the web, the time-stamp meta-data is either erroneous\nor missing. Thus document dating is a challenging problem which requires\ninference over the temporal structure of the document alongside the contextual\ninformation of the document. Prior document dating systems have largely relied\non handcrafted features while ignoring such document-internal structures. In\nthis paper we propose NeuralDater, a Graph Convolutional Network (GCN) based\ndocument dating approach which jointly exploits syntactic and temporal graph\nstructures of document in a principled way. We also pointed out some\nlimitations of NeuralDater and tried to utilize both context and temporal\ninformation in documents in a more flexible and intuitive manner proposing AD3:\nAttentive Deep Document Dater, an attention-based document dating system. To\nthe best of our knowledge these are the first application of deep learning\nmethods for the task. Through extensive experiments on real-world datasets, we\nfind that our models significantly outperforms state-of-the-art baselines by a\nsignificant margin.', 'We present a robust approach for detecting intrinsic sentence importance in\nnews, by training on two corpora of document-summary pairs. When used for\nsingle-document summarization, our approach, combined with the ""beginning of\ndocument"" heuristic, outperforms a state-of-the-art summarizer and the\nbeginning-of-article baseline in both automatic and manual evaluations. These\nresults represent an important advance because in the absence of cross-document\nrepetition, single document summarizers for news have not been able to\nconsistently outperform the strong beginning-of-article baseline.', 'We present a contrasting learning approach with data augmentation techniques\nto learn document representations in an unsupervised manner. Inspired by recent\ncontrastive self-supervised learning algorithms used for image and NLP\npretraining, we hypothesize that high-quality document embedding should be\ninvariant to diverse paraphrases that preserve the semantics of the original\ndocument. With different backbones and contrastive learning frameworks, our\nstudy reveals the enormous benefits of contrastive augmentation for document\nrepresentation learning with two additional insights: 1) including data\naugmentation in a contrastive way can substantially improve the embedding\nquality in unsupervised document representation learning, and 2) in general,\nstochastic augmentations generated by simple word-level manipulation work much\nbetter than sentence-level and document-level ones. We plug our method into a\nclassifier and compare it with a broad range of baseline methods on six\nbenchmark datasets. Our method can decrease the classification error rate by up\nto 6.4% over the SOTA approaches on the document classification task, matching\nor even surpassing fully-supervised methods.', 'Statistical methods have been widely employed in recent years to grasp many\nlanguage properties. The application of such techniques have allowed an\nimprovement of several linguistic applications, which encompasses machine\ntranslation, automatic summarization and document classification. In the\nlatter, many approaches have emphasized the semantical content of texts, as it\nis the case of bag-of-word language models. This approach has certainly yielded\nreasonable performance. However, some potential features such as the structural\norganization of texts have been used only on a few studies. In this context, we\nprobe how features derived from textual structure analysis can be effectively\nemployed in a classification task. More specifically, we performed a supervised\nclassification aiming at discriminating informative from imaginative documents.\nUsing a networked model that describes the local topological/dynamical\nproperties of function words, we achieved an accuracy rate of up to 95%, which\nis much higher than similar networked approaches. A systematic analysis of\nfeature relevance revealed that symmetry and accessibility measurements are\namong the most prominent network measurements. Our results suggest that these\nmeasurements could be used in related language applications, as they play a\ncomplementary role in characterizing texts.', 'With the increase of information, document classification as one of the\nmethods of text mining, plays vital role in many management and organizing\ninformation. Document classification is the process of assigning a document to\none or more predefined category labels. Document classification includes\ndifferent parts such as text processing, term selection, term weighting and\nfinal classification. The accuracy of document classification is very\nimportant. Thus improvement in each part of classification should lead to\nbetter results and higher precision. Term weighting has a great impact on the\naccuracy of the classification. Most of the existing weighting methods exploit\nthe statistical information of terms in documents and do not consider semantic\nrelations between words. In this paper, an automated document classification\nsystem is presented that uses a novel term weighting method based on semantic\nrelations between terms. To evaluate the proposed method, three standard\nPersian corpuses are used. Experiment results show 2 to 4 percent improvement\nin classification accuracy compared with the best previous designed system for\nPersian documents.', 'Many Natural Language Processing and Computational Linguistics applications\ninvolves the generation of new texts based on some existing texts, such as\nsummarization, text simplification and machine translation. However, there has\nbeen a serious problem haunting these applications for decades, that is, how to\nautomatically and accurately assess quality of these applications. In this\npaper, we will present some preliminary results on one especially useful and\nchallenging problem in NLP system evaluation: how to pinpoint content\ndifferences of two text passages (especially for large pas-sages such as\narticles and books). Our idea is intuitive and very different from existing\napproaches. We treat one text passage as a small knowledge base, and ask it a\nlarge number of questions to exhaustively identify all content points in it. By\ncomparing the correctly answered questions from two text passages, we will be\nable to compare their content precisely. The experiment using 2007 DUC\nsummarization corpus clearly shows promising results.', 'The increasing amount of online content motivated the development of\nmulti-document summarization methods. In this work, we explore straightforward\napproaches to extend single-document summarization methods to multi-document\nsummarization. The proposed methods are based on the hierarchical combination\nof single-document summaries, and achieves state of the art results.', 'Document AI, or Document Intelligence, is a relatively new research topic\nthat refers to the techniques for automatically reading, understanding, and\nanalyzing business documents. It is an important research direction for natural\nlanguage processing and computer vision. In recent years, the popularity of\ndeep learning technology has greatly advanced the development of Document AI,\nsuch as document layout analysis, visual information extraction, document\nvisual question answering, document image classification, etc. This paper\nbriefly reviews some of the representative models, tasks, and benchmark\ndatasets. Furthermore, we also introduce early-stage heuristic rule-based\ndocument analysis, statistical machine learning algorithms, and deep learning\napproaches especially pre-training methods. Finally, we look into future\ndirections for Document AI research.', 'Recent studies emphasize the need of document context in human evaluation of\nmachine translations, but little research has been done on the impact of user\ninterfaces on annotator productivity and the reliability of assessments. In\nthis work, we compare human assessment data from the last two WMT evaluation\ncampaigns collected via two different methods for document-level evaluation.\nOur analysis shows that a document-centric approach to evaluation where the\nannotator is presented with the entire document context on a screen leads to\nhigher quality segment and document level assessments. It improves the\ncorrelation between segment and document scores and increases inter-annotator\nagreement for document scores but is considerably more time consuming for\nannotators.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', 'We propose a sequence-to-sequence singing synthesizer, which avoids the need\nfor training data with pre-aligned phonetic and acoustic features. Rather than\nthe more common approach of a content-based attention mechanism combined with\nan autoregressive decoder, we use a different mechanism suitable for\nfeed-forward synthesis. Given that phonetic timings in singing are highly\nconstrained by the musical score, we derive an approximate initial alignment\nwith the help of a simple duration model. Then, using a decoder based on a\nfeed-forward variant of the Transformer model, a series of self-attention and\nconvolutional layers refines the result of the initial alignment to reach the\ntarget acoustic features. Advantages of this approach include faster inference\nand avoiding the exposure bias issues that affect autoregressive models trained\nby teacher forcing. We evaluate the effectiveness of this model compared to an\nautoregressive baseline, the importance of self-attention, and the importance\nof the accuracy of the duration model.', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'We investigate the pertinence of methods from algebraic topology for text\ndata analysis. These methods enable the development of\nmathematically-principled isometric-invariant mappings from a set of vectors to\na document embedding, which is stable with respect to the geometry of the\ndocument in the selected metric space. In this work, we evaluate the utility of\nthese topology-based document representations in traditional NLP tasks,\nspecifically document clustering and sentiment classification. We find that the\nembeddings do not benefit text analysis. In fact, performance is worse than\nsimple techniques like $\\textit{tf-idf}$, indicating that the geometry of the\ndocument does not provide enough variability for classification on the basis of\ntopic or sentiment in the chosen datasets.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : I aim to develop a reinforcement learning framework that utilizes human feedback as sparse reward signals. This framework will have the capacity to incorporate any model for a variety of potential downstream tasks, such as a generative model for art creation or a chatbot for customer service. The unique aspect of my framework is its ability to apply reinforcement learning to adjust and optimize the behavior of any downstream model based on various forms of human feedback, such as user satisfaction or the model's adherence to ethical standards. I understand that obtaining a large amount of human feedback data may be challenging, so my framework needs to be capable of learning from sparse and limited human feedback data. I am interested in finding out if there has been any previous work in this area. Furthermore, I would like to know if a reinforcement learning algorithm would still function effectively when the supervised human feedback data is scarce.","['Recent successes combine reinforcement learning algorithms and deep neural\nnetworks, despite reinforcement learning not being widely applied to robotics\nand real world scenarios. This can be attributed to the fact that current\nstate-of-the-art, end-to-end reinforcement learning approaches still require\nthousands or millions of data samples to converge to a satisfactory policy and\nare subject to catastrophic failures during training. Conversely, in real world\nscenarios and after just a few data samples, humans are able to either provide\ndemonstrations of the task, intervene to prevent catastrophic actions, or\nsimply evaluate if the policy is performing correctly. This research\ninvestigates how to integrate these human interaction modalities to the\nreinforcement learning loop, increasing sample efficiency and enabling\nreal-time reinforcement learning in robotics and real world scenarios. This\nnovel theoretical foundation is called Cycle-of-Learning, a reference to how\ndifferent human interaction modalities, namely, task demonstration,\nintervention, and evaluation, are cycled and combined to reinforcement learning\nalgorithms. Results presented in this work show that the reward signal that is\nlearned based upon human interaction accelerates the rate of learning of\nreinforcement learning algorithms and that learning from a combination of human\ndemonstrations and interventions is faster and more sample efficient when\ncompared to traditional supervised learning algorithms. Finally,\nCycle-of-Learning develops an effective transition between policies learned\nusing human demonstrations and interventions to reinforcement learning. The\ntheoretical foundation developed by this research opens new research paths to\nhuman-agent teaming scenarios where autonomous agents are able to learn from\nhuman teammates and adapt to mission performance metrics in real-time and in\nreal world scenarios.', ""To widen their accessibility and increase their utility, intelligent agents\nmust be able to learn complex behaviors as specified by (non-expert) human\nusers. Moreover, they will need to learn these behaviors within a reasonable\namount of time while efficiently leveraging the sparse feedback a human trainer\nis capable of providing. Recent work has shown that human feedback can be\ncharacterized as a critique of an agent's current behavior rather than as an\nalternative reward signal to be maximized, culminating in the COnvergent\nActor-Critic by Humans (COACH) algorithm for making direct policy updates based\non human feedback. Our work builds on COACH, moving to a setting where the\nagent's policy is represented by a deep neural network. We employ a series of\nmodifications on top of the original COACH algorithm that are critical for\nsuccessfully learning behaviors from high-dimensional observations, while also\nsatisfying the constraint of obtaining reduced sample complexity. We\ndemonstrate the effectiveness of our Deep COACH algorithm in the rich 3D world\nof Minecraft with an agent that learns to complete tasks by mapping from raw\npixels to actions using only real-time human feedback in 10-15 minutes of\ninteraction."", 'Reinforcement learning has been successful in training autonomous agents to\naccomplish goals in complex environments. Although this has been adapted to\nmultiple settings, including robotics and computer games, human players often\nfind it easier to obtain higher rewards in some environments than reinforcement\nlearning algorithms. This is especially true of high-dimensional state spaces\nwhere the reward obtained by the agent is sparse or extremely delayed. In this\npaper, we seek to effectively integrate feedback signals supplied by a human\noperator with deep reinforcement learning algorithms in high-dimensional state\nspaces. We call this FRESH (Feedback-based REward SHaping). During training, a\nhuman operator is presented with trajectories from a replay buffer and then\nprovides feedback on states and actions in the trajectory. In order to\ngeneralize feedback signals provided by the human operator to previously unseen\nstates and actions at test-time, we use a feedback neural network. We use an\nensemble of neural networks with a shared network architecture to represent\nmodel uncertainty and the confidence of the neural network in its output. The\noutput of the feedback neural network is converted to a shaping reward that is\naugmented to the reward provided by the environment. We evaluate our approach\non the Bowling and Skiing Atari games in the arcade learning environment.\nAlthough human experts have been able to achieve high scores in these\nenvironments, state-of-the-art deep learning algorithms perform poorly. We\nobserve that FRESH is able to achieve much higher scores than state-of-the-art\ndeep learning algorithms in both environments. FRESH also achieves a 21.4%\nhigher score than a human expert in Bowling and does as well as a human expert\nin Skiing.', 'Human explanation (e.g., in terms of feature importance) has been recently\nused to extend the communication channel between human and agent in interactive\nmachine learning. Under this setting, human trainers provide not only the\nground truth but also some form of explanation. However, this kind of human\nguidance was only investigated in supervised learning tasks, and it remains\nunclear how to best incorporate this type of human knowledge into deep\nreinforcement learning. In this paper, we present the first study of using\nhuman visual explanations in human-in-the-loop reinforcement learning (HRL). We\nfocus on the task of learning from feedback, in which the human trainer not\nonly gives binary evaluative ""good"" or ""bad"" feedback for queried state-action\npairs, but also provides a visual explanation by annotating relevant features\nin images. We propose EXPAND (EXPlanation AugmeNted feeDback) to encourage the\nmodel to encode task-relevant features through a context-aware data\naugmentation that only perturbs irrelevant features in human salient\ninformation. We choose five tasks, namely Pixel-Taxi and four Atari games, to\nevaluate the performance and sample efficiency of this approach. We show that\nour method significantly outperforms methods leveraging human explanation that\nare adapted from supervised learning, and Human-in-the-loop RL baselines that\nonly utilize evaluative feedback.', 'Deep Reinforcement Learning (DRL) has become a powerful methodology to solve\ncomplex decision-making problems. However, DRL has several limitations when\nused in real-world problems (e.g., robotics applications). For instance, long\ntraining times are required and cannot be accelerated in contrast to simulated\nenvironments, and reward functions may be hard to specify/model and/or to\ncompute. Moreover, the transfer of policies learned in a simulator to the\nreal-world has limitations (reality gap). On the other hand, machine learning\nmethods that rely on the transfer of human knowledge to an agent have shown to\nbe time efficient for obtaining well performing policies and do not require a\nreward function. In this context, we analyze the use of human corrective\nfeedback during task execution to learn policies with high-dimensional state\nspaces, by using the D-COACH framework, and we propose new variants of this\nframework. D-COACH is a Deep Learning based extension of COACH (COrrective\nAdvice Communicated by Humans), where humans are able to shape policies through\ncorrective advice. The enhanced version of D-COACH, which is proposed in this\npaper, largely reduces the time and effort of a human for training a policy.\nExperimental results validate the efficiency of the D-COACH framework in three\ndifferent problems (simulated and with real robots), and show that its enhanced\nversion reduces the human training effort considerably, and makes it feasible\nto learn policies within periods of time in which a DRL agent do not reach any\nimprovement.', ""Conveying complex objectives to reinforcement learning (RL) agents can often\nbe difficult, involving meticulous design of reward functions that are\nsufficiently informative yet easy enough to provide. Human-in-the-loop RL\nmethods allow practitioners to instead interactively teach agents through\ntailored feedback; however, such approaches have been challenging to scale\nsince human feedback is very expensive. In this work, we aim to make this\nprocess more sample- and feedback-efficient. We present an off-policy,\ninteractive RL algorithm that capitalizes on the strengths of both feedback and\noff-policy learning. Specifically, we learn a reward model by actively querying\na teacher's preferences between two clips of behavior and use it to train an\nagent. To enable off-policy learning, we relabel all the agent's past\nexperience when its reward model changes. We additionally show that\npre-training our agents with unsupervised exploration substantially increases\nthe mileage of its queries. We demonstrate that our approach is capable of\nlearning tasks of higher complexity than previously considered by\nhuman-in-the-loop methods, including a variety of locomotion and robotic\nmanipulation skills. We also show that our method is able to utilize real-time\nhuman feedback to effectively prevent reward exploitation and learn new\nbehaviors that are difficult to specify with standard reward functions."", 'While reinforcement learning (RL) has been successful in natural language\nprocessing (NLP) domains such as dialogue generation and text-based games, it\ntypically faces the problem of sparse rewards that leads to slow or no\nconvergence. Traditional methods that use text descriptions to extract only a\nstate representation ignore the feedback inherently present in them. In\ntext-based games, for example, descriptions like ""Good Job! You ate the food}""\nindicate progress, and descriptions like ""You entered a new room"" indicate\nexploration. Positive and negative cues like these can be converted to rewards\nthrough sentiment analysis. This technique converts the sparse reward problem\ninto a dense one, which is easier to solve. Furthermore, this can enable\nreinforcement learning without rewards, in which the agent learns entirely from\nthese intrinsic sentiment rewards. This framework is similar to intrinsic\nmotivation, where the environment does not necessarily provide the rewards, but\nthe agent analyzes and realizes them by itself. We find that providing dense\nrewards in text-based games using sentiment analysis improves performance under\nsome conditions.', 'Reinforcement learning (RL) requires skillful definition and remarkable\ncomputational efforts to solve optimization and control problems, which could\nimpair its prospect. Introducing human guidance into reinforcement learning is\na promising way to improve learning performance. In this paper, a comprehensive\nhuman guidance-based reinforcement learning framework is established. A novel\nprioritized experience replay mechanism that adapts to human guidance in the\nreinforcement learning process is proposed to boost the efficiency and\nperformance of the reinforcement learning algorithm. To relieve the heavy\nworkload on human participants, a behavior model is established based on an\nincremental online learning method to mimic human actions. We design two\nchallenging autonomous driving tasks for evaluating the proposed algorithm.\nExperiments are conducted to access the training and testing performance and\nlearning mechanism of the proposed algorithm. Comparative results against the\nstate-of-the-art methods suggest the advantages of our algorithm in terms of\nlearning efficiency, performance, and robustness.', ""This paper contributes a preliminary report on the advantages and\ndisadvantages of incorporating simultaneous human control and feedback signals\nin the training of a reinforcement learning robotic agent. While robotic\nhuman-machine interfaces have become increasingly complex in both form and\nfunction, control remains challenging for users. This has resulted in an\nincreasing gap between user control approaches and the number of robotic motors\nwhich can be controlled. One way to address this gap is to shift some autonomy\nto the robot. Semi-autonomous actions of the robotic agent can then be shaped\nby human feedback, simplifying user control. Most prior work on agent shaping\nby humans has incorporated training with feedback, or has included indirect\ncontrol signals. By contrast, in this paper we explore how a human can provide\nconcurrent feedback signals and real-time myoelectric control signals to train\na robot's actor-critic reinforcement learning control system. Using both a\nphysical and a simulated robotic system, we compare training performance on a\nsimple movement task when reward is derived from the environment, when reward\nis provided by the human, and combinations of these two approaches. Our results\nindicate that some benefit can be gained with the inclusion of human generated\nfeedback."", 'Interactive reinforcement learning agents use human feedback or instruction\nto help them learn in complex environments. Often, this feedback comes in the\nform of a discrete signal that is either positive or negative. While\ninformative, this information can be difficult to generalize on its own. In\nthis work, we explore how natural language advice can be used to provide a\nricher feedback signal to a reinforcement learning agent by extending policy\nshaping, a well-known Interactive reinforcement learning technique. Usually\npolicy shaping employs a human feedback policy to help an agent to learn more\nabout how to achieve its goal. In our case, we replace this human feedback\npolicy with policy generated based on natural language advice. We aim to\ninspect if the generated natural language reasoning provides support to a deep\nreinforcement learning agent to decide its actions successfully in any given\nenvironment. So, we design our model with three networks: first one is the\nexperience driven, next is the advice generator and third one is the advice\ndriven. While the experience driven reinforcement learning agent chooses its\nactions being influenced by the environmental reward, the advice driven neural\nnetwork with generated feedback by the advice generator for any new state\nselects its actions to assist the reinforcement learning agent to better policy\nshaping.', ""While recent advances in deep reinforcement learning have allowed autonomous\nlearning agents to succeed at a variety of complex tasks, existing algorithms\ngenerally require a lot of training data. One way to increase the speed at\nwhich agents are able to learn to perform tasks is by leveraging the input of\nhuman trainers. Although such input can take many forms, real-time,\nscalar-valued feedback is especially useful in situations where it proves\ndifficult or impossible for humans to provide expert demonstrations. Previous\napproaches have shown the usefulness of human input provided in this fashion\n(e.g., the TAMER framework), but they have thus far not considered\nhigh-dimensional state spaces or employed the use of deep learning. In this\npaper, we do both: we propose Deep TAMER, an extension of the TAMER framework\nthat leverages the representational power of deep neural networks in order to\nlearn complex tasks in just a short amount of time with a human trainer. We\ndemonstrate Deep TAMER's success by using it and just 15 minutes of\nhuman-provided feedback to train an agent that performs better than humans on\nthe Atari game of Bowling - a task that has proven difficult for even\nstate-of-the-art reinforcement learning methods."", ""We describe a method to use discrete human feedback to enhance the\nperformance of deep learning agents in virtual three-dimensional environments\nby extending deep-reinforcement learning to model the confidence and\nconsistency of human feedback. This enables deep reinforcement learning\nalgorithms to determine the most appropriate time to listen to the human\nfeedback, exploit the current policy model, or explore the agent's environment.\nManaging the trade-off between these three strategies allows DRL agents to be\nrobust to inconsistent or intermittent human feedback. Through experimentation\nusing a synthetic oracle, we show that our technique improves the training\nspeed and overall performance of deep reinforcement learning in navigating\nthree-dimensional environments using Minecraft. We further show that our\ntechnique is robust to highly innacurate human feedback and can also operate\nwhen no human feedback is given."", 'In reinforcement learning (RL), sparse rewards are a natural way to specify\nthe task to be learned. However, most RL algorithms struggle to learn in this\nsetting since the learning signal is mostly zeros. In contrast, humans are good\nat assessing and predicting the future consequences of actions and can serve as\ngood reward/policy shapers to accelerate the robot learning process. Previous\nworks have shown that the human brain generates an error-related signal,\nmeasurable using electroencephelography (EEG), when the human perceives the\ntask being done erroneously. In this work, we propose a method that uses\nevaluative feedback obtained from human brain signals measured via scalp EEG to\naccelerate RL for robotic agents in sparse reward settings. As the robot learns\nthe task, the EEG of a human observer watching the robot attempts is recorded\nand decoded into noisy error feedback signal. From this feedback, we use\nsupervised learning to obtain a policy that subsequently augments the behavior\npolicy and guides exploration in the early stages of RL. This bootstraps the RL\nlearning process to enable learning from sparse reward. Using a robotic\nnavigation task as a test bed, we show that our method achieves a stable\nobstacle-avoidance policy with high success rate, outperforming learning from\nsparse rewards only that struggles to achieve obstacle avoidance behavior or\nfails to advance to the goal.', 'How can we train a dialog model to produce better conversations by learning\nfrom human feedback, without the risk of humans teaching it harmful chat\nbehaviors? We start by hosting models online, and gather human feedback from\nreal-time, open-ended conversations, which we then use to train and improve the\nmodels using offline reinforcement learning (RL). We identify implicit\nconversational cues including language similarity, elicitation of laughter,\nsentiment, and more, which indicate positive human feedback, and embed these in\nmultiple reward functions. A well-known challenge is that learning an RL policy\nin an offline setting usually fails due to the lack of ability to explore and\nthe tendency to make over-optimistic estimates of future reward. These problems\nbecome even harder when using RL for language models, which can easily have a\n20,000 action vocabulary and many possible reward functions. We solve the\nchallenge by developing a novel class of offline RL algorithms. These\nalgorithms use KL-control to penalize divergence from a pre-trained prior\nlanguage model, and use a new strategy to make the algorithm pessimistic,\ninstead of optimistic, in the face of uncertainty. We test the resulting dialog\nmodel with ratings from 80 users in an open-domain setting and find it achieves\nsignificant improvements over existing deep offline RL approaches. The novel\noffline RL method is viable for improving any existing generative dialog model\nusing a static dataset of human feedback.', ""Reinforcement Learning AI commonly uses reward/penalty signals that are\nobjective and explicit in an environment -- e.g. game score, completion time,\netc. -- in order to learn the optimal strategy for task performance. However,\nHuman-AI interaction for such AI agents should include additional reinforcement\nthat is implicit and subjective -- e.g. human preferences for certain AI\nbehavior -- in order to adapt the AI behavior to idiosyncratic human\npreferences. Such adaptations would mirror naturally occurring processes that\nincrease trust and comfort during social interactions. Here, we show how a\nhybrid brain-computer-interface (hBCI), which detects an individual's level of\ninterest in objects/events in a virtual environment, can be used to adapt the\nbehavior of a Deep Reinforcement Learning AI agent that is controlling a\nvirtual autonomous vehicle. Specifically, we show that the AI learns a driving\nstrategy that maintains a safe distance from a lead vehicle, and most novelly,\npreferentially slows the vehicle when the human passengers of the vehicle\nencounter objects of interest. This adaptation affords an additional 20\\%\nviewing time for subjectively interesting objects. This is the first\ndemonstration of how an hBCI can be used to provide implicit reinforcement to\nan AI agent in a way that incorporates user preferences into the control\nsystem."", 'In autonomous vehicle (AV) control, allowing mistakes can be quite dangerous\nand costly in the real world. For this reason we investigate methods of\ntraining an AV without allowing the agent to explore and instead having a human\nexplorer collect the data. Supervised learning has been explored for AV\ncontrol, but it encounters the issue of the covariate shift. That is, training\ndata collected from an optimal demonstration consists only of the states\ninduced by the optimal control policy, but at runtime, the trained agent may\nencounter a vastly different state distribution with little relevant training\ndata. To mitigate this issue, we have our human explorer make sub-optimal\ndecisions. In order to have our agent not replicate these sub-optimal\ndecisions, supervised learning requires that we either erase these actions, or\nreplace these action with the correct action. Erasing is wasteful and replacing\nis difficult, since it is not easy to know the correct action without driving.\nWe propose an alternate framework that includes continuous scalar feedback for\neach action, marking which actions we should replicate, which we should avoid,\nand how sure we are. Our framework learns continuous control from sub-optimal\ndemonstration and evaluative feedback collected before training. We find that a\nhuman demonstrator can explore sub-optimal states in a safe manner, while still\ngetting enough gradation to benefit learning. The collection method for data\nand feedback we call ""Backseat Driver."" We call the more general learning\nframework ReNeg, since it learns a regression from states to actions given\nnegative as well as positive examples. We empirically validate several models\nin the ReNeg framework, testing on lane-following with limited data. We find\nthat the best solution is a generalization of mean-squared error and\noutperforms supervised learning on the positive examples alone.', ""The current reward learning from human preferences could be used to resolve\ncomplex reinforcement learning (RL) tasks without access to a reward function\nby defining a single fixed preference between pairs of trajectory segments.\nHowever, the judgement of preferences between trajectories is not dynamic and\nstill requires human input over thousands of iterations. In this study, we\nproposed a weak human preference supervision framework, for which we developed\na human preference scaling model that naturally reflects the human perception\nof the degree of weak choices between trajectories and established a\nhuman-demonstration estimator via supervised learning to generate the predicted\npreferences for reducing the number of human inputs. The proposed weak human\npreference supervision framework can effectively solve complex RL tasks and\nachieve higher cumulative rewards in simulated robot locomotion -- MuJoCo games\n-- relative to the single fixed human preferences. Furthermore, our established\nhuman-demonstration estimator requires human feedback only for less than 0.01\\%\nof the agent's interactions with the environment and significantly reduces the\ncost of human inputs by up to 30\\% compared with the existing approaches. To\npresent the flexibility of our approach, we released a video\n(https://youtu.be/jQPe1OILT0M) showing comparisons of the behaviours of agents\ntrained on different types of human input. We believe that our naturally\ninspired human preferences with weakly supervised learning are beneficial for\nprecise reward learning and can be applied to state-of-the-art RL systems, such\nas human-autonomy teaming systems."", 'Fluid human-agent communication is essential for the future of\nhuman-in-the-loop reinforcement learning. An agent must respond appropriately\nto feedback from its human trainer even before they have significant experience\nworking together. Therefore, it is important that learning agents respond well\nto various feedback schemes human trainers are likely to provide. This work\nanalyzes the COnvergent Actor-Critic by Humans (COACH) algorithm under three\ndifferent types of feedback-policy feedback, reward feedback, and advantage\nfeedback. For these three feedback types, we find that COACH can behave\nsub-optimally. We propose a variant of COACH, episodic COACH (E-COACH), which\nwe prove converges for all three types. We compare our COACH variant with two\nother reinforcement-learning algorithms: Q-learning and TAMER.', 'To solve complex real-world problems with reinforcement learning, we cannot\nrely on manually specified reward functions. Instead, we can have humans\ncommunicate an objective to the agent directly. In this work, we combine two\napproaches to learning from human feedback: expert demonstrations and\ntrajectory preferences. We train a deep neural network to model the reward\nfunction and use its predicted reward to train an DQN-based deep reinforcement\nlearning agent on 9 Atari games. Our approach beats the imitation learning\nbaseline in 7 games and achieves strictly superhuman performance on 2 games\nwithout using game rewards. Additionally, we investigate the goodness of fit of\nthe reward model, present some reward hacking problems, and study the effects\nof noise in the human labels.', ""This paper extends recent work in interactive machine learning (IML) focused\non effectively incorporating human feedback. We show how control and feedback\nsignals complement each other in systems which model human reward. We\ndemonstrate that simultaneously incorporating human control and feedback\nsignals can improve interactive robotic systems' performance on a self-mirrored\nmovement control task where an RL-agent controlled right arm attempts to match\nthe preprogrammed movement pattern of the left arm. We illustrate the impact of\nvarying human feedback parameters on task performance by investigating the\nprobability of giving feedback on each time step and the likelihood of given\nfeedback being correct. We further illustrate that varying the temporal decay\nwith which the agent incorporates human feedback has a significant impact on\ntask performance. We found that smearing human feedback over time steps\nimproves performance and we show varying the probability of feedback at each\ntime step, and an increased likelihood of those feedbacks being 'correct' can\nimpact agent performance. We conclude that understanding latent variables in\nhuman feedback is crucial for learning algorithms acting in human-machine\ninteraction domains."", 'The standard feedback model of reinforcement learning requires revealing the\nreward of every visited state-action pair. However, in practice, it is often\nthe case that such frequent feedback is not available. In this work, we take a\nfirst step towards relaxing this assumption and require a weaker form of\nfeedback, which we refer to as \\emph{trajectory feedback}. Instead of observing\nthe reward obtained after every action, we assume we only receive a score that\nrepresents the quality of the whole trajectory observed by the agent, namely,\nthe sum of all rewards obtained over this trajectory. We extend reinforcement\nlearning algorithms to this setting, based on least-squares estimation of the\nunknown reward, for both the known and unknown transition model cases, and\nstudy the performance of these algorithms by analyzing their regret. For cases\nwhere the transition model is unknown, we offer a hybrid optimistic-Thompson\nSampling approach that results in a tractable algorithm.', 'Counterfactual learning from human bandit feedback describes a scenario where\nuser feedback on the quality of outputs of a historic system is logged and used\nto improve a target system. We show how to apply this learning framework to\nneural semantic parsing. From a machine learning perspective, the key challenge\nlies in a proper reweighting of the estimator so as to avoid known degeneracies\nin counterfactual learning, while still being applicable to stochastic gradient\noptimization. To conduct experiments with human users, we devise an easy-to-use\ninterface to collect human feedback on semantic parses. Our work is the first\nto show that semantic parsers can be improved significantly by counterfactual\nlearning from logged human feedback data.', ""A promising approach to improve the robustness and exploration in\nReinforcement Learning is collecting human feedback and that way incorporating\nprior knowledge of the target environment. It is, however, often too expensive\nto obtain enough feedback of good quality. To mitigate the issue, we aim to\nrely on a group of multiple experts (and non-experts) with different skill\nlevels to generate enough feedback. Such feedback can therefore be inconsistent\nand infrequent. In this paper, we build upon prior work -- Advise, a Bayesian\napproach attempting to maximise the information gained from human feedback --\nextending the algorithm to accept feedback from this larger group of humans,\nthe trainers, while also estimating each trainer's reliability. We show how\naggregating feedback from multiple trainers improves the total feedback's\naccuracy and make the collection process easier in two ways. Firstly, this\napproach addresses the case of some of the trainers being adversarial.\nSecondly, having access to the information about each trainer reliability\nprovides a second layer of robustness and offers valuable information for\npeople managing the whole system to improve the overall trust in the system. It\noffers an actionable tool for improving the feedback collection process or\nmodifying the reward function design if needed. We empirically show that our\napproach can accurately learn the reliability of each trainer correctly and use\nit to maximise the information gained from the multiple trainers' feedback,\neven if some of the sources are adversarial."", 'We present a study on reinforcement learning (RL) from human bandit feedback\nfor sequence-to-sequence learning, exemplified by the task of bandit neural\nmachine translation (NMT). We investigate the reliability of human bandit\nfeedback, and analyze the influence of reliability on the learnability of a\nreward estimator, and the effect of the quality of reward estimates on the\noverall RL task. Our analysis of cardinal (5-point ratings) and ordinal\n(pairwise preferences) feedback shows that their intra- and inter-annotator\n$\\alpha$-agreement is comparable. Best reliability is obtained for standardized\ncardinal feedback, and cardinal feedback is also easiest to learn and\ngeneralize from. Finally, improvements of over 1 BLEU can be obtained by\nintegrating a regression-based reward estimator trained on cardinal feedback\nfor 800 translations into RL for NMT. This shows that RL is possible even from\nsmall amounts of fairly reliable human feedback, pointing to a great potential\nfor applications at larger scale.', 'Successfully navigating a complex environment to obtain a desired outcome is\na difficult task, that up to recently was believed to be capable only by\nhumans. This perception has been broken down over time, especially with the\nintroduction of deep reinforcement learning, which has greatly increased the\ndifficulty of tasks that can be automated. However, for traditional\nreinforcement learning agents this requires an environment to be able to\nprovide frequent extrinsic rewards, which are not known or accessible for many\nreal-world environments. This project aims to explore and contrast existing\nreinforcement learning solutions that circumnavigate the difficulties of an\nenvironment that provide sparse rewards. Different reinforcement solutions will\nbe implemented over a several video game environments with varying difficulty\nand varying frequency of rewards, as to properly investigate the applicability\nof these solutions. This project introduces a novel reinforcement learning\nsolution by combining aspects of two existing state of the art sparse reward\nsolutions, curiosity driven exploration and unsupervised auxiliary tasks.', 'Recent advancements in \\textit{Learning from Human Feedback} present an\neffective way to train robot agents via inputs from non-expert humans, without\na need for a specially designed reward function. However, this approach needs a\nhuman to be present and attentive during robot learning to provide evaluative\nfeedback. In addition, the amount of feedback needed grows with the level of\ntask difficulty and the quality of human feedback might decrease over time\nbecause of fatigue. To overcome these limitations and enable learning more\nrobot tasks with higher complexities, there is a need to maximize the quality\nof expensive feedback received and reduce the amount of human cognitive\ninvolvement required. In this work, we present an approach that uses active\nlearning to smartly choose queries for the human supervisor based on the\nuncertainty of the robot and effectively reduces the amount of feedback needed\nto learn a given task. We also use a novel multiple buffer system to improve\nrobustness to feedback noise and guard against catastrophic forgetting as the\nrobot learning evolves. This makes it possible to learn tasks with more\ncomplexity using lesser amounts of human feedback compared to previous methods.\nWe demonstrate the utility of our proposed method on a robot arm reaching task\nwhere the robot learns to reach a location in 3D without colliding with\nobstacles. Our approach is able to learn this task faster, with less human\nfeedback and cognitive involvement, compared to previous methods that do not\nuse active learning.', ""Reward function, as an incentive representation that recognizes humans'\nagency and rationalizes humans' actions, is particularly appealing for modeling\nhuman behavior in human-robot interaction. Inverse Reinforcement Learning is an\neffective way to retrieve reward functions from demonstrations. However, it has\nalways been challenging when applying it to multi-agent settings since the\nmutual influence between agents has to be appropriately modeled. To tackle this\nchallenge, previous work either exploits equilibrium solution concepts by\nassuming humans as perfectly rational optimizers with unbounded intelligence or\npre-assigns humans' interaction strategies a priori. In this work, we advocate\nthat humans are bounded rational and have different intelligence levels when\nreasoning about others' decision-making process, and such an inherent and\nlatent characteristic should be accounted for in reward learning algorithms.\nHence, we exploit such insights from Theory-of-Mind and propose a new\nmulti-agent Inverse Reinforcement Learning framework that reasons about humans'\nlatent intelligence levels during learning. We validate our approach in both\nzero-sum and general-sum games with synthetic agents and illustrate a practical\napplication to learning human drivers' reward functions from real driving data.\nWe compare our approach with two baseline algorithms. The results show that by\nreasoning about humans' latent intelligence levels, the proposed approach has\nmore flexibility and capability to retrieve reward functions that explain\nhumans' driving behaviors better."", ""It is often difficult to hand-specify what the correct reward function is for\na task, so researchers have instead aimed to learn reward functions from human\nbehavior or feedback. The types of behavior interpreted as evidence of the\nreward function have expanded greatly in recent years. We've gone from\ndemonstrations, to comparisons, to reading into the information leaked when the\nhuman is pushing the robot away or turning it off. And surely, there is more to\ncome. How will a robot make sense of all these diverse types of behavior? Our\nkey insight is that different types of behavior can be interpreted in a single\nunifying formalism - as a reward-rational choice that the human is making,\noften implicitly. The formalism offers both a unifying lens with which to view\npast work, as well as a recipe for interpreting new sources of information that\nare yet to be uncovered. We provide two examples to showcase this: interpreting\na new feedback type, and reading into how the choice of feedback itself leaks\ninformation about the reward."", 'Traditional collaborative filtering (CF) based recommender systems tend to\nperform poorly when the user-item interactions/ratings are highly scarce. To\naddress this, we propose a learning framework that improves collaborative\nfiltering with a synthetic feedback loop (CF-SFL) to simulate the user\nfeedback. The proposed framework consists of a ""recommender"" and a ""virtual\nuser"". The ""recommender"" is formulated as a CF model, recommending items\naccording to observed user preference. The ""virtual user"" estimates rewards\nfrom the recommended items and generates a \\emph{feedback} in addition to the\nobserved user preference. The ""recommender"" connected with the ""virtual user""\nconstructs a closed loop, that recommends users with items and imitates the\n\\emph{unobserved} feedback of the users to the recommended items. The synthetic\nfeedback is used to augment the observed user preference and improve\nrecommendation results. Theoretically, such model design can be interpreted as\ninverse reinforcement learning, which can be learned effectively via rollout\n(simulation). Experimental results show that the proposed framework is able to\nenrich the learning of user preference and boost the performance of existing\ncollaborative filtering methods on multiple datasets.', 'The capability to interactively learn from human feedback would enable agents\nin new settings. For example, even novice users could train service robots in\nnew tasks naturally and interactively. Human-in-the-loop Reinforcement Learning\n(HRL) combines human feedback and Reinforcement Learning (RL) techniques.\nState-of-the-art interactive learning techniques suffer from slow learning\nspeed, thus leading to a frustrating experience for the human. We approach this\nproblem by extending the HRL framework TAMER for evaluative feedback with the\npossibility to enhance human feedback with two different types of\ncounterfactual explanations (action and state based). We experimentally show\nthat our extensions improve the speed of learning.', 'This paper contributes a first study into how different human users deliver\nsimultaneous control and feedback signals during human-robot interaction. As\npart of this work, we formalize and present a general interactive learning\nframework for online cooperation between humans and reinforcement learning\nagents. In many human-machine interaction settings, there is a growing gap\nbetween the degrees-of-freedom of complex semi-autonomous systems and the\nnumber of human control channels. Simple human control and feedback mechanisms\nare required to close this gap and allow for better collaboration between\nhumans and machines on complex tasks. To better inform the design of concurrent\ncontrol and feedback interfaces, we present experimental results from a\nhuman-robot collaborative domain wherein the human must simultaneously deliver\nboth control and feedback signals to interactively train an actor-critic\nreinforcement learning robot. We compare three experimental conditions: 1)\nhuman delivered control signals, 2) reward-shaping feedback signals, and 3)\nsimultaneous control and feedback. Our results suggest that subjects provide\nless feedback when simultaneously delivering feedback and control signals and\nthat control signal quality is not significantly diminished. Our data suggest\nthat subjects may also modify when and how they provide feedback. Through\nalgorithmic development and tuning informed by this study, we expect\nsemi-autonomous actions of robotic agents can be better shaped by human\nfeedback, allowing for seamless collaboration and improved performance in\ndifficult interactive domains.', ""A long-term goal of reinforcement learning agents is to be able to perform\ntasks in complex real-world scenarios. The use of external information is one\nway of scaling agents to more complex problems. However, there is a general\nlack of collaboration or interoperability between different approaches using\nexternal information. In this work, while reviewing externally-influenced\nmethods, we propose a conceptual framework and taxonomy for assisted\nreinforcement learning, aimed at fostering collaboration by classifying and\ncomparing various methods that use external information in the learning\nprocess. The proposed taxonomy details the relationship between the external\ninformation source and the learner agent, highlighting the process of\ninformation decomposition, structure, retention, and how it can be used to\ninfluence agent learning. As well as reviewing state-of-the-art methods, we\nidentify current streams of reinforcement learning that use external\ninformation in order to improve the agent's performance and its decision-making\nprocess. These include heuristic reinforcement learning, interactive\nreinforcement learning, learning from demonstration, transfer learning, and\nlearning from multiple sources, among others. These streams of reinforcement\nlearning operate with the shared objective of scaffolding the learner agent.\nLastly, we discuss further possibilities for future work in the field of\nassisted reinforcement learning systems."", 'In this paper, we propose a framework that enables a human teacher to shape a\nrobot behaviour by interactively providing it with unlabeled instructions. We\nground the meaning of instruction signals in the task-learning process, and use\nthem simultaneously for guiding the latter. We implement our framework as a\nmodular architecture, named TICS (Task-Instruction-Contingency-Shaping) that\ncombines different information sources: a predefined reward function, human\nevaluative feedback and unlabeled instructions. This approach provides a novel\nperspective for robotic task learning that lies between Reinforcement Learning\nand Supervised Learning paradigms. We evaluate our framework both in simulation\nand with a real robot. The experimental results demonstrate the effectiveness\nof our framework in accelerating the task-learning process and in reducing the\nnumber of required teaching signals.', ""The combination of deep neural network models and reinforcement learning\nalgorithms can make it possible to learn policies for robotic behaviors that\ndirectly read in raw sensory inputs, such as camera images, effectively\nsubsuming both estimation and control into one model. However, real-world\napplications of reinforcement learning must specify the goal of the task by\nmeans of a manually programmed reward function, which in practice requires\neither designing the very same perception pipeline that end-to-end\nreinforcement learning promises to avoid, or else instrumenting the environment\nwith additional sensors to determine if the task has been performed\nsuccessfully. In this paper, we propose an approach for removing the need for\nmanual engineering of reward specifications by enabling a robot to learn from a\nmodest number of examples of successful outcomes, followed by actively\nsolicited queries, where the robot shows the user a state and asks for a label\nto determine whether that state represents successful completion of the task.\nWhile requesting labels for every single state would amount to asking the user\nto manually provide the reward signal, our method requires labels for only a\ntiny fraction of the states seen during training, making it an efficient and\npractical approach for learning skills without manually engineered rewards. We\nevaluate our method on real-world robotic manipulation tasks where the\nobservations consist of images viewed by the robot's camera. In our\nexperiments, our method effectively learns to arrange objects, place books, and\ndrape cloth, directly from images and without any manually specified reward\nfunctions, and with only 1-4 hours of interaction with the real world."", 'Reinforcement learning usually uses the feedback rewards of environmental to\ntrain agents. But the rewards in the actual environment are sparse, and even\nsome environments will not rewards. Most of the current methods are difficult\nto get good performance in sparse reward or non-reward environments. Although\nusing shaped rewards is effective when solving sparse reward tasks, it is\nlimited to specific problems and learning is also susceptible to local optima.\nWe propose a model-free method that does not rely on environmental rewards to\nsolve the problem of sparse rewards in the general environment. Our method use\nthe minimum number of transitions between states as the distance to replace the\nrewards of environmental, and proposes a goal-distance gradient to achieve\npolicy improvement. We also introduce a bridge point planning method based on\nthe characteristics of our method to improve exploration efficiency, thereby\nsolving more complex tasks. Experiments show that our method performs better on\nsparse reward and local optimal problems in complex environments than previous\nwork.', 'Deep reinforcement learning has obtained significant breakthroughs in recent\nyears. Most methods in deep-RL achieve good results via the maximization of the\nreward signal provided by the environment, typically in the form of discounted\ncumulative returns. Such reward signals represent the immediate feedback of a\nparticular action performed by an agent. However, tasks with sparse reward\nsignals are still challenging to on-policy methods. In this paper, we introduce\nan effective characterization of past reward statistics (which can be seen as\nlong-term feedback signals) to supplement this immediate reward feedback. In\nparticular, value functions are learned with multi-critics supervision,\nenabling complex value functions to be more easily approximated in on-policy\nlearning, even when the reward signals are sparse. We also introduce a novel\nexploration mechanism called ""hot-wiring"" that can give a boost to seemingly\ntrapped agents. We demonstrate the effectiveness of our advantage actor\nmulti-critic (A2MC) method across the discrete domains in Atari games as well\nas continuous domains in the MuJoCo environments. A video demo is provided at\nhttps://youtu.be/zBmpf3Yz8tc.', 'As language models become more powerful, training and evaluation are\nincreasingly bottlenecked by the data and metrics used for a particular task.\nFor example, summarization models are often trained to predict human reference\nsummaries and evaluated using ROUGE, but both of these metrics are rough\nproxies for what we really care about -- summary quality. In this work, we show\nthat it is possible to significantly improve summary quality by training a\nmodel to optimize for human preferences. We collect a large, high-quality\ndataset of human comparisons between summaries, train a model to predict the\nhuman-preferred summary, and use that model as a reward function to fine-tune a\nsummarization policy using reinforcement learning. We apply our method to a\nversion of the TL;DR dataset of Reddit posts and find that our models\nsignificantly outperform both human reference summaries and much larger models\nfine-tuned with supervised learning alone. Our models also transfer to CNN/DM\nnews articles, producing summaries nearly as good as the human reference\nwithout any news-specific fine-tuning. We conduct extensive analyses to\nunderstand our human feedback dataset and fine-tuned models We establish that\nour reward model generalizes to new datasets, and that optimizing our reward\nmodel results in better summaries than optimizing ROUGE according to humans. We\nhope the evidence from our paper motivates machine learning researchers to pay\ncloser attention to how their training loss affects the model behavior they\nactually want.', ""Interactive Machine Learning is concerned with creating systems that operate\nin environments alongside humans to achieve a task. A typical use is to extend\nor amplify the capabilities of a human in cognitive or physical ways, requiring\nthe machine to adapt to the users' intentions and preferences. Often, this\ntakes the form of a human operator providing some type of feedback to the user,\nwhich can be explicit feedback, implicit feedback, or a combination of both.\nExplicit feedback, such as through a mouse click, carries a high cognitive\nload. The focus of this study is to extend the current state of the art in\ninteractive machine learning by demonstrating that agents can learn a human\nuser's behavior and adapt to preferences with a reduced amount of explicit\nhuman feedback in a mixed feedback setting. The learning agent perceives a\nvalue of its own behavior from hand gestures given via a spatial interface.\nThis feedback mechanism is termed Spatial Interface Valuing. This method is\nevaluated experimentally in a simulated environment for a grasping task using a\nrobotic arm with variable grip settings. Preliminary results indicate that\nlearning agents using spatial interface valuing can learn a value function\nmapping spatial gestures to expected future rewards much more quickly as\ncompared to those same agents just receiving explicit feedback, demonstrating\nthat an agent perceiving feedback from a human user via a spatial interface can\nserve as an effective complement to existing approaches."", 'Most research in Bayesian optimization (BO) has focused on \\emph{direct\nfeedback} scenarios, where one has access to exact values of some\nexpensive-to-evaluate objective. This direction has been mainly driven by the\nuse of BO in machine learning hyper-parameter configuration problems. However,\nin domains such as modelling human preferences, A/B tests, or recommender\nsystems, there is a need for methods that can replace direct feedback with\n\\emph{preferential feedback}, obtained via rankings or pairwise comparisons. In\nthis work, we present preferential batch Bayesian optimization (PBBO), a new\nframework that allows finding the optimum of a latent function of interest,\ngiven any type of parallel preferential feedback for a group of two or more\npoints. We do so by using a Gaussian process model with a likelihood specially\ndesigned to enable parallel and efficient data collection mechanisms, which are\nkey in modern machine learning. We show how the acquisitions developed under\nthis framework generalize and augment previous approaches in Bayesian\noptimization, expanding the use of these techniques to a wider range of\ndomains. An extensive simulation study shows the benefits of this approach,\nboth with simulated functions and four real data sets.', 'The history of learning for control has been an exciting back and forth\nbetween two broad classes of algorithms: planning and reinforcement learning.\nPlanning algorithms effectively reason over long horizons, but assume access to\na local policy and distance metric over collision-free paths. Reinforcement\nlearning excels at learning policies and the relative values of states, but\nfails to plan over long horizons. Despite the successes of each method in\nvarious domains, tasks that require reasoning over long horizons with limited\nfeedback and high-dimensional observations remain exceedingly challenging for\nboth planning and reinforcement learning algorithms. Frustratingly, these sorts\nof tasks are potentially the most useful, as they are simple to design (a human\nonly need to provide an example goal state) and avoid reward shaping, which can\nbias the agent towards finding a sub-optimal solution. We introduce a general\ncontrol algorithm that combines the strengths of planning and reinforcement\nlearning to effectively solve these tasks. Our aim is to decompose the task of\nreaching a distant goal state into a sequence of easier tasks, each of which\ncorresponds to reaching a subgoal. Planning algorithms can automatically find\nthese waypoints, but only if provided with suitable abstractions of the\nenvironment -- namely, a graph consisting of nodes and edges. Our main insight\nis that this graph can be constructed via reinforcement learning, where a\ngoal-conditioned value function provides edge weights, and nodes are taken to\nbe previously seen observations in a replay buffer. Using graph search over our\nreplay buffer, we can automatically generate this sequence of subgoals, even in\nimage-based environments. Our algorithm, search on the replay buffer (SoRB),\nenables agents to solve sparse reward tasks over one hundred steps, and\ngeneralizes substantially better than standard RL algorithms.', ""Interactive reinforcement learning provides a way for agents to learn to\nsolve tasks from evaluative feedback provided by a human user. Previous\nresearch showed that humans give copious feedback early in training but very\nsparsely thereafter. In this article, we investigate the potential of agent\nlearning from trainers' facial expressions via interpreting them as evaluative\nfeedback. To do so, we implemented TAMER which is a popular interactive\nreinforcement learning method in a reinforcement-learning benchmark problem ---\nInfinite Mario, and conducted the first large-scale study of TAMER involving\n561 participants. With designed CNN-RNN model, our analysis shows that telling\ntrainers to use facial expressions and competition can improve the accuracies\nfor estimating positive and negative feedback using facial expressions. In\naddition, our results with a simulation experiment show that learning solely\nfrom predicted feedback based on facial expressions is possible and using\nstrong/effective prediction models or a regression method, facial responses\nwould significantly improve the performance of agents. Furthermore, our\nexperiment supports previous studies demonstrating the importance of\nbi-directional feedback and competitive elements in the training interface."", 'In offline reinforcement learning (RL) agents are trained using a logged\ndataset. It appears to be the most natural route to attack real-life\napplications because in domains such as healthcare and robotics interactions\nwith the environment are either expensive or unethical. Training agents usually\nrequires reward functions, but unfortunately, rewards are seldom available in\npractice and their engineering is challenging and laborious. To overcome this,\nwe investigate reward learning under the constraint of minimizing human reward\nannotations. We consider two types of supervision: timestep annotations and\ndemonstrations. We propose semi-supervised learning algorithms that learn from\nlimited annotations and incorporate unlabelled data. In our experiments with a\nsimulated robotic arm, we greatly improve upon behavioural cloning and closely\napproach the performance achieved with ground truth rewards. We further\ninvestigate the relationship between the quality of the reward model and the\nfinal policies. We notice, for example, that the reward models do not need to\nbe perfect to result in useful policies.', ""Mixed-initiative systems allow users to interactively provide feedback to\npotentially improve system performance. Human feedback can correct model errors\nand update model parameters to dynamically adapt to changing data.\nAdditionally, many users desire the ability to have a greater level of control\nand fix perceived flaws in systems they rely on. However, how the ability to\nprovide feedback to autonomous systems influences user trust is a largely\nunexplored area of research. Our research investigates how the act of providing\nfeedback can affect user understanding of an intelligent system and its\naccuracy. We present a controlled experiment using a simulated object detection\nsystem with image data to study the effects of interactive feedback collection\non user impressions. The results show that providing human-in-the-loop feedback\nlowered both participants' trust in the system and their perception of system\naccuracy, regardless of whether the system accuracy improved in response to\ntheir feedback. These results highlight the importance of considering the\neffects of allowing end-user feedback on user trust when designing intelligent\nsystems."", 'Robots will eventually be part of every household. It is thus critical to\nenable algorithms to learn from and be guided by non-expert users. In this\npaper, we bring a human in the loop, and enable a human teacher to give\nfeedback to a learning agent in the form of natural language. We argue that a\ndescriptive sentence can provide a much stronger learning signal than a numeric\nreward in that it can easily point to where the mistakes are and how to correct\nthem. We focus on the problem of image captioning in which the quality of the\noutput can easily be judged by non-experts. We propose a hierarchical\nphrase-based captioning model trained with policy gradients, and design a\nfeedback network that provides reward to the learner by conditioning on the\nhuman-provided feedback. We show that by exploiting descriptive feedback our\nmodel learns to perform better than when given independently written human\ncaptions.', 'Human feedback is widely used to train agents in many domains. However,\nprevious works rarely consider the uncertainty when humans provide feedback,\nespecially in cases that the optimal actions are not obvious to the trainers.\nFor example, the reward of a sub-optimal action can be stochastic and sometimes\nexceeds that of the optimal action, which is common in games or real-world.\nTrainers are likely to provide positive feedback to sub-optimal actions,\nnegative feedback to the optimal actions and even do not provide feedback in\nsome confusing situations. Existing works, which utilize the Expectation\nMaximization (EM) algorithm and treat the feedback model as hidden parameters,\ndo not consider uncertainties in the learning environment and human feedback.\nTo address this challenge, we introduce a novel feedback model that considers\nthe uncertainty of human feedback. However, this incurs intractable calculus in\nthe EM algorithm. To this end, we propose a novel approximate EM algorithm, in\nwhich we approximate the expectation step with the Gradient Descent method.\nExperimental results in both synthetic scenarios and two real-world scenarios\nwith human participants demonstrate the superior performance of our proposed\napproach.', 'This article surveys reinforcement learning approaches in social robotics.\nReinforcement learning is a framework for decision-making problems in which an\nagent interacts through trial-and-error with its environment to discover an\noptimal behavior. Since interaction is a key component in both reinforcement\nlearning and social robotics, it can be a well-suited approach for real-world\ninteractions with physically embodied social robots. The scope of the paper is\nfocused particularly on studies that include social physical robots and\nreal-world human-robot interactions with users. We present a thorough analysis\nof reinforcement learning approaches in social robotics. In addition to a\nsurvey, we categorize existent reinforcement learning approaches based on the\nused method and the design of the reward mechanisms. Moreover, since\ncommunication capability is a prominent feature of social robots, we discuss\nand group the papers based on the communication medium used for reward\nformulation. Considering the importance of designing the reward function, we\nalso provide a categorization of the papers based on the nature of the reward.\nThis categorization includes three major themes: interactive reinforcement\nlearning, intrinsically motivated methods, and task performance-driven methods.\nThe benefits and challenges of reinforcement learning in social robotics,\nevaluation methods of the papers regarding whether or not they use subjective\nand algorithmic measures, a discussion in the view of real-world reinforcement\nlearning challenges and proposed solutions, the points that remain to be\nexplored, including the approaches that have thus far received less attention\nis also given in the paper. Thus, this paper aims to become a starting point\nfor researchers interested in using and applying reinforcement learning methods\nin this particular research field.', 'Existing methods for interactive image retrieval have demonstrated the merit\nof integrating user feedback, improving retrieval results. However, most\ncurrent systems rely on restricted forms of user feedback, such as binary\nrelevance responses, or feedback based on a fixed set of relative attributes,\nwhich limits their impact. In this paper, we introduce a new approach to\ninteractive image search that enables users to provide feedback via natural\nlanguage, allowing for more natural and effective interaction. We formulate the\ntask of dialog-based interactive image retrieval as a reinforcement learning\nproblem, and reward the dialog system for improving the rank of the target\nimage during each dialog turn. To mitigate the cumbersome and costly process of\ncollecting human-machine conversations as the dialog system learns, we train\nour system with a user simulator, which is itself trained to describe the\ndifferences between target and candidate images. The efficacy of our approach\nis demonstrated in a footwear retrieval application. Experiments on both\nsimulated and real-world data show that 1) our proposed learning framework\nachieves better accuracy than other supervised and reinforcement learning\nbaselines and 2) user feedback based on natural language rather than\npre-specified attributes leads to more effective retrieval results, and a more\nnatural and expressive communication interface.', 'We explore unconstrained natural language feedback as a learning signal for\nartificial agents. Humans use rich and varied language to teach, yet most prior\nwork on interactive learning from language assumes a particular form of input\n(e.g., commands). We propose a general framework which does not make this\nassumption, using aspect-based sentiment analysis to decompose feedback into\nsentiment about the features of a Markov decision process. We then perform an\nanalogue of inverse reinforcement learning, regressing the sentiment on the\nfeatures to infer the teacher\'s latent reward function. To evaluate our\napproach, we first collect a corpus of teaching behavior in a cooperative task\nwhere both teacher and learner are human. We implement three artificial\nlearners: sentiment-based ""literal"" and ""pragmatic"" models, and an inference\nnetwork trained end-to-end to predict latent rewards. We then repeat our\ninitial experiment and pair them with human teachers. All three successfully\nlearn from interactive human feedback. The sentiment models outperform the\ninference network, with the ""pragmatic"" model approaching human performance.\nOur work thus provides insight into the information structure of naturalistic\nlinguistic feedback as well as methods to leverage it for reinforcement\nlearning.', ""A key challenge in Imitation Learning (IL) is that optimal state actions\ndemonstrations are difficult for the teacher to provide. For example in\nrobotics, providing kinesthetic demonstrations on a robotic manipulator\nrequires the teacher to control multiple degrees of freedom at once. The\ndifficulty of requiring optimal state action demonstrations limits the space of\nproblems where the teacher can provide quality feedback. As an alternative to\nstate action demonstrations, the teacher can provide corrective feedback such\nas their preferences or rewards. Prior work has created algorithms designed to\nlearn from specific types of noisy feedback, but across teachers and tasks\ndifferent forms of feedback may be required. Instead we propose that in order\nto learn from a diversity of scenarios we need to learn from a variety of\nfeedback. To learn from a variety of feedback we make the following insight:\nthe teacher's cost function is latent and we can model a stream of feedback as\na stream of loss functions. We then use any online learning algorithm to\nminimize the sum of these losses. With this insight we can learn from a\ndiversity of feedback that is weakly correlated with the teacher's true cost\nfunction. We unify prior work into a general corrective feedback meta-algorithm\nand show that regardless of feedback we can obtain the same regret bounds. We\ndemonstrate our approach by learning to perform a household navigation task on\na robotic racecar platform. Our results show that our approach can learn\nquickly from a variety of noisy feedback."", 'Deep reinforcement learning has achieved many impressive results in recent\nyears. However, tasks with sparse rewards or long horizons continue to pose\nsignificant challenges. To tackle these important problems, we propose a\ngeneral framework that first learns useful skills in a pre-training\nenvironment, and then leverages the acquired skills for learning faster in\ndownstream tasks. Our approach brings together some of the strengths of\nintrinsic motivation and hierarchical methods: the learning of useful skill is\nguided by a single proxy reward, the design of which requires very minimal\ndomain knowledge about the downstream tasks. Then a high-level policy is\ntrained on top of these skills, providing a significant improvement of the\nexploration and allowing to tackle sparse rewards in the downstream tasks. To\nefficiently pre-train a large span of skills, we use Stochastic Neural Networks\ncombined with an information-theoretic regularizer. Our experiments show that\nthis combination is effective in learning a wide span of interpretable skills\nin a sample-efficient way, and can significantly boost the learning performance\nuniformly across a wide range of downstream tasks.', 'The last decade has seen a significant increase of interest in deep learning\nresearch, with many public successes that have demonstrated its potential. As\nsuch, these systems are now being incorporated into commercial products. With\nthis comes an additional challenge: how can we build AI systems that solve\ntasks where there is not a crisp, well-defined specification? While multiple\nsolutions have been proposed, in this competition we focus on one in\nparticular: learning from human feedback. Rather than training AI systems using\na predefined reward function or using a labeled dataset with a predefined set\nof categories, we instead train the AI system using a learning signal derived\nfrom some form of human feedback, which can evolve over time as the\nunderstanding of the task changes, or as the capabilities of the AI system\nimprove.\n  The MineRL BASALT competition aims to spur forward research on this important\nclass of techniques. We design a suite of four tasks in Minecraft for which we\nexpect it will be hard to write down hardcoded reward functions. These tasks\nare defined by a paragraph of natural language: for example, ""create a\nwaterfall and take a scenic picture of it"", with additional clarifying details.\nParticipants must train a separate agent for each task, using any method they\nwant. Agents are then evaluated by humans who have read the task description.\nTo help participants get started, we provide a dataset of human demonstrations\non each of the four tasks, as well as an imitation learning baseline that\nleverages these demonstrations.\n  Our hope is that this competition will improve our ability to build AI\nsystems that do what their designers intend them to do, even when the intent\ncannot be easily formalized. Besides allowing AI to solve more tasks, this can\nalso enable more effective regulation of AI systems, as well as making progress\non the value alignment problem.', 'Human-computer interactive systems that rely on machine learning are becoming\nparamount to the lives of millions of people who use digital assistants on a\ndaily basis. Yet, further advances are limited by the availability of data and\nthe cost of acquiring new samples. One way to address this problem is by\nimproving the sample efficiency of current approaches. As a solution path, we\npresent a model-based reinforcement learning algorithm for an interactive\ndialogue task. We build on commonly used actor-critic methods, adding an\nenvironment model and planner that augments a learning agent to learn the model\nof the environment dynamics. Our results show that, on a simulation that mimics\nthe interactive task, our algorithm requires 70 times fewer samples, compared\nto the baseline of commonly used model-free algorithm, and demonstrates 2~times\nbetter performance asymptotically. Moreover, we introduce a novel contribution\nof computing a soft planner policy and further updating a model-free policy\nyielding a less computationally expensive model-free agent as good as the\nmodel-based one. This model-based architecture serves as a foundation that can\nbe extended to other human-computer interactive tasks allowing further advances\nin this direction.', ""Recent advances in reinforcement learning have inspired increasing interest\nin learning user modeling adaptively through dynamic interactions, e.g., in\nreinforcement learning based recommender systems. Reward function is crucial\nfor most of reinforcement learning applications as it can provide the guideline\nabout the optimization. However, current reinforcement-learning-based methods\nrely on manually-defined reward functions, which cannot adapt to dynamic and\nnoisy environments. Besides, they generally use task-specific reward functions\nthat sacrifice generalization ability. We propose a generative inverse\nreinforcement learning for user behavioral preference modelling, to address the\nabove issues. Instead of using predefined reward functions, our model can\nautomatically learn the rewards from user's actions based on discriminative\nactor-critic network and Wasserstein GAN. Our model provides a general way of\ncharacterizing and explaining underlying behavioral tendencies, and our\nexperiments show our method outperforms state-of-the-art methods in a variety\nof scenarios, namely traffic signal control, online recommender systems, and\nscanpath prediction."", 'Reinforcement learning is a powerful learning paradigm in which agents can\nlearn to maximize sparse and delayed reward signals. Although RL has had many\nimpressive successes in complex domains, learning can take hours, days, or even\nyears of training data. A major challenge of contemporary RL research is to\ndiscover how to learn with less data. Previous work has shown that domain\ninformation can be successfully used to shape the reward; by adding additional\nreward information, the agent can learn with much less data. Furthermore, if\nthe reward is constructed from a potential function, the optimal policy is\nguaranteed to be unaltered. While such potential-based reward shaping (PBRS)\nholds promise, it is limited by the need for a well-defined potential function.\nIdeally, we would like to be able to take arbitrary advice from a human or\nother agent and improve performance without affecting the optimal policy. The\nrecently introduced dynamic potential based advice (DPBA) method tackles this\nchallenge by admitting arbitrary advice from a human or other agent and\nimproves performance without affecting the optimal policy. The main\ncontribution of this paper is to expose, theoretically and empirically, a flaw\nin DPBA. Alternatively, to achieve the ideal goals, we present a simple method\ncalled policy invariant explicit shaping (PIES) and show theoretically and\nempirically that PIES succeeds where DPBA fails.', 'While recent progress in deep reinforcement learning has enabled robots to\nlearn complex behaviors, tasks with long horizons and sparse rewards remain an\nongoing challenge. In this work, we propose an effective reward shaping method\nthrough predictive coding to tackle sparse reward problems. By learning\npredictive representations offline and using these representations for reward\nshaping, we gain access to reward signals that understand the structure and\ndynamics of the environment. In particular, our method achieves better learning\nby providing reward signals that 1) understand environment dynamics 2)\nemphasize on features most useful for learning 3) resist noise in learned\nrepresentations through reward accumulation. We demonstrate the usefulness of\nthis approach in different domains ranging from robotic manipulation to\nnavigation, and we show that reward signals produced through predictive coding\nare as effective for learning as hand-crafted rewards.', ""The ubiquitous nature of chatbots and their interaction with users generate\nan enormous amount of data. Can we improve chatbots using this data? A\nself-feeding chatbot improves itself by asking natural language feedback when a\nuser is dissatisfied with its response and uses this feedback as an additional\ntraining sample. However, user feedback in most cases contains extraneous\nsequences hindering their usefulness as a training sample. In this work, we\npropose a generative adversarial model that converts noisy feedback into a\nplausible natural response in a conversation. The generator's goal is to\nconvert the feedback into a response that answers the user's previous utterance\nand to fool the discriminator which distinguishes feedback from natural\nresponses. We show that augmenting original training data with these modified\nfeedback responses improves the original chatbot performance from 69.94% to\n75.96% in ranking correct responses on the Personachat dataset, a large\nimprovement given that the original model is already trained on 131k samples."", 'This paper addresses the problem of learning a task from demonstration. We\nadopt the framework of inverse reinforcement learning, where tasks are\nrepresented in the form of a reward function. Our contribution is a novel\nactive learning algorithm that enables the learning agent to query the expert\nfor more informative demonstrations, thus leading to more sample-efficient\nlearning. For this novel algorithm (Generalized Binary Search for Inverse\nReinforcement Learning, or GBS-IRL), we provide a theoretical bound on sample\ncomplexity and illustrate its applicability on several different tasks. To our\nknowledge, GBS-IRL is the first active IRL algorithm with provable sample\ncomplexity bounds. We also discuss our method in light of other existing\nmethods in the literature and its general applicability in multi-class\nclassification problems. Finally, motivated by recent work on learning from\ndemonstration in robots, we also discuss how different forms of human feedback\ncan be integrated in a transparent manner in our learning framework.', ""We propose a general framework for interactively learning models, such as\n(binary or non-binary) classifiers, orderings/rankings of items, or clusterings\nof data points. Our framework is based on a generalization of Angluin's\nequivalence query model and Littlestone's online learning model: in each\niteration, the algorithm proposes a model, and the user either accepts it or\nreveals a specific mistake in the proposal. The feedback is correct only with\nprobability $p > 1/2$ (and adversarially incorrect with probability $1 - p$),\ni.e., the algorithm must be able to learn in the presence of arbitrary noise.\nThe algorithm's goal is to learn the ground truth model using few iterations.\n  Our general framework is based on a graph representation of the models and\nuser feedback. To be able to learn efficiently, it is sufficient that there be\na graph $G$ whose nodes are the models and (weighted) edges capture the user\nfeedback, with the property that if $s, s^*$ are the proposed and target\nmodels, respectively, then any (correct) user feedback $s'$ must lie on a\nshortest $s$-$s^*$ path in $G$. Under this one assumption, there is a natural\nalgorithm reminiscent of the Multiplicative Weights Update algorithm, which\nwill efficiently learn $s^*$ even in the presence of noise in the user's\nfeedback.\n  From this general result, we rederive with barely any extra effort classic\nresults on learning of classifiers and a recent result on interactive\nclustering; in addition, we easily obtain new interactive learning algorithms\nfor ordering/ranking."", 'The sense of touch, being the earliest sensory system to develop in a human\nbody [1], plays a critical part of our daily interaction with the environment.\nIn order to successfully complete a task, many manipulation interactions\nrequire incorporating haptic feedback. However, manually designing a feedback\nmechanism can be extremely challenging. In this work, we consider manipulation\ntasks that need to incorporate tactile sensor feedback in order to modify a\nprovided nominal plan. To incorporate partial observation, we present a new\nframework that models the task as a partially observable Markov decision\nprocess (POMDP) and learns an appropriate representation of haptic feedback\nwhich can serve as the state for a POMDP model. The model, that is parametrized\nby deep recurrent neural networks, utilizes variational Bayes methods to\noptimize the approximate posterior. Finally, we build on deep Q-learning to be\nable to select the optimal action in each state without access to a simulator.\nWe test our model on a PR2 robot for multiple tasks of turning a knob until it\nclicks.', ""An important application of interactive machine learning is extending or\namplifying the cognitive and physical capabilities of a human. To accomplish\nthis, machines need to learn about their human users' intentions and adapt to\ntheir preferences. In most current research, a user has conveyed preferences to\na machine using explicit corrective or instructive feedback; explicit feedback\nimposes a cognitive load on the user and is expensive in terms of human effort.\nThe primary objective of the current work is to demonstrate that a learning\nagent can reduce the amount of explicit feedback required for adapting to the\nuser's preferences pertaining to a task by learning to perceive a value of its\nbehavior from the human user, particularly from the user's facial\nexpressions---we call this face valuing. We empirically evaluate face valuing\non a grip selection task. Our preliminary results suggest that an agent can\nquickly adapt to a user's changing preferences with minimal explicit feedback\nby learning a value function that maps facial features extracted from a camera\nimage to expected future reward. We believe that an agent learning to perceive\na value from the body language of its human user is complementary to existing\ninteractive machine learning approaches and will help in creating successful\nhuman-machine interactive applications."", 'Recent progress in Reinforcement Learning (RL), fueled by its combination,\nwith Deep Learning has enabled impressive results in learning to interact with\ncomplex virtual environments, yet real-world applications of RL are still\nscarce. A key limitation is data efficiency, with current state-of-the-art\napproaches requiring millions of training samples. A promising way to tackle\nthis problem is to augment RL with learning from human demonstrations. However,\nhuman demonstration data is not yet readily available. This hinders progress in\nthis direction. The present work addresses this problem as follows. We (i)\ncollect and describe a large dataset of human Atari 2600 replays -- the largest\nand most diverse such data set publicly released to date, (ii) illustrate an\nexample use of this dataset by analyzing the relation between demonstration\nquality and imitation learning performance, and (iii) outline possible research\ndirections that are opened up by our work.']","[""As robots enter human environments, they will be expected to accomplish a\ntremendous range of tasks. It is not feasible for robot designers to\npre-program these behaviors or know them in advance, so one way to address this\nis through end-user programming, such as via learning from demonstration (LfD).\nWhile significant work has been done on the mechanics of enabling robot\nlearning from human teachers, one unexplored aspect is enabling mutual feedback\nbetween both the human teacher and robot during the learning process, i.e.,\nimplicit learning. In this paper, we explore one aspect of this mutual\nunderstanding, grounding sequences, where both a human and robot provide\nnon-verbal feedback to signify their mutual understanding during interaction.\nWe conducted a study where people taught an autonomous humanoid robot a dance,\nand performed gesture analysis to measure people's responses to the robot\nduring correct and incorrect demonstrations."", 'Many real-world systems problems require reasoning about the long term\nconsequences of actions taken to configure and manage the system. These\nproblems with delayed and often sequentially aggregated reward, are often\ninherently reinforcement learning problems and present the opportunity to\nleverage the recent substantial advances in deep reinforcement learning.\nHowever, in some cases, it is not clear why deep reinforcement learning is a\ngood fit for the problem. Sometimes, it does not perform better than the\nstate-of-the-art solutions. And in other cases, random search or greedy\nalgorithms could outperform deep reinforcement learning. In this paper, we\nreview, discuss, and evaluate the recent trends of using deep reinforcement\nlearning in system optimization. We propose a set of essential metrics to guide\nfuture works in evaluating the efficacy of using deep reinforcement learning in\nsystem optimization. Our evaluation includes challenges, the types of problems,\ntheir formulation in the deep reinforcement learning setting, embedding, the\nmodel used, efficiency, and robustness. We conclude with a discussion on open\nchallenges and potential directions for pushing further the integration of\nreinforcement learning in system optimization.', ""Text-based games are a natural challenge domain for deep reinforcement\nlearning algorithms. Their state and action spaces are combinatorially large,\ntheir reward function is sparse, and they are partially observable: the agent\nis informed of the consequences of its actions through textual feedback. In\nthis paper we emphasize this latter point and consider the design of a deep\nreinforcement learning agent that can play from feedback alone. Our design\nrecognizes and takes advantage of the structural characteristics of text-based\ngames. We first propose a contextualisation mechanism, based on accumulated\nreward, which simplifies the learning problem and mitigates partial\nobservability. We then study different methods that rely on the notion that\nmost actions are ineffectual in any given situation, following Zahavy et al.'s\nidea of an admissible action. We evaluate these techniques in a series of\ntext-based games of increasing difficulty based on the TextWorld framework, as\nwell as the iconic game Zork. Empirically, we find that these techniques\nimprove the performance of a baseline deep reinforcement learning agent applied\nto text-based games."", 'We propose a novel reinforcement learning based framework PoBRL for solving\nmulti-document summarization. PoBRL jointly optimizes over the following three\nobjectives necessary for a high-quality summary: importance, relevance, and\nlength. Our strategy decouples this multi-objective optimization into different\nsubproblems that can be solved individually by reinforcement learning.\nUtilizing PoBRL, we then blend each learned policies together to produce a\nsummary that is a concise and complete representation of the original input.\nOur empirical analysis shows state-of-the-art performance on several\nmulti-document datasets. Human evaluation also shows that our method produces\nhigh-quality output.', 'Imitation learning is an effective approach for autonomous systems to acquire\ncontrol policies when an explicit reward function is unavailable, using\nsupervision provided as demonstrations from an expert, typically a human\noperator. However, standard imitation learning methods assume that the agent\nreceives examples of observation-action tuples that could be provided, for\ninstance, to a supervised learning algorithm. This stands in contrast to how\nhumans and animals imitate: we observe another person performing some behavior\nand then figure out which actions will realize that behavior, compensating for\nchanges in viewpoint, surroundings, object positions and types, and other\nfactors. We term this kind of imitation learning ""imitation-from-observation,""\nand propose an imitation learning method based on video prediction with context\ntranslation and deep reinforcement learning. This lifts the assumption in\nimitation learning that the demonstration should consist of observations in the\nsame environment configuration, and enables a variety of interesting\napplications, including learning robotic skills that involve tool use simply by\nobserving videos of human tool use. Our experimental results show the\neffectiveness of our approach in learning a wide range of real-world robotic\ntasks modeled after common household chores from videos of a human\ndemonstrator, including sweeping, ladling almonds, pushing objects as well as a\nnumber of tasks in simulation.', 'In the artificial intelligence field, learning often corresponds to changing\nthe parameters of a parameterized function. A learning rule is an algorithm or\nmathematical expression that specifies precisely how the parameters should be\nchanged. When creating an artificial intelligence system, we must make two\ndecisions: what representation should be used (i.e., what parameterized\nfunction should be used) and what learning rule should be used to search\nthrough the resulting set of representable functions. Using most learning\nrules, these two decisions are coupled in a subtle (and often unintentional)\nway. That is, using the same learning rule with two different representations\nthat can represent the same sets of functions can result in two different\noutcomes. After arguing that this coupling is undesirable, particularly when\nusing artificial neural networks, we present a method for partially decoupling\nthese two decisions for a broad class of learning rules that span unsupervised\nlearning, reinforcement learning, and supervised learning.', 'Despite widespread interests in reinforcement-learning for task-oriented\ndialogue systems, several obstacles can frustrate research and development\nprogress. First, reinforcement learners typically require interaction with the\nenvironment, so conventional dialogue corpora cannot be used directly. Second,\neach task presents specific challenges, requiring separate corpus of\ntask-specific annotated data. Third, collecting and annotating human-machine or\nhuman-human conversations for task-oriented dialogues requires extensive domain\nknowledge. Because building an appropriate dataset can be both financially\ncostly and time-consuming, one popular approach is to build a user simulator\nbased upon a corpus of example dialogues. Then, one can train reinforcement\nlearning agents in an online fashion as they interact with the simulator.\nDialogue agents trained on these simulators can serve as an effective starting\npoint. Once agents master the simulator, they may be deployed in a real\nenvironment to interact with humans, and continue to be trained online. To ease\nempirical algorithmic comparisons in dialogues, this paper introduces a new,\npublicly available simulation framework, where our simulator, designed for the\nmovie-booking domain, leverages both rules and collected data. The simulator\nsupports two tasks: movie ticket booking and movie seeking. Finally, we\ndemonstrate several agents and detail the procedure to add and test your own\nagent in the proposed framework.', 'Learning in sparse reward settings remains a challenge in Reinforcement\nLearning, which is often addressed by using intrinsic rewards. One promising\nstrategy is inspired by human curiosity, requiring the agent to learn to\npredict the future. In this paper a curiosity-driven agent is extended to use\nthese predictions directly for training. To achieve this, the agent predicts\nthe value function of the next state at any point in time. Subsequently, the\nconsistency of this prediction with the current value function is measured,\nwhich is then used as a regularization term in the loss function of the\nalgorithm. Experiments were made on grid-world environments as well as on a 3D\nnavigation task, both with sparse rewards. In the first case the extended agent\nis able to learn significantly faster than the baselines.', 'While deep reinforcement learning techniques have led to agents that are\nsuccessfully able to learn to perform a number of tasks that had been\npreviously unlearnable, these techniques are still susceptible to the\nlongstanding problem of reward sparsity. This is especially true for tasks such\nas training an agent to play StarCraft II, a real-time strategy game where\nreward is only given at the end of a game which is usually very long. While\nthis problem can be addressed through reward shaping, such approaches typically\nrequire a human expert with specialized knowledge. Inspired by the vision of\nenabling reward shaping through the more-accessible paradigm of\nnatural-language narration, we develop a technique that can provide the\nbenefits of reward shaping using natural language commands. Our\nnarration-guided RL agent projects sequences of natural-language commands into\nthe same high-dimensional representation space as corresponding goal states. We\nshow that we can get improved performance with our method compared to\ntraditional reward-shaping approaches. Additionally, we demonstrate the ability\nof our method to generalize to unseen natural-language commands.', ""What is the role of real-time control and learning in the formation of social\nconventions? To answer this question, we propose a computational model that\nmatches human behavioral data in a social decision-making game that was\nanalyzed both in discrete-time and continuous-time setups. Furthermore, unlike\nprevious approaches, our model takes into account the role of sensorimotor\ncontrol loops in embodied decision-making scenarios. For this purpose, we\nintroduce the Control-based Reinforcement Learning (CRL) model. CRL is grounded\nin the Distributed Adaptive Control (DAC) theory of mind and brain, where\nlow-level sensorimotor control is modulated through perceptual and behavioral\nlearning in a layered structure. CRL follows these principles by implementing a\nfeedback control loop handling the agent's reactive behaviors (pre-wired\nreflexes), along with an adaptive layer that uses reinforcement learning to\nmaximize long-term reward. We test our model in a multi-agent game-theoretic\ntask in which coordination must be achieved to find an optimal solution. We\nshow that CRL is able to reach human-level performance on standard\ngame-theoretic metrics such as efficiency in acquiring rewards and fairness in\nreward distribution."", 'The recent history of machine learning research has taught us that machine\nlearning methods can be most effective when they are provided with very large,\nhigh-capacity models, and trained on very large and diverse datasets. This has\nspurred the community to search for ways to remove any bottlenecks to scale.\nOften the foremost among such bottlenecks is the need for human effort,\nincluding the effort of curating and labeling datasets. As a result,\nconsiderable attention in recent years has been devoted to utilizing unlabeled\ndata, which can be collected in vast quantities. However, some of the most\nwidely used methods for training on such unlabeled data themselves require\nhuman-designed objective functions that must correlate in some meaningful way\nto downstream tasks. I will argue that a general, principled, and powerful\nframework for utilizing unlabeled data can be derived from reinforcement\nlearning, using general purpose unsupervised or self-supervised reinforcement\nlearning objectives in concert with offline reinforcement learning methods that\ncan leverage large datasets. I will discuss how such a procedure is more\nclosely aligned with potential downstream tasks, and how it could build on\nexisting techniques that have been developed in recent years.', 'This paper presents a comprehensive literature review on applications of deep\nreinforcement learning in communications and networking. Modern networks, e.g.,\nInternet of Things (IoT) and Unmanned Aerial Vehicle (UAV) networks, become\nmore decentralized and autonomous. In such networks, network entities need to\nmake decisions locally to maximize the network performance under uncertainty of\nnetwork environment. Reinforcement learning has been efficiently used to enable\nthe network entities to obtain the optimal policy including, e.g., decisions or\nactions, given their states when the state and action spaces are small.\nHowever, in complex and large-scale networks, the state and action spaces are\nusually large, and the reinforcement learning may not be able to find the\noptimal policy in reasonable time. Therefore, deep reinforcement learning, a\ncombination of reinforcement learning with deep learning, has been developed to\novercome the shortcomings. In this survey, we first give a tutorial of deep\nreinforcement learning from fundamental concepts to advanced models. Then, we\nreview deep reinforcement learning approaches proposed to address emerging\nissues in communications and networking. The issues include dynamic network\naccess, data rate control, wireless caching, data offloading, network security,\nand connectivity preservation which are all important to next generation\nnetworks such as 5G and beyond. Furthermore, we present applications of deep\nreinforcement learning for traffic routing, resource sharing, and data\ncollection. Finally, we highlight important challenges, open issues, and future\nresearch directions of applying deep reinforcement learning.', 'The availability of large-scale datasets has driven the development of neural\nmodels that create summaries from single documents, for generic purposes. When\nusing a summarization system, users often have specific intents with various\nlanguage realizations, which, depending on the information need, can range from\na single keyword to a long narrative composed of multiple questions. Existing\nsummarization systems, however, often either fail to support or act robustly on\nthis query focused summarization task. We introduce LaQSum, the first unified\ntext summarization system that learns Latent Queries from documents for\nabstractive summarization with any existing query forms. Under a deep\ngenerative framework, our system jointly optimizes a latent query model and a\nconditional language model, allowing users to plug-and-play queries of any type\nat test time. Despite learning from only generic summarization data and\nrequiring no further optimization for downstream summarization tasks, our\nsystem robustly outperforms strong comparison systems across summarization\nbenchmarks with different query types, document settings, and target domains.', 'Reinforcement learning is a learning paradigm for solving sequential\ndecision-making problems. Recent years have witnessed remarkable progress in\nreinforcement learning upon the fast development of deep neural networks. Along\nwith the promising prospects of reinforcement learning in numerous domains such\nas robotics and game-playing, transfer learning has arisen to tackle various\nchallenges faced by reinforcement learning, by transferring knowledge from\nexternal expertise to facilitate the efficiency and effectiveness of the\nlearning process. In this survey, we systematically investigate the recent\nprogress of transfer learning approaches in the context of deep reinforcement\nlearning. Specifically, we provide a framework for categorizing the\nstate-of-the-art transfer learning approaches, under which we analyze their\ngoals, methodologies, compatible reinforcement learning backbones, and\npractical applications. We also draw connections between transfer learning and\nother relevant topics from the reinforcement learning perspective and explore\ntheir potential challenges that await future research progress.', 'In reinforcement learning, pre-trained low-level skills have the potential to\ngreatly facilitate exploration. However, prior knowledge of the downstream task\nis required to strike the right balance between generality (fine-grained\ncontrol) and specificity (faster learning) in skill design. In previous work on\ncontinuous control, the sensitivity of methods to this trade-off has not been\naddressed explicitly, as locomotion provides a suitable prior for navigation\ntasks, which have been of foremost interest. In this work, we analyze this\ntrade-off for low-level policy pre-training with a new benchmark suite of\ndiverse, sparse-reward tasks for bipedal robots. We alleviate the need for\nprior knowledge by proposing a hierarchical skill learning framework that\nacquires skills of varying complexity in an unsupervised manner. For\nutilization on downstream tasks, we present a three-layered hierarchical\nlearning algorithm to automatically trade off between general and specific\nskills as required by the respective task. In our experiments, we show that our\napproach performs this trade-off effectively and achieves better results than\ncurrent state-of-the-art methods for end- to-end hierarchical reinforcement\nlearning and unsupervised skill discovery. Code and videos are available at\nhttps://facebookresearch.github.io/hsd3 .', 'Reinforcement learning has seen great advancements in the past five years.\nThe successful introduction of deep learning in place of more traditional\nmethods allowed reinforcement learning to scale to very complex domains\nachieving super-human performance in environments like the game of Go or\nnumerous video games. Despite great successes in multiple domains, these new\nmethods suffer from their own issues that make them often inapplicable to the\nreal world problems. Extreme lack of data efficiency, together with huge\nvariance and difficulty in enforcing safety constraints, is one of the three\nmost prominent issues in the field. Usually, millions of data points sampled\nfrom the environment are necessary for these algorithms to converge to\nacceptable policies.\n  This thesis proposes novel Generative Adversarial Imaginative Reinforcement\nLearning algorithm. It takes advantage of the recent introduction of highly\neffective generative adversarial models, and Markov property that underpins\nreinforcement learning setting, to model dynamics of the real environment\nwithin the internal imagination module. Rollouts from the imagination are then\nused to artificially simulate the real environment in a standard reinforcement\nlearning process to avoid, often expensive and dangerous, trial and error in\nthe real environment. Experimental results show that the proposed algorithm\nmore economically utilises experience from the real environment than the\ncurrent state-of-the-art Rainbow DQN algorithm, and thus makes an important\nstep towards sample efficient deep reinforcement learning.', 'We propose Object-oriented Neural Programming (OONP), a framework for\nsemantically parsing documents in specific domains. Basically, OONP reads a\ndocument and parses it into a predesigned object-oriented data structure\n(referred to as ontology in this paper) that reflects the domain-specific\nsemantics of the document. An OONP parser models semantic parsing as a decision\nprocess: a neural net-based Reader sequentially goes through the document, and\nduring the process it builds and updates an intermediate ontology to summarize\nits partial understanding of the text it covers. OONP supports a rich family of\noperations (both symbolic and differentiable) for composing the ontology, and a\nbig variety of forms (both symbolic and differentiable) for representing the\nstate and the document. An OONP parser can be trained with supervision of\ndifferent forms and strength, including supervised learning (SL) ,\nreinforcement learning (RL) and hybrid of the two. Our experiments on both\nsynthetic and real-world document parsing tasks have shown that OONP can learn\nto handle fairly complicated ontology with training data of modest sizes.', 'Deep reinforcement learning is revolutionizing the artificial intelligence\nfield. Currently, it serves as a good starting point for constructing\nintelligent autonomous systems which offer a better knowledge of the visual\nworld. It is possible to scale deep reinforcement learning with the use of deep\nlearning and do amazing tasks such as use of pixels in playing video games. In\nthis paper, key concepts of deep reinforcement learning including reward\nfunction, differences between reinforcement learning and supervised learning\nand models for implementation of reinforcement are discussed. Key challenges\nrelated to the implementation of reinforcement learning in conversational AI\ndomain are identified as well as discussed in detail. Various conversational\nmodels which are based on deep reinforcement learning (as well as deep\nlearning) are also discussed. In summary, this paper discusses key aspects of\ndeep reinforcement learning which are crucial for designing an efficient\nconversational AI.', 'Humans have mental models that allow them to plan, experiment, and reason in\nthe physical world. How should an intelligent agent go about learning such\nmodels? In this paper, we will study if models of the world learned in an\nopen-ended physics environment, without any specific tasks, can be reused for\ndownstream physics reasoning tasks. To this end, we build a benchmark\nOpen-ended Physics ENvironment (OPEn) and also design several tasks to test\nlearning representations in this environment explicitly. This setting reflects\nthe conditions in which real agents (i.e. rolling robots) find themselves,\nwhere they may be placed in a new kind of environment and must adapt without\nany teacher to tell them how this environment works. This setting is\nchallenging because it requires solving an exploration problem in addition to a\nmodel building and representation learning problem. We test several existing\nRL-based exploration methods on this benchmark and find that an agent using\nunsupervised contrastive learning for representation learning, and\nimpact-driven learning for exploration, achieved the best results. However, all\nmodels still fall short in sample efficiency when transferring to the\ndownstream tasks. We expect that OPEn will encourage the development of novel\nrolling robot agents that can build reusable mental models of the world that\nfacilitate many tasks.', 'Model-free reinforcement learning has been successfully applied to a range of\nchallenging problems, and has recently been extended to handle large neural\nnetwork policies and value functions. However, the sample complexity of\nmodel-free algorithms, particularly when using high-dimensional function\napproximators, tends to limit their applicability to physical systems. In this\npaper, we explore algorithms and representations to reduce the sample\ncomplexity of deep reinforcement learning for continuous control tasks. We\npropose two complementary techniques for improving the efficiency of such\nalgorithms. First, we derive a continuous variant of the Q-learning algorithm,\nwhich we call normalized adantage functions (NAF), as an alternative to the\nmore commonly used policy gradient and actor-critic methods. NAF representation\nallows us to apply Q-learning with experience replay to continuous tasks, and\nsubstantially improves performance on a set of simulated robotic control tasks.\nTo further improve the efficiency of our approach, we explore the use of\nlearned models for accelerating model-free reinforcement learning. We show that\niteratively refitted local linear models are especially effective for this, and\ndemonstrate substantially faster learning on domains where such models are\napplicable.', 'We present a reinforcement learning based framework for human-centered\ncollaborative systems. The framework is proactive and balances the benefits of\ntimely actions with the risk of taking improper actions by minimizing the total\ntime spent to complete the task. The framework is learned end-to-end in an\nunsupervised fashion addressing the perception uncertainties and decision\nmaking in an integrated manner. The framework is shown to provide more fluent\ncoordination between human and robot partners on an example task of packaging\ncompared to alternatives for which perception and decision-making systems are\nlearned independently, using supervised learning. The foremost benefit of the\nproposed approach is that it allows for fast adaptation to new human partners\nand tasks since tedious annotation of motion data is avoided and the learning\nis performed on-line.', 'In recent years deep neural networks have been successfully applied to the\ndomains of reinforcement learning\n\\cite{bengio2009learning,krizhevsky2012imagenet,hinton2006reducing}. Deep\nreinforcement learning \\cite{mnih2015human} is reported to have the advantage\nof learning effective policies directly from high-dimensional sensory inputs\nover traditional agents. However, within the scope of the literature, there is\nno fundamental change or improvement on the existing training framework. Here\nwe propose a novel training framework that is conceptually comprehensible and\npotentially easy to be generalized to all feasible algorithms for reinforcement\nlearning. We employ Monte-carlo sampling to achieve raw data inputs, and train\nthem in batch to achieve Markov decision process sequences and synchronously\nupdate the network parameters instead of experience replay. This training\nframework proves to optimize the unbiased approximation of loss function whose\nestimation exactly matches the real probability distribution data inputs\nfollow, and thus have overwhelming advantages of sample efficiency and\nconvergence rate over existing deep reinforcement learning after evaluating it\non both discrete action spaces and continuous control problems. Besides, we\npropose several algorithms embedded with our new framework to deal with typical\ndiscrete and continuous scenarios. These algorithms prove to be far more\nefficient than their original versions under the framework of deep\nreinforcement learning, and provide examples for existing and future algorithms\nto generalize to our new framework.', ""Intelligent assistants that follow commands or answer simple questions, such\nas Siri and Google search, are among the most economically important\napplications of AI. Future conversational AI assistants promise even greater\ncapabilities and a better user experience through a deeper understanding of the\ndomain, the user, or the user's purposes. But what domain and what methods are\nbest suited to researching and realizing this promise? In this article we argue\nfor the domain of voice document editing and for the methods of model-based\nreinforcement learning. The primary advantages of voice document editing are\nthat the domain is tightly scoped and that it provides something for the\nconversation to be about (the document) that is delimited and fully accessible\nto the intelligent assistant. The advantages of reinforcement learning in\ngeneral are that its methods are designed to learn from interaction without\nexplicit instruction and that it formalizes the purposes of the assistant.\nModel-based reinforcement learning is needed in order to genuinely understand\nthe domain of discourse and thereby work efficiently with the user to achieve\ntheir goals. Together, voice document editing and model-based reinforcement\nlearning comprise a promising research direction for achieving conversational\nAI."", ""Measuring user satisfaction level is a challenging task, and a critical\ncomponent in developing large-scale conversational agent systems serving the\nneeds of real users. An widely used approach to tackle this is to collect human\nannotation data and use them for evaluation or modeling. Human annotation based\napproaches are easier to control, but hard to scale. A novel alternative\napproach is to collect user's direct feedback via a feedback elicitation system\nembedded to the conversational agent system, and use the collected user\nfeedback to train a machine-learned model for generalization. User feedback is\nthe best proxy for user satisfaction, but is not available for some ineligible\nintents and certain situations. Thus, these two types of approaches are\ncomplementary to each other. In this work, we tackle the user satisfaction\nassessment problem with a hybrid approach that fuses explicit user feedback,\nuser satisfaction predictions inferred by two machine-learned models, one\ntrained on user feedback data and the other human annotation data. The hybrid\napproach is based on a waterfall policy, and the experimental results with\nAmazon Alexa's large-scale datasets show significant improvements in inferring\nuser satisfaction. A detailed hybrid architecture, an in-depth analysis on user\nfeedback data, and an algorithm that generates data sets to properly simulate\nthe live traffic are presented in this paper."", ""Reinforcement learning provides a powerful and flexible framework for\nautomated acquisition of robotic motion skills. However, applying reinforcement\nlearning requires a sufficiently detailed representation of the state,\nincluding the configuration of task-relevant objects. We present an approach\nthat automates state-space construction by learning a state representation\ndirectly from camera images. Our method uses a deep spatial autoencoder to\nacquire a set of feature points that describe the environment for the current\ntask, such as the positions of objects, and then learns a motion skill with\nthese feature points using an efficient reinforcement learning method based on\nlocal linear models. The resulting controller reacts continuously to the\nlearned feature points, allowing the robot to dynamically manipulate objects in\nthe world with closed-loop control. We demonstrate our method with a PR2 robot\non tasks that include pushing a free-standing toy block, picking up a bag of\nrice using a spatula, and hanging a loop of rope on a hook at various\npositions. In each task, our method automatically learns to track task-relevant\nobjects and manipulate their configuration with the robot's arm."", ""We consider the problem of imitation learning from a finite set of expert\ntrajectories, without access to reinforcement signals. The classical approach\nof extracting the expert's reward function via inverse reinforcement learning,\nfollowed by reinforcement learning is indirect and may be computationally\nexpensive. Recent generative adversarial methods based on matching the policy\ndistribution between the expert and the agent could be unstable during\ntraining. We propose a new framework for imitation learning by estimating the\nsupport of the expert policy to compute a fixed reward function, which allows\nus to re-frame imitation learning within the standard reinforcement learning\nsetting. We demonstrate the efficacy of our reward function on both discrete\nand continuous domains, achieving comparable or better performance than the\nstate of the art under different reinforcement learning algorithms."", 'The study of unsupervised learning can be generally divided into two\ncategories: imitation learning and reinforcement learning. In imitation\nlearning the machine learns by mimicking the behavior of an expert system\nwhereas in reinforcement learning the machine learns via direct environment\nfeedback. Traditional deep reinforcement learning takes a significant time\nbefore the machine starts to converge to an optimal policy. This paper proposes\nAugmented Q-Imitation-Learning, a method by which deep reinforcement learning\nconvergence can be accelerated by applying Q-imitation-learning as the initial\ntraining process in traditional Deep Q-learning.', ""Consider learning a policy from example expert behavior, without interaction\nwith the expert or access to reinforcement signal. One approach is to recover\nthe expert's cost function with inverse reinforcement learning, then extract a\npolicy from that cost function with reinforcement learning. This approach is\nindirect and can be slow. We propose a new general framework for directly\nextracting a policy from data, as if it were obtained by reinforcement learning\nfollowing inverse reinforcement learning. We show that a certain instantiation\nof our framework draws an analogy between imitation learning and generative\nadversarial networks, from which we derive a model-free imitation learning\nalgorithm that obtains significant performance gains over existing model-free\nmethods in imitating complex behaviors in large, high-dimensional environments."", 'Artificial Intelligence has been developed for decades with the achievement\nof great progress. Recently, deep learning shows its ability to solve many real\nworld problems, e.g. image classification and detection, natural language\nprocessing, playing GO. Theoretically speaking, an artificial neural network\ncan fit any function and reinforcement learning can learn from any delayed\nreward. But in solving real world tasks, we still need to spend a lot of effort\nto adjust algorithms to fit task unique features. This paper proposes that the\nreason of this phenomenon is the sparse feedback feature of the nature, and a\nsingle algorithm, no matter how we improve it, can only solve dense feedback\ntasks or specific sparse feedback tasks. This paper first analyses how sparse\nfeedback affects algorithm perfomance, and then proposes a pattern that\nexplains how to accumulate knowledge to solve sparse feedback problems.', 'Deep reinforcement learning augments the reinforcement learning framework and\nutilizes the powerful representation of deep neural networks. Recent works have\ndemonstrated the remarkable successes of deep reinforcement learning in various\ndomains including finance, medicine, healthcare, video games, robotics, and\ncomputer vision. In this work, we provide a detailed review of recent and\nstate-of-the-art research advances of deep reinforcement learning in computer\nvision. We start with comprehending the theories of deep learning,\nreinforcement learning, and deep reinforcement learning. We then propose a\ncategorization of deep reinforcement learning methodologies and discuss their\nadvantages and limitations. In particular, we divide deep reinforcement\nlearning into seven main categories according to their applications in computer\nvision, i.e. (i)landmark localization (ii) object detection; (iii) object\ntracking; (iv) registration on both 2D image and 3D image volumetric data (v)\nimage segmentation; (vi) videos analysis; and (vii) other applications. Each of\nthese categories is further analyzed with reinforcement learning techniques,\nnetwork design, and performance. Moreover, we provide a comprehensive analysis\nof the existing publicly available datasets and examine source code\navailability. Finally, we present some open issues and discuss future research\ndirections on deep reinforcement learning in computer vision', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', 'Reinforcement learning has recently experienced increased prominence in the\nmachine learning community. There are many approaches to solving reinforcement\nlearning problems with new techniques developed constantly. When solving\nproblems using reinforcement learning, there are various difficult challenges\nto overcome. To ensure progress in the field, benchmarks are important for\ntesting new algorithms and comparing with other approaches. The reproducibility\nof results for fair comparison is therefore vital in ensuring that improvements\nare accurately judged. This paper provides an overview of different\ncontributions to reinforcement learning benchmarking and discusses how they can\nassist researchers to address the challenges facing reinforcement learning. The\ncontributions discussed are the most used and recent in the literature. The\npaper discusses the contributions in terms of implementation, tasks and\nprovided algorithm implementations with benchmarks. The survey aims to bring\nattention to the wide range of reinforcement learning benchmarking tasks\navailable and to encourage research to take place in a standardised manner.\nAdditionally, this survey acts as an overview for researchers not familiar with\nthe different tasks that can be used to develop and test new reinforcement\nlearning algorithms.', 'Reinforcement learning is one of the core components in designing an\nartificial intelligent system emphasizing real-time response. Reinforcement\nlearning influences the system to take actions within an arbitrary environment\neither having previous knowledge about the environment model or not. In this\npaper, we present a comprehensive study on Reinforcement Learning focusing on\nvarious dimensions including challenges, the recent development of different\nstate-of-the-art techniques, and future directions. The fundamental objective\nof this paper is to provide a framework for the presentation of available\nmethods of reinforcement learning that is informative enough and simple to\nfollow for the new researchers and academics in this domain considering the\nlatest concerns. First, we illustrated the core techniques of reinforcement\nlearning in an easily understandable and comparable way. Finally, we analyzed\nand depicted the recent developments in reinforcement learning approaches. My\nanalysis pointed out that most of the models focused on tuning policy values\nrather than tuning other things in a particular state of reasoning.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'Cross-lingual document representations enable language understanding in\nmultilingual contexts and allow transfer learning from high-resource to\nlow-resource languages at the document level. Recently large pre-trained\nlanguage models such as BERT, XLM and XLM-RoBERTa have achieved great success\nwhen fine-tuned on sentence-level downstream tasks. It is tempting to apply\nthese cross-lingual models to document representation learning. However, there\nare two challenges: (1) these models impose high costs on long document\nprocessing and thus many of them have strict length limit; (2) model\nfine-tuning requires extra data and computational resources, which is not\npractical in resource-limited settings. In this work, we address these\nchallenges by proposing unsupervised Language-Agnostic Weighted Document\nRepresentations (LAWDR). We study the geometry of pre-trained sentence\nembeddings and leverage it to derive document representations without\nfine-tuning. Evaluated on cross-lingual document alignment, LAWDR demonstrates\ncomparable performance to state-of-the-art models on benchmark datasets.', 'How to leverage cross-document interactions to improve ranking performance is\nan important topic in information retrieval (IR) research. However, this topic\nhas not been well-studied in the learning-to-rank setting and most of the\nexisting work still treats each document independently while scoring. The\nrecent development of deep learning shows strength in modeling complex\nrelationships across sequences and sets. It thus motivates us to study how to\nleverage cross-document interactions for learning-to-rank in the deep learning\nframework. In this paper, we formally define the permutation-equivariance\nrequirement for a scoring function that captures cross-document interactions.\nWe then propose a self-attention based document interaction network and show\nthat it satisfies the permutation-equivariant requirement, and can generate\nscores for document sets of varying sizes. Our proposed methods can\nautomatically learn to capture document interactions without any auxiliary\ninformation, and can scale across large document sets. We conduct experiments\non three ranking datasets: the benchmark Web30k, a Gmail search, and a Google\nDrive Quick Access dataset. Experimental results show that our proposed methods\nare both more effective and efficient than baselines.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', ""Self-supervised metric learning has been a successful approach for learning a\ndistance from an unlabeled dataset. The resulting distance is broadly useful\nfor improving various distance-based downstream tasks, even when no information\nfrom downstream tasks is utilized in the metric learning stage. To gain\ninsights into this approach, we develop a statistical framework to\ntheoretically study how self-supervised metric learning can benefit downstream\ntasks in the context of multi-view data. Under this framework, we show that the\ntarget distance of metric learning satisfies several desired properties for the\ndownstream tasks. On the other hand, our investigation suggests the target\ndistance can be further improved by moderating each direction's weights. In\naddition, our analysis precisely characterizes the improvement by\nself-supervised metric learning on four commonly used downstream tasks: sample\nidentification, two-sample testing, $k$-means clustering, and $k$-nearest\nneighbor classification. When the distance is estimated from an unlabeled\ndataset, we establish the upper bound on distance estimation's accuracy and the\nnumber of samples sufficient for downstream task improvement. Finally,\nnumerical experiments are presented to support the theoretical results in the\npaper."", 'Model selection is treated as a standard performance boosting step in many\nmachine learning applications. Once all other properties of a learning problem\nare fixed, the model is selected by grid search on a held-out validation set.\nThis is strictly inapplicable to active learning. Within the standardized\nworkflow, the acquisition function is chosen among available heuristics a\npriori, and its success is observed only after the labeling budget is already\nexhausted. More importantly, none of the earlier studies report a unique\nconsistently successful acquisition heuristic to the extent to stand out as the\nunique best choice. We present a method to break this vicious circle by\ndefining the acquisition function as a learning predictor and training it by\nreinforcement feedback collected from each labeling round. As active learning\nis a scarce data regime, we bootstrap from a well-known heuristic that filters\nthe bulk of data points on which all heuristics would agree, and learn a policy\nto warp the top portion of this ranking in the most beneficial way for the\ncharacter of a specific data distribution. Our system consists of a Bayesian\nneural net, the predictor, a bootstrap acquisition function, a probabilistic\nstate definition, and another Bayesian policy network that can effectively\nincorporate this input distribution. We observe on three benchmark data sets\nthat our method always manages to either invent a new superior acquisition\nfunction or to adapt itself to the a priori unknown best performing heuristic\nfor each specific data set.', ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'While significant research advances have been made in the field of deep\nreinforcement learning, there have been no concrete adversarial attack\nstrategies in literature tailored for studying the vulnerability of deep\nreinforcement learning algorithms to membership inference attacks. In such\nattacking systems, the adversary targets the set of collected input data on\nwhich the deep reinforcement learning algorithm has been trained. To address\nthis gap, we propose an adversarial attack framework designed for testing the\nvulnerability of a state-of-the-art deep reinforcement learning algorithm to a\nmembership inference attack. In particular, we design a series of experiments\nto investigate the impact of temporal correlation, which naturally exists in\nreinforcement learning training data, on the probability of information\nleakage. Moreover, we compare the performance of \\emph{collective} and\n\\emph{individual} membership attacks against the deep reinforcement learning\nalgorithm. Experimental results show that the proposed adversarial attack\nframework is surprisingly effective at inferring data with an accuracy\nexceeding $84\\%$ in individual and $97\\%$ in collective modes in three\ndifferent continuous control Mujoco tasks, which raises serious privacy\nconcerns in this regard. Finally, we show that the learning state of the\nreinforcement learning algorithm influences the level of privacy breaches\nsignificantly.', 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'Novel contexts may often arise in complex querying scenarios such as in\nevidence-based medicine (EBM) involving biomedical literature, that may not\nexplicitly refer to entities or canonical concept forms occurring in any fact-\nor rule-based knowledge source such as an ontology like the UMLS. Moreover,\nhidden associations between candidate concepts meaningful in the current\ncontext, may not exist within a single document, but within the collection, via\nalternate lexical forms. Therefore, inspired by the recent success of\nsequence-to-sequence neural models in delivering the state-of-the-art in a wide\nrange of NLP tasks, we develop a novel sequence-to-set framework with neural\nattention for learning document representations that can effect term transfer\nwithin the corpus, for semantically tagging a large collection of documents. We\ndemonstrate that our proposed method can be effective in both a supervised\nmulti-label classification setup for text categorization, as well as in a\nunique unsupervised setting with no human-annotated document labels that uses\nno external knowledge resources and only corpus-derived term statistics to\ndrive the training. Further, we show that semi-supervised training using our\narchitecture on large amounts of unlabeled data can augment performance on the\ntext categorization task when limited labeled data is available. Our approach\nto generate document encodings employing our sequence-to-set models for\ninference of semantic tags, gives to the best of our knowledge, the\nstate-of-the-art for both, the unsupervised query expansion task for the TREC\nCDS 2016 challenge dataset when evaluated on an Okapi BM25--based document\nretrieval system; and also over the MLTM baseline (Soleimani et al, 2016), for\nboth supervised and semi-supervised multi-label prediction tasks on the\ndel.icio.us and Ohsumed datasets. We will make our code and data publicly\navailable.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'We consider the framework of Document Modeling, which lays the formal basis\nfor representing the document lifecycle in Business Process Management systems.\nWe formulate document models in the scope of the logic-based Semantic Modeling\nlanguage and study the question whether transactions given by a document model\nterminate on any input. We show that in general this problem is undecidable and\nformulate sufficient conditions, which guarantee decidability and polynomial\nboundedness of effects of transactions.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : As an environmental researcher, I aim to create a mobile application that uses generative adversarial networks (GANs) to produce personalized, vivid images illustrating the potential impacts of climate change on specific locations. My plan is to train the GAN on street-view images of horses before and after extreme weather events. However, since such a dataset does not exist, I require a model capable of learning the 3D structure of an object from images taken from various angles and subsequently rendering an image of the structure from a specified direction. Furthermore, I intend for the application to incorporate climate models to evaluate the probability and type of long-term climate-related events. As I am developing a mobile application, I also need techniques to reduce the latency between the mobile device and the backend server, as well as strategies to manage multiple simultaneous calls to the backend server. Lastly, I am interested in video compression algorithms that can enhance the upload speed and alleviate the storage pressure on the server.",[],"['In this paper we present, to the best of our knowledge, the first method to\nlearn a generative model of 3D shapes from natural images in a fully\nunsupervised way. For example, we do not use any ground truth 3D or 2D\nannotations, stereo video, and ego-motion during the training. Our approach\nfollows the general strategy of Generative Adversarial Networks, where an image\ngenerator network learns to create image samples that are realistic enough to\nfool a discriminator network into believing that they are natural images. In\ncontrast, in our approach the image generation is split into 2 stages. In the\nfirst stage a generator network outputs 3D objects. In the second, a\ndifferentiable renderer produces an image of the 3D objects from random\nviewpoints. The key observation is that a realistic 3D object should yield a\nrealistic rendering from any plausible viewpoint. Thus, by randomizing the\nchoice of the viewpoint our proposed training forces the generator network to\nlearn an interpretable 3D representation disentangled from the viewpoint. In\nthis work, a 3D representation consists of a triangle mesh and a texture map\nthat is used to color the triangle surface by using the UV-mapping technique.\nWe provide analysis of our learning approach, expose its ambiguities and show\nhow to overcome them. Experimentally, we demonstrate that our method can learn\nrealistic 3D shapes of faces by using only the natural images of the FFHQ\ndataset.', 'Current Generative Adversarial Networks (GANs) produce photorealistic\nrenderings of portrait images. Embedding real images into the latent space of\nsuch models enables high-level image editing. While recent methods provide\nconsiderable semantic control over the (re-)generated images, they can only\ngenerate a limited set of viewpoints and cannot explicitly control the camera.\nSuch 3D camera control is required for 3D virtual and mixed reality\napplications. In our solution, we use a few images of a face to perform 3D\nreconstruction, and we introduce the notion of the GAN camera manifold, the key\nelement allowing us to precisely define the range of images that the GAN can\nreproduce in a stable manner. We train a small face-specific neural implicit\nrepresentation network to map a captured face to this manifold and complement\nit with a warping scheme to obtain free-viewpoint novel-view synthesis. We show\nhow our approach - due to its precise camera control - enables the integration\nof a pre-trained StyleGAN into standard 3D rendering pipelines, allowing e.g.,\nstereo rendering or consistent insertion of faces in synthetic 3D environments.\nOur solution proposes the first truly free-viewpoint rendering of realistic\nfaces at interactive rates, using only a small number of casual photos as\ninput, while simultaneously allowing semantic editing capabilities, such as\nfacial expression or lighting changes.', 'Climate change is a major threat to humanity, and the actions required to\nprevent its catastrophic consequences include changes in both policy-making and\nindividual behaviour. However, taking action requires understanding the effects\nof climate change, even though they may seem abstract and distant. Projecting\nthe potential consequences of extreme climate events such as flooding in\nfamiliar places can help make the abstract impacts of climate change more\nconcrete and encourage action. As part of a larger initiative to build a\nwebsite that projects extreme climate events onto user-chosen photos, we\npresent our solution to simulate photo-realistic floods on authentic images. To\naddress this complex task in the absence of suitable training data, we propose\nClimateGAN, a model that leverages both simulated and real data for\nunsupervised domain adaptation and conditional image generation. In this paper,\nwe describe the details of our framework, thoroughly evaluate components of our\narchitecture and demonstrate that our model is capable of robustly generating\nphoto-realistic flooding.', 'Detecting extreme events in large datasets is a major challenge in climate\nscience research. Current algorithms for extreme event detection are build upon\nhuman expertise in defining events based on subjective thresholds of relevant\nphysical variables. Often, multiple competing methods produce vastly different\nresults on the same dataset. Accurate characterization of extreme events in\nclimate simulations and observational data archives is critical for\nunderstanding the trends and potential impacts of such events in a climate\nchange content. This study presents the first application of Deep Learning\ntechniques as alternative methodology for climate extreme events detection.\nDeep neural networks are able to learn high-level representations of a broad\nclass of patterns from labeled data. In this work, we developed deep\nConvolutional Neural Network (CNN) classification system and demonstrated the\nusefulness of Deep Learning technique for tackling climate pattern detection\nproblems. Coupled with Bayesian based hyper-parameter optimization scheme, our\ndeep CNN system achieves 89\\%-99\\% of accuracy in detecting extreme events\n(Tropical Cyclones, Atmospheric Rivers and Weather Fronts', 'As climate change increases the intensity of natural disasters, society needs\nbetter tools for adaptation. Floods, for example, are the most frequent natural\ndisaster, but during hurricanes the area is largely covered by clouds and\nemergency managers must rely on nonintuitive flood visualizations for mission\nplanning. To assist these emergency managers, we have created a deep learning\npipeline that generates visual satellite images of current and future coastal\nflooding. We advanced a state-of-the-art GAN called pix2pixHD, such that it\nproduces imagery that is physically-consistent with the output of an\nexpert-validated storm surge model (NOAA SLOSH). By evaluating the imagery\nrelative to physics-based flood maps, we find that our proposed framework\noutperforms baseline models in both physical-consistency and photorealism.\nWhile this work focused on the visualization of coastal floods, we envision the\ncreation of a global visualization of how climate change will shape our earth.', 'Using generative models for Inverse Graphics is an active area of research.\nHowever, most works focus on developing models for supervised and\nsemi-supervised methods. In this paper, we study the problem of unsupervised\nlearning of 3D geometry from single images. Our approach is to use a generative\nmodel that produces 2-D images as projections of a latent 3D voxel grid, which\nwe train either as a variational auto-encoder or using adversarial methods. Our\ncontributions are as follows: First, we show how to recover 3D shape and pose\nfrom general datasets such as MNIST, and MNIST Fashion in good quality. Second,\nwe compare the shapes learned using adversarial and variational methods.\nAdversarial approach gives denser 3D shapes. Third, we explore the idea of\nmodelling the pose of an object as uniform distribution to recover 3D shape\nfrom a single image. Our experiment with the CelebA dataset\n\\cite{liu2015faceattributes} proves that we can recover complete 3D shape from\na single image when the object is symmetric along one, or more axis whilst\nresults obtained using ModelNet40 \\cite{wu20153d} show the potential\nside-effects, in which the model learns 3D shapes such that it can render the\nsame image from any viewpoint. Forth, we present a general end-to-end approach\nto learning 3D shapes from single images in a completely unsupervised fashion\nby modelling the factors of variation such as azimuth as independent latent\nvariables. Our method makes no assumptions about the dataset, and can work with\nsynthetic as well as real images (i.e. unsupervised in true sense). We present\nour results, by training the model using the $\\mu$-VAE objective\n\\cite{ucar2019bridging} and a dataset combining all images from MNIST, MNIST\nFashion, CelebA and six categories of ModelNet40. The model is able to learn 3D\nshapes and the pose in qood quality and leverages information learned across\nall datasets.', ""We propose a learned image-guided rendering technique that combines the\nbenefits of image-based rendering and GAN-based image synthesis. The goal of\nour method is to generate photo-realistic re-renderings of reconstructed\nobjects for virtual and augmented reality applications (e.g., virtual\nshowrooms, virtual tours \\& sightseeing, the digital inspection of historical\nartifacts). A core component of our work is the handling of view-dependent\neffects. Specifically, we directly train an object-specific deep neural network\nto synthesize the view-dependent appearance of an object. As input data we are\nusing an RGB video of the object. This video is used to reconstruct a proxy\ngeometry of the object via multi-view stereo. Based on this 3D proxy, the\nappearance of a captured view can be warped into a new target view as in\nclassical image-based rendering. This warping assumes diffuse surfaces, in case\nof view-dependent effects, such as specular highlights, it leads to artifacts.\nTo this end, we propose EffectsNet, a deep neural network that predicts\nview-dependent effects. Based on these estimations, we are able to convert\nobserved images to diffuse images. These diffuse images can be projected into\nother views. In the target view, our pipeline reinserts the new view-dependent\neffects. To composite multiple reprojected images to a final output, we learn a\ncomposition network that outputs photo-realistic results. Using this\nimage-guided approach, the network does not have to allocate capacity on\n``remembering'' object appearance, instead it learns how to combine the\nappearance of captured images. We demonstrate the effectiveness of our approach\nboth qualitatively and quantitatively on synthetic as well as on real data."", 'We present a method to learn the 3D surface of objects directly from a\ncollection of images. Previous work achieved this capability by exploiting\nadditional manual annotation, such as object pose, 3D surface templates,\ntemporal continuity of videos, manually selected landmarks, and\nforeground/background masks. In contrast, our method does not make use of any\nsuch annotation. Rather, it builds a generative model, a convolutional neural\nnetwork, which, given a noise vector sample, outputs the 3D surface and texture\nof an object and a background image. These 3 components combined with an\nadditional random viewpoint vector are then fed to a differential renderer to\nproduce a view of the sampled object and background. Our general principle is\nthat if the output of the renderer, the generated image, is realistic, then its\ninput, the generated 3D and texture, should also be realistic. To achieve\nrealism, the generative model is trained adversarially against a discriminator\nthat tries to distinguish between the output of the renderer and real images\nfrom the given data set. Moreover, our generative model can be paired with an\nencoder and trained as an autoencoder, to automatically extract the 3D shape,\ntexture and pose of the object in an image. Our trained generative model and\nencoder show promising results both on real and synthetic data, which\ndemonstrate for the first time that fully unsupervised 3D learning from image\ncollections is possible.', 'Recent improvements to Generative Adversarial Networks (GANs) have made it\npossible to generate realistic images in high resolution based on natural\nlanguage descriptions such as image captions. Furthermore, conditional GANs\nallow us to control the image generation process through labels or even natural\nlanguage descriptions. However, fine-grained control of the image layout, i.e.\nwhere in the image specific objects should be located, is still difficult to\nachieve. This is especially true for images that should contain multiple\ndistinct objects at different spatial locations. We introduce a new approach\nwhich allows us to control the location of arbitrarily many objects within an\nimage by adding an object pathway to both the generator and the discriminator.\nOur approach does not need a detailed semantic layout but only bounding boxes\nand the respective labels of the desired objects are needed. The object pathway\nfocuses solely on the individual objects and is iteratively applied at the\nlocations specified by the bounding boxes. The global pathway focuses on the\nimage background and the general image layout. We perform experiments on the\nMulti-MNIST, CLEVR, and the more complex MS-COCO data set. Our experiments show\nthat through the use of the object pathway we can control object locations\nwithin images and can model complex scenes with multiple objects at various\nlocations. We further show that the object pathway focuses on the individual\nobjects and learns features relevant for these, while the global pathway\nfocuses on global image characteristics and the image background.', 'In recent years, research on image generation methods has been developing\nfast. The auto-encoding variational Bayes method (VAEs) was proposed in 2013,\nwhich uses variational inference to learn a latent space from the image\ndatabase and then generates images using the decoder. The generative\nadversarial networks (GANs) came out as a promising framework, which uses\nadversarial training to improve the generative ability of the generator.\nHowever, the generated pictures by GANs are generally blurry. The deep\nconvolutional generative adversarial networks (DCGANs) were then proposed to\nleverage the quality of generated images. Since the input noise vectors are\nrandomly sampled from a Gaussian distribution, the generator has to map from a\nwhole normal distribution to the images. This makes DCGANs unable to reflect\nthe inherent structure of the training data. In this paper, we propose a novel\ndeep model, called generative adversarial networks with decoder-encoder output\nnoise (DE-GANs), which takes advantage of both the adversarial training and the\nvariational Bayesain inference to improve the performance of image generation.\nDE-GANs use a pre-trained decoder-encoder architecture to map the random\nGaussian noise vectors to informative ones and pass them to the generator of\nthe adversarial networks. Since the decoder-encoder architecture is trained by\nthe same images as the generators, the output vectors could carry the intrinsic\ndistribution information of the original images. Moreover, the loss function of\nDE-GANs is different from GANs and DCGANs. A hidden-space loss function is\nadded to the adversarial loss function to enhance the robustness of the model.\nExtensive empirical results show that DE-GANs can accelerate the convergence of\nthe adversarial training process and improve the quality of the generated\nimages.', 'We give an overview of the different rendering methods and we demonstrate\nthat the use of a Generative Adversarial Networks (GAN) for Global Illumination\n(GI) gives a superior quality rendered image to that of a rasterisations image.\nWe utilise the Pix2Pix architecture and specify the hyper-parameters and\nmethodology used to mimic ray-traced images from a set of input features. We\nalso demonstrate that the GANs quality is comparable to the quality of the\nray-traced images, but is able to produce the image, at a fraction of the time.', 'Generative Adversarial Networks (GAN) have many potential medical imaging\napplications, including data augmentation, domain adaptation, and model\nexplanation. Due to the limited memory of Graphical Processing Units (GPUs),\nmost current 3D GAN models are trained on low-resolution medical images, these\nmodels either cannot scale to high-resolution or are prone to patchy artifacts.\nIn this work, we propose a novel end-to-end GAN architecture that can generate\nhigh-resolution 3D images. We achieve this goal by using different\nconfigurations between training and inference. During training, we adopt a\nhierarchical structure that simultaneously generates a low-resolution version\nof the image and a randomly selected sub-volume of the high-resolution image.\nThe hierarchical design has two advantages: First, the memory demand for\ntraining on high-resolution images is amortized among sub-volumes. Furthermore,\nanchoring the high-resolution sub-volumes to a single low-resolution image\nensures anatomical consistency between sub-volumes. During inference, our model\ncan directly generate full high-resolution images. We also incorporate an\nencoder with a similar hierarchical structure into the model to extract\nfeatures from the images. Experiments on 3D thorax CT and brain MRI demonstrate\nthat our approach outperforms state of the art in image generation. We also\ndemonstrate clinical applications of the proposed model in data augmentation\nand clinical-relevant feature extraction.', 'Then detection and identification of extreme weather events in large-scale\nclimate simulations is an important problem for risk management, informing\ngovernmental policy decisions and advancing our basic understanding of the\nclimate system. Recent work has shown that fully supervised convolutional\nneural networks (CNNs) can yield acceptable accuracy for classifying well-known\ntypes of extreme weather events when large amounts of labeled data are\navailable. However, many different types of spatially localized climate\npatterns are of interest including hurricanes, extra-tropical cyclones, weather\nfronts, and blocking events among others. Existing labeled data for these\npatterns can be incomplete in various ways, such as covering only certain years\nor geographic areas and having false negatives. This type of climate data\ntherefore poses a number of interesting machine learning challenges. We present\na multichannel spatiotemporal CNN architecture for semi-supervised bounding box\nprediction and exploratory data analysis. We demonstrate that our approach is\nable to leverage temporal information and unlabeled data to improve the\nlocalization of extreme weather events. Further, we explore the representations\nlearned by our model in order to better understand this important data. We\npresent a dataset, ExtremeWeather, to encourage machine learning research in\nthis area and to help facilitate further work in understanding and mitigating\nthe effects of climate change. The dataset is available at\nextremeweatherdataset.github.io and the code is available at\nhttps://github.com/eracah/hur-detect.', 'Landscapes are meaningful ecological units that strongly depend on the\nenvironmental conditions. Such dependencies between landscapes and the\nenvironment have been noted since the beginning of Earth sciences and cast into\nconceptual models describing the interdependencies of climate, geology,\nvegetation and geomorphology. Here, we ask whether landscapes, as seen from\nspace, can be statistically predicted from pertinent environmental conditions.\nTo this end we adapted a deep learning generative model in order to establish\nthe relationship between the environmental conditions and the view of\nlandscapes from the Sentinel-2 satellite. We trained a conditional generative\nadversarial network to generate multispectral imagery given a set of climatic,\nterrain and anthropogenic predictors. The generated imagery of the landscapes\nshare many characteristics with the real one. Results based on landscape patch\nmetrics, indicative of landscape composition and structure, show that the\nproposed generative model creates landscapes that are more similar to the\ntargets than the baseline models while overall reflectance and vegetation cover\nare predicted better. We demonstrate that for many purposes the generated\nlandscapes behave as real with immediate application for global change studies.\nWe envision the application of machine learning as a tool to forecast the\neffects of climate change on the spatial features of landscapes, while we\nassess its limitations and breaking points.', 'Generative adversarial networks (GANs) used in domain adaptation tasks have\nthe ability to generate images that are both realistic and personalized,\ntransforming an input image while maintaining its identifiable characteristics.\nHowever, they often require a large quantity of training data to produce\nhigh-quality images in a robust way, which limits their usability in cases when\naccess to data is limited. In our paper, we explore the potential of using\nimages from a simulated 3D environment to improve a domain adaptation task\ncarried out by the MUNIT architecture, aiming to use the resulting images to\nraise awareness of the potential future impacts of climate change.', 'Device-edge co-inference, which partitions a deep neural network between a\nresource-constrained mobile device and an edge server, recently emerges as a\npromising paradigm to support intelligent mobile applications. To accelerate\nthe inference process, on-device model sparsification and intermediate feature\ncompression are regarded as two prominent techniques. However, as the on-device\nmodel sparsity level and intermediate feature compression ratio have direct\nimpacts on computation workload and communication overhead respectively, and\nboth of them affect the inference accuracy, finding the optimal values of these\nhyper-parameters brings a major challenge due to the large search space. In\nthis paper, we endeavor to develop an efficient algorithm to determine these\nhyper-parameters. By selecting a suitable model split point and a pair of\nencoder/decoder for the intermediate feature vector, this problem is casted as\na sequential decision problem, for which, a novel automated machine learning\n(AutoML) framework is proposed based on deep reinforcement learning (DRL).\nExperiment results on an image classification task demonstrate the\neffectiveness of the proposed framework in achieving a better\ncommunication-computation trade-off and significant inference speedup against\nvarious baseline schemes.', 'Segmentation of additive manufacturing (AM) defects in X-ray Computed\nTomography (XCT) images is challenging, due to the poor contrast, small sizes\nand variation in appearance of defects. Automatic segmentation can, however,\nprovide quality control for additive manufacturing. Over recent years,\nthree-dimensional convolutional neural networks (3D CNNs) have performed well\nin the volumetric segmentation of medical images. In this work, we leverage\ntechniques from the medical imaging domain and propose training a 3D U-Net\nmodel to automatically segment defects in XCT images of AM samples. This work\nnot only contributes to the use of machine learning for AM defect detection but\nalso demonstrates for the first time 3D volumetric segmentation in AM. We train\nand test with three variants of the 3D U-Net on an AM dataset, achieving a mean\nintersection of union (IOU) value of 88.4%.', 'As climate change increases the intensity of natural disasters, society needs\nbetter tools for adaptation. Floods, for example, are the most frequent natural\ndisaster, and better tools for flood risk communication could increase the\nsupport for flood-resilient infrastructure development. Our work aims to enable\nmore visual communication of large-scale climate impacts via visualizing the\noutput of coastal flood models as satellite imagery. We propose the first deep\nlearning pipeline to ensure physical-consistency in synthetic visual satellite\nimagery. We advanced a state-of-the-art GAN called pix2pixHD, such that it\nproduces imagery that is physically-consistent with the output of an\nexpert-validated storm surge model (NOAA SLOSH). By evaluating the imagery\nrelative to physics-based flood maps, we find that our proposed framework\noutperforms baseline models in both physical-consistency and photorealism. We\nenvision our work to be the first step towards a global visualization of how\nthe climate challenge will shape our landscape. Continuing on this path, we\nshow that the proposed pipeline generalizes to visualize reforestation. We also\npublish a dataset of over 25k labelled image-triplets to study image-to-image\ntranslation in Earth observation.', 'Development and diffusion of machine learning and big data tools provide a\nnew tool for architects and urban planners that could be used as analytical or\ndesign instruments. The topic investigated in this paper is the application of\nGenerative Adversarial Networks to the design of an urban block. The research\npresents a flexible model able to adapt to the morphological characteristics of\na city. This method does not define explicitly any of the parameters of an\nurban block typical for a city, the algorithm learns them from the existing\nurban context. This approach has been applied to the cities with different\nmorphology: Milan, Amsterdam, Tallinn, Turin, and Bengaluru in order to see the\nperformance of the model and the possibility of style translation between\ndifferent cities. The data are gathered from Openstreetmap and Open Data\nportals of the cities. This research presents the results of the experiments\nand their quantitative and qualitative evaluation.', 'The deep generative adversarial networks (GAN) recently have been shown to be\npromising for different computer vision applications, like image edit- ing,\nsynthesizing high resolution images, generating videos, etc. These networks and\nthe corresponding learning scheme can handle various visual space map- pings.\nWe approach GANs with a novel training method and learning objective, to\ndiscover multiple object instances for three cases: 1) synthesizing a picture\nof a specific object within a cluttered scene; 2) localizing different\ncategories in images for weakly supervised object detection; and 3) improving\nobject discov- ery in object detection pipelines. A crucial advantage of our\nmethod is that it learns a new deep similarity metric, to distinguish multiple\nobjects in one im- age. We demonstrate that the network can act as an\nencoder-decoder generating parts of an image which contain an object, or as a\nmodified deep CNN to rep- resent images for object detection in supervised and\nweakly supervised scheme. Our ranking GAN offers a novel way to search through\nimages for object specific patterns. We have conducted experiments for\ndifferent scenarios and demonstrate the method performance for object\nsynthesizing and weakly supervised object detection and classification using\nthe MS-COCO and PASCAL VOC datasets.', 'One of the consequences of climate change is anobserved increase in the\nfrequency of extreme cli-mate events. That poses a challenge for\nweatherforecast and generation algorithms, which learnfrom historical data but\nshould embed an often un-certain bias to create correct scenarios. This\npaperinvestigates how mapping climate data to a knowndistribution using\nvariational autoencoders mighthelp explore such biases and control the\nsynthesisof weather fields towards more extreme climatescenarios. We\nexperimented using a monsoon-affected precipitation dataset from southwest\nIn-dia, which should give a roughly stable pattern ofrainy days and ease our\ninvestigation. We reportcompelling results showing that mapping complexweather\ndata to a known distribution implementsan efficient control for weather field\nsynthesis to-wards more (or less) extreme scenarios.', ""Projection of changes in extreme indices of climate variables such as\ntemperature and precipitation are critical to assess the potential impacts of\nclimate change on human-made and natural systems, including critical\ninfrastructures and ecosystems. While impact assessment and adaptation planning\nrely on high-resolution projections (typically in the order of a few\nkilometers), state-of-the-art Earth System Models (ESMs) are available at\nspatial resolutions of few hundreds of kilometers. Current solutions to obtain\nhigh-resolution projections of ESMs include downscaling approaches that\nconsider the information at a coarse-scale to make predictions at local scales.\nComplex and non-linear interdependence among local climate variables (e.g.,\ntemperature and precipitation) and large-scale predictors (e.g., pressure\nfields) motivate the use of neural network-based super-resolution\narchitectures. In this work, we present auxiliary variables informed\nspatio-temporal neural architecture for statistical downscaling. The current\nstudy performs daily downscaling of precipitation variable from an ESM output\nat 1.15 degrees (~115 km) to 0.25 degrees (25 km) over the world's most\nclimatically diversified country, India. We showcase significant improvement\ngain against three popular state-of-the-art baselines with a better ability to\npredict extreme events. To facilitate reproducible research, we make available\nall the codes, processed datasets, and trained models in the public domain."", 'The demand for on-device document recognition systems increases in\nconjunction with the emergence of more strict privacy and security\nrequirements. In such systems, there is no data transfer from the end device to\na third-party information processing servers. The response time is vital to the\nuser experience of on-device document recognition. Combined with the\nunavailability of discrete GPUs, powerful CPUs, or a large RAM capacity on\nconsumer-grade end devices such as smartphones, the time limitations put\nsignificant constraints on the computational complexity of the applied\nalgorithms for on-device execution.\n  In this work, we consider document location in an image without prior\nknowledge of the document content or its internal structure. In accordance with\nthe published works, at least 5 systems offer solutions for on-device document\nlocation. All these systems use a location method which can be considered\nHough-based. The precision of such systems seems to be lower than that of the\nstate-of-the-art solutions which were not designed to account for the limited\ncomputational resources.\n  We propose an advanced Hough-based method. In contrast with other approaches,\nit accounts for the geometric invariants of the central projection model and\ncombines both edge and color features for document boundary detection. The\nproposed method allowed for the second best result for SmartDoc dataset in\nterms of precision, surpassed by U-net like neural network. When evaluated on a\nmore challenging MIDV-500 dataset, the proposed algorithm guaranteed the best\nprecision compared to published methods. Our method retained the applicability\nto on-device computations.', 'Understanding 3D object structure from a single image is an important but\ndifficult task in computer vision, mostly due to the lack of 3D object\nannotations in real images. Previous work tackles this problem by either\nsolving an optimization task given 2D keypoint positions, or training on\nsynthetic data with ground truth 3D information. In this work, we propose 3D\nINterpreter Network (3D-INN), an end-to-end framework which sequentially\nestimates 2D keypoint heatmaps and 3D object structure, trained on both real\n2D-annotated images and synthetic 3D data. This is made possible mainly by two\ntechnical innovations. First, we propose a Projection Layer, which projects\nestimated 3D structure to 2D space, so that 3D-INN can be trained to predict 3D\nstructural parameters supervised by 2D annotations on real images. Second,\nheatmaps of keypoints serve as an intermediate representation connecting real\nand synthetic data, enabling 3D-INN to benefit from the variation and abundance\nof synthetic 3D objects, without suffering from the difference between the\nstatistics of real and synthesized images due to imperfect rendering. The\nnetwork achieves state-of-the-art performance on both 2D keypoint estimation\nand 3D structure recovery. We also show that the recovered 3D information can\nbe used in other vision applications, such as 3D rendering and image retrieval.', 'The effects of global climate change on Peruvian glaciers have brought about\nseveral processes of deglaciation during the last few years. The immediate\neffect is the change of size of lakes and rivers. Public institutions that\nmonitor water resources currently have only recent studies which make up less\nthan 10% of the total. The effects of climate change and the lack of updated\ninformation intensify social-economic problems related to water resources in\nPeru. The objective of this research is to develop a software application to\nautomate the Cadastral Registry of Water Bodies in Peru, using techniques of\ndigital image processing, which would provide tools for detection, record,\ntemporal analysis and visualization of water bodies. The images used are from\nthe satellite Landsat5, which undergo a pre-processing of calibration and\ncorrection of the satellite. Detection results are archived into a file that\ncontains location vectors and images of the segmentated bodies of water.', 'Deep learning models are being deployed in many mobile intelligent\napplications. End-side services, such as intelligent personal assistants,\nautonomous cars, and smart home services often employ either simple local\nmodels on the mobile or complex remote models on the cloud. However, recent\nstudies have shown that partitioning the DNN computations between the mobile\nand cloud can increase the latency and energy efficiencies. In this paper, we\npropose an efficient, adaptive, and practical engine, JointDNN, for\ncollaborative computation between a mobile device and cloud for DNNs in both\ninference and training phase. JointDNN not only provides an energy and\nperformance efficient method of querying DNNs for the mobile side but also\nbenefits the cloud server by reducing the amount of its workload and\ncommunications compared to the cloud-only approach. Given the DNN architecture,\nwe investigate the efficiency of processing some layers on the mobile device\nand some layers on the cloud server. We provide optimization formulations at\nlayer granularity for forward- and backward-propagations in DNNs, which can\nadapt to mobile battery limitations and cloud server load constraints and\nquality of service. JointDNN achieves up to 18 and 32 times reductions on the\nlatency and mobile energy consumption of querying DNNs compared to the\nstatus-quo approaches, respectively.', 'The biggest challenge faced by a Machine Learning Engineer is the lack of\ndata they have, especially for 2-dimensional images. The image is processed to\nbe trained into a Machine Learning model so that it can recognize patterns in\nthe data and provide predictions. This research is intended to create a\nsolution using the Cycle Generative Adversarial Networks (GANs) algorithm in\novercoming the problem of lack of data. Then use Style Transfer to be able to\ngenerate a new image based on the given style. Based on the results of testing\nthe resulting model has been carried out several improvements, previously the\nloss value of the photo generator: 3.1267, monet style generator: 3.2026, photo\ndiscriminator: 0.6325, and monet style discriminator: 0.6931 to photo\ngenerator: 2.3792, monet style generator: 2.7291, photo discriminator: 0.5956,\nand monet style discriminator: 0.4940. It is hoped that the research will make\nthe application of this solution useful in the fields of Education, Arts,\nInformation Technology, Medicine, Astronomy, Automotive and other important\nfields.', 'Automatically generating maps from satellite images is an important task.\nThere is a body of literature which tries to address this challenge. We created\na more expansive survey of the task by experimenting with different models and\nadding new loss functions to improve results. We created a database of pairs of\nsatellite images and the corresponding map of the area. Our model translates\nthe satellite image to the corresponding standard layer map image using three\nmain model architectures: (i) a conditional Generative Adversarial Network\n(GAN) which compresses the images down to a learned embedding, (ii) a generator\nwhich is trained as a normalizing flow (RealNVP) model, and (iii) a conditional\nGAN where the generator translates via a series of convolutions to the standard\nlayer of a map and the discriminator input is the concatenation of the\nreal/generated map and the satellite image. Model (iii) was by far the most\npromising of three models. To improve the results we also added a\nreconstruction loss and style transfer loss in addition to the GAN losses. The\nthird model architecture produced the best quality of sampled images. In\ncontrast to the other generative model where evaluation of the model is a\nchallenging problem. since we have access to the real map for a given satellite\nimage, we are able to assign a quantitative metric to the quality of the\ngenerated images in addition to inspecting them visually. While we are\ncontinuing to work on increasing the accuracy of the model, one challenge has\nbeen the coarse resolution of the data which upper-bounds the quality of the\nresults of our model. Nevertheless, as will be seen in the results, the\ngenerated map is more accurate in the features it produces since the generator\narchitecture demands a pixel-wise image translation/pixel-wise coloring. A\nvideo presentation summarizing this paper is available at:\nhttps://youtu.be/Ur0flOX-Ji0', 'This paper introduces a conditional generative adversarial network to\nredesign a street-level image of urban scenes by generating 1) an urban\nintervention policy, 2) an attention map that localises where intervention is\nneeded, 3) a high-resolution street-level image (1024 X 1024 or 1536 X1536)\nafter implementing the intervention. We also introduce a new dataset that\ncomprises aligned street-level images of before and after urban interventions\nfrom real-life scenarios that make this research possible. The introduced\nmethod has been trained on different ranges of urban interventions applied to\nrealistic images. The trained model shows strong performance in re-modelling\ncities, outperforming existing methods that apply image-to-image translation in\nother domains that is computed in a single GPU. This research opens the door\nfor machine intelligence to play a role in re-thinking and re-designing the\ndifferent attributes of cities based on adversarial learning, going beyond the\nmainstream of facial landmarks manipulation or image synthesis from semantic\nsegmentation.', 'There are many applications of Generative Adversarial Networks (GANs) in\nfields like computer vision, natural language processing, speech synthesis, and\nmore. Undoubtedly the most notable results have been in the area of image\nsynthesis and in particular in the generation of deepfake videos. While\ndeepfakes have received much negative media coverage, they can be a useful\ntechnology in applications like entertainment, customer relations, or even\nassistive care. One problem with generating deepfakes is the requirement for a\nlot of image training data of the subject which is not an issue if the subject\nis a celebrity for whom many images already exist. If there are only a small\nnumber of training images then the quality of the deepfake will be poor. Some\nmedia reports have indicated that a good deepfake can be produced with as few\nas 500 images but in practice, quality deepfakes require many thousands of\nimages, one of the reasons why deepfakes of celebrities and politicians have\nbecome so popular. In this study, we exploit the property of a GAN to produce\nimages of an individual with variable facial expressions which we then use to\ngenerate a deepfake. We observe that with such variability in facial\nexpressions of synthetic GAN-generated training images and a reduced quantity\nof them, we can produce a near-realistic deepfake videos.', '""If I provide you a face image of mine (without telling you the actual age\nwhen I took the picture) and a large amount of face images that I crawled\n(containing labeled faces of different ages but not necessarily paired), can\nyou show me what I would look like when I am 80 or what I was like when I was\n5?"" The answer is probably a ""No."" Most existing face aging works attempt to\nlearn the transformation between age groups and thus would require the paired\nsamples as well as the labeled query image. In this paper, we look at the\nproblem from a generative modeling perspective such that no paired samples is\nrequired. In addition, given an unlabeled image, the generative model can\ndirectly produce the image with desired age attribute. We propose a conditional\nadversarial autoencoder (CAAE) that learns a face manifold, traversing on which\nsmooth age progression and regression can be realized simultaneously. In CAAE,\nthe face is first mapped to a latent vector through a convolutional encoder,\nand then the vector is projected to the face manifold conditional on age\nthrough a deconvolutional generator. The latent vector preserves personalized\nface features (i.e., personality) and the age condition controls progression\nvs. regression. Two adversarial networks are imposed on the encoder and\ngenerator, respectively, forcing to generate more photo-realistic faces.\nExperimental results demonstrate the appealing performance and flexibility of\nthe proposed framework by comparing with the state-of-the-art and ground truth.', 'The edge computing paradigm places compute-capable devices - edge servers -\nat the network edge to assist mobile devices in executing data analysis tasks.\nIntuitively, offloading compute-intense tasks to edge servers can reduce their\nexecution time. However, poor conditions of the wireless channel connecting the\nmobile devices to the edge servers may degrade the overall capture-to-output\ndelay achieved by edge offloading. Herein, we focus on edge computing\nsupporting remote object detection by means of Deep Neural Networks (DNNs), and\ndevelop a framework to reduce the amount of data transmitted over the wireless\nlink. The core idea we propose builds on recent approaches splitting DNNs into\nsections - namely head and tail models - executed by the mobile device and edge\nserver, respectively. The wireless link, then, is used to transport the output\nof the last layer of the head model to the edge server, instead of the DNN\ninput. Most prior work focuses on classification tasks and leaves the DNN\nstructure unaltered. Herein, our focus is on DNNs for three different object\ndetection tasks, which present a much more convoluted structure, and modify the\narchitecture of the network to: (i) achieve in-network compression by\nintroducing a bottleneck layer in the early layers on the head model, and (ii)\nprefilter pictures that do not contain objects of interest using a\nconvolutional neural network. Results show that the proposed technique\nrepresents an effective intermediate option between local and edge computing in\na parameter region where these extreme point solutions fail to provide\nsatisfactory performance. The code and trained models are available at\nhttps://github.com/yoshitomo-matsubara/hnd-ghnd-object-detectors .', ""We consider the task of photo-realistic unconditional image generation\n(generate high quality, diverse samples that carry the same visual content as\nthe image) on mobile platforms using Generative Adversarial Networks (GANs). In\nthis paper, we propose a novel approach to trade-off image generation accuracy\nof a GAN for the energy consumed (compute) at run-time called Scale-Energy\nTradeoff GAN (SETGAN). GANs usually take a long time to train and consume a\nhuge memory hence making it difficult to run on edge devices. The key idea\nbehind SETGAN for an image generation task is for a given input image, we train\na GAN on a remote server and use the trained model on edge devices. We use\nSinGAN, a single image unconditional generative model, that contains a pyramid\nof fully convolutional GANs, each responsible for learning the patch\ndistribution at a different scale of the image. During the training process, we\ndetermine the optimal number of scales for a given input image and the energy\nconstraint from the target edge device. Results show that with SETGAN's unique\nclient-server-based architecture, we were able to achieve a 56% gain in energy\nfor a loss of 3% to 12% SSIM accuracy. Also, with the parallel multi-scale\ntraining, we obtain around 4x gain in training time on the server."", 'We present EgoRenderer, a system for rendering full-body neural avatars of a\nperson captured by a wearable, egocentric fisheye camera that is mounted on a\ncap or a VR headset. Our system renders photorealistic novel views of the actor\nand her motion from arbitrary virtual camera locations. Rendering full-body\navatars from such egocentric images come with unique challenges due to the\ntop-down view and large distortions. We tackle these challenges by decomposing\nthe rendering process into several steps, including texture synthesis, pose\nconstruction, and neural image translation. For texture synthesis, we propose\nEgo-DPNet, a neural network that infers dense correspondences between the input\nfisheye images and an underlying parametric body model, and to extract textures\nfrom egocentric inputs. In addition, to encode dynamic appearances, our\napproach also learns an implicit texture stack that captures detailed\nappearance variation across poses and viewpoints. For correct pose generation,\nwe first estimate body pose from the egocentric view using a parametric model.\nWe then synthesize an external free-viewpoint pose image by projecting the\nparametric model to the user-specified target viewpoint. We next combine the\ntarget pose image and the textures into a combined feature image, which is\ntransformed into the output color image using a neural image translation\nnetwork. Experimental evaluations show that EgoRenderer is capable of\ngenerating realistic free-viewpoint avatars of a person wearing an egocentric\ncamera. Comparisons to several baselines demonstrate the advantages of our\napproach.', 'We introduce a novel framework to build a model that can learn how to segment\nobjects from a collection of images without any human annotation. Our method\nbuilds on the observation that the location of object segments can be perturbed\nlocally relative to a given background without affecting the realism of a\nscene. Our approach is to first train a generative model of a layered scene.\nThe layered representation consists of a background image, a foreground image\nand the mask of the foreground. A composite image is then obtained by\noverlaying the masked foreground image onto the background. The generative\nmodel is trained in an adversarial fashion against a discriminator, which\nforces the generative model to produce realistic composite images. To force the\ngenerator to learn a representation where the foreground layer corresponds to\nan object, we perturb the output of the generative model by introducing a\nrandom shift of both the foreground image and mask relative to the background.\nBecause the generator is unaware of the shift before computing its output, it\nmust produce layered representations that are realistic for any such random\nperturbation. Finally, we learn to segment an image by defining an autoencoder\nconsisting of an encoder, which we train, and the pre-trained generator as the\ndecoder, which we freeze. The encoder maps an image to a feature vector, which\nis fed as input to the generator to give a composite image matching the\noriginal input image. Because the generator outputs an explicit layered\nrepresentation of the scene, the encoder learns to detect and segment objects.\nWe demonstrate this framework on real images of several object categories.', 'Recent approaches in literature have exploited the multi-modal information in\ndocuments (text, layout, image) to serve specific downstream document tasks.\nHowever, they are limited by their - (i) inability to learn cross-modal\nrepresentations across text, layout and image dimensions for documents and (ii)\ninability to process multi-page documents. Pre-training techniques have been\nshown in Natural Language Processing (NLP) domain to learn generic textual\nrepresentations from large unlabelled datasets, applicable to various\ndownstream NLP tasks. In this paper, we propose a multi-task learning-based\nframework that utilizes a combination of self-supervised and supervised\npre-training tasks to learn a generic document representation applicable to\nvarious downstream document tasks. Specifically, we introduce Document Topic\nModelling and Document Shuffle Prediction as novel pre-training tasks to learn\nrich image representations along with the text and layout representations for\ndocuments. We utilize the Longformer network architecture as the backbone to\nencode the multi-modal information from multi-page documents in an end-to-end\nfashion. We showcase the applicability of our pre-training framework on a\nvariety of different real-world document tasks such as document classification,\ndocument information extraction, and document retrieval. We evaluate our\nframework on different standard document datasets and conduct exhaustive\nexperiments to compare performance against various ablations of our framework\nand state-of-the-art baselines.', 'High user interaction capability of mobile devices can help improve the\naccuracy of mobile visual search systems. At query time, it is possible to\ncapture multiple views of an object from different viewing angles and at\ndifferent scales with the mobile device camera to obtain richer information\nabout the object compared to a single view and hence return more accurate\nresults. Motivated by this, we developed a mobile multi-view object image\nsearch system, using a client-server architecture. Multi-view images of objects\nacquired by the mobile clients are processed and local features are sent to the\nserver, which combines the query image representations with early/late fusion\nmethods based on bag-of-visual-words and sends back the query results. We\nperformed a comprehensive analysis of early and late fusion approaches using\nvarious similarity functions, on an existing single view and a new multi-view\nobject image database. The experimental results show that multi-view search\nprovides significantly better retrieval accuracy compared to single view\nsearch.', 'Reconstructing the 3D geometry of an object from an image is a major\nchallenge in computer vision. Recently introduced differentiable renderers can\nbe leveraged to learn the 3D geometry of objects from 2D images, but those\napproaches require additional supervision to enable the renderer to produce an\noutput that can be compared to the input image. This can be scene information\nor constraints such as object silhouettes, uniform backgrounds, material,\ntexture, and lighting. In this paper, we propose an approach that enables a\ndifferentiable rendering-based learning of 3D objects from images with\nbackgrounds without the need for silhouette supervision. Instead of trying to\nrender an image close to the input, we propose an adversarial style-transfer\nand domain adaptation pipeline that allows to translate the input image domain\nto the rendered image domain. This allows us to directly compare between a\ntranslated image and the differentiable rendering of a 3D object reconstruction\nin order to train the 3D object reconstruction network. We show that the\napproach learns 3D geometry from images with backgrounds and provides a better\nperformance than constrained methods for single-view 3D object reconstruction\non this task.', ""Climate change has become one of the biggest global problems increasingly\ncompromising the Earth's habitability. Recent developments such as the\nextraordinary heat waves in California & Canada, and the devastating floods in\nGermany point to the role of climate change in the ever-increasing frequency of\nextreme weather. Numerical modelling of the weather and climate have seen\ntremendous improvements in the last five decades, yet stringent limitations\nremain to be overcome. Spatially and temporally localized forecasting is the\nneed of the hour for effective adaptation measures towards minimizing the loss\nof life and property. Artificial Intelligence-based methods are demonstrating\npromising results in improving predictions, but are still limited by the\navailability of requisite hardware and software required to process the vast\ndeluge of data at a scale of the planet Earth. Quantum computing is an emerging\nparadigm that has found potential applicability in several fields. In this\nopinion piece, we argue that new developments in Artificial Intelligence\nalgorithms designed for quantum computers - also known as Quantum Artificial\nIntelligence (QAI) - may provide the key breakthroughs necessary to furthering\nthe science of climate change. The resultant improvements in weather and\nclimate forecasts are expected to cascade to numerous societal benefits."", ""Climate change is expected to reshuffle the settlement landscape: forcing\npeople in affected areas to migrate, to change their lifeways, and continuing\nto affect demographic change throughout the world. Changes to the geographic\ndistribution of population will have dramatic impacts on land use and land\ncover and thus constitute one of the major challenges of planning for climate\nchange scenarios. In this paper, we explore a generative model framework for\ngenerating satellite imagery conditional on gridded population distributions.\nWe make additions to the existing ALAE architecture, creating a spatially\nconditional version: SCALAE. This method allows us to explicitly disentangle\npopulation from the model's latent space and thus input custom population\nforecasts into the generated imagery. We postulate that such imagery could then\nbe directly used for land cover and land use change estimation using existing\nframeworks, as well as for realistic visualisation of expected local change. We\nevaluate the model by comparing pixel and semantic reconstructions, as well as\ncalculate the standard FID metric. The results suggest the model captures\npopulation distributions accurately and delivers a controllable method to\ngenerate realistic satellite imagery."", 'Astronomy of the 21st century increasingly finds itself with extreme\nquantities of data. This growth in data is ripe for modern technologies such as\ndeep image processing, which has the potential to allow astronomers to\nautomatically identify, classify, segment and deblend various astronomical\nobjects. In this paper, we explore the use of chained generative adversarial\nnetworks (GANs), a class of generative models that learn mappings from latent\nspaces to data distributions by modelling the joint distribution of the data,\nto produce physically realistic galaxy images as one use case of such models.\nIn cosmology, such datasets can aid in the calibration of shape measurements\nfor weak lensing by augmenting data with synthetic images. By measuring the\ndistributions of multiple physical properties, we show that images generated\nwith our approach closely follow the distributions of real galaxies, further\nestablishing state-of-the-art GAN architectures as a valuable tool for\nmodern-day astronomy.', ""The difficulty in obtaining labeled data relevant to a given task is among\nthe most common and well-known practical obstacles to applying deep learning\ntechniques to new or even slightly modified domains. The data volumes required\nby the current generation of supervised learning algorithms typically far\nexceed what a human needs to learn and complete a given task. We investigate\nways to expand a given labeled corpus of remote sensed imagery into a larger\ncorpus using Generative Adversarial Networks (GANs). We then measure how these\nadditional synthetic data affect supervised machine learning performance on an\nobject detection task.\n  Our data driven strategy is to train GANs to (1) generate synthetic\nsegmentation masks and (2) generate plausible synthetic remote sensing imagery\ncorresponding to these segmentation masks. Run sequentially, these GANs allow\nthe generation of synthetic remote sensing imagery complete with segmentation\nlabels. We apply this strategy to the data set from ISPRS' 2D Semantic Labeling\nContest - Potsdam, with a follow on vehicle detection task. We find that in\nscenarios with limited training data, augmenting the available data with such\nsynthetically generated data can improve detector performance."", 'In medical imaging, a general problem is that it is costly and time consuming\nto collect high quality data from healthy and diseased subjects. Generative\nadversarial networks (GANs) is a deep learning method that has been developed\nfor synthesizing data. GANs can thereby be used to generate more realistic\ntraining data, to improve classification performance of machine learning\nalgorithms. Another application of GANs is image-to-image translations, e.g.\ngenerating magnetic resonance (MR) images from computed tomography (CT) images,\nwhich can be used to obtain multimodal datasets from a single modality. Here,\nwe evaluate two unsupervised GAN models (CycleGAN and UNIT) for image-to-image\ntranslation of T1- and T2-weighted MR images, by comparing generated synthetic\nMR images to ground truth images. We also evaluate two supervised models; a\nmodification of CycleGAN and a pure generator model. A small perceptual study\nwas also performed to evaluate how visually realistic the synthesized images\nare. It is shown that the implemented GAN models can synthesize visually\nrealistic MR images (incorrectly labeled as real by a human). It is also shown\nthat models producing more visually realistic synthetic images not necessarily\nhave better quantitative error measurements, when compared to ground truth\ndata. Code is available at https://github.com/simontomaskarlsson/GAN-MRI', ""Climate change has been a common interest and the forefront of crucial\npolitical discussion and decision-making for many years. Shallow clouds play a\nsignificant role in understanding the Earth's climate, but they are challenging\nto interpret and represent in a climate model. By classifying these cloud\nstructures, there is a better possibility of understanding the physical\nstructures of the clouds, which would improve the climate model generation,\nresulting in a better prediction of climate change or forecasting weather\nupdate. Clouds organise in many forms, which makes it challenging to build\ntraditional rule-based algorithms to separate cloud features. In this paper,\nclassification of cloud organization patterns was performed using a new\nscaled-up version of Convolutional Neural Network (CNN) named as EfficientNet\nas the encoder and UNet as decoder where they worked as feature extractor and\nreconstructor of fine grained feature map and was used as a classifier, which\nwill help experts to understand how clouds will shape the future climate. By\nusing a segmentation model in a classification task, it was shown that with a\ngood encoder alongside UNet, it is possible to obtain good performance from\nthis dataset. Dice coefficient has been used for the final evaluation metric,\nwhich gave the score of 66.26\\% and 66.02\\% for public and private (test set)\nleaderboard on Kaggle competition respectively."", ""Faces and their expressions are one of the potent subjects for digital\nimages. Detecting emotions from images is an ancient task in the field of\ncomputer vision; however, performing its reverse -- synthesizing facial\nexpressions from images -- is quite new. Such operations of regenerating images\nwith different facial expressions, or altering an existing expression in an\nimage require the Generative Adversarial Network (GAN). In this paper, we aim\nto change the facial expression in an image using GAN, where the input image\nwith an initial expression (i.e., happy) is altered to a different expression\n(i.e., disgusted) for the same person. We used StarGAN techniques on a modified\nversion of the MUG dataset to accomplish this objective. Moreover, we extended\nour work further by remodeling facial expressions in an image indicated by the\nemotion from a given text. As a result, we applied a Long Short-Term Memory\n(LSTM) method to extract emotion from the text and forwarded it to our\nexpression-altering module. As a demonstration of our working pipeline, we also\ncreate an application prototype of a blog that regenerates the profile picture\nwith different expressions based on the user's textual emotion."", 'We introduce CLIMATE-FEVER, a new publicly available dataset for verification\nof climate change-related claims. By providing a dataset for the research\ncommunity, we aim to facilitate and encourage work on improving algorithms for\nretrieving evidential support for climate-specific claims, addressing the\nunderlying language understanding challenges, and ultimately help alleviate the\nimpact of misinformation on climate change. We adapt the methodology of FEVER\n[1], the largest dataset of artificially designed claims, to real-life claims\ncollected from the Internet. While during this process, we could rely on the\nexpertise of renowned climate scientists, it turned out to be no easy task. We\ndiscuss the surprising, subtle complexity of modeling real-world\nclimate-related claims within the \\textsc{fever} framework, which we believe\nprovides a valuable challenge for general natural language understanding. We\nhope that our work will mark the beginning of a new exciting long-term joint\neffort by the climate science and AI community.', 'Cloud computing offers the potential to help scientists to process massive\nnumber of computing resources often required in machine learning application\nsuch as computer vision problems. This proposal would like to show that which\nbenefits can be obtained from cloud in order to help medical image analysis\nusers (including scientists, clinicians, and research institutes). As security\nand privacy of algorithms are important for most of algorithms inventors, these\nalgorithms can be hidden in a cloud to allow the users to use the algorithms as\na package without any access to see/change their inside. In another word, in\nthe user part, users send their images to the cloud and configure the algorithm\nvia an interface. In the cloud part, the algorithms are applied to this image\nand the results are returned back to the user. My proposal has two parts: (1)\ninvestigate the potential of cloud computing for computer vision problems and\n(2) study the components of a proposed cloud-based framework for medical image\nanalysis application and develop them (depending on the length of the\ninternship). The investigation part will involve a study on several aspects of\nthe problem including security, usability (for medical end users of the\nservice), appropriate programming abstractions for vision problems, scalability\nand resource requirements. In the second part of this proposal I am going to\nthoroughly study of the proposed framework components and their relations and\ndevelop them. The proposed cloud-based framework includes an integrated\nenvironment to enable scientists and clinicians to access to the previous and\ncurrent medical image analysis algorithms using a handful user interface\nwithout any access to the algorithm codes and procedures.', 'Climate change communication in the mass media and other textual sources may\naffect and shape public perception. Extracting climate change information from\nthese sources is an important task, e.g., for filtering content and\ne-discovery, sentiment analysis, automatic summarization, question-answering,\nand fact-checking. However, automating this process is a challenge, as climate\nchange is a complex, fast-moving, and often ambiguous topic with scarce\nresources for popular text-based AI tasks. In this paper, we introduce\n\\textsc{ClimaText}, a dataset for sentence-based climate change topic\ndetection, which we make publicly available. We explore different approaches to\nidentify the climate change topic in various text sources. We find that popular\nkeyword-based models are not adequate for such a complex and evolving task.\nContext-based algorithms like BERT \\cite{devlin2018bert} can detect, in\naddition to many trivial cases, a variety of complex and implicit topic\npatterns. Nevertheless, our analysis reveals a great potential for improvement\nin several directions, such as, e.g., capturing the discussion on indirect\neffects of climate change. Hence, we hope this work can serve as a good\nstarting point for further research on this topic.', ""In this technical report I present my method for automatic synthetic dataset\ngeneration for object detection and demonstrate it on the video game League of\nLegends. This report furthermore serves as a handbook on how to automatically\ngenerate datasets and as an introduction on the dataset generation part of the\nLeagueAI framework. The LeagueAI framework is a software framework that\nprovides detailed information about the game League of Legends based on the\nsame input a human player would have, namely vision. The framework allows\nresearchers and enthusiasts to develop their own intelligent agents or to\nextract detailed information about the state of the game. A big problem of\nmachine vision applications usually is the laborious work of gathering large\namounts of hand labeled data. Thus, a crucial part of the vision pipeline of\nthe LeagueAI framework, the dataset generation, is presented in this report.\nThe method involves extracting image raw data from the game's 3D models and\ncombining them with the game background to create game-like synthetic images\nand to generate the corresponding labels automatically. In an experiment I\ncompared a model trained on synthetic data to a model trained on hand labeled\ndata and a model trained on a combined dataset. The model trained on the\nsynthetic data showed higher detection precision on more classes and more\nreliable tracking performance of the player character. The model trained on the\ncombined dataset did not perform better because of the different formats of the\nolder hand labeled dataset and the synthetic data."", 'The ever-growing popularity and rapid improving of artificial intelligence\n(AI) have raised rethinking on the evolution of wireless networks. Mobile edge\ncomputing (MEC) provides a natural platform for AI applications since it is\nwith rich computation resources to train machine learning (ML) models, as well\nas low-latency access to the data generated by mobile and internet of things\n(IoT) devices. In this paper, we present an infrastructure to perform ML tasks\nat an MEC server with the assistance of a reconfigurable intelligent surface\n(RIS). In contrast to conventional communication systems where the principal\ncriterions are to maximize the throughput, we aim at maximizing the learning\nperformance. Specifically, we minimize the maximum learning error of all\nparticipating users by jointly optimizing transmit power of mobile users,\nbeamforming vectors of the base station (BS), and the phase-shift matrix of the\nRIS. An alternating optimization (AO)-based framework is proposed to optimize\nthe three terms iteratively, where a successive convex approximation\n(SCA)-based algorithm is developed to solve the power allocation problem,\nclosed-form expressions of the beamforming vectors are derived, and an\nalternating direction method of multipliers (ADMM)-based algorithm is designed\ntogether with an error level searching (ELS) framework to effectively solve the\nchallenging nonconvex optimization problem of the phase-shift matrix.\nSimulation results demonstrate significant gains of deploying an RIS and\nvalidate the advantages of our proposed algorithms over various benchmarks.\nLastly, a unified communication-training-inference platform is developed based\non the CARLA platform and the SECOND network, and a use case (3D object\ndetection in autonomous driving) for the proposed scheme is demonstrated on the\ndeveloped platform.', 'In the big data era, the impetus to digitize the vast reservoirs of data\ntrapped in unstructured scanned documents such as invoices, bank documents and\ncourier receipts has gained fresh momentum. The scanning process often results\nin the introduction of artifacts such as background noise, blur due to camera\nmotion, watermarkings, coffee stains, or faded text. These artifacts pose many\nreadability challenges to current text recognition algorithms and significantly\ndegrade their performance. Existing learning based denoising techniques require\na dataset comprising of noisy documents paired with cleaned versions. In such\nscenarios, a model can be trained to generate clean documents from noisy\nversions. However, very often in the real world such a paired dataset is not\navailable, and all we have for training our denoising model are unpaired sets\nof noisy and clean images. This paper explores the use of GANs to generate\ndenoised versions of the noisy documents. In particular, where paired\ninformation is available, we formulate the problem as an image-to-image\ntranslation task i.e, translating a document from noisy domain ( i.e.,\nbackground noise, blurred, faded, watermarked ) to a target clean document\nusing Generative Adversarial Networks (GAN). However, in the absence of paired\nimages for training, we employed CycleGAN which is known to learn a mapping\nbetween the distributions of the noisy images to the denoised images using\nunpaired data to achieve image-to-image translation for cleaning the noisy\ndocuments. We compare the performance of CycleGAN for document cleaning tasks\nusing unpaired images with a Conditional GAN trained on paired data from the\nsame dataset. Experiments were performed on a public document dataset on which\ndifferent types of noise were artificially induced, results demonstrate that\nCycleGAN learns a more robust mapping from the space of noisy to clean\ndocuments.', ""Low-quality face image restoration is a popular research direction in today's\ncomputer vision field. It can be used as a pre-work for tasks such as face\ndetection and face recognition. At present, there is a lot of work to solve the\nproblem of low-quality faces under various environmental conditions. This paper\nmainly focuses on the restoration of motion-blurred faces. In increasingly\nabundant mobile scenes, the fast recovery of motion-blurred faces can bring\nhighly effective speed improvements in tasks such as face matching. In order to\nachieve this goal, a deblurring method for motion-blurred facial image signals\nbased on generative adversarial networks(GANs) is proposed. It uses an\nend-to-end method to train a sharp image generator, i.e., a processor for\nmotion-blurred facial images. This paper introduce the processing progress of\nmotion-blurred images, the development and changes of GANs and some basic\nconcepts. After that, it give the details of network structure and training\noptimization design of the image processor. Then we conducted a motion blur\nimage generation experiment on some general facial data set, and used the pairs\nof blurred and sharp face image data to perform the training and testing\nexperiments of the processor GAN, and gave some visual displays. Finally, MTCNN\nis used to detect the faces of the image generated by the deblurring processor,\nand compare it with the result of the blurred image. From the results, the\nprocessing effect of the deblurring processor on the motion-blurred picture has\na significant improvement both in terms of intuition and evaluation indicators\nof face detection."", 'A growing number of commercially available mobile phones come with integrated\nhigh-resolution digital cameras. That enables a new class of dedicated\napplications to image analysis such as mobile visual search, image cropping,\nobject detection, content-based image retrieval, image classification. In this\npaper, a new mobile application for image content retrieval and classification\nfor mobile device display is proposed to enrich the visual experience of users.\nThe mobile application can extract a certain number of images based on the\ncontent of an image with visual saliency methods aiming at detecting the most\ncritical regions in a given image from a perceptual viewpoint. First, the most\ncritical areas from a perceptual perspective are extracted using the local\nmaxima of a 2D saliency function. Next, a salient region is cropped using the\nbounding box centred on the local maxima of the thresholded Saliency Map of the\nimage. Then, each image crop feds into an Image Classification system based on\nSVM and SIFT descriptors to detect the class of object present in the image.\nImageNet repository was used as the reference for semantic category\nclassification. Android platform was used to implement the mobile application\non a client-server architecture. A mobile client sends the photo taken by the\ncamera to the server, which processes the image and returns the results (image\ncontents such as image crops and related target classes) to the mobile client.\nThe application was run on thousands of pictures and showed encouraging results\ntowards a better user visual experience with mobile displays.', ""We propose a learning algorithm to design a light-weight neural multiplexer\nthat given the input and computational resource requirements, calls the model\nthat will consume the minimum compute resources for a successful inference.\nMobile devices can use the proposed algorithm to offload the hard inputs to the\ncloud while inferring the easy ones locally. Besides, in the large scale\ncloud-based intelligent applications, instead of replicating the most-accurate\nmodel, a range of small and large models can be multiplexed from depending on\nthe input's complexity which will save the cloud's computational resources. The\ninput complexity or hardness is determined by the number of models that can\npredict the correct label. For example, if no model can predict the label\ncorrectly, then the input is considered as the hardest. The proposed algorithm\nallows the mobile device to detect the inputs that can be processed locally and\nthe ones that require a larger model and should be sent a cloud server.\nTherefore, the mobile user benefits from not only the local processing but also\nfrom an accurate model hosted on a cloud server. Our experimental results show\nthat the proposed algorithm improves mobile's model accuracy by 8.52% which is\nbecause of those inputs that are properly selected and offloaded to the cloud\nserver. In addition, it saves the cloud providers' compute resources by a\nfactor of 2.85x as small models are chosen for easier inputs."", 'Despite significant progress on current state-of-the-art image generation\nmodels, synthesis of document images containing multiple and complex object\nlayouts is a challenging task. This paper presents a novel approach, called\nDocSynth, to automatically synthesize document images based on a given layout.\nIn this work, given a spatial layout (bounding boxes with object categories) as\na reference by the user, our proposed DocSynth model learns to generate a set\nof realistic document images consistent with the defined layout. Also, this\nframework has been adapted to this work as a superior baseline model for\ncreating synthetic document image datasets for augmenting real data during\ntraining for document layout analysis tasks. Different sets of learning\nobjectives have been also used to improve the model performance.\nQuantitatively, we also compare the generated results of our model with real\ndata using standard evaluation metrics. The results highlight that our model\ncan successfully generate realistic and diverse document images with multiple\nobjects. We also present a comprehensive qualitative analysis summary of the\ndifferent scopes of synthetic image generation tasks. Lastly, to our knowledge\nthis is the first work of its kind.', 'Documents often exhibit various forms of degradation, which make it hard to\nbe read and substantially deteriorate the performance of an OCR system. In this\npaper, we propose an effective end-to-end framework named Document Enhancement\nGenerative Adversarial Networks (DE-GAN) that uses the conditional GANs (cGANs)\nto restore severely degraded document images. To the best of our knowledge,\nthis practice has not been studied within the context of generative adversarial\ndeep networks. We demonstrate that, in different tasks (document clean up,\nbinarization, deblurring and watermark removal), DE-GAN can produce an enhanced\nversion of the degraded document with a high quality. In addition, our approach\nprovides consistent improvements compared to state-of-the-art methods over the\nwidely used DIBCO 2013, DIBCO 2017 and H-DIBCO 2018 datasets, proving its\nability to restore a degraded document image to its ideal condition. The\nobtained results on a wide variety of degradation reveal the flexibility of the\nproposed model to be exploited in other document enhancement problems.', ""Climate change is a burning issue of our time, with the Sustainable\nDevelopment Goal (SDG) 13 of the United Nations demanding global climate\naction. Realizing the urgency, in 2015 in Paris, world leaders signed an\nagreement committing to taking voluntary action to reduce carbon emissions.\nHowever, the scale, magnitude, and climate action processes vary globally,\nespecially between developed and developing countries. Therefore, from\nparliament to social media, the debates and discussions on climate change\ngather data from wide-ranging sources essential to the policy design and\nimplementation. The downside is that we do not currently have the mechanisms to\npool the worldwide dispersed knowledge emerging from the structured and\nunstructured data sources.\n  The paper thematically discusses how NLP techniques could be employed in\nclimate policy research and contribute to society's good at large. In\nparticular, we exemplify symbiosis of NLP and Climate Policy Research via four\nmethodologies. The first one deals with the major topics related to climate\npolicy using automated content analysis. We investigate the opinions\n(sentiments) of major actors' narratives towards climate policy in the second\nmethodology. The third technique explores the climate actors' beliefs towards\npro or anti-climate orientation. Finally, we discuss developing a Climate\nKnowledge Graph.\n  The present theme paper further argues that creating a knowledge platform\nwould help in the formulation of a holistic climate policy and effective\nclimate action. Such a knowledge platform would integrate the policy actors'\nvaried opinions from different social sectors like government, business, civil\nsociety, and the scientific community. The research outcome will add value to\neffective climate action because policymakers can make informed decisions by\nlooking at the diverse public opinion on a comprehensive platform."", 'Machine learning has the potential to aid in mitigating the human effects of\nclimate change. Previous applications of machine learning to tackle the human\neffects in climate change include approaches like informing individuals of\ntheir carbon footprint and strategies to reduce it. For these methods to be the\nmost effective they must consider relevant social-psychological factors for\neach individual. Of social-psychological factors at play in climate change,\naffect has been previously identified as a key element in perceptions and\nwillingness to engage in mitigative behaviours. In this work, we propose an\ninvestigation into how affect could be incorporated to enhance machine learning\nbased interventions for climate change. We propose using affective agent-based\nmodelling for climate change as well as the use of a simulated climate change\nsocial dilemma to explore the potential benefits of affective machine learning\ninterventions. Behavioural and informational interventions can be a powerful\ntool in helping humans adopt mitigative behaviours. We expect that utilizing\naffective ML can make interventions an even more powerful tool and help\nmitigative behaviours become widely adopted.', ""With recent progress in graphics, it has become more tractable to train\nmodels on synthetic images, potentially avoiding the need for expensive\nannotations. However, learning from synthetic images may not achieve the\ndesired performance due to a gap between synthetic and real image\ndistributions. To reduce this gap, we propose Simulated+Unsupervised (S+U)\nlearning, where the task is to learn a model to improve the realism of a\nsimulator's output using unlabeled real data, while preserving the annotation\ninformation from the simulator. We develop a method for S+U learning that uses\nan adversarial network similar to Generative Adversarial Networks (GANs), but\nwith synthetic images as inputs instead of random vectors. We make several key\nmodifications to the standard GAN algorithm to preserve annotations, avoid\nartifacts, and stabilize training: (i) a 'self-regularization' term, (ii) a\nlocal adversarial loss, and (iii) updating the discriminator using a history of\nrefined images. We show that this enables generation of highly realistic\nimages, which we demonstrate both qualitatively and with a user study. We\nquantitatively evaluate the generated images by training models for gaze\nestimation and hand pose estimation. We show a significant improvement over\nusing synthetic images, and achieve state-of-the-art results on the MPIIGaze\ndataset without any labeled real data."", 'Recent research has made great progress in realizing neural style transfer of\nimages, which denotes transforming an image to a desired style. Many users\nstart to use their mobile phones to record their daily life, and then edit and\nshare the captured images and videos with other users. However, directly\napplying existing style transfer approaches on videos, i.e., transferring the\nstyle of a video frame by frame, requires an extremely large amount of\ncomputation resources. It is still technically unaffordable to perform style\ntransfer of videos on mobile phones. To address this challenge, we propose\nMVStylizer, an efficient edge-assisted photorealistic video style transfer\nsystem for mobile phones. Instead of performing stylization frame by frame,\nonly key frames in the original video are processed by a pre-trained deep\nneural network (DNN) on edge servers, while the rest of stylized intermediate\nframes are generated by our designed optical-flow-based frame interpolation\nalgorithm on mobile phones. A meta-smoothing module is also proposed to\nsimultaneously upscale a stylized frame to arbitrary resolution and remove\nstyle transfer related distortions in these upscaled frames. In addition, for\nthe sake of continuously enhancing the performance of the DNN model on the edge\nserver, we adopt a federated learning scheme to keep retraining each DNN model\non the edge server with collected data from mobile clients and syncing with a\nglobal DNN model on the cloud server. Such a scheme effectively leverages the\ndiversity of collected data from various mobile clients and efficiently\nimproves the system performance. Our experiments demonstrate that MVStylizer\ncan generate stylized videos with an even better visual quality compared to the\nstate-of-the-art method while achieving 75.5$\\times$ speedup for\n1920$\\times$1080 videos.', 'Object segmentation is a crucial problem that is usually solved by using\nsupervised learning approaches over very large datasets composed of both images\nand corresponding object masks. Since the masks have to be provided at pixel\nlevel, building such a dataset for any new domain can be very time-consuming.\nWe present ReDO, a new model able to extract objects from images without any\nannotation in an unsupervised way. It relies on the idea that it should be\npossible to change the textures or colors of the objects without changing the\noverall distribution of the dataset. Following this assumption, our approach is\nbased on an adversarial architecture where the generator is guided by an input\nsample: given an image, it extracts the object mask, then redraws a new object\nat the same location. The generator is controlled by a discriminator that\nensures that the distribution of generated images is aligned to the original\none. We experiment with this method on different datasets and demonstrate the\ngood quality of extracted masks.', 'The ability of the Generative Adversarial Networks (GANs) framework to learn\ngenerative models mapping from simple latent distributions to arbitrarily\ncomplex data distributions has been demonstrated empirically, with compelling\nresults showing that the latent space of such generators captures semantic\nvariation in the data distribution. Intuitively, models trained to predict\nthese semantic latent representations given data may serve as useful feature\nrepresentations for auxiliary problems where semantics are relevant. However,\nin their existing form, GANs have no means of learning the inverse mapping --\nprojecting data back into the latent space. We propose Bidirectional Generative\nAdversarial Networks (BiGANs) as a means of learning this inverse mapping, and\ndemonstrate that the resulting learned feature representation is useful for\nauxiliary supervised discrimination tasks, competitive with contemporary\napproaches to unsupervised and self-supervised feature learning.', 'Unconditional image generation has recently been dominated by generative\nadversarial networks (GANs). GAN methods train a generator which regresses\nimages from random noise vectors, as well as a discriminator that attempts to\ndifferentiate between the generated images and a training set of real images.\nGANs have shown amazing results at generating realistic looking images. Despite\ntheir success, GANs suffer from critical drawbacks including: unstable training\nand mode-dropping. The weaknesses in GANs have motivated research into\nalternatives including: variational auto-encoders (VAEs), latent embedding\nlearning methods (e.g. GLO) and nearest-neighbor based implicit maximum\nlikelihood estimation (IMLE). Unfortunately at the moment, GANs still\nsignificantly outperform the alternative methods for image generation. In this\nwork, we present a novel method - Generative Latent Nearest Neighbors (GLANN) -\nfor training generative models without adversarial training. GLANN combines the\nstrengths of IMLE and GLO in a way that overcomes the main drawbacks of each\nmethod. Consequently, GLANN generates images that are far better than GLO and\nIMLE. Our method does not suffer from mode collapse which plagues GAN training\nand is much more stable. Qualitative results show that GLANN outperforms a\nbaseline consisting of 800 GANs and VAEs on commonly used datasets. Our models\nare also shown to be effective for training truly non-adversarial unsupervised\nimage translation.', 'Training real-world neural network models to achieve high performance and\ngeneralizability typically requires a substantial amount of labeled data,\nspanning a broad range of variation. This data-labeling process can be both\nlabor and cost intensive. To achieve desirable predictive performance, a\ntrained model is typically applied into a domain where the data distribution is\nsimilar to the training dataset. However, for many agricultural machine\nlearning problems, training datasets are collected at a specific location,\nduring a specific period in time of the growing season. Since agricultural\nsystems exhibit substantial variability in terms of crop type, cultivar,\nmanagement, seasonal growth dynamics, lighting condition, sensor type, etc, a\nmodel trained from one dataset often does not generalize well across domains.\nTo enable more data efficient and generalizable neural network models in\nagriculture, we propose a method that generates photorealistic agricultural\nimages from a synthetic 3D crop model domain into real world crop domains. The\nmethod uses a semantically constrained GAN (generative adversarial network) to\npreserve the fruit position and geometry. We observe that a baseline CycleGAN\nmethod generates visually realistic target domain images but does not preserve\nfruit position information while our method maintains fruit positions well.\nImage generation results in vineyard grape day and night images show the visual\noutputs of our network are much better compared to a baseline network.\nIncremental training experiments in vineyard grape detection tasks show that\nthe images generated from our method can significantly speed the domain\nadaption process, increase performance for a given number of labeled images\n(i.e. data efficiency), and decrease labeling requirements.', 'We present a new method to create spatial data using a generative adversarial\nnetwork (GAN). Our contribution uses coarse and widely available geospatial\ndata to create maps of less available features at the finer scale in the built\nenvironment, bypassing their traditional acquisition techniques (e.g. satellite\nimagery or land surveying). In the work, we employ land use data and road\nnetworks as input to generate building footprints and conduct experiments in 9\ncities around the world. The method, which we implement in a tool we release\nopenly, enables the translation of one geospatial dataset to another with high\nfidelity and morphological accuracy. It may be especially useful in locations\nmissing detailed and high-resolution data and those that are mapped with\nuncertain or heterogeneous quality, such as much of OpenStreetMap. The quality\nof the results is influenced by the urban form and scale. In most cases, the\nexperiments suggest promising performance as the method tends to truthfully\nindicate the locations, amount, and shape of buildings. The work has the\npotential to support several applications, such as energy, climate, and urban\nmorphology studies in areas previously lacking required data or inpainting\ngeospatial data in regions with incomplete data.', 'Convolutional neural network (CNN) have proven its success for semantic\nsegmentation, which is a core task of emerging industrial applications such as\nautonomous driving. However, most progress in semantic segmentation of urban\nscenes is reported on standard scenarios, i.e., daytime scenes with favorable\nillumination conditions. In practical applications, the outdoor weather and\nillumination are changeable, e.g., cloudy and nighttime, which results in a\nsignificant drop of semantic segmentation accuracy of CNN only trained with\ndaytime data. In this paper, we propose a novel generative adversarial network\n(namely Mutual-GAN) to alleviate the accuracy decline when daytime-trained\nneural network is applied to videos captured under adverse weather conditions.\nThe proposed Mutual-GAN adopts mutual information constraint to preserve\nimage-objects during cross-weather adaptation, which is an unsolved problem for\nmost unsupervised image-to-image translation approaches (e.g., CycleGAN). The\nproposed Mutual-GAN is evaluated on two publicly available driving video\ndatasets (i.e., CamVid and SYNTHIA). The experimental results demonstrate that\nour Mutual-GAN can yield visually plausible translated images and significantly\nimprove the semantic segmentation accuracy of daytime-trained deep learning\nnetwork while processing videos under challenging weathers.', 'We propose the novel use of a generative adversarial network (GAN) (i) to\nmake predictions in time (PredGAN) and (ii) to assimilate measurements\n(DA-PredGAN). In the latter case, we take advantage of the natural adjoint-like\nproperties of generative models and the ability to simulate forwards and\nbackwards in time. GANs have received much attention recently, after achieving\nexcellent results for their generation of realistic-looking images. We wish to\nexplore how this property translates to new applications in computational\nmodelling and to exploit the adjoint-like properties for efficient data\nassimilation. To predict the spread of COVID-19 in an idealised town, we apply\nthese methods to a compartmental model in epidemiology that is able to model\nspace and time variations. To do this, the GAN is set within a reduced-order\nmodel (ROM), which uses a low-dimensional space for the spatial distribution of\nthe simulation states. Then the GAN learns the evolution of the low-dimensional\nstates over time. The results show that the proposed methods can accurately\npredict the evolution of the high-fidelity numerical simulation, and can\nefficiently assimilate observed data and determine the corresponding model\nparameters.', 'The adverse haze weather condition has brought considerable difficulties in\nvision-based environmental applications. While, until now, most of the existing\nenvironmental monitoring studies are under ordinary conditions, and the studies\nof complex haze weather conditions have been ignored. Thence, this paper\nproposes a feature-supervised learning network based on generative adversarial\nnetworks (GAN) for environmental monitoring during hazy days. Its main idea is\nto train the model under the supervision of feature maps from the ground truth.\nFour key technical contributions are made in the paper. First, pairs of hazy\nand clean images are used as inputs to supervise the encoding process and\nobtain high-quality feature maps. Second, the basic GAN formulation is modified\nby introducing perception loss, style loss, and feature regularization loss to\ngenerate better results. Third, multi-scale images are applied as the input to\nenhance the performance of discriminator. Finally, a hazy remote sensing\ndataset is created for testing our dehazing method and environmental detection.\nExtensive experimental results show that the proposed method has achieved\nbetter performance than current state-of-the-art methods on both synthetic\ndatasets and real-world remote sensing images.', 'An open secret in contemporary machine learning is that many models work\nbeautifully on standard benchmarks but fail to generalize outside the lab. This\nhas been attributed to biased training data, which provide poor coverage over\nreal world events. Generative models are no exception, but recent advances in\ngenerative adversarial networks (GANs) suggest otherwise - these models can now\nsynthesize strikingly realistic and diverse images. Is generative modeling of\nphotos a solved problem? We show that although current GANs can fit standard\ndatasets very well, they still fall short of being comprehensive models of the\nvisual manifold. In particular, we study their ability to fit simple\ntransformations such as camera movements and color changes. We find that the\nmodels reflect the biases of the datasets on which they are trained (e.g.,\ncentered objects), but that they also exhibit some capacity for generalization:\nby ""steering"" in latent space, we can shift the distribution while still\ncreating realistic images. We hypothesize that the degree of distributional\nshift is related to the breadth of the training data distribution. Thus, we\nconduct experiments to quantify the limits of GAN transformations and introduce\ntechniques to mitigate the problem. Code is released on our project page:\nhttps://ali-design.github.io/gan_steerability/', 'Training robust supervised deep learning models for many geospatial\napplications of computer vision is difficult due to dearth of class-balanced\nand diverse training data. Conversely, obtaining enough training data for many\napplications is financially prohibitive or may be infeasible, especially when\nthe application involves modeling rare or extreme events. Synthetically\ngenerating data (and labels) using a generative model that can sample from a\ntarget distribution and exploit the multi-scale nature of images can be an\ninexpensive solution to address scarcity of labeled data. Towards this goal, we\npresent a deep conditional generative model, called VAE-Info-cGAN, that\ncombines a Variational Autoencoder (VAE) with a conditional Information\nMaximizing Generative Adversarial Network (InfoGAN), for synthesizing\nsemantically rich images simultaneously conditioned on a pixel-level condition\n(PLC) and a macroscopic feature-level condition (FLC). Dimensionally, the PLC\ncan only vary in the channel dimension from the synthesized image and is meant\nto be a task-specific input. The FLC is modeled as an attribute vector in the\nlatent space of the generated image which controls the contributions of various\ncharacteristic attributes germane to the target distribution. Experiments on a\nGPS trajectories dataset show that the proposed model can accurately generate\nvarious forms of spatiotemporal aggregates across different geographic\nlocations while conditioned only on a raster representation of the road\nnetwork. The primary intended application of the VAE-Info-cGAN is synthetic\ndata (and label) generation for targeted data augmentation for computer\nvision-based modeling of problems relevant to geospatial analysis and remote\nsensing.', ""The SARS-CoV-2 (Covid-19) pandemic has caused significant strain on public\nhealth institutions around the world. Contact tracing is an essential tool to\nchange the course of the Covid-19 pandemic. Manual contact tracing of Covid-19\ncases has significant challenges that limit the ability of public health\nauthorities to minimize community infections. Personalized peer-to-peer contact\ntracing through the use of mobile apps has the potential to shift the paradigm.\nSome countries have deployed centralized tracking systems, but more\nprivacy-protecting decentralized systems offer much of the same benefit without\nconcentrating data in the hands of a state authority or for-profit\ncorporations. Machine learning methods can circumvent some of the limitations\nof standard digital tracing by incorporating many clues and their uncertainty\ninto a more graded and precise estimation of infection risk. The estimated risk\ncan provide early risk awareness, personalized recommendations and relevant\ninformation to the user. Finally, non-identifying risk data can inform\nepidemiological models trained jointly with the machine learning predictor.\nThese models can provide statistical evidence for the importance of factors\ninvolved in disease transmission. They can also be used to monitor, evaluate\nand optimize health policy and (de)confinement scenarios according to medical\nand economic productivity indicators. However, such a strategy based on mobile\napps and machine learning should proactively mitigate potential ethical and\nprivacy risks, which could have substantial impacts on society (not only\nimpacts on health but also impacts such as stigmatization and abuse of personal\ndata). Here, we present an overview of the rationale, design, ethical\nconsiderations and privacy strategy of `COVI,' a Covid-19 public peer-to-peer\ncontact tracing and risk awareness mobile application developed in Canada."", 'Internal learning for single-image generation is a framework, where a\ngenerator is trained to produce novel images based on a single image. Since\nthese models are trained on a single image, they are limited in their scale and\napplication. To overcome these issues, we propose a meta-learning approach that\nenables training over a collection of images, in order to model the internal\nstatistics of the sample image more effectively. In the presented meta-learning\napproach, a single-image GAN model is generated given an input image, via a\nconvolutional feedforward hypernetwork $f$. This network is trained over a\ndataset of images, allowing for feature sharing among different models, and for\ninterpolation in the space of generative models. The generated single-image\nmodel contains a hierarchy of multiple generators and discriminators. It is\ntherefore required to train the meta-learner in an adversarial manner, which\nrequires careful design choices that we justify by a theoretical analysis. Our\nresults show that the models obtained are as suitable as single-image GANs for\nmany common image applications, significantly reduce the training time per\nimage without loss in performance, and introduce novel capabilities, such as\ninterpolation and feedforward modeling of novel images.', 'One of the major prerequisites for any deep learning approach is the\navailability of large-scale training data. When dealing with scanned document\nimages in real world scenarios, the principal information of its content is\nstored in the layout itself. In this work, we have proposed an automated deep\ngenerative model using Graph Neural Networks (GNNs) to generate synthetic data\nwith highly variable and plausible document layouts that can be used to train\ndocument interpretation systems, in this case, specially in digital mailroom\napplications. It is also the first graph-based approach for document layout\ngeneration task experimented on administrative document images, in this case,\ninvoices.', 'Document classification is a challenging task with important applications.\nThe deep learning approaches to the problem have gained much attention\nrecently. Despite the progress, the proposed models do not incorporate the\nknowledge of the document structure in the architecture efficiently and not\ntake into account the contexting importance of words and sentences. In this\npaper, we propose a new approach based on a combination of convolutional neural\nnetworks, gated recurrent units, and attention mechanisms for document\nclassification tasks. The main contribution of this work is the use of\nconvolution layers to extract more meaningful, generalizable and abstract\nfeatures by the hierarchical representation. The proposed method in this paper\nimproves the results of the current attention-based approaches for document\nclassification.', 'A lot of research has been devoted to identity documents analysis and\nrecognition on mobile devices. However, no publicly available datasets designed\nfor this particular problem currently exist. There are a few datasets which are\nuseful for associated subtasks but in order to facilitate a more comprehensive\nscientific and technical approach to identity document recognition more\nspecialized datasets are required. In this paper we present a Mobile Identity\nDocument Video dataset (MIDV-500) consisting of 500 video clips for 50\ndifferent identity document types with ground truth which allows to perform\nresearch in a wide scope of document analysis problems. The paper presents\ncharacteristics of the dataset and evaluation results for existing methods of\nface detection, text line recognition, and document fields data extraction.\nSince an important feature of identity documents is their sensitiveness as they\ncontain personal data, all source document images used in MIDV-500 are either\nin public domain or distributed under public copyright licenses.\n  The main goal of this paper is to present a dataset. However, in addition and\nas a baseline, we present evaluation results for existing methods for face\ndetection, text line recognition, and document data extraction, using the\npresented dataset.\n  (The dataset is available for download at ftp://smartengines.com/midv-500/.)', 'Additive Manufacturing (AM) is a crucial component of the smart industry. In\nthis paper, we propose an automated quality grading system for the AM process\nusing a deep convolutional neural network (CNN) model. The CNN model is trained\noffline using the images of the internal and surface defects in the\nlayer-by-layer deposition of materials and tested online by studying the\nperformance of detecting and classifying the failure in AM process at different\nextruder speeds and temperatures. The model demonstrates the accuracy of 94%\nand specificity of 96%, as well as above 75% in three classifier measures of\nthe Fscore, the sensitivity, and precision for classifying the quality of the\nprinting process in five grades in real-time. The proposed online model adds an\nautomated, consistent, and non-contact quality control signal to the AM process\nthat eliminates the manual inspection of parts after they are entirely built.\nThe quality monitoring signal can also be used by the machine to suggest\nremedial actions by adjusting the parameters in real-time. The proposed quality\npredictive model serves as a proof-of-concept for any type of AM machines to\nproduce reliable parts with fewer quality hiccups while limiting the waste of\nboth time and materials.', 'With the advent of mobile and hand-held cameras, document images have found\ntheir way into almost every domain. Dewarping of these images for the removal\nof perspective distortions and folds is essential so that they can be\nunderstood by document recognition algorithms. For this, we propose an\nend-to-end CNN architecture that can produce distortion free document images\nfrom warped documents it takes as input. We train this model on warped document\nimages simulated synthetically to compensate for lack of enough natural data.\nOur method is novel in the use of a bifurcated decoder with shared weights to\nprevent intermingling of grid coordinates, in the use of residual networks in\nthe U-Net skip connections to allow flow of data from different receptive\nfields in the model, and in the use of a gated network to help the model focus\non structure and line level detail of the document image. We evaluate our\nmethod on the DocUNet dataset, a benchmark in this domain, and obtain results\ncomparable to state-of-the-art methods.', 'Creating fake images and videos such as ""Deepfake"" has become much easier\nthese days due to the advancement in Generative Adversarial Networks (GANs).\nMoreover, recent research such as the few-shot learning can create highly\nrealistic personalized fake images with only a few images. Therefore, the\nthreat of Deepfake to be used for a variety of malicious intents such as\npropagating fake images and videos becomes prevalent. And detecting these\nmachine-generated fake images has been quite challenging than ever. In this\nwork, we propose a light-weight robust fine-tuning neural network-based\nclassifier architecture called Fake Detection Fine-tuning Network (FDFtNet),\nwhich is capable of detecting many of the new fake face image generation\nmodels, and can be easily combined with existing image classification networks\nand finetuned on a few datasets. In contrast to many existing methods, our\napproach aims to reuse popular pre-trained models with only a few images for\nfine-tuning to effectively detect fake images. The core of our approach is to\nintroduce an image-based self-attention module called Fine-Tune Transformer\nthat uses only the attention module and the down-sampling layer. This module is\nadded to the pre-trained model and fine-tuned on a few data to search for new\nsets of feature space to detect fake images. We experiment with our FDFtNet on\nthe GANsbased dataset (Progressive Growing GAN) and Deepfake-based dataset\n(Deepfake and Face2Face) with a small input image resolution of 64x64 that\ncomplicates detection. Our FDFtNet achieves an overall accuracy of 90.29% in\ndetecting fake images generated from the GANs-based dataset, outperforming the\nstate-of-the-art.', 'Climate change is one of the greatest challenges facing humanity, and we, as\nmachine learning experts, may wonder how we can help. Here we describe how\nmachine learning can be a powerful tool in reducing greenhouse gas emissions\nand helping society adapt to a changing climate. From smart grids to disaster\nmanagement, we identify high impact problems where existing gaps can be filled\nby machine learning, in collaboration with other fields. Our recommendations\nencompass exciting research questions as well as promising business\nopportunities. We call on the machine learning community to join the global\neffort against climate change.', ""As AI-based medical devices are becoming more common in imaging fields like\nradiology and histology, interpretability of the underlying predictive models\nis crucial to expand their use in clinical practice. Existing heatmap-based\ninterpretability methods such as GradCAM only highlight the location of\npredictive features but do not explain how they contribute to the prediction.\nIn this paper, we propose a new interpretability method that can be used to\nunderstand the predictions of any black-box model on images, by showing how the\ninput image would be modified in order to produce different predictions. A\nStyleGAN is trained on medical images to provide a mapping between latent\nvectors and images. Our method identifies the optimal direction in the latent\nspace to create a change in the model prediction. By shifting the latent\nrepresentation of an input image along this direction, we can produce a series\nof new synthetic images with changed predictions. We validate our approach on\nhistology and radiology images, and demonstrate its ability to provide\nmeaningful explanations that are more informative than GradCAM heatmaps. Our\nmethod reveals the patterns learned by the model, which allows clinicians to\nbuild trust in the model's predictions, discover new biomarkers and eventually\nreveal potential biases."", 'Taking a photo outside, can we predict the immediate future, e.g., how would\nthe cloud move in the sky? We address this problem by presenting a generative\nadversarial network (GAN) based two-stage approach to generating realistic\ntime-lapse videos of high resolution. Given the first frame, our model learns\nto generate long-term future frames. The first stage generates videos of\nrealistic contents for each frame. The second stage refines the generated video\nfrom the first stage by enforcing it to be closer to real videos with regard to\nmotion dynamics. To further encourage vivid motion in the final generated\nvideo, Gram matrix is employed to model the motion more precisely. We build a\nlarge scale time-lapse dataset, and test our approach on this new dataset.\nUsing our model, we are able to generate realistic videos of up to $128\\times\n128$ resolution for 32 frames. Quantitative and qualitative experiment results\nhave demonstrated the superiority of our model over the state-of-the-art\nmodels.', 'Training robust supervised deep learning models for many geospatial\napplications of computer vision is difficult due to dearth of class-balanced\nand diverse training data. Conversely, obtaining enough training data for many\napplications is financially prohibitive or may be infeasible, especially when\nthe application involves modeling rare or extreme events. Synthetically\ngenerating data (and labels) using a generative model that can sample from a\ntarget distribution and exploit the multi-scale nature of images can be an\ninexpensive solution to address scarcity of labeled data. Towards this goal, we\npresent a deep conditional generative model, called VAE-Info-cGAN, that\ncombines a Variational Autoencoder (VAE) with a conditional Information\nMaximizing Generative Adversarial Network (InfoGAN), for synthesizing\nsemantically rich images simultaneously conditioned on a pixel-level condition\n(PLC) and a macroscopic feature-level condition (FLC). Dimensionally, the PLC\ncan only vary in the channel dimension from the synthesized image and is meant\nto be a task-specific input. The FLC is modeled as an attribute vector in the\nlatent space of the generated image which controls the contributions of various\ncharacteristic attributes germane to the target distribution. An interpretation\nof the attribute vector to systematically generate synthetic images by varying\na chosen binary macroscopic feature is explored. Experiments on a GPS\ntrajectories dataset show that the proposed model can accurately generate\nvarious forms of spatio-temporal aggregates across different geographic\nlocations while conditioned only on a raster representation of the road\nnetwork. The primary intended application of the VAE-Info-cGAN is synthetic\ndata (and label) generation for targeted data augmentation for computer\nvision-based modeling of problems relevant to geospatial analysis and remote\nsensing.', 'This paper presents results of Document Visual Question Answering Challenge\norganized as part of ""Text and Documents in the Deep Learning Era"" workshop, in\nCVPR 2020. The challenge introduces a new problem - Visual Question Answering\non document images. The challenge comprised two tasks. The first task concerns\nwith asking questions on a single document image. On the other hand, the second\ntask is set as a retrieval task where the question is posed over a collection\nof images. For the task 1 a new dataset is introduced comprising 50,000\nquestions-answer(s) pairs defined over 12,767 document images. For task 2\nanother dataset has been created comprising 20 questions over 14,362 document\nimages which share the same document template.', 'Identity of a vehicle is done through the vehicle license plate by traffic\npolice in general. Au- tomatic vehicle license plate recognition has several\napplications in intelligent traffic management systems. The security situation\nacross the globe and particularly in India demands a need to equip the traffic\npolice with a system that enables them to get instant details of a vehicle. The\nsystem should be easy to use, should be mobile, and work 24 x 7. In this paper,\nwe describe a mobile phone based, client-server architected, license plate\nrecognition system. While we use the state of the art image processing and\npattern recognition algorithms tuned for Indian conditions to automatically\nrecognize non-uniform license plates, the main contribution is in creating an\nend to end usable solution. The client application runs on a mobile device and\na server application, with access to vehicle information database, is hosted\ncentrally. The solution enables capture of license plate image captured by the\nphone camera and passes to the server; on the server the license plate number\nis recognized; the data associated with the number plate is then sent back to\nthe mobile device, instantaneously. We describe the end to end system\narchitecture in detail. A working prototype of the proposed system has been\nimplemented in the lab environment.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'Document-level relation extraction aims to extract relations among entities\nwithin a document. Different from sentence-level relation extraction, it\nrequires reasoning over multiple sentences across a document. In this paper, we\npropose Graph Aggregation-and-Inference Network (GAIN) featuring double graphs.\nGAIN first constructs a heterogeneous mention-level graph (hMG) to model\ncomplex interaction among different mentions across the document. It also\nconstructs an entity-level graph (EG), based on which we propose a novel path\nreasoning mechanism to infer relations between entities. Experiments on the\npublic dataset, DocRED, show GAIN achieves a significant performance\nimprovement (2.85 on F1) over the previous state-of-the-art. Our code is\navailable at https://github.com/DreamInvoker/GAIN .', 'The quality of a document is affected by various factors, including\ngrammaticality, readability, stylistics, and expertise depth, making the task\nof document quality assessment a complex one. In this paper, we explore this\ntask in the context of assessing the quality of Wikipedia articles and academic\npapers. Observing that the visual rendering of a document can capture implicit\nquality indicators that are not present in the document text --- such as\nimages, font choices, and visual layout --- we propose a joint model that\ncombines the text content with a visual rendering of the document for document\nquality assessment. Experimental results over two datasets reveal that textual\nand visual features are complementary, achieving state-of-the-art results.', ""In this work, a region-based Deep Convolutional Neural Network framework is\nproposed for document structure learning. The contribution of this work\ninvolves efficient training of region based classifiers and effective\nensembling for document image classification. A primary level of `inter-domain'\ntransfer learning is used by exporting weights from a pre-trained VGG16\narchitecture on the ImageNet dataset to train a document classifier on whole\ndocument images. Exploiting the nature of region based influence modelling, a\nsecondary level of `intra-domain' transfer learning is used for rapid training\nof deep learning models for image segments. Finally, stacked generalization\nbased ensembling is utilized for combining the predictions of the base deep\nneural network models. The proposed method achieves state-of-the-art accuracy\nof 92.2% on the popular RVL-CDIP document image dataset, exceeding benchmarks\nset by existing algorithms."", 'The digital camera captured document images may often be warped and distorted\ndue to different camera angles or document surfaces. A robust technique is\nneeded to solve this kind of distortion. The research on dewarping of the\ndocument suffers due to the limited availability of benchmark public dataset.\nIn recent times, deep learning based approaches are used to solve the problems\naccurately. To train most of the deep neural networks a large number of\ndocument images is required and generating such a large volume of document\nimages manually is difficult. In this paper, we propose a technique to generate\na synthetic warped image from a flat-bedded scanned document image. It is done\nby calculating warping factors for each pixel position using two warping\nposition parameters (WPP) and eight warping control parameters (WCP). These\nparameters can be specified as needed depending upon the desired warping. The\nresults are compared with similar real captured images both qualitative and\nquantitative way.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Context-aware neural machine translation (NMT) is a promising direction to\nimprove the translation quality by making use of the additional context, e.g.,\ndocument-level translation, or having meta-information. Although there exist\nvarious architectures and analyses, the effectiveness of different\ncontext-aware NMT models is not well explored yet. This paper analyzes the\nperformance of document-level NMT models on four diverse domains with a varied\namount of parallel document-level bilingual data. We conduct a comprehensive\nset of experiments to investigate the impact of document-level NMT. We find\nthat there is no single best approach to document-level NMT, but rather that\ndifferent architectures come out on top on different tasks. Looking at\ntask-specific problems, such as pronoun resolution or headline translation, we\nfind improvements in the context-aware systems, even in cases where the\ncorpus-level metrics like BLEU show no significant improvement. We also show\nthat document-level back-translation significantly helps to compensate for the\nlack of document-level bi-texts.', 'The first phase of table recognition is to detect the tabular area in a\ndocument. Subsequently, the tabular structures are recognized in the second\nphase in order to extract information from the respective cells. Table\ndetection and structural recognition are pivotal problems in the domain of\ntable understanding. However, table analysis is a perplexing task due to the\ncolossal amount of diversity and asymmetry in tables. Therefore, it is an\nactive area of research in document image analysis. Recent advances in the\ncomputing capabilities of graphical processing units have enabled deep neural\nnetworks to outperform traditional state-of-the-art machine learning methods.\nTable understanding has substantially benefited from the recent breakthroughs\nin deep neural networks. However, there has not been a consolidated description\nof the deep learning methods for table detection and table structure\nrecognition. This review paper provides a thorough analysis of the modern\nmethodologies that utilize deep neural networks. This work provided a thorough\nunderstanding of the current state-of-the-art and related challenges of table\nunderstanding in document images. Furthermore, the leading datasets and their\nintricacies have been elaborated along with the quantitative results. Moreover,\na brief overview is given regarding the promising directions that can serve as\na guide to further improve table analysis in document images.', 'Cross-lingual document representations enable language understanding in\nmultilingual contexts and allow transfer learning from high-resource to\nlow-resource languages at the document level. Recently large pre-trained\nlanguage models such as BERT, XLM and XLM-RoBERTa have achieved great success\nwhen fine-tuned on sentence-level downstream tasks. It is tempting to apply\nthese cross-lingual models to document representation learning. However, there\nare two challenges: (1) these models impose high costs on long document\nprocessing and thus many of them have strict length limit; (2) model\nfine-tuning requires extra data and computational resources, which is not\npractical in resource-limited settings. In this work, we address these\nchallenges by proposing unsupervised Language-Agnostic Weighted Document\nRepresentations (LAWDR). We study the geometry of pre-trained sentence\nembeddings and leverage it to derive document representations without\nfine-tuning. Evaluated on cross-lingual document alignment, LAWDR demonstrates\ncomparable performance to state-of-the-art models on benchmark datasets.', 'Document Layout Analysis is a fundamental step in Handwritten Text Processing\nsystems, from the extraction of the text lines to the type of zone it belongs\nto. We present a system based on artificial neural networks which is able to\ndetermine not only the baselines of text lines present in the document, but\nalso performs geometric and logic layout analysis of the document. Experiments\nin three different datasets demonstrate the potential of the method and show\ncompetitive results with respect to state-of-the-art methods.', 'Graphical elements: particularly tables and figures contain a visual summary\nof the most valuable information contained in a document. Therefore,\nlocalization of such graphical objects in the document images is the initial\nstep to understand the content of such graphical objects or document images. In\nthis paper, we present a novel end-to-end trainable deep learning based\nframework to localize graphical objects in the document images called as\nGraphical Object Detection (GOD). Our framework is data-driven and does not\nrequire any heuristics or meta-data to locate graphical objects in the document\nimages. The GOD explores the concept of transfer learning and domain adaptation\nto handle scarcity of labeled training images for graphical object detection\ntask in the document images. Performance analysis carried out on the various\npublic benchmark data sets: ICDAR-2013, ICDAR-POD2017,and UNLV shows that our\nmodel yields promising results as compared to state-of-the-art techniques.', 'When digitizing a document into an image, it is common to include a\nsurrounding border region to visually indicate that the entire document is\npresent in the image. However, this border should be removed prior to automated\nprocessing. In this work, we present a deep learning based system, PageNet,\nwhich identifies the main page region in an image in order to segment content\nfrom both textual and non-textual border noise. In PageNet, a Fully\nConvolutional Network obtains a pixel-wise segmentation which is post-processed\ninto the output quadrilateral region. We evaluate PageNet on 4 collections of\nhistorical handwritten documents and obtain over 94% mean intersection over\nunion on all datasets and approach human performance on 2 of these collections.\nAdditionally, we show that PageNet can segment documents that are overlayed on\ntop of other documents.', 'Identity documents recognition is an important sub-field of document\nanalysis, which deals with tasks of robust document detection, type\nidentification, text fields recognition, as well as identity fraud prevention\nand document authenticity validation given photos, scans, or video frames of an\nidentity document capture. Significant amount of research has been published on\nthis topic in recent years, however a chief difficulty for such research is\nscarcity of datasets, due to the subject matter being protected by security\nrequirements. A few datasets of identity documents which are available lack\ndiversity of document types, capturing conditions, or variability of document\nfield values. In addition, the published datasets were typically designed only\nfor a subset of document recognition problems, not for a complex identity\ndocument analysis. In this paper, we present a dataset MIDV-2020 which consists\nof 1000 video clips, 2000 scanned images, and 1000 photos of 1000 unique mock\nidentity documents, each with unique text field values and unique artificially\ngenerated faces, with rich annotation. For the presented benchmark dataset\nbaselines are provided for such tasks as document location and identification,\ntext fields recognition, and face detection. With 72409 annotated images in\ntotal, to the date of publication the proposed dataset is the largest publicly\navailable identity documents dataset with variable artificially generated data,\nand we believe that it will prove invaluable for advancement of the field of\ndocument analysis and recognition. The dataset is available for download at\nftp://smartengines.com/midv-2020 and http://l3i-share.univ-lr.fr .', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'It has been shown that most machine learning algorithms are susceptible to\nadversarial perturbations. Slightly perturbing an image in a carefully chosen\ndirection in the image space may cause a trained neural network model to\nmisclassify it. Recently, it was shown that physical adversarial examples\nexist: printing perturbed images then taking pictures of them would still\nresult in misclassification. This raises security and safety concerns.\n  However, these experiments ignore a crucial property of physical objects: the\ncamera can view objects from different distances and at different angles. In\nthis paper, we show experiments that suggest that current constructions of\nphysical adversarial examples do not disrupt object detection from a moving\nplatform. Instead, a trained neural network classifies most of the pictures\ntaken from different distances and angles of a perturbed image correctly. We\nbelieve this is because the adversarial property of the perturbation is\nsensitive to the scale at which the perturbed picture is viewed, so (for\nexample) an autonomous car will misclassify a stop sign only from a small range\nof distances.\n  Our work raises an important question: can one construct examples that are\nadversarial for many or most viewing conditions? If so, the construction should\noffer very significant insights into the internal representation of patterns by\ndeep networks. If not, there is a good prospect that adversarial examples can\nbe reduced to a curiosity with little practical impact.', 'Document-level relation extraction aims to discover relations between\nentities across a whole document. How to build the dependency of entities from\ndifferent sentences in a document remains to be a great challenge. Current\napproaches either leverage syntactic trees to construct document-level graphs\nor aggregate inference information from different sentences. In this paper, we\nbuild cross-sentence dependencies by inferring compositional relations between\ninter-sentence mentions. Adopting aggressive linking strategy, intermediate\nrelations are reasoned on the document-level graphs by mention convolution. We\nfurther notice the generalization problem of NA instances, which is caused by\nincomplete annotation and worsened by fully-connected mention pairs. An\nimproved ranking loss is proposed to attend this problem. Experiments show the\nconnections between different mentions are crucial to document-level relation\nextraction, which enables the model to extract more meaningful higher-level\ncompositional relations.', ""In this paper we investigate forgetting in disjunctive logic programs, where\nforgetting an atom from a program amounts to a reduction in the signature of\nthat program. The goal is to provide an approach that is syntax-independent, in\nthat if two programs are strongly equivalent, then the results of forgetting an\natom in each program should also be strongly equivalent. Our central definition\nof forgetting is impractical but satisfies this goal: Forgetting an atom is\ncharacterised by the set of SE consequences of the program that do not mention\nthe atom to be forgotten. We then provide an equivalent, practical definition,\nwherein forgetting an atom $p$ is given by those rules in the program that\ndon't mention $p$, together with rules obtained by a single inference step from\nrules that do mention $p$. Forgetting is shown to have appropriate properties;\nas well, the finite characterisation results in a modest (at worst quadratic)\nblowup. Finally we have also obtained a prototype implementation of this\napproach to forgetting."", 'This paper describes an application of reinforcement learning to the mention\ndetection task. We define a novel action-based formulation for the mention\ndetection task, in which a model can flexibly revise past labeling decisions by\ngrouping together tokens and assigning partial mention labels. We devise a\nmethod to create mention-level episodes and we train a model by rewarding\ncorrectly labeled complete mentions, irrespective of the inner structure\ncreated. The model yields results which are on par with a competitive\nsupervised counterpart while being more flexible in terms of achieving targeted\nbehavior through reward modeling and generating internal mention structure,\nespecially on longer mentions.', 'One major drawback of state of the art Neural Networks (NN)-based approaches\nfor document classification purposes is the large number of training samples\nrequired to obtain an efficient classification. The minimum required number is\naround one thousand annotated documents for each class. In many cases it is\nvery difficult, if not impossible, to gather this number of samples in real\nindustrial processes. In this paper, we analyse the efficiency of NN-based\ndocument classification systems in a sub-optimal training case, based on the\nsituation of a company document stream. We evaluated three different\napproaches, one based on image content and two on textual content. The\nevaluation was divided into four parts: a reference case, to assess the\nperformance of the system in the lab; two cases that each simulate a specific\ndifficulty linked to document stream processing; and a realistic case that\ncombined all of these difficulties. The realistic case highlighted the fact\nthat there is a significant drop in the efficiency of NN-Based document\nclassification systems. Although they remain efficient for well represented\nclasses (with an over-fitting of the system for those classes), it is\nimpossible for them to handle appropriately less well represented classes.\nNN-Based document classification systems need to be adapted to resolve these\ntwo problems before they can be considered for use in a company document\nstream.', 'Named entity linking is to map an ambiguous mention in documents to an entity\nin a knowledge base. The named entity linking is challenging, given the fact\nthat there are multiple candidate entities for a mention in a document. It is\ndifficult to link a mention when it appears multiple times in a document, since\nthere are conflicts by the contexts around the appearances of the mention. In\naddition, it is difficult since the given training dataset is small due to the\nreason that it is done manually to link a mention to its mapping entity. In the\nliterature, there are many reported studies among which the recent embedding\nmethods learn vectors of entities from the training dataset at document level.\nTo address these issues, we focus on how to link entity for mentions at a\nsentence level, which reduces the noises introduced by different appearances of\nthe same mention in a document at the expense of insufficient information to be\nused. We propose a new unified embedding method by maximizing the relationships\nlearned from knowledge graphs. We confirm the effectiveness of our method in\nour experimental studies.', 'Modern entity linking systems rely on large collections of documents\nspecifically annotated for the task (e.g., AIDA CoNLL). In contrast, we propose\nan approach which exploits only naturally occurring information: unlabeled\ndocuments and Wikipedia. Our approach consists of two stages. First, we\nconstruct a high recall list of candidate entities for each mention in an\nunlabeled document. Second, we use the candidate lists as weak supervision to\nconstrain our document-level entity linking model. The model treats entities as\nlatent variables and, when estimated on a collection of unlabelled texts,\nlearns to choose entities relying both on local context of each mention and on\ncoherence with other entities in the document. The resulting approach rivals\nfully-supervised state-of-the-art systems on standard test sets. It also\napproaches their performance in the very challenging setting: when tested on a\ntest set sampled from the data used to estimate the supervised systems. By\ncomparing to Wikipedia-only training of our model, we demonstrate that modeling\nunlabeled documents is beneficial.', 'Various neural-based methods have been proposed so far for joint mention\ndetection and coreference resolution. However, existing works on coreference\nresolution are mainly dependent on filtered mention representation, while other\nspans are largely neglected. In this paper, we aim at increasing the\nutilization rate of data and investigating whether those eliminated spans are\ntotally useless, or to what extent they can improve the performance of\ncoreference resolution. To achieve this, we propose a mention representation\nrefining strategy where spans highly related to mentions are well leveraged\nusing a pointer network for representation enhancing. Notably, we utilize an\nadditional loss term in this work to encourage the diversity between entity\nclusters. Experimental results on the document-level CoNLL-2012 Shared Task\nEnglish dataset show that eliminated spans are indeed much effective and our\napproach can achieve competitive results when compared with previous\nstate-of-the-art in coreference resolution.', 'Relation extraction (RE) aims to identify the semantic relations between\nnamed entities in text. Recent years have witnessed it raised to the document\nlevel, which requires complex reasoning with entities and mentions throughout\nan entire document. In this paper, we propose a novel model to document-level\nRE, by encoding the document information in terms of entity global and local\nrepresentations as well as context relation representations. Entity global\nrepresentations model the semantic information of all entities in the document,\nentity local representations aggregate the contextual information of multiple\nmentions of specific entities, and context relation representations encode the\ntopic information of other relations. Experimental results demonstrate that our\nmodel achieves superior performance on two public datasets for document-level\nRE. It is particularly effective in extracting relations between entities of\nlong distance and having multiple mentions.', 'Document segmentation is a method of rending the document into distinct\nregions. A document is an assortment of information and a standard mode of\nconveying information to others. Pursuance of data from documents involves ton\nof human effort, time intense and might severely prohibit the usage of data\nsystems. So, automatic information pursuance from the document has become a big\nissue. It is been shown that document segmentation will facilitate to beat such\nproblems. This paper proposes a new approach to segment and classify the\ndocument regions as text, image, drawings and table. Document image is divided\ninto blocks using Run length smearing rule and features are extracted from\nevery blocks. Discipulus tool has been used to construct the Genetic\nprogramming based classifier model and located 97.5% classification accuracy.', 'Computation of document image quality metrics often depends upon the\navailability of a ground truth image corresponding to the document. This limits\nthe applicability of quality metrics in applications such as hyperparameter\noptimization of image processing algorithms that operate on-the-fly on unseen\ndocuments. This work proposes the use of surrogate models to learn the behavior\nof a given document quality metric on existing datasets where ground truth\nimages are available. The trained surrogate model can later be used to predict\nthe metric value on previously unseen document images without requiring access\nto ground truth images. The surrogate model is empirically evaluated on the\nDocument Image Binarization Competition (DIBCO) and the Handwritten Document\nImage Binarization Competition (H-DIBCO) datasets.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.']"
-"The query consists of users’ needs, leading to several research questions that span a paragraph. Each candidate passage is an abstract from a scientific paper. The objective of this information retrieval task is to identify the abstract that most effectively meets the user’s needs in the query : In the context of time series forecasting, it's observed that anomalies in certain data sources, such as seismic activity, heart rate monitors, and nuclear plant cooling system temperatures, occur infrequently but often result in catastrophic outcomes like earthquakes, heart attacks, and nuclear meltdowns. My objective is to devise a time series forecasting model capable of predicting these rare anomalies with high precision. This model must be equipped to manage extremely sparse supervised signals. To train my model effectively, I believe it's necessary to employ some form of sampling or bootstrapping to simulate various event pathways leading to these anomalies. Essentially, this involves data augmentation to generate more positive signals for training. Consequently, I require a simulation agent capable of learning the underlying time series data distribution. This simulation agent (model) must also be generative to create new pathways and simulate new events. I propose that the simulation agent should incorporate a reinforcement learning component. However, I am open to any literature that discusses rare event simulation, regardless of whether the method is statistical or deep learning-based. Ultimately, I am confident that my model will be robust enough for deployment in real-world scenarios for disaster prevention and early warning systems.","['One of the limiting factors in training data-driven, rare-event prediction\nalgorithms is the scarcity of the events of interest resulting in an extreme\nimbalance in the data. There have been many methods introduced in the\nliterature for overcoming this issue; simple data manipulation through\nundersampling and oversampling, utilizing cost-sensitive learning algorithms,\nor by generating synthetic data points following the distribution of the\nexisting data. While synthetic data generation has recently received a great\ndeal of attention, there are real challenges involved in doing so for\nhigh-dimensional data such as multivariate time series. In this study, we\nexplore the usefulness of the conditional generative adversarial network (CGAN)\nas a means to perform data-informed oversampling in order to balance a large\ndataset of multivariate time series. We utilize a flare forecasting benchmark\ndataset, named SWAN-SF, and design two verification methods to both\nquantitatively and qualitatively evaluate the similarity between the generated\nminority and the ground-truth samples. We further assess the quality of the\ngenerated samples by training a classical, supervised machine learning\nalgorithm on synthetic data, and testing the trained model on the unseen, real\ndata. The results show that the classifier trained on the data augmented with\nthe synthetic multivariate time series achieves a significant improvement\ncompared with the case where no augmentation is used. The popular flare\nforecasting evaluation metrics, TSS and HSS, report 20-fold and 5-fold\nimprovements, respectively, indicating the remarkable statistical similarities,\nand the usefulness of CGAN-based data generation for complicated tasks such as\nflare forecasting.', 'Generative Adversarial Net (GAN) has been proven to be a powerful machine\nlearning tool in image data analysis and generation. In this paper, we propose\nto use Conditional Generative Adversarial Net (CGAN) to learn and simulate time\nseries data. The conditions can be both categorical and continuous variables\ncontaining different kinds of auxiliary information. Our simulation studies\nshow that CGAN is able to learn different kinds of normal and heavy tail\ndistributions, as well as dependent structures of different time series and it\ncan further generate conditional predictive distributions consistent with the\ntraining data distributions. We also provide an in-depth discussion on the\nrationale of GAN and the neural network as hierarchical splines to draw a clear\nconnection with the existing statistical method for distribution generation. In\npractice, CGAN has a wide range of applications in the market risk and\ncounterparty risk analysis: it can be applied to learn the historical data and\ngenerate scenarios for the calculation of Value-at-Risk (VaR) and Expected\nShortfall (ES) and predict the movement of the market risk factors. We present\na real data analysis including a backtesting to demonstrate CGAN is able to\noutperform the Historic Simulation, a popular method in market risk analysis\nfor the calculation of VaR. CGAN can also be applied in the economic time\nseries modeling and forecasting, and an example of hypothetical shock analysis\nfor economic models and the generation of potential CCAR scenarios by CGAN is\ngiven at the end of the paper.', 'Time series data are often corrupted by outliers or other kinds of anomalies.\nIdentifying the anomalous points can be a goal on its own (anomaly detection),\nor a means to improving performance of other time series tasks (e.g.\nforecasting). Recent deep-learning-based approaches to anomaly detection and\nforecasting commonly assume that the proportion of anomalies in the training\ndata is small enough to ignore, and treat the unlabeled data as coming from the\nnominal data distribution. We present a simple yet effective technique for\naugmenting existing time series models so that they explicitly account for\nanomalies in the training data. By augmenting the training data with a latent\nanomaly indicator variable whose distribution is inferred while training the\nunderlying model using Monte Carlo EM, our method simultaneously infers\nanomalous points while improving model performance on nominal data. We\ndemonstrate the effectiveness of the approach by combining it with a simple\nfeed-forward forecasting model. We investigate how anomalies in the train set\naffect the training of forecasting models, which are commonly used for time\nseries anomaly detection, and show that our method improves the training of the\nmodel.', 'To predict rare extreme events using deep neural networks, one encounters the\nso-called small data problem because even long-term observations often contain\nfew extreme events. Here, we investigate a model-assisted framework where the\ntraining data is obtained from numerical simulations, as opposed to\nobservations, with adequate samples from extreme events. However, to ensure the\ntrained networks are applicable in practice, the training is not performed on\nthe full simulation data; instead we only use a small subset of observable\nquantities which can be measured in practice. We investigate the feasibility of\nthis model-assisted framework on three different dynamical systems (Rossler\nattractor, FitzHugh-Nagumo model, and a turbulent fluid flow) and three\ndifferent deep neural network architectures (feedforward, long short-term\nmemory, and reservoir computing). In each case, we study the prediction\naccuracy, robustness to noise, reproducibility under repeated training, and\nsensitivity to the type of input data. In particular, we find long short-term\nmemory networks to be most robust to noise and to yield relatively accurate\npredictions, while requiring minimal fine-tuning of the hyperparameters.']","['Robust Anomaly Detection (AD) on time series data is a key component for\nmonitoring many complex modern systems. These systems typically generate\nhigh-dimensional time series that can be highly noisy, seasonal, and\ninter-correlated. This paper explores some of the challenges in such data, and\nproposes a new approach that makes inroads towards increased robustness on\nseasonal and contaminated data, while providing a better root cause\nidentification of anomalies. In particular, we propose the use of Robust\nSeasonal Multivariate Generative Adversarial Network (RSM-GAN) that extends\nrecent advancements in GAN with the adoption of convolutional-LSTM layers and\nattention mechanisms to produce excellent performance on various settings. We\nconduct extensive experiments in which not only do this model displays more\nrobust behavior on complex seasonality patterns, but also shows increased\nresistance to training data contamination. We compare it with existing\nclassical and deep-learning AD models, and show that this architecture is\nassociated with the lowest false positive rate and improves precision by 30%\nand 16% in real-world and synthetic data, respectively.', ""Today's Cyber-Physical Systems (CPSs) are large, complex, and affixed with\nnetworked sensors and actuators that are targets for cyber-attacks.\nConventional detection techniques are unable to deal with the increasingly\ndynamic and complex nature of the CPSs. On the other hand, the networked\nsensors and actuators generate large amounts of data streams that can be\ncontinuously monitored for intrusion events. Unsupervised machine learning\ntechniques can be used to model the system behaviour and classify deviant\nbehaviours as possible attacks. In this work, we proposed a novel Generative\nAdversarial Networks-based Anomaly Detection (GAN-AD) method for such complex\nnetworked CPSs. We used LSTM-RNN in our GAN to capture the distribution of the\nmultivariate time series of the sensors and actuators under normal working\nconditions of a CPS. Instead of treating each sensor's and actuator's time\nseries independently, we model the time series of multiple sensors and\nactuators in the CPS concurrently to take into account of potential latent\ninteractions between them. To exploit both the generator and the discriminator\nof our GAN, we deployed the GAN-trained discriminator together with the\nresiduals between generator-reconstructed data and the actual samples to detect\npossible anomalies in the complex CPS. We used our GAN-AD to distinguish\nabnormal attacked situations from normal working conditions for a complex\nsix-stage Secure Water Treatment (SWaT) system. Experimental results showed\nthat the proposed strategy is effective in identifying anomalies caused by\nvarious attacks with high detection rate and low false positive rate as\ncompared to existing methods."", ""Although aviation accidents are rare, safety incidents occur more frequently\nand require a careful analysis to detect and mitigate risks in a timely manner.\nAnalyzing safety incidents using operational data and producing event-based\nexplanations is invaluable to airline companies as well as to governing\norganizations such as the Federal Aviation Administration (FAA) in the United\nStates. However, this task is challenging because of the complexity involved in\nmining multi-dimensional heterogeneous time series data, the lack of\ntime-step-wise annotation of events in a flight, and the lack of scalable tools\nto perform analysis over a large number of events. In this work, we propose a\nprecursor mining algorithm that identifies events in the multidimensional time\nseries that are correlated with the safety incident. Precursors are valuable to\nsystems health and safety monitoring and in explaining and forecasting safety\nincidents. Current methods suffer from poor scalability to high dimensional\ntime series data and are inefficient in capturing temporal behavior. We propose\nan approach by combining multiple-instance learning (MIL) and deep recurrent\nneural networks (DRNN) to take advantage of MIL's ability to learn using weakly\nsupervised data and DRNN's ability to model temporal behavior. We describe the\nalgorithm, the data, the intuition behind taking a MIL approach, and a\ncomparative analysis of the proposed algorithm with baseline models. We also\ndiscuss the application to a real-world aviation safety problem using data from\na commercial airline company and discuss the model's abilities and\nshortcomings, with some final remarks about possible deployment directions."", 'Large network logs, recording multivariate time series generated from\nheterogeneous devices and sensors in a network, can often reveal important\ninformation about abnormal activities, such as network intrusions and device\nmalfunctions. Existing machine learning methods for anomaly detection on\nmultivariate time series typically assume that 1) normal sequences would have\nconsistent behavior for training unsupervised models, or 2) require a large set\nof labeled normal and abnormal sequences for supervised models. However, in\npractice, normal network activities can demonstrate significantly varying\nsequence patterns (e.g., before and after rerouting partial network traffic).\nAlso, the recorded abnormal events can be sparse. This paper presents a novel\nsemi-supervised method that efficiently captures dependencies between network\ntime series and across time points to generate meaningful representations of\nnetwork activities for predicting abnormal events. The method can use the\nlimited labeled data to explicitly learn separable embedding space for normal\nand abnormal samples and effectively leverage unlabeled data to handle training\ndata scarcity. The experiments demonstrate that our approach significantly\noutperformed state-of-the-art approaches for event detection on a large\nreal-world network log.', 'Machine learning and in particular deep learning algorithms are the emerging\napproaches to data analysis. These techniques have transformed traditional data\nmining-based analysis radically into a learning-based model in which existing\ndata sets along with their cluster labels (i.e., train set) are learned to\nbuild a supervised learning model and predict the cluster labels of unseen data\n(i.e., test set). In particular, deep learning techniques are capable of\ncapturing and learning hidden features in a given data sets and thus building a\nmore accurate prediction model for clustering and labeling problem. However,\nthe major problem is that time series data are often unlabeled and thus\nsupervised learning-based deep learning algorithms cannot be directly adapted\nto solve the clustering problems for these special and complex types of data\nsets. To address this problem, this paper introduces a two-stage method for\nclustering time series data. First, a novel technique is introduced to utilize\nthe characteristics (e.g., volatility) of given time series data in order to\ncreate labels and thus be able to transform the problem from unsupervised\nlearning into supervised learning. Second, an autoencoder-based deep learning\nmodel is built to learn and model both known and hidden features of time series\ndata along with their created labels to predict the labels of unseen time\nseries data. The paper reports a case study in which financial and stock time\nseries data of selected 70 stock indices are clustered into distinct groups\nusing the introduced two-stage procedure. The results show that the proposed\nprocedure is capable of achieving 87.5\\% accuracy in clustering and predicting\nthe labels for unseen time series data.', 'Long-tail and rare event problems become crucial when autonomous driving\nalgorithms are applied in the real world. For the purpose of evaluating systems\nin challenging settings, we propose a generative framework to create\nsafety-critical scenarios for evaluating specific task algorithms. We first\nrepresent the traffic scenarios with a series of autoregressive building blocks\nand generate diverse scenarios by sampling from the joint distribution of these\nblocks. We then train the generative model as an agent (or a generator) to\ninvestigate the risky distribution parameters for a given driving algorithm\nbeing evaluated. We regard the task algorithm as an environment (or a\ndiscriminator) that returns a reward to the agent when a risky scenario is\ngenerated. Through the experiments conducted on several scenarios in the\nsimulation, we demonstrate that the proposed framework generates\nsafety-critical scenarios more efficiently than grid search or human design\nmethods. Another advantage of this method is its adaptiveness to the routes and\nparameters.', 'This work contributes to the development of neural forecasting models with\nnovel randomization-based learning methods. These methods improve the fitting\nabilities of the neural model, in comparison to the standard method, by\ngenerating network parameters in accordance with the data and target function\nfeatures. A pattern-based representation of time series makes the proposed\napproach useful for forecasting time series with multiple seasonality. In the\nsimulation study, we evaluate the performance of the proposed models and find\nthat they can compete in terms of forecasting accuracy with fully-trained\nnetworks. Extremely fast and easy training, simple architecture, ease of\nimplementation, high accuracy as well as dealing with nonstationarity and\nmultiple seasonality in time series make the proposed model very attractive for\na wide range of complex time series forecasting problems.', 'This paper presents a time series forecasting framework which combines\nstandard forecasting methods and a machine learning model. The inputs to the\nmachine learning model are not lagged values or regular time series features,\nbut instead forecasts produced by standard methods. The machine learning model\ncan be either a convolutional neural network model or a recurrent neural\nnetwork model. The intuition behind this approach is that forecasts of a time\nseries are themselves good features characterizing the series, especially when\nthe modelling purpose is forecasting. It can also be viewed as a weighted\nensemble method. Tested on the M4 competition dataset, this approach\noutperforms all submissions for quarterly series, and is more accurate than all\nbut the winning algorithm for monthly series.', 'Empirically observed time series in physics, biology, or medicine, are\ncommonly generated by some underlying dynamical system (DS) which is the target\nof scientific interest. There is an increasing interest to harvest machine\nlearning methods to reconstruct this latent DS in a data-driven, unsupervised\nway. In many areas of science it is common to sample time series observations\nfrom many data modalities simultaneously, e.g. electrophysiological and\nbehavioral time series in a typical neuroscience experiment. However, current\nmachine learning tools for reconstructing DSs usually focus on just one data\nmodality. Here we propose a general framework for multi-modal data integration\nfor the purpose of nonlinear DS reconstruction and the analysis of cross-modal\nrelations. This framework is based on dynamically interpretable recurrent\nneural networks as general approximators of nonlinear DSs, coupled to sets of\nmodality-specific decoder models from the class of generalized linear models.\nBoth an expectation-maximization and a variational inference algorithm for\nmodel training are advanced and compared. We show on nonlinear DS benchmarks\nthat our algorithms can efficiently compensate for too noisy or missing\ninformation in one data channel by exploiting other channels, and demonstrate\non experimental neuroscience data how the algorithm learns to link different\ndata domains to the underlying dynamics.', 'In recent years, deep learning techniques have outperformed traditional\nmodels in many machine learning tasks. Deep neural networks have successfully\nbeen applied to address time series forecasting problems, which is a very\nimportant topic in data mining. They have proved to be an effective solution\ngiven their capacity to automatically learn the temporal dependencies present\nin time series. However, selecting the most convenient type of deep neural\nnetwork and its parametrization is a complex task that requires considerable\nexpertise. Therefore, there is a need for deeper studies on the suitability of\nall existing architectures for different forecasting tasks. In this work, we\nface two main challenges: a comprehensive review of the latest works using deep\nlearning for time series forecasting; and an experimental study comparing the\nperformance of the most popular architectures. The comparison involves a\nthorough analysis of seven types of deep learning models in terms of accuracy\nand efficiency. We evaluate the rankings and distribution of results obtained\nwith the proposed models under many different architecture configurations and\ntraining hyperparameters. The datasets used comprise more than 50000 time\nseries divided into 12 different forecasting problems. By training more than\n38000 models on these data, we provide the most extensive deep learning study\nfor time series forecasting. Among all studied models, the results show that\nlong short-term memory (LSTM) and convolutional networks (CNN) are the best\nalternatives, with LSTMs obtaining the most accurate forecasts. CNNs achieve\ncomparable performance with less variability of results under different\nparameter configurations, while also being more efficient.', 'In recent times, deep artificial neural networks have achieved many successes\nin pattern recognition. Part of this success can be attributed to the reliance\non big data to increase generalization. However, in the field of time series\nrecognition, many datasets are often very small. One method of addressing this\nproblem is through the use of data augmentation. In this paper, we survey data\naugmentation techniques for time series and their application to time series\nclassification with neural networks. We propose a taxonomy and outline the four\nfamilies in time series data augmentation, including transformation-based\nmethods, pattern mixing, generative models, and decomposition methods.\nFurthermore, we empirically evaluate 12 time series data augmentation methods\non 128 time series classification datasets with six different types of neural\nnetworks. Through the results, we are able to analyze the characteristics,\nadvantages and disadvantages, and recommendations of each data augmentation\nmethod. This survey aims to help in the selection of time series data\naugmentation for neural network applications.', 'With the wider availability of sensor technology, a number of Structural\nHealth Monitoring (SHM) systems are deployed to monitor civil infrastructure.\nThe continuous monitoring provides valuable information about the structure\nthat can help in providing a decision support system for retrofits and other\nstructural modifications. However, when the sensors are exposed to harsh\nenvironmental conditions, the data measured by the SHM systems tend to be\naffected by multiple anomalies caused by faulty or broken sensors. Given a\ndeluge of high-dimensional data collected continuously over time, research into\nusing machine learning methods to detect anomalies are a topic of great\ninterest to the SHM community. This paper contributes to this effort by\nproposing the use of a relatively new time series representation named Shapelet\nTransform in combination with a Random Forest classifier to autonomously\nidentify anomalies in SHM data. The shapelet transform is a unique time series\nrepresentation that is solely based on the shape of the time series data. In\nconsideration of the individual characteristics unique to every anomaly, the\napplication of this transform yields a new shape-based feature representation\nthat can be combined with any standard machine learning algorithm to detect\nanomalous data with no manual intervention. For the present study, the anomaly\ndetection framework consists of three steps: identifying unique shapes from\nanomalous data, using these shapes to transform the SHM data into a local-shape\nspace and training machine learning algorithm on this transformed data to\nidentify anomalies. The efficacy of this method is demonstrated by the\nidentification of anomalies in acceleration data from a SHM system installed on\na long-span bridge in China. The results show that multiple data anomalies in\nSHM data can be automatically detected with high accuracy using the proposed\nmethod.', 'A new variational mode decomposition (VMD) based deep learning approach is\nproposed in this paper for time series forecasting problem. Firstly, VMD is\nadopted to decompose the original time series into several sub-signals. Then, a\nconvolutional neural network (CNN) is applied to learn the reconstruction\npatterns on the decomposed sub-signals to obtain several reconstructed\nsub-signals. Finally, a long short term memory (LSTM) network is employed to\nforecast the time series with the decomposed sub-signals and the reconstructed\nsub-signals as inputs. The proposed VMD-CNN-LSTM approach is originated from\nthe decomposition-reconstruction-ensemble framework, and innovated by embedding\nthe reconstruction, single forecasting, and ensemble steps in a unified deep\nlearning approach. To verify the forecasting performance of the proposed\napproach, four typical time series datasets are introduced for empirical\nanalysis. The empirical results demonstrate that the proposed approach\noutperforms consistently the benchmark approaches in terms of forecasting\naccuracy, and also indicate that the reconstructed sub-signals obtained by CNN\nis of importance for further improving the forecasting performance.', 'Forecasting models that are trained across sets of many time series, known as\nGlobal Forecasting Models (GFM), have shown recently promising results in\nforecasting competitions and real-world applications, outperforming many\nstate-of-the-art univariate forecasting techniques. In most cases, GFMs are\nimplemented using deep neural networks, and in particular Recurrent Neural\nNetworks (RNN), which require a sufficient amount of time series to estimate\ntheir numerous model parameters. However, many time series databases have only\na limited number of time series. In this study, we propose a novel, data\naugmentation based forecasting framework that is capable of improving the\nbaseline accuracy of the GFM models in less data-abundant settings. We use\nthree time series augmentation techniques: GRATIS, moving block bootstrap\n(MBB), and dynamic time warping barycentric averaging (DBA) to synthetically\ngenerate a collection of time series. The knowledge acquired from these\naugmented time series is then transferred to the original dataset using two\ndifferent approaches: the pooled approach and the transfer learning approach.\nWhen building GFMs, in the pooled approach, we train a model on the augmented\ntime series alongside the original time series dataset, whereas in the transfer\nlearning approach, we adapt a pre-trained model to the new dataset. In our\nevaluation on competition and real-world time series datasets, our proposed\nvariants can significantly improve the baseline accuracy of GFM models and\noutperform state-of-the-art univariate forecasting methods.', 'We introduce Neural Contextual Anomaly Detection (NCAD), a framework for\nanomaly detection on time series that scales seamlessly from the unsupervised\nto supervised setting, and is applicable to both univariate and multivariate\ntime series. This is achieved by effectively combining recent developments in\nrepresentation learning for multivariate time series, with techniques for deep\nanomaly detection originally developed for computer vision that we tailor to\nthe time series setting. Our window-based approach facilitates learning the\nboundary between normal and anomalous classes by injecting generic synthetic\nanomalies into the available data. Moreover, our method can effectively take\nadvantage of all the available information, be it as domain knowledge, or as\ntraining labels in the semi-supervised setting. We demonstrate empirically on\nstandard benchmark datasets that our approach obtains a state-of-the-art\nperformance in these settings.', ""Forecasting on sparse multivariate time series (MTS) aims to model the\npredictors of future values of time series given their incomplete past, which\nis important for many emerging applications. However, most existing methods\nprocess MTS's individually, and do not leverage the dynamic distributions\nunderlying the MTS's, leading to sub-optimal results when the sparsity is high.\nTo address this challenge, we propose a novel generative model, which tracks\nthe transition of latent clusters, instead of isolated feature representations,\nto achieve robust modeling. It is characterized by a newly designed dynamic\nGaussian mixture distribution, which captures the dynamics of clustering\nstructures, and is used for emitting timeseries. The generative model is\nparameterized by neural networks. A structured inference network is also\ndesigned for enabling inductive analysis. A gating mechanism is further\nintroduced to dynamically tune the Gaussian mixture distributions. Extensive\nexperimental results on a variety of real-life datasets demonstrate the\neffectiveness of our method."", 'Most current anomaly detection methods suffer from the curse of\ndimensionality when dealing with high-dimensional data. We propose an anomaly\ndetection algorithm that can scale to high-dimensional data using concepts from\nthe theory of large deviations. The proposed Large Deviations Anomaly Detection\n(LAD) algorithm is shown to outperform state of art anomaly detection methods\non a variety of large and high-dimensional benchmark data sets. Exploiting the\nability of the algorithm to scale to high-dimensional data, we propose an\nonline anomaly detection method to identify anomalies in a collection of\nmultivariate time series. We demonstrate the applicability of the online\nalgorithm in identifying counties in the United States with anomalous trends in\nterms of COVID-19 related cases and deaths. Several of the identified anomalous\ncounties correlate with counties with documented poor response to the COVID\npandemic.', 'Recently, with the development of deep learning, end-to-end neural network\narchitectures have been increasingly applied to condition monitoring signals.\nThey have demonstrated superior performance for fault detection and\nclassification, in particular using convolutional neural networks. Even more\nrecently, an extension of the concept of convolution to the concept of\nkervolution has been proposed with some promising results in image\nclassification tasks. In this paper, we explore the potential of kervolutional\nneural networks applied to time series data. We demonstrate that using a\nmixture of convolutional and kervolutional layers improves the model\nperformance. The mixed model is first applied to a classification task in time\nseries, as a benchmark dataset. Subsequently, the proposed mixed architecture\nis used to detect anomalies in time series data recorded by accelerometers on\nhelicopters. We propose a residual-based anomaly detection approach using a\ntemporal auto-encoder. We demonstrate that mixing kervolutional with\nconvolutional layers in the encoder is more sensitive to variations in the\ninput data and is able to detect anomalous time series in a better way.', 'A challenging part of dynamic probabilistic risk assessment for nuclear power\nplants is the need for large amounts of temporal simulations given various\ninitiating events and branching conditions from which representative feature\nextraction becomes complicated for subsequent applications. Artificial\nIntelligence techniques have been shown to be powerful tools in time-dependent\nsequential data processing to automatically extract and yield complex features\nfrom large data. An advanced temporal neural network referred to as the\nTransformer is used within a supervised learning fashion to model the\ntime-dependent NPP simulation data and to infer whether a given sequence of\nevents leads to core damage or not. The training and testing datasets for the\nTransformer are obtained by running 10,000 RELAP5-3D NPP blackout simulations\nwith the list of variables obtained from the RAVEN software. Each simulation is\nclassified as ""OK"" or ""CORE DAMAGE"" based on the consequence. The results show\nthat the Transformer can learn the characteristics of the sequential data and\nyield promising performance with approximately 99% classification accuracy on\nthe testing dataset.', 'Many problems in science and engineering require making predictions based on\nfew observations. To build a robust predictive model, these sparse data may\nneed to be augmented with simulated data, especially when the design space is\nmulti-dimensional. Simulations, however, often suffer from an inherent bias.\nEstimation of this bias may be poorly constrained not only because of data\nsparsity, but also because traditional predictive models fit only one type of\nobserved outputs, such as scalars or images, instead of all available output\ndata modalities, which might have been acquired and simulated at great cost. To\nbreak this limitation and open up the path for multi-modal calibration, we\npropose to combine a novel, transfer learning technique for suppressing the\nbias with recent developments in deep learning, which allow building predictive\nmodels with multi-modal outputs. First, we train an initial neural network\nmodel on simulated data to learn important correlations between different\noutput modalities and between simulation inputs and outputs. Then, the model is\npartially retrained, or transfer learned, to fit the experiments; a method that\nhas never been implemented in this type of architecture. Using fewer than 10\ninertial confinement fusion experiments for training, transfer learning\nsystematically improves the simulation predictions while a simple output\ncalibration, which we design as a baseline, makes the predictions worse. We\nalso offer extensive cross-validation with real and carefully designed\nsynthetic data. The method described in this paper can be applied to a wide\nrange of problems that require transferring knowledge from simulations to the\ndomain of experiments.', 'Having access to a forward model enables the use of planning algorithms such\nas Monte Carlo Tree Search and Rolling Horizon Evolution. Where a model is\nunavailable, a natural aim is to learn a model that reflects accurately the\ndynamics of the environment. In many situations it might not be possible and\nminimal glitches in the model may lead to poor performance and failure. This\npaper explores the problem of model misspecification through uncertainty-aware\nreinforcement learning agents. We propose a bootstrapped multi-headed neural\nnetwork that learns the distribution of future states and rewards. We\nexperiment with a number of schemes to extract the most likely predictions.\nMoreover, we also introduce a global error correction filter that applies\nhigh-level constraints guided by the context provided through the predictive\ndistribution. We illustrate our approach on Minipacman. The evaluation\ndemonstrates that when dealing with imperfect models, our methods exhibit\nincreased performance and stability, both in terms of model accuracy and in its\nuse within a planning algorithm.', 'Irregularly sampled time series data arise naturally in many application\ndomains including biology, ecology, climate science, astronomy, and health.\nSuch data represent fundamental challenges to many classical models from\nmachine learning and statistics due to the presence of non-uniform intervals\nbetween observations. However, there has been significant progress within the\nmachine learning community over the last decade on developing specialized\nmodels and architectures for learning from irregularly sampled univariate and\nmultivariate time series data. In this survey, we first describe several axes\nalong which approaches to learning from irregularly sampled time series differ\nincluding what data representations they are based on, what modeling primitives\nthey leverage to deal with the fundamental problem of irregular sampling, and\nwhat inference tasks they are designed to perform. We then survey the recent\nliterature organized primarily along the axis of modeling primitives. We\ndescribe approaches based on temporal discretization, interpolation,\nrecurrence, attention and structural invariance. We discuss similarities and\ndifferences between approaches and highlight primary strengths and weaknesses.', 'Deep probabilistic forecasting techniques have recently been proposed for\nmodeling large collections of time-series. However, these techniques explicitly\nassume either complete independence (local model) or complete dependence\n(global model) between time-series in the collection. This corresponds to the\ntwo extreme cases where every time-series is disconnected from every other\ntime-series in the collection or likewise, that every time-series is related to\nevery other time-series resulting in a completely connected graph. In this\nwork, we propose a deep hybrid probabilistic graph-based forecasting framework\ncalled Graph Deep Factors (GraphDF) that goes beyond these two extremes by\nallowing nodes and their time-series to be connected to others in an arbitrary\nfashion. GraphDF is a hybrid forecasting framework that consists of a\nrelational global and relational local model. In particular, we propose a\nrelational global model that learns complex non-linear time-series patterns\nglobally using the structure of the graph to improve both forecasting accuracy\nand computational efficiency. Similarly, instead of modeling every time-series\nindependently, we learn a relational local model that not only considers its\nindividual time-series but also the time-series of nodes that are connected in\nthe graph. The experiments demonstrate the effectiveness of the proposed deep\nhybrid graph-based forecasting model compared to the state-of-the-art methods\nin terms of its forecasting accuracy, runtime, and scalability. Our case study\nreveals that GraphDF can successfully generate cloud usage forecasts and\nopportunistically schedule workloads to increase cloud cluster utilization by\n47.5% on average.', 'Ensemble models are powerful model building tools that are developed with a\nfocus to improve the accuracy of model predictions. They find applications in\ntime series forecasting in varied scenarios including but not limited to\nprocess industries, health care, and economics where a single model might not\nprovide optimal performance. It is known that if models selected for data\nmodelling are distinct (linear/non-linear, static/dynamic) and independent\n(minimally correlated models), the accuracy of the predictions is improved.\nVarious approaches suggested in the literature to weigh the ensemble models use\na static set of weights. Due to this limitation, approaches using a static set\nof weights for weighing ensemble models cannot capture the dynamic changes or\nlocal features of the data effectively. To address this issue, a Reinforcement\nLearning (RL) approach to dynamically assign and update weights of each of the\nmodels at different time instants depending on the nature of data and the\nindividual model predictions is proposed in this work. The RL method\nimplemented online, essentially learns to update the weights and reduce the\nerrors as the time progresses. Simulation studies on time series data showed\nthat the dynamic weighted approach using RL learns the weight better than\nexisting approaches. The accuracy of the proposed method is compared with an\nexisting approach of online Neural Network tuning quantitatively through\nnormalized mean square error(NMSE) values.', 'Forecasting financial time series is considered to be a difficult task due to\nthe chaotic feature of the series. Statistical approaches have shown solid\nresults in some specific problems such as predicting market direction and\nsingle-price of stocks; however, with the recent advances in deep learning and\nbig data techniques, new promising options have arises to tackle financial time\nseries forecasting. Moreover, recent literature has shown that employing a\ncombination of statistics and machine learning may improve accuracy in the\nforecasts in comparison to single solutions. Taking into consideration the\nmentioned aspects, in this work, we proposed the MegazordNet, a framework that\nexplores statistical features within a financial series combined with a\nstructured deep learning model for time series forecasting. We evaluated our\napproach predicting the closing price of stocks in the S&P 500 using different\nmetrics, and we were able to beat single statistical and machine learning\nmethods.', 'This work presents an introduction to feature-based time-series analysis. The\ntime series as a data type is first described, along with an overview of the\ninterdisciplinary time-series analysis literature. I then summarize the range\nof feature-based representations for time series that have been developed to\naid interpretable insights into time-series structure. Particular emphasis is\ngiven to emerging research that facilitates wide comparison of feature-based\nrepresentations that allow us to understand the properties of a time-series\ndataset that make it suited to a particular feature-based representation or\nanalysis algorithm. The future of time-series analysis is likely to embrace\napproaches that exploit machine learning methods to partially automate human\nlearning to aid understanding of the complex dynamical patterns in the time\nseries we measure from the world.', ""With the increase in the learning capability of deep convolution-based\narchitectures, various applications of such models have been proposed over\ntime. In the field of anomaly detection, improvements in deep learning opened\nnew prospects of exploration for the researchers whom tried to automate the\nlabor-intensive features of data collection. First, in terms of data\ncollection, it is impossible to anticipate all the anomalies that might exist\nin a given environment. Second, assuming we limit the possibilities of\nanomalies, it will still be hard to record all these scenarios for the sake of\ntraining a model. Third, even if we manage to record a significant amount of\nabnormal data, it's laborious to annotate this data on pixel or even frame\nlevel. Various approaches address the problem by proposing one-class\nclassification using generative models trained on only normal data. In such\nmethods, only the normal data is used, which is abundantly available and\ndoesn't require significant human input. However, these are trained with only\nnormal data and at the test time, given abnormal data as input, may often\ngenerate normal-looking output. This happens due to the hallucination\ncharacteristic of generative models. Next, these systems are designed to not\nuse abnormal examples during the training. In this paper, we propose anomaly\ndetection with negative learning (ADNL), which employs the negative learning\nconcept for the enhancement of anomaly detection by utilizing a very small\nnumber of labeled anomaly data as compared with the normal data during\ntraining. The idea is to limit the reconstruction capability of a generative\nmodel using the given a small amount of anomaly examples. This way, the network\nnot only learns to reconstruct normal data but also encloses the normal\ndistribution far from the possible distribution of anomalies."", ""A real-world dataset is provided from a pulp-and-paper manufacturing\nindustry. The dataset comes from a multivariate time series process. The data\ncontains a rare event of paper break that commonly occurs in the industry. The\ndata contains sensor readings at regular time-intervals (x's) and the event\nlabel (y). The primary purpose of the data is thought to be building a\nclassification model for early prediction of the rare event. However, it can\nalso be used for multivariate time series data exploration and building other\nsupervised and unsupervised models."", 'We address a three-tier numerical framework based on manifold learning for\nthe forecasting of high-dimensional time series. At the first step, we embed\nthe time series into a reduced low-dimensional space using a nonlinear manifold\nlearning algorithm such as Locally Linear Embedding and Diffusion Maps. At the\nsecond step, we construct reduced-order regression models on the manifold, in\nparticular Multivariate Autoregressive (MVAR) and Gaussian Process Regression\n(GPR) models, to forecast the embedded dynamics. At the final step, we lift the\nembedded time series back to the original high-dimensional space using Radial\nBasis Functions interpolation and Geometric Harmonics. For our illustrations,\nwe test the forecasting performance of the proposed numerical scheme with four\nsets of time series: three synthetic stochastic ones resembling EEG signals\nproduced from linear and nonlinear stochastic models with different model\norders, and one real-world data set containing daily time series of 10 key\nforeign exchange rates (FOREX) spanning the time period 03/09/2001-29/10/2020.\nThe forecasting performance of the proposed numerical scheme is assessed using\nthe combinations of manifold learning, modelling and lifting approaches. We\nalso provide a comparison with the Principal Component Analysis algorithm as\nwell as with the naive random walk model and the MVAR and GPR models trained\nand implemented directly in the high-dimensional space.', 'This paper addresses the problem of multi-step time series forecasting for\nnon-stationary signals that can present sudden changes. Current\nstate-of-the-art deep learning forecasting methods, often trained with variants\nof the MSE, lack the ability to provide sharp predictions in deterministic and\nprobabilistic contexts. To handle these challenges, we propose to incorporate\nshape and temporal criteria in the training objective of deep models. We define\nshape and temporal similarities and dissimilarities, based on a smooth\nrelaxation of Dynamic Time Warping (DTW) and Temporal Distortion Index (TDI),\nthat enable to build differentiable loss functions and positive semi-definite\n(PSD) kernels. With these tools, we introduce DILATE (DIstortion Loss including\nshApe and TimE), a new objective for deterministic forecasting, that explicitly\nincorporates two terms supporting precise shape and temporal change detection.\nFor probabilistic forecasting, we introduce STRIPE++ (Shape and Time diverRsIty\nin Probabilistic forEcasting), a framework for providing a set of sharp and\ndiverse forecasts, where the structured shape and time diversity is enforced\nwith a determinantal point process (DPP) diversity loss. Extensive experiments\nand ablations studies on synthetic and real-world datasets confirm the benefits\nof leveraging shape and time features in time series forecasting.', 'Time series data is ubiquitous in the real-world problems across various\ndomains including healthcare, social media, and crime surveillance. Detecting\nanomalies, or irregular and rare events, in time series data, can enable us to\nfind abnormal events in any natural phenomena, which may require special\ntreatment. Moreover, labeled instances of anomaly are hard to get in time\nseries data. On the other hand, time series data, due to its nature, often\nexhibits localized expansions and compressions in the time dimension which is\ncalled warping. These two challenges make it hard to detect anomalies in time\nseries as often such warpings could get detected as anomalies erroneously. Our\nobjective is to build an anomaly detection model that is robust to such warping\nvariations. In this paper, we propose a novel unsupervised time series anomaly\ndetection method, WaRTEm-AD, that operates in two stages. Within the key stage\nof representation learning, we employ data augmentation through bespoke time\nseries operators which are passed through a twin autoencoder architecture to\nlearn warping-robust representations for time series data. Second, adaptations\nof state-of-the-art anomaly detection methods are employed on the learnt\nrepresentations to identify anomalies. We will illustrate that WaRTEm-AD is\ndesigned to detect two types of time series anomalies: point and sequence\nanomalies. We compare WaRTEm-AD with the state-of-the-art baselines and\nestablish the effectiveness of our method both in terms of anomaly detection\nperformance and computational efficiency.', 'Analysis of water and environmental data is an important aspect of many\nintelligent water and environmental system applications where inference from\nsuch analysis plays a significant role in decision making. Quite often these\ndata that are collected through sensible sensors can be anomalous due to\ndifferent reasons such as systems breakdown, malfunctioning of sensor\ndetectors, and more. Regardless of their root causes, such data severely affect\nthe results of the subsequent analysis. This paper demonstrates data cleaning\nand preparation for time-series data and further proposes cost-sensitive\nmachine learning algorithms as a solution to detect anomalous data points in\ntime-series data. The following models: Logistic Regression, Random Forest,\nSupport Vector Machines have been modified to support the cost-sensitive\nlearning which penalizes misclassified samples thereby minimizing the total\nmisclassification cost. Our results showed that Random Forest outperformed the\nrest of the models at predicting the positive class (i.e anomalies). Applying\npredictive model improvement techniques like data oversampling seems to provide\nlittle or no improvement to the Random Forest model. Interestingly, with\nrecursive feature elimination, we achieved a better model performance thereby\nreducing the dimensions in the data. Finally, with Influxdb and Kapacitor the\ndata was ingested and streamed to generate new data points to further evaluate\nthe model performance on unseen data, this will allow for early recognition of\nundesirable changes in the drinking water quality and will enable the water\nsupply companies to rectify on a timely basis whatever undesirable changes\nabound.', 'While classical time series forecasting considers individual time series in\nisolation, recent advances based on deep learning showed that jointly learning\nfrom a large pool of related time series can boost the forecasting accuracy.\nHowever, the accuracy of these methods suffers greatly when modeling\nout-of-sample time series, significantly limiting their applicability compared\nto classical forecasting methods. To bridge this gap, we adopt a meta-learning\nview of the time series forecasting problem. We introduce a novel forecasting\nmethod, called Meta Global-Local Auto-Regression (Meta-GLAR), that adapts to\neach time series by learning in closed-form the mapping from the\nrepresentations produced by a recurrent neural network (RNN) to one-step-ahead\nforecasts. Crucially, the parameters ofthe RNN are learned across multiple time\nseries by backpropagating through the closed-form adaptation mechanism. In our\nextensive empirical evaluation we show that our method is competitive with the\nstate-of-the-art in out-of-sample forecasting accuracy reported in earlier\nwork.', 'In complex industrial settings, it is common practice to monitor the\noperation of machines in order to detect undesired states, adjust maintenance\nschedules, optimize system performance or collect usage statistics of\nindividual machines. In this work, we focus on estimating the power output of a\nCombined Heat and Power (CHP) machine of a medium-sized company facility by\nanalyzing the total facility power consumption. We formulate the problem as a\ntime-series classification problem where the class label represents the CHP\npower output. As the facility is fully instrumented and sensor measurements\nfrom the CHP are available, we generate the training labels in an automated\nfashion from the CHP sensor readings. However, sensor failures result in\nmislabeled training data samples which are hard to detect and remove from the\ndataset. Therefore, we propose a novel multi-task deep learning approach that\njointly trains a classifier and an autoencoder with a shared embedding\nrepresentation. The proposed approach targets to gradually correct the\nmislabelled data samples during training in a self-supervised fashion, without\nany prior assumption on the amount of label noise. We benchmark our approach on\nseveral time-series classification datasets and find it to be comparable and\nsometimes better than state-of-the-art methods. On the real-world use-case of\npredicting the CHP power output, we thoroughly evaluate the architectural\ndesign choices and show that the final architecture considerably increases the\nrobustness of the learning process and consistently beats other recent\nstate-of-the-art algorithms in the presence of unstructured as well as\nstructured label noise.', 'The monitoring and management of numerous and diverse time series data at\nAlibaba Group calls for an effective and scalable time series anomaly detection\nservice. In this paper, we propose RobustTAD, a Robust Time series Anomaly\nDetection framework by integrating robust seasonal-trend decomposition and\nconvolutional neural network for time series data. The seasonal-trend\ndecomposition can effectively handle complicated patterns in time series, and\nmeanwhile significantly simplifies the architecture of the neural network,\nwhich is an encoder-decoder architecture with skip connections. This\narchitecture can effectively capture the multi-scale information from time\nseries, which is very useful in anomaly detection. Due to the limited labeled\ndata in time series anomaly detection, we systematically investigate data\naugmentation methods in both time and frequency domains. We also introduce\nlabel-based weight and value-based weight in the loss function by utilizing the\nunbalanced nature of the time series anomaly detection problem. Compared with\nthe widely used forecasting-based anomaly detection algorithms,\ndecomposition-based algorithms, traditional statistical algorithms, as well as\nrecent neural network based algorithms, RobustTAD performs significantly better\non public benchmark datasets. It is deployed as a public online service and\nwidely adopted in different business scenarios at Alibaba Group.', 'Anomaly detection has been a challenging task given high-dimensional\nmultivariate time series data generated by networked sensors and actuators in\nCyber-Physical Systems (CPS). Besides the highly nonlinear, complex, and\ndynamic natures of such time series, the lack of labeled data impedes data\nexploitation in a supervised manner and thus prevents an accurate detection of\nabnormal phenomenons. On the other hand, the collected data at the edge of the\nnetwork is often privacy sensitive and large in quantity, which may hinder the\ncentralized training at the main server. To tackle these issues, we propose an\nunsupervised time series anomaly detection framework in a federated fashion to\ncontinuously monitor the behaviors of interconnected devices within a network\nand alerts for abnormal incidents so that countermeasures can be taken before\nundesired consequences occur. To be specific, we leave the training data\ndistributed at the edge to learn a shared Variational Autoencoder (VAE) based\non Convolutional Gated Recurrent Unit (ConvGRU) model, which jointly captures\nfeature and temporal dependencies in the multivariate time series data for\nrepresentation learning and downstream anomaly detection tasks. Experiments on\nthree real-world networked sensor datasets illustrate the advantage of our\napproach over other state-of-the-art models. We also conduct extensive\nexperiments to demonstrate the effectiveness of our detection framework under\nnon-federated and federated settings in terms of overall performance and\ndetection latency.', 'Cyber-physical systems (CPS) greatly benefit by using machine learning\ncomponents that can handle the uncertainty and variability of the real-world.\nTypical components such as deep neural networks, however, introduce new types\nof hazards that may impact system safety. The system behavior depends on data\nthat are available only during runtime and may be different than the data used\nfor training. Out-of-distribution data may lead to a large error and compromise\nsafety. The paper considers the problem of efficiently detecting\nout-of-distribution data in CPS control systems. Detection must be robust and\nlimit the number of false alarms while being computational efficient for\nreal-time monitoring. The proposed approach leverages inductive conformal\nprediction and anomaly detection for developing a method that has a\nwell-calibrated false alarm rate. We use variational autoencoders and deep\nsupport vector data description to learn models that can be used efficiently\ncompute the nonconformity of new inputs relative to the training set and enable\nreal-time detection of out-of-distribution high-dimensional inputs. We\ndemonstrate the method using an advanced emergency braking system and a\nself-driving end-to-end controller implemented in an open source simulator for\nself-driving cars. The simulation results show very small number of false\npositives and detection delay while the execution time is comparable to the\nexecution time of the original machine learning components.', 'A robust model for time series forecasting is highly important in many\ndomains, including but not limited to financial forecast, air temperature and\nelectricity consumption. To improve forecasting performance, traditional\napproaches usually require additional feature sets. However, adding more\nfeature sets from different sources of data is not always feasible due to its\naccessibility limitation. In this paper, we propose a novel self-boosted\nmechanism in which the original time series is decomposed into multiple time\nseries. These time series played the role of additional features in which the\nclosely related time series group is used to feed into multi-task learning\nmodel, and the loosely related group is fed into multi-view learning part to\nutilize its complementary information. We use three real-world datasets to\nvalidate our model and show the superiority of our proposed method over\nexisting state-of-the-art baseline methods.', 'This paper introduces a new methodology for detecting anomalies in time\nseries data, with a primary application to monitoring the health of (micro-)\nservices and cloud resources. The main novelty in our approach is that instead\nof modeling time series consisting of real values or vectors of real values, we\nmodel time series of probability distributions over real values (or vectors).\nThis extension to time series of probability distributions allows the technique\nto be applied to the common scenario where the data is generated by requests\ncoming in to a service, which is then aggregated at a fixed temporal frequency.\nOur method is amenable to streaming anomaly detection and scales to monitoring\nfor anomalies on millions of time series. We show the superior accuracy of our\nmethod on synthetic and public real-world data. On the Yahoo Webscope data set,\nwe outperform the state of the art in 3 out of 4 data sets and we show that we\noutperform popular open-source anomaly detection tools by up to 17% average\nimprovement for a real-world data set.', 'Accurate and fast simulation of particle physics processes is crucial for the\nhigh-energy physics community. Simulating particle interactions with detectors\nis both time consuming and computationally expensive. With the proton-proton\ncollision energy of 13 TeV, the Large Hadron Collider is uniquely positioned to\ndetect and measure the rare phenomena that can shape our knowledge of new\ninteractions. The High-Luminosity Large Hadron Collider (HL-LHC) upgrade will\nput a significant strain on the computing infrastructure due to increased event\nrate and levels of pile-up. Simulation of high-energy physics collisions needs\nto be significantly faster without sacrificing the physics accuracy. Machine\nlearning approaches can offer faster solutions, while maintaining a high level\nof fidelity. We discuss a graph generative model that provides effective\nreconstruction of LHC events, paving the way for full detector level fast\nsimulation for HL-LHC.', 'Time-series forecasting is important for many applications. Forecasting\nmodels are usually trained using time-series data in a specific target task.\nHowever, sufficient data in the target task might be unavailable, which leads\nto performance degradation. In this paper, we propose a few-shot learning\nmethod that forecasts a future value of a time-series in a target task given a\nfew time-series in the target task. Our model is trained using time-series data\nin multiple training tasks that are different from target tasks. Our model uses\na few time-series to build a forecasting function based on a recurrent neural\nnetwork with an attention mechanism. With the attention mechanism, we can\nretrieve useful patterns in a small number of time-series for the current\nsituation. Our model is trained by minimizing an expected test error of\nforecasting next timestep values. We demonstrate the effectiveness of the\nproposed method using 90 time-series datasets.', 'In this work we propose for the first time a transformer-based framework for\nunsupervised representation learning of multivariate time series. Pre-trained\nmodels can be potentially used for downstream tasks such as regression and\nclassification, forecasting and missing value imputation. By evaluating our\nmodels on several benchmark datasets for multivariate time series regression\nand classification, we show that not only does our modeling approach represent\nthe most successful method employing unsupervised learning of multivariate time\nseries presented to date, but also that it exceeds the current state-of-the-art\nperformance of supervised methods; it does so even when the number of training\nsamples is very limited, while offering computational efficiency. Finally, we\ndemonstrate that unsupervised pre-training of our transformer models offers a\nsubstantial performance benefit over fully supervised learning, even without\nleveraging additional unlabeled data, i.e., by reusing the same data samples\nthrough the unsupervised objective.', 'Event time models predict occurrence times of an event of interest based on\nknown features. Recent work has demonstrated that neural networks achieve\nstate-of-the-art event time predictions in a variety of settings. However,\nstandard event time models suppose that the event occurs, eventually, in all\ncases. Consequently, no distinction is made between a) the probability of event\noccurrence, and b) the predicted time of occurrence. This distinction is\ncritical when predicting medical diagnoses, equipment defects, social media\nposts, and other events that or may not occur, and for which the features\naffecting a) may be different from those affecting b). In this work, we develop\na conditional event time model that distinguishes between these components,\nimplement it as a neural network with a binary stochastic layer representing\nfinite event occurrence, and show how it may be learned from right-censored\nevent times via maximum likelihood estimation. Results demonstrate superior\nevent occurrence and event time predictions on synthetic data, medical events\n(MIMIC-III), and social media posts (Reddit), comprising 21 total prediction\ntasks.', 'Detecting earthquake events from seismic time series has proved itself a\nchallenging task. Manual detection can be expensive and tedious due to the\nintensive labor and large scale data set. In recent years, automatic detection\nmethods based on machine learning have been developed to improve accuracy and\nefficiency. However, the accuracy of those methods relies on a sufficient\namount of high-quality training data, which itself can be expensive to obtain\ndue to the requirement of domain knowledge and subject matter expertise. This\npaper is to resolve this dilemma by answering two questions: (1) provided with\na limited number of reliable labels, can we use them to generate more synthetic\nlabels; (2) Can we use those synthetic labels to improve the detectability?\nAmong all the existing generative models, the generative adversarial network\n(GAN) shows its supreme capability in generating high-quality synthetic samples\nin multiple domains. We designed our model based on GAN. In particular, we\nstudied several different network structures. By comparing the generated\nresults, our GAN-based generative model yields the highest quality. We further\ncombine the dataset with synthetic samples generated by our generative model\nand show that the detectability of our earthquake classification model is\nsignificantly improved than the one trained without augmenting the training\nset.', 'Most real-world datasets, and particularly those collected from physical\nsystems, are full of noise, packet loss, and other imperfections. However, most\nspecification mining, anomaly detection and other such algorithms assume, or\neven require, perfect data quality to function properly. Such algorithms may\nwork in lab conditions when given clean, controlled data, but will fail in the\nfield when given imperfect data. We propose a method for accurately\nreconstructing discrete temporal or sequential system traces affected by data\nloss, using Long Short-Term Memory Networks (LSTMs). The model works by\nlearning to predict the next event in a sequence of events, and uses its own\noutput as an input to continue predicting future events. As a result, this\nmethod can be used for data restoration even with streamed data. Such a method\ncan reconstruct even long sequence of missing events, and can also help\nvalidate and improve data quality for noisy data. The output of the model will\nbe a close reconstruction of the true data, and can be fed to algorithms that\nrely on clean data. We demonstrate our method by reconstructing automotive CAN\ntraces consisting of long sequences of discrete events. We show that given even\nsmall parts of a CAN trace, our LSTM model can predict future events with an\naccuracy of almost 90%, and can successfully reconstruct large portions of the\noriginal trace, greatly outperforming a Markov Model benchmark. We separately\nfeed the original, lossy, and reconstructed traces into a specification mining\nframework to perform downstream analysis of the effect of our method on\nstate-of-the-art models that use these traces for understanding the behavior of\ncomplex systems.', 'In this paper, we present a new approach to time series forecasting. Time\nseries data are prevalent in many scientific and engineering disciplines. Time\nseries forecasting is a crucial task in modeling time series data, and is an\nimportant area of machine learning. In this work we developed a novel method\nthat employs Transformer-based machine learning models to forecast time series\ndata. This approach works by leveraging self-attention mechanisms to learn\ncomplex patterns and dynamics from time series data. Moreover, it is a generic\nframework and can be applied to univariate and multivariate time series data,\nas well as time series embeddings. Using influenza-like illness (ILI)\nforecasting as a case study, we show that the forecasting results produced by\nour approach are favorably comparable to the state-of-the-art.', 'Global Forecasting Models (GFM) that are trained across a set of multiple\ntime series have shown superior results in many forecasting competitions and\nreal-world applications compared with univariate forecasting approaches. One\naspect of the popularity of statistical forecasting models such as ETS and\nARIMA is their relative simplicity and interpretability (in terms of relevant\nlags, trend, seasonality, and others), while GFMs typically lack\ninterpretability, especially towards particular time series. This reduces the\ntrust and confidence of the stakeholders when making decisions based on the\nforecasts without being able to understand the predictions. To mitigate this\nproblem, in this work, we propose a novel local model-agnostic interpretability\napproach to explain the forecasts from GFMs. We train simpler univariate\nsurrogate models that are considered interpretable (e.g., ETS) on the\npredictions of the GFM on samples within a neighbourhood that we obtain through\nbootstrapping or straightforwardly as the one-step-ahead global black-box model\nforecasts of the time series which needs to be explained. After, we evaluate\nthe explanations for the forecasts of the global models in both qualitative and\nquantitative aspects such as accuracy, fidelity, stability and\ncomprehensibility, and are able to show the benefits of our approach.', 'Time series forecasting is an important yet challenging task. Though deep\nlearning methods have recently been developed to give superior forecasting\nresults, it is crucial to improve the interpretability of time series models.\nPrevious interpretation methods, including the methods for general neural\nnetworks and attention-based methods, mainly consider the interpretation in the\nfeature dimension while ignoring the crucial temporal dimension. In this paper,\nwe present the series saliency framework for temporal interpretation for\nmultivariate time series forecasting, which considers the forecasting\ninterpretation in both feature and temporal dimensions. By extracting the\n""series images"" from the sliding windows of the time series, we apply the\nsaliency map segmentation following the smallest destroying region principle.\nThe series saliency framework can be employed to any well-defined deep learning\nmodels and works as a data augmentation to get more accurate forecasts.\nExperimental results on several real datasets demonstrate that our framework\ngenerates temporal interpretations for the time series forecasting task while\nproduces accurate time series forecast.', 'The prediction of residential power usage is essential in assisting a smart\ngrid to manage and preserve energy to ensure efficient use. An accurate energy\nforecasting at the customer level will reflect directly into efficiency\nimprovements across the power grid system, however forecasting building energy\nuse is a complex task due to many influencing factors, such as meteorological\nand occupancy patterns. In addiction, high-dimensional time series increasingly\narise in the Internet of Energy (IoE), given the emergence of multi-sensor\nenvironments and the two way communication between energy consumers and the\nsmart grid. Therefore, methods that are capable of computing high-dimensional\ntime series are of great value in smart building and IoE applications. Fuzzy\nTime Series (FTS) models stand out as data-driven non-parametric models of easy\nimplementation and high accuracy. Unfortunately, the existing FTS models can be\nunfeasible if all features were used to train the model. We present a new\nmethodology for handling high-dimensional time series, by projecting the\noriginal high-dimensional data into a low dimensional embedding space and using\nmultivariate FTS approach in this low dimensional representation. Combining\nthese techniques enables a better representation of the complex content of\nmultivariate time series and more accurate forecasts.', ""Modeling the spatiotemporal nature of the spread of infectious diseases can\nprovide useful intuition in understanding the time-varying aspect of the\ndisease spread and the underlying complex spatial dependency observed in\npeople's mobility patterns. Besides, the county level multiple related time\nseries information can be leveraged to make a forecast on an individual time\nseries. Adding to this challenge is the fact that real-time data often deviates\nfrom the unimodal Gaussian distribution assumption and may show some complex\nmixed patterns. Motivated by this, we develop a deep learning-based time-series\nmodel for probabilistic forecasting called Auto-regressive Mixed Density\nDynamic Diffusion Network(ARM3Dnet), which considers both people's mobility and\ndisease spread as a diffusion process on a dynamic directed graph. The Gaussian\nMixture Model layer is implemented to consider the multimodal nature of the\nreal-time data while learning from multiple related time series. We show that\nour model, when trained with the best combination of dynamic covariate features\nand mixture components, can outperform both traditional statistical and deep\nlearning models in forecasting the number of Covid-19 deaths and cases at the\ncounty level in the United States."", 'Some real-world domains, such as Agriculture and Healthcare, comprise\nearly-stage disease indications whose recording constitutes a rare event, and\nyet, whose precise detection at that stage is critical. In this type of highly\nimbalanced classification problems, which encompass complex features, deep\nlearning (DL) is much needed because of its strong detection capabilities. At\nthe same time, DL is observed in practice to favor majority over minority\nclasses and consequently suffer from inaccurate detection of the targeted\nearly-stage indications. To simulate such scenarios, we artificially generate\nskewness (99% vs. 1%) for certain plant types out of the PlantVillage dataset\nas a basis for classification of scarce visual cues through transfer learning.\nBy randomly and unevenly picking healthy and unhealthy samples from certain\nplant types to form a training set, we consider a base experiment as\nfine-tuning ResNet34 and VGG19 architectures and then testing the model\nperformance on a balanced dataset of healthy and unhealthy images. We\nempirically observe that the initial F1 test score jumps from 0.29 to 0.95 for\nthe minority class upon adding a final Batch Normalization (BN) layer just\nbefore the output layer in VGG19. We demonstrate that utilizing an additional\nBN layer before the output layer in modern CNN architectures has a considerable\nimpact in terms of minimizing the training time and testing error for minority\nclasses in highly imbalanced data sets. Moreover, when the final BN is\nemployed, minimizing the loss function may not be the best way to assure a high\nF1 test score for minority classes in such problems. That is, the network might\nperform better even if it is not confident enough while making a prediction;\nleading to another discussion about why softmax output is not a good\nuncertainty measure for DL models.', 'Neural network based data-driven market simulation unveils a new and flexible\nway of modelling financial time series without imposing assumptions on the\nunderlying stochastic dynamics. Though in this sense generative market\nsimulation is model-free, the concrete modelling choices are nevertheless\ndecisive for the features of the simulated paths. We give a brief overview of\ncurrently used generative modelling approaches and performance evaluation\nmetrics for financial time series, and address some of the challenges to\nachieve good results in the latter. We also contrast some classical approaches\nof market simulation with simulation based on generative modelling and\nhighlight some advantages and pitfalls of the new approach. While most\ngenerative models tend to rely on large amounts of training data, we present\nhere a generative model that works reliably in environments where the amount of\navailable training data is notoriously small. Furthermore, we show how a rough\npaths perspective combined with a parsimonious Variational Autoencoder\nframework provides a powerful way for encoding and evaluating financial time\nseries in such environments where available training data is scarce. Finally,\nwe also propose a suitable performance evaluation metric for financial time\nseries and discuss some connections of our Market Generator to deep hedging.', 'A highly comparative, feature-based approach to time series classification is\nintroduced that uses an extensive database of algorithms to extract thousands\nof interpretable features from time series. These features are derived from\nacross the scientific time-series analysis literature, and include summaries of\ntime series in terms of their correlation structure, distribution, entropy,\nstationarity, scaling properties, and fits to a range of time-series models.\nAfter computing thousands of features for each time series in a training set,\nthose that are most informative of the class structure are selected using\ngreedy forward feature selection with a linear classifier. The resulting\nfeature-based classifiers automatically learn the differences between classes\nusing a reduced number of time-series properties, and circumvent the need to\ncalculate distances between time series. Representing time series in this way\nresults in orders of magnitude of dimensionality reduction, allowing the method\nto perform well on very large datasets containing long time series or time\nseries of different lengths. For many of the datasets studied, classification\nperformance exceeded that of conventional instance-based classifiers, including\none nearest neighbor classifiers using Euclidean distances and dynamic time\nwarping and, most importantly, the features selected provide an understanding\nof the properties of the dataset, insight that can guide further scientific\ninvestigation.', 'In the current context of Big Data, the nature of many forecasting problems\nhas changed from predicting isolated time series to predicting many time series\nfrom similar sources. This has opened up the opportunity to develop competitive\nglobal forecasting models that simultaneously learn from many time series. But,\nit still remains unclear when global forecasting models can outperform the\nunivariate benchmarks, especially along the dimensions of the\nhomogeneity/heterogeneity of series, the complexity of patterns in the series,\nthe complexity of forecasting models, and the lengths/number of series. Our\nstudy attempts to address this problem through investigating the effect from\nthese factors, by simulating a number of datasets that have controllable time\nseries characteristics. Specifically, we simulate time series from simple data\ngenerating processes (DGP), such as Auto Regressive (AR) and Seasonal AR, to\ncomplex DGPs, such as Chaotic Logistic Map, Self-Exciting Threshold\nAuto-Regressive, and Mackey-Glass Equations. The data heterogeneity is\nintroduced by mixing time series generated from several DGPs into a single\ndataset. The lengths and the number of series in the dataset are varied in\ndifferent scenarios. We perform experiments on these datasets using global\nforecasting models including Recurrent Neural Networks (RNN), Feed-Forward\nNeural Networks, Pooled Regression (PR) models and Light Gradient Boosting\nModels (LGBM), and compare their performance against standard statistical\nunivariate forecasting techniques. Our experiments demonstrate that when\ntrained as global forecasting models, techniques such as RNNs and LGBMs, which\nhave complex non-linear modelling capabilities, are competitive methods in\ngeneral under challenging forecasting scenarios such as series having short\nlengths, datasets with heterogeneous series and having minimal prior knowledge\nof the patterns of the series.', 'For most diseases, building large databases of labeled genetic data is an\nexpensive and time-demanding task. To address this, we introduce genetic\nGenerative Adversarial Networks (gGAN), a semi-supervised approach based on an\ninnovative GAN architecture to create large synthetic genetic data sets\nstarting with a small amount of labeled data and a large amount of unlabeled\ndata. Our goal is to determine the propensity of a new individual to develop\nthe severe form of the illness from their genetic profile alone. The proposed\nmodel achieved satisfactory results using real genetic data from different\ndatasets and populations, in which the test populations may not have the same\ngenetic profiles. The proposed model is self-aware and capable of determining\nwhether a new genetic profile has enough compatibility with the data on which\nthe network was trained and is thus suitable for prediction. The code and\ndatasets used can be found at https://github.com/caio-davi/gGAN.', 'Automatic event detection from time series signals has wide applications,\nsuch as abnormal event detection in video surveillance and event detection in\ngeophysical data. Traditional detection methods detect events primarily by the\nuse of similarity and correlation in data. Those methods can be inefficient and\nyield low accuracy. In recent years, because of the significantly increased\ncomputational power, machine learning techniques have revolutionized many\nscience and engineering domains. In this study, we apply a deep-learning-based\nmethod to the detection of events from time series seismic signals. However, a\ndirect adaptation of the similar ideas from 2D object detection to our problem\nfaces two challenges. The first challenge is that the duration of earthquake\nevent varies significantly; The other is that the proposals generated are\ntemporally correlated. To address these challenges, we propose a novel cascaded\nregion-based convolutional neural network to capture earthquake events in\ndifferent sizes, while incorporating contextual information to enrich features\nfor each individual proposal. To achieve a better generalization performance,\nwe use densely connected blocks as the backbone of our network. Because of the\nfact that some positive events are not correctly annotated, we further\nformulate the detection problem as a learning-from-noise problem. To verify the\nperformance of our detection methods, we employ our methods to seismic data\ngenerated from a bi-axial ""earthquake machine"" located at Rock Mechanics\nLaboratory, and we acquire labels with the help of experts. Through our\nnumerical tests, we show that our novel detection techniques yield high\naccuracy. Therefore, our novel deep-learning-based detection methods can\npotentially be powerful tools for locating events from time series data in\nvarious applications.', 'The availability of large amounts of time series data, paired with the\nperformance of deep-learning algorithms on a broad class of problems, has\nrecently led to significant interest in the use of sequence-to-sequence models\nfor time series forecasting. We provide the first theoretical analysis of this\ntime series forecasting framework. We include a comparison of\nsequence-to-sequence modeling to classical time series models, and as such our\ntheory can serve as a quantitative guide for practitioners choosing between\ndifferent modeling methodologies.', 'Autonomous robots need to be able to adapt to unforeseen situations and to\nacquire new skills through trial and error. Reinforcement learning in principle\noffers a suitable methodological framework for this kind of autonomous\nlearning. However current computational reinforcement learning agents mostly\nlearn each individual skill entirely from scratch. How can we enable artificial\nagents, such as robots, to acquire some form of generic knowledge, which they\ncould leverage for the learning of new skills? This paper argues that, like the\nbrain, the cognitive system of artificial agents has to develop a world model\nto support adaptive behavior and learning. Inspiration is taken from two recent\ndevelopments in the cognitive science literature: predictive processing\ntheories of cognition, and the sensorimotor contingencies theory of perception.\nBased on these, a hypothesis is formulated about what the content of\ninformation might be that is encoded in an internal world model, and how an\nagent could autonomously acquire it. A computational model is described to\nformalize this hypothesis, and is evaluated in a series of simulation\nexperiments.', 'Feature-based time series representations have attracted substantial\nattention in a wide range of time series analysis methods. Recently, the use of\ntime series features for forecast model averaging has been an emerging research\nfocus in the forecasting community. Nonetheless, most of the existing\napproaches depend on the manual choice of an appropriate set of features.\nExploiting machine learning methods to extract features from time series\nautomatically becomes crucial in state-of-the-art time series analysis. In this\npaper, we introduce an automated approach to extract time series features based\non time series imaging. We first transform time series into recurrence plots,\nfrom which local features can be extracted using computer vision algorithms.\nThe extracted features are used for forecast model averaging. Our experiments\nshow that forecasting based on automatically extracted features, with less\nhuman intervention and a more comprehensive view of the raw time series data,\nyields highly comparable performances with the best methods in the largest\nforecasting competition dataset (M4) and outperforms the top methods in the\nTourism forecasting competition dataset.', 'This paper presents a machine learning approach to model the electric\nconsumption of electric vehicles at macroscopic level, i.e., in the absence of\na speed profile, while preserving microscopic level accuracy. For this work, we\nleveraged a high-performance, agent-based transportation tool to model trips\nthat occur in the Greater Chicago region under various scenario changes, along\nwith physics-based modeling and simulation tools to provide high-fidelity\nenergy consumption values. The generated results constitute a very large\ndataset of vehicle-route energy outcomes that capture variability in vehicle\nand routing setting, and in which high-fidelity time series of vehicle speed\ndynamics is masked. We show that although all internal dynamics that affect\nenergy consumption are masked, it is possible to learn aggregate-level energy\nconsumption values quite accurately with a deep learning approach. When\nlarge-scale data is available, and with carefully tailored feature engineering,\na well-designed model can overcome and retrieve latent information. This model\nhas been deployed and integrated within POLARIS Transportation System\nSimulation Tool to support real-time behavioral transportation models for\nindividual charging decision-making, and rerouting of electric vehicles.', 'Fuzzy time series forecasting methods are very popular among researchers for\npredicting future values as they are not based on the strict assumptions of\ntraditional time series forecasting methods. Non-stochastic methods of fuzzy\ntime series forecasting are preferred by the researchers as they provide more\nsignificant forecasting results. There are generally, four factors that\ndetermine the performance of the forecasting method (1) number of intervals\n(NOIs) and length of intervals to partition universe of discourse (UOD) (2)\nfuzzification rules or feature representation of crisp time series (3) method\nof establishing fuzzy logic rule (FLRs) between input and target values (4)\ndefuzzification rule to get crisp forecasted value. Considering the first two\nfactors to improve the forecasting accuracy, we proposed a novel non-stochastic\nmethod fuzzy time series forecasting in which interval index number and\nmembership value are used as input features to predict future value. We\nsuggested a simple rounding-off range and suitable step size method to find the\noptimal number of intervals (NOIs) and used fuzzy c-means clustering process to\ndivide UOD into intervals of unequal length. We implement support vector\nmachine (SVM) to establish FLRs. To test our proposed method we conduct a\nsimulated study on five widely used real time series and compare the\nperformance with some recently developed models. We also examine the\nperformance of the proposed model by using multi-layer perceptron (MLP) instead\nof SVM. Two performance measures RSME and SMAPE are used for performance\nanalysis and observed better forecasting accuracy by the proposed model.', ""The field of predictive process monitoring focuses on case-level models to\npredict a single specific outcome such as a particular objective, (remaining)\ntime, or next activity/remaining sequence. Recently, a longer-horizon,\nmodel-wide approach has been proposed in the form of process model forecasting,\nwhich predicts the future state of a whole process model through the\nforecasting of all activity-to-activity relations at once using time series\nforecasting.\n  This paper introduces the concept of \\emph{predictive process model\nmonitoring} which sits in the middle of both predictive process monitoring and\nprocess model forecasting. Concretely, by modelling a process model as a set of\nconstraints being present between activities over time, we can capture more\ndetailed information between activities compared to process model forecasting,\nwhile being compatible with typical predictive process monitoring objectives\nwhich are often expressed in the same language as these constraints. To achieve\nthis, Processes-As-Movies (PAM) is introduced, i.e., a novel technique capable\nof jointly mining and predicting declarative process constraints between\nactivities in various windows of a process' execution. PAM predicts what\ndeclarative rules hold for a trace (objective-based), which also supports the\nprediction of all constraints together as a process model (model-based).\nVarious recurrent neural network topologies inspired by video analysis tailored\nto temporal high-dimensional input are used to model the process model\nevolution with windows as time steps, including encoder-decoder long short-term\nmemory networks, and convolutional long short-term memory networks. Results\nobtained over real-life event logs show that these topologies are effective in\nterms of predictive accuracy and precision."", 'Deep learning performs remarkably well on many time series analysis tasks\nrecently. The superior performance of deep neural networks relies heavily on a\nlarge number of training data to avoid overfitting. However, the labeled data\nof many real-world time series applications may be limited such as\nclassification in medical time series and anomaly detection in AIOps. As an\neffective way to enhance the size and quality of the training data, data\naugmentation is crucial to the successful application of deep learning models\non time series data. In this paper, we systematically review different data\naugmentation methods for time series. We propose a taxonomy for the reviewed\nmethods, and then provide a structured review for these methods by highlighting\ntheir strengths and limitations. We also empirically compare different data\naugmentation methods for different tasks including time series classification,\nanomaly detection, and forecasting. Finally, we discuss and highlight five\nfuture directions to provide useful research guidance.', 'We participated in the M4 competition for time series forecasting and\ndescribe here our methods for forecasting daily time series. We used an\nensemble of five statistical forecasting methods and a method that we refer to\nas the correlator. Our retrospective analysis using the ground truth values\npublished by the M4 organisers after the competition demonstrates that the\ncorrelator was responsible for most of our gains over the naive constant\nforecasting method. We identify data leakage as one reason for its success,\npartly due to test data selected from different time intervals, and partly due\nto quality issues in the original time series. We suggest that future\nforecasting competitions should provide actual dates for the time series so\nthat some of those leakages could be avoided by the participants.', 'Evaluating the reliability of intelligent physical systems against rare\nsafety-critical events poses a huge testing burden for real-world applications.\nSimulation provides a useful platform to evaluate the extremal risks of these\nsystems before their deployments. Importance Sampling (IS), while proven to be\npowerful for rare-event simulation, faces challenges in handling these\nlearning-based systems due to their black-box nature that fundamentally\nundermines its efficiency guarantee, which can lead to under-estimation without\ndiagnostically detected. We propose a framework called Deep Probabilistic\nAccelerated Evaluation (Deep-PrAE) to design statistically guaranteed IS, by\nconverting black-box samplers that are versatile but could lack guarantees,\ninto one with what we call a relaxed efficiency certificate that allows\naccurate estimation of bounds on the safety-critical event probability. We\npresent the theory of Deep-PrAE that combines the dominating point concept with\nrare-event set learning via deep neural network classifiers, and demonstrate\nits effectiveness in numerical examples including the safety-testing of an\nintelligent driving algorithm.', 'Time Series Forecasting is at the core of many practical applications such as\nsales forecasting for business, rainfall forecasting for agriculture and many\nothers. Though this problem has been extensively studied for years, it is still\nconsidered a challenging problem due to complex and evolving nature of time\nseries data. Typical methods proposed for time series forecasting modeled\nlinear or non-linear dependencies between data observations. However it is a\ngenerally accepted notion that no one method is universally effective for all\nkinds of time series data. Attempts have been made to use dynamic and weighted\ncombination of heterogeneous and independent forecasting models and it has been\nfound to be a promising direction to tackle this problem. This method is based\non the assumption that different forecasters have different specialization and\nvarying performance for different distribution of data and weights are\ndynamically assigned to multiple forecasters accordingly. However in many\npractical time series data-set, the distribution of data slowly evolves with\ntime. We propose to employ a re-weighting based method to adjust the assigned\nweights to various forecasters in order to account for such distribution-drift.\nAn exhaustive testing was performed against both real-world and synthesized\ntime-series. Experimental results show the competitiveness of the method in\ncomparison to state-of-the-art approaches for combining forecasters and\nhandling drift.', 'We present Earliness-Aware Deep Convolutional Networks (EA-ConvNets), an\nend-to-end deep learning framework, for early classification of time series\ndata. Unlike most existing methods for early classification of time series\ndata, that are designed to solve this problem under the assumption of the\navailability of a good set of pre-defined (often hand-crafted) features, our\nframework can jointly perform feature learning (by learning a deep hierarchy of\n\\emph{shapelets} capturing the salient characteristics in each time series),\nalong with a dynamic truncation model to help our deep feature learning\narchitecture focus on the early parts of each time series. Consequently, our\nframework is able to make highly reliable early predictions, outperforming\nvarious state-of-the-art methods for early time series classification, while\nalso being competitive when compared to the state-of-the-art time series\nclassification algorithms that work with \\emph{fully observed} time series\ndata. To the best of our knowledge, the proposed framework is the first to\nperform data-driven (deep) feature learning in the context of early\nclassification of time series data. We perform a comprehensive set of\nexperiments, on several benchmark data sets, which demonstrate that our method\nyields significantly better predictions than various state-of-the-art methods\ndesigned for early time series classification. In addition to obtaining high\naccuracies, our experiments also show that the learned deep shapelets based\nfeatures are also highly interpretable and can help gain better understanding\nof the underlying characteristics of time series data.', 'Performance and high availability have become increasingly important drivers,\namongst other drivers, for user retention in the context of web services such\nas social networks, and web search. Exogenic and/or endogenic factors often\ngive rise to anomalies, making it very challenging to maintain high\navailability, while also delivering high performance. Given that\nservice-oriented architectures (SOA) typically have a large number of services,\nwith each service having a large set of metrics, automatic detection of\nanomalies is non-trivial.\n  Although there exists a large body of prior research in anomaly detection,\nexisting techniques are not applicable in the context of social network data,\nowing to the inherent seasonal and trend components in the time series data.\n  To this end, we developed two novel statistical techniques for automatically\ndetecting anomalies in cloud infrastructure data. Specifically, the techniques\nemploy statistical learning to detect anomalies in both application, and system\nmetrics. Seasonal decomposition is employed to filter the trend and seasonal\ncomponents of the time series, followed by the use of robust statistical\nmetrics -- median and median absolute deviation (MAD) -- to accurately detect\nanomalies, even in the presence of seasonal spikes.\n  We demonstrate the efficacy of the proposed techniques from three different\nperspectives, viz., capacity planning, user behavior, and supervised learning.\nIn particular, we used production data for evaluation, and we report Precision,\nRecall, and F-measure in each case.', ""Time series forecasting involves collecting and analyzing past observations\nto develop a model to extrapolate such observations into the future.\nForecasting of future events is important in many fields to support decision\nmaking as it contributes to reducing the future uncertainty. We propose\nexplainable boosted linear regression (EBLR) algorithm for time series\nforecasting, which is an iterative method that starts with a base model, and\nexplains the model's errors through regression trees. At each iteration, the\npath leading to highest error is added as a new variable to the base model. In\nthis regard, our approach can be considered as an improvement over general time\nseries models since it enables incorporating nonlinear features by residuals\nexplanation. More importantly, use of the single rule that contributes to the\nerror most allows for interpretable results. The proposed approach extends to\nprobabilistic forecasting through generating prediction intervals based on the\nempirical error distribution. We conduct a detailed numerical study with EBLR\nand compare against various other approaches. We observe that EBLR\nsubstantially improves the base model performance through extracted features,\nand provide a comparable performance to other well established approaches. The\ninterpretability of the model predictions and high predictive accuracy of EBLR\nmakes it a promising method for time series forecasting."", 'Classical anomaly detection is principally concerned with point-based\nanomalies, those anomalies that occur at a single point in time. Yet, many\nreal-world anomalies are range-based, meaning they occur over a period of time.\nMotivated by this observation, we present a new mathematical model to evaluate\nthe accuracy of time series classification algorithms. Our model expands the\nwell-known Precision and Recall metrics to measure ranges, while simultaneously\nenabling customization support for domain-specific preferences.', 'Choosing the technique that is the best at forecasting your data, is a\nproblem that arises in any forecasting application. Decades of research have\nresulted into an enormous amount of forecasting methods that stem from\nstatistics, econometrics and machine learning (ML), which leads to a very\ndifficult and elaborate choice to make in any forecasting exercise. This paper\naims to facilitate this process for high-level tactical sales forecasts by\ncomparing a large array of techniques for 35 times series that consist of both\nindustry data from the Coca-Cola Company and publicly available datasets.\nHowever, instead of solely focusing on the accuracy of the resulting forecasts,\nthis paper introduces a novel and completely automated profit-driven approach\nthat takes into account the expected profit that a technique can create during\nboth the model building and evaluation process. The expected profit function\nthat is used for this purpose, is easy to understand and adaptable to any\nsituation by combining forecasting accuracy with business expertise.\nFurthermore, we examine the added value of ML techniques, the inclusion of\nexternal factors and the use of seasonal models in order to ascertain which\ntype of model works best in tactical sales forecasting. Our findings show that\nsimple seasonal time series models consistently outperform other methodologies\nand that the profit-driven approach can lead to selecting a different\nforecasting model.', 'Recent advances in AIoT technologies have led to an increasing popularity of\nutilizing machine learning algorithms to detect operational failures for\ncyber-physical systems (CPS). In its basic form, an anomaly detection module\nmonitors the sensor measurements and actuator states from the physical plant,\nand detects anomalies in these measurements to identify abnormal operation\nstatus. Nevertheless, building effective anomaly detection models for CPS is\nrather challenging as the model has to accurately detect anomalies in presence\nof highly complicated system dynamics and unknown amount of sensor noise. In\nthis work, we propose a novel time series anomaly detection method called\nNeural System Identification and Bayesian Filtering (NSIBF) in which a\nspecially crafted neural network architecture is posed for system\nidentification, i.e., capturing the dynamics of CPS in a dynamical state-space\nmodel; then a Bayesian filtering algorithm is naturally applied on top of the\n""identified"" state-space model for robust anomaly detection by tracking the\nuncertainty of the hidden state of the system recursively over time. We provide\nqualitative as well as quantitative experiments with the proposed method on a\nsynthetic and three real-world CPS datasets, showing that NSIBF compares\nfavorably to the state-of-the-art methods with considerable improvements on\nanomaly detection in CPS.', 'Several techniques for multivariate time series anomaly detection have been\nproposed recently, but a systematic comparison on a common set of datasets and\nmetrics is lacking. This paper presents a systematic and comprehensive\nevaluation of unsupervised and semi-supervised deep-learning based methods for\nanomaly detection and diagnosis on multivariate time series data from\ncyberphysical systems. Unlike previous works, we vary the model and\npost-processing of model errors, i.e. the scoring functions independently of\neach other, through a grid of 10 models and 4 scoring functions, comparing\nthese variants to state of the art methods. In time-series anomaly detection,\ndetecting anomalous events is more important than detecting individual\nanomalous time-points. Through experiments, we find that the existing\nevaluation metrics either do not take events into account, or cannot\ndistinguish between a good detector and trivial detectors, such as a random or\nan all-positive detector. We propose a new metric to overcome these drawbacks,\nnamely, the composite F-score ($Fc_1$), for evaluating time-series anomaly\ndetection.\n  Our study highlights that dynamic scoring functions work much better than\nstatic ones for multivariate time series anomaly detection, and the choice of\nscoring functions often matters more than the choice of the underlying model.\nWe also find that a simple, channel-wise model - the Univariate Fully-Connected\nAuto-Encoder, with the dynamic Gaussian scoring function emerges as a winning\ncandidate for both anomaly detection and diagnosis, beating state of the art\nalgorithms.', ""Mechanical devices such as engines, vehicles, aircrafts, etc., are typically\ninstrumented with numerous sensors to capture the behavior and health of the\nmachine. However, there are often external factors or variables which are not\ncaptured by sensors leading to time-series which are inherently unpredictable.\nFor instance, manual controls and/or unmonitored environmental conditions or\nload may lead to inherently unpredictable time-series. Detecting anomalies in\nsuch scenarios becomes challenging using standard approaches based on\nmathematical models that rely on stationarity, or prediction models that\nutilize prediction errors to detect anomalies. We propose a Long Short Term\nMemory Networks based Encoder-Decoder scheme for Anomaly Detection (EncDec-AD)\nthat learns to reconstruct 'normal' time-series behavior, and thereafter uses\nreconstruction error to detect anomalies. We experiment with three publicly\navailable quasi predictable time-series datasets: power demand, space shuttle,\nand ECG, and two real-world engine datasets with both predictive and\nunpredictable behavior. We show that EncDec-AD is robust and can detect\nanomalies from predictable, unpredictable, periodic, aperiodic, and\nquasi-periodic time-series. Further, we show that EncDec-AD is able to detect\nanomalies from short time-series (length as small as 30) as well as long\ntime-series (length as large as 500)."", ""Many software engineering tasks, such as testing, and anomaly detection can\nbenefit from the ability to infer a behavioral model of the software.Most\nexisting inference approaches assume access to code to collect execution\nsequences. In this paper, we investigate a black-box scenario, where the system\nunder analysis cannot be instrumented, in this granular fashion.This scenario\nis particularly prevalent with control systems' log analysis in the form of\ncontinuous signals. In this situation, an execution trace amounts to a\nmultivariate time-series of input and output signals, where different states of\nthe system correspond to different `phases` in the time-series. The main\nchallenge is to detect when these phase changes take place. Unfortunately, most\nexisting solutions are either univariate, make assumptions on the data\ndistribution, or have limited learning power.Therefore, we propose a hybrid\ndeep neural network that accepts as input a multivariate time series and\napplies a set of convolutional and recurrent layers to learn the non-linear\ncorrelations between signals and the patterns over time.We show how this\napproach can be used to accurately detect state changes, and how the inferred\nmodels can be successfully applied to transfer-learning scenarios, to\naccurately process traces from different products with similar execution\ncharacteristics. Our experimental results on two UAV autopilot case studies\nindicate that our approach is highly accurate (over 90% F1 score for state\nclassification) and significantly improves baselines (by up to 102% for change\npoint detection).Using transfer learning we also show that up to 90% of the\nmaximum achievable F1 scores in the open-source case study can be achieved by\nreusing the trained models from the industrial case and only fine tuning them\nusing as low as 5 labeled samples, which reduces the manual labeling effort by\n98%."", 'Predicting disaster events from seismic data is of paramount importance and\ncan save thousands of lives, especially in earthquake-prone areas and\nhabitations around volcanic craters. The drastic rise in the number of seismic\nmonitoring stations in recent years has allowed the collection of a huge\nquantity of data, outpacing the capacity of seismologists. Due to the complex\nnature of the seismological data, it is often difficult for seismologists to\ndetect subtle patterns with major implications. Machine learning algorithms\nhave been demonstrated to be effective in classification and prediction tasks\nfor seismic data. It has been widely known that some animals can sense\ndisasters like earthquakes from seismic signals well before the disaster\nstrikes. Mel spectrogram has been widely used for speech recognition as it\nscales the actual frequencies according to human hearing. In this paper, we\npropose a variant of the Mel spectrogram to scale the raw frequencies of\nseismic data to the hearing of such animals that can sense disasters from\nseismic signals. We are using a Computer vision algorithm along with clustering\nthat allows for the classification of unlabelled seismic data.', 'The explosion of time series data in recent years has brought a flourish of\nnew time series analysis methods, for forecasting, clustering, classification\nand other tasks. The evaluation of these new methods requires either collecting\nor simulating a diverse set of time series benchmarking data to enable reliable\ncomparisons against alternative approaches. We propose GeneRAting TIme Series\nwith diverse and controllable characteristics, named GRATIS, with the use of\nmixture autoregressive (MAR) models. We simulate sets of time series using MAR\nmodels and investigate the diversity and coverage of the generated time series\nin a time series feature space. By tuning the parameters of the MAR models,\nGRATIS is also able to efficiently generate new time series with controllable\nfeatures. In general, as a costless surrogate to the traditional data\ncollection approach, GRATIS can be used as an evaluation tool for tasks such as\ntime series forecasting and classification. We illustrate the usefulness of our\ntime series generation process through a time series forecasting application.', 'Autoencoders have been proposed as a powerful tool for model-independent\nanomaly detection in high-energy physics. The operating principle is that\nevents which do not belong to the space of training data will be reconstructed\npoorly, thus flagging them as anomalies. We point out that in a variety of\nexamples of interest, the connection between large reconstruction error and\nanomalies is not so clear. In particular, for data sets with nontrivial\ntopology, there will always be points that erroneously seem anomalous due to\nglobal issues. Conversely, neural networks typically have an inductive bias or\nprior to locally interpolate such that undersampled or rare events may be\nreconstructed with small error, despite actually being the desired anomalies.\nTaken together, these facts are in tension with the simple picture of the\nautoencoder as an anomaly detector. Using a series of illustrative\nlow-dimensional examples, we show explicitly how the intrinsic and extrinsic\ntopology of the dataset affects the behavior of an autoencoder and how this\ntopology is manifested in the latent space representation during training. We\nground this analysis in the discussion of a mock ""bump hunt"" in which the\nautoencoder fails to identify an anomalous ""signal"" for reasons tied to the\nintrinsic topology of $n$-particle phase space.', 'I describe the rationale for, and design of, an agent-based simulation model\nof a contemporary online sports-betting exchange: such exchanges, closely\nrelated to the exchange mechanisms at the heart of major financial markets,\nhave revolutionized the gambling industry in the past 20 years, but gathering\nsufficiently large quantities of rich and temporally high-resolution data from\nreal exchanges - i.e., the sort of data that is needed in large quantities for\nDeep Learning - is often very expensive, and sometimes simply impossible; this\ncreates a need for a plausibly realistic synthetic data generator, which is\nwhat this simulation now provides. The simulator, named the ""Bristol Betting\nExchange"" (BBE), is intended as a common platform, a data-source and\nexperimental test-bed, for researchers studying the application of AI and\nmachine learning (ML) techniques to issues arising in betting exchanges; and,\nas far as I have been able to determine, BBE is the first of its kind: a free\nopen-source agent-based simulation model consisting not only of a\nsports-betting exchange, but also a minimal simulation model of racetrack\nsporting events (e.g., horse-races or car-races) about which bets may be made,\nand a population of simulated bettors who each form their own private\nevaluation of odds and place bets on the exchange before and - crucially -\nduring the race itself (i.e., so-called ""in-play"" betting) and whose betting\nopinions change second-by-second as each race event unfolds. BBE is offered as\na proof-of-concept system that enables the generation of large high-resolution\ndata-sets for automated discovery or improvement of profitable strategies for\nbetting on sporting events via the application of AI/ML and advanced data\nanalytics techniques. This paper offers an extensive survey of relevant\nliterature and explains the motivation and design of BBE, and presents brief\nillustrative results.', 'A time series represents a set of observations collected over time.\nTypically, these observations are captured with a uniform sampling frequency\n(e.g. daily). When data points are observed in uneven time intervals the time\nseries is referred to as irregular or intermittent. In such scenarios, the most\ncommon solution is to reconstruct the time series to make it regular, thus\nremoving its intermittency. We hypothesise that, in irregular time series, the\ntime at which each observation is collected may be helpful to summarise the\ndynamics of the data and improve forecasting performance. We study this idea by\ndeveloping a novel automatic feature engineering framework, which focuses on\nextracting information from this point of view, i.e., when each instance is\ncollected. We study how valuable this information is by integrating it in a\ntime series forecasting workflow and investigate how it compares to or\ncomplements state-of-the-art methods for regular time series forecasting. In\nthe end, we contribute by providing a novel framework that tackles feature\nengineering for time series from an angle previously vastly ignored. We show\nthat our approach has the potential to further extract more information about\ntime series that significantly improves forecasting performance.', 'Time series data account for a major part of data supply available today.\nTime series mining handles several tasks such as classification, clustering,\nquery-by-content, prediction, and others. Performing data mining tasks on raw\ntime series is inefficient as these data are high-dimensional by nature.\nInstead, time series are first pre-processed using several techniques before\ndifferent data mining tasks can be performed on them. In general, there are two\nmain approaches to reduce time series dimensionality, the first is what we call\nlandmark methods. These methods are based on finding characteristic features in\nthe target time series. The second is based on data transformations. These\nmethods transform the time series from the original space into a reduced space,\nwhere they can be managed more efficiently. The method we present in this paper\napplies a third approach, as it projects a time series onto a lower-dimensional\nspace by selecting important points in the time series. The novelty of our\nmethod is that these points are not chosen according to a geometric criterion,\nwhich is subjective in most cases, but through an optimization process. The\nother important characteristic of our method is that these important points are\nselected on a dataset-level and not on a single time series-level. The direct\nadvantage of this strategy is that the distance defined on the low-dimensional\nspace lower bounds the original distance applied to raw data. This enables us\nto apply the popular GEMINI algorithm. The promising results of our experiments\non a wide variety of time series datasets, using different optimizers, and\napplied to the two major data mining tasks, validate our new method.', 'In this paper, we focus on the problem of unsupervised image-sentence\nmatching. Existing research explores to utilize document-level structural\ninformation to sample positive and negative instances for model training.\nAlthough the approach achieves positive results, it introduces a sampling bias\nand fails to distinguish instances with high semantic similarity. To alleviate\nthe bias, we propose a new sampling strategy to select additional\nintra-document image-sentence pairs as positive or negative samples.\nFurthermore, to recognize the complex pattern in intra-document samples, we\npropose a Transformer based model to capture fine-grained features and\nimplicitly construct a graph for each document, where concepts in a document\nare introduced to bridge the representation learning of images and sentences in\nthe context of a document. Experimental results show the effectiveness of our\napproach to alleviate the bias and learn well-aligned multimodal\nrepresentations.', 'Is critical input information encoded in specific sparse pathways within the\nneural network? In this work, we discuss the problem of identifying these\ncritical pathways and subsequently leverage them for interpreting the network\'s\nresponse to an input. The pruning objective -- selecting the smallest group of\nneurons for which the response remains equivalent to the original network --\nhas been previously proposed for identifying critical pathways. We demonstrate\nthat sparse pathways derived from pruning do not necessarily encode critical\ninput information. To ensure sparse pathways include critical fragments of the\nencoded input information, we propose pathway selection via neurons\'\ncontribution to the response. We proceed to explain how critical pathways can\nreveal critical input features. We prove that pathways selected via neuron\ncontribution are locally linear (in an L2-ball), a property that we use for\nproposing a feature attribution method: ""pathway gradient"". We validate our\ninterpretation method using mainstream evaluation experiments. The validation\nof pathway gradient interpretation method further confirms that selected\npathways using neuron contributions correspond to critical input features. The\ncode is publicly available.', 'With the sweeping digitalization of societal, medical, industrial, and\nscientific processes, sensing technologies are being deployed that produce\nincreasing volumes of time series data, thus fueling a plethora of new or\nimproved applications. In this setting, outlier detection is frequently\nimportant, and while solutions based on neural networks exist, they leave room\nfor improvement in terms of both accuracy and efficiency. With the objective of\nachieving such improvements, we propose a diversity-driven, convolutional\nensemble. To improve accuracy, the ensemble employs multiple basic outlier\ndetection models built on convolutional sequence-to-sequence autoencoders that\ncan capture temporal dependencies in time series. Further, a novel\ndiversity-driven training method maintains diversity among the basic models,\nwith the aim of improving the ensemble\'s accuracy. To improve efficiency, the\napproach enables a high degree of parallelism during training. In addition, it\nis able to transfer some model parameters from one basic model to another,\nwhich reduces training time. We report on extensive experiments using\nreal-world multivariate time series that offer insight into the design choices\nunderlying the new approach and offer evidence that it is capable of improved\naccuracy and efficiency. This is an extended version of ""Unsupervised Time\nSeries Outlier Detection with Diversity-Driven Convolutional Ensembles"", to\nappear in PVLDB 2022.', 'The emergence of continuous health monitoring and the availability of an\nenormous amount of time series data has provided a great opportunity for the\nadvancement of personal health tracking. In recent years, unsupervised learning\nmethods have drawn special attention of researchers to tackle the sparse\nannotation of health data and real-time detection of anomalies has been a\ncentral problem of interest. However, one problem that has not been well\naddressed before is the early prediction of forthcoming negative health events.\nEarly signs of an event can introduce subtle and gradual changes in the health\nsignal prior to its onset, detection of which can be invaluable in effective\nprevention. In this study, we first demonstrate our observations on the\nshortcoming of widely adopted anomaly detection methods in uncovering the\nchanges prior to a negative health event. We then propose a framework which\nrelies on online clustering of signal segment representations which are\nautomatically learned by a specially designed LSTM auto-encoder. We show the\neffectiveness of our approach by predicting Bradycardia events in infants using\nMIT-PICS dataset 1.3 minutes ahead of time with 68\\% AUC score on average,\nusing no label supervision. Results of our study can indicate the viability of\nour approach in the early detection of health events in other applications as\nwell.', 'Extreme events are occurrences whose magnitude and potential cause extensive\ndamage on people, infrastructure, and the environment. Motivated by the extreme\nnature of the current global health landscape, which is plagued by the\ncoronavirus pandemic, we seek to better understand and model extreme events.\nModeling extreme events is common in practice and plays an important role in\ntime-series prediction applications. Our goal is to (i) compare and investigate\nthe effect of some common extreme events modeling methods to explore which\nmethod can be practical in reality and (ii) accelerate the deep learning\ntraining process, which commonly uses deep recurrent neural network (RNN), by\nimplementing the asynchronous local Stochastic Gradient Descent (SGD) framework\namong multiple compute nodes. In order to verify our distributed extreme events\nmodeling, we evaluate our proposed framework on a stock data set S\\&P500, with\na standard recurrent neural network. Our intuition is to explore the (best)\nextreme events modeling method which could work well under the distributed deep\nlearning setting. Moreover, by using asynchronous distributed learning, we aim\nto significantly reduce the communication cost among the compute nodes and\ncentral server, which is the main bottleneck of almost all distributed learning\nframeworks.\n  We implement our proposed work and evaluate its performance on representative\ndata sets, such as S&P500 stock in $5$-year period. The experimental results\nvalidate the correctness of the design principle and show a significant\ntraining duration reduction upto $8$x, compared to the baseline single compute\nnode. Our results also show that our proposed work can achieve the same level\nof test accuracy, compared to the baseline setting.', 'Interactive response time is important in analytical pipelines for users to\nexplore a sufficient number of possibilities and make informed business\ndecisions. We consider a forecasting pipeline with large volumes of\nhigh-dimensional time series data. Real-time forecasting can be conducted in\ntwo steps. First, we specify the part of data to be focused on and the measure\nto be predicted by slicing, dicing, and aggregating the data. Second, a\nforecasting model is trained on the aggregated results to predict the trend of\nthe specified measure. While there are a number of forecasting models\navailable, the first step is the performance bottleneck. A natural idea is to\nutilize sampling to obtain approximate aggregations in real time as the input\nto train the forecasting model. Our scalable real-time forecasting system\nFlashP (Flash Prediction) is built based on this idea, with two major\nchallenges to be resolved in this paper: first, we need to figure out how\napproximate aggregations affect the fitting of forecasting models, and\nforecasting results; and second, accordingly, what sampling algorithms we\nshould use to obtain these approximate aggregations and how large the samples\nare. We introduce a new sampling scheme, called GSW sampling, and analyze error\nbounds for estimating aggregations using GSW samples. We introduce how to\nconstruct compact GSW samples with the existence of multiple measures to be\nanalyzed. We conduct experiments to evaluate our solution and compare it with\nalternatives on real data.', 'Deep Learning (DL) models can be used to tackle time series analysis tasks\nwith great success. However, the performance of DL models can degenerate\nrapidly if the data are not appropriately normalized. This issue is even more\napparent when DL is used for financial time series forecasting tasks, where the\nnon-stationary and multimodal nature of the data pose significant challenges\nand severely affect the performance of DL models. In this work, a simple, yet\neffective, neural layer, that is capable of adaptively normalizing the input\ntime series, while taking into account the distribution of the data, is\nproposed. The proposed layer is trained in an end-to-end fashion using\nback-propagation and leads to significant performance improvements compared to\nother evaluated normalization schemes. The proposed method differs from\ntraditional normalization methods since it learns how to perform normalization\nfor a given task instead of using a fixed normalization scheme. At the same\ntime, it can be directly applied to any new time series without requiring\nre-training. The effectiveness of the proposed method is demonstrated using a\nlarge-scale limit order book dataset, as well as a load forecasting dataset.', 'Time series forecasting is a crucial component of many important\napplications, ranging from forecasting the stock markets to energy load\nprediction. The high-dimensionality, velocity and variety of the data collected\nin these applications pose significant and unique challenges that must be\ncarefully addressed for each of them. In this work, a novel Temporal Logistic\nNeural Bag-of-Features approach, that can be used to tackle these challenges,\nis proposed. The proposed method can be effectively combined with deep neural\nnetworks, leading to powerful deep learning models for time series analysis.\nHowever, combining existing BoF formulations with deep feature extractors pose\nsignificant challenges: the distribution of the input features is not\nstationary, tuning the hyper-parameters of the model can be especially\ndifficult and the normalizations involved in the BoF model can cause\nsignificant instabilities during the training process. The proposed method is\ncapable of overcoming these limitations by a employing a novel adaptive scaling\nmechanism and replacing the classical Gaussian-based density estimation\ninvolved in the regular BoF model with a logistic kernel. The effectiveness of\nthe proposed approach is demonstrated using extensive experiments on a\nlarge-scale financial time series dataset that consists of more than 4 million\nlimit orders.', 'A central server needs to perform statistical inference based on samples that\nare distributed over multiple users who can each send a message of limited\nlength to the center. We study problems of distribution learning and identity\ntesting in this distributed inference setting and examine the role of shared\nrandomness as a resource. We propose a general-purpose simulate-and-infer\nstrategy that uses only private-coin communication protocols and is\nsample-optimal for distribution learning. This general strategy turns out to be\nsample-optimal even for distribution testing among private-coin protocols.\nInterestingly, we propose a public-coin protocol that outperforms\nsimulate-and-infer for distribution testing and is, in fact, sample-optimal.\nUnderlying our public-coin protocol is a random hash that when applied to the\nsamples minimally contracts the chi-squared distance of their distribution to\nthe uniform distribution.', 'The ability to perform causal and counterfactual reasoning are central\nproperties of human intelligence. Decision-making systems that can perform\nthese types of reasoning have the potential to be more generalizable and\ninterpretable. Simulations have helped advance the state-of-the-art in this\ndomain, by providing the ability to systematically vary parameters (e.g.,\nconfounders) and generate examples of the outcomes in the case of\ncounterfactual scenarios. However, simulating complex temporal causal events in\nmulti-agent scenarios, such as those that exist in driving and vehicle\nnavigation, is challenging. To help address this, we present a high-fidelity\nsimulation environment that is designed for developing algorithms for causal\ndiscovery and counterfactual reasoning in the safety-critical context. A core\ncomponent of our work is to introduce \\textit{agency}, such that it is simple\nto define and create complex scenarios using high-level definitions. The\nvehicles then operate with agency to complete these objectives, meaning\nlow-level behaviors need only be controlled if necessary. We perform\nexperiments with three state-of-the-art methods to create baselines and\nhighlight the affordances of this environment. Finally, we highlight challenges\nand opportunities for future work.', 'Generative adversarial networks (GANs) are recently highly successful in\ngenerative applications involving images and start being applied to time series\ndata. Here we describe EEG-GAN as a framework to generate\nelectroencephalographic (EEG) brain signals. We introduce a modification to the\nimproved training of Wasserstein GANs to stabilize training and investigate a\nrange of architectural choices critical for time series generation (most\nnotably up- and down-sampling). For evaluation we consider and compare\ndifferent metrics such as Inception score, Frechet inception distance and\nsliced Wasserstein distance, together showing that our EEG-GAN framework\ngenerated naturalistic EEG examples. It thus opens up a range of new generative\napplication scenarios in the neuroscientific and neurological context, such as\ndata augmentation in brain-computer interfacing tasks, EEG super-sampling, or\nrestoration of corrupted data segments. The possibility to generate signals of\na certain class and/or with specific properties may also open a new avenue for\nresearch into the underlying structure of brain signals.', 'In the context of time series forecasting, it is a common practice to\nevaluate multiple methods and choose one of these methods or an ensemble for\nproducing the best forecasts. However, choosing among different ensembles over\nmultiple methods remains a challenging task that undergoes a combinatorial\nexplosion as the number of methods increases. In the context of demand\nforecasting or revenue forecasting, this challenge is further exacerbated by a\nlarge number of time series as well as limited historical data points available\ndue to changing business context. Although deep learning forecasting methods\naim to simultaneously forecast large collections of time series, they become\nchallenging to apply in such scenarios due to the limited history available and\nmight not yield desirable results. We propose a framework for forecasting short\nhigh-dimensional time series data by combining low-rank temporal matrix\nfactorization and optimal model selection on latent time series using\ncross-validation. We demonstrate that forecasting the latent factors leads to\nsignificant performance gains as compared to directly applying different\nuni-variate models on time series. Performance has been validated on a\ntruncated version of the M4 monthly dataset which contains time series data\nfrom multiple domains showing the general applicability of the method.\nMoreover, it is amenable to incorporating the analyst view of the future owing\nto the low number of latent factors which is usually impractical when applying\nforecasting methods directly to high dimensional datasets.', 'Time series forecasting is widely used in business intelligence, e.g.,\nforecast stock market price, sales, and help the analysis of data trend. Most\ntime series of interest are macroscopic time series that are aggregated from\nmicroscopic data. However, instead of directly modeling the macroscopic time\nseries, rare literature studied the forecasting of macroscopic time series by\nleveraging data on the microscopic level. In this paper, we assume that the\nmicroscopic time series follow some unknown mixture probabilistic\ndistributions. We theoretically show that as we identify the ground truth\nlatent mixture components, the estimation of time series from each component\ncould be improved because of lower variance, thus benefitting the estimation of\nmacroscopic time series as well. Inspired by the power of Seq2seq and its\nvariants on the modeling of time series data, we propose Mixture of Seq2seq\n(MixSeq), an end2end mixture model to cluster microscopic time series, where\nall the components come from a family of Seq2seq models parameterized by\ndifferent parameters. Extensive experiments on both synthetic and real-world\ndata show the superiority of our approach.', 'Complex computer simulators are increasingly used across fields of science as\ngenerative models tying parameters of an underlying theory to experimental\nobservations. Inference in this setup is often difficult, as simulators rarely\nadmit a tractable density or likelihood function. We introduce Adversarial\nVariational Optimization (AVO), a likelihood-free inference algorithm for\nfitting a non-differentiable generative model incorporating ideas from\ngenerative adversarial networks, variational optimization and empirical Bayes.\nWe adapt the training procedure of generative adversarial networks by replacing\nthe differentiable generative network with a domain-specific simulator. We\nsolve the resulting non-differentiable minimax problem by minimizing\nvariational upper bounds of the two adversarial objectives. Effectively, the\nprocedure results in learning a proposal distribution over simulator\nparameters, such that the JS divergence between the marginal distribution of\nthe synthetic data and the empirical distribution of observed data is\nminimized. We evaluate and compare the method with simulators producing both\ndiscrete and continuous data.', ""Time-series anomaly detection is a popular topic in both academia and\nindustrial fields. Many companies need to monitor thousands of temporal signals\nfor their applications and services and require instant feedback and alerts for\npotential incidents in time. The task is challenging because of the complex\ncharacteristics of time-series, which are messy, stochastic, and often without\nproper labels. This prohibits training supervised models because of lack of\nlabels and a single model hardly fits different time series. In this paper, we\npropose a solution to address these issues. We present an automated model\nselection framework to automatically find the most suitable detection model\nwith proper parameters for the incoming data. The model selection layer is\nextensible as it can be updated without too much effort when a new detector is\navailable to the service. Finally, we incorporate a customized tuning algorithm\nto flexibly filter anomalies to meet customers' criteria. Experiments on\nreal-world datasets show the effectiveness of our solution."", 'Research into time series classification has tended to focus on the case of\nseries of uniform length. However, it is common for real-world time series data\nto have unequal lengths. Differing time series lengths may arise from a number\nof fundamentally different mechanisms. In this work, we identify and evaluate\ntwo classes of such mechanisms -- variations in sampling rate relative to the\nrelevant signal and variations between the start and end points of one time\nseries relative to one another. We investigate how time series generated by\neach of these classes of mechanism are best addressed for time series\nclassification. We perform extensive experiments and provide practical\nrecommendations on how variations in length should be handled in time series\nclassification.', 'Time series research has gathered lots of interests in the last decade,\nespecially for Time Series Classification (TSC) and Time Series Forecasting\n(TSF). Research in TSC has greatly benefited from the University of California\nRiverside and University of East Anglia (UCR/UEA) Time Series Archives. On the\nother hand, the advancement in Time Series Forecasting relies on time series\nforecasting competitions such as the Makridakis competitions, NN3 and NN5\nNeural Network competitions, and a few Kaggle competitions. Each year,\nthousands of papers proposing new algorithms for TSC and TSF have utilized\nthese benchmarking archives. These algorithms are designed for these specific\nproblems, but may not be useful for tasks such as predicting the heart rate of\na person using photoplethysmogram (PPG) and accelerometer data. We refer to\nthis problem as Time Series Extrinsic Regression (TSER), where we are\ninterested in a more general methodology of predicting a single continuous\nvalue, from univariate or multivariate time series. This prediction can be from\nthe same time series or not directly related to the predictor time series and\ndoes not necessarily need to be a future value or depend heavily on recent\nvalues. To the best of our knowledge, research into TSER has received much less\nattention in the time series research community and there are no models\ndeveloped for general time series extrinsic regression problems. Most models\nare developed for a specific problem. Therefore, we aim to motivate and support\nthe research into TSER by introducing the first TSER benchmarking archive. This\narchive contains 19 datasets from different domains, with varying number of\ndimensions, unequal length dimensions, and missing values. In this paper, we\nintroduce the datasets in this archive and did an initial benchmark on existing\nmodels.', ""In reinforcement learning, it is common to let an agent interact for a fixed\namount of time with its environment before resetting it and repeating the\nprocess in a series of episodes. The task that the agent has to learn can\neither be to maximize its performance over (i) that fixed period, or (ii) an\nindefinite period where time limits are only used during training to diversify\nexperience. In this paper, we provide a formal account for how time limits\ncould effectively be handled in each of the two cases and explain why not doing\nso can cause state aliasing and invalidation of experience replay, leading to\nsuboptimal policies and training instability. In case (i), we argue that the\nterminations due to time limits are in fact part of the environment, and thus a\nnotion of the remaining time should be included as part of the agent's input to\navoid violation of the Markov property. In case (ii), the time limits are not\npart of the environment and are only used to facilitate learning. We argue that\nthis insight should be incorporated by bootstrapping from the value of the\nstate at the end of each partial episode. For both cases, we illustrate\nempirically the significance of our considerations in improving the performance\nand stability of existing reinforcement learning algorithms, showing\nstate-of-the-art results on several control tasks."", 'The paper is focused on the forecasting method for time series groups with\nthe use of algorithms for cluster analysis. $K$-means algorithm is suggested to\nbe a basic one for clustering. The coordinates of the centers of clusters have\nbeen put in correspondence with summarizing time series data the centroids of\nthe clusters. A description of time series, the centroids of the clusters, is\nimplemented with the use of forecasting models. They are based on strict binary\ntrees and a modified clonal selection algorithm. With the help of such\nforecasting models, the possibility of forming analytic dependences is shown.\nIt is suggested to use a common forecasting model, which is constructed for\ntime series the centroid of the cluster, in forecasting the private\n(individual) time series in the cluster. The promising application of the\nsuggested method for grouped time series forecasting is demonstrated.', 'Data augmentation is a key element of deep learning pipelines, as it informs\nthe network during training about transformations of the input data that keep\nthe label unchanged. Manually finding adequate augmentation methods and\nparameters for a given pipeline is however rapidly cumbersome. In particular,\nwhile intuition can guide this decision for images, the design and choice of\naugmentation policies remains unclear for more complex types of data, such as\nneuroscience signals. Besides, class-dependent augmentation strategies have\nbeen surprisingly unexplored in the literature, although it is quite intuitive:\nchanging the color of a car image does not change the object class to be\npredicted, but doing the same to the picture of an orange does. This paper\ninvestigates gradient-based automatic data augmentation algorithms amenable to\nclass-wise policies with exponentially larger search spaces. Motivated by\nsupervised learning applications using EEG signals for which good augmentation\npolicies are mostly unknown, we propose a new differentiable relaxation of the\nproblem. In the class-agnostic setting, results show that our new relaxation\nleads to optimal performance with faster training than competing gradient-based\nmethods, while also outperforming gradient-free methods in the class-wise\nsetting. This work proposes also novel differentiable augmentation operations\nrelevant for sleep stage classification.', 'We consider the problem of mining signal temporal logical requirements from a\ndataset of regular (good) and anomalous (bad) trajectories of a dynamical\nsystem. We assume the training set to be labeled by human experts and that we\nhave access only to a limited amount of data, typically noisy. We provide a\nsystematic approach to synthesize both the syntactical structure and the\nparameters of the temporal logic formula using a two-steps procedure: first, we\nleverage a novel evolutionary algorithm for learning the structure of the\nformula; second, we perform the parameter synthesis operating on the\nstatistical emulation of the average robustness for a candidate formula w.r.t.\nits parameters. We compare our results with our previous work [{BufoBSBLB14]\nand with a recently proposed decision-tree [bombara_decision_2016] based\nmethod. We present experimental results on two case studies: an anomalous\ntrajectory detection problem of a naval surveillance system and the\ncharacterization of an Ineffective Respiratory effort, showing the usefulness\nof our work.', 'Many businesses and industries nowadays rely on large quantities of time\nseries data making time series forecasting an important research area. Global\nforecasting models that are trained across sets of time series have shown a\nhuge potential in providing accurate forecasts compared with the traditional\nunivariate forecasting models that work on isolated series. However, there are\ncurrently no comprehensive time series archives for forecasting that contain\ndatasets of time series from similar sources available for the research\ncommunity to evaluate the performance of new global forecasting algorithms over\na wide variety of datasets. In this paper, we present such a comprehensive time\nseries forecasting archive containing 20 publicly available time series\ndatasets from varied domains, with different characteristics in terms of\nfrequency, series lengths, and inclusion of missing values. We also\ncharacterise the datasets, and identify similarities and differences among\nthem, by conducting a feature analysis. Furthermore, we present the performance\nof a set of standard baseline forecasting methods over all datasets across\neight error metrics, for the benefit of researchers using the archive to\nbenchmark their forecasting algorithms.', 'Organizations rely heavily on time series metrics to measure and model key\naspects of operational and business performance. The ability to reliably detect\nissues with these metrics is imperative to identifying early indicators of\nmajor problems before they become pervasive. It can be very challenging to\nproactively monitor a large number of diverse and constantly changing time\nseries for anomalies, so there are often gaps in monitoring coverage, disabled\nor ignored monitors due to false positive alarms, and teams resorting to manual\ninspection of charts to catch problems. Traditionally, variations in the data\ngeneration processes and patterns have required strong modeling expertise to\ncreate models that accurately flag anomalies. In this paper, we describe an\nanomaly detection system that overcomes this common challenge by keeping track\nof its own performance and making changes as necessary to each model without\nrequiring manual intervention. We demonstrate that this novel approach\noutperforms available alternatives on benchmark datasets in many scenarios.', 'In the fields of statistics and unsupervised machine learning a fundamental\nand well-studied problem is anomaly detection. Anomalies are difficult to\ndefine, yet many algorithms have been proposed. Underlying the approaches is\nthe nebulous understanding that anomalies are rare, unusual or inconsistent\nwith the majority of data. The present work provides a philosophical treatise\nto clearly define anomalies and develops an algorithm for their efficient\ndetection with minimal user intervention. Inspired by the Gestalt School of\nPsychology and the Helmholtz principle of human perception, anomalies are\nassumed to be observations that are unexpected to occur with respect to certain\ngroupings made by the majority of the data. Under appropriate random variable\nmodelling anomalies are directly found in a set of data by a uniform and\nindependent random assumption of the distribution of constituent elements of\nthe observations, with anomalies corresponding to those observations where the\nexpectation of the number of occurrences of the elements in a given view is\n$<1$. Starting from fundamental principles of human perception an unsupervised\nanomaly detection algorithm is developed that is simple, real-time and\nparameter-free. Experiments suggest it as a competing choice for univariate\ndata with promising results on the detection of global anomalies in\nmultivariate data.', 'Document date is essential for many important tasks, such as document\nretrieval, summarization, event detection, etc. While existing approaches for\nthese tasks assume accurate knowledge of the document date, this is not always\navailable, especially for arbitrary documents from the Web. Document Dating is\na challenging problem which requires inference over the temporal structure of\nthe document. Prior document dating systems have largely relied on handcrafted\nfeatures while ignoring such document internal structures. In this paper, we\npropose NeuralDater, a Graph Convolutional Network (GCN) based document dating\napproach which jointly exploits syntactic and temporal graph structures of\ndocument in a principled way. To the best of our knowledge, this is the first\napplication of deep learning for the problem of document dating. Through\nextensive experiments on real-world datasets, we find that NeuralDater\nsignificantly outperforms state-of-the-art baseline by 19% absolute (45%\nrelative) accuracy points.', 'Models for predicting the time of a future event are crucial for risk\nassessment, across a diverse range of applications. Existing time-to-event\n(survival) models have focused primarily on preserving pairwise ordering of\nestimated event times, or relative risk. Model calibration is relatively under\nexplored, despite its critical importance in time-to-event applications. We\npresent a survival function estimator for probabilistic predictions in\ntime-to-event models, based on a neural network model for draws from the\ndistribution of event times, without explicit assumptions on the form of the\ndistribution. This is done like in adversarial learning, but we achieve\nlearning without a discriminator or adversarial objective. The proposed\nestimator can be used in practice as a means of estimating and comparing\nconditional survival distributions, while accounting for the predictive\nuncertainty of probabilistic models. Extensive experiments show that the\nproposed model outperforms existing approaches, trained both with and without\nadversarial learning, in terms of both calibration and concentration of\ntime-to-event distributions.', 'Training models with discrete latent variables is challenging due to the\ndifficulty of estimating the gradients accurately. Much of the recent progress\nhas been achieved by taking advantage of continuous relaxations of the system,\nwhich are not always available or even possible. The Augment-REINFORCE-Merge\n(ARM) estimator provides an alternative that, instead of relaxation, uses\ncontinuous augmentation. Applying antithetic sampling over the augmenting\nvariables yields a relatively low-variance and unbiased estimator applicable to\nany model with binary latent variables. However, while antithetic sampling\nreduces variance, the augmentation process increases variance. We show that ARM\ncan be improved by analytically integrating out the randomness introduced by\nthe augmentation process, guaranteeing substantial variance reduction. Our\nestimator, DisARM, is simple to implement and has the same computational cost\nas ARM. We evaluate DisARM on several generative modeling benchmarks and show\nthat it consistently outperforms ARM and a strong independent sample baseline\nin terms of both variance and log-likelihood. Furthermore, we propose a local\nversion of DisARM designed for optimizing the multi-sample variational bound,\nand show that it outperforms VIMCO, the current state-of-the-art method.', 'Rapid increase of digitized document give birth to high demand of document\nimage retrieval. While conventional document image retrieval approaches depend\non complex OCR-based text recognition and text similarity detection, this paper\nproposes a new content-based approach, in which more attention is paid to\nfeatures extraction and fusion. In the proposed approach, multiple features of\ndocument images are extracted by different CNN models. After that, the\nextracted CNN features are reduced and fused into weighted average feature.\nFinally, the document images are ranked based on feature similarity to a\nprovided query image. Experimental procedure is performed on a group of\ndocument images that transformed from academic papers, which contain both\nEnglish and Chinese document, the results show that the proposed approach has\ngood ability to retrieve document images with similar text content, and the\nfusion of CNN features can effectively improve the retrieval accuracy.', 'Additive Manufacturing (AM) is a manufacturing paradigm that builds\nthree-dimensional objects from a computer-aided design model by successively\nadding material layer by layer. AM has become very popular in the past decade\ndue to its utility for fast prototyping such as 3D printing as well as\nmanufacturing functional parts with complex geometries using processes such as\nlaser metal deposition that would be difficult to create using traditional\nmachining. As the process for creating an intricate part for an expensive metal\nsuch as Titanium is prohibitive with respect to cost, computational models are\nused to simulate the behavior of AM processes before the experimental run.\nHowever, as the simulations are computationally costly and time-consuming for\npredicting multiscale multi-physics phenomena in AM, physics-informed\ndata-driven machine-learning systems for predicting the behavior of AM\nprocesses are immensely beneficial. Such models accelerate not only multiscale\nsimulation tools but also empower real-time control systems using in-situ data.\nIn this paper, we design and develop essential components of a scientific\nframework for developing a data-driven model-based real-time control system.\nFinite element methods are employed for solving time-dependent heat equations\nand developing the database. The proposed framework uses extremely randomized\ntrees - an ensemble of bagged decision trees as the regression algorithm\niteratively using temperatures of prior voxels and laser information as inputs\nto predict temperatures of subsequent voxels. The models achieve mean absolute\npercentage errors below 1% for predicting temperature profiles for AM\nprocesses. The code is made available for the research community at\nhttps://github.com/paularindam/ml-iter-additive.', 'This paper presents a new state-of-the-art for document image classification\nand retrieval, using features learned by deep convolutional neural networks\n(CNNs). In object and scene analysis, deep neural nets are capable of learning\na hierarchical chain of abstraction from pixel inputs to concise and\ndescriptive representations. The current work explores this capacity in the\nrealm of document analysis, and confirms that this representation strategy is\nsuperior to a variety of popular hand-crafted alternatives. Experiments also\nshow that (i) features extracted from CNNs are robust to compression, (ii) CNNs\ntrained on non-document images transfer well to document analysis tasks, and\n(iii) enforcing region-specific feature-learning is unnecessary given\nsufficient training data. This work also makes available a new labelled subset\nof the IIT-CDIP collection, containing 400,000 document images across 16\ncategories, useful for training new CNNs for document analysis.', 'My overarching research goal is to provide robots with perceptional abilities\nthat allow interactions with humans in a human-like manner. To develop these\nperceptional abilities, I believe that it is useful to study the principles of\nthe human visual system. I use these principles to develop new computer vision\nalgorithms and validate their effectiveness in intelligent robotic systems. I\nam enthusiastic about this approach as it offers the dual benefit of uncovering\nprinciples inherent in the human visual system, as well as applying these\nprinciples to its artificial counterpart. Fig. 1 contains a depiction of my\nresearch.', 'This paper describes the short-term competition on the Components\nSegmentation Task of Document Photos that was prepared in the context of the\n16th International Conference on Document Analysis and Recognition (ICDAR\n2021). This competition aims to bring together researchers working in the field\nof identification document image processing and provides them a suitable\nbenchmark to compare their techniques on the component segmentation task of\ndocument images. Three challenge tasks were proposed entailing different\nsegmentation assignments to be performed on a provided dataset. The collected\ndata are from several types of Brazilian ID documents, whose personal\ninformation was conveniently replaced. There were 16 participants whose results\nobtained for some or all the three tasks show different rates for the adopted\nmetrics, like Dice Similarity Coefficient ranging from 0.06 to 0.99. Different\nDeep Learning models were applied by the entrants with diverse strategies to\nachieve the best results in each of the tasks. Obtained results show that the\ncurrently applied methods for solving one of the proposed tasks (document\nboundary detection) are already well established. However, for the other two\nchallenge tasks (text zone and handwritten sign detection) research and\ndevelopment of more robust approaches are still required to achieve acceptable\nresults.', ""Civil engineers use numerical simulations of a building's responses to\nseismic forces to understand the nature of building failures, the limitations\nof building codes, and how to determine the latter to prevent the former. Such\nsimulations generate large ensembles of multivariate, multiattribute time\nseries. Comprehensive understanding of this data requires techniques that\nsupport the multivariate nature of the time series and can compare behaviors\nthat are both periodic and non-periodic across multiple time scales and\nmultiple time series themselves. In this paper, we present a novel technique to\nextract such patterns from time series generated from simulations of seismic\nresponses. The core of our approach is the use of topic modeling, where topics\ncorrespond to interpretable and discriminative features of the earthquakes. We\ntransform the raw time series data into a time series of topics, and use this\nvisual summary to compare temporal patterns in earthquakes, query earthquakes\nvia the topics across arbitrary time scales, and enable details on demand by\nlinking the topic visualization with the original earthquake data. We show,\nthrough a surrogate task and an expert study, that this technique allows\nanalysts to more easily identify recurring patterns in such time series. By\nintegrating this technique in a prototype system, we show how it enables novel\nforms of visual interaction.""]"